This page intentionally left blank

Optimality Theory

This is an introduction to Optimality Theory, the central idea of which

is that surface forms of language reect resolutions of conicts be-
tween competing constraints. A surface form is optimal if it incurs
the least serious violations of a set of constraints, taking into account
their hierarchical ranking. Languages differ in the ranking of con-
straints; and any violations must be minimal. The book does not limit
its empirical scope to phonological phenomena, but also contains
chapters on the learnability of OT grammars; OTs implications for
syntax; and other issues such as opacity. It also reviews in detail a
selection of the considerable research output which OT has already
produced. Exercises accompany chapters 17, and there are sections
on further reading. Optimality Theory will be welcomed by any lin-
guist with a basic knowledge of derivational Generative Phonology.

R EN KAGE R teaches linguistics at Utrecht University, the

Netherlands.
CAMBRIDGE TEXTBOOKS IN LINGUISTICS

General editors: s . r . a n d e r s o n , j . b r e s n a n , b . c o m r i e , w.
d r e s s l e r , c . e w e n, r . h u d d l e s t o n, r . l a s s , d. l i g h t f o o t,
j . ly o n s , p . h . m a t t h e w s , r . p o s n e r , s . r o m a i n e , n . v.
s m i t h , n. v i n c e n t

O P T I M A L I T Y T H E O RY
In this series

p . h . m a t t h e w s Morphology Second edition

b . c o m r i e Aspect
r . m . k e m p s o n Semantic Theory
t . b y n o n Historical Linguistics
j . a l l w o o d, l . - g . a n d e r s o n and . d a h l Logic in Linguistics
d . b . f r y The Physics of Speech
r . a . h u d s o n Sociolinguistics Second edition
a . j . e l l i o t t Child Language
p . h . m a t t h e w s Syntax
a . r a d f o r d Transformational Syntax
l . b a u e r English Word-Formation
s . c . l e v i n s o n Pragmatics
g . b r o w n and g . y u l e Discourse Analysis
r . h u d d l e s t o n Introduction to the Grammar of English
r . l a s s Phonology
b . c o m r i e Tense
w. k l e i n Second Language Acquisition
a . j . w o o d s , p . e t c h e r and a . h u g h e s Statistics in Language
Studies
d . a . c r u s e Lexical Semantics
f. r . p a l m e r Mood and Modality
a . r a d f o r d Transformational Grammar
m . g a r m a n Psycholinguistics
w. c r o f t Typology and Universals
g . g . c o r b e t t Gender
h . j . g i e g e r i c h English Phonology
r . c a n n Formal Semantics
p . j . h o p p e r and e . c . t r a u g o t t Grammaticalization
j . l av e r Principles of Phonetics
f. r . p a l m e r Grammatical Roles and Relations
b . b l a k e Case
m . a . j o n e s Foundations of French Syntax
a . r a d f o r d Syntactic Theory and the Structure of English: a Minimalist
Approach
r . d . va n va l i n , j r. and r . j . l a p o l l a Syntax: Structure, Meaning
Sand Function
a . d u r a n t i Linguistic Anthropology
a . c r u t t e n d e n Intonation Second edition
j . k . c h a m b e r s a n d p e t e r t r u d g i l l Dialectology Second edition
c . ly o n s Deniteness
r . k a g e r Optimality Theory
OPTIMALITY THEORY

ren kager

The Pitt Building, Trumpington Street, Cambridge, United Kingdom

The Edinburgh Building, Cambridge CB2 2RU, UK
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
Ruiz de Alarcn 13, 28014 Madrid, Spain
Dock House, The Waterfront, Cape Town 8001, South Africa

http://www.cambridge.org

Cambridge University Press 2004

First published in printed format 1999

ISBN 0-511-03977-8 eBook (netLibrary)

ISBN 0-521-58019-6 hardback
ISBN 0-521-58980-0 paperback
 CONTENTS

Preface page xi

1 Conicts in grammars 1
1.1 Introduction: goals of linguistic theory 1
1.2 Basic concepts of OT 3
1.3 Examples of constraint interaction 14
1.4 The architecture of an OT grammar 18
1.5 Interactions of markedness and faithfulness 27
1.6 Lexicon Optimization 32
1.7 A factorial typology of markedness and faithfulness 34
1.8 On dening segment inventories 43
1.9 Conclusion 47

2 The typology of structural changes 52

2.1 Introduction 52
2.2 Nasal substitution and related effects 59
2.3 The typology of *Nt effects 78
2.4 Conspiracies of nasal substitution and
other processes 83
2.5 Conclusion: a comparison with rule-based theory 86

3 Syllable structure and economy 91

3.1 Introduction 91
3.2 The basic syllable typology 92
3.3 Epenthesis and the conict of well-formedness and
faithfulness 98
3.4 Generalized Alignment 117
3.5 The quality of epenthetic segments 124
3.6 Coda conditions 130
3.7 Conclusion 139
vii
Contents

4 Metrical structure and parallelism 142

4.1 Introduction 142
4.2 Word stress: general background 143
4.3 Case-study: rhythmic lengthening in Hixkaryana 148
4.4 A set of metrical constraints 161
4.5 Case-study: rhythmic syncope in
Southeastern Tepehuan 177
4.6 Conclusions 189

5 Correspondence in reduplication 194

5.1 Introduction 194
5.2 Reduplicative identity: the constraints 201
5.3 From classical templates to generalized templates 216
5.4 From circumscription to alignment 223
5.5 Classical versus OT-based prosodic morphology:
conclusions 229
5.6 Overapplication and underapplication in reduplication 230
5.7 Summary of Correspondence Theory 248

6 Output-to-output correspondence 257

6.1 Introduction 257
6.2 Identity effects in truncation 259
6.3 Identity effects in stem-based afxation 273
6.4 The cycle versus base-identity 277
6.5 Output-to-output correspondence: conclusions 293

7 Learning OT grammars 296

7.1 Introduction 296
7.2 Learning constraint rankings 297
7.3 Learning the Pintupi grammar of stress 300
7.4 The learning algorithm: discussion 321
7.5 Learning alternations and input representations 324

8 Extensions to syntax 341

8.1 Introduction 341
8.2 OT and syntax 341
8.3 The structure of extended verbal projections
in English 353
8.4 Typological consequences 366
8.5 Conclusions 369
viii
 Contents

9 Residual issues 372

9.1 Introduction 372
9.2 Opacity 372
9.3 Absolute ungrammaticality 400
9.4 Free variation 404
9.5 Positional faithfulness 407
9.6 Underlying Representations versus allomorphy 413
9.7 Conclusion: future perspectives 420

References 425
Index of languages 445
Index of subjects 447
Index of constraints 451

ix
 PREFACE

This book presents an introduction to Optimality Theory, a grammatical framework

of recent origin (Prince and Smolensky 1993, McCarthy and Prince 1993a, b).
The central idea of Optimality Theory (OT) is that surface forms of language reect
resolutions of conicts between competing demands or constraints. A surface
form is optimal in the sense that it incurs the least serious violations of a set of
violable constraints, ranked in a language-specic hierarchy. Constraints are uni-
versal, and directly encode markedness statements and principles enforcing the
preservation of contrasts. Languages differ in the ranking of constraints, giving
priorities to some constraints over others. Such rankings are based on strict domina-
tion: if one constraint outranks another, the higher-ranked constraint has priority,
regardless of violations of the lower-ranked one. However, such violation must be
minimal, which predicts the economy property of grammatical processes. OTs
basic assumptions and the architecture of OT grammars will be dealt with in
chapters 1 and 2.
Optimality Theory is a development of Generative Grammar, a theory sharing
its focus on formal description and quest for universal principles, on the basis
of empirical research of linguistic typology and (rst) language acquisition.
However, OT radically differs from earlier generative models in various ways. To
accommodate cross-linguistic variation within a theory of Universal Grammar, OT
assumes that universal constraints are violable, while earlier models assumed
parametric variation of inviolate principles. Moreover, OT is surface-based in
the sense that well-formedness constraints evaluate surface forms only no struc-
tural conditions are placed on lexical forms. Earlier models had assumed Morph-
eme Structure Constraints, resulting in the duplication of static and dynamic rules
in phonotactics. In contrast, OT entirely abandons the notion of rewrite rule,
dissociating triggers and repairs. This serves to explain conspiracies: multiple
processes triggered by a single output-oriented goal. Finally, OT also eliminates
derivations, replacing these by parallelism: all constraints pertaining to some type
of structure are evaluated within a single hierarchy. The comparison of OT and

xi
Preface

its generative ancestors will be the topic of chapter 2, although the issue will
reoccur in later chapters (specically 4, 5, and 9).
Optimality Theory is not a theory of representations, but a theory of inter-
actions of grammatical principles. More accurately, the issue of representations is
orthogonal to that of constraint interaction. Therefore the divergence from earlier
generative models is less clear-cut in this respect. Most OT literature on phono-
logy, for example, assumes the representational alphabet of non-linear (metrical
and autosegmental) phonology. In this book, the emphasis will be on prosodic
phenomena, partly reecting a tendency in the eld, and partly the research inter-
ests of the author. Some of OTs most striking results have been reached in the
domain of prosodically governed phenomena, such as syllable-dependent epenth-
esis (chapter 3), interactions of syllable weight and metrical structure (chapter 4),
and prosodic targets in reduplication (chapter 5). However, our discussion of these
phenomena serves to highlight results of OT that are relevant beyond prosody. To
support this point, a range of segmental phenomena will be analysed throughout
the book. Finally, OT has consequences for representational issues which are more
closely connected with grammatical interactions, in particular for featural under-
specication, as will be shown in chapters 1, 3, and 9.
Optimality Theory is a general theory of grammar, rather than one of phono-
logy. Therefore this book is not limited in its empirical scope to phonological
phenomena, but it also contains chapters on the learnability of OT grammars
(chapter 7) and extensions to syntax (chapter 8). Finally, chapter 9 will address a
number of important residual issues in OT, focussing on opacity, and discussing
current developments in assumptions on lexical representations (versus allomor-
phy), optionality, absolute ungrammaticality, and various functionally oriented
approaches to phonology.
During its brief period of existence, OT has sparked off a large output of
articles, PhD dissertations, and volumes. Here we will review a selection of this
research output, in a way that maximally highlights the theorys contribution to
insights into language. In chapters 2 and 58, one particular piece of research
will be focussed on, while placing it against a broad theoretical background.
Chapter 2 focusses on the analysis of post-nasal-obstruent-voicing effects by
Pater (forthcoming), and serves to highlight factorial typology, OTs explanation
of conspiracies, and to introduce Correspondence Theory. Chapter 5 is devoted to
the Correspondence Theory of reduplication by McCarthy and Prince (1995a,
forthcoming), emphasizing the emergence of the unmarked and parallelism of
evaluation, and also extending the notion of correspondence to relations between
outputs. Chapter 6 discusses Benuas (1995) paper on output-to-output correspond-
ence in truncation, and its extensions to stem-based afxation, while comparing
OT and derivational theory for cyclic phenomena. Chapter 7 discusses work by
xii
 Preface

Tesar and Smolensky (1993, 1998) on the learnability of OT grammars, and its
dependence on basic OT notions, such as strict domination, minimal violation,
and assumptions on lexical forms. Chapter 8 is devoted to the analysis of Wh-
movement and its relation with auxiliary inversion and do-support in English by
Grimshaw (1997), pointing out the relevance of OT outside phonology.
This book is not a general introduction to phonology, and the reader should
come equipped with a basic knowledge of derivational Generative Phonology,
including rules and representations, and some knowledge of Minimalist Syntax
for chapter 8. Exercises have been added to chapters 17 to increase analytic
skills and reection on theoretical issues. Moreover, each chapter contains a list
of suggestions for further reading.
The idea for this book arose during a course I taught at the LOT summer school
at the University of Amsterdam in 1995. Stephen Anderson, who was present at
this course, suggested basing an OT textbook on its contents. For his role in
originating this book, I owe him special thanks.
Parts of this book are based on research reported on earlier occasions. Chapter
4 is partly based on Kager (1997a), rst presented at the workshop on Derivations
and Constraints in Phonology, held at the University of Essex, September 1995.
Chapter 6 contains results from Kager (forthcoming), presented at the conference
on the Derivational Residue in Phonology, Tilburg University, October 1995. I
wish to thank the organizers of these events: Iggy Roca, Ben Hermans, and Marc
van Oostendorp. Research for this book was partly sponsored by the Dutch Royal
Academy of Sciences (KNAW), whose support is gratefully acknowledged.
For their comments on earlier versions of chapters I wish to thank Peter Ackema,
Stephen Anderson, Roger Billerey, Gabriel Drachman, Nine Elenbaas, Bruce
Hayes, Claartje Levelt, Ad Neeleman, Joe Pater, Bruce Tesar, Wim Zonneveld,
and an anonymous reviewer. These comments have led to a number of substantial
improvements. Needless to say, I take the blame for any mistakes in content or
my presentation of other researchers ideas. Thanks to Martin Everaert for supply-
ing the child language data discussed in chapter 7.
Finally, this book would not have been nished without the encouragement and
patience of my colleagues, friends, and family. Jacqueline, this book is dedicated
to you.

xiii
 1
Conicts in grammars

1.1 Introduction: goals of linguistic theory

1.1.1 Universality
The central goal of linguistic theory is to shed light on the core of grammatical
principles that is common to all languages. Evidence for the assumption that there
should be such a core of principles comes from two domains: language typology
and language acquisition. Over the past decades our knowledge of linguistic
typology has become more and more detailed, due to extensive eldwork and
ne-grained analysis of data from languages of different families. From this large
body of research a broad picture emerges of unity in variety: core properties of
grammars (with respect to the subsystems of sounds, words, phrases, and mean-
ing) instantiate a set of universal properties. Grammars of individual languages
draw their basic options from this limited set, which many researchers identify
as Universal Grammar (UG). Each language thus reects, in a specic way, the
structure of language. A second source of evidence for universal grammatical
principles comes from the universally recurring patterns of rst language acqui-
sition. It is well known that children acquiring their rst language proceed in
remarkably similar ways, going through developmental stages that are (to a large
extent) independent of the language being learnt. By hypothesis, the innateness
of UG is what makes grammars so much alike in their basic designs, and what
causes the observed developmental similarities.
The approach to universality sketched above implies that linguistic theory
should narrow down the class of universally possible grammars by imposing
restrictions on the notions of possible grammatical process and possible inter-
action of processes. In early Generative Grammar (Chomsky 1965, Chomsky and
Halle 1968), processes took the shape of rewrite rules, while the major mode of
interaction was linear ordering. Rewrite rules take as their input a linguistic
representation, part of which is modied in the output. Rules apply one after
another, where one rules output is the next rules input. It was soon found that
this rule-based theory hardly imposes any limits on the notion of possible rule,
1
Conicts in grammars

nor on the notion of possible rule interaction. In the late 1970s and early 1980s,
considerable efforts were put into constraining both rule typology and interac-
tions. The broad idea was to factor out universal properties of rules in the form
of conditions.1 While rules themselves may differ between languages, they must
always respect a xed set of universal principles. Gradually more and more
properties were factored out of rules and attributed to universal conditions on
rules and representations. Developments came to their logical conclusion in
Principles-and-Parameters Theory (Chomsky 1981b, Hayes 1980), which has as
its central claim that grammars of individual languages are built on a central core
of xed universal properties ( principles), plus a specication of a limited num-
ber of universal binary choices ( parameters). Examples of parameters are the side
of the head (left or right) in syntactic phrases, or the obligatoriness (yes/no)
of an onset in a syllable. At the same time, considerable interest developed in
representations, as a way of constraining rule application, mainly with respect to
locality (examples are trace theory in syntax, and underspecication theory in
phonology). Much attention was also devoted to constraining rule interactions,
resulting in sophisticated theories of the architecture of UG (the T-model) and
its components (e.g. Lexical Phonology, Kiparsky 1982b).

1.1.2 Markedness
What all these efforts to constrain rules and rule interactions share, either implic-
itly or explicitly, is the assumption that universal principles can only be universal
if they are actually inviolate in every language. This interpretation of universal-
ity leads to a sharp increase in the abstractness of both linguistic representations
and rule interactions. When some universal principle is violated in the output of
the grammar, then the characteristic way of explaining this was to set up an
intermediate level of representation at which it is actually satised. Each gram-
matical principle thus holds at a specic level of description, and may be switched
off at other levels.
This absolute interpretation of universality is not the only one possible, how-
ever. In structuralist linguistics (Hjelmslev 1935, Trubetzkoy 1939, Jakobson
1941; cf. Anderson 1985), but also in Generative Phonology (Chomsky and Halle
1968, Kean 1975, Kiparsky 1985) and Natural Phonology (Stampe 1972, Hooper
1976), a notion of markedness plays a key role, which embodies universality in
a soft sense. The idea is that all types of linguistic structure have two values,
one of which is marked, the other unmarked. Unmarked values are cross-
linguistically preferred and basic in all grammars, while marked values are cross-
linguistically avoided and used by grammars only to create contrast. For example,
1
For example, Subjacency was proposed as a universal condition on syntactic movement rules
and the Obligatory Contour Principle as a universal condition on phonological rules.

2
 1.2 Basic concepts of OT

all languages have unrounded front vowels such as [i] and [e], but only a subset
of languages contrast these vowels with rounded front vowels such as [y] and [].
Hence, the unmarked value of the distinctive feature [round] is [round] in
front vowels. At a suprasegmental level, markedness affects prosodic categories.
For example, the unmarked value for syllable closure is open since all lan-
guages have open syllables (CV, V), while only a subset of languages allow closed
syllables (CVC, VC).2 The notion of markedness is not only relevant to sound
systems. Markedness principles have been proposed for morphological and
syntactic systems as well (Chomsky 1981a).
The markedness approach of linguistic universality is built on two assump-
tions. First, markedness is inherently a relative concept: that is, a marked linguis-
tic element is not ill-formed per se, but only in comparison to other linguistic
elements. Second, what is marked and unmarked for some structural distinc-
tion is not an arbitrary formal choice, but rooted in the articulatory and perceptual
systems. By this combination of two factors, markedness allows an interpretation
of universality that is fundamentally different from Principles-and-Parameters
Theory, in which markedness has no substantive status in the grammar, but func-
tions as an external system of annotations on parameter values, evaluating a
grammars complexity.3

1.2 Basic concepts of OT

Optimality Theory (Prince and Smolensky 1993, McCarthy and Prince
1993a,b) turns markedness statements into the actual substance of grammars.
Markedness is built into grammars in the form of universal output constraints
which directly state marked or unmarked patterns, for example: front vowels are
unrounded or syllables are open. The universal interpretation of markedness
constraints is reconciled with the observation that languages, to a certain extent
at least, tolerate marked types of structures. Universal markedness constraints can
be literally untrue for a grammars output, or to phrase it in optimality-theoretic
terms: constraints are violable. Violation of a constraint is not a direct cause of
ungrammaticality, nor is absolute satisfaction of all constraints essential to the
grammars outputs. Instead what determines the best output of a grammar is the
least costly violation of the constraints. Constraints are intrinsically in conict,
hence every logically possible output of any grammar will necessarily violate at
least some constraint. Grammars must be able to regulate conicts between uni-
versal constraints, in order to select the most harmonic or optimal output form.
2
Markedness may also involve scales. For example, the higher a consonants sonority value, the
more likely its occurrence in the syllable coda.
3
For the view of markedness as a criterion external to the grammar, evaluating its complexity,
see Chomsky and Halle (1968) and Kean (1975, 1981).

3
Conicts in grammars

This conict-regulating mechanism consists of a ranking of universal con-

straints. Languages basically differ in their ranking of constraints. Each violation
of a constraint is avoided; yet the violation of higher-ranked constraints is avoided
more forcefully than the violation of lower-ranked constraints. Accordingly, the
notion of grammatical well-formedness becomes a relative one, which is equiv-
alent to the degree of satisfaction of the constraint hierarchy, or harmony.
OTs viewpoint of UG is fundamentally different from that of classical rule-
based generative theory, where UG is dened as a set of inviolate principles and
rule schemata (or parameters). OT denes UG as a set of universal constraints
(markedness relations and other types of constraints, as we will see below), and
a basic alphabet of linguistic representational categories. In its interactions, it is
limited to a single device: constraint ranking. OT still shares with its rule-based
generative ancestors the central position taken by UG, as described above. OT is
a theory of the human language capacity.

The remainder of this chapter is organized as follows. Section 1.2 will introduce
basic notions of OT: conict, constraints, and domination, which will be exem-
plied in section 1.3. In section 1.4, we will discuss the architecture of an OT
grammar. Section 1.5 will deal with interactions of markedness and faithfulness,
relating these to the lexicon in section 1.6. A factorial typology of constraint
interactions will be developed in section 1.7 and applied to segment inventories
in section 1.8. Finally, section 1.9 presents conclusions.

1.2.1 Language as a system of conicting universal forces

At the heart of Optimality Theory lies the idea that language, and in fact every
grammar, is a system of conicting forces. These forces are embodied by con-
straints, each of which makes a requirement about some aspect of grammatical
output forms. Constraints are typically conicting, in the sense that to satisfy one
constraint implies the violation of another. Given the fact that no form can satisfy
all constraints simultaneously, there must be some mechanism selecting forms that
incur lesser constraint violations from others that incur more serious ones.
This selectional mechanism involves hierarchical ranking of constraints, such
that higher-ranked constraints have priority over lower-ranked ones. While con-
straints are universal, the rankings are not: differences in ranking are the source
of cross-linguistic variation.
But before discussing actual constraints and their rankings, let us rst nd out
in a general way about the two major forces embodied by constraints. Two forces
are engaged in a fundamental conict in every grammar. The rst is markedness,
which we use here as a general denominator for the grammatical factors that exert
pressure toward unmarked types of structure. This force is counterbalanced by
4
 1.2 Basic concepts of OT

faithfulness, understood here as the combined grammatical factors preserving

lexical contrasts. Let us focus on both general forces to nd out why they are
inherently conicting.

In sound systems, certain types of structure segments, segment combinations,

or prosodic structures are universally favoured over others. For example, front
unrounded vowels are unmarked as compared to front rounded vowels, open
syllables as compared to closed syllables, short vowels as compared to long
vowels, and voiceless obstruents compared to voiced obstruents. As was observed
above, marked structures are avoided by all languages, while they are completely
banned by some languages. Therefore the notion of markedness is inherently
asymmetrical.
Most phonologists agree that phonological markedness is ultimately
grounded in factors outside of the grammatical system proper. In particular,
the systems of articulation and perception naturally impose limitations on which
sounds (or sound sequences) should be favoured. Yet explaining markedness rela-
tions by phonetic factors does not amount to denying the basis of phonology as
a grammatical system, for two reasons. The rst reason is that phonetic factors
are gradient, and add up to numerical patterns, while phonological factors are
categorical, producing patterns whose boundaries are clearly cut by categorical
distinctions. The symmetry of phonological systems cannot be captured by the
interaction of raw phonetic factors. The second reason is that the relative
strength of the individual markedness factors varies from language to language,
which entails that there must be a language-specic system dening the balance
of factors. This is the grammar, a system of ranked constraints, of which phono-
logy is an integral part.

The major force counterbalancing markedness is faithfulness to lexical contrasts.

A grammar that is maximally faithful to a lexical contrast is one in which output
forms are completely congruent with their lexical inputs with respect to some
featural opposition. Or to put it differently, the total amount of lexically contrast-
ive variation of some feature is realized in all of the grammars output forms. For
example, a lexical contrast of voicing in obstruents is preserved in output forms
regardless of their phonological context (at the end of a word, between vowels,
etc.). Thus one may think of faithfulness as the general requirement for linguistic
forms to be realized as close as possible to their lexical basic forms. From a
functional angle, the importance of faithfulness is clear: to express contrasts of
meaning, any language needs a minimal amount of formal contrast. Formal con-
trasts should be preserved in realizations of lexical items, and not be eroded (or
at least, not too much) by factors reducing markedness. In the realm of sound
5
Conicts in grammars

systems (or phonologies), lexical contrasts are carried by oppositions between

sounds, as well as by their combinations. Phonological elements are not the only
carriers of lexical contrast. (Although phonology is what we will focus on in this
book.) Lexical contrasts are also expressible by word structure (morphology) or
phrase structure (syntax).
Closely related to faithfulness (or preservation of lexical contrasts) is the pres-
sure towards the shape invariability of lexically related items in various gram-
matical contexts. This was known in pre-generative linguistics as paradigm
uniformity. Shape invariance of lexical items is understandable as another prior-
ity of linguistic communication: there should be a one-to-one relation between
lexical items, the atoms of meaning, and the shapes which encode them.

1.2.2 Conicts between markedness and faithfulness

Markedness and faithfulness are inherently conicting. Whenever some lexical
contrast is being preserved, there will be some cost associated in terms of mark-
edness since in every opposition one member is marked. For example, consider
the fact that English limits the possible contrasts in its vowels with respect to the
dimensions of backness and rounding: no rounded front vowels stand in contrast
to unrounded front vowels. This correlation of rounding and backness in vowels
is not idiosyncratic to English, but it reoccurs in a great majority of the worlds
languages. In fact it is grounded in properties of the articulatory and perceptual
systems. Yet this restriction is certainly not universal in the sense that all of
the worlds languages respect it. Many languages do allow a contrast of rounding
in front vowels, thus increasing the potential amount of lexical contrast at the
expense of an increase in markedness.
Generally we nd that the larger the array of means of encoding lexical con-
trasts, the larger the complexity of the sound system, either in terms of segmental
complexity, or in terms of the combinatory possibilities between segments (pho-
notactics). A language can be maximally faithful to meaningful sound contrasts
only at the expense of an enormous increase in phonological markedness. Con-
versely, a language can decrease phonological markedness only at the expense of
giving up valuable means to express lexical contrast.

First consider what a hypothetical language would look like at one extreme of the
spectrum: a language giving maximal priority to the expression of lexical con-
trasts, while imposing no markedness restrictions. We endow this language with
the combined segment inventories of the worlds languages, roughly 50 conson-
ants and 30 vowels (Ladefoged and Maddieson 1996). We drop combinatory
markedness restrictions, allowing all logically possible segment combinations to
form a lexical item. Permutation of these 80 segments into lexical items of two
6
 1.2 Basic concepts of OT

segments already produces some 6,400 items, including [ph], [
x], and [.],
all highly marked. But why stop at two segments per item? By sheer lack of
phonotactic limitations, nothing rules out lexical items of 37 or 4,657 segments,
or even longer. Now consider the fact that the number of possible lexical items
increases exponentially with the number of segments (80n) so that at segmental
length 6 we already approximate an awesome 300 billion potential lexical items.
Clearly no human language requires this number of lexical contrasts, hence there
is room to impose markedness restrictions on segments and their combinations in
lexical items. Since such restrictions make sense from an articulatory and percep-
tual point of view, we expect to nd them.
Let us now turn the tables to nd out what a language at the other extreme
would look like, a language giving maximal priority to markedness, and minimal
priority to the expression of lexical contrasts. Let us assume that this language
limits its lexical items to the general shape of CV* (sequences of consonant
vowel), with C {p,t,k} and V {,a}.4 The complete set of potential monosylla-
bles contains 6 items {p, pa; t, ta; k, ka}, the set of disyllables contains 36 (or
62) items ({pp, pap, kp . . . }), trisyllables 216 (or 63), etc. But stop! We are
overlooking the fact that the unmarked length of lexical item is two syllables (this
is the minimum size in many languages and by far the most frequent size in most
languages). Since we are assuming that this language is maximally concerned
about markedness, we should limit word size to two syllables. The bitter conse-
quence is a mini-lexicon containing at most 36 items. Now consider the fact that
the lexicon of an average natural language contains some 100,000 items.5 It is
clear that giving maximal priority to markedness implies an acute shortage of
lexical contrasts, which no language can afford.
This comparison of two extremes shows that languages may, in principle at
least, go astray in either of two ways: by giving blind priority to expression of
lexical contrast, resulting in massive costs in terms of markedness or, at the other
end of the spectrum, by giving unlimited priority to markedness reduction, result-
ing in a fatal lack of contrast.
4
These limitations are actually grounded in speech production and perception: every consonant
is maximally different from a vowel (hence, all consonants are voiceless stops). Every vowel
is maximally different from other vowels (a 2-vowel set,  a). Every consonant is maximally
different from other consonants (place of articulation restricted to labial, alveolar, and velar).
Every vowel is preceded by a consonant (no word-initial vowels, no hiatus). Every consonant
precedes a vowel for optimal release (hence no consonant clusters nor word-nal Cs).
5
Suppose that our hypothetical language would not respect word size restrictions, having at its
disposition all possible CV*-shaped items. Here, with a maximal density of lexical contrast,
all potential items up to seven syllables long would not sufce to build the required size of
lexicon. This would only reach to a moderate total of (46,656 + 7,776 + 1296 + 216 + 36 +
6) = 55,986 lexical items. The average item in this language would be over six syllables long.
Without doubt, speaking would become a rather time-consuming activity.

7
Conicts in grammars

In sum, we have seen that every grammar must reconcile the inherently compet-
ing forces of faithfulness to lexical contrasts (the inertness which draws output
forms back to their basic lexical shapes) and markedness (minimization of
marked forms). However, as we are about to nd out, Optimality Theory recog-
nizes no unitary or monolithic forces of faithfulness or markedness: the picture
is more fragmented. In the grammars of individual languages, the overall conict
between both forces assumes the form of ner-grained interactions of individual
constraints. At this level, where individual constraints compete, languages are
quite diverse in their resolutions of conicts between markedness and faithful-
ness. A language may give priority to faithfulness over markedness with respect
to some opposition, but reverse its priorities for another opposition.
Let us now turn to the implementation of these basic ideas in Optimality
Theory.

1.2.3 The OT grammar as an inputoutput device

The basic assumption of OT is that each linguistic output form is optimal, in the
sense that it incurs the least serious violations of a set of conicting constraints.
For a given input, the grammar generates and then evaluates an innite set of
output candidates, from which it selects the optimal candidate, which is the actual
output. Evaluation takes place by a set of hierarchically ranked constraints (C1

C2
. . . Cn), each of which may eliminate some candidate outputs, until a point
is reached at which only one output candidate survives. This elimination process
is represented schematically:6

(1) Mapping of input to output in OT grammar

C1
C2
. . . Cn
Candidate a
Candidate b
Input Candidate c
Candidate d Output
Candidate . . .

The optimal output candidate is the one that is most harmonic with respect to
the set of ranked constraints. Harmony is a kind of relative well-formedness,
taking into account the severity of the violations of individual constraints, as
determined by their hierarchical ranking. That is, violation of a higher-ranked

6
Elimination of less-harmonic candidates is portrayed in (1) as a serial ltering process, but we
will learn to view it as a parallel process, with higher-ranked constraints taking priority over
lower-ranked constraints.

8
 1.2 Basic concepts of OT

constraint incurs a greater cost to harmony than violation of a lower-ranked con-

straint. Some violations must occur in every output candidate, as constraints
impose conicting requirements. Accordingly, a lower-ranked constraint can be
violated to avoid the violation of a higher-ranked one, but violation is always kept
to a minimum, given the requirement of maximal harmony.
With the basic assumptions of OT in our minds, let us now turn to a ner-
grained discussion of the core notions constraints, conict, domination, and
optimality.

1.2.4 Constraints: universality and violability

Our preliminary denition of constraint is: a structural requirement that may
be either satised or violated by an output form. A form satises a constraint if
it fully meets the structural requirement, while any form not meeting this require-
ment is said to violate it. For the moment we will assume no degrees of viola-
tion, so that output forms are simply categorized by a crude binary criterion as
either satisfying or violating a constraint. Forms may satisfy constraints vacu-
ously, which is the case if a constraint makes a requirement about some structural
element that is not present in a particular candidate.
OT recognizes two types of constraints, faithfulness constraints and markedness
constraints. Each individual constraint evaluates one specic aspect of output
markedness or faithfulness. Let us now look into the general properties of both
types of constraints, and into their functions in the grammar.

Markedness constraints require that output forms meet some criterion of struc-
tural well-formedness. As the examples below illustrate, such requirements may
take the form of prohibitions of marked phonological structures, including seg-
ment types (2a), prosodic structures (2b), or occurrences of segment types in
specic positions (2c).

(2) Examples of markedness constraints

a. Vowels must not be nasal
b. Syllables must not have codas
c. Obstruents must not be voiced in coda position
d. Sonorants must be voiced
e. Syllables must have onsets
f. Obstruents must be voiced after nasals

However, markedness constraints may just as well be stated positively, as in

(2df). Note that markedness constraints refer to output forms only and are blind
to the (lexical) input.
9
Conicts in grammars

As we have seen in section 1.1, markedness is an inherently asymmetrical

notion. Hence, the universal constraint inventory lacks the antagonist constraints
of (1ae), which make opposite requirements syllables must have codas, sonor-
ants must be voiceless, etc.7

Faithfulness constraints require that outputs preserve the properties of their basic
(lexical) forms, requiring some kind of similarity between the output and its
input.

(3) Examples of faithfulness constraints

a. The output must preserve all segments present in the input
b. The output must preserve the linear order of segments in the input
c. Output segments must have counterparts in the input
d. Output segments and input segments must share values for [voice]

Faithfulness constraints are, strictly speaking, not pure output constraints, since
they take into account elements at two levels: input and output. In contrast,
markedness constraints never take into account elements in the input.8 The
important thing is, however, that both kinds of constraints refer to the output
(exclusively so in markedness, and in relation to the input in faithfulness). OT has
no constraints that exclusively refer to the input. (This is a crucial difference from
classical generative phonology, as we will see in chapter 2.)
From a functional viewpoint, faithfulness constraints protect the lexical items
of a language against the eroding powers of markedness constraints, and thereby
serve two major communicative functions. First, they preserve lexical contrasts,
making it possible for languages to have sets of formally distinct lexical items to
express different meanings. Phrasing it slightly differently, with an emphasis on
contrast, we may say that faithfulness is what keeps the shapes of different lexical
items apart. Second, by limiting the distance between input and output, faithful-
ness constraints restrict the shape variability of lexical items. Faithfulness thus
keeps the contextual realizations of a single morpheme (called its alternants)
from drifting too far apart. This enhances the one-to-one relations of meaning and
form. In sum, the overall function of faithfulness is to enforce the phonological
shape of lexical forms in the output, as a sort of inertness limiting the distance
between outputs and their basic shapes.
Two more assumptions are to be made about constraints in OT: they are uni-
versal and violable requirements on some aspect of linguistic output forms. Let
us now focus on each of these properties of constraints. The rst property is

7
We will see later that some markedness constraints do have antagonists.
8
See chapter 9 for OT models which weaken this assumption.

10
 1.2 Basic concepts of OT

(4) Universality: constraints are universal.

In its strongest interpretation, by which all constraints are part of UG, this implies
that all constraints are part of the grammars of all natural languages. This is not
to say that every constraint will be equally active in all languages. Due to the
language-specic ranking of constraints, a constraint that is never violated in one
language may be violated but still be active in a second language, and be totally
inactive in yet a third language. This strong interpretation, which leaves no room
for language-specic constraints, nor for constraint variability, will be slightly
relativized below.
For phonological markedness constraints, universality may be established by a
variety of factors, ideally in combination. The rst sense of universality is typological:
a constraint states a preference for certain structures over other types of struc-
tures, which reoccurs in a range of unrelated languages. Segmental markedness
constraints, for example, may be validated by inspecting the relative markedness
of segments in inventories on a cross-linguistic basis. (Such an overview is pre-
sented in Maddieson 1984.) However, any exclusively typology-based deni-
tion of universality runs the risk of circularity: certain properties are posited as
unmarked simply because they occur in sound systems with greater frequency
than other marked properties.
Hence, a second (non-circular) criterion of universality should ideally accom-
pany typological criteria: phonological markedness constraints should be phonetic-
ally grounded in some property of articulation or perception. That is, phonetic
evidence from production or perception should support a cross-linguistic prefer-
ence for a segment (or feature value) to others in certain contexts. For example,
there is articulatory evidence (to be reviewed in chapter 2) that voiced obstruents
are preferred to voiceless obstruents in a position immediately following a nasal.
Indeed many languages avoid or disallow voiceless post-nasal obstruents, neu-
tralizing voicing contrasts in this position.9 Even though a growing number of
constraints has been phonetically grounded (see the suggested readings at the end
of this chapter), such grounding is still lacking for others.
It should be clear from this discussion that we should be very careful about
positing any constraint lacking both typological motivation and phonetic ground-
ing, even if there is compelling motivation for it from the language data under
analysis. Nevertheless, not all constraints that have been proposed in the OT lit-
erature satisfy both criteria, indicating that the major issue of universality of
constraints has not yet been resolved, since analysts do not share the same criteria.
In this book, whenever we employ a constraint that strikes us as parochial or

9
Post-nasal voicing and its typological consequences will be discussed in detail in chapter 2.

11
Conicts in grammars

language-specic (since it lacks both phonetic grounding and cross-linguistic

motivation), this will be indicated.
However, the universality of constraints should directly be relativized some-
what. We will nd that in special cases, language-specic elements may occur in
constraints of otherwise universal formats. This option is typical for a class of
constraints dening the interface of morphology and phonology, so-called align-
ment constraints, matching up the edges of specic morphemes and prosodic
categories. (See chapters 3 and 5.) Such interface constraints dene schemata in
which individual languages may substitute their specic morphemes.

We now move on to the second major property of OT constraints: their softness,

or violability. Violability of constraints must be understood in a specic way: the
general requirement is that it must be minimal:

(5) Violability: constraints are violable, but violation must be minimal.

No constraint is violated without a compelling reason: avoiding the violation of

another higher-ranked constraint. And even if a constraint is violated, violation
must be kept to a minimum. Everything else being equal, forms with lesser
violations are more harmonic than forms with greater violations. (Exactly how
degree of violation is determined will be taken up in section 1.4.3.) Violability of
constraints is an essential property of OT, representing a radical break away from
derivational models, as well as from constraint-based theories, such as Declarat-
ive Phonology (Bird 1990, Scobbie 1991), which assume that constraints are hard
or inviolate. (For a broad comparison with derivational theory, see chapter 2.)
This discussion of violable constraints gives rise to an important new question,
to which we now turn: what is the optimal candidate?

1.2.5 Optimality: domination and conict

As mentioned before, optimality is the status of being most harmonic with respect
to a set of conicting constraints. It is now time to take a closer look at the
concept of optimal in OT. The general idea is that the grammar evaluates an
innite set of candidate output forms, all analyses of a given input. From this
candidate set it selects the optimal output, the one which best matches the set
of conicting constraints. But what precisely does it mean for an output to be
optimal? Does it involve some sort of compromise between constraints of dif-
ferent strengths? Or is it perhaps the case that weaker constraints are rendered
inactive when they come into conict with stronger constraints?
In fact optimality involves neither compromise nor suppression of constraints,
but instead it is built on (strict) domination of constraints in a hierarchy.

12
 1.2 Basic concepts of OT

(6) Optimality: an output is optimal when it incurs the least serious

violations of a set of constraints, taking into account their hierarchi-
cal ranking.

So we assume that each output form of the grammar is by denition the best
possible in terms of the hierarchy of constraints, rather than the form which
matches all constraints at the same time. Perfect output forms are principally
non-existent, as every output form will violate at least some constraints. There-
fore the selection of the optimal output form involves setting priorities.
This is where a hierarchy comes into play. Conicts are resolved by domination:

(7) Domination: the higher-ranked of a pair of conicting constraints

takes precedence over the lower-ranked one.

This tentative denition will be rened below in section 1.4, on the basis of more
complex cases.
The ranking of constraints can be demonstrated by a tableau: this lists two
(or any number of ) output candidates vertically in random order, and constraints
horizontally, in a descending ranking from left to right. The cells contain violation
marks * incurred by each candidate for the constraint heading the column.
Schematically:

(8) A tableau for simple domination

C1 C2

a. + candidate a *
b. candidate b *!

The optimal candidate is marked by the index +. This candidate is (8a), which
has no violations of the higher-ranked constraint C1, a constraint violated by its
competitor (8b). Note that the optimal candidate (8a) is actually not impeccable
itself: it has a violation of C2, but this aw is insignicant to the outcome.
Although the pattern of violations for C2 is the reverse of that for C1, this does
not help candidate b. Its violation of C1 is already fatal, indicated by the accom-
panying exclamation mark ! and the shading of cells whose violation content is
no longer relevant. In sum, candidate (a) is optimal as no candidate is available
that fares better, satisfying both constraints at the same time. A violation of C2 is
taken for granted, as long as C1 can be satised.
We now turn to exemplication of the ideas that have been introduced thus far.

13
Conicts in grammars

1.3 Examples of constraint interaction

1.3.1 Neutralization of voicing contrast in Dutch

Among the universal and violable constraints is the following:

(9) *Voiced-Coda
Obstruents must not be voiced in coda position.

This is a typical markedness constraint, which bans a marked segment type (here:
voiced obstruents) from the syllable coda (which is itself a marked position).10
Coda obstruents are voiceless in Dutch, as illustrated by the following
alternation:

(10) a. /bd / bt bed

b. /bd-n/ bdn beds

Dutch has no voicing contrast in nal obstruents, neutralizing it towards voice-

lessness.
Next consider the evaluation of two candidate outputs for the input / bd /, [bt],
and [bd], with respect to *Voiced-Coda:

(11) Evaluation of two candidates with respect to *Voiced-Coda

a. [bt] satises *Voiced-Coda
(since [t] is an obstruent in a syllable coda, and [t] is
voiceless)
b. [bd] violates *Voiced-Coda
(since [d] is an obstruent in a syllable coda, and [d] is
voiced)

If this constraint were the only one relevant for these forms, then things would
be simple. Violators could be dismissed without second thoughts. But in actual
grammars things are not that simple since constraints may make conicting
requirements about output forms.
A second constraint of the universal inventory is a typical faithfulness con-
straint, requiring that the input value of the feature [voice] be preserved in the
output.

(12) Ident-IO(voice)
The specication for the feature [voice] of an input segment must be
preserved in its output correspondent.

10
Actually *Voiced-Coda can be interpreted as the conjunction of two markedness statements,
an idea to which we will return in chapter 9.

14
 1.3 Examples of constraint interaction

This faithfulness constraint mentions a notion correspondent, which is tentat-

ively (and very informally) dened as follows:

(13) Correspondent: the output segment that is the realization of an

input segment.

This informal denition is precise enough for our present purposes. (We will
return to the important notion of correspondence, particularly in chapters 2 and
5.)
In a correspondence diagram of the Dutch word [bt] bed the input and out-
put segments that are correspondents of one another are connected by vertical lines.

(14) Correspondence diagram of [bt]

/b d/ Input

[b t] Output

This diagram indicates that Ident-IO(voice) is violated in [bt]. Violation arises

since [t], a voiceless segment in the output, corresponds with a voiced segment /d/
in the input, and both segments have conicting values for voice. But at the same
time, [bt] satises the markedness constraint *Voiced-Coda, as [t] is a voiceless
obstruent in coda position.
We are, of course, looking at a simple conict between two constraints,
*Voiced-Coda and Ident-IO(voice). Both constraints make incompatible re-
quirements about the value of voice for any coda obstruent whose input is spe-
cied as [+voice]. An evaluation of both candidate outputs, [bd] and [bt], by
the conicting constraints is shown in (15):

(15) Evaluation of two candidate outputs for the input /bd /

a. [bd] satises Ident-IO(voice), but violates *Voiced-Coda
b. [bt] violates Ident-IO(voice), but satises *Voiced-Coda

Observe the conict: the evaluation of both output forms is different for each
constraint.
This conict requires resolution, which is the task of the constraint hierarchy.
The form [bt] emerges as the optimal output of the grammar, given the following
fragment of the phonology of the language:

(16) Coda devoicing in Dutch

*Voiced-Coda
Ident-IO(voice)

The symbol
connecting both constraints is to be read as dominates. Hence

we read (16) as follows: *Voiced-Coda dominates Ident-IO(voice). Domina-
tion ensures that the candidate outputs, [bd] and [bt], differ in their relative
15
Conicts in grammars

well-formedness with respect to the ranking in (16). Or stated differently, [bt] is

more harmonic than [bd] with respect to the ranking in (16).
(17) Harmonic ranking of two output candidates for the input / bd / in
Dutch
[bt]
[bed]
Since we are only considering two candidates here, the harmonic ranking directly
gives us the optimal output: [bt].
The correctness of this constraint ranking can be represented in a tableau-format:

(18) Tableau for the input /bd/, assuming the Dutch ranking
Candidates: *Voiced-Coda Ident-IO(voice)

a. + [bt] *
b. [bd] *!

The optimal candidate in the top row, [bt], incurs a violation of Ident-IO(voice)
while it satises *Voiced-Coda. Suboptimal [bd] has exactly the reverse pattern
of violations: it has a violation mark for *Voiced-Coda, but none for Ident-
IO(voice).
Being presented with these two output candidates, the grammar (whose only
goal is selecting an optimal output) must settle for a candidate that has a violation
of a lower-ranked constraint, simply because no perfect output candidate is avail-
able, satisfying both constraints. This point can be made more general: constraints
are intrinsically conicting, hence perfect output candidates will never occur in
any tableau:
(19) Fallacy of perfection: no output form is possible that satises all
constraints.
An output is optimal since there is no such thing as a perfect output: all that
grammars may accomplish is to select the most harmonic output, the one which
incurs the minimal violation of constraints, taking into account their ranking.
Nothing better is available.
Observe that the result of the constraint interaction in Dutch is a neutralization
of the voicing contrast in a specic context: the syllable coda. That neutralization
indeed takes place can be easily shown by the following set of examples:
(20) a.i /bd / bt bed
a.ii /bd-n/ b.dn beds
b.i /bt/ bt (I) dab
b.ii /bt-n/ b.tn (we) dab
16
 1.3 Examples of constraint interaction

Neutralization of the sound shapes of two lexical items is the ultimate conse-
quence of the domination of markedness over faithfulness. The lexical contrast
between /bd / and /bt/, residing in the value of voicing of their nal stem con-
sonants, might (in principle at least) have been preserved in all morphological
contexts in which they occur. But this is not the case, and a complete neutraliza-
tion occurs, into [bt].

1.3.2 Preservation of voicing contrast in English

In English, as opposed to Dutch, an analogous input /bd/ is mapped to an output
[bd], preserving the voicing in the nal consonant. Accordingly, English con-
trasts words such as bed and bet. This is due to the following fragment of the
phonology of this language:

(21) Preservation of voicing contrast in English

Ident-IO(voice)
*Voiced-Coda

In English, Ident-IO(voice) dominates *Voiced-Coda, which is the reverse

ranking of the one we established for Dutch. Accordingly, the harmonic ranking
of the output candidates under discussion is reversed, as compared to the one of
Dutch:

(22) Harmonic ranking of two output candidates for the input /bd / in
English
[bd]
[bt]

That is, assuming an input /bd/, [bd] is more harmonic than [bt] with respect
to the ranking in (21).
Again, we illustrate this ranking with the help of a tableau, evaluating the same
candidates as we used in tableau (18) for Dutch. Observe that Ident-IO(voice)
and *Voiced-Coda have changed places:

(23) Tableau for the input /bd/, assuming the English ranking
Candidates: Ident-IO(voice) *Voiced-Coda

a. [bt] *!
b. + [bd] *

The net result of this ranking is that the index pointing at the optimal output has
shifted downwards (as compared to tableau 18) to the second candidate under
consideration, that is, [bed]. Note that by this ranking, English preserves the
phonological contrast between distinct lexical items, as in bed [bd] versus bet
17
Conicts in grammars

[bt]. (This contrast is actually reinforced by a subsidiary vowel length difference

between both words: [b]d] versus [bt].)

1.3.3 The relation between universal and language-specic

What we have just witnessed in the examples from Dutch and English is the
universal pan-grammatical conict of markedness and faithfulness taking place
on a micro-scale. In both languages, the same conict arises with respect to
preservation of a contrastive property (the feature [voice]), and its neutralization
in a specic context (syllable coda). However, the outcome of this conict is
different for both languages. Dutch resolves it in the favour of markedness
whereas English favours faithfulness. This shows that universal constraints are
ranked in language-specic ways. OT clearly marks off the universal from the
language-specic. Both constraints and the general principles of their interaction
are universal, while constraint hierarchies are language-specic.
Speaking of forces of faithfulness and markedness is somewhat misleading,
since this suggests that conicts between these forces are resolved on a super-
ordinate level in the grammar of a single language. This is clearly not the case.
For example, the fact that Dutch ranks markedness above faithfulness with
respect to voice in coda obstruents does not imply that it selects the same ranking
(M
F) with respect to voice in other contexts, nor that it selects this ranking
with respect to other features in the syllable coda. In Dutch, voice is contrastive
in obstruents in onsets (even though voiced obstruents are universally marked).
Also, place features are contrastive in obstruents and nasals in codas (in spite of
the markedness of labials and velars). This shows that there are no monolithic
forces of faithfulness and markedness, but that instead ner-grained interactions
occur between the context- and feature-specic versions of these classes of con-
straints. Still, for expository purposes, the classication of constraints into faith-
fulness and markedness constraints remains useful, as are shorthand notations
such as M
F.
These remarks bring us back to our starting point in this section: the concep-
tion of universal grammar in OT. But what exactly do we mean by grammar in
the rst place? The OT grammar, and its architecture, will be the topic of the next
section.

1.4 The architecture of an OT grammar

The OT grammar is an inputoutput mechanism that pairs an output form to an
input form (such that each input has precisely one output). To accomplish this
function, the grammar contains a division of labour between a component
which maps the input onto an innite set of candidate output forms, and another
component that is burdened with evaluating the candidate output forms by a set
18
 1.4 The architecture of an OT grammar

of ranked constraints, and selecting the optimal output among these. These two
components are known under the names of Generator (or Gen) and Evaluator
(or Eval ). This grammatical organization is schematically represented in a func-
tion notation as follows:

(24) The grammar as an inputoutput mechanism

Gen (input) {cand1, cand2 . . . candn}
Eval {cand1, cand2 . . . candn} output

That is, Gen is a function that, when applied to some input, produces a set of
candidates, all of which are logically possible analyses of this input. Similarly,
Eval is a function that, when applied to a set of output candidates, produces an
output, the optimal analysis of the input. In addition to Gen and Eval, the gram-
mar contains a lexicon storing all lexical forms that are input to Gen. Recapit-
ulating, we nd the following model of the grammar:

(25) Components of the OT grammar

Lexicon: contains lexical representations (or underlying forms) of
morphemes, which form the input to:
Generator: generates output candidates for some input, and sub-
mits these to:
Evaluator: the set of ranked constraints, which evaluates out-
put candidates as to their harmonic values, and selects the optimal
candidate.

Let us now focus on some properties of the different components.

1.4.1 The LEXICON, and Richness of the Base

The Lexicon contains all contrastive properties of morphemes (roots, stems,
and afxes) of a language, including phonological, morphological, syntactic, and
semantic properties. The Lexicon provides the input specications which are to
be submitted to the Generator. In this connection, perhaps the most striking prop-
erty of the Lexicon, as conceived of in OT, is that no specic property can be
stated at the level of underlying representations:

(26) Richness of the Base: no constraints hold at the level of underlying

forms.

In OT grammatical generalizations are expressed as interactions of constraints

at the level of the output, never at the input level. Markedness constraints always
state requirements of output forms. Faithfulness constraints also evaluate output
forms, although they refer to the input level in stating their requirements. The
notion of contrast, which derivational theory locates at the level of the lexical
19
Conicts in grammars

representation, is attributed to interactions at the output level in OT. Whether or

not a feature is contrastive in some language depends on interactions of output-
oriented markedness and faithfulness constraints, either preserving or overruling
input specications (see section 1.5).
OT thus abandons Morpheme Structure Constraints (MSCs), which in classical
generative phonology (Chomsky and Halle 1968) account for prohibitions against
specic types of structure at the level of the morpheme, in specic languages.
MSCs were used, for example, to express prohibitions against front rounded vowels,
or sequences of three or more consonants, or two labial consonants occurring
within a morpheme. In the early 1970s MSCs were argued to be theoretically
problematic in the sense that they duplicate information which is, independently,
expressed by phonological rewrite rules, or that they globally guide the applica-
tion of rules, a property called structure-preservingness.11 By locating the burden
of explanation of the lack of specic kinds of structure at the level of the output,
OT, in principle at least, circumvents this Duplication Problem.

1.4.2 The GENERATOR , and Freedom of Analysis

The essential property of the Generator is that it is free to generate any conceiv-
able output candidate for some input. This property is called Freedom of Analysis.

(27) Freedom of Analysis: Any amount of structure may be posited.

The only true restriction imposed on all output candidates generated by Gen is
that these are made up of licit elements from the universal vocabularies of lin-
guistic representation, such as segmental structure (features and their grouping
below the level of the segment), prosodic structure (mora, syllable, foot, prosodic
word, etc.), morphology (root, stem, word, afx, etc.), and syntax (X-bar struc-
ture, heads/complements/speciers, etc.). Within these limits, anything goes.
Since Gen generates all logically possible candidate analyses of a given input,
the OT grammar needs no rewrite rules to map inputs onto outputs. All structural
changes are applied in one step, in parallel. The evaluation of these candidate
analyses is the function of the Evaluator, the component of ranked constraints,
discussed in section 1.4.3. There we will also discuss the issue of whether or not
Eval is able to deal with an innite candidate space.

1.4.3 The EVALUATOR: economy, strict domination, and parallelism

The Evaluator (henceforth Eval ) is undoubtedly the central component of
the grammar since it is burdened with the responsibility of accounting for all

11
For example, rewrite rules may be blocked if their output would violate a MSC, or may be
triggered to repair a violation of a MSC.

20
 1.4 The architecture of an OT grammar

observable regularities of surface forms. Although any candidate output can be

posited by Gen, the crucial role of Eval is to assess the harmony of outputs with
respect to a given ranking of constraints.
Eval is structured as a (language-specic) hierarchy of universal constraints,
plus devices for evaluation. The latter include the means to assess violation marks
on candidate outputs for every constraint, and the means to rank an innite set
of candidate outputs for harmony with respect to the hierarchy of constraints,
and select the most harmonic one of these as optimal the actual output of the
grammar. Let us now take a closer look at each of these devices: the constraint
hierarchy, marking of violations, and harmony evaluation.

First, the constraint hierarchy contains all universal constraints (a set called Con),
which are ranked in a language-specic way. We (tentatively) assume that all
constraints are ranked with respect to each other, so as to exclude variable and
undetermined rankings. (For cases in which two constraints cannot be ranked with
respect to each other, due to a trivial lack of interaction, we nevertheless assume
some ranking, arbitrarily one or the other.)
Moreover, within the hierarchy, dominance relations are transitive:

(28) Transitivity of ranking: If C1
C2 and C2
C3 then C1
C3

This property of ranking will allow us to construct ranking arguments, as we will

see below.

Second, with respect to violation marks, we assume that each output candidate is
provided with as many marks as it has violations for a constraint. This number of
marks potentially ranges from zero until innite. However, for purposes of deter-
mining optimal outputs, an innite number of marks is never practically relevant.
The essence of minimal violation of constraints is that every violation of a con-
straint serves a purpose: to avoid a violation of some higher-ranked constraint.
This is a property which is stated by Prince and Smolensky (1993: 27):

(29) Economy: banned options are available only to avoid violations of

higher-ranked constraints and can only be banned minimally.

For example, the Generator component is free to submit any kind of analysis of
(English) /bd/ that is couched within the universal alphabet of representational
options, including excessively unfaithful candidates such as [plow] and [mtrs].
But these candidates will be (hopefully!) ruled out regardless of constraint rank-
ing, since they violate faithfulness constraints without compensation from reduc-
tions in markedness. This economy property of OT will be discussed in more
detail in section 1.7.5.
21
Conicts in grammars

Third, we have not yet precisely formulated in which way the evaluation of output
candidates by ranked constraints proceeds. Eval determines the harmonic status
of output candidates, and eventually the most harmonic or optimal candidate. To
this end, it uses a process by which the set of candidates is reduced until the point
is reached at which one output remains. This is a multi-step process, schematic-
ally repeated below from (1):

(30)
C1
C2
. . . Cn
Candidate a
Candidate b
Input Candidate c
Candidate d Output
Candidate . . .

The major property of this evaluation process is that it applies from one state to
another without looking ahead to following steps. That is, the elimination of
candidate outputs by a constraint Cn is never affected by a lower-ranked constraint
Cn+m. Stated in a non-serial manner, this implies:

(31) Strict domination: violation of higher-ranked constraints cannot be

compensated for by satisfaction of lower-ranked constraints.

Optimality does not involve any kind of compromise between constraints of dif-
ferent ranks.
To illustrate strict domination, let us return to tableau (8) the cases of simple
domination and ask what would have happened if the rst candidate had had
not one but two violations of C2. The following tableau shows that even here, the
rst candidate would still be optimal, even though its total number of violations
is greater:
(32) Strict domination: multiple violations of a lower-ranked constraint
C1 C2

a. + candidate a **
b. candidate b *!

No smaller amount of violations can compensate for ranking of constraints. Dom-

ination is strict: any candidate that incurs a violation of some higher-ranked
constraint (on which another candidate incurs no violations) is mercilessly ex-
cluded, regardless of its relative well-formedness with respect to any lower-ranked
constraints.
22
 1.4 The architecture of an OT grammar

There is yet another sense in which domination is strict, which is not illustrated
by (32) constraint violations are never added for different constraints. The
added violations of two lower-ranked constraints (C2 and C3) are not able to
cancel out a single violation of a higher-ranked constraint (C1):

(33) Strict domination: violations of multiple lower-ranked constraints

C1 C2 C3

a. + candidate a * *
b. candidate b *!

That is, lower-ranked constraints cannot team up against a higher-ranked

constraint.
We see that there is no element of compromise in the notion optimal: evalu-
ation of candidates by the set of constraints is based on strict domination, and
accordingly, satisfaction of higher-ranked constraints has uncompromised priority
over satisfaction of lower-ranked ones. Uncompromised, since no possible degree
of satisfaction of lower-ranked constraints can compensate for the violation of a
single high-ranked constraint.
Not all interactions of constraints are of this relatively simple kind, where an
optimal candidate satises a high-ranked constraint that is violated by all com-
petitors. Actually most interactions involve some degree of violation in the
optimal candidate. How can this occur? Violation of a constraint is, by itself,
an insufcient ground for ungrammaticality. Recall that the goal of evaluation
is to single out one unique form as the most harmonic one. Elimination of all
candidates in the set under consideration is therefore not allowed. This is shown
in diagram below, where C1 functions as a no-pass lter:

(34)
C1
C2
Candidate a
Candidate b
Candidate c
Candidate d
Candidate . . .

This must be avoided. Hence for a violation of some constraint C1 to be fatal

(eliminating from the candidate set any forms which incur it) at least one other
form must occur in the candidate set that satises C1 (without being less harmonic
on higher-ranked constraints, of course). If no such form can be found, some
violation must be taken for granted.
23
Conicts in grammars

In such a situation, in which all remaining candidate outputs violate a con-

straint (due to higher-ranked constraints), the seriousness of violation must be
taken into account for each individual form. That is, forms with fewer violation
marks of C1 are preferred to forms with more violation marks for C1. This situ-
ation may still produce a ranking argument for C1 and C2, as tableau (35) shows:
(35) Amount of violation decisive
C1 C2

a. + candidate a * *
b. candidate b **!

Finally, if multiple candidates have the same number of violations for C1 (and this
equals the minimal violation in the set), then all survive and are passed on for
evaluation by the next constraint down the hierarchy, C2.12
(36) Tie between candidates (with lower-ranking constraint decisive)
C1 C2

a. + candidate a * *!
b. candidate b *

This situation can be represented as an all-pass lter C1:

(37)
C1
C2
Candidate a
Candidate b
Input Candidate c
Candidate d
Candidate . . .

Of course, ties between candidates may also arise between forms that have
no violations at all, or between forms that have two, three, or any number of
violations.

Finally, we emphasize that lower-ranked constraints are not rendered inact-

ive, or switched off by higher-ranked constraints, but that their violation is
only avoided with less priority. Lower-ranked constraints may be violated by the
12
Note that in the case of a tie, the ranking of constraints C1 and C2 becomes indeterminable
from the actual form however, we assume that this ranking may be established from other
forms.

24
 1.4 The architecture of an OT grammar

optimal output, but their violation must be minimal. Given the chance, any con-
straint (regardless of its position in the hierarchy) will be active in determining
the optimal output.

(38) Activity of a dominated constraint

C1 C2 C3

a. + candidate a * *
b. candidate b **!
c. candidate c *!

That C2 is dominated is apparent from the fact that candidate (38c) is less har-
monic than (38a), even though it has no violations of C2 . But C2 is still active,
since it dominates C3 .

The nal property of Eval to be briey discussed here is:

(39) Parallelism: all constraints pertaining to some type of structure
interact in a single hierarchy.
In a trivial sense, it is parallelism which predicts that faithfulness constraints
may interact with markedness constraints in a single hierarchy. But at a higher
level of sophistication, parallelism is also the basis of explanation of phenomena
involving interface properties. In particular, we will see many examples in this
book showing that morphological and phonological properties of an output form
are mutually dependent. The most spectacular cases will come from the area of
prosodic morphology, that is, types of morphology that depend on aspects of
syllabication and metrical structure (examples being reduplication, inxation,
and truncation). It is parallelism that makes information ow back and forth
between morphological and prosodic aspects in such cases. Striking conse-
quences of parallelism will be discussed in later chapters of this book, in par-
ticular in chapter 4 (on interactions of quantity and stress) and chapter 5 (on
reduplication).

1.4.4 Fear of innity

Freedom of Analysis may seem to pose an overwhelming computational problem
for the basic function of a grammar, which is to provide a mapping between input
and output. Perhaps the most apparent fear is that an innite candidate space is
computationally intractable. Reactions to this point focus on the nature of candid-
ate space, on evaluation strategies which assure a more efcient processing, and
on computational results booked so far in modelling OT. For an elaboration of the
25
Conicts in grammars

arguments below, and for some others, see chapter 10 of Prince and Smolensky
(1993).
Firstly, it is a well-accepted assumption among linguists that there is a distinc-
tion between the grammar (competence) and its cognitive implementation (per-
formance). This distinction is assumed in most formal theories of grammar, and
particularly in generative linguistics (Chomsky 1965). Therefore a model of
grammar is adequate to the extent that it explains observed systematicities in
natural languages, and the grammatical judgements of speakers. Explaining the
actual processing of linguistic knowledge by the human mind is not the goal of the
formal theory of grammar, but that of linguistic disciplines (such as psycholin-
guistics, neurolinguistics, and computional linguistics). The central point is that a
grammatical model should not be equated with its computational implementation.
Secondly, turning now to computational plausibility, the fact that candidate
space is innite does not imply that the problem is logically unsolvable. You may
convince yourself of this by thinking of arithmetic or any kind of numerical
problem. For example, there is a unique solution to the equation 3n2 3 = 45,
which you will be able to nd after a moments thought, even though the can-
didate set (let us say, all integers) is innite. From a computational viewpoint, the
decisive factor is that a guaranteed method (an algorithm) exists that will cer-
tainly produce a solution for any input. Therefore, no simple argument against OT
as being computationally intractable can be based on the observation that can-
didate space is innite.13
Thirdly, smart computational strategies may eliminate suboptimal candidates
by classes, rather than on a one-by-one basis. As soon as a candidate has been
excluded due to its violation of some constraint C, the evaluation process can
immediately eliminate all other candidates that violate this constraint C more
severely. This leads us to yet another property of candidate space that might be
put to use in computational evaluation models. By far the great majority of
candidates proposed by Gen can never be selected as optimal, under any possible
ranking of constraints. Such intrinsically suboptimal candidates can be readily
identied as follows: they share with another candidate (of the same input) some
set of violation marks, but have at least one additional violation of some other
constraint (an example will be discussed in section 1.7.5). Sophisticated evalua-
tion strategies may capitalize on this. Since the identication of intrinsically sub-
optimal candidates involves no ranked constraints, innite candidate space may
be drastically reduced by eliminating the worst-of-the-worst of candidates by pre-
processing prior to the evaluation by ranked constraints. Since this preprocessing

13
Conversely, a nite set of candidates does not guarantee that a problem is logically solvable.
This argument is due to Alan Prince (presentation at Utrecht University, January 1994).

26
 1.5 Interactions of markedness and faithfulness

would eliminate the great majority of candidates, the ultimately relevant remain-
ing part of candidate space may well have quite manageable proportions, and
perhaps even reduce to a nite set (Hammond 1997).
Whether or not a computational method can be established for the evaluation
of an innite candidate space in OT grammars is still largely an open issue, but
encouraging results are available. For example, Karttunen (1998) demonstrates
that the computation of the most optimal surface realization of any input string
can be carried out entirely within a nite-state calculus, subject to the limitation
(Frank and Satta (1998)) that the maximal number of violations is bounded.
Karttunen adds that [i]t is not likely that this limitation is a serious obstacle to
practical optimality computations with nite-state systems as the number of con-
straint violations that need to be taken into account is generally small.

1.5 Interactions of markedness and faithfulness

This section will deal with important types of interactions of markedness and
faithfulness constraints, building on key insights of Prince and Smolensky (1993),
Kirchner (1995), and Steriade (1995b). In section 1.3.3 we studied nal devoicing
in Dutch as a case of positional neutralization: the feature [voice] is neutralized
in a specic context (the syllable coda), as a result of a markedness constraint
dominating a faithfulness constraint. Here we will rst extend this simple inter-
action of markedness and faithfulness to a new phenomenon: allophonic varia-
tion. In section 1.6 we will look into the notion of contrast as it is dened in OT,
and its consequences for lexical representation. This will allow us to dene more
clearly the ranking schemata of faithfulness and markedness constraints that are
responsible for the various attested situations (contrast, neutralization, and
allophonic variation). In section 1.7 we will summarize these results in the form
of a factorial typology. In section 1.8 we will see how segment inventories
follow from interactions of faithfulness and markedness.

1.5.1 Allophonic variation

Consider a language that has no lexical contrast of oral and nasal vowels. In this
language oral and nasal vowels are allophones, variants of one another which are
fully predictable from the phonological contexts. For example, vowels are gener-
ally oral except when they directly precede a tautosyllabic nasal stop, in which
case they are nasal. This allophonic pattern occurs in many dialects of English;
see the examples below:
(40) a.i cat [kt] b.i cant [kNnt]
a.ii sad [sd] b.ii sand [sNnd]
a.iii met [mt] b.iii meant [mknt]
a.iv lick [lk] b.iv link [lPk]
27
Conicts in grammars

When we say that English lacks a contrast of oral and nasal vowels, we do not
imply that English completely lacks either kind of vowels, but only that no word
pairs occur that are distinguished by orality/nasality of their vowels. Whatever
variation there is between oral and nasal vowels is totally conditioned by the
context and does not reect lexical specication. Vowels are nasal when they
precede a tautosyllabic nasal, and are oral in all other contexts. This complement-
ary distribution, and the corresponding lack of word pairs that differ only in the
specication of some feature, is what denes an allophonic pattern. How can the
allophonic pattern in (40) be stated in terms of violable constraints?
In order to answer this question, we must rst identify the set of constraints
which are involved. Universally, nasal vowels are marked as compared to oral
vowels. Most of the worlds languages completely lack nasal vowels, having oral
vowels only (Maddieson 1984). Languages may have both oral and nasal vowels,
but no languages have only nasal vowels. In sum, when a language has nasal
vowels, it must also have oral vowels. The marked status of nasal vowels is
expressed by the context-free markedness constraint in (41), which militates
against nasal vowels:

(41) *VNASAL
Vowels must not be nasal.

When this constraint is undominated in some language, then all of its vowels will
be oral, regardless of their lexical specication, or their position in the syllable
(before an oral or nasal).
Moreover, many languages tend to nasalize vowels in precisely the position
where they are nasal in English: before a tautosyllabic nasal stop. The vowel thus
anticipates the nasality of the following stop, a preferred state of affairs from the
viewpoint of perception and articulation (Cohn 1993a).14 Again, a markedness
constraint expresses the universal markedness, ruling out oral vowels that precede
a tautosyllabic nasal:

(42) *VORALN
Before a tautosyllabic nasal, vowels must not be oral.

Observe that this constraint is context-sensitive, since it states a connection

between the nasality of a vowel and a nasal stop in its context. More precisely,
it is violated by an oral vowel that stands directly before a tautosyllabic nasal:

14
Cohn (1993a) argues that nasalization in English vowels is gradient, and has no phonological
status, as in French. For the sake of the argument, we will assume here that English nasaliza-
tion is in fact categorical, although it is crucially non-contrastive.

28
 1.5 Interactions of markedness and faithfulness

(43) a. *VORALN satised b. *VORALN violated

i. Nn] n]
ii. d]

If this constraint is undominated, underlying contrasts between oral and nasal

vowels (if any) will be neutralized in positions before a tautosyllabic nasal.

1.5.2 Neutralization and contrast as constraint rankings

Now consider the consequences of the OT assumption of the Richness of the
Base, which was stated in section 1.4.1. This says that no constraints restrict the
input, or to put it differently, that lexical representations in any language are free
to contain any kind of phonological contrast. Whether some surface phonetic
contrast (such as that between oral and nasal vowels) is allophonic or lexically
distinctive in a language depends on interactions of two basic kinds of constraints:
markedness constraints, which express markedness statements, and faithfulness
constraints, which penalize deviations of the surface form (output) from its lexical
form (input). When markedness dominates faithfulness, the language achieves
outputs that are minimally marked, at the expense of a neutralization of lexical
contrasts. But when faithfulness dominates markedness, the language makes the
reverse choice, realizing its input contrasts at the expense of output markedness:

(44) a. Markedness
Faithfulness lexical contrasts are neutralized

b. Faithfulness
Markedness lexical contrasts are expressed

Richness of the Base implies that English (as any other language) is allowed the
option of setting up a contrast of oral and nasal vowels in its underlying repres-
entations. However, this hypothetical contrast is never realized at the surface,
because with respect to nasality/orality in vowels, English happens to be a lan-
guage of the type (44a), which gives priority to markedness over faithfulness.
Whatever lexical contrast of nasality there might be in vowels will be obscured
by effects of markedness. The input faithfulness constraint that is crucially dom-
inated in English requires that surface values of nasality in vowels are identical
to their underlying values:

(45) Ident-IO(nasal)
Correspondent segments in input and output have identical values for
[nasal].

In a language in which Ident-IO(nasal) is undominated, any lexical contrast of

nasality in vowels will be allowed to surface, uninhibited by the markedness con-
straints (412). Such a language is free to set up and preserve any lexical contrast
between oral and nasal vowels anywhere, that is, without any neutralization. This
29
Conicts in grammars

situation corresponds to the interaction (44b). But in a language in which Ident-

IO(nasal) is dominated by both of the markedness constraints (41) and (42), any
(potential) contrast of orality/nasality in vowels will be fully neutralized, as is the
case in allophonic variation. This is the situation (44a), found in English.

Let us now return to the allophonic pattern (40) and nd out how this results from
the interaction of the three constraints that were introduced earlier. In terms of
constraint interaction, faithfulness to the lexical specication of a vowel is com-
pletely dominated by markedness constraints reecting markedness of orality/
nasality in vowels. In terms of ranking, Ident-IO(nasal) is dominated by both
markedness constraints:

(46) Neutralization of lexical contrast

Markedness
Faithfulness
*VNASAL, *VORALN
Ident-IO(nasal)

This is an instantiation of the schema in (44a), where markedness constraints

completely dominate faithfulness.

The question which arises next is how both markedness constraints, *VORALN and
*VNASAL, are ranked with respect to each other. As we observed earlier in connec-
tion with the context-free constraint *VNASAL, any language in which this is undom-
inated will totally lack nasal vowels in its surface patterns. This is not the case in
English, however, where nasal vowels do occur (as allophones of oral vowels) in
specic positions, that is, before tautosyllabic nasal stops. We must therefore
rene the ranking in (46) to that in (47):

(47) Allophonic variation

Contextual markedness
Context-free markedness
Faithfulness
*VORALN
*VNASAL
Ident-
IO(nasal)

This ranking states that nasal realization of vowels before tautosyllabic nasal
consonants takes priority over a total lack of nasality in vowels. In sum, both
nasal and oral vowels occur at the surface, but their distribution is xed, rather
than free.

This ranking is illustrated in the tableaux (4851). First consider the case of an
oral vowel in the actual output, for example sad [sd]. When we assume that this
has an oral vowel in its lexical representation, e.g. /sd/, matching its surface
status, we arrive at the rst tableau (48). Candidate (48a) is optimal as it violates
30
 1.5 Interactions of markedness and faithfulness

none of the constraints in the tableau, regardless of ranking. It satises *VORALN

since this constraint has nothing to say about vowels that stand before oral stops.
It also satises *VNASAL since it has no nasal vowel. Finally it satises Ident-
IO(nasal) because the input and output agree in nasality.

(48) Input: /sd/ *VORALN *VNASAL Ident-IO(nasal)

a. + [sd]
b. [sNd] *! *

The losing candidate [sNd] (48b) is less harmonic than (48a) in two respects. It
contains a nasal vowel, fatally violating the markedness constraint *VNASAL. It
violates Ident-IO(nasal) as well, as the nasal vowel in the output fails to match
its oral correspondent in the input.
Because of Richness of the Base, we must guarantee that this correct result is
not negatively affected when we make different assumptions about the nasality of
vowels in the input. Indeed, the same candidate [sd] is selected when the input
would contain a nasal vowel, e.g. /sNd/, here in deance of its surface form. This
is shown in tableau (49). Again markedness uniquely determines the outcome,
without interference on the part of the faithfulness constraint Ident-IO(nasal).

(49) Input: /sNd/ *VORALN *VNASAL Ident-IO(nasal)

a. + [sd] *
b. [sNd] *!

Note that in this case Ident-IO(nasal) is violated in the optimal candidate. This
motivates the ranking *VNASAL
Ident-IO(nasal), a markedness constraint dom-
inating faithfulness. That is, even if the input of sad were to contain a nasal vowel,
its nasality would be wiped out in the surface form by markedness constraints.
This is of course the central result that we need to account for allophonic varia-
tion, in a theory which assumes Richness of the Base.
We can only rightfully claim to have captured the complementary distribution
of oral and nasal vowels if we can prove the total irrelevance of the input for
words which surface with nasal vowels, for example sand [sNnd]. Again we
consider two underlying forms, one with an oral vowel and the other with a nasal
vowel. Tableau (50) shows that an underlying form with an oral vowel /snd/
results in an optimal output with a nasal vowel, [sNnd]. This is due to the un-
dominated context-sensitive markedness constraint *VORALN, which requires that
vowels are nasal before a tautosyllabic nasal stop:
31
Conicts in grammars

(50) Input: /snd/ *VORALN *VNASAL Ident-IO(nasal)

a. [snd] *!
b. + [sNnd] * *

Observe that two markedness constraints, *VORALN and *VNASAL, are in conict
here. The former requires a nasal vowel in the output whereas the latter militates
against it. The fact that the actual output [sNnd] has a nasal vowel shows that
*VORALN dominates *VNASAL. (If the ranking had been reverse, the result would
have been in favour of candidate 50a, which has an oral vowel.) Observe also that
the underlying orality of the vowel in sand does not affect the outcome. Ident-
IO(nasal) is violated in the optimal output, since it contains a nasal vowel whereas
the input contains an oral vowel. This conclusion is essential to the argument
that faithfulness is dominated by both markedness constraints. We have already
reached this conclusion for *VNASAL in tableau (49), and now we conrm it for
*VORALN.

The argument for the irrelevance of inputs in allophonic patterns is completed by

an inspection of tableau (51), which has an underlying form with a nasal vowel,
/sNnd/. In this tableau, the same optimal candidate is selected as in the previous
one, simply because markedness uniquely determines the outcome.

(51) Input: /sNnd/ *VORALN *VNASAL Ident-IO(nasal)

a. [snd] *! *
b. + [sNnd] *

A comparison of tableaux (50) and (51) reveals the complete inactivity of the
faithfulness constraint Ident-IO(nasal). We conclude that the orality/nasality of
the underlying vowel is completely irrelevant to the surface distribution of oral
and nasal vowels.

1.6 Lexicon Optimization

The main result of the preceding section is that lexical specications for [nasal]
in vowels in English are totally irrelevant to their surface realization. Should we
then conclude that the English lexicon is completely unstructured for nasality in
vowels, in the sense that the vowels in lexical items sad and sand are randomly
specied for this feature? Or should we still insist that the lexicon be kept clean
from featural noise, and contain only feature values that are actually related to
32
 1.6 Lexicon Optimization

output values? An answer to this question is potentially relevant to language

acquisition. In order to build a lexicon, the learner must somehow be able to
determine underlying forms, for example to infer the underlying form of sad on
the basis of its surface form [sd]. When we concentrate on possible values for
nasality in the vowel, there is a choice of two lexical representations, /sd/ and
/sNd/. We have seen that, given the ranking of *VNASAL over Ident-IO(nasal), both
lexical representations result in identical outputs. This ranking completely masks
the input, obscuring empirical evidence that the learner might use to base his/her
choice of an underlying form on.
It has been proposed that in the absence of empirical evidence for one input
form over another, the input should be selected that is closest to the output, in
this case /sd/. That is, wherever the learner has no evidence (from surface
forms) to postulate a specic diverging lexical form, (s)he will assume that the
input is identical to the surface form. In terms of constraint violations, this strat-
egy has the advantage of minimizing the violation of faithfulness, as compared
to any other hypothetical inputs producing the same output. This strategy is called
Lexicon Optimization in Prince and Smolensky (1993: 192):

(52) Lexicon Optimization: suppose that several different inputs I1,

I2 . . . , In when parsed by a grammar G lead to corresponding outputs
O1, O2 . . . , On, all of which are realized as the same phonetic form
these inputs are phonetically equivalent with respect to G. Now
one of these outputs must be the most harmonic, by virtue of incur-
ring the least signicant violation marks: suppose this optimal one is
labelled Ok. Then the learner should choose, as the underlying form
for , the input Ik.

This principle is, in its turn, an elaboration of an idea of Stampe (1972), who
suggested that underlying forms should always match surface forms in the
absence of evidence to the contrary. (The masking effect of one underlying form,
/sNd/, by another, /sd/, is called Stampean occultation in Prince and Smolen-
sky 1993.)
An alternative to Lexicon Optimization is to assume that certain inputs contain
no specication with respect to a feature (Kiparsky 1985, Steriade 1987, Arch-
angeli 1988). This underspecication analysis of nasality in vowels is based on
the idea that the burden of explanation for contrastive versus allophonic patterns
is in the underlying form, rather than in the relationship between underlying form
and surface form, as is the case in OT.15

15
See Smolensky (1993), Inkelas (1995) and It, Mester, and Padgett (1995) for comments on
underspecication in OT.

33
Conicts in grammars

Importantly, Lexicon Optimization does not contradict the assumption of Rich-

ness of the Base, even though it may give rise to lexicons that are impoverished
in terms of featural noise. The burden of explanation still remains on the inter-
action of markedness and faithfulness. More specically, the ranking markedness

faithfulness implies that it is not an accidental observation that nasality is never
distinctive in vowels in English on the contrary, this is a solid effect of con-
straint interactions in the English grammar.

1.7 A factorial typology of markedness and faithfulness

In the remainder of this chapter we will consider the consequences of reranking
the three constraint types which we have assumed in the analysis of allophonic
variation, that is: context-free markedness, context-sensitive markedness, and
faithfulness. We will see that by reranking these three constraint types into differ-
ent hierarchies, a factorial typology arises which exactly matches the attested
cross-linguistic variation in terms of allophonic variation, positional neutraliza-
tion, and free contrast. This will provide further support for the markedness
approach of allophonic variation which we have used so far, as well as for the
assumption of Richness of the Base.

1.7.1 Typological goals of OT

The important notion of factorial typology requires some explanation, before we
actually construct one. The key assumption of OT is that grammars are means to
resolve conicts between universal constraints. More specically, the grammar
of an individual language is a specic way, out of many possible, to rank a set of
universal and violable constraints. Differences between languages must therefore
be due to different rankings of a single set of universal constraints. To state it
differently, we can build one grammar out of another by a rearrangement of its
basic universal material, that is, by reranking the constraints.
The relative success of any theory of grammar should be measured by its
ability to characterize the notion of possible grammar (see again the remarks in
the introduction of this chapter). Constructing grammars (constraint hierarchies)
of individual languages may tell us much about the ways in which linguistic
properties are interconnected within a single linguistic system. But what we are
eventually interested in are typological results of the theory, that is, the predictions
it makes about clusterings of linguistic properties, on a broad cross-linguistic
basis. For example, the theory should explain why no languages occur that have
a contrast of oral and nasal vowels, but contextually restrict this contrast to vowels
immediately preceding nasal stops (neutralizing it everywhere else). Languages

34
 1.7 A factorial typology of markedness and faithfulness

of this kind are logically possible, yet unattested. Can this situation be described
by a reranking of the constraints governing nasality in vowels? Conversely, we
should ask what language types would arise by reranking a number of constraints
that are motivated in the analysis of an individual language. Does reranking of
these constraints produce attested languages as well?
Taken quite literally, the reranking approach would predict that any new gram-
mar that arises from a reranking of any pair of constraints will precisely cor-
relate with one of the worlds languages. This prediction is based on the deeply
naive assumption that every possible ranking should be instantiated by some
attested language. This is naive, just as it is deeply naive to expect that all logic-
ally possible permutations of genetic material in the human genome are actually
attested in individual humans. Therefore, in order to test the typological predic-
tions of the theory of contrast and contextual neutralization presented in this
section, we will rerank types of constraints (rather than individual constraints) of
the general types faithfulness, contextual markedness, and context-free mark-
edness. The resulting factorial typology will be matched with broad typological
diversity between languages, along the dimensions that these constraint types
represent. (Of course, this is not to deny that cases can occur in which it is more
useful to compute factorial typologies of individual constraints.)

1.7.2 Constructing a factorial typology

To construct a factorial typology of a set of constraints, we sum up all logically
possible rankings of this set of constraints, and compute the different outcomes.
With large sets of constraints the number of possible rankings rises steeply, as
with a constraint set of size n, we must consider all n! rankings. (This equals 2
rankings for 2 constraints, 6 rankings for 3, 24 for 4, 120 for 5, 720 for 6, etc.)
Fortunately, many of the individual rankings in a factorial typology produce iden-
tical surface patterns. Therefore the number of predicted patterns is much smaller
than the total number of logically possible rankings. Keeping these remarks in
mind, we now turn to a real case.
Our goal is to construct a factorial typology of the constraint types introduced
so far (faithfulness constraints, context-free and context-sensitive markedness
constraints). One proviso has to be made: actual grammars rank individual con-
straints, not constraint types. (For example, no grammar ranks all faithfulness
constraints above all markedness constraints, for reasons discussed in section
1.2.2.) Nevertheless, we will generalize the factorial typology by a reranking of
constraint types. Accordingly, the emerging factorial typology should be taken as
a catalogue of general effects, which may be instantiated in an individual gram-
mar with respect to specic features.

35
Conicts in grammars

A factorial typology of markedness and faithfulness is presented below, result-

ing from a reranking of both markedness constraint types with faithfulness. We
will abbreviate these types of markedness constraints as MC-free (for context-
free markedness constraint) and MC-sensitive (for context-sensitive markedness
constraint).

(53) A factorial typology of markedness and faithfulness

a. MC-free
MC-sensitive, Faithfulness Lack of variation
(unmarked)
b. MC-sensitive
MC-free
Faithfulness Allophonic variation
c. MC-sensitive
Faithfulness
MC-free Positional neutralization
d. Faithfulness
MC-sensitive, MC-free Full contrast

The attentive reader may have noted that we have only four rankings here, rather
than the predicted six (or 3!). This reduction is due to the fact that in rankings
(53a) and (53d), the mutual ranking of the bottom two constraints is of no import-
ance to the outcome.
The following subsections discuss how these situations arise from these rank-
ings, and also illustrate each ranking by tableaux for contrastive nasality in vowels.

1.7.3 Neutralization: lack of variation versus allophonic variation

Both (53a) and (53b) are situations of complete neutralization, since in both cases
there is a total lack of activity of the faithfulness constraint, which is at the very
bottom of the hierarchy. The difference between the rankings resides in whether
or not the neutralized feature is contextually coloured, that is, subject to allo-
phonic variation.
First consider the situation of total lack of variation, which is produced by
ranking (53a), due to an undominated context-free markedness constraint for
some feature [F]. Accordingly the unmarked value of this feature (for segments
of a given type), [uF], will always appear at the surface, regardless of its under-
lying specication, and regardless of the context. This results in the complete
neutralization of this feature in the direction of the unmarked value. Such total
lack of variation for a given feature (in all segments of some type) is widely
attested for different features among the worlds languages.
For example, if the constraint *VNASAL is undominated, then all surface vowels
are oral, even those vowels which are underlyingly nasal, and even those vowels
which are adjacent to a nasal consonant. This is illustrated in the set of four
tableaux below. Each of these tableaux takes as its input one of the four possible
combinations of input nasality in vowels (nasal versus oral) and output context
36
 1.7 A factorial typology of markedness and faithfulness

of the vowel (preceding a nasal [n] or an oral [l]). All four possible inputs {/pan /
~ /pn/ ~ /pal / ~ /pl /} map onto oral output vowels:

(54) Lack of variation of nasality in vowels (total orality)

*VNASAL
*VORALN, Ident-IO(nasal)
(i) Input: /pan/ *VNASAL *VORALN Ident-IO(nasal)

a. pn *!
b. + pan *

(ii) Input: /pn/ *VNASAL *VORALN Ident-IO(nasal)

a. pn *!
b. + pan * *

(iii) Input: /pal/ *VNASAL *VORALN Ident-IO(nasal)

a. pl *!
b. + pal

(iv) Input: /pl/ *VNASAL *VORALN Ident-IO(nasal)

a. pl *!
b. + pal *

Observe that the ranking of *VORALN and Ident-IO(nasal) with respect to one
another is totally irrelevant to the outcome, since the orality of the vowel is
uniquely determined by *VNASAL.
Ranking (53b) produces the typologically common case of allophonic varia-
tion, of which we have already encountered an example in the form of vowel
nasalization before tautosyllabic nasals in English. As compared to the previous
ranking, (53a), this ranking maintains complete neutralization, yet it allows for
some variation in output values for the relevant feature. For example, both values
of nasality in vowels do occur in surface forms, although their distribution is
totally determined by the context. Vowels are nasal before nasal consonants
(regardless of their input specication), and they are oral in all other contexts
(regardless of their input specication). Although tableaux of English examples
have already been presented in section 1.5.2, we include new tableaux here for
maximal clarity:
37
Conicts in grammars

(55) Allophonic variation of nasality in vowels

*VORALN
*VNASAL
Ident-IO(nasal)
(i) Input: /pan/ *VORALN *VNASAL Ident-IO(nasal)

a. + pn * *
b. pan *!

(ii) Input: /pn / *VORALN *VNASAL Ident-IO(nasal)

a. + pn *
b. pan *! *

(iii) Input: /pal/ *VORALN *VNASAL Ident-IO(nasal)

a. pl *! *
b. + pal

(iv) Input: /pl/ *VORALN *VNASAL Ident-IO(nasal)

a. pl *!
b. + pal *

The two remaining rankings in the factorial typology, (53c) and (53d), produce
varying degrees of contrastiveness, as we will see below.

1.7.4 Contrast: positional neutralization versus full contrast

Ranking (53c) produces a positional neutralization of underlying feature values.
This is a situation in which an underlying contrast is freely realized in most
contexts, but where it is neutralized in a specic context. For example, nasality is
contrastive in vowels, except in the context before a nasal consonant, where all
vowels are nasal:
(56) Positional neutralization of nasality in vowels before nasal consonants
*VORALN
Ident-IO(nasal)
*VNASAL
(i) Input: /pan/ *VORALN Ident-IO(nasal) *VNASAL

a. + pn * *
b. pan *!

38
 1.7 A factorial typology of markedness and faithfulness

(ii) Input: /pn/ *VORALN Ident-IO(nasal) *VNASAL

a. + pn *
b. pan *! *

(iii) Input: /pal/ *VORALN Ident-IO(nasal) *VNASAL

a. pl *! *
b. + pal

(iv) Input: /pl/ *VORALN Ident-IO(nasal) *VNASAL

a. + pl *
b. pal *!

Finally, the logically opposite situation of (53a) is that produced by ranking

(53d), where a faithfulness constraint governing a feature dominates all marked-
ness constraints (governing this feature). This produces a pattern in which input
feature specications are freely realized, that is, a situation of full contrast for the
relevant feature.
For nasality, this ranking produces a situation in which underlying specications
in vowels are realized at the surface, regardless of their adjacency to nasal consonants:
(57) Full contrast of nasality in vowels
Ident-IO(nasal)
*VNASAL, *VORALN
(i) Input: /pan/ Ident-IO(nasal) *VNASAL *VORALN

a. pn *! *
b. + pan *

(ii) Input: /pn/ Ident-IO(nasal) *VNASAL *VORALN

a. + pn *
b. pan *! *

(iii) Input: /pal/ Ident-IO(nasal) *VNASAL *VORALN

a. pl *! *
b. + pal

39
Conicts in grammars

(iv) Input: /pl/ Ident-IO(nasal) *VNASAL *VORALN

a. + pl *
b. pal *!

Again, this free contrast is a cross-linguistically common type of situation.

1.7.5 Positional neutralization of voice in Dutch

It will now be clear that Dutch nal devoicing is an example of positional neu-
tralization: it produces a neutralization of the feature [voice] in obstruents in the
specic context of a syllable coda, while leaving unaffected the lexical distri-
bution of [voice] in obstruents in other contexts. To t the Dutch case into the
ranking schemata of the previous section, we must rst determine which context-
free markedness constraint is involved in the ranking.
The unmarked value for the feature [voice] in obstruents is [voice], as stated
in Voiced Obstruent Prohibition (58c, henceforth VOP, after It and Mester
1998), which is accompanied by the other two constraints relevant to the Dutch
devoicing pattern:

(58) a. *Voiced-Coda (context-sensitive markedness constraint)

Coda obstruents are voiceless.
b. Ident-IO(voice) ( faithfulness constraint)
The value of the feature [voice] of an input segment must be
preserved in its output correspondent.
c. VOP (context-free markedness constraint)
*[+voi, son]
No obstruent must be voiced.

These three constraints are ranked in the following way in Dutch, instantiating the
pattern of positional neutralization (53c) with respect to the feature [voice]:

(59) Ranking producing positional neutralization of voice in Dutch

MC-sensitive
Faithfulness
MC-free
*Voiced-Coda
Ident-IO(voice)
VOP

This ranking states that a voiceless realization of obstruents in coda position takes
priority over preservation of [voice] in coda obstruents. However, preservation of
input values of [voice] takes priority over the complete devoicing of obstruents.
In sum, the contrast of voiced and voiceless obstruents is positionally neutralized
in the syllable coda. Elsewhere, a contrast is possible input values of [voice] are
preserved in the output.
40
 1.7 A factorial typology of markedness and faithfulness

In terms of concrete examples, this ranking correctly predicts that the output
of /bd/ is [bet], which is unfaithful to input values for [voice] only in its coda
consonant. But the voiced onset consonant /b/ is protected from the complete
devoicing of obstruents required by the context-free markedness constraint VOP
(hence, *[pet]). This interaction is shown by tableau (60), containing all four
logically possible combinations of [voice] in the onset and coda consonants:

(60) Input: / bd/ *Voiced-Coda Ident-IO(voice) VOP

a. + [bt] * *
b. [pt] **!
c. [bd] *! **
d. [pd] *! * *

Two candidates (60cd) are eliminated by undominated *Voiced-Coda, as each

contains a voiced obstruent in coda position. Both remaining candidates (60ab)
satisfy *Voiced-Coda, hence both are passed on for evaluation by the next-lower-
ranked constraint in the hierarchy, Ident-IO(voice). Although both (60a) and
(60b) violate Ident-IO(voice), the former is selected since it violates Ident-
IO(voice) minimally. It has only one violation, while its contestant (60b) incurs
two violations, one more than is strictly necessary. This result reects an import-
ant property of the architecture of OT: a constraint can be active even when it
is dominated by one or more other constraints. Constraints may be violated, but
violation must be minimal. This property of constraint interactions will reoccur
many times in this book.
Another major property of constraint interaction is also illustrated by tableau
(60). This is that some candidates can never emerge as optimal, regardless of the
ranking of constraints. To see this, consider output candidate (60d), [pd], which
preserves the input value for [voice] in its coda consonant, but is unfaithful to
[voice] in its onset consonant. This incurs violations for each of the three con-
straints in the tableau: it violates *Voiced-Coda as it has a voiced coda obstruent
[d], it violates Ident-IO(voice) as it is unfaithful to the input value of [voice] in
one of its consonants (the onset [p]), and nally it violates VOP because it con-
tains a voiced obstruent [d]. Under what constraint ranking might this candidate
be selected as optimal? The surprising answer is: under no ranking, since all
logically possible rankings of the three constraints evaluate (60d) as suboptimal
to some other candidate. To prove this point, we need not go through all tableaux
of all possible rankings, although this method will certainly lead to the same
conclusion. A more general proof is available. To mark a candidate cand1 as
41
Conicts in grammars

intrinsically suboptimal, it sufces to identify a rivalling candidate cand2 which

shares with cand1 the violation marks for every constraint, except for at least one
constraint C, on which cand2 is more harmonic. If such a constraint C exists, then
cand2 must be a better candidate than cand1 regardless of the ranking of C, since
the minimal difference in violation marks always works in its favour, even if C
were to dwell at the very bottom of the hierarchy.
There happens to be such a candidate cand2 in tableau (60): candidate (60a),
[bt], shares with candidate [pd] (60d) one violation mark for Ident-IO(voice),
and one for VOP, yet it minimally improves over (60d) with respect to *Voiced-
Coda. Hence:

(61) [bt]
[pd] For input / bd/, irrespective of ranking.

This does not imply that [bt] is the optimal candidate under any ranking: it
clearly is not (only consider rankings in which either *Voiced-Coda or VOP is
undominated). It does imply, however, that [pd] is intrinsically suboptimal
which means that it will never be selected as optimal under any logically possible
ranking of the three constraints under consideration.
This result, although apparently limited to the interaction of the three con-
straints in tableau (60), in fact has broader typological implications. A prediction
follows from it, which is stated in general terms as follows. Assume a context-
free markedness constraint banning one value of a feature [F], and another
context-sensitive markedness constraint banning the same value [F] in a specic
context. The prediction is that no language can have a contrast of [F] exclusively
in the context where a context-sensitive markedness constraint bans [F]. This
seems to be correct, although further testing may be required.
One particular language type excluded is one that has a lexical contrast of
voicing exclusively in syllable codas. See the following hypothetical pattern of
contrast:

(62) A hypothetical language that is predicted not to occur

a. a contrast of voice in syllable codas
lap ~ lab, pot ~ pod, muuk ~ muug
b. but no contrast of voice elsewhere
paa (*baa), ma.tol (*ma.dol), tol.ku (*tol.gu)

Such a language would preserve a contrast of voice in the coda, but neutralize it
elsewhere. That is, it would map an input /bd/ onto an optimal output [pd]. But
we have just seen that such a mapping is ruled out on principled grounds, since
it involves the selection of an intrinsically suboptimal candidate. The asymmetry
between onsets and codas is due to a context-sensitive markedness constraint
42
 1.8 On dening segment inventories

*Voiced-Coda which rules out [+voice] in the syllable coda, while there is no
analogous context-free markedness constraint which rules out any feature of voice
specically in the onset.
Alternative theories which do not assume markedness to be the actual sub-
stance of the grammar fail to derive this general prediction. For example, a rule-
based theory in which the notion of markedness is an external criterion fails to
predict that hypothetical languages such as (62) should not exist. This is because
phonological rules that neutralize a contrast of voice are natural in any context,
regardless of whether they apply in onset or in coda position. A rule neutralizing
voicing in onsets is natural in this general sense, and no language that has this
rule is committed to having a second rule neutralizing voicing in codas as well.
Therefore a grammar which neutralizes a voicing contrast in all contexts except
in codas should be possible, even though it would be complex (in the sense that
different rules would be employed, instead of a single general one).

1.7.6 Typology: some preliminary conclusions

To wind up this section let us now summarize the results. At the heart of OT is
the notion that grammars of individual languages instantiate general ranking
schemata of constraints of different types. The basic method of checking the
typological predictions made by the theory is that of constructing a factorial
typology by the reranking of constraints of different types. In this section we have
constructed a basic factorial typology of faithfulness and markedness, and found
that all predicted types of inputoutput relationships are attested. By varying the
ranking of faithfulness with respect to (context-free and context-sensitive) mark-
edness constraints, we found a factorial typology which ranged from a situation
of total neutralization on the one hand, to that of total freedom of contrast on the
other hand. In between these extremes, we identied two intermediate situations:
allophonic variation (a specic kind of neutralization which allows two values of
some feature in the output), and positional neutralization (a situation in which a
feature is contrastive, except in a specic context, in which it is neutralized). We
elaborated on positional neutralization of voice in Dutch to demonstrate two
typical properties of constraint interaction in OT. In the rst place, we found that
dominated constraints may still be active, in the sense that a constraint, even when
it is violated, must be minimally violated. Secondly, we found that some output
candidates are intrinsically suboptimal to others, regardless of ranking. This cap-
tures certain typological observations with respect to positional neutralization.

1.8 On dening segment inventories

This section will show how segment inventories result from interactions of faith-
fulness constraints and markedness constraints. The discussion is related to the
43
Conicts in grammars

notion of Lexicon Optimization (section 1.6). As before, we will draw heavily on

Prince and Smolensky (1993).
Recall that the main type of conict in OT is that between markedness con-
straints and faithfulness constraints. Faithfulness constraints militate against any
loss of contrast, enforcing identity between the input and the output. Markedness
constraints are natural antagonists of faithfulness constraints, militating against
marked structures in the output. They may produce the effect of a loss of a feature
value present in the input, if that value is a marked value.

1.8.1 Markedness constraints on scales

Markedness is intrinsically a relative notion. That is, which types of elements are
marked and which are unmarked can be established only in comparison to
other elements. For example, we cannot say that nasalized vowels are intrinsically
marked; they are marked in relation to oral vowels only. Moreover, markedness
often involves a hierarchy of segment types, each member of which is more
marked than successive members of the hierarchy. For example, it has been
argued in the literature on segmental markedness that coronals are universally less
marked than labials (Paradis and Prunet 1991).
To capture this relative markedness of segment types, Prince and Smolensky
(1993) introduce the idea that markedness relations can be organized in a scalar
fashion, as in the harmony scale Cor
Lab. Moreover, the ranking of con-
straints that govern markedness relations along a single dimension (such as place
of articulation) is universally xed. In the case in hand the constraints governing
the relative markedness of labials and coronals are intrinsically ranked as below:

(63) Universal ranking for markedness constraints governing place of

articulation
*[lab]
*[cor]

Presumably this ranking is universal, hence respected by every constraint hier-

archy. Note that, by itself, this partial ranking makes no predictions about which
segments are actually attested in a segment inventory of a particular language,
and which are not. Inventories emerge from interactions of markedness constraint
hierarchies (as in 63) with faithfulness constraints, which serve to preserve input
segmental contrasts in the output. Generally speaking, the higher faithfulness
constraints are ranked, the larger the segment inventory will be. And vice versa,
the lower faithfulness is ranked, the smaller the inventory.

1.8.2 Interaction of markedness scales and faithfulness

Now consider the faithfulness constraint militating against differences of place
features in the input and output:
44
 1.8 On dening segment inventories

(64) Ident-IO(Place)
The specication for place of articulation of an input segment must
be preserved in its output correspondent.

This correspondence constraint is satised by (65a); but it is violated by (65b):

(65) a. /p/ Input b. /p/ Input

[p] Output [t] Output

Let us now consider two grammars differing only in the ranking of Ident-
IO(Place) with respect to the markedness constraints of (63). One grammar that
we will consider ranks Ident-IO(Place) above both markedness constraints:

(66) A grammar that is maximally faithful to place of articulation

Ident-IO(Place)
*[lab]
*[cor]

This grammar is maximally faithful to its input place of articulation, due to high-
ranked Ident-IO(Place). It is more important to be faithful to the input place of
articulation of a segment than to its output markedness. The tableaux (67.iii),
one for each input segment /p/ and /t/, illustrate this:

(67.i) /. . . p . . ./ Ident-IO(Place) *[lab] *[cor]

a. + [. . . p . . .] *
b. [. . . t . . .] *! *

(67.ii) /. . . t . . ./ Ident-IO(Place) *[lab] *[cor]

a. [. . . p . . .] *! *
b. + [. . . t . . .] *

These tableaux can be summarized as follows: whatever place of articulation is

specied at the lexical level will reach the surface level. Accordingly, the seg-
ment inventory of this language will contain two places of articulation: {labial,
coronal}.
Next consider a second grammar, which ranks Ident-IO(Place) more modestly:

(68) A grammar that is less faithful to place of articulation

*[lab]
Ident-IO(Place)
*[cor]

This grammar is less faithful (than the one in 66) to its input place features. It
blocks the surfacing of any labial in the input since the markedness constraint
45
Conicts in grammars

militating against this segment type outranks the faithfulness constraint Ident-
IO(Place). Accordingly, any input labial surfaces as a coronal, if it surfaces at
all:16

(69.i) /. . . p . . ./ *[lab] Ident-IO(Place) *[cor]

a. [. . . p . . .] *!
b. + [. . . t . . .] * *

(69.ii) /. . . t . . ./ *[lab] Ident-IO(Place) *[cor]

a. [. . . p . . .] *! *
b. + [. . . t . . .] *

This amounts to a neutralization (loss of contrast) of place-of-articulation fea-

tures, in the favour of coronals. In fact the surface level of such a language would
present no evidence whatsoever for input labials. For the learner, it would be
senseless to set up a lexical contrast among labials and coronals, since such a
contrast would be entirely overruled by markedness effects. Prince and Smolen-
sky (1993) refer to the obscuring of input contrasts as Stampean occultation,
stating it as a principle of Lexicon Optimization (see section 1.6).

1.8.3 Conclusions and predictions

In sum, we have found that the resulting consonant inventories depend on the
ranking of faithfulness constraints with respect to the markedness constraints.
With respect to place of articulation, we have found the following relation be-
tween ranking and inventory:

(70) Ranking Inventory

a. Ident-IO(Place)
*[lab]
*[cor] {p, t}
b. *[lab]
Ident-IO(Place)
*[cor] {t}

The interesting consequence of this approach is that segment inventories need no

longer be stipulated at the level of lexical forms, as in derivational theory. Instead,
the ranking of faithfulness constraints amongst markedness constraints is part
of the grammar, hence it is automatically respected at the level of the output.

16
Of course one might also consider the logical possibility that input labials are simply deleted
there is no way of telling the difference between both possibilities since it is impossible to
establish the presence of input labials.

46
 1.9 Conclusion

Derivational theory arrives at the same result by stipulating that the output of
(lexical) rules must contain no segments that are not part of the input inventory:
this is referred to as the structure-preserving property of phonological rules
(Kiparsky 1985).
Another interesting consequence of this theory of segmental markedness is
that it predicts that unmarked segments will emerge wherever faithfulness con-
straints are put out of control. Some segment types in every inventory (for
example, coronals) are less marked than other segments, even though the gram-
mar allows both more and less marked segments at the surface level, due to high-
ranked faithfulness constraints. However, given the chance, the grammar will still
favour unmarked segments over marked segments. This occurs in special situa-
tions in which, for some reason, input faithfulness requirements no longer hold.
This prediction has been conrmed robustly for a wide range of situations in a
wide range of languages.
Consider, for example, epenthesis: a segment appears in the output exclusively
for phonotactic reasons, as in the case of a vowel inserted to break up a consonant
cluster. By its very nature, the inserted vowel lacks a counterpart in the input. It
cannot be subject to input faithfulness, so that its featural content is fully deter-
mined by markedness factors. The prediction is that epenthetic segments are
segmentally unmarked or easily inuenced by segments in their contexts. This,
and other related observations, is presented in (71).

(71) a. Epenthetic segments are less marked than lexically sponsored

segments.
(Explanation: epenthetic segments have no input counterparts,
hence they are free of faithfulness constraints. See chapter 3.)
b. Segments in reduplicants are less marked than segments in their
bases.
(Explanation: copied segments have no input counterparts,
hence they are free of faithfulness constraints. See chapter 5.)
c. Segments in afxes are less marked than segments in roots.
(Explanation: afx faithfulness is intrinsically lower-ranked than
root faithfulness. See chapters 5 and 9.)

The observations in (71ac) will all be conrmed in later chapters, in the light
of insights into faithfulnessmarkedness interactions.

1.9 Conclusion
In this chapter we have laid out the foundations on which OT is built, and pointed
out the most important linguistic phenomena that fall in its scope. In the discus-
sions throughout this chapter, one aspect of OT stood out: the interaction of
47
Conicts in grammars

faithfulness and markedness. We have seen that all general phonological phe-
nomena discussed here are variations on this theme: the notion of contrast, and,
related to this, allophonic variation, neutralization, and lexical representations,
and nally, the notion of segment inventory. In every case, a phonological pattern
resulted from (more or less complex) interactions of constraints that preserve
lexical input properties (faithfulness), and others that reduce output markedness.
In later chapters of this book, we will maintain this perspective, and generalize it
to a range of other linguistic phenomena.

The following eight chapters of this book will each be devoted to a research topic
in which OT has left its marks. Chapter 2 addresses functional relations among
processes, comparing OT with rule-based theory from this perspective. Chapter 3
discusses syllable structure and related phenomena, such as syllabically governed
epenthesis and deletion. Chapter 4 deals with metrical phenomena, particularly
word stress and quantity effects. Chapter 5 addresses the morphologyphonology
interface, focussing on reduplication, and also extends the notion of correspond-
ence beyond relations of input and output. Chapter 6 further extends correspond-
ence to relations between morphologically related output forms, covering
paradigm regularity. Chapter 7 is devoted to the issues of learnability and acqui-
sition. Chapter 8 contains applications of OT outside phonology in syntax.
Finally, chapter 9 discusses residual issues, focussing on issues that deserve further
research (in particular, opacity) as well as on current theoretical developments.

SUGGESTIONS FOR FUR THER READING

General introductions to phonology

Goldsmith, John (1990) Autosegmental and metrical phonology. Oxford: Basil
Blackwell.
Kaye, Jonathan (1990) Phonology: a cognitive view. Hillsdale: LEA.
Kenstowicz, Michael (1994a) Phonology in generative grammar. Oxford: Basil
Blackwell.
Kenstowicz, Michael and Charles Kisseberth (1979) Generative phonology:
description and theory. San Diego: Academic Press.
Roca, Iggy (1994) Generative phonology. London and New York: Routledge.

Constraints in phonology
Hayes, Bruce (1986) Inalterability in CV Phonology. Language 62. 321 51.
McCarthy, John (1986) OCP effects: gemination and antigemination. Linguistic
Inquiry 17. 20763.
48
 1.9 Further
1.9 Conclusion
Reading

Odden, David (1986) On the role of the Obligatory Contour Principle in phono-
logical theory. Language 62. 353 83.
Paradis, Carole (1988) On constraints and repair strategies. The Linguistic Review
6. 71 97.
Yip, Moira (1988) The Obligatory Contour Principle and phonological rules: a
loss of identity. Linguistic Inquiry 19. 65100.

General introductions to OT
Archangeli, Diana (1997) Optimality Theory: an introduction to linguistics in the
1990s. In D. Archangeli and D. T. Langendoen (eds.), Optimality Theory: an
introduction. Oxford: Blackwell. 132.
Burzio, Luigi (1995) The rise of Optimality Theory. Glot International 1:6. 3 7.
Prince, Alan and Paul Smolensky (1997) Optimality: from neural networks to
universal grammar. Science 275. 1604 10.
Sherrard, Nicholas (1997) Questions of priorities: an introductory overview of
Optimality Theory in phonology. In Roca. 4389.

Founding papers of OT
McCarthy, John (1993) A case of surface constraint violation. In C. Paradis and
D. LaCharite (eds.), Constraint-based theories in multilinear phonology, spe-
cial issue of Canadian Journal of Linguistics 38. 169 95.
Prince, Alan and Paul Smolensky (1993) Optimality Theory: constraint interac-
tion in generative grammar. Ms., Rutgers University, New Brunswick, and
University of Colorado, Boulder. RuCCS-TR-2. [To appear, Cambridge,
Mass.: MIT Press.]

Phonetic grounding of constraints

Archangeli, Diana and Douglas Pulleyblank (1994) Grounded phonology. Cam-
bridge, Mass.: MIT Press.
Hayes, Bruce (1996a) Phonetically driven phonology: the role of Optimality The-
ory and inductive grounding. Ms., UCLA. [ROA-158, http://ruccs.rutgers.edu /
roa.html]
Myers, Scott (1997b) Expressing phonetic naturalness in phonology. In Roca.
125 52.
Steriade, Donca. (1995b) Positional neutralization. Ms., University of California,
Los Angeles.

Markedness and underspecication

Archangeli, Diana (1988) Aspects of underspecication theory. Phonology 5.183
207.
49
Conicts in grammars

Inkelas, Sharon (1995) The consequences of optimization for underspecica-

tion. In J. Beckman (ed.), Proceedings of NELS 25. 287302. [ROA-40,
http://ruccs.rutgers.edu/roa.html]
It, Junko, R. Armin Mester, and Jaye Padgett (1995) Licensing and under-
specication in Optimality Theory. Linguistic Inquiry 26. 571 613.
Steriade, Donca (1995a) Underspecication and markedness. In Goldsmith. 114
74.

OT in computational phonology
Eisner, Jason (1997) Efcient generation in primitive Optimality Theory. Pro-
ceedings of the 35th Annual Meeting of the Association for Computational
Linguistics. [ROA-206, http://ruccs.rutgers.edu/roa.html]
Ellison, Mark T. (1994) Phonological derivation in Optimality Theory. Proceed-
ings of the fteenth International Conference on Computational Linguistics,
II. Kyoto, Japan. 100713. [ROA-75, http://ruccs.rutgers.edu /roa.html]
Hammond, Michael (1997) Parsing syllables: modeling OT computationally. Ms.,
University of Arizona. [ROA-222, http://ruccs.rutgers.edu /roa.html, parser
code available from author, http://www.u.arizona.edu /~hammond]
Walther, Markus (1996) ot simple A construction-kit approach to Optimal-
ity Theory implementation. Ms., Heinrich Heine Universitt, Dsseldorf.
[Software and paper can be downloaded from http://www.phil-fak.uni-
duesseldorf.de/~walther/otsimple.html. Also ROA-152, http://ruccs.rutgers.
edu /roa.html]

EXERCISES

1 Japanese
Consider the following distribution of [g] and [] in Japanese (It and Mester
1997):

(i) geta *eta clogs

giri *iri duty
guchi *uchi complaint
go *o (game of) Go
(ii) kai *kagi key
kao *kago basket
kaae *kagae thought
tokae *tokage lizard

50
 1.9
1.9Conclusion
Exercises

a. State the generalization for the distribution of [g] and []. Categorize
this distribution in terms of the typology discussed in section 1.7.
b. Account for this generalization by a set of ranked constraints.
c. Support your analysis by tableaux of geta and kafi.

2 English
Consider the following English word pairs, some of which display alternations of
voice:

(i) cat [kt] cats [kts]

dog [d
] dogs [d
z]
hen [hen] hens [henz]
(ii) twelve [twelv] twelfth [twelf]
eight [et] eighth [et]
ten [ten] tenth [ten]

a. What are the underlying forms of the sufxes in (i) and (ii)?
b. To account for these alternations, you need a new constraint. State
this constraint (as generally as possible). To what extent is this con-
straint phonetically grounded?
c. Rank the constraints, motivating each individual ranking by at least
one form. Support your analysis by tableaux of cats, dogs, hens,
twelfth, and eighth.

51
 2
The typology of structural changes

2.1 Introduction
In this chapter we will learn how a range of phonological processes can be
triggered by a single markedness constraint, depending on its interactions with
faithfulness constraints. Unity in diversity among processes is predicted by OT,
due to its surface-oriented nature. To satisfy a markedness constraint in the out-
put, various repair strategies can be applied to the input; whichever is chosen
depends on the relative ranking of different faithfulness constraints. The main
result of this chapter will be a typological one: functionally related phonological
processes arise by reranking universal constraints. This factorial typology of pro-
cesses is based on Correspondence Theory (McCarthy and Prince 1995a, forthcom-
ing), a subtheory of faithfulness constraints allowing a limited set of structural
changes, such as deletions, insertions, fusions, and featural changes.
This chapter is organized as follows. Section 2.1 will compare OT and rule-
based theory with respect to their core devices: constraints versus rewrite rules.
Section 2.2 will discuss the analysis of Pater (forthcoming) of nasal substitu-
tion in Indonesian, a process avoiding Nt (a nasal-and-voiceless-obstruent se-
quence) and introduce major correspondence constraints. Section 2.3 will develop
the factorial typology of Nt effects, while section 2.4 will focus on conspirac-
ies of related processes within a language. Finally, section 2.5 will present
conclusions.

2.1.1 An initial comparison with rule-based theory

In classical generative phonology the grammar is an input output mechanism,
just as it is in OT. However, the key mechanism is the rewrite rule, rather than
the output constraint. An example of a rewrite rule is devoicing of obstruents in
the coda (Dutch, chapter 1):

(1) Coda devoicing as a rewrite rule

[son] [voice] / ___ ]

52
 2.1 Introduction

Although rewrite rules are language-specic in the sense that not all grammars
share the same set of rules, their format is universal. A, B, X, and Y are natural
classes of elements:

(2) Format of context-sensitive rewrite rule

A B / X ___ Y

Element A (the focus) is rewritten as element B in the context of elements X and

Y. Each rule makes precisely one structural change (A B) to the input. For
a rule to apply to an input, the input must match the rules structural description
(XAY). The structural change eliminates the conguration XAY, resulting into the
output conguration (XBY).

(3) Application of a rewrite rule

Input: . . . XAY . . .

Output: . . . XBY . . .

The repertory of structural changes to be described by rewrite rules includes

permutations (replacement of A by B, both non-null), insertions (where A is
null), and deletions (where B is null).

2.1.1.1 Triggers and changes

We may distinguish two aspects in each context-sensitive rewrite rule:

(4) Two aspects of context-sensitive rewrite rules

a. a structural description, XAY, which denes the rules input, a
trigger
b. a structural change, A B, which denes the rules output
(XBY)

Each aspect has a counterpart in OT. First, the trigger is a negative constraint
(*XAY), dening a conguration to be avoided. Second, the structural change
becomes a context-free rewrite operation (A B), which is attributed to the Gen-
erator component, one out of many changes possible under Freedom of Analysis.

Since Gen is contextually blind, application of a process fully depends on Eval,

more specically on the language-specic interaction of a markedness constraint
(*XAY) and a faithfulness constraint militating against this change (*A B):

53
The typology of structural changes

(5) Trigger and anti-change in OT

a. *XAY Avoid the conguration XAY (markedness)
b. *A B A must not be realized as B (faithfulness)
For the structural change to apply, the faithfulness constraint must be dominated
by the markedness constraint the trigger. In the reverse ranking, the change
will not apply.
(6) Application and non-application in OT
a. Ranking for application: Markedness
Faithfulness
*XAY
*A B
b. Ranking for non-application: Faithfulness
Markedness
*A B
*XAY
For example, Ident-IO(voice) is dominated by *Voiced-Coda in Dutch, while
English has the reverse ranking. We have now distinguished two aspects in rewrite
rules, trigger and change, and seen how OT decomposes processes into pairs
of antagonistic constraints.

2.1.1.2 The notion of structural change

Until here the OT and rule-based scenarios match fairly well: both identify ele-
ment /A/ as a focus of change, triggered by a structural description or a negative
constraint *XAY. However, we now arrive at an important point of difference. OT
predicts that the output goal of avoiding *XAY is attainable in various ways. That
is, whatever structural change A B may represent, it is only one out of a range
of changes generated by Gen, none of which is bound to a specic structural
description. On Freedom of Analysis, Gen is free to propose any kind of change
to the input element /A/, including no change at all:
(7) Some candidate analyses of an input / . . . XAY . . . /
/ . . . XAY. . . / [ . . . XAY . . . ] (no change)
[ . . . XBY . . . ] (change/A/into [B])
[ . . . XCY . . . ] (change/A/into [C])
[ . . . XY . . . ] (delete/A/)
For example, *Voiced-Coda in Dutch is satised by obstruent devoicing,
but potentially it might also have been satised by obstruent deletion or coda
nasalization.
Satisfaction of *XAY is not uniquely achieved by changes in /A/, but may
equally well arise as a result of changes in any of its neighbouring segments X
or Y, by insertion of an element B, or more generally by any change eliminating
[XAY]:

54
 2.1 Introduction

(8) More candidate analyses of an input / . . . XAY . . . /, which leave A

unaffected
/ . . . XAY . . . / [AY] (delete /X/)
[XAB] (change /Y/ into [B])
[XABY] (insert [B] between A and Y)
[XA] (delete /Y/)

All analyses proposed by Gen are submitted to Eval for parallel evaluation. Cru-
cially, all candidate analyses of (7) which change /A/ are evaluated in a single
pool of candidates together with those of (8), which employ different types of
changes. Each analysis has a price tag attached, in terms of constraint violations.
The single analysis that is perfectly faithful to the input, /XAY/, is the canonical
violator of the markedness constraint *XAY. All candidate analyses changing the
input violate one or the other faithfulness constraint, depending on the nature of
the change. Hence, the relative success of each candidate fully depends on the
interaction of *XAY with other constraints.
If *XAY is undominated, some change must be made. The nature of this change
(deletion, insertion, feature change, etc.) depends upon the relative ranking of
faithfulness constraints. Generally speaking, the lowest-ranking faithfulness con-
straint is the one to be violated in the actual output. For example, any grammar
in which segment preservation (Max-IO) outranks featural identity (Ident-
IO) will prefer a featural change A B to the deletion of A. Other structural
changes arise under different rankings of faithfulness constraints, yielding a fac-
torial typology of structural changes. (See section 2.3.)

2.1.2 Two differences between OT and rule-based theory

2.1.2.1 Conspiracies and the functional unity of processes

In the preceding discussion, we have learned that rule-based and constraint-based
theories differ in the following way. In rule-based theory, the structural con-
dition and structural change are linked in the rule; a rule uniquely determines the
structural change in response to the structural condition. In OT, structural con-
dition and structural changes must always be evaluated among other possible
resolutions of constraint violations. Therefore OT predicts that a markedness con-
straint may trigger various types of structural changes, depending on its inter-
action with faithfulness constraints. Different languages should therefore pursue
different repair strategies in attaining identical output goals. (In section 2.4 we
will see that different repair strategies may even occur within a single language
as a result of a single constraint ranking.)

55
The typology of structural changes

In contrast, rule-based theory fails to make this prediction of the functional

unity of processes because it has no formal means of expressing the notion of
output goal of a phonological rule. The functional unity among rules attaining
the same goal, either within a single language or across languages, remains un-
explained. This problem was signalled as early as Kisseberth (1970) in a paper
carrying the meaningful title On the functional unity of phonological rules. As
an example, consider the set of rules in (9). All function to avoid the conguration
*XAY, yet these rules cannot be formally related:

(9) A set of functionally coherent rules

a. A B / X ___ Y d. Y Z / XA ___
b. A C / X ___ Y e. B / XA ___Y
c. A / X ___ Y f. X / ___ AY

This reoccurrence of a common output factor which guides different rules, with-
out being explicitly stated in the rules, is called a conspiracy (Kisseberth 1970).
As we will see in section 2.2, functionally related processes are straightforwardly
dealt with by OT.

Another observation difcult to explain by rule-based theory is that the dynamic

phonology of a language (structural changes brought about by the collective
rewrite rules) is closely related to the static phonology (the structural conditions
holding for all lexical items: Morpheme Structure Constraints). For example, a
rule of vowel epenthesis breaking up a consonant cluster arising by morpheme
concatenation satises a goal generally respected by all morphemes of the lan-
guage. Output goals of rules are mirrored by the structure of morphemes, without
any formal recognition of this similarity in the grammar. This defect is known as
the Duplication Problem (Kenstowicz and Kisseberth 1977).

Before OT, phonologists had already realized that output constraints are neces-
sary ingredients of grammatical theory. As a response to rule conspiracies and
the Duplication Problem, they introduced output constraints to block or trigger
the application of rules. Among the rst output constraints were the OCP in
autosegmental theory (no identical adjacent autosegments, Goldsmith 1976),
and the No-Clash constraint in metrical theory (Liberman 1975). Such additions
resulted in mixed models, containing both rules and output constraints. Various
proposals were made for interactions of rules and constraints, such as the Theory
of Constraints and Repair Strategies (Paradis 1988), and Persistent Rule Theory
(Myers 1991).
Mixed models naturally lead to overlapping functions of rules and constraints,
as rules preserve specic structural conditions. These state congurations to be
56
 2.1 Introduction

repaired by a structural change, hence they are interpretable as rule-specic

negative constraints. Output targets are stated both in the rules and in the con-
straints, producing an overlap in theoretical machinery.
But problems of mixed models do not end with a certain amount of conceptual
overlap. Such models involve an extremely complicated interaction of rules and
constraints. A rule may apply in violation of a constraint, which violation is later
repaired by some subsequent rule. Therefore a mixed model must not only
stipulate structural conditions of the rules and the linear ordering of the rules,
but also interactions of rules and output constraints, dening the conditions
under which output constraints can be temporarily violated. OT avoids such
interactional complexity by limiting grammatical interactions to constraints. This
unication of interaction makes OT, both conceptually and computationally, a
much simpler theory than any mixed model.
The next question, to be addressed below, is: are rule-based grammars and OT
grammars equivalent in the sense that they may perform the same kinds of input
output mappings?

2.1.2.2 Intermediate levels in derivations

The second major difference from rule-based theory resides in the notion of
derivation. In rule-based theory, application of rules follows the principle of
linear (serial) ordering, rather than a parallel mapping according to a hierarchy
of importance, as in OT. Linearly ordered rewrite rules map inputs into outputs
by a derivation:

(10) A derivational grammar as an inputoutput mechanism

Lexical representation (Input to Rule 1)
Rule 1
Intermediate representation (Output of Rule 1,
Input to Rule 2)
Rule 2
Intermediate representation (Output of Rule 2,
Input to Rule 3)
...
Rule n
Surface representation (Output of Rule n)

What is essential about a derivation is that each step is a miniature inputoutput

mapping. That is, a rewrite rule takes as its input a state of affairs left by the
previous rule, makes a change, and passes its output on to the next rewrite rule.
In a derivation the application of a rule solely depends on whether the structural
description is met by the output of the immediately preceding rule. Rules are
57
The typology of structural changes

blind to their own outputs, which they produce mechanically. Moreover, each rule
is blind to the output of the derivation as a whole, which arises only after the last
rule has applied. It is thus predicted that the application of a rule can never
depend on its eventual consequences at the surface.
To appreciate the unlimited potential of intermediate representations in rule-
based theory, consider the following hypothetical derivation.
(11) Lexical form: /XAY/
Rule 1 A B / X ___ XBY
Rule 2 Y Z / B ___ XBZ
Rule 3 B C / X ___ Z XCZ
Surface form: [XCZ]
Segment B is rst introduced by Rule 1, then triggers a change in a neighbouring
segment Y due to Rule 2, and is nally deleted by Rule 3. Segment B, which plays
an active role in the derivation, is not present in the input nor in the output. It is
abstract, being present only at an intermediate stage of the representation between
the application of Rules 1 and 3. Rule-based theory predicts that such situations
should occur in natural languages, since they arise as a natural consequence of
rule interaction.
In contrast, OT attributes major importance to the surface level in the interac-
tion of constraints, disallowing access to intermediary levels between the input
and output. Constraints refer to either the output alone (in markedness con-
straints), or the input and output in combination (in faithfulness constraints of the
correspondence format). This predicts that no property of phonological forms
depends on information that is not present in the output either in the output
alone, or in the relation between the input and output.
It is, of course, an empirical matter to decide on the correctness of this prediction.
Do phonologies of natural languages ever show the degree of abstractness pre-
dicted by rule-based theory, as illustrated in derivation (11)? It has been argued,
on the basis of evidence from well-studied cases, that phonological rules may
refer to elements at a level of representation that does not coincide with the output
(see Kisseberth 1969 on Yokuts). In most of such cases this level was identied
as the input, which is not an intermediate level. But in other cases the abstract
element was indeed argued to be introduced by rule (see Brame 1974 on Pales-
tinian Arabic). Both types of opacity are potential problems for surface-oriented
OT we will return to this issue in chapter 9. Meanwhile, however, it seems fair
to conclude that the radical abstractness of the type (11) that rule-based theory
predicts to be possible in natural languages is rarely if ever attested. Over-
whelmingly, phonological generalizations refer to the output, possibly in relation
to the input.
58
 2.2 Nasal substitution and related effects

2.2 Nasal substitution and related effects

In this section we will look into evidence from segmental phonology which sup-
ports both predictions made in section 2.1. First we will nd that a single mark-
edness constraint can trigger an array of structural changes, even within a single
language. Discussion will be based on the analysis of nasal substitution and
related effects by Pater (forthcoming). This analysis is couched in the frame-
work of Correspondence Theory, a subtheory of OT which allows a limited set
of structural changes (McCarthy and Prince 1995a). Second, we will nd that the
surface level is essential in capturing phonological generalizations.

2.2.1 Nasal substitution in Indonesian

The well-known pattern of nasal substitution in Indonesian (e.g., Halle and
Clements 1983) appears in verbs prexed by /mN-/ (where /N/ indicates a
nasal that is not specied for place of articulation):

(12) Nasal substitution in Indonesian

a.i /mN-pilih/ mmilih to choose, to vote
a.ii /mN-tulis/ mnulis to write
a.iii /mN-kasih/ masih to give
b.i /mN-blih/ mmblih to buy
b.ii /mN-dapat/ mndapat to get, to receive
b.iii /mN-
anti/ m
anti to change

The generalization can be stated as follows: when the input contains a sequence
of a nasal followed by a voiceless obstruent, the latter is deleted, leaving its place
of articulation on the nasal. This is arguably a case of coalescence (segment
fusion), since the consonant in the output has the characteristics of both input
counterparts: the nasality of the lefthand consonant, and the place of articulation
of the righthand consonant. Turning to (12b), we see that a voiced obstruent is
retained in the output form, as part of a nasal-plus-voiced-obstruent cluster. At
the lexical level, the forms in (12a) differ from those in (12b) only in their voicing
of the rst stem consonant.
To reinforce the arguments for the coalescence analysis, let us rst try to side-
step it by breaking the substitution pattern into two ordered rules. Such an ana-
lysis is actually possible, but we will soon nd that it is inadequate. It runs as
follows. Nasal assimilation copies the place of articulation of the obstruent onto
the nasal; then a second rule deletes a voiceless consonant which follows a nasal:1

1
This is a so-called counterbleeding rule order: deletion might have eliminated the context of
application for assimilation by applying rst, but failed to do so. The reverse bleeding rule
order is discussed below. See chapter 9 for further discussion of counterbleeding in OT.

59
The typology of structural changes

(13) Nasal substitution as a sequence of rules

Lexical form: /mN-pilih/ /mN-blih/
Nasal assimilation mmpilih mmblih
Post-nasal voiceless consonant deletion mmilih ---
Surface form: [mmilih] [mmblih]

Pater (following Stahlke 1976) points out that an ordered rule analysis predicts
that both rules should occur independently. While nasal assimilation is a cross-
linguistically highly common process, no typological evidence seems to exist for
the second rule in isolation. That is, post-nasal voiceless consonant deletion
seems always to apply in combination with nasal assimilation. By splitting the
process of nasal substitution into two parts, the incorrect prediction is made that
assimilation and deletion may apply in the reverse order in different languages,
as in the derivation below:

(14) Nasal substitution: the dumb rule order

Lexical form: /mN-pilih/ /mN-blih/
Post-nasal voiceless consonant deletion mNilih ---
Nasal assimilation --- mmblih
Surface form: [milih] [mmblih]

In sum, the fact that the two changes cannot be typologically disentangled, and
the fact that the nasal always preserves the place of articulation of the deleted
consonant, both point to a coalescence: the input nasal and voiceless stop merge
into a single nasal of the place of articulation of the input stop. An ordered rule
analysis entirely misses this point. (See de Haas 1988 for similar arguments about
vowel coalescence.)
There is independent evidence for the coalescence analysis within the phono-
logy of Indonesian. Consider the reduplication pattern of prexed roots (Lapoliwa
1981):

(15) a. /mN-kata-red -i/ mata- ata -i to speak ill of someone

b. /mN-
rak-red/ m
rak-
rak to move something
repeatedly
c. /mN-lu-red-kan/ mlu-lu-kan to praise

Examples (15b c) show that the nasal of a prex is not normally copied by
reduplication, and that reduplication is limited to segments of the root. The prex-
nal nasal in (15c) is [], a nasal with the default place of articulation. But
example (15a) shows that a prexal nasal that is the result of nasal substitution
is copied along. The criterion for the copying of a nasal cannot be that it is
60
 2.2 Nasal substitution and related effects

assimilated to the place of articulation of a root consonant. This is because both

(15a) and (15b) involve assimilation, while only the nasal in (15a) is copied.
Therefore the criterion for copying must be stronger: a nasal is copied only if it
results from fusion with the root consonant. In copying the material of the root,
the nasal must be copied along, simply because it is (partly) made up of a root
segment.
Assuming the correctness of the coalescence analysis, the rst question which
we must ask is this: why are voiced obstruents preserved, but voiceless obstruents
coalesced? Pater argues that the coalescence is triggered by a prohibition against
sequences of a nasal and a voiceless obstruent, stated as a contextual markedness
constraint:

(16) *Nh
No nasal plus voiceless obstruent sequences.

This markedness constraint *Nt is grounded in articulatory mechanisms as fol-

lows (see, among others, Huffman 1993, Hayes and Stivers 1995).2 Nasal conson-
ants involve airow through the nose, which is achieved by lowering the velum,
the valve separating the oral and nasal cavities. Going from a nasal consonant into
an obstruent involves a raising of the velum, cutting off the air from passing
through the nose. However, the raising of the velum takes some time and is not
yet completed at the point where the obstruent begins. At that point (which is
actually perceptually determined), there is still a bit of air owing out of the nose,
so-called nasal leak. As long as air is owing, this will facilitate voicing. There
is a second factor facilitating voicing throughout the nasal-plus-obstruent cluster.
Once the velum is raised high enough to cut off nasal leak, it continues to rise
toward the high position characteristic of obstruents. Consequently the volume
of the oral cavity expands, rarefying the air column sitting above the glottis.
This causes a small leak of air through the glottis, again facilitating prolongued
voicing. Both mechanisms, nasal leak and velum raising, cooperate to facilitate
continuous voicing throughout a nasal-obstruent cluster. In sum, a nasal-plus-
voiced obstruent cluster is more phonetically natural than a nasal-plus-voiceless-
obstruent cluster.
The post-nasal voicing effect is wide-spread in the worlds languages. More-
over, it occurs in child language (Drachman and Malikouti-Drachman 1973,
Smith 1973, Vogel 1976, cited in Pater forthcoming). The markedness constraint
*Nt (15) directly encodes the phonetic basis of the effect. Pater (forthcoming)
shows that *Nt produces a range of effects in various languages, of which nasal

2
There is a complementary perceptual explanation due to Ohala and Ohala (1993).

61
The typology of structural changes

substitution is only one. Moreover, the full range of effects is due to interactions
of *Nt with other (mostly faithfulness) constraints.

We rst consider the analysis of nasal substitution. In Indonesian, *Nt is ranked

sufciently high so as to disallow outputs containing a nasal plus voiceless
obstruent.3 Indonesian uses nasal substitution (or fusion) as a strategy to resolve
violations of *Nt. Clearly fusion involves some divergence from the input form,
hence a violation of some faithfulness constraint. To nd out which faithfulness
constraint is violated under fusion, let us consider the correspondence diagram of
the fusion of the input sequence /Np/:

(17) Correspondence diagram for fusion

Input: N1 p2

Output: m1,2

Fusion is a split correspondence between a pair of input segments and a single

output segment. This involves violation of some faithfulness constraint(s), as the
input segments /N, p/ do not stand in a perfect one-to-one relation with two output
segments. The broken one-to-one relation between the input and output has two
aspects: feature identity and segment correspondence.

With respect to feature identity it will be clear that the fusion of two non-identical
segments typically involves the violation of some Ident-IO constraints. This is
because a single output segment cannot preserve the feature specications of both
input segments if these are conicting. In a fusion of /Np/ into [m], for example,
featural specications of both input segments /N, p/ are preserved, except for the
manner features of input /p/. For /p/, the features specifying sonority class, nasal-
ity, and voicing fail to reoccur in output [m], leading to violation of various
Ident-IO constraints. On the other hand, some feature specications of the input
segments are preserved. Place of articulation of input /p/ is preserved in output
[m], as well as manner of input /N/. Nor is there any loss of identity for place of
articulation for input /N/, which is arguably unspecied for place features. On the
whole, we nd that violations of featural identity are minimal: they are limited to
manner of the input obstruent.
As Pater points out, OT predicts that segment fusion should be more likely
when the input segments are more similar: that is, when the violation of faithful-
ness constraints involved is relatively small. We will return to the role of featural

3
Below we will nd that Nt-sequences do occur root-internally in Indonesian. Then we will
address the issue of what makes nasal substitution apply only across morpheme boundaries.

62
 2.2 Nasal substitution and related effects

faithfulness in fusion in more detail in section 2.2.2. There we will also address
the question of what causes the obstruent (rather than the nasal) to give up its
manner specications.

The second aspect of fusion is pure segment correspondence, disregarding feature

content. Consider again the correspondence diagram of fusion in (17), in which
two input segments correspond to a single output segment. Observe that no input
segment is deleted since both input segments have output correspondents. Nor is
any segment inserted in the output (that is not present in the input). However,
some information is lost in the output that was still present in the input: the linear
ordering of the input segments.
Accordingly Pater adopts the following correspondence constraint:4

(18) Linearity-IO
The output reects the precedence structure of the input, and vice
versa.

This constraint is violated by diagram (17) in the following way. In the input the
segment /N/ precedes the segment /p/. Both segments share a correspondent [m]
in the output. But no precedence relation is dened on the output segment [m],
since it is not a sequence of segments but a single segment. (Of course, the
subscripts on input and output segments have nothing to do with linear ordering.)
The output has lost information that was present in the input (its precedence
structure) a clear violation of faithfulness.
Pater argues that Linearity-IO is independently needed in correspondence
theory as the constraint militating against metathesis. Metathesis is the phenom-
enon of two segments reversing their linear order:

(19) Correspondence diagram for metathesis

Input: C1 C2

Output: C2 C1

4
McCarthy and Prince (1995a) argue for a separate anti-coalescence constraint: Uniformity:
(i) Uniformity-IO
No element of the output has multiple correspondents in the input.
Contrary to Linearity, this makes no demands on the linear ordering of morphemes in the
input. This is a genuine advantage, as the assumption that morphemes are linearly ordered in
the input is problematic, as we will see in chapter 5 on reduplication. (For linear ordering of
morphemes may well depend on phonological information that is not present in the input.) See
also Lamontagne and Rice (1995) for analysis of coalescence in Correspondence Theory.

63
The typology of structural changes

Again, no segments are deleted or inserted, but the precedence structure is not
preserved in the output.

We now return to the analysis of nasal substitution in Indonesian and consider the
ranking of the constraints introduced so far. The very fact that Linearity-IO
is violated in nasal substitution (i.e. coalescence) gives evidence that *Nt ranks
above Linearity-IO. The ranking argument has the following logical form:

(20) Faithful  Coalesced Markedness
Faithfulness

candidate candidate
mm1p2ilih  mm1,2ilih therefore *Nt
Linearity-IO

That is, the faithful candidate [mm1p2ilih] is less harmonic than the coalesced
candidate [mm1,2ilih]. From this it follows that *Nt (the constraint that is vio-
lated in the losing candidate [mm1p2ilih]) dominates Linearity-IO (the con-
straint that is violated in the winning candidate [mm1,2ilih]).
This ranking argument is proved by the following tableau:

(21) Input: /mN1-p2ilih/ *Nt Linearity-IO

a. mm1p2ilih *!
b. + mm1,2ilih *

This is the basis of the analysis of nasal substitution: a contextual markedness

constraint (*Nt) dominates a faithfulness constraint (Linearity-IO).

2.2.2 Alternative strategies to satisfy *Nj

Next we consider the question of what logically possible strategies other than
nasal substitution might have satised *Nt. The alternative strategies are: dele-
tion, epenthesis, post-nasal voicing, and denasalization. We will nd that all of
these processes, except epenthesis, are attested in some language other than Indo-
nesian. In a rule format, these processes are given below:

(22) a. Deletion [+nas] / ___ [voice]

b. Epenthesis / [+nas] ___ [voice]
c. Post-nasal voicing [son] [+voice] / [+nas] ___
d. Denasalization [+nas] [nas] / ___ [voice]

Written in the standard rule format, these four processes seem to have little in
common. But a generalization lurks beneath the surface: all four processes apply
in the context of a nasal plus voiceless consonant. To put it even more strongly: all
four processes have the effect of undoing this context: they avoid the violation
64
 2.2 Nasal substitution and related effects

of *Nt. However, this functional unity is left unexpressed in the rule format.
Consider the following hypothetical rule, applying in the same context of nasal
plus voiceless consonant:

(23) Post-nasal spirantization [voice] [+continuant] / [+nas] ___

Judging by its context of application, this hypothetical rule falls into the same
class of rules as those in (22). Yet it has a completely different status, since it
fails to avoid a violation of *Nt. In order to express the typologically favoured
status of these rules over other rules that apply in the same context, such as (23),
the markedness criterion of rule-based theory must have the power to detect the
output goals of rules. However, the goal of a rule can be detected only by taking
into consideration both its structural context (here: Nt) and its structural change
(here: undoing Nt). To detect a rules goal is, by itself, in contradiction to the
fundamental assumption of rule-based theory that rules apply blindly, ignorant of
their eventual effects at the surface. Building surface-sensitivity into rules would
move rule-based theory closer to surface-oriented OT see again section 2.1.2.1.
The fact that all four processes in (22) except vowel epenthesis are typologic-
ally attested is by itself sufcient reason to favour a theory which expresses their
functional unity. But an even stronger case is made by Pater on the basis of the
observation that some of these processes cooccur within a single language. While
the occurrence of functionally related rules in different languages might still be
taken as coincidental, their cooccurrence in the same language is an even stronger
argument for OT. These cases will be discussed in section 2.3.

The actual response to *Nt in a particular language follows from the interac-
tion of *Nt with the set of faithfulness constraints. The logic of this interaction
can be stated as follows. Assume some high-ranked markedness constraint, such
as *Nt. To satisfy it, a change is necessary, which will inevitably violate some
faithfulness constraint. By itself the fact that some change is to be made does not
predict which change will be made. But we know beforehand that the grammar
will militate against any change regardless of what it may be, simply because any
change will violate at least some faithfulness constraint. It may seem as if every
grammar makes the following contradictory requirement:

(24) Repair the conguration *Nt, but

do not change a segment (nasal or obstruent) into another segment;
do not fuse two segments (nasal and obstruent) into a single
segment;
do not delete a segment (nasal or obstruent);
do not insert a segment.
65
The typology of structural changes

This is like making the impossible requirement: remove the stain from the carpet,
but do not use water, or foam, or acid, or scissors, or . . . How to break the spell?
However, the requirement in (24) is contradictory only when each of the logic-
ally possible changes is equally expensive in terms of violations of faithful-
ness. Of course, in OT structural changes are never equally expensive, due to
strict domination. Therefore, the relative ranking of the faithfulness constraints
determines the change which serves as repair of *Nt. This is the central insight
which lies at the heart of OTs ability to deal with functionally related processes
that is, with process typology.

Now let us assume a schematic hierarchy as in (25). The top-ranking markedness

constraint *Nt is, potentially at least, satisable by the violation of any constraint
of the set {Faith-IO1 . . . Faith-IOn}:

(25) *Nt, Faith-IO1
Faith-IO2 . . .
Faith-IOn

Violation must be minimal, hence the grammar will select output forms which
satisfy the higher-ranking constraints, at the expense of violations of lower-
ranking constraints. This means that the lowest-ranking faithfulness constraint
will be violated. In sum, the specic method which a language selects to satisfy
some high-ranking markedness constraint only depends on the lowest-ranking
faithfulness constraint. (In terms of the stained carpet analogue used above, one
might say that using scissors or acid to remove the stain would ruin the carpet,
outweighing the gains, so that the least expensive option is using foam, at least if
water does not do the job.)
Next observe that the relative ranking of the faithfulness constraints that dom-
inate Faith-IOn is immaterial to the outcome. The violation of Faith-IOn alone
sufces to avoid violation of all other constraints; this is to say that no further
violations of higher-ranking faithfulness constraints are necessary. Since unnec-
essary violations are always avoided, none of the faithfulness constraints domin-
ating Faith-IOn will in fact be violated. This, in its turn, means that we will
typically nd no positive evidence for the mutual ranking of the set of higher-
ranking faithfulness constraints. Positive evidence for ranking is always contin-
gent upon the violation of a constraint in the actual form. Therefore all we know
is that Faith-IOn is ranked below all the other faithfulness constraints. The ranking
of the set of remaining faithfulness constraints cannot be established.
For Indonesian we already know that the dominated constraint Faith-IOn must
be Linearity-IO. We now substitute Linearity-IO in (25), and also replace the
abstract set of faithfulness constraints {Faith-IO1, Faith-IO2 . . .} by the set of
informal faithfulness constraints of (24). After doing this, we arrive at:

66
 2.2 Nasal substitution and related effects

(26) *Nt, No-deletion, No-epenthesis, No-denasalization, No-voicing

Linearity-IO

But this is, of course, not the end of the story. We must now formalize the
constraints that are loosely referred to as No-deletion, No-epenthesis, etc. in
(26) as correspondence constraints, which militate against divergences of input
and output along one dimension. Informally, the general format of this type of
constraint is as follows:

(27) Correspondence constraint: the output equals the input for some
property P.

We already know two correspondence constraints from chapter 1: Ident-

IO(voice) and Ident-IO(nasal), both members of the Identity family of con-
straints. In both of these constraints the relevant property P was featural identity
of corresponding segments. We now see that Linearity-IO is another example
of a correspondence constraint, matching linear order of corresponding segments.
The anti-deletion and anti-epenthesis statements of (24) require new types of
correspondence constraints. Let us now see what these are.

2.2.2.1 Deletion and MAX-IO

The rst strategy that Indonesian might have employed to satisfy *Nt is to simply
delete one of the input segments (nasal or voiceless obstruent). For example,
the deletion of the nasal in /mN1-p2ilih/ would produce an unattested output
*[mp2ilih], where *Nt would be (trivially) satised since the nasal is no longer
present. However, such deletion would violate the correspondence constraint
Max-IO.

(28) Max-IO
Input segments must have output correspondents.
(No deletion)

The violation of Max-IO in the suboptimal output [mp2ilih] is illustrated by the

diagram below. Note that input /N1/ lacks an output correspondent:

(29) Correspondence diagram for nasal deletion

Input: N1 p2

Output: p2

The very fact that this strategy is left unemployed in Indonesian points to the
conclusion that Max-IO dominates Linearity-IO:

67
The typology of structural changes

(30) Max-IO
Linearity-IO

mp2ilih  mm1,2ilih

That is, the preservation of input segments in the output has a higher priority than
the preservation of precedence relations of the input. This ranking argument is
proved correct by tableau (31):5

(31) Input: /mN1-p2ilih/ Max-IO *Nt Linearity-IO

a. mp2ilih *!
b. + mm1,2ilih *

We simply add Max-IO to the set of undominated constraints, since there is no

evidence for its ranking with respect to *Nt. (Since we focus here on the candid-
ates relevant to the ranking of Max-IO and Linearity-IO, we omit the fully
faithful analysis [mm1p2ilih] from this tableau and those that follow. This is of
course ruled out by undominated *Nt.) Later we will see that some languages
actually select nasal deletion as a strategy to satisfy *Nt. The difference between
Indonesian and these deleting languages will be construed as a difference in
constraint ranking.

2.2.2.2 Epenthesis and DEP-IO

A second strategy to satisfy *Nt, one which simultaneously satises Linearity-
IO, is epenthesis of a segment between the nasal and obstruent, undoing their
adjacency. With schwa as epenthetic vowel, this would produce an incorrect out-
put *[m1p2ilih]. Epenthesis violates another important segment correspond-
ence constraint, Dep-IO. This militates against segments in the output that have
no correspondents in the input.

(32) Dep-IO
Output segments must have input correspondents.
(No epenthesis)

A diagram illustrates the violation of Dep-IO in [m1p2ilih]:

5
To complete formally the argument that Max-IO dominates Linearity-IO, we must show that
a third candidate [mp2ilih], which deletes input / N/, does not violate Linearity-IO. (If both
31a and 31b were to violate Linearity-IO, then Max-IO would be decisive regardless of its
ranking.) On a more formal interpretation of Linearity-IO, this constraint cannot be violated
under deletion. That is, precedence relations can only be distinct if both input segments have
correspondents in the output.

68
 2.2 Nasal substitution and related effects

(33) Correspondence diagram for schwa epenthesis

Input: N1 p2

Output: 1 p2
The output segment [] lacks an input correspondent, in violation of Dep-IO.
The observation that the epenthetic output *[m1p2ilih] is ill-formed again
leads to a ranking argument of a type that is now familiar:
(34) Dep-IO
Linearity-IO
m1p2ilih  mm1,2ilih
Tableau (35) conrms the correctness of this ranking:

(35) Input: /mN1-p2ilih/ Dep-IO *Nt Linearity-IO

a. m1p2ilih *!
b. + mm1,2ilih *

We therefore add yet another correspondence constraint (Dep-IO) to the set of

undominated constraints.

2.2.2.3 Post-nasal voicing and IDENT-IO(voice)

Both alternatives to coalescence that we have discussed above, deletion and epen-
thesis, involved pure segment correspondence, and no featural identity. How-
ever, considering the full range of possible repairs of *Nt leads us to segmental
changes, that is, into issues of featural identity. The third alternative strategy to
satisfy *Nt involves such a featural change of input segments. This strategy is
post-nasal voicing. The output *[mm1b2ilih], which results from post-nasal voic-
ing, violates identity with respect to the feature [voice], since this is the single
difference between input /p/ and its output correspondent [b].
(36) Correspondence diagram for post-nasal voicing
Input: N1p2 [voice]

Output: m1b2 [+voice]

However, we cannot identify the relevant constraint as Ident-IO(voice), for the
following reason. If it were indeed Ident-IO(voice) that ruled out a candidate
*[mm1b2ilih], then this constraint would incorrectly rule out the (actual) coa-
lesced form [mm1,2ilih] as well. Here voiceless /p/ in the input corresponds with
voiced [m] in the output, which produces a (phonetic) voicing contrast between
both correspondents:
69
The typology of structural changes

(37) Correspondence diagram for fusion

Input: N1p2 [voice]

Output: m1,2 [+voice]

Pater proposes that the relevant identity constraint is more specic than Ident-
IO(voice), in the sense that it is restricted to voicing in obstruents. This more
specic constraint is Ident-IO(ObstruentVoice):
(38) Ident-IO(ObsVce)
Correspondent obstruents are identical in their specication for
voice.
(No changes in the voicing of obstruents)
Note that Ident-IO(ObsVce) applies equally well in the analysis of devoicing
that was proposed in section 2.1.4, where Ident-IO(voice) functioned to militate
against alternations of voice in obstruents.6
In Indonesian Ident-IO(ObsVce) clearly dominates Linearity-IO:
(39) Ident-IO(ObsVce)
Linearity-IO
mm1b2ilih  mm1,2ilih

The relevant candidates are evaluated by the following tableau:

(40) Input: /mN1-p2ilih/ Ident-IO *Nt Linearity-IO

(ObsVce)

a. mm1b2ilih *!
b. + mm1,2ilih *

2.2.2.4 Denasalization and IDENT-IO(nasal)

The nal alternative (to nasal substitution) that we will consider is denasalization.
By changing the input nasal of /Np/ into an oral consonant, producing [pp], the
violation of *Nt is equally well avoided. This strategy, as applied to our example,
would produce an unattested output [mp1p2ilih]. This would violate a featural
identity constraint Ident-IO(nasal) requiring that corresponding segments have
the same specication for [nasal]. See the diagram below:
6
It may seem that we can dispense with reference to obstruents by Ident-IO(ObsVce) if we
adopt contrastive underspecication (Steriade 1987), a theory in which features are only
specied in segment types in which they are contrastive, for example, [voice] in obstruents.
However, we would still need some constraint militating against the neutralization of voicing
contrasts in sonorants, since languages exist that have contrastive voicing in nasals, such as
Burmese (Ladefoged and Maddieson 1996).

70
 2.2 Nasal substitution and related effects

(41) Correspondence diagram for denasalization

Input: [+nasal] N1 p2

Output: [nasal] p1 p2

Again a subtle technical problem arises that must be solved to make the analysis
work. If denasalization is to be effectively excluded by Ident-IO(nasal), then this
constraint must not be violated by the optimal coalescence candidate [mm1,2ilih].
We repeat the diagram for coalescence with the relevant feature specications for
[nasal] added:

(42) Correspondence diagram for fusion

Input: [+nasal] N1p2 [nasal]

Output: [+nasal] m1,2

Apparently coalescence violates Ident-IO(nasal) with respect to the input

obstruent /p/, whose specication [nasal] fails to reoccur in the output.

Pater argues, though, that there is a major difference between denasalization (41)
and fusion (42). It is known from many processes involving [nasal] that only the
positive specication of this feature is ever phonologically active. This observa-
tion has led several researchers (Cohn 1993b, Piggott 1993, Rice 1993, Steriade
1993, Trigo 1993) to assume that [nasal] is a monovalent feature, one that has
only one (positive) value. Under denasalization, [nasal] is lost from input /N/ (see
43a). But under nasal substitution there is no such loss of [nasal] from input /N/,
nor from the input obstruent (see 43b).

(43) Correspondence diagrams for denasalization and fusion, with mono-

valent [nasal]
a. Input: [nasal] N1 p2 b. Input: [nasal] N1p2

Output: p1 p2 Output: [nasal] m1,2

At worst, nasal substitution involves an insertion (rather than loss) of [nasal]:

input /p2/ lacks [nasal], while its output correspondent [m2] is specied as [nasal].
(Arguably this is not a faithfulness violation at all because the feature [nasal] in
output [m] is not inserted in any sense: it simply corresponds to the nasality of
one of [m]s input correspondents.)
Pater proposes to capture the distinction between the deletion and insertion
of a feature in the format of featural identity constraints. This will now include a
71
The typology of structural changes

direction. For example, if the input specication of some feature must be pre-
served in the output, then the direction is from I to O, as in (44):

(44) Ident-IO(nasal)
Any correspondent of an input segment specied as F must be F.
(No denasalization)

This constraint militates against denasalization. Since denasalization is not

employed as a strategy to avoid violations of *Nt in Indonesian, Ident-
IO(nasal) must dominate the constraint that is violated under pressure of *Nt.
This is Linearity-IO. Again, we arrive at a ranking argument similar to the one
we encountered on various earlier occasions:

(45) Ident-IO(nasal)
Linearity-IO

mp1p2ilih  mm1,2ilih

The correctness of this ranking appears from the following tableau:

(46) Input: /mN1-p2ilih/ Ident-IO *Nt Linearity-IO

(nasal)

a. mp1p2ilih *!
b. + mm1,2ilih *

As we will see in section 2.3.4, denasalization is employed as a strategy to satisfy

*Nt in a number of languages.

2.2.2.5 A summary of *Ni effects

We have now discussed four alternative strategies to satisfy *Nt (or *Nt effects)
which, logically speaking, Indonesian might have employed instead of nasal sub-
stitution. Results so far are summarized in the following table:

(47) Candidate Strategy employed Ranking

a. mm1,2ilih Nasal substitution *Nt
Linearity-IO
b. *mm1b2ilih Post-nasal voicing Ident-IO(ObsVce)

Linearity-IO
c. *mp1p2ilih Denasalization Ident-IO(nasal)

Linearity-IO
d. *mp2ilih Nasal deletion Max-IO
Linearity-IO
e. *m1p2ilih Vowel epenthesis Dep-IO
Linearity-IO

This produces the following ranking for Indonesian nasal substitution:

72
 2.2 Nasal substitution and related effects

(48) Indonesian nasal substitution

*Nt, Dep-IO, Max-IO, Ident-IO(nasal), Ident-IO(ObsVce)

Linearity-IO

Tableau (49) evaluates all candidates discussed thus far:7

(49)
Input: /mN1-p2ilih/ *Nt Dep Max- Ident-IO Ident-IO Linearity-
-IO IO (nasal) (ObsVce) IO

a. + mm1,2ilih *
b. mm1b2ilih *!
c. mp1p2ilih *!
d. mp2ilih *!
e. m1p2ilih *!
f. mm1p2ilih *!

Of course, the fact that each candidate violates precisely one constraint is not a
logical necessity, but a restriction on the candidate set made for presentational
reasons only. Gen is free to propose any candidate analyses violating more than
a single constraint of the set in tableau (49). These candidates would, of course,
be suboptimal to those in (49).
For example, in addition to the candidate set evaluated in (49) we may consider
a candidate having violations of both Linearity-IO and Ident-IO(nasal). This
candidate would have both coalescence and loss of [nasal] from the input, as in
*[mp1,2ilih] (50c). We might even consider another candidate having, on top of
violations of Linearity-IO and Ident-IO(nasal), an additional violation of
Ident-IO(ObsVce), as in *[mb1,2ilih] (50d). These overly-unfaithful candidates
are simply excluded by the current hierarchy:

7
Yet another candidate, discussed by Pater at the end of his paper, involves nasalization of the
post-nasal obstruent, as in *[mm1m2ilih] with a geminate nasal. Given the current constraint
ranking this candidate is actually optimal, as it has no violations of any constraints in tableau
(49). (According to Pater this pattern is attested in Konjo, where it is partially conditioned by
the type of prex.) Note that this gemination candidate [mm1m2ilih] minimally differs from
the fusion candidate [mm1,2ilih] in the following way. In the former, the output segment [m2]
is specied as [nasal], contrary to its input correspondent /p/. In the fusion candidate, the
specication [nasal] of the output segment [m1,2] reoccurs in one of its input correspondents,
/N1/. We therefore assume that the gemination candidate is ruled out by a constraint militating
against [nasal] in an output segment whose input correspondent is not specied as [nasal]. A
constraint that has this effect is Ident-OI(nasal), that is, the reverse of Ident-IO(nasal).

73
The typology of structural changes

(50)
Input: /mN1-p2ilih / *Nt Dep Max- Ident-IO Ident-IO Linearity-
-IO IO (nasal) (ObsVce) IO

a. + mm1,2ilih *
b. mp1p2ilih *!
c. mp1,2ilih *! *
d. mb1,2ilih *! * *

Why even pay attention to these candidates? Is it not merely logical that the
worst of the worst is excluded? But actually we have arrived at an important
result here: OT predicts that the simplest change should be made to avoid the
violation of some markedness constraint. Although two, or in fact any possible
number of changes could be made in order to avoid violation of *Nt, only one
change sufces: nasal substitution (or coalescence). No more changes occur than
are strictly necessary. Prince and Smolensky (1993) have stated this as the Eco-
nomy principle of OT, repeated below from chapter 1 (29):
(51) Economy: banned options are available only to avoid violations of
higher-ranked constraints and can only be banned minimally.
In chapter 1 we have already arrived at a similar result in the discussion of
devoicing in Dutch, but then we did not take into account any alternative strat-
egies of avoiding violation of obstruent voicing in codas. Here we have found
evidence for the economy property from a case that involves a consideration of
various possible changes, rather than just one.

Let us now nd out if rule-based theory makes the same prediction of economy.
The notion of faithfulness is absent from rule-based theory altogether, so it is
difcult to see how application of a rule might be blocked in a case where an
earlier rule has already accomplished a surface target. However, rule-based the-
ory happens to arrive at the same prediction of economy for rules sharing the
same structural description. This prediction follows from rule ordering. Consider,
for example, the functionally related rules of Post-nasal voicing and Denasaliza-
tion (before voiceless obstruents). Whichever applies rst will destroy the context
of application for the other rule:
(52) Mutual blocking of functionally related rules by ordering
a. Input: /mp/ b. Input: /mp/
Post-nasal voicing mb Denasalization pp
Denasalization --- Post-nasal voicing ---
Output: [mb] Output: [pp]
74
 2.2 Nasal substitution and related effects

Both cases exemplify bleeding rule orders: one rule destroys the context of
application for another rule. Rule-ordering is indeed crucial to economical out-
puts, as simultaneous application of both rules would produce *[bb], a non-
economical output. The equivalence of rule-ordering theory and OT with respect
to the mutual blocking of functionally related processes is not generally extend-
able to all cases of economy, however. In later chapters (for example, chapter 8
on syntax), we will nd that OT offers a more principled account of economy
phenomena than rule-based theory. But also see chapter 9 for a discussion of
opacity, a phenomenon which apparently falsies economy.

The next subsection will rene the analysis of nasal substitution in Indonesian by
showing the relevance of morphological context. This discussion will be based on
data of nasal plus obstruent sequences which we have not looked into so far.

2.2.3 Morphological effects

2.2.3.1 Root-internal blocking

All cases thus far involved a nasal in a prex before a root-initial obstruent. Pater
points out, however, that Indonesian does not ban sequences of nasal plus voice-
less consonant across the board. Consider the following root-internal occurrences
of this sequence:

(53) Root-internal occurrences of Nt in Indonesian

a. mpat four
b. untuk for
c. mukin possible

The same sequence that is excluded across a prexroot boundary8 is apparently

allowed root-internally. Pater observes that the root-internal prohibition against
nasal substitution exemplies a cross-linguistic tendency for root-internal seg-
ments to be more resistant to phonological processes than segments in other
positions. There is a well-known class of processes that apply only across mor-
pheme boundaries, but fail to apply inside a morpheme, displaying the derived
environment effect (Mascar 1976, Kiparsky 1982b, 1993a).
To account for this cross-linguistic tendency in OT, it has been recently pro-
posed that faithfulness requirements are enforced more strictly within the root
than in non-root morphemes, such as afxes (McCarthy and Prince 1995a). This

8
Nasal substitution does not apply across all prexroot junctures. Certain prex boundaries are
opaque to nasal substitution, for example /mN-pr-besar/ [mmprbesar] to enlarge.

75
The typology of structural changes

is captured by root-specic versions of faithfulness constraints, ranked above

general versions of these constraints.9

(54) Root-Faithfulness
Faithfulness

Application of a phonological process across a morpheme boundary shows that

general faithfulness is ranked below some markedness constraint. And non-
application in a root-internal context shows that root-faithfulness dominates this
markedness constraint. We thus nd sandwiching of a markedness constraint
between a root-particular faithfulness constraint and the general faithfulness
constraint:

(55) Non-application within the root, application elsewhere in the word

Root-Faithfulness
Markedness
Faithfulness

This is the general ranking schema for non-application of a process within a root.
(When the markedness constraint would dominate both others, no root-effect
would occur.)
Pater goes on to propose a root-particular version of Linearity-IO:

(56) RootLin-IO
The output reects the precedence structure of the input segments of
the root, and vice versa.

Sandwiching of the phonotactic constraint *Nt in between RootLin-IO and

Linearity-IO produces the following (partial) ranking:

(57) Blocking of root-internal fusion

RootLin-IO
*Nt
Linearity-IO

This ranking correctly accounts for the blocking of root-internal fusion, as shown
in the following tableau:

(58) Input: /m1p2at/ RootLin-IO *Nt Linearity-IO

a. + m1p2at *
b. m1,2at *! *

9
This ranking has been argued to be universal, but this is difcult to establish. It is only when
root-faithfulness (the more specic constraint) dominates the general constraint that any root-
faithfulness effects become apparent. Under the reverse ranking, with the general constraint
taking priority, root-faithfulness is simply suppressed.

76
 2.2 Nasal substitution and related effects

Finally we need to integrate this partial ranking with the other correspondence
constraints argued for in the previous section. This in fact is quite simple. Within
roots no alternative strategies (neither deletion, epenthesis, post-nasal voicing, nor
denasalization) are ever employed to avoid violation of *Nt. Hence, all corres-
pondence constraints militating against these strategies dominate *Nt. The com-
plete ranking is given in (59).

(59) Indonesian nasal substitution, nal ranking

Dep-IO, Max-IO, Ident-IO(nasal), Ident-IO(ObsVce),
RootLin-IO
*Nt
Linearity-IO

2.2.3.2 An alternative analysis

It may be worth exploring an alternative to this analysis that is not based on root-
specic faithfulness constraints, but rather on the notion that morphemes are not
linearly ordered in the input. (See also note 4.) On this account, there is no input
linear ordering between the prex and the root, hence no linear ordering is present
in the input between the prex nasal and root obstruent. In contrast, all root-
internal occurrences of nasal and obstruent do have an input linear ordering. What
blocks nasal substitution in root-internal contexts is Linearity-IO, which must
therefore be ranked above *Nt.

(60) Input: /m1p2at / Linearity-IO *Nt

a. + m1p2at *
b. m1,2at *!

Nasal substitution across a prexroot boundary cannot involve violation of

Linearity, but it does violate the correspondence constraint Uniformity-IO
(McCarthy and Prince 1995a).

(61) Uniformity-IO
No element of the output has multiple correspondents in the input.
(No coalescence)

This anti-coalescence constraint must be dominated by *Nt, as tableau (62)

shows:

(62) Input: /mN1, p2ilih/ Linearity-IO *Nt Uniformity-IO

a. mm1p2ilih *!
b. + mm1,2ilih *

77
The typology of structural changes

Let us now balance the pros and cons of this analysis. On the positive side, it
avoids root-specic faithfulness constraints, thereby restricting the machinery
of OT. Root-particular faithfulness constraints may eventually turn out to be
independently required but the null-hypothesis is that faithfulness constraints
are blind to morphological structure. Another advantage is avoidance of the
assumption that morphemes must be linearly ordered in the input. (In chapter 5
we will see that this assumption is problematic because of afxations whose
linear position with respect to the stem depends on output phonological structure,
such as inxation and reduplication.) On the negative side, it sets up a corres-
pondence constraint for coalescence, losing the unied account of coalescence
and metathesis.

2.3 The typology of *Nh effects

This section will deal with the cross-linguistic status of *Nt and strategies that
languages may employ to resolve its violation. Recall from section 2.2 that the
type of change (in response to some phonotactic constraint) depends on the rank-
ing of correspondence constraints militating against specic changes. Specically,
the observed effect involves violation of the lowest-ranking correspondence con-
straint in the hierarchy. This constraint is crucially dominated by the markedness
constraint (the trigger), as well as by other faithfulness constraints, whose vio-
lation is more expensive. For example, nasal substitution emerges if Linearity-
IO is lowest-ranking, nasal deletion if Max-IO is lowest-ranking, etc. Abstracting
away from interactions with other faithfulness constraints in the hierarchy, we
arrive at the following typology of *Nt effects:

(63) A typology of *Nt effects

a. Nasal substitution *Nt
Linearity-IO No coalescence
b. Nasal deletion *Nt
Max-IO No deletion
c. Vowel epenthesis *Nt
Dep-IO No epenthesis
d. Post-nasal voicing *Nt
Ident-IO(ObsVce)
No obstruent
voicing
e. Denasalization *Nt
Ident-IO(nasal) No
denasalization

The theory of correspondence constraints is, in a certain sense, the theory of

phonological changes. It predicts the possible effects that may occur cross-
linguistically in response to phonotactic constraints. What we are specically
interested in in this section is whether the full range of *Nt effects predicted by
this typology is indeed attested cross-linguistically. As we will see, the empirical
coverage is quite high: all *Nt effects occur except (vowel) epenthesis.

78
 2.3 The typology of *Nh
*Ni effects

2.3.1 Nasal deletion in Kelantan Malay

The rst *Nt effect to be discussed is nasal deletion, the deletion of a nasal
before a voiceless consonant. This avoids violation of *Nt by simply removing
the segment that induces voicing in the following obstruent.
Among the languages which Pater mentions as examples of nasal deletion
are the Kelantan dialect of Malay, and the African languages Venda, Swahili,
and Maore. Nasal deletion has also been reported for child language, in English,
Greek, and Spanish. All in all, this is an impressive list of cases, which gives
additional force to the observation that it is always the nasal that is deleted, rather
than the obstruent.10 This is an issue about which Pater makes no conclusive
statements, but only some interesting suggestions.11
Pater gives no examples of nasal deletion in Kelantan-like languages, and
hence we will follow his schematic account. An input sequence of a nasal (N1)
plus a voiceless consonant (T2) maps onto a single output segment that is identical
to T2. That is, an input segment N1 has no output correspondent:

(64) Correspondence diagram for nasal deletion

Input: N1T2

Output: T2
This involves the violation of Max-IO. Hence we may infer that Max-IO is the
lowest-ranking faithfulness constraint in the hierarchy:

(65) Ranking for nasal deletion in Kelantan-like languages

*Nt, Dep-IO, Ident-IO(nasal), Ident-IO(ObsVce), Linearity-
IO
Max-IO

This ranking predicts nasal deletion, as illustrated below for the candidate set
previously evaluated in the Indonesian tableau (49). Participating segments are
indicated by symbols (N = nasal, T = voiceless obstruent, D = voiced
obstruent).

10
Recall that the Indonesian case is analysed as a fusion rather than the deletion of the obstruent
on the basis of various kinds of evidence: the stability of the obstruents place of articulation,
as well as the behaviour of substitutions under reduplication.
11
Pater suggests that nasals are weaker than obstruents in various respects. First, nasals tend
to assimilate the place of articulation of the obstruent, rather than the reverse. Second, the
deletion of nasals in Nt (rather than obstruents) suggests a similar weakness. Apparently,
faithfulness requirements for obstruents are stricter than those for nasals. To implement this
idea, Pater proposes to relativize faithfulness constraints with respect to major class, branch-
ing out Max-IO into ObsMax (No deletion of obstruents) and NasMax (No deletion of
nasals), with the former universally outranking the former.

79
The typology of structural changes

(66)
Input: / N1 T2/ *Nt Dep Ident-IO Ident-IO Linearity- Max-
-IO (nasal) (ObsVce) IO IO

a. N1,2 *!
b. N1 D2 *!
c. T1 T2 *!
d. + T2 *
e. N1T2 *!
f. N1 T2 *!

2.3.2 Vowel epenthesis

The second predicted *Nt effect is vowel epenthesis, the insertion of a segment
between the nasal and voiceless consonant to undo their adjacency. Pater reports
that this *Nt effect is not attested in any natural language, as far is currently
known. He makes various suggestions towards an explanation of this potential
gap in the typology, for which we refer the reader to the paper.

2.3.3 Post-nasal voicing in Puyu Pungo Quechua

The third predicted *Nt effect is post-nasal voicing, the neutralization of the
voicing contrast in obstruents following nasals. According to Pater, post-nasal
voicing is by far the most common of all *Nt effects. It occurs in various lan-
guages ranging from Zoque to the Puyu Pungo dialect of Quechua. The pattern
of post-nasal voicing in Puyu Pungo Quechua (Orr 1962, Rice 1993, as cited in
Pater forthcoming) is exemplied below.
(67) Post-nasal voicing in Puyu Pungo Quechua
a.i sinik-pa porcupines a.ii kam-ba yours
b.i wasi-ta in the jungle b.ii wakin-da the others
Post-nasal voicing amounts to the following correspondence diagram:
(68) Correspondence diagram for post-nasal voicing
Input: m1p2 [voice]

Output: m1b2 [+voice]

Different values of the feature [voice] in the input obstruent and its output corres-
pondent incur a violation of the featural identity constraint Ident-IO(ObsVce).
Accordingly, this must be the lowest-ranking faithfulness constraint in the hierarchy:
80
 2.3 The typology of *Nh
*Ni effects

(69) Ranking for post-nasal voicing in Puyu Pungo-like languages

*Nt, Dep-IO, Max-IO, Ident-IO (nasal), Linearity-IO

Ident-IO(ObsVce)

Tableau (70) evaluates various candidates that match the alternative *Nt effects.

(70)
Input: / kam1-p2a/ *Nt Dep Max- Ident- Linearity- Ident-IO
-IO IO IO(nasal) IO (ObsVce)

a. kam1,2a *!
b. + kam1b2a *
c. kap2p2a *!
d. kap2a *!
e. kam1p2a *!
f. kam1p2a *!

Puyo Pungo Quechua shares with Indonesian the phenomenon that the ban on
Nt sequences is relaxed root-internally. (An example presented by Pater is [9iki]
soot.) Of course this points to the undominated position of RootLin-IO, sim-
ilarly to Indonesian.

2.3.4 Denasalization in Mandar

The fourth and nal *Nt effect to be discussed is denasalization of the nasal
preceding the voiceless obstruent. Pater reports that denasalization is attested in
at least three languages: Toba Batak, Kaingang, and Mandar. The Mandar pattern
(Mills 1975, as cited in Pater forthcoming) is exemplied below:

(71) Denasalization in Mandar

a. /maN-dundu / mandundu to drink
b. /maN-tunu / mattunu to burn

Denasalization involves a loss of the feature [+nasal], which is present on the

input nasal, in its output correspondent, an obstruent.

(72) Correspondence diagram for post-nasal voicing

Input: [+nasal] N1 t2

Output: [nasal] t1 t2
81
The typology of structural changes

This amounts to violation of Ident-IO(nasal). Hence this must be the lowest-

ranking faithfulness constraint in the hierarchy:

(73) Ranking for denasalization in Mandar-like languages

*Nt, Dep-IO, Max-IO, Ident-IO(ObsVce), Linearity-IO

Ident-IO(nasal)

That denasalization is indeed the correct outcome of constraint evaluation is

shown in the following tableau:

(74)
Input: /maN-tunu / *Nt Dep Max- Ident-IO Linearity- Ident-
-IO IO (ObsVce) IO IO(nasal)

a. man1,2unu *!
b. man1d2unu *!
c. + mat1t2unu *
d. mat2unu *!
e. man1t2unu *!
f. man1t2unu *!

Mandar, unlike Indonesian and Puyo Pungo Quechua, completely disallows

Nt sequences, extending the ban to root-internal contexts (where post-nasal
obstruents are always voiced, resulting in a lack of alternations). In Paters
approach to root-specic faithfulness, this property of Mandar can be accounted
for by a reranking of *Nt, so that it comes to dominate RootLin-IO.

2.3.5 The typology of *Nj effects: conclusions

The conclusion of this typology section is that theoretical predictions about *Nt
effects are (mostly) correct. All predicted effects, with the exception of vowel
epenthesis, are attested in some language. The architecture of OT brings out the
functional unity of these effects: all are resolutions of a potential violation of a
phonotactic constraint, *Nt, militating against sequences of nasal plus voiceless
obstruent. Correspondence Theory has the ability of predicting typologies, factor-
ing out universally possible responses to some phonotactic constraint. Whether
or not a language employs a specic effect as a response to *Nt depends on a
language-specic ranking of universal correspondence constraints. The nature of
the *Nt effect shows which correspondence constraint is violated; this in its turn
pinpoints the correspondence constraint which is violated as the lowest ranking
82
 2.4 Conspiracies of nasal substitution and other processes

in the hierarchy. Finally, we repeat the conclusion from section 2.2.2 that OT
predicts that languages make economical use of the available repair strategies.
This prediction follows from the minimization of violation of constraints, one of
the corner stones of OT.
In the next section we will strengthen these conclusions by a discussion of the
co-occurrence of multiple *Nt effects within a single language.

2.4 Conspiracies of nasal substitution and other processes

So far we have discussed cases in which *Nt is satised by a single process per
language (with the option of blocking this process root-internally). The argument
made on the basis of these cases is that OT captures the functional unity of these
processes by separating the trigger (*Nt) and the effect (the relative ranking of
faithfulness constraints). Here we will consider various strategies to satisfy *Nt
within a single language, OshiKwanyama, a western Bantu language (Steinbergs
1985, as cited in Pater forthcoming).
The rst *Nt effect is a phonotactic pattern which holds as a static general-
ization for all roots of the language: post-nasal voicing is a morpheme-structure
constraint. But a dynamic view of post-nasal voicing is warranted by patterns of
loan words from English:12

(75) Post-nasal voicing in OshiKwanyama loan words

a. sitamba stamp
b. pelanda print
c. oinga ink

OshiKwanyama loan words display no voicing alternations in stops to directly

support the process interpretation, as the nasal-plus-stop cluster occurs internal to
the root, rather than across a morpheme boundary. Nevertheless a dynamic view
is supported by the fact that post-nasal voicing is productive: it affects loan words
having a nasal-plus-voiceless-stop cluster in their source language, amounting to
a positional neutralization.
The other *Nt effect found in OshiKwanyama is nasal substitution. This
occurs in root-initial position, following a prex ending in a nasal (precisely as
in Indonesian):

(76) Root-initial nasal substitution in OshiKwanyama

a. /eN-pati/ emati ribs
b. /oN-pote/ omote good-for-nothing
c. /oN-tana/ onana calf

12
For approaches to loan word phonology in OT, see Yip (1993) and It and Mester (1995).

83
The typology of structural changes

This process is supported by alternations, naturally, because it applies across a

morpheme boundary.
There is, of course, a clear functional unity between post-nasal voicing and
nasal substitution. Both serve to resolve violations of *Nt. Such functional unity
of processes within a language is known as a conspiracy (Kisseberth 1970) and,
as we learned in section 2.1.2, rule-based theory cannot formally capture this
situation. Observe that even under a static (rather than dynamic) interpretation
of post-nasal voicing, its functional unity with nasal substitution still remains to
be accounted for. (In that case, OshiKwanyama can be argued to exemplify the
duplication problem.)
Pater gives a straightforward analysis of the OshiKwanyama *Nt conspiracy,
one which uses no other constraints than those independently motivated by pre-
vious cases. In a nutshell, the analytic problem that must be resolved is the
following. So far we have assumed that the nature of the effect was tantamount
to pinpointing the dominated correspondence constraint in the hierarchy. But here
we have two *Nt effects, and we naturally wonder if both can be related to a
violation of some correspondence constraint without running into a constraint
ranking paradox. Post-nasal voicing violates Ident-IO(ObsVce), while nasal sub-
stitution violates Linearity-IO. But how could both (at the same time) be the
lowest-ranking correspondence constraint in the hierarchy?
Pater answers this question quite simply. Both processes differ in their
domains. Post-nasal voicing applies root-internally, that is, in a context where
resolutions of *Nt violations are hampered by root-specic faithfulness (Root-
Lin-IO). In contrast, fusion (or nasal substitution) applies across the prexroot
boundary, escaping from any (possibly) detrimental effects of RootLin-IO. This
opens up the possibility of designating fusion as the primary *Nt effect, while
blocking it root-internally by undominated RootLin-IO, where it is supplanted
by the secondary *Nt effect: post-nasal voicing. Once we see the outlines of
the analysis, it is fairly easy to state it in a constraint interaction. We will build
the ranking piece by piece, making full use of analytic insights inherited from
discussions of fusion and post-nasal voicing in Indonesian and Puyu Pungo
Quechua.

The rst building block is the ranking that accounts for nasal substitution, as well
as its root-internal blocking. This is copied straight away from the analysis of
Indonesian:

(77) Nasal substitution, and its root-internal blocking

RootLin-IO
*Nt
Linearity-IO

84
 2.4 Conspiracies of nasal substitution and other processes

That is, Linearity-IO can be violated under pressure of *Nt, while RootLin-
IO cannot.
The second building block is post-nasal voicing, which we copy from the
analysis of Puyo Pungo Quechua:

(78) Post-nasal voicing

*Nt
Ident-IO(ObsVce)

That is, featural identity of voice in obstruents can be violated under pressure of
*Nt.
One question remains: what is the relative ranking of Linearity-IO and
Ident-IO(ObsVce)? This is in fact a specic form of a general analytic prob-
lem stated above: how could both Linearity-IO and Ident-IO(ObsVce) be the
lowest-ranking correspondence constraint in the hierarchy? Now we see the rele-
vance of the idea of designating fusion as the primary *Nt effect, and post-
nasal voicing as the secondary *Nt effect. This has a direct implication for
the relative ranking of both correspondence constraints involved. This is stated
below:

(79) Fusion preferred to post-nasal voicing

*Nt
Ident-IO(ObsVce)
Linearity-IO

Both correspondence constraints are dominated by *Nt. Therefore, in principle,

both are violable to resolve *Nt violations. Yet the language will select a viola-
tion of the lowest-ranking constraint (Linearity-IO) wherever this is possible.
This is equivalent to saying that fusion is the primary *Nt effect. It will only
resort to the secondary *Nt effect of post-nasal voicing in contexts where fusion
fails, due to some higher-ranking constraint.
All we need to do now is put the rankings together:

(80) The OshiKwanyama *Nt conspiracy

RootLin-IO, *Nt
Ident-IO(ObsVce)
Linearity-IO

It remains to illustrate the correctness of this analysis by two tableaux, one of a

case of fusion, and the other of a case of post-nasal voicing.
The fusion tableau (81) serves to illustrate the status of fusion as the primary
effect on post-nasal voicing. The correspondence constraint which is violated by
fusion (Linearity-IO) is ranked at the bottom of the hierarchy. Therefore its
violation (that is, a fusion) is preferred to the violation of Ident-IO(ObsVce), that
is, post-nasal voicing:

85
The typology of structural changes

(81)
Input: /eN1-p2ati/ RootLin- *Nt Ident- Linearity-
IO IO(ObsVce) IO

a. + em1,2ati *
b. em1b2ati *!
c. em1p2ati *!

Even though fusion disrupts the inputs linear ordering of segments, no fatal
violation of RootLin-IO follows from this. This is because the violation involves
segments belonging to different morphemes. The violation of Linearity-IO in
the optimal form [em1,2ati] is taken for granted since this is the least expensive
violation of a correspondence constraint in the hierarchy.13
Turning to the post-nasal voicing tableau, we see that RootLin-IO precludes
the option of fusion, which is otherwise the primary *Nt effect. With fusion
ruled out, the grammar is thrown back onto its secondary resources in dealing
with Nt sequences, post-nasal voicing:

(82)
Input: /sitam1p2a/ RootLin- *Nt Ident- Linearity-
IO IO(ObsVce) IO

a. sitam1,2a *! *
b. + sitam1b2a *
c. sitam1p2a *!

The fusion candidate, which was still a winner in the previous tableau, is now
taken out of the race by RootLin-IO. Of the remaining candidates, the one with
post-nasal voicing incurs the least expensive violation.

2.5 Conclusion: a comparison with rule-based theory

Let us now turn to a nal comparison with rule-based theory. Consider the fol-
lowing set of rules, all of which (except epenthesis) state processes which have
gured prominently in the preceding discussions:

13
The difference with tableau (58) for Indonesian is that Ident-IO(ObsVce) is now included, in
a position below *Nt.

86
 2.5 Conclusion: a comparison with rule-based theory

(83)
Language Strategy employed Rule
a. Indonesian Nasal substitution [son, voice] / [+nasal] ___
b. Kelantan Nasal deletion [+nasal] / ___ [son, voice]
c. (unattested) Vowel epenthesis V / [+nasal] ___ [son, voice]
d. Puyo Pungo Post-nasal voicing [son] [+voice] / [+nasal] ___
e. Mandar Denasalization [+nasal] [nasal] /___[son, voice]

As we have observed earlier, the rules of this set are functionally related, but the
nature of the relationship is very difcult to capture in rule-based theory. The
diagram below portrays four of these rules as changes from an initial state
[+nasal][voice]. The four output states are placed next to the diagram.

(84) [+nasal] a. [+nasal] (nasal substitution)

b. [voice] (nasal deletion)
[+nasal] [voice]
d. [+nasal] [+voice] (post-nasal voicing)
[nasal] [+voice] e. [nasal] [voice] (denasalization)

It is clear that this set of output states share no positive characteristics: they do
not form a natural class. All that unites them is the fact that they are not
[+nasal][voice]. However, this is a negative characteristic, which also includes
many more states than those of (84ae). Aiming at a positive common denition,
one might note that all are derivationally related to a common single input state
[+nasal][voice]. Again, this denition is not sufciently precise, as it includes
many more output states than the above set. The single denition to capture
accurately the relevant set of output states combines both criteria: it denes the
set as (i) being derivationally related to the input state [+nasal][voice], as well
as (ii) not matching the output state [+nasal][voice]. Or to put it in plain prose,
these are precisely the rules that repair the ill-formed input conguration *Nt.
Rule-based theory is able to capture the functional unity of rules only at the
expense of becoming teleological that is, by acknowledging the insight that
structural changes function to undo their structural descriptions.
The same point can be made with even more force on the basis of the func-
tionally related rules of post-nasal voicing and nasal substitution in OshiKwan-
yama. The shared goal of both rules is precisely to avoid the sequence [+nasal]
[voice]. Again, this goal is not one that can be simply inferred from the statements
of both rules. Their structural changes may differ, but both function to avoid *Nt.
Conspiracies cannot be explained by rule-based theory, but they are within the
scope of explanation of OT. This success is due to two factors. First, OT denes
structural well-formedness on surface representations, rather than on input (or
87
The typology of structural changes

intermediate) representations. Second, OT separates the triggers of changes

from the actual changes.

SUGGESTIONS FOR FUR THER READING

Rule-ordering theory
Bromberger, Sylvain and Morris Halle (1989) Why phonology is different. Lin-
guistic Inquiry 20. 5170.
Kenstowicz, Michael and Charles Kisseberth (1979) Generative phonology:
description and theory. New York: Academic Press. [chapter 2, Phonological
rules and representations, and chapter 8, Rule interaction]

Rules versus constraints

Blevins, Juliette (1997) Rules in Optimality Theory: two case studies. In Roca.
227 60.
Noyer, Rolf (1997) Attic Greek accentuation and intermediate derivational repres-
entations. In Roca. 50126.

Segmental versus featural identity

Lombardi, Linda (1995a) Laryngeal neutralization and alignment. In Beckman,
Walsh Dickey, and Urbanczyk. 225 48.
Zoll, Cheryl (1996) Parsing below the segment in a constraint based frame-
work. PhD dissertation, University of California, Berkeley. [ROA-143,
http://ruccs.rutgers.edu/roa.html]

Factorial typology
Myers, Scott (1997a) OCP effects in Optimality Theory. Natural Language and
Linguistic Theory 15. 84792.

EXERCISES

1 A *Nh conspiracy in Modern Greek

Pater (1996) discusses a conspiracy of *Nt effects in Modern Greek (Newton
1972). The effects are twofold: post-nasal voicing occurs in (i), while nasal dele-
tion before a cluster of voiceless obstruents occurs in (ii):

88
 2.5 Conclusion: a comparison with rule-based
2.5 Exercises
theory

(i.a) /pemp-o/ pembo I send

(i.b) /ton#topo/ tondopo the place
(ii.a) /e-pemp-sa/ epepsa I send aorist
(ii.b) /ton#psefti / topsefti the liar (Cypriot dialect)14

a. Propose an analysis of this conspiracy, extending the analysis of *Nt

effects given in section 2.2. You will need at least one new constraint.
Motivate this constraint, and rank it.
b. Is your current analysis capable of excluding the suboptimal form
*epembza? If not, can you repair this defect by an additional con-
straint? Motivate and rank it.
c. Present tableaux of pembo and epepsa.

2 Kikuyu verbs
Kikuyu (Clements 1985b) has alternations due to interacting processes which
affect place of articulation, continuancy, and voicing, as illustrated by the follow-
ing verb forms:

Imperative 1 sg. imperfect

ur-a m-bur-eet lop off
tm-a n-dm-t cut
reh-a n-deh-eet pay
cin-a -jin-eet burn
kom-a -
om-t sleep
or-a -
or-eet buy

a. For each of these three alternations, decide which pairs of marked-

ness and faithfulness constraints are relevant. Rank each pair of
constraints.
b. Try to explain the fact that only prex consonants alternate in place
of articulation, while initial root consonants remain stable. What
assumptions do you make about the input form of the prex? Discuss
any consequences for Lexicon Optimization.
c. Integrate the three rankings found under (ab) into a total ranking.
Are any of the pairs of constraints found in (a) dependent upon other
pairs? Motivate each subranking by a form.
d. Present tableaux of zur-a and m-bur-eetb.

14
Pater, based on Newton (1972), reports that [i]n all dialects, the nasal is deleted within the
word [ii.a], and in most dialects, including Cypriot, it is deleted in an article preceding a noun,
except in slow, deliberate speech [ii.b].

89
The typology of structural changes

3 Voicing assimilation in Dutch

Consider the pattern of voicing assimilation in Dutch (Trommelen and Zonneveld
1979):

(i) /p+d / bd stropdas tie

/t+b/ db witboek white book
/k+d /
d zakdoek handkerchief
(ii) /d+k/ tk bloedkoraal red coral
/d+p/ tp huidplooi skin crease
/b+k / pk slobkous gaiter
(iii) /s+b/ zb kasboek cash book
/x+b/ b lachbui t of laughter
/f+b/ vb lafbek coward
(iv) /+t / xt hoogtij heyday
/z+p/ sp kaaspers cheese press
/v+k/ fk lijfknecht serf
(v) /k+v/ kf boekvorm book form
/p+z/ ps diepzee deep-sea
/t+z/ ts hartzeer heartache
(vi) /s+v/ sf bosveen peat
/f+z/ fs strafzaak trial
/x+v/ xf pechvogel unlucky person
(vii) /d+v/ tf handvat handle
/d+z/ ts Noordzee North Sea
/b+z/ ps krabzeer scratching sore
(viii) /z+v/ sf kaasvorm cheese mould
/v+z/ fs drijfzand quicksand
/+v/ xf hoogvlakte plateau

Two types of voicing assimilation interact in this pattern.

a. Which are these? State the generalization(s) concerning voicing

assimilation.
b. To account for the contrast between (i) and (ii), you need a new
constraint. State this constraint. Does it t in with the format of
constraints that has been assumed so far? Hint: is this a faithfulness
constraint or a contextual markedness constraint, or neither?
c. To account for the forms in (viiviii) you need yet another constraint.
State this constraint and rank it with respect to the other constraints.
d. Support your nal analysis by tableaux of kasboek and handvat.

90
 3
Syllable structure and economy

3.1 Introduction
The syllable is a major ingredient of phonological generalizations. It is crucial
in dening phonotactic patterns: well-formed sequences of segments, in particu-
lar of consonants and vowels. The syllable also governs patterns of epenthesis
and deletion, as discussed below. It supplies a level of prosodic organiza-
tion between segments and higher-level prosodic units: the foot and the pro-
sodic word (chapter 4). Finally, the syllable functions in the demarcation of
morpheme edges, as we will see below, and in dening the position and shape
of afxes (particularly inxes, discussed here, and reduplication, to be discussed
in chapter 5).
The syllable is dened as a prosodic category organizing segments in
sequences according to their sonority values. Each syllable has a sonority peak
(nucleus), usually a vowel, possibly surrounded on both sides by margin segments
of lower sonority, usually consonants (onset, coda). We make the slightly simpli-
fying assumption that syllables are organized into constituents in the following
way:

(1)

O N C O N C

t e m ple t
This at structure denes the syllable as a constituent that has a tripartite organ-
ization, consisting of an onset, a nucleus, and a coda, of which only the nucleus
is universally obligatory. We assume that consonants must be syllabied into
margins, that is onset or coda, while vowels must be syllabied as nuclei. Yet this
simple at syllabic model can be used to explain important cross-linguistic obser-
vations. Languages differ with respect to syllable structure along various dimen-
sions related to the tripartite structure of (1), the onset, nucleus, and coda, as we

91
Syllable structure and economy

will see shortly in section 3.2.1 The model will serve as the basis of the analysis
of epenthesis and deletion in section 3.35.
Syllabic well-formedness will turn out to be yet another instantiation of the
basic conict between faithfulness and markedness (or structural well-formed-
ness). The range of syllable types allowed by individual languages arise by
interactions of syllabic well-formedness constraints and segmental faithfulness
constraints. The precise nature of these interactions will be the primary con-
cern of this chapter. Again we will pay much attention to the issue of language
typology, as we did in both earlier chapters. Finally, by dening syllable well-
formedness in terms of constraint interactions, this chapter also presents the
point of departure for further applications of the syllable in stress systems and in
prosodic morphology in later chapters.

In this chapter we will review results of research on syllable structure, and place
these in the overall perspectives of the goals and issues in OT. First, in section
3.2 we will look into basic typological observations on syllable structure, and see
how OT captures these by well-formedness constraints. Section 3.2 will place
syllable-governed epenthesis and deletion in the perspective of conicts between
well-formedness and faithfulness. Our discussion will draw heavily on the typo-
logical results of Prince and Smolensky (1993), transposed into a correspondence
format. After studying epenthesis patterns in various languages, we will develop
a factorial typology of syllabic well-formedness and faithfulness constraints in
section 3.3. This will explain some major cross-linguistic properties of epenthesis,
in particular its economy. Next section 3.4 will deal with the cross-linguistically
common case of blocking of epenthesis and deletion at word edges. This will lead
us to the introduction of the concept of alignment (McCarthy and Prince 1993a),
the matching of (morphological and prosodic) edges, enforced by constraints.
Section 3.5 addresses the quality of epenthetic segments, focussing on their min-
imal markedness and contextual colouring, relating these properties to the discus-
sion of segment inventories of chapter 1. Phonotactic functions of the syllable, in
particular coda conditions, will be the topic of section 3.6. Finally, section 3.7
will summarize the results.

3.2 The basic syllable typology

3.2.1 Onsets and codas

Typological studies of syllable structure (for example, Jakobson 1962, Greenberg
1978, Kaye and Lowenstamm 1981, It 1986; see for an overview Blevins 1995)
1
In chapter 4 we will nd evidence for a more sophisticated view of syllable structure, one that
accommodates the notion of syllable weight.

92
 3.2 The basic syllable typology

have revealed solid cross-linguistic preferences for certain syllable types over
other types. In this section we will focus on a well-known cross-linguistic asym-
metry of onsets and codas: syllables prefer to begin with a consonant, whereas
they prefer to end in a vowel.
Following standard terminology, we will refer to syllable-initial consonants as
the onset of the syllable. The important typological nding with respect to
onsets is that all languages allow them, and none exclude them. That is:

(2) Implicational universal for syllable onsets

If a language has syllables that lack an onset, then it also has syl-
lables that have an onset.

Accordingly languages fall into two large classes: those allowing onset-less syl-
lables, such as Japanese, Diola-Fogny, Ponapean, and English, and those that do
not allow onset-less syllables, such as Temiar, Axininca Campa, and Arabic (It
1989). The crucial point is that no languages are known that disallow onsets.
Some languages actually supply an onset when no consonant is available in
the input. Such epenthesis of a consonant in onset position occurs in Axininca
Campa (Payne 1981, It 1989, McCarthy and Prince 1993b), see the examples in
(3).

(3) Consonant epenthesis in Axininca Campa

a. /no-N-koma-i/ nokomati he will paddle
b. /no-N-0hik-i / no0hiki (*no0hikti) he will cut

Note that [t]-epenthesis applies only in (3a), where two vowels are adjacent in
the input, while there is no need for epenthesis in (3b), where the input supplies
a consonant as an onset to the nal syllable. We will return to this pattern of
consonant epenthesis in section 3.3.2, where we will analyse it as an interaction
of syllabic well-formedness and faithfulness constraints. Here we will focus
on the syllabic well-formedness aspect, the avoidance of onset-less syllables, to
which we now turn.
We conclude from the typological results and epenthesis data that the presence
of an onset is an unmarked situation as compared to its absence. This is expressed
in the structural well-formedness constraint Onset (It 1989, Prince and Smolen-
sky 1993):

(4) Onset
*[ V (Syllables must have onsets.)

This constraint requires that syllables must not begin with vowels; it is satised
only by syllables that have an initial consonant, or onset. Therefore languages
in which Onset is undominated have obligatory onsets. Finally, Onset is
93
Syllable structure and economy

grounded in the articulatory and perceptual systems: the best starting point for
a vowel is a preceding consonant (rather than another vowel).

At the right margin of the syllable, the unmarked situation is the reverse, because
here the lack of a consonant is universally preferred to the presence of one:

(5) Implicational universal for syllable-nal segments

If a language has closed syllables, then it also has open syllables.

Again, languages fall into two large classes, according to whether they allow or
disallow codas. For example, Arabic, Tunica, and English allow codas, while
Fijian, Mazateco, and Cayuvava disallow codas (Blevins 1995). Crucially, no
languages are known in which syllables must have codas.
As before, languages may employ strategies to repair imperfect syllables,
such as the epenthesis of a vowel in order to avoid a coda. Such epenthesis occurs
in Boumaa Fijian (Dixon 1988), a language in which syllables strictly conform to
the shapes CV or V. Consider the following words, all realizations of English
loans ending in consonants:

(6) Vowel epenthesis in Boumaa Fijian

a. kaloko clock
b. aapolo apple
c. tone John

Again faithfulness is violated because vowel epenthesis causes a divergence

between the input and the output.

Both language typology and the wide-spread occurrence of processes which avoid
codas suggest that the unmarked situation is for syllables to lack codas. This
unmarked situation is encoded in the following well-formedness constraint.

(7) No-Coda
*C ] (Syllables are open.)

This constraint requires that syllables must not end in a consonant, or coda.
Languages in which No-Coda is undominated have open syllables only. Like
Onset, this constraint is grounded in the perceptual system: coda consonants,
particularly those standing before another consonant, tend to be unreleased, and
hence lack perceptual cues that are present in prevocalic consonants, which are
released (Ohala 1990, Steriade 1995b).
In sum, the two syllable form constraints Onset and No-Coda evaluate the
four logically possible syllable types as below:
94
 3.2 The basic syllable typology

(8) Syllable type Onset No-Coda

a. CV 3 3
b. CVC 3 *
c. V * 3
d. VC * *

This is not a tableau: observe that the constraints are not in a conicting relation-
ship, but merely cross-classify the four logical possibilities. The conclusion to be
drawn from this table is that CV is the perfect syllable shape, while all remain-
ing shapes {CVC, V, VC} are less perfect, by having a coda, or lacking an onset,
or having both defects.

We are already able to account for an important cross-linguistic observation:

single intervocalic consonants are universally syllabied as onsets, rather than
codas.2

(9) Universal syllabication of single intervocalic consonants

CV.CV
CVC.V

This follows from the fact that both well-formedness constraints, Onset and
No-Coda, favour an onset syllabication of (10a) over the alternative coda
syllabication of (10b):

(10) Input: / baba/ Onset No-Coda

a. + ba.ba
b. bab.a *! *!

Regardless of the ranking of both constraints, a single intervocalic consonant is

always optimally syllabied as an onset to the second syllable, rather than as a
coda to the rst. This reects the asymmetrical nature of both syllabic well-
formedness constraints.

3.2.2 Complex onsets and codas

So far our syllable typology has only considered the mere presence or absence of
onsets and codas. But languages also differ along the dimension of complexity
of syllable margins. For example, a language may allow codas, but still disallow

2
In the literature two languages have been reported to be exceptional to this generalization, by
syllabifying a single intervocalic consonant backward as the coda of the preceding syllable.
Both cases, Oykangand (Sommer 1981) and Barra Gaelic (Borgstrm 1937), are doubtful,
however. See for discussion Clements (1986), McCarthy and Prince (1993b), Blevins (1995).

95
Syllable structure and economy

complex codas, that is, codas consisting of two or more consonants. In such a
language, codas must be simple, that is, consist of one consonant.
Observe that the use of the terminology complex and simple already sug-
gests a markedness relation in this respect: indeed many languages restrict the
complexity of syllable margins complex onsets are universally marked as com-
pared to simple onsets; complex codas are marked as compared to simple codas.
The unmarked status of simple onsets and codas is supported by various kinds of
evidence.
First consider the implicational universals below:

(11) Implicational universal for onset complexity

If a language allows complex onsets, then it also allows simple
onsets.

(12) Implicational universal for coda complexity

If a language allows complex codas, then it also allows simple codas.

Crucially, no languages have syllables with complex margins, while disallowing

syllables with simple margins.
Another type of evidence for the marked status of complex margins is the fact
that many languages actively avoid complex onsets and codas. Typical avoidance
strategies are vowel epenthesis and consonant deletion. Some languages, such
as Lenakel (Lynch 1974, Blevins 1995), avoid both types of complex margins
(onsets and codas) by a single strategy, that of vowel epenthesis:3

(13) Avoidance of complex margins by vowel epenthesis in Lenakel

a. /t-n-ak-ol/ tp.na.
l you (sg.) will do it
b. /t-r-kn/ tpr.
n he will eat it
c. /ark-ark/ ar.
a.rpkh to growl
d. /kam-n-Ran-n/ kam.np.Ra.npn for her brother4

An epenthetic vowel [] breaks up clusters of two consonants at the beginning

(13ab) or end (13cd) of a word, as well as clusters of three consonants in
medial position (13d). A medial cluster of two consonants (such as /rk / in 13c)
is not broken up by epenthesis, as this can be split between two syllables without
need for a complex margin: the rst consonant syllabies as a simple coda, and
the second as a simple onset.5

3
Lynch (1974: 87) reports that word-initial consonant-glide-clusters are permitted.
4
The actual form is [kam.np.R.npn]. Here /a / has undergone a process of rounding into []
applying when /a / is adjacent to a velarized labial (Lynch 1974: 97).
5
We focus on obligatory epenthetic vowels, abstracting away from optional epenthetic vowels
that may be inserted in clusters of two consonants before unstressed vowels (Lynch 1974: 84).

96
 3.2 The basic syllable typology

In spite of the appearance that may be given by the Lenakel examples, no

intrinsic typological connection has been established between complexity of
onsets and codas. The following table shows that all combinations of onset and
coda complexity occur in the worlds languages (It 1986, Blevins 1995):

(14) A typology of margin complexity

Simple codas only Complex codas allowed

Simple onsets only CV, CVC CV, CVC, CVCC

(Japanese, Yokuts) (Finnish, Tunica)

Complex onsets CV, CVC, CCV, CCVC CV, CVC, CVCC, CCV, CCVC, CCVCC
allowed (Spanish,6 Sedang) (English, Totonac)

This shows that two structural well-formedness constraints are needed to capture
margin complexity: one for onsets, and another for codas. These constraints are
stated below:

(15) *ComplexOns
*[ CC (Onsets are simple)

(16) *ComplexCod
*CC ] (Codas are simple)

Any language in which *ComplexOns is undominated has simple onsets only (this
is the case in languages such as Japanese, Yokuts, Finnish, and Tunica). And any
language in which *ComplexCod is undominated allows no complex codas, if it
allows codas at all (as is the case in Japanese, Yokuts, Spanish, and Sedang).7

We have seen that syllables are subject to structural well-formedness constraints.

Syllable inventories of languages behave symmetrically with respect to these con-
straints. At this point the following question naturally arises: why do imperfect
syllable types such as CVC, V, VC, or VCC, occur at all in the worlds languages?
Why do languages not radically eliminate any syllabic imperfection, to become
perfect CV-languages? That is, why are not all languages like Boumaa Fijian? The
answer to this question, to be given in the next section, is characteristic of OT:
input faithfulness may prevent the segmental changes necessary to create perfect
syllables. That is, languages may tolerate imperfect syllables, but only under

6
Spanish rarely allows complex codas word-nally, e.g. biceps (Harris 1983). Another language
of this type is Canela-Krah (Popjes and Popjes 1986).
7
Whether or not a language has codas depends on the ranking of No-Coda, the constraint ruling
out codas regardless of their complexity.

97
Syllable structure and economy

duress of faithfulness. In the next section we will take a detailed look at the
interaction of faithfulness and syllabic well-formedness.

3.3 Epenthesis and the conict of well-formedness and faithfulness

Stated in optimality-theoretic terms, epenthesis involves a violation of faithful-
ness: the epenthetic segment has no counterpart in the input. This violation of
faithfulness is due to dominant syllable well-formedness constraints, such as
Onset and No-Coda, discussed in the previous section. Reconsider consonant
epenthesis in the onset in Axininca Campa, an example of which is repeated
below from (3):

(17) /no-N-koma-i/ no.ko.ma.ti
no.ko.ma.i

The epenthetic output [no.ko.ma.ti] contains a segment that has no input coun-
terpart, a violation of faithfulness. Still, this is preferred to the faithful candidate
*[no.ko.ma.i], which has a syllable without an onset, thus violating Onset. We
recognize a now-familiar theme, the conict of well-formedness (here, avoid-
ance of marked syllable structure) and faithfulness (here avoidance of lexically
unsponsored segments). In sum, epenthesis involves a resolution of this conict
at the expense of faithfulness: the costs of inserting a (non-underlying) segment
are less than those of imperfect syllable structure. This section will develop an
optimality model of epenthesis capturing this insight, which has a range of typo-
logical consequences. Before looking into the model, we need to discuss briey
its theoretical origins.

3.3.1 A brief history of epenthesis

The key insight that epenthesis and syllabication are inextricably connected
already appeared in derivational phonology (Selkirk 1981, It 1986, 1989).
According to Selkirk and It an epenthetic segment is an empty structural posi-
tion whose presence is required by a language-specic syllable template. This
template denes whether or not an onset is obligatory, whether or not a coda is
allowed, etc. In this model, epenthesis is driven by an imperfect match between
input segments and the syllable template. An imperfect match may arise, for
example, when the input contains a sequence of two vowels, without an inter-
vening consonant, whereas the syllable template requires an obligatory onset.
Syllabication then enforces the template by providing an empty onset position:
one that is not lled by an input consonant. Empty syllabic positions must receive
some phonetic interpretation (spelling out their features) at the end of the deri-
vation. For example, the empty onset of the nal syllable in the Axininca example
[nokomati] is spelled out as [t].

98
 3.3 Epenthesis and the conict of well-formedness and faithfulness

(18)

O N C O N O N O N

n o k o m a t i [nokomati]
In this model epenthesis is a fully automatic consequence of syllabication, a
spell-out of an empty templatic position. Accordingly there are no rules of epen-
thesis, but only rules of syllabication enforcing the syllable template. Whether
or not epenthesis applies in a language only depends on the template: this may
require specic positions to be obligatory. Anticipating OT constraint ranking, It
(1989) argues that templates enforce universal syllabication principles with dif-
ferent strengths. For example, languages such as Axininca Campa (which enforce
onsets by consonant epenthesis) use a strengthened version of the Onset Principle
(avoid onset-less syllables). This version It dubs the Strict Onset Principle:
onset-less syllables are impossible. Thus, cross-linguistic variation in onset
epenthesis is due to a single binary parameter: languages select either the weak
Onset Principle or the strong Strict Onset Principle.

Prince and Smolensky (1993) take these ideas to their logical conclusion, deriving
relative strength of templatic principles from constraint ranking. In their Opti-
mality model, output constraints assume the function of dening syllabication
and epenthesis. The grammar is an evaluation device selecting the most harmonic
output from an innite candidate set, according to a language-specic ranking
of universal constraints. OT does away with language-specic templates: cross-
linguistic variation in epenthesis patterns is explained by variation in constraint
ranking. For example, Prince and Smolensky argue that in languages enforcing
syllable onsets by consonant epenthesis, Onset takes precedence over the faith-
fulness constraint Fill, requiring that syllable positions must be lled with
underlying segments. Note that Fill is indeed violated in the output (18). This
output is unfaithful to its underlying segments as it contains a syllabic position
that is left unlled by an underlying segment.
Prince and Smolenskys model is known as Containment Theory, after its cen-
tral assumption that no element may be literally removed from the input form.
The input is thus contained in every candidate output, including those candidate
outputs in which it is left unpronounced. In Containment Theory, the deletion of
a segment does not involve its removal from the output (this is ruled out by
containment). Instead the input segment is left unparsed by a syllable, hence
phonetically uninterpreted (compare Stray Erasure in McCarthy 1979, Steriade
1982, It 1986). Analogously to Fill, deletion is penalized by a faithfulness

99
Syllable structure and economy

constraint Parse underlying segments must be parsed into syllable structure.

See the hypothetical example of a deletion of a consonant /p/ in (19):

(19) Deletion and epenthesis in Containment Theory (O = onset, N =

nucleus, C = coda)
a. b.

O N O N C

m a p [ma] a p [tap]

In (19a), a segment /p/ is left phonetically uninterpreted as it is not syllabied.

Epenthesis of [t] (19b) is the exact opposite situation: a syllabic position is
phonetically interpreted even though it is not lled by a segment. Prince and
Smolenskys OT model of epenthesis thus preserves the central assumption of
the earlier (SelkirkIt) derivational theory, that epenthetic segments are empty
syllable positions.

Correspondence Theory (McCarthy and Prince 1995a) drops Prince and Smolen-
skys notion of containment. That is, it is no longer required that the input be
contained in each output candidate. Consequently epenthetic segments are no
longer viewed as the phonetic interpretations of empty positions, but rather as
(full-edged) output segments that have no counterparts in the input. This cor-
rectly predicts that the feature contents of epenthetic segments may participate
in phonological processes (Davis 1995). The faithfulness role of containment,
restricting divergence between input and output (penalizing overparsing and
underparsing), is transferred to correspondence. Correspondence, as we learned
in chapter 1, is a relation between pairs of segments in the input and output. It is
subject to constraints requiring various kinds of congruence between segments in
input and output (such as identity, linearity, etc.). In spite of the differences
between Containment Theory and Correspondence Theory, the latter preserves the
insight that epenthesis and deletion both reect the dominance of syllabic well-
formedness over faithfulness. In the following subsections we will develop a
correspondence model of epenthesis capturing this insight.

3.3.2 Introducing two correspondence constraints

3.3.2.1 The anti-epenthesis constraint DEP-IO

Epenthesis involves a violation of faithfulness: the output diverges from the input
by the presence of an epenthetic segment, one that is not sponsored by lexical
100
 3.3 Epenthesis and the conict of well-formedness and faithfulness

representation. The faithfulness constraint militating against epenthesis is

Dependency-IO (or Dep-IO), after McCarthy and Prince (1995a):8

(20) Dep-IO
Output segments must have input correspondents. (No epenthesis)

This constraint is violated by any output segment that lacks a correspondent

in the input. Such a violation is incurred by epenthetic [t] in the example
[nokomati] from Axininca Campa, introduced in the preceding section. This
violation is shown in a diagram below that indicates corresponding segments
through vertical lines. Note that the input contains no segment corresponding to
output [t]:

(21) Violation of Dep-IO in [no.ko.ma.ti]

Input n o- N- k o m a -i

Output n o k o m a t i

Other examples of Axininca Campa show that epenthetic [t] occurs between every
pair of vowels that are in adjacent morphemes in the input:

(22) a.i /no-N-koma-i/ nokomati he will paddle

a.ii /no-N-koma-aa-i/ nokomataati
he will paddle
again
a.iii /no-N-koma-ako-i/ nokomatakoti he will paddle
for
a.iv /no-N-koma-ako-aa-i-ro/ nokomatakotaatiro he will paddle
for it again
b.i /no-N-0hik-i/ no0hiki he will cut
b.ii /no-N-0hik-aa-i/ no0hikaati he will cut
again
b.iii /no-N-0hik-ako-i/ no0hikakoti he will cut for
b.iv /no-N-0hik-ako-aa-i-ro/ no0hikakotaatiro he will cut for
it again

Epenthesis in onsets shows that Axininca ranks Dep-IO below Onset, the struc-
tural well-formedness constraint whose violation epenthesis avoids. Onset epen-
thesis involves the following ranking:

(23) Epenthesis in onset

Onset
Dep-IO

8
Dep-IO was introduced in chapter 1, but left unillustrated.

101
Syllable structure and economy

This ranking is demonstrated by the following tableau. It contains two candidates,

which differ only in the presence versus absence of an epenthetic consonant.

(24) Input: /no-N-koma-i/ Onset Dep-IO

a. + no.ko.ma.ti *
b. no.ko.ma.i *!

In sum, epenthesis is the assertion of syllabic well-formedness at the expense of

input faithfulness.

3.3.2.2 The anti-deletion constraint MAX-IO

Next consider the possibility of a second strategy that might have been employed
to avoid onset-less syllables in Axininca Campa: the deletion of a vowel. This
might have resulted in *[no.ko.ma], by deletion of the sufx vowel /i / following
the root vowel /a/.9 Vowel deletion is a cross-linguistically common reaction to
hiatus; however, not so in Axininca Campa, which prefers consonant epenthesis:

(25) Consonant epenthesis preferred to vowel deletion

[no.ko.ma.ti]
*[no.ko.ma]

Segment deletion is a violation of faithfulness, just as epenthesis is. It is the

mirror-image case: an input segment has no counterpart in the output. The
constraint that enforces the preservation of input segments in the output is
Maximality-IO (or Max-IO):

(26) Max-IO
Input segments must have output correspondents. (No deletion)

Importantly, Max-IO does not require that an input segment and its correspond-
ing output segment have identical feature content. (This requirement is made by
another faithfulness constraint introduced in chapter 1: Ident-IO.) All that is
required by Max-IO is that for every output segment there is some input segment
corresponding to it. This is tantamount to a ban on the deletion of input segments
hence Max-IO is the anti-deletion constraint.
Consider the following diagram, which shows that Max-IO is actually violated
in the output candidate *[no.ko.ma], as the input segment /i/ has no output
counterpart:

9
Rather arbitrarily, we consider the deletion of a sufx vowel, rather than a root vowel both
options are equivalent with respect to the violation of faithfulness involved. But see chapter 2
for McCarthy and Princes proposal that root faithfulness outranks afx faithfulness.

102
 3.3 Epenthesis and the conict of well-formedness and faithfulness

(27) Violation of Max-IO in *[no.ko.ma]

Input n o- N- k o m a -i

Output n o k o m a

Compare this diagram with the one in (21), to see that epenthesis and deletion
are mirror-image cases from the viewpoint of input correspondence.
The fact that Axininca Campa prefers consonant epenthesis to vowel deletion
tells us that Max-IO dominates Dep-IO. The former is not violable under duress
of Onset, whereas the latter is. Combining our conclusions, we arrive at the
following total ranking:

(28) Consonant epenthesis preferred to vowel deletion in hiatus resolution

Onset, Max-IO
Dep-IO

This ranking is illustrated by the following tableau, which contains both candid-
ates of the preceding tableau, plus the deletion candidate (29b):

(29) Input: /no-N-koma-i/ Onset Max-IO Dep-IO

a. + no.ko.ma.ti *
b. no.ko.ma *!
c. no.ko.ma.i *!

Note that each candidate incurs a violation of one constraint. As we may predict
from the discussion in chapter 2, the optimal output (29a) violates the lowest-
ranking faithfulness constraint, here Dep-IO. The lowest-ranking faithfulness con-
straint is, by denition, the one whose violation is least expensive. It is violated
in order to avoid violation of both higher-ranking constraints, Onset and Max-
IO, whose mutual ranking is irrelevant here.
The insight that the lowest-ranking constraint determines the outcome allows
us to construct a factorial typology of these three constraints. Only three rankings
need to be considered, one for each constraint in lowest-ranking position:

(30) A factorial typology of onset-driven epenthesis and deletion

a. Onset, Max-IO
Dep-IO Consonant epenthesis to create onset.
b. Onset, Dep-IO
Max-IO Vowel deletion to avoid onset-less
syllable.
c. Max-IO, Dep-IO
Onset No deletion, nor epenthesis to create
onset.
103
Syllable structure and economy

Ranking (30a) produces the case of onset epenthesis, where Dep-IO is violated,
of which Axininca Campa is an example. Ranking (30b) is the case of vowel
deletion in hiatus, to avoid an onset-less syllable, violating Max-IO. This is also
a cross-linguistically common strategy occurring in languages such as Modern
Greek and Yokuts. Finally, ranking (30c) represents languages that disallow dele-
tion and epenthesis as strategies to avoid onset-less syllables: when all strategies
to attain unmarked structures are blocked, marked structures are necessarily
allowed. This happens at the expense of syllabic well-formedness (Onset).
This is also a cross-linguistically common case, occurring in English and other
languages.

In sum, we have found that the range of syllable structures of a language depends
on the interaction between syllabic well-formedness and faithfulness. In the next
section we will extend the factorial typology to all four constraints that were used
so far: Onset, No-Coda, Dep-IO, and Max-IO. This extended typology will
explain some major cross-linguistic properties of epenthesis and deletion.

3.3.3 Economy, or against excessive epenthesis

A general consequence of minimal violation is that changes of the input take
place only under duress. Without proper incentive, no change will take place. The
logic of economy is shown in tableau (31). Two realizations of an input /A/ are
compared as to their relative harmony. The rst realization of /A/, the output
candidate, [A], incurs no violation of the well-formedness constraint, and is also
faithful to its input. The second realization of /A/, the output candidate, [B], is
also structurally well-formed, but incurs a minimal violation of faithfulness.
Faithfulness will always decide in favour of the most faithful candidate, [A], at
the expense of less faithful candidate [B], regardless of the ranking of constraints.

(31) /A/ Well-formedness Faithfulness

a. + [A]
b. [B] *!

It is only when a structural well-formedness constraint militates against output

[A] (e.g. a constraint *[A] ) that the output may be forced to diverge from the
input, given the ranking Well-formedness
Faithfulness.
This principle of inertness-by-default, or economy, is a direct consequence
of minimal violation. Or stated differently (Prince and Smolensky 1993),

(32) do only when necessary.

104
 3.3 Epenthesis and the conict of well-formedness and faithfulness

Now consider what do only when necessary predicts for epenthesis. As we

saw earlier, epenthesis incurs a violation of faithfulness, as adding a segment
entails some divergence between output and input. Given the presence of the
anti-epenthesis constraint Dep-IO in all grammars, no excess epenthesis will
take place, regardless of its ranking. Pressure to violate Dep-IO may only reside
in dominant well-formedness constraints, specically Onset and No-Coda.
It has been argued by Prince and Smolensky (1993) that regardless of the
ranking of Dep-IO, Onset, and No-Coda, specic kinds of epenthesis are gen-
erally disfavoured. For example, the epenthesis of a consonant as a syllable
coda cannot be modelled under any logically possible ranking of syllabic well-
formedness and faithfulness constraints, as we will see below. Such typological
results are essential since they show that OT, a theory of violable constraints, is
still capable of ruling out logically possible but cross-linguistically unattested
situations. Or to state it differently, an empirically restrictive linguistic theory
requires no inviolable constraints.

3.3.3.1 Predictions about epenthesis

Here we will consider the following three predictions:
(33) Three predictions derived from minimal violation
a. Epenthesis never results in an increase of structural markedness.
b. Epenthesis always results in a decrease of structural markedness.
c. Epenthesis always applies minimally, that is, precisely to the
extent that is necessary to improve structural markedness.
Let us now see how these typological predictions follow from constraint
interaction.

The rst prediction is that epenthesis never results in an increase of markedness,

as compared to a non-epenthetic output. For example, epenthesis never creates a
marked structural position, such as a coda. By denition, marked positions cannot
be required by structural constraints. Hence, epenthesis of a coda consonant is
inherently anti-harmonic, regardless of the mutual ranking of syllabic well-
formedness (No-Coda) and faithfulness (Dep-IO). We consider a hypothetical
example / baba/, with an epenthetic consonant []:
(34) No consonant epenthesis ( []) in coda position:
Input: /baba/ Onset No-Coda Dep-IO

a. + ba.ba
b. ba.ba * *

105
Syllable structure and economy

The epenthetic candidate (34b) is less harmonic than its non-epenthetic compet-
itor (34a) in two respects, as it violates both No-Coda and Dep-IO. Conversely
no well-formedness constraints are satised by (34b) which (34a) violates. This
means that (34b) will always be less harmonic than (34a), in any grammar,
regardless of the ranking of the constraints. (Since any violation will be fatal to
34b, no exclamation marks have been added.)
By the same reasoning, no vowel will be epenthesized adjacent to another
vowel, since that produces unnecessary violations of both Onset and Dep-IO:

(35) No vowel epenthesis ( [i]) next to another vowel:

Input: / baba/ Onset No-Coda Dep-IO

a. + ba.ba
b. ba.i.ba * *

Again the ranking of the constraints involved is irrelevant to the outcome.

The second prediction is that epenthesis is excluded whenever it yields no gains

in terms of structural well-formedness. For example, the epenthesis of an entire
CV syllable, even though this violates no well-formedness constraints, is intrin-
sically non-harmonic, since nothing is gained by it.10 Assuming an epenthetic
onset [] plus a nuclear [i], this would produce (36b):

(36) No epenthesis of an entire syllable ([i]):

Input: / baba/ Onset No-Coda Dep-IO

a. + ba.ba
b. ba.i.ba *!*

Considering the shape of the epenthetic syllable, we nd that this is structurally

perfect, since CV has an onset and lacks a coda. Then why is epenthesis of such
a perfect syllable nevertheless excluded? This reects a fundamental property of
constraint evaluation. The criterion for evaluation is minimization of ill-formed
structures, rather than maximization of well-formed structures. Nothing is gained
by adding any structurally perfect elements without an accompanying reduction
of ill-formedness. When no structural improvements are attained by adding an

10
In fact, languages are known in which the epenthesis of an entire syllable is forced by
superordinate prosodic well-formedness constraints. For example, syllable epenthesis may
augment a monosyllabic form in order to satisfy a high-ranking minimal word constraint to
the effect that every word be disyllabic; cf. McCarthy and Prince (1993b). See chapter 4.

106
 3.3 Epenthesis and the conict of well-formedness and faithfulness

element, its addition is ruled out by faithfulness constraints. This is what occurs
in tableau (36). There is no bonus for addition of the structurally perfect epen-
thetic syllable in the columns below the well-formedness constraints Onset and
No-Coda, but simply a lack of marks, indicating an absence of violation of these
constraints. In this respect, the epenthetic candidate (36b) fares no better than the
faithful candidate (36a). In the end, it is faithfulness that swings the balance in
favour of (36a).

The third prediction is that epenthesis always applies minimally. That is, no more
segments are epenthesized than are strictly required to minimize the violation of
syllabic well-formedness. This prediction is actually a corollary of both predic-
tions made above. Epenthesis serves to minimize the violation of well-formedness
constraints. Any additional epenthesis on top of this minimum, which yields no
further gains in well-formedness, is unnecessary hence it is blocked by faith-
fulness, as we will illustrate in the next section.

3.3.3.2 Minimal epenthesis in Lenakel

Consider the epenthetic examples from Lenakel (Lynch 1974, Blevins 1995),
presented earlier in (13). Epenthesis of [] breaks up any consonant clusters that
would otherwise surface as complex syllable margins:
(37) a. /t-n-ak-ol/ tp.na.
l you (sg.) will do it
*tna.
l
b. /ark-ark / ar.
a.r pkh to growl
*ar.
arkh
c. /kam-n-Ran-n/ kam.np.Ra.npn for her brother
*kam.nRann, *kamn.Rann
No complex margins occur in Lenakel, hence *ComplexOns and *ComplexCod
are both undominated. We combine both into a cover constraint *Complex
no complex syllable margins. Since epenthesis applies to avoid violations of
*Complex, the latter constraint must dominate Dep-IO:
(38) Vowel epenthesis triggered to avoid complex margins
*Complex
Dep-IO
The relevant observation is that no more vowels are inserted than are necessary
to avoid violations of *Complex:
(39) Vowel epenthesis as a minimal adjustment in Lenakel
ar.
a.rpkh
a.rp.
a.rp.
p
tp.na.
l
tp.na.
.lp
kam.np.Ra.npn
*ka.mp.np.Ra.np.np
107
Syllable structure and economy

Note that all suboptimal righthand forms are perfectly structured strings of CV
syllables. All syllables have (simple) onsets, while none have codas. But in spite
of their structural perfection, these candidates fall short of being selected. Appar-
ently, vowel epenthesis is not available in Lenakel as a general strategy to avoid
codas, but only as a more restricted strategy to avoid complex codas and onsets.
To put it differently, input faithfulness takes priority over avoidance of codas:
(40) Vowel epenthesis unavailable to avoid codas
Dep-IO
No-Coda
In sum, the faithfulness constraint Dep-IO is sandwiched between two syllabic
structural well-formedness constraints, *Complex and No-Coda:
(41) Preliminary ranking for Lenakel vowel epenthesis
*Complex
Dep-IO
No-Coda
Tableau (42) considers ve candidates differing only in the number of epenthetic
vowels. At the top we nd the most faithful candidate (42a), which has no viola-
tions of Dep-IO; at the bottom is the structurally perfect candidate (42e), which
has no violations of well-formedness constraints *Complex and No-Coda. The
winner, (42c), stands in between both extremes. Still there is no trade-off
between faithfulness and well-formedness, as everything follows from strict
domination.

(42) Input: / kam-n-Ran-n/ *Complex Dep-IO No-Coda

a. kamn.Rann *!* **
b. kam.np.Rann *! * **
c. + kam.np.Ra.npn ** **
d. ka.mp.np.Ra.npn ***! *
e. ka.mp.np.Ra.np.np ***!*

Candidate (42c) is optimal since it has the minimal number of violations of Dep-
IO that are necessary to avoid violations of *Complex. This minimum number of
violations is two: fewer would not do to avoid complex margins (42ab), while
more violations are ruled out by faithfulness (42de).11 In conclusion, epenthesis
is economical no more vowels are inserted than are strictly necessary to avoid
violations of syllable well-formedness.
11
Of course, the prediction that epenthesis applies minimally is relativized to the grammar. In
Lenakel minimal has a different meaning than in a language which completely avoids codas
(due to undominated No-Coda).

108
 3.3 Epenthesis and the conict of well-formedness and faithfulness

That No-Coda is still active despite the fact that it is dominated is shown by
the tableau of [tp.na.
l]:

(43) Input: /t-n-ak-ol/ *Complex Dep-IO No-Coda

a. tna.
l *! *
b. + tp.na.
l * *
c. pt.na.
l * **!
d. tp.na.
.lp **!

The faithful non-epenthetic candidate (43a) is eliminated by undominated *Com-

plex. Initial epenthesis, as in (43c), successfully avoids a complex onset, at the
expense of only one violation of Dep-IO. But there is an even more successful
competitor, which has both of these virtues, plus the additional advantage of
having one fewer violation of No-Coda. This is (43b), where smart epenthesis
breaks up consonant clusters in such a way that both consonants form onsets in
the ouput.12
The attentive reader may have noticed that avoidance of complex margins may
be achieved in more than one way, even for two epenthetic vowels. Three output
candidates that were left unconsidered in tableau (42) are presented in (44). All
are suboptimal as compared to [kam.np.Ra.npn]:

(44) Positional preferences in epenthesis, unaccounted for by present

analysis:
kam.np.Ra.npn
*ka.mpn.Ra.npn, *kam.np.Ran.np, *ka.mpn.Ran.np
The model of epenthesis that we have developed so far has no means to make
structural distinctions among these candidates. Observe that all four incur iden-
tical violations of syllabic well-formedness constraints (in particular, of No-
Coda). Positional preferences for epenthetic elements will be discussed in the
next section.

3.3.4 Alignment and epenthesis

The simple model of epenthesis that we have developed so far does not account
for any morphological inuence over epenthesis. Such inuence has been pointed
out for various languages by Kenstowicz (1994b), Spencer (1994), and Blevins
(1995). Cross-linguistic evidence shows that epenthesis applies in such a way
that morphological constituents are maximally respected. For example, epenthesis
12
In section 3.3.4 we will nd that there is yet another constraint favouring non-initial epen-
thesis, one requiring the coincidence of morphological and prosodic edges.

109
Syllable structure and economy

tends to respect the contiguity of segments by applying between morphemes,

rather than inside morphemes. Epenthesis also tends to highlight the edges of
morphemes, for example by making the beginning of a morpheme coincide with
the beginning of a syllable. Such morphological factors in epenthesis make sense
from the viewpoint of the processing system: the processing of lexical elements
is improved when their realizations are constant and signalled by prosodic edges.

3.3.4.1 Alignment and epenthesis in Axininca Campa

As an example of the relevance of word edges to epenthesis, reconsider the
pattern of [t]-epenthesis in Axininca Campa, a pattern analysed earlier in section
3.3.2. There we concluded that epenthesis serves to avoid onset-less syllables.
McCarthy and Prince (1993a, b) draw attention to the fact that epenthesis fails to
apply in word-initial position:

(45) Lack of initial epenthesis in Axininca Campa

a. /osampi/ osampi *tosampi ask
b. /i-N-koma-i/ ikomati *tikomati he will paddle

What blocks epenthesis initially? Naively, one might be tempted to relax the
constraint Onset in word-initial syllables. This would lead to a specic version
of this constraint:

(46) Onset (naive version, to be rejected)

Syllables must have onsets, except word-initially.

However, (46) is not a very illuminating way to approach this pattern. The rst
problem is its relation to the general constraint Onset, which has no such excep-
tion clause. If this is replaced by (46), then we can no longer explain the fact
that in some languages all syllables must have onsets, regardless of their position
in the word. Alternatively, if (46) is interpreted as independent of the general
constraint, then a redundancy problem arises: the same requirement (syllables
must have onsets) is stated by two separate constraints in the universal inventory.
Such a redundancy should be avoided, as it amounts to the loss of generalization.
A second problem is that (46) fails to explain why the exception clause for onsets
holds for word-initial syllables, that is, for word-initial position. There is much
evidence (from a range of phenomena other than epenthesis) that the left edge of
the word is respected by phonological changes. This inertness of the left word
edge is expressed by the constraint Align-L (after McCarthy and Prince 1993a):13

13
McCarthy and Prince (1993a) state Align-L on the Stem, rather than on the Grammatical
Word.

110
 3.3 Epenthesis and the conict of well-formedness and faithfulness

(47) Align-L
The left edge of the Grammatical Word coincides with the left edge
of the PrWd.

Alignment is a much more general phenomenon than may be suggested by this

constraint, which is only one member of a constraint type of which we will see
many more examples in this book. In section 3.4 we will generalize the format
of alignment constraints.14
The misaligned nature of the suboptimal candidate [tosampi] is shown in dia-
gram (48), representing both its prosodic and morphological structure:
(48) PrWd
Prosodic structure

tosampi
Morphological structure
GrWd
Note that the left edge of the Grammatical Word is dened by the initial segment
of the stem [o], whereas the left edge of the Prosodic Word is dened by the
epenthetic segment [t]. Hence both edges do not coincide, in violation of Align-L.

The relation between [t]-epenthesis and its blocking in word-initial position now
becomes clear. It is just another interaction of conicting, violable constraints:
alignment takes priority over avoidance of onset-less syllables. This is stated in
the ranking in (49), which includes the subranking Onset
Dep-IO from (23)
above:

(49) Onset epenthesis respecting left word edge

Align-L
Onset
Dep-IO

Having factored out the left-edge effect, we may now dispense with the odd
exception clause in the Onset constraint (46), restoring this constraint to its
original generality.
A translation of (49) into a prose statement would read as follows:

(50) Avoid epenthesis,

except when this provides an onset to an otherwise onset-less syllable,
except when this syllable stands in initial position.
14
McCarthy and Prince (1995a) restate alignment as a correspondence constraint Anchoring,
requiring that corresponding segments in the input and output both stand at an edge. See
section 3.6.3.

111
Syllable structure and economy

As Prince and Smolensky (1993) point out, such exception clauses are highly
common in generalizations about linguistic patterns. OT is ideally suited to cap-
ture such situations. Constraints are the substance of the theory, expressing lin-
guistic generalizations without qualication about exception clauses. The latter
function, encoding the relative scopes of generalizations, is attributed to the con-
straint hierarchy. Therefore, constraints can be highly general, as long as they are
violable as well.

A full account of the Axininca epenthetic pattern should also address the ranking
of Max-IO. Earlier we found that Max-IO and Onset were both undominated,
but that was before the initial blocking data in (45) had entered the picture, telling
us that Onset is dominated. However, we still have no reason to demote Max-
IO, and we will keep it at the top of the hierarchy, where it sits together with
Align-L:

(51) Input: /osampi/ Align-L Max-IO Onset Dep-IO

a. + o.sam.pi *
b. to.sam.pi *! *
c. sam.pi *!

The ranking Max-IO
Onset has no negative consequences for the analysis
presented in section 3.3.2. This ranking correctly predicts that vowel deletion is
not allowed as a strategy to avoid onset-less syllables in Axininca Campa.

3.3.4.2 Alignment and epenthesis in Lenakel

We now return to the cases of Lenakel epenthesis that we left slightly unresolved
in section 3.3.3.2. All forms in (52ab) have the minimal number of epenthetic
vowels to avoid violations of *Complex. And all have the same number of
violations of No-Coda. Then what causes the ill-formedness of the starred
forms?

(52) a. /ark-ark / ar.

a.rpkh to growl
*ar.
ar.
p
b. / kam-n-Ran-n / kam.np.Ra.npn for her brother
*kam.np.Ran.np

This question is now answered as follows: both starred output forms violate
alignment at their right edges. A word-nal epenthetic vowel is misaligning since
it breaks PrWds coincidence with the right edge of the Grammatical Word. This
112
 3.3 Epenthesis and the conict of well-formedness and faithfulness

is shown once more in the form of hierarchical representations in (53) which

compare the optimal forms (which satisfy right-edge alignment) and suboptimal
forms (violating right-edge alignment):
(53) Coincidence of morphological edges and prosodic edges in Lenakel
a. /ark-ark/ b. /kam-n-Ran-n/
PrWd PrWd PrWd PrWd

ar
a r kh
ar
ar k kam n Ra n  n
kam n Ran n 

GrWd GrWd GrWd GrWd

The righthand candidate of each pair is misaligned, because the epenthetic vowel
stands at the right edge of the PrWd, thus obstructing the coincidence of morpho-
logical and prosodic edges.
In addition to the left-edge aligning constraint motivated for Axininca Campa
we also need a right-edge aligning constraint. It is stated in (54), after McCarthy
and Prince (1993a):
(54) Align-R
The right edge of a Grammatical Word coincides with the right edge
of a syllable.
Observe that both alignment constraints introduced thus far, Align-L and Align-
R, share a common format, in the sense that two categories (a morphological and
a phonological) are related by one of their edges (left or right). We will return to
this generalized format of alignment constraints in section 3.4.
Align-R may be considered undominated in Lenakel because it is never vio-
lated by epenthesis. We turn to the tableau of [ar.
a.rpkh] to illustrate its role.

(55) Input: /ark-ark/ Align-R *Complex Dep-IO No-Coda

a. ar.
arkh *! **
b. + ar.
a.rpkh * **
c. ar.
ar.
p *! * **
d. a.rp.
a.rpkh **! *
e. a.rp.
a.rp.
p *! ***

113
Syllable structure and economy

The non-epenthetic (fully faithful) candidate (55a) is naturally ruled out by

*Complex. Of the remaining candidates, two (55c, e) violate Align-R. The
choice among the remaining two candidates (55b, d) is made on the basis of
economy of epenthesis. Only one vowel sufces to avoid complex syllable mar-
gins, as in (55b).
Is there any evidence for left-alignment in Lenakel on the basis of epenthesis?
In fact Align-L is never violated, as epenthesis never applies word-initially. With
Align-L undominated, the effects attributed earlier to No-Coda (see tableau 43)
may now be reattributed to Align-L. See the diagram in (56):

(56) /t-n-ak-ol/
PrWd PrWd
Prosodic organization

tna
l
tna
l

Morphological organization
GrWd GrWd

A tableau showing the activity of Align-L is given below:

(57)
Input: /t-n-ak-ol / Align-L Align-R *Complex Dep-IO No-Coda

a. tna.
l *! *
b. + tp.na.
l * *
c. pt.na.
l *! **
d. tp.na.
.lp *! **

(This does not affect the ranking argument given earlier for Dep-IO
No-Coda,
because epenthesis never applies to avoid codas, regardless of position.)

Having ruled out epenthesis at both edges of the word, we now turn to yet another
potential output form for (52b), *[ka.mpn.Ra.npn], which differs from the optimal
output [kam.np.Ra.npn] only in the position of the leftmost epenthetic vowel. The
diagram below shows both candidates:
114
 3.3 Epenthesis and the conict of well-formedness and faithfulness

(58) /kam-n-Ran-n/

kam n Ra n  n
ka mn Ra n  n

What we see is that epenthesis harmonizes with the morphological structure
of the word in the following way: ideally each morpheme begins a new syl-
lable; epenthesis helps to attain this goal by inserting a vowel directly after a
morpheme-initial consonant, rather than before it.15 This is yet another form of
morpho-prosodic alignment, now between the edges of morphemes and those
of syllables. More precisely, we may state this as:
(59) Align-Morph-L
The left edge of a morpheme coincides with the left edge of a
syllable.
If this were undominated, then no syllable would contain segmental material
belonging to different morphemes. (For maximally one morpheme can be left-
aligned with a syllable.) Considering the syllabication of the optimal output
[kam.np.Ra.npn], we see that Align-Morph-L cannot be undominated. Here the
nal morpheme /-n / is not left-aligned with a syllable, but instead shares a syl-
lable with the nal consonant of the preceding morpheme /-Ran-/.

To nd out by which constraint Align-Morph-L is dominated, we must imagine

what the ideal syllabication of this word would be, if it were undominated.
Whatever rules out this form must be the dominant constraint that we are looking
for. It is easy to see that this is *[kam-np-Ran-np], a form in which all morphemes
are left-aligned with a syllable. This output crucially differs from the optimal
(attested) output in the position of the second epenthetic vowel. This stands in
nal position of the word, violating Align-R. The ranking argument immediately
emerges:
15
Lynch (1974: 89) states the generalization about medial vowel epenthesis as follows: Clusters
of three non-syllabics are not permitted in Lenakel surface representations. When they do
arise, generally at morpheme boundaries, schwa is obligatorily inserted between the second
and third member of the cluster of non-syllabics . . . Now when the morphemes with an initial
two-consonant cluster are immediately preceded by a consonant-nal morpheme, schwa is
inserted between the second and third consonants of the resultant cluster. All examples of
medial epenthesis in triconsonantal clusters which Lynch presents indeed have a morpholog-
ical structure of the following kind: the rst consonant belongs to a different morpheme from
the other consonants, for example / kam-n-Ran-n / for her brother, /s-t-pn-aan/ dont go
up there!, or / k-ar-pkom / theyre heavy. These three forms surface as [kam.n.Ra.nn],
[s.d.b.nan], and [kar.b.
m], respectively.

115
Syllable structure and economy

(60) Align-R
Align-Morph-L

kam.np.Ra.npn
kam.np.Ran.np
In the competition among morphemes for prosodic alignment, the grammatical
word is stronger than the morpheme.

Next, Align-Morph-L can also be ranked with respect to Dep-IO. The argument
is based on the forms [ar.
a.rpkh] and [tp.na.
l]. Both forms violate Align-
Morph-L to some extent. To repair the misalignment of morphemes and
syllable edges, one might attempt to add epenthetic syllables, resulting in [a.rpkh-
a.rpkh] and [tp-np-akh-l]. But this leads to fatal additional violations of Dep-IO:

(61) Dep-IO
Align-Morph-L

/ark-ark/ ar.
a.rpkh
a.rp
.a.rpkh
/t-n-ak-ol / tp.na.
l
tp.np.a
.l
That is, economy of epenthesis has higher priority than morpho-syllabic alignment.
It is not possible to rank Align-Morph-L with respect to No-Coda. Therefore
we rank both constraints at the bottom of the hierarchy, which now takes its nal
form:
(62) Final ranking for Lenakel epenthesis
Align-L, Align-R, *Complex
Dep-IO
Align-Morph-L, No-
Coda
In sum, three constraint types interact in the Lenakel pattern of epenthesis: (i)
alignment constraints, (ii) syllable well-formedness constraints, and (iii) faithful-
ness constraints.
The correctness of this ranking is demonstrated by the tableaux (635). The
rst tableau evaluates candidates for the input /kam-n-Ran-n /. The non-epenthetic
candidate (63a) fatally violates *Complex. Its competitors all violate Dep-IO to
the same extent in having two epenthetic vowels, but their evaluation is drastically
different.
(63)
Input: / kam-n-Ran-n / Align- Align- *Complex Dep-IO Align- No-
L R Morph-L Coda

a. kamn.Rann *!* ** **

b. + kam.np.Ra.npn ** * **

c. ka.mpn.Ra.npn ** **! **

d. kam.np.Ran.np *! ** **

116
 3.4 Generalized Alignment

The single candidate that has no violations of Align-Morph-L is eliminated by

Align-R (63d). Among the remaining candidates (63bc), the former is selected
as it has one fewer violation of Align-Morph-L than its competitor.
Both following tableaux demonstrate the ranking of Dep-IO over Align-
Morph-L. The topmost pairs of candidates of both tableaux are eliminated by
*Complex. Of the remaining set of candidates, the most economical ones in
terms of Dep-IO, (64c) and (65c), are selected at the expense of candidates (64d)
and (65d), which are perfectly aligned.
(64)
Input: /ark-ark/ Align- Align- *Complex Dep-IO Align- No-
L R Morph-L Coda

a. ar.
arkh *! * **
b. ar
.arkh **! **
c. + ar.
a.rpk h
* * **
d. a.rp
.a.rpkh **! **

(65)
Input: /t-n-ak-ol/ Align- Align- *Complex Dep-IO Align- No-
L R Morph-L Coda

a. tna.
l *! *** *
b. tna
.l *! ** **
c. + tp.na.
l * ** *
d. tp.np.a
.l **! **

This concludes the analysis of epenthesis in Lenakel. We have seen how the
interaction of three types of constraints (syllabic well-formedness, faithfulness,
and morpho-prosodic alignment) produces a complex pattern of epenthesis. Our
most important ndings are economy (epenthesis only when necessary) and
alignment (the edges of morphological constituents and prosodic constituents
must coincide). The next subsection elaborates on the second nding, extending
the notion of alignment to new applications.

3.4 Generalized Alignment

The notion of alignment originates in phrasal phonology, where it was devel-
oped in the analysis of relations between syntactic constituents and prosodic phrases
117
Syllable structure and economy

(Selkirk 1986, Selkirk and Shen 1990). This edge-based theory of the syntax
phonology interface was adopted into OT by McCarthy and Prince (1993a), who
claim that the prosodymorphology interface should be dened exclusively in
terms of alignment constraints. Yet, alignment not only serves to match edges of
morphological and prosodic categories, but also edges of phonological constitu-
ents (syllable, foot, PrWd). In recent OT the notion of alignment has assumed a
range of applications that goes far beyond the types of cases seen above.
In all alignment constraints seen so far, Align-L, Align-R, and Align-
Morph-L, the edge of a grammatical category (grammatical word or morpheme)
is paired with the edge of a prosodic category (prosodic word or syllable). See
(66):

(66) GramCat ProsCat Edge

a. Align-L GrWd PrWd Left
b. Align-R GrWd Syllable Right
c. Align-Morph-L Morpheme Syllable Left

In principle, alignment requirements may be stated on other morphological con-

stituents (for example, stem, root, or afx) and/or prosodic constituents (for ex-
ample, foot or mora). This generalized theory states that alignment constraints
come in different types, differing with respect to the type of categories and edge
to which they refer.

3.4.1 The format of alignment constraints

In order to full all these diverse functions, alignment must have a very general
format, which McCarthy and Prince (1993a) name Generalized Alignment. This
is stated below:

(67) Generalized Alignment

Align (Cat1, Edge1, Cat2, Edge2) =def
Cat1 Cat2 such that Edge1 of Cat1 and Edge2 of Cat2
coincide.
Where Cat1, Cat2 ProsCat GramCat
Edge1, Edge2 {Right, Left}

Four aspects of this constraint format merit discussion: (i) categories that may
appear in alignment constraints, (ii) the choice of the edges, (iii) the order in
which the categories appear in the constraint, and (iv) how to evaluate violations
of alignment constraints.

First, the categories Cat1 and Cat2 range over the alphabets of grammatical and
phonological categories, for example:
118
 3.4 Generalized Alignment

(68) GramCat: {Word, Stem, Root, Afx . . .}

ProsCat: {PrWd, Foot, Syllable, Mora . . .}

These categories can also be lled by specic morphemes in the grammars of

individual languages. This involves a weakening of the assumption that con-
straints are universal, at least on its strictest interpretation. However, such weak-
ening seems to be tolerable for the following reasons. First, the format of
alignment constraints is universal, which imposes strong limitations on what ele-
ments afx-specic constraints can refer to (that is, edges of morphological or
prosodic categories). Second, relativization to specic morphemes is limited to
alignment constraints, prohibiting morpheme-specic markedness and faithful-
ness constraints. Third, alignment constraints may assume the function of desig-
nating an afx as either a prex or a sufx, depending on the edge of the word
(left/right) with which it aligns. Such a view is strongly supported by inxation,
as we will see in section 3.4.2.

The second aspect of Generalized Alignment concerns the pairing of edges. In

(67), Edge1 and Edge2 range over two values: Left and Right. Although it is
not logically necessary that alignment constraints relate pairs of identical edges
(that is, both left, or both right) for both categories, such matching is implicitly
assumed by most researchers. Nevertheless, alignment constraints have been pro-
posed which pair opposite edges.16

A third aspect of the format of Generalized Alignment which deserves attention

is the asymmetrical relation between both categories involved. By asymmetrical
we mean that the order in which both categories occur in the constraint state-
ment is not random. The rst category mentioned is connected with a universal
quantier , and the second category with an existential quantier . Accord-
ingly there exists a difference in logical interpretation between the following
alignment types:

(69) a. Align (Stem, R, , R)

For every stem there must be some syllable such that the right
edge of the stem matches the right edge of the syllable.
b. Align (, R, Stem, R)
For every syllable there must be some stem such that the right
edge of the syllable matches the right edge of the stem.

16
An example is Align-Afx (McCarthy and Prince 1993b) which aligns the left edge of a sufx
with the right edge of a PrWd.

119
Syllable structure and economy

The scope difference will turn out to have considerable empirical effects which,
however, we will not be able to illustrate until chapter 4 on metrical theory.
However, here we may already fathom the potential relevance of the asymmetry
by spelling out the conditions under which the different constraints in (69ab) are
satised or violated.
Consider the schematic examples in (70). The rst constraint, Align (Stem, R,
, R), is violated by any stem whose right edge fails to coincide with the right
edge of some syllable. This is the case in (70c), where the rightmost syllable is
separated from the right edge of the stem by one or more extrasyllabic segments.
In (70d) no syllable is present, in which case violation is self-explanatory. The
constraint is satised in (70a) and (70b).

(70) Align (Stem, R, , R) Align (, R, Stem, R)

a. []Stem satised satised
b. [12]Stem satised violated (by 1)
c. [ . . . ]Stem violated violated
d. [ . . . ]Stem violated satised (vacuously)

In contrast, the constraint Align (, R, Stem, R) is violated by any syllable whose

right edge fails to coincide with the right edge of some stem. This is violated
when a syllable is separated from the right edge of a stem by any elements, either
a syllable (70b) or a segment (70c). This constraint may be satised in one of two
ways: a stem is either monosyllabic (70a), or has no syllables. We nd that all
four possible combinations of satisfaction and violation are instantiated.

The fourth issue is that of how to evaluate violations of alignment constraints.

In the literature a number of subissues have been identied, and discussion is
ongoing. In McCarthy and Prince (1993a) it is suggested that evaluation of
alignment constraints has two dimensions, which may be specied for each indi-
vidual constraint.
One dimension, which also applies outside alignment constraints proper, is that
of absolute versus gradient evaluation. Absolute constraints are either satised or
violated, and fail to take into account degree of violation. These either incur one
violation mark, or none. (Nevertheless, there may be multiple independent viola-
tions of an absolute constraint in a single candidate, which are added up in the
relevant cell.) Gradient constraints are different in that they take into account
degree of violation. Their evaluation is always scalar, ranking all candidates by
their relative (dis)satisfaction of a goal. (For example, a constraint may pick out
the candidate in which stress falls on the heaviest syllable, comparing this with
the heaviness of stressed syllables in other candidates.) Thus far in this book, we
have not come across any examples of gradient constraints. When a cell in some
120
 3.4 Generalized Alignment

tableau contained multiple violation marks, this was always due to multiple
independent violations of an absolute constraint in a single candidate, rather than
to gradient evaluation. Alignment constraints are typically evaluated gradiently,
measuring the distance between both edges referred to in the constraint.
Gradient evaluation implies that some unit must be specied by which viola-
tions are measured. If an alignment constraint requires two edges to match, then
are violations of the constraint measured in segments, syllables, or perhaps fea-
tures? No general answer to this question has been given in the literature, and it
has been proposed that the element type is established on a constraint-by-
constraint basis. However, the segment is by far the most common element meas-
ured by alignment constraints, and it seems to be a default. Apparently distances
are measured in syllables only by alignment constraints which refer to feet, locat-
ing stresses with respect to edges, which will be discussed in chapter 4.17

Alignment constraints form a family of constraints. A constraint family is a set

of constraints sharing a general format. Another example of a family of con-
straints is the featural identity constraint Identity-IO[F], requiring identity of the
values of some feature (for example, [voice]), between correspondents in input and
output. Yet another example is the set of context-free markedness constraints that
ban different places of articulation, such as *[lab], *[cor], introduced in chapter 1.
Other examples of constraint families will be introduced later in this book.
The next subsection will illustrate the use of alignment as a gradient constraint
in accounting for inxation in Tagalog. This analysis directly connects to the main
topic of this chapter, syllable well-formedness.

3.4.2 Alignment and inxation in Tagalog

Consider the pattern of -um-inxation in Tagalog (French 1988), which serves as
a further illustration of the role of alignment constraints, as well as their gradient
evaluation. The generalization is that -um- is inxed immediately after the onset
of the word (see 71bc), if it has one; if the word has no onset, -um- becomes a
prex (see 71a):

(71) Tagalog prexal inxation

a. /um + alis/ um-alis leave
b. /um + tawag/ t-um-awag call, pf., actor trigger
c. /um + gradwet/ gr-um-adwet graduate

17
It has been suggested that alignment constraints which refer to units of a prosodic category
Cn are evaluated by units of the next-lower category in the prosodic hierarchy, Cn1. Syllable
and foot indeed stand in such a relation. See chapter 4 for the analysis of word edge effects
in metrical systems, and its relation with extrametricality.

121
Syllable structure and economy

The prex skips over the initial onset, settling itself between a consonant and
vowel in the base, with which it integrates into a well-formed sequence of sylla-
bles. (Tagalog is a language allowing for complex onsets and for codas.) What
inxation apparently avoids is an output in which the vowel of -um- lacks an
onset, while its consonant syllabies as a coda, as in *um-tawag. Avoidance of
ill-formed syllables is what drives -um-inxation.
Apparently the morphology (um-afxation) depends on the prosody of the
output. This is an important insight and we will see that it can be easily expressed
in OT, because of its parallelism. Constraints of morphological well-formedness
are ranked in a single hierarchy together with constraints of syllabic well-formed-
ness. Both types of constraint evaluate the same candidate set, of which a single
candidate is selected as optimal with respect to the total, integrated hierarchy.
Here we can discern a morphological constraint making requirements on the
position of -um- in the word, which conicts with constraints of syllabic well-
formedness. As always, conicts are resolved in conformity with the core prin-
ciples of strict domination and minimal violation.
The analysis by Prince and Smolensky (1993) of this pattern is strikingly sim-
ple, and directly encodes the dominance of prosody over morphological require-
ments. The prosodic constraint that comes into play here is No-Coda, repeated
below from (7):

(72) No-Coda
*C ] (Syllables are open.)

This interacts with a morphological alignment constraint expressing the status of

-um- as an afx that must be as close as possible to the left edge of the word:18

(73) Align-um-L
Align the left edge of -um- with the left edge of the PrWd.

In fact, this is nothing but the requirement that -um- is a prex. This require-
ment is stated in the form of an alignment of -um- with the Prosodic Word. Like
other constraints, this morphological constraint is in principle violable. As we will
see, the violability of Align-um-L is the key to its distribution. All we need to
assume is that this is a gradient constraint, that evaluates candidates as to their
relative success in placing -um- at the left edge.

The following ranking of No-Coda and Align-um-L produces the mixed effect
of inxation in some contexts, and of prexation in other contexts:

18
Align-um-L renames Prince and Smolenskys Edgemost(um, L).

122
 3.4 Generalized Alignment

(74) P-constraint
M-constraint

No-Coda
Align-um-L
The ranking states that Tagalog assigns greater importance to the avoidance of
codas than to the prexal status of the afx -um-. However, -um- asserts itself
maximally as a prex, in the sense that it stands as close as possible to the left
edge of the word, respecting the superordinate requirements of prosody. This is
illustrated by the tableau of gr-um-adwet.

(75) Input: {um, gradwet} No-Coda Align-um-L

a. um.grad.wet ***!
b. gum.rad.wet ***! g
c. + gru.mad.wet ** gr
d. gra.um.dwet ** g r a!
e. gra.dum.wet ** g r a! d
f. grad.wu.met ** g r a! d w
g. grad.we.umt ** g r a! d w e
h. grad.we.tum ** g r a! d w e t

Candidate (75a), where -um- is a genuine prex, has three violations of No-
Coda, one more than necessary. (The present tableau does not evaluate any
candidates having fewer than two violations of No-Coda, since these would
violate other top-ranking faithfulness constraints, such as Max-IO and Dep-IO.)
No improvement with respect to No-Coda is reached in candidate (75b), which
inxes -um- between the rst and second consonants of the base. Each of the
remaining candidates (75ch) incurs two violations of No-Coda. Among these,
the one is selected that minimally violates the next constraint down the hierarchy,
Align-um-L. This is a gradient constraint, violations of which are measured by
segments lying between the left word edge and the afx -um-. The optimal output,
(75c), minimally violates Align-um-L, hence maximally respects the prexal
status of -um-. All remaining candidates (75dh) violate Align-um-L to a larger
degree than is strictly necessary.
As mentioned earlier, this analysis is based on parallel evaluation of prosodic
and morphological constraints. Any theory that separates out the evaluation of
morphological and prosodic well-formedness to different levels cannot capture
the same generalization. For example, compare the way in which rule theory might
deal with the Tagalog pattern. This analysis would involve extraprosodicity of
123
Syllable structure and economy

the word onset, rendering it invisible to afxation. Next, -um- is prexed to the
residue (here the word minus the word onset), which begins with a vowel. Finally,
the word onset is rendered visible again, producing the effect of um-inxation.

(76) a. extraprosodicity gradwet gradwet

b. prexation to residue: gradwet gr-um-adwet
c. undoing extraprosodicity: gr-um-adwet gr-um-adwet

This analysis leaves unexpressed the relation between the prosodic shape of the
afx, -um- (VC), and its surface distribution, immediately after the word onset.
Extraprosodicity is unable to make this connection, since it is formally separ-
ated from the afxation that it serves to potentiate. Under this analysis the fact
that the inx has the skeletal shape VC is an arbitrary property. For example,
an extraprosodicity analysis handles with equal ease an imaginary pattern for a
hypothetical inx with the shape CV, say -mu-, for example in mu-abot, t-mu-
awag, gr-mu-adwet.

(77) a. extraprosodicity gradwet gradwet

b. prexation to residue: gradwet gr-mu-adwet
c. undoing extraprosodicity: gr-mu-adwet gr-mu-adwet

One might argue that such inxations never arise because they would seriously
violate the phonotactic principles of the language (more specically, the require-
ments that onsets be maximally binary, and that hiatus is ill-formed). However,
this is precisely the point to be made against the extraprosodicity analysis: it fails
to express the overall contribution of prosodic well-formedness to the distribution
of the inx -um-.

3.5 The quality of epenthetic segments

Thus far in this chapter we have only considered the distribution of epenthetic
segments, not their phonetic content. But from the viewpoint of Correspondence
Theory, the quality of epenthetic segments is just as interesting as their distribu-
tion. Epenthetic segments are special in that they lack input counterparts. Hence
we expect that their featural realization is due to output factors only. This is
precisely what we nd in most languages: epenthetic segments tend to be realized
as the minimally marked segment, and tend to be subject to contextual colouring
(Selkirk 1981, It 1986, Lowenstamm and Kaye 1986):

(78) a. Epenthetic segments tend to be minimally marked qua feature

composition.
b. Epenthetic segments tend to be contextually coloured.

124
 3.5 The quality of epenthetic segments

There is a close connection between both properties: an epenthetic segment is

featurally dependent. Given the fact that an epenthetic segment has no input
features to be faithful to, their feature content is delegated to markedness con-
straints. Among these are two sets of constraints that come into play.

3.5.1 Minimal markedness and contextual colouring

The rst type of constraints determining segment quality are those expressing
pure featural markedness. The very essence of an epenthetic segment is that it
does not occur in the input, hence it has no lexical feature specication to be
faithful to. With input faithfulness constraints (especially, Ident-IO) irrelevant,
the feature content of epenthetic segments depends on the context-free marked-
ness constraints. This correctly predicts that cross-linguistically, featurally unmarked
vowels such as [i], [], and [] are often selected as epenthetic vowels. In conson-
ants, one nds a similar preference for featural unmarkedness. Many languages
select [], a stop that lacks supralaryngeal specications. Other languages select
the semivowel [j], the consonantal counterpart of [i], or alveolar [t], which has
the least marked place of articulation [cor].
The second type of constraints that can inuence the choice of epenthetic
segments are contextual markedness constraints. Featural values of segments,
either epenthetic or non-epenthetic, may depend on neighbouring segments by
assimilation. Assimilation in a non-epenthetic segment occurs at the expense
of faithfulness to its input specication. Here assimilatory forces are intrinsically
in conict with contrast-preserving forces: the result depends on interactions
between contextual constraints (striving for assimilated outputs), and faithfulness
constraints (striving for maximally faithful outputs). In contrast, when an epen-
thetic segment assimilates to a neighbouring segment, input faithfulness never
comes into play, as the epenthetic segment lacks an input. With faithfulness con-
straints excluded as anti-assimilatory factors, it is predicted that the quality
of epenthetic segments strongly depends on contextual markedness constraints.
In many languages, epenthetic segments indeed undergo considerable colouring
by their contexts.
Is there a contradiction? Both predictions cannot be true at the same time. That
is, assimilation of an epenthetic segment to its context produces a segment that is
no longer minimally marked due to its assimilated marked features. As always,
we expect that the outcome is determined by the interaction of both types of
constraints, context-free and contextual markedness. When the former outrank the
latter, as in (79a), the epenthetic segment displays no contextual variation and is
minimally marked. But under the reverse ranking, in (79b), minimal markedness
is violated in the interest of contextual colouring.

125
Syllable structure and economy

(79) Factorial typology of quality of epenthetic segments

a. Context-free markedness
Contextual markedness
Epenthetic segment is minimally marked.
b. Contextual markedness
Context-free markedness
Epenthetic segment is contextually coloured.

Of course, this is only a schematic typology. In reality the factors in this typo-
logy are not unary, but composite. Context-free markedness comprises the set
of individual featural markedness constraints, grouped in hierarchies that are to
some extent universal. And the factor context-sensitive markedness denotes a
set of constraints whose ranking is, again, partly universal. The interaction will
usually not result in a complete victory of one factor over the other, but a more
fragmented picture will be presented, with contributions made by both context-
free and contextual factors. Once again we turn to Lenakel for exemplication.

3.5.2 The epenthetic vowels of Lenakel

Lenakel, the language analysed in some detail above, actually has two epenthetic
vowels: [] and its allophone []. Both vowels are contextually predictable, and
occur in complementary distribution. So far we have only considered examples
containing epenthetic [], a high central unrounded vowel. This occurs after coro-
nal consonants, dened by the set {t, s, n, r, l}.19

(80) Epenthetic [] appears after coronals

a. /t-n-ak-ol/ tp.na.
l you (sg.) will do it
b. /ark-ark/ ar.
a.rpkh to growl
c. /kam-n-Ran-n/ kam.np.Ra.npn for her brother
d. /r-n-ol/ rp.nl he has done it

Cross-linguistically, high central [] is a fairly frequent choice for an epenthetic

vowel. A contextual relation between coronals and high vowels is cross-linguistic-
ally less common than between coronals and front vowels (Hume 1992). How-
ever, the Lenakel data clearly support such a relation.
Its counterpart is a mid central unrounded vowel [], differing from [] only in
its vowel height. This epenthetic vowel occurs after non-coronal consonants.

(81) Epenthetic [] appears after non-coronals (from Lynch 1974)

a. /to-rm-n/ tr.mqn to his father (p. 87)
b. /apn-apn/ ab.na.bqn free (p. 88)

19
As well as after /v/. This surprising fact becomes more natural upon considering the phonetics
of /v/: a high central glide with varying, though weak, amounts of bilabial articulation
(Lynch 1974: 21). That is, /v/ shares both its centrality and height with [].

126
 3.5 The quality of epenthetic segments

c. /k-ar-pkom/ kar.bq.
m theyre heavy (p. 90)
d. /r-m-n/ r.m.qn he was afraid of him/it (p. 88)

Again, Lenakel is far from unique in having [] as an epenthetic vowel. (Other

languages, such as Dutch, have [] as their only epenthetic vowel.) In Lenakel,
epenthetic [] appears after non-coronals, that is, in the elsewhere context, com-
plementary to the context of []. This accords with schwas status as minimally
marked: a mid central unrounded vowel.
The position of epenthetic vowels in the Lenakel vowel inventory is stated
below:

(82) Lenakel vowel inventory (with epenthetic vowels circled)

front central back i p u e q o a
high i  u high + + +
mid e o low +
low a back + + + + +
round + +

First note that both epenthetic vowels are [low, +back, round], all featurally
unmarked values. More formally, epenthetic vowels never violate any of the fol-
lowing context-free markedness constraints:

(83) + low], *[+

*[+ + round], *[
back]

To pinpoint the difference between epenthetic and lexical vowels, let us address
the issue of what licenses lexical vowels such as low [a], round [o][u], or front
[e][i] in Lenakel. The answer resides in featural faithfulness to the input. What
makes an epenthetic vowel different from a lexical vowel is that only the latter
has input specications, among which are values such as [+low], [+round], or
[back]. Lexical vowels are faithful to their input as Ident-IO constraints for
different features outrank context-free markedness constraints:

(84) Non-reduction of input vowels in Lenakel

Ident-IO(low), Ident-IO(round), Ident-IO(back)
*[+low],
*[+round], *[back]

Ident-IO constraints are irrelevant to epenthetic vowels (because these have

no lexical inputs), and therefore epenthetic vowels must fully comply with context-
free markedness.20

20
Unmarkedness of epenthetic vowels arises in underspecication theory (Archangeli 1988) by
featural emptiness. That is, an epenthetic segment is maximally underspecied.

127
Syllable structure and economy

Next, what causes the contextual variation in vowel height among the epen-
thetic vowels? This variation reects interaction of a context-free markedness
constraint with a contextual markedness constraint. This interaction, which was
schematically presented in (79b) as the ranking of contextual colouring,
involves a pair of markedness constraints. In Lenakel, the dominated member
of the pair is a context-free markedness constraint, violated in contexts where
[] appears. This is a constraint militating against high vowels:

(85) + high]
*[+

This is in competition with the dominant member of the pair, a contextual

markedness constraint requiring that coronals be followed by high vowels:

(86) Cor-high
Coronals are followed by high vowels.

Cor-high sets up the pressure under which *[+high] may be violated. Accord-
ingly, it must dominate *[+high]:

(87) Cor-high
*[+high]

Integrating this interaction with the current ranking, we arrive at the complete
hierarchy:

(88) Contextual variation between epenthetic [] and [] in Lenakel

Ident-IO(low/round/back)
*[+low], *[+round], *[back]
Cor-
high
*[+high]

Strictly speaking, we have no evidence to rank Cor-high below context-free

markedness constraints. This ranking was chosen only because of gains in perspi-
cuity in the tableaux.
Tableau (89) evaluates candidates for an input /t-n-ak-ol/. Five candidates (89a
e) vary in the choice of epenthetic vowel, ranging from [] to [a]. All ve are
faithful to the input vowels, /a/ and /o/. In the columns of context-free marked-
ness constraints, each cell holds the combined violation marks for non-epenthetic
and epenthetic vowels. Therefore each of the candidates (89a e) has at least one
mark for *[+low], due to [a], and another for *[+round], due to []. Additional
marks for *[+low] and *[+round], which are fatal, are due to epenthetic vowels ([a]
in 89e, [u] in 89d). Candidate (89c), with epenthetic [i], fatally violates *[back].
The remaining candidates (89a) and (89b) have epenthetic [] and [], respect-
ively. The former fatally violates context-sensitive Cor-high, incurring two viola-
tions, for t and na, even though it satises context-free *[+high]. The optimal
candidate (89b) has the reverse pattern, violating *[+high] while violating Cor-
high only once (for na):
128
 3.5 The quality of epenthetic segments

(89)
Input: /t-n-ak-ol/ Ident-IO *[+low] *[+round] *[back] Cor-high *[+high]

a. t.na.
l * * **!

b. + tp.na.
l * * * *

c. ti.na.
l * * *! * *

d. tu.na.
l * **! * *

e. ta.na.
l **! * **

f. tp.n.
l *!* **

The bottom candidate (89f) represents the reduction of the input vowels /a/ and
/o/ to [] and [], respectively. These are the least marked vowels, taking into
account contextual factors for this form. Such a reduction is excluded by Ident-
IO, a shorthand notation for three Ident-IO constraints with respect to individual
features [low][round][back].
Next consider tableau (90) of the form [tr.mn]. Again a subset of candidates
(90ce), containing either front, rounded, or low epenthetic vowels, are excluded
by the context-free constraints. The bottom candidate (90f) fatally violates featu-
ral faithfulness. The remaining pair (90ab) are evaluated vacuously by Cor-high
since only the lexical vowel violates it, while the epenthetic vowel is not in the
relevant context (following a coronal). These candidates are submitted to the next-
lower constraint, *[+high], which selects the former (90a) because its epenthetic
vowel [] is non-high:
(90)
Input: /to-rm-n/ Ident- *[+low] *[+round] *[back] Cor-high *[+high]
IO

a. + tr.mqn * *
b. tr.mpn * * *!
c. tr.min * *! * *
d. tr.mun **! * *
e. tr.man *! * *
f. tr.mqn *! *

This tableau shows that schwa [] is the minimally unmarked vowel with respect
to the context-free markedness constraints. Hence, it appears as the epenthetic
vowel wherever possible: wherever contextual constraints do not stand in its way.
129
Syllable structure and economy

Our discussion of the factors governing the quality of epenthetic segments

leads to the following conclusions. The quality of epenthetic segments is min-
imized, to avoid unnecessary violations of featural markedness constraints. This
follows straightforwardly from the natural assumption that epenthetic segments
are not represented in the input, and from the principle of minimal violation, a
cornerstone of OT. The choice of epenthetic segments depends on interactions of
contextual and context-free markedness constraints.
The second conclusion is that markedness constraints jump into activity in
special situations where faithfulness constraints are irrelevant. This situation
arises for a segment lacking an input counterpart, that is, an epenthetic seg-
ment. The minimized markedness of epenthetic segments is due to the activity
of featural markedness constraints, which are normally dormant due to do-
minating featural faithfulness constraints. Such a situation has been referred to
by McCarthy and Prince (1994a) as the emergence of the unmarked. (See
chapter 5 for extensive discussion of this effect.)

3.5.3 Epenthetic morphemes

Languages occur in which the choice of epenthetic segments is not completely a
matter of the emergence of the unmarked, but is to some extent morphologically
governed.21 For example, Mohawk (Piggott 1995) has three epenthetic vowels [e],
[a], [i], which have a partly morphological, and partly phonological distribution.
We may call such elements epenthetic morphemes (or augments) since they have
minimal or zero semantic content. Lexical input specication of augments renders
faithfulness constraints relevant to their realization, for which reason augments
behave like lexical vowels. The other interesting issue is the factors governing
the distribution of the contextual variants of augments, their allomorphs. We
assume an allomorph to be a lexically listed shape of a morpheme, which cannot
be derived by the productive phonology of the language. We then predict that
the different allomorphs, being semantically non-distinct from one another, have
a distribution that is uniquely governed by phonological markedness. Such phono-
logically-driven allomorphy is attested in many languages (McCarthy and Prince
1993b, Mester 1994, Kager 1996). See chapter 9 for further discussion.

3.6 Coda conditions

To wind up this chapter, we consider in more detail the phonotactic role of the
syllable. Many languages restrict the type of consonant that may occur in the
syllable coda (Prince 1984, It 1986, 1989, Clements 1990, Goldsmith 1990). In

21
Steriade (1995a) mentions Hindi, Hungarian, Basque, Maltese Arabic, and Tiberian Hebrew
as languages in which epenthetic vowels have lexical properties.

130
 3.6 Coda conditions

chapter 1 we encountered a case of coda neutralization in Dutch, a language

disallowing voiced obstruents in codas. But such coda conditions are much
more general: in this section we will look into cases from other languages. Our
goal will be to strengthen a conclusion reached in chapter 2: that phonological
processes may be functionally related, which is captured straightforwardly in OT.
We will discuss three strategies which languages employ to resolve violations of
a coda-condition, a constraint militating against place features in the syllable
coda.

3.6.1 Homorganic clusters in Japanese

A fairly common type of restriction on coda consonants is that these must share
the place of articulation with the immediately following onset consonant. This coda
condition is in force in Japanese (It 1989). In intervocalic clusters, the coda con-
sonant is either the rst half of a geminate, or a nasal preceding a homorganic stop:

(91) Homorganic clusters in Japanese

a. Geminates b. Homorganic Nasal +
Obstruent
kap.pa a legendary being tom.bo dragony
kit.te stamp non.do tranquil
gak.koo school ka.gae thought
c. Excluded hypothetical cases: *kap.ta, *tog.ba, *pa.kap, etc.

The constraint militating against codas having independent place of articulation

is stated below, after It (1989):

(92) Coda-Cond
*Place]

The intended interpretation of this constraint is: a coda cannot license place
features.22 That is, a coda can contain place features only when these are simul-
taneously associated with a syllable position outside the coda, such as the second
syllables onset in (93a):

(93) a. b.

C V C C V C V C C V

k a p a k a k p a
22
This slightly simplies the discussion, abstracting away from the Linking Constraint, which
in It (1989) accounts for the distinction between (93a) and (93b). Prince (1984), Goldsmith
(1990), and It and Mester (forthcoming) propose various technical explanations of homorg-
anicity. See chapter 9 for a radically different approach, known as positional faithfulness.

131
Syllable structure and economy

Shared place of articulation between the coda and the following onset is therefore
not in violation of the Coda-Cond. In contrast, the /k/ in the coda of structure
(93b) violates it.
Although no alternations are present, we must allow the possibility of con-
sonant clusters in the input whose members have different places of articulation,
e.g. /kakpa/. This follows from the assumption of Richness of the Base: OT has
no constraints on inputs. Any (hypothetical) heterorganic inputs should be over-
ruled by Coda-Cond. This implies that one consonant in the cluster must lose its
place features, thereby violating:

(94) Ident-IO(Place)
Correspondents in input and output have identical place features.

This constraint is dominated by Coda-Cond:

(95) Place assimilation to resolve violations of the coda condition

Coda-Cond
Ident-IO(Place)

Now assume a hypothetical input /kakpa/. A strictly faithful analysis of this input
would be [kak.pa], with the rst syllables coda violating Coda-Cond. The opti-
mal analysis eliminates the place features of one of the consonants in the cluster
/kp/. See tableau (96).

(96) Input: / kakpa/ Coda-Cond Ident-IO(Place)

a. + kap.pa *
b. kak.pa *!

Observe that a morpheme-internal input cluster such as /kp/ would be completely

obscured by this ranking. Since there would be no evidence for the input quality
of the consonant cluster, there is no more suitable input for the output form
[kap.pa] than / kappa/ itself, which is identical to the output. Lexicon Optimiza-
tion would select / kappa/ since its mapping relation to the output [kap.pa]
involves no violation of Ident-IO(Place).
This observation actually points to a hidden weakness in this analysis: it is
unable to predict which of the two input consonants is charged for violation of
the coda condition. (That is, [kak.ka] would be an equally harmonic outcome.)
However, typological evidence suggests that the choice is not random. Cases
from other languages involving alternations display a strong preference for the
rst consonant in the cluster (the coda) to lose its place features, rather than the
second one. (In section 3.6.3 we will look into a case involving deletion.)
132
 3.6 Coda conditions

3.6.2 Vowel epenthesis in Ponapean

We have now seen a static coda condition, one unaccompanied by alternations.
Let us now look into a case which involves alternations, as well as a second
strategy to resolve violations of the coda condition. Ponapean (Rehg and Sohl
1981, It 1989) epenthesizes a vowel ([e] or [i]) into heterorganic consonant
clusters which arise by the morphology, to avoid violation of the coda condition.
Like Japanese, Ponapean enforces coda consonants which are homorganic with
the following onset consonant.

(97) Vowel insertion in Ponapean

a. /ak-dei/ a.ke.dei *ak.dei a throwing contest
b. /kitik-men/ ki.ti.ki.men *ki.tik.men rat indef.

The epenthetic vowel makes the lefthand consonant in the cluster to syllabify
as an onset, rather than as a coda. As an onset, this consonant can maintain its
place features without violating Coda-Cond, while at the same time satisfying
Ident-IO(Place). But of course, these benets come at a cost: a violation of the
anti-epenthesis constraint Dep-IO. This scenario is expressed by the following
ranking:

(98) Vowel epenthesis to resolve violations of the coda condition

Coda-Cond, Ident-IO(Place)
Dep-IO

The violation of Dep-IO, the lower-ranked of both faithfulness constraints, is

the least expensive way to satisfy the coda condition. (Japanese has reversed
priorities.)
Consider three output candidates for the input /ak-dei/ in tableau (99). The fully
faithful candidate (99a), which preserves place of articulation while avoiding
epenthesis, fatally violates Coda-Cond. To avoid such a violation, two strategies
are considered, as represented by candidates (99bc). Both violate some faithful-
ness constraint. The former, (99b), with a homorganic cluster, violates Ident-
IO(Place). Its competitor (99c), with an epenthetic vowel, violates Dep-IO.
Because of strict domination, the lower-ranking of the two faithfulness constraints
gives in:

(99) Input: /ak-dei/ Coda-Cond Ident-IO(Place) Dep-IO

a. ak.dei *!
b. ad.dei *!
c. + a.ke.dei *

133
Syllable structure and economy

Observe that although vowel epenthesis leads an output syllabication that has no
coda, it is not triggered by No-Coda. (Ponapean allows for homorganic coda
onset sequences.)
Finally we must address the fact that the coda condition is violable in Ponapean
in a specic context: the nal consonant of the word, a coda itself. What blocks
epenthesis in word-nal position? Arguably, this blocking effect is due to Align-
R, as in Lenakel (see section 3.3.4.2). This implies the subranking Align-R

Coda-Cond, which is integrated in the nal ranking below:

(100) Final ranking for Ponapean

Align-R, Ident-IO(Place)
Coda-Cond
Dep-IO

A tableau shows the interaction of these four constraints for [na.kep] inlet:

(101)
Input: /nakep/ Align-R Ident-IO(Place) Coda-Cond Dep-IO

a. na.ke.pi *! *
b. na.ke *!
c. + na.kep *

Candidate (101a) fatally violates Align-R. Another strategy to satisfy the coda
condition, seen in candidate (101b), is the deletion of place features from the
coda consonant, which yields a glottal stop []. This runs into a fatal violation of
Ident-IO(Place), however.23

3.6.3 Consonant deletion in Diola-Fogny

A third logically possible strategy to avoid violation of the coda condition is the
deletion of a consonant in a heterorganic cluster. This strategy is employed by
Diola-Fogny (Sapir 1965, Steriade 1982, It 1986):

(102) Consonant deletion in Diola-Fogny

a. /let-ku-jaw/ le.ku.jaw *let.ku.jaw they wont go
b. /jaw-bu-ar/ ja.bu.ar *jaw.bu.ar voyager

It readily appears that the core of the analysis must be a domination of Coda-
Cond over the anti-deletion constraint Max-IO:

23
Ponapean allows homorganic nal clusters, for example, mand tame, emp coconut crab.
Here alignment blocks nal epenthesis (*[em.pi]), while economy rules out medial epenthesis
(*[em.ip]), which fails to cancel the violation of the coda condition. Note that homorganicity
of nal clusters is not predicted under an extrametricality analysis (It 1989).

134
 3.6 Coda conditions

(103) Consonant deletion to resolve violations of the coda condition

Coda-Cond
Max-IO
le.ku.jaw
let.ku.jaw

We cannot present the complete analysis before having addressed two interesting
issues, each of which merits detailed discussion: the choice of the deleted con-
sonant in a cluster, and the immunity of the word-nal consonant. Both issues
have consequences for the view of the morphologyphonology interface.
The rst issue is the choice of the deleted consonant in the cluster. Observe
that the lefthand consonant is deleted, while the righthand consonant is preserved.
A possible interpretation of this is that a coda consonant, rather than an onset
consonant, is deleted. But how can a coda-onset asymmetry on deletion be
accounted for in a theory based on output constraints? Should we rst syllabify
the input to identify the deleted consonant as a coda, before we can delete it?

(104)

/let-ku-jaw/ let.ku.jaw le.ku.jaw

This would reintroduce a derivation, thereby giving up the idea that constraints
evaluate surface forms only.
But another interpretation of the deletion asymmetry is possible, one requiring
no intermediate level of representation.24 Observe that the deleted consonant is
nal in its morpheme, while the preserved consonant is morpheme-initial. This
deletion pattern may serve to keep syllable structure maximally isomorphic to
morphological structure. To see this, we must consider the two output forms
which arise by deletion of either the lefthand consonant (105a) or the righthand
consonant (105b). The syllable structure is represented on top of the segments,
while morphological structure is represented below:

(105) a. b.

le ku jaw (perfectly aligned) le tu jaw (mis-aligned)

The preference for the lefthand candidate is another case of morpho-prosodic

alignment, of which we have seen earlier cases in this chapter. (See again the
discussion of Lenakel.) The constraint enforcing alignment of morphological
edges and syllable edges is Align-Morph-L (repeated below from 59):

24
This analysis misses the generalization that a coda consonant is deleted, rather than an onset
one. Chapter 9 will sketch an alternative analysis based on positional faithfulness.

135
Syllable structure and economy

(106) Align-Morph-L
The left edge of every morpheme coincides with the left edge of a
syllable.

The ranking of this constraint with respect to Coda-Cond
Max-IO cannot be

exactly established. This is due to the fact that Align-Morph-L is not in conict
with either of the two constraints: it only serves as a tie-breaker between two
deletion candidates (107a) and (107b) which equally satisfy (or violate) Coda-
Cond and Max-IO.

(107) Input: /let-ku-jaw/ Align-Morph-L Coda-Cond Max-IO

a. let.ku.jaw **!
b. + le.ku.jaw * *
c. le.tu.jaw *! * *

Since Align-Morph-L is not violated in any example presented earlier, we will

assume it to be undominated.

The second issue is the preservation of the word-nal consonant. Apparently the
grammar of Diola-Fogny exerts pressure to leave the right edge of the grammat-
ical word unaffected. This actually overrides the coda condition. What constraint
is responsible for the preservation of the rightmost segment in the word?
In earlier sections (section 3.3.4.2 on Lenakel and section 3.6.2 on Ponapean)
we encountered an alignment constraint that might, in principle at least, do the
job. This constraint is Align-R, requiring that the right edge of every grammat-
ical word coincide with the right edge of a PrWd. See the diagrams below:

(108) a. PrWd b. PrWd

le ku jaw (right-aligned) le ku ja (right-aligned)

GrWd GrWd

But now observe that both structures are perfectly aligned since the rightmost
segment of each GrWd matches the rightmost segment of PrWd. In (108a) this
rightmost segment is [w], and in (105b) it is [a]. Both are the rightmost remaining
136
 3.6 Coda conditions

segments of the grammatical word, which are rightmost in the PrWd. Therefore
standard alignment does not work.
What we are looking for is a novel type of alignment, which actually inspects
the correspondence relation between the rightmost segments in the input and
output. Input output diagrams of the two major candidates clarify the kind of
correspondence relation which is relevant. This relation is intact in (109a), while
it is broken in (109b):

(109) a. Input: l e t ku j aw b. Input: l e t ku j aw

Output: l e ku j aw Output: l e ku j a
The relevant correspondence constraint focusses on the rightmost segments in the
input and output. It identies the rightmost segment of the output grammatical
word, and then it identies the rightmost segment of the input grammatical word.
When the segments are correspondents, as in (109a), then the constraint is
satised. But when the segments are not correspondents, as in (109b), then the
constraint is violated.
This correspondence constraint is Anchoring-IO (McCarthy and Prince
1995a):

(110) Anchoring-IO(GrWd, R)
Any segment at the right periphery of the output GrWd has a cor-
respondent at the right periphery of the input GrWd.
(No deletion/epenthesis at the edge.)

This adds another correspondence constraint to the current inventory, which

already has four: Dep-IO, Max-IO, Ident-IO, and Linearity-IO.25 The function
of Anchoring-IO is to block deletion at the edge of a grammatical word.
Interestingly, Anchoring-IO militates against epenthesis at word edges as
well. Any non-correspondence of segments in the input and output at the right
edge will violate Anchoring-IO, regardless of whether it is caused by deletion
or insertion of a segment. To block epenthesis at word edges in Axininca Campa,
Lenakel, and Ponapean, we have thus far assumed Generalized Alignment con-
straints, Align-L and Align-R. It now seems that our theoretical machinery is
becoming redundant, as the two tools may do the same job. McCarthy and Prince
(1995a, forthcoming) propose to merge both into a single constraint format, argu-
ing that Anchoring subsumes Generalized Alignment.26 We will not pursue this
issue here, but return to the analysis of Diola-Fogny.

25
Not counting Uniformity-IO, which was briey discussed at the end of chapter 2.
26
See chapter 5 (section 5.7), for more on this unication.

137
Syllable structure and economy

Having dened the constraint which blocks deletion of word-nal segments, we

must now rank it. The ranking of Anchoring-IO with respect to the core inter-
action of Coda-Cond
Max-IO is readily established. As we observed above,
its function is to block consonant deletion at the right word edge, resisting the
pressure from Coda-Cond. Hence we nd the following interaction:

(111) Blocking of coda deletion at the word end

Anchoring-IO(GrWd, R)
Coda-Cond
le.ku.jaw
le.ku.ja
This result is now integrated into the previously assumed ranking:

(112) Final ranking for Diola-Fogny

Anchoring-IO(GrWd, R)
Coda-Cond
Max-IO

The tableau of [le.ku.jaw] shows its two main competitors, [let.ku.jaw] and
[le.ku.ja], as well as their evaluation under the hierarchy.

(113) Input: /let-ku-jaw/ Anchoring-IO Coda-Cond Max-IO

(GrWd, R)

a. let.ku.jaw **!
b. + le.ku.jaw * *
c. le.ku.ja *! **

Note that this exemplies a perfect twofold domination: each lower ranking con-
straint is violated to avoid violation of a higher-ranking constraint. The nested
logical structure of the generalization is stated below:
(114) Input consonants are preserved,
except when these have place features and are syllabiable as codas,
except when these stand at the right edge of a grammatical word.
This nishes the analysis of Diola-Fogny consonant deletion. In the discussion
section below we will return to the relation of this pattern to the other two
patterns (assimilation and epenthesis) discussed earlier in this section.

3.6.4 Coda conditions: a comparison with rule-based theory

In this section we will address the functional relation among the three processes
which were discussed in this section, that is: place assimilation in Japanese, vowel
epenthesis in Ponapean, and consonant deletion in Diola-Fogny. All three pro-
cesses are depicted below in rule format, using the notational means of auto-
segmental phonology:
138
 3.7 Conclusion

(115) a. Place assimilation: Coda Onset

=
[Place] [Place]

b. Vowel epenthesis: V/ Coda Onset

[Place] [Place]

c. Consonant deletion: Coda / Onset

[Place] [Place]

All three rules share a common structural description: all are triggered by a coda
that fails to share its place of articulation with the following onset. All function
to eliminate this conguration, by different structural changes. The rst rule
assimilates the onsets place features back to the preceding coda, making it share
its features. The second rule inserts a vowel between the onset and the coda,
turning the offending coda consonant into an onset. And the third rule atly
deletes the coda consonant.
However, there is no formal recognition of this functional unity among the
rules. The argument is analogous to the one made in chapter 2 for the processes
applying in the *Nt (nasal-plus-voiceless-consonant) context. To capture the
functional unity of a set of rules, it does not sufce for these rules to share a
structural description. Nor does it sufce for their outputs to lack codas with
place features. The proper generalization combines both aspects: all rules function
so as to eliminate the conguration that forms their structural description. Any
theory that fails to recognize the output as a level at which phonological general-
izations hold fails to capture the functional unity of these phonological processes.
In contrast, OT captures this functional unity straightforwardly, thereby creating
unity in typological diversity.

3.7 Conclusion
Syllables are key factors in stating phonological generalizations. In this chapter
we have focussed on two types of phenomena in which syllables play a major
role: epenthesis and the phonotactics of coda consonants. Throughout this chap-
ter we have kept an eye on a set of more general issues in OT:
Process format. We have learned that syllable-dependent processes, like any
other phonological processes, involve interactions of markedness constraints
and faithfulness constraints. A change (deletion, insertion, or featural change)
is made at the cost of some faithfulness constraint, but only to avoid the violation
139
Syllable structure and economy

of higher-ranking well-formedness constraints. Correspondence Theory denes a

set of constraint families that militate against changes of various kinds dele-
tions, insertions, featural changes, etc. The way in which the grammar reacts in
order to avoid a violation of some well-formedness constraint depends on the
relative ranking of the correspondence constraints.
Process economy. Changes are always economical faithfulness rules out any
change on top of those which are strictly necessary. We have seen this economy
property in two aspects of epenthesis: its distribution and the quality of the epen-
thetic segment.
Language typology. Languages differ in terms of constraint rankings. Therefore
to evaluate a set of constraints on a cross-linguistic scale, constructing factorial
typologies is essential. We have found valuable typological results with respect to
syllable structure. In a nutshell, languages all strive toward the ideal CV syllable
format; more complex types of syllable structure involve a cost in markedness
constraints which is only enforceable through faithfulness constraints.
The following chapter will generalize this approach one level up in the pro-
sodic hierarchy.

SUGGESTIONS FOR FUR THER READING

The syllable in phonological theory

Blevins, Juliette (1995) The syllable in phonological theory. In Goldsmith. 206
44.
Clements, George N. and Samuel J. Keyser (1983) CV phonology: a generative
theory of the syllable. Cambridge, Mass.: MIT Press.
Hayes, Bruce (1989) Compensatory lengthening in moraic phonology. Linguistic
Inquiry 20. 253306.

Syllables in OT
Broselow, Ellen, Su-I Chen, and Marie Huffman (1997) Syllable weight: conver-
gence of phonology and phonetics. Phonology 14. 4782.
Rosenthall, Sam (1997) The distribution of prevocalic vowels. Natural Language
and Linguistic Theory 15. 13980.

Alignment and syllabication

Kenstowicz, Michael (1994b) Syllabication in Chukchee: a constraints-based
analysis. In A. Davison, N. Maier, G. Silva, and W. S. Yan (eds.), Proceed-
ings of the Formal Linguistics Society of the Midwest 4. Iowa City: Univer-
sity of Iowa. 16081.
140
 3.7
3.7Conclusion
Exercises

McCarthy, John and Alan Prince (1993a) Generalized Alignment. In G. Booij

and J. van Marle (eds.), Yearbook of Morphology 1993. Dordrecht: Kluwer.
79153.
Merchant, Jason (1996) Alignment and fricative assimilation in German. Lin-
guistic Inquiry 27. 70919.

EXERCISES

1 -um-inxation in Ilokano
Ilokano (Hayes and Abad 1989) has a pattern of um-inxation that is almost
identical to that of Tagalog, discussed in section 3.4.2. For Ilokano it is reported
that speakers have the free variant g-um-radwet next to gr-um-adwet.

a. What ranking of constraints would produce the output g-um-radwet?

Are No-Coda and Align-um-L sufcient, or do we need any addi-
tional constraint(s)?
b. Compute a factorial typology of the constraints that you have used
in your analysis.

2 Epenthesis in Harari
Consider the following two paradigms of the root /sbr/ break in Harari (Leslau
1958, Kenstowicz and Kisseberth 1979):

Imperfect Negative imperfect

2 masc. tisbri /t-sbr/ ztsibr /z-t-sbr/
2 fem. tisbri /t-sbr-i/ ztsibri /z-t-sbr-i/
3 masc. yisbri /y-sbr/ zysibr /z-y-sbr/
3 fem. tisbri /t-sbr/ ztsibr /z-t-sbr/
1 pl. nisbri /n-sbr/ znsibr /z-n-sbr/
2 pl. tisbru /t-sbr-u / ztsibru /z-t-sbr-u /
3 pl. yisbru /y-sbr-u / zysibru /z-y-sbr-u /

For the purposes of the problem, you are free to ignore the variation in the stem
shapes (/sbr/ versus /sbr/), which is under morphological control.

a. What is the ranking of constraints underlying this pattern of epen-

thesis? Do you need any new constraints? (Hint: why not *tisbir
instead of tisbri?)
b. Support your analysis by tableaux of tisbri and ztsibr.

141
 4
Metrical structure and parallelism

4.1 Introduction
This chapter deals with word stress patterns in the framework of OT. Within
phonology, word stress patterns stand out as being typically governed by conict-
ing forces, as was recognized in studies in pre-OT metrical phonology. Before we
can actually point out the interactions of conicting metrical constraints, we must
rst become familiar with some results of the metrical theory of word stress.

Our representational basis is metrical phonology, a theory developed in the 1970s,

the central assumption of which is that stress is a rhythmic phenomenon, encoded
by strongweak relations between syllables (Liberman and Prince 1977, Hayes
1980, 1995, Halle and Vergnaud 1987). Consider the stress pattern of the word
Alabama, which contains two pairs of syllables standing in such a strongweak
relationship [F.l.bE.m]. We analyse this word rhythmically into two metrical
feet of two syllables each [.l] and [b.m]:

(1) (. * ) PrWd-level
(* .)( * .) Foot-level
.l.b.m Syllable-level

The rhythmically strong syllables are initial in the foot: such feet are called
trochees. Each foot is represented by a pair of parentheses, and an asterisk on top
of the strong syllable, while the weak syllable has a dot. We will refer to the
strong syllable of a foot as its head, and to the weak syllable as its non-head. Note
that the strong syllable of the nal foot is more prominent than that of the initial
foot. This weakstrong relationship between feet is represented by yet another
metrical constituent, on the next-higher prosodic level that of the PrWd. Again,
relative prominence is marked by a dot and an asterisk, now sitting on top of
heads of feet.
Extensive research in the typology of word stress (e.g. Hayes 1995) has shown
that stress patterns are a domain of potentially conicting forces, among which
are rhythm (the pressure toward regularly alternating distributions of strong and
142
 4.2 Word stress: general background

weak syllables), quantity-sensitivity (the pressure to match syllable weight and

prominence), and edge-marking (the pressure to mark the edges of morphological
domains by strong syllables). OT is naturally equipped to capture interactions of
conicting forces, and to establish interactions between prosodic levels, by its
parallelism. Finally, OTs reranking approach to typology makes it a highly suit-
able basis to capture cross-linguistic variation between metrical systems. This
chapter will develop the analytical tools by which OT achieves these ends.

We will proceed as follows. In section 4.2, we will become familiar with the
major cross-linguistic forces in word stress, among which are rhythm, quantity-
sensitivity, and edge-marking. This section will also introduce the representational
tools of metrical theory, the universal alphabet of prosodic categories, including
the syllable, the foot, and the PrWd. In section 4.3 we will present a case-study
of Hixkaryana, and see how this language accommodates the conicts between
metrical forces. In the course of this case-study, metrical forces will be trans-
lated one by one into a small set of metrical constraints. These constraints will be
placed in a broader typological perspective in section 4.4, where a number of
basic interaction schemata will be introduced. Section 4.4 also develops a facto-
rial typology by reranking a subset of metrical constraints, matching it against
what is known about cross-linguistic variation in stress systems. A case-study of
another language, Southeastern Tepehuan, will serve to highlight a comparison
with rule-based parametric metrical theory in section 4.5. Finally, section 4.6 will
deal with the interactions between metrical constraints and faithfulness con-
straints, which point to parallellism and a typical OT perspective.

4.2 Word stress: general background

4.2.1 Four cross-linguistic preferences

Before developing an OT theory of word stress, we must rst familiarize our-
selves with a number of important cross-linguistically common properties of
stress languages. The best known of these properties are the following four.

The culminative property. It is typical for stress languages that morphological or

syntactic constituents (stem, word, phrase, etc.) have a single prosodic peak.
This is known as the culminative property of stress. Many languages impose this
stressability requirement on content words (nouns, verbs, adjectives, adverbs)
only; function words (articles, pronouns, prepositions, etc.) need not be stressed,
and are prosodically dependent on content words.
Closely related to stressability, languages may require content words to have
some minimum size, for example two syllables, or a single heavy syllable. This
143
Metrical structure and parallelism

minimal word typically equals a single foot, a rhythmic unit consisting of two
syllables or two moras (length units). Languages may actively reinforce a binary
word minimum by prosodically expanding any word that would otherwise fall
below the minimum, by adding an augment of a mora or vowel to a subminimal
word. This is illustrated by examples from Levantine Arabic (lengthening, 2a) and
Iraqi Arabic (epenthesis, 2b), both from Broselow (1995):

(2) Epenthesis driven by the minimal word

a. /sal/ saal ask, masc. sing.
b. /drus/ idrus study

The binary foot size is related to the rhythmic property, which will be discussed
below.

The demarcative property. Stress tends to be placed near edges of constituents

(phrases, words, stems, etc.). This is the demarcative property, which has been
argued to facilitate the processing of grammatical units in perception. Cross-
linguistically favoured positions for primary word stress are the initial syllable
(3a: Pintupi, Australian; Hansen and Hansen 1969), the prenal (or penultimate)
syllable (3b: Warao, South American; Osborn 1966), and the nal syllable (3c:
Weri, New Guinea; Boxwell and Boxwell 1966).

(3) Three cross-linguistically favoured positions of primary stress

a. initial plikalatju we (sat) on the hill
b. prenal yapurukitanehse verily to climb
c. nal aknetepl times

These positions are ranked in decreasing order of popularity among the worlds
languages. The low ranking of pattern (3c) illustrates a dispreference against nal
stress, evidence for which will be adduced in section 4.3. (Secondary stresses are
left unmarked in 3.)
Some languages mark both edges of a word, while other languages mark edges
of morphemes contained within a word (root, stem, afx). Both strategies are
exemplied by Sibutu Sama, a language of the southern Philippines (Allison
1979). The word is marked on its prenal syllable, and the stem and every prex
on their initial syllables:

(4) Demarcative stress highlighting edges of morphemes

a. pDna-bDssalhan to be persuaded
b. mka-pgba-bDssalhan able to cause persuasion

This example illustrates how word stress may serve to highlight morphological
structure.
144
 4.2 Word stress: general background

The rhythmic property. Stress languages have a clear tendency towards rhythmic
patterns, with strong and weak syllables spaced apart at regular intervals. Rhyth-
mic alternation is manifested in a number of ways: by avoidance of adjacent
stressed syllables (clashes) or long strings of unstressed syllables (lapses).
Rhythmic alternation is directional, that is, oriented with respect to an edge of
the word, beginning or end. For example, Pintupi (5a) stresses the initial syllable
plus following alternate non-nal syllables, while Warao (5b) stresses the prenal
syllable and alternating preceding syllables.

(5) Binary alternation of stresses

a.i (tj.i).(r.u).(lm.pa).tju the re for our benet ared
up
a.ii (y.ma).(.ka).(m.ra).(tj.a).ka because of mother-in-law
b.i (y.pu).(r.ki).(t.ne).(h.se) verily to climb
b.ii e.(n.ho).(r.a).(h.ku).(t.i) the one who caused him to
eat

Bracketed two-syllable portions are metrical feet: the smallest units of linguistic
rhythm. All feet in (5) are trochees: here the initial syllable is strong and the
second weak. As we will see below, languages may also select iambs (whose
second syllable is strong).
The rhythmic spectrum is occupied at one end by languages that have a max-
imally dense rhythmic organization, like Pintupi and Warao. Yidi (Dixon 1977)
even enforces binary rhythm by apocopating a vowel of words with three or ve
syllables, to make it t the binary mould:

(6) Strict parsing of PrWd by feet in Yidi

a. /dur
uu-mu / (dur.
um) mopoke owl-abl
b. /
alambaraa-mu/ (
a.lm).(ba.ram) march y-abl

At the opposite end of the rhythmic spectrum we nd unbounded languages,

which have no alternating rhythm, allowing long strings of unstressed syllables
with a single stress per word. (Examples are Khalka Mongolian, Classical Arabic,
and Eastern Cheremis.) In this chapter we will focus on bounded, rhythmically
alternating, stress systems.

Quantity-sensitivity. Stress naturally prefers to fall on elements which have some

intrinsic prominence. For example, stress tends to be attracted by long vowels
(rather than by short vowels), by diphthongs (rather than monophthongs), and by
closed syllables (rather than open syllables). Attraction of stress by heavy syl-
lables is known as quantity-sensitivity. This phenomenon is exemplied by the
145
Metrical structure and parallelism

Australian language Nunggubuyu (Hore 1981; 7a) and the South American lan-
guage Hixkaryana (Derbyshire 1979; 7b):

(7) Heavy syllables attracting stresses

a. (2a).(2ag) almost
b. (nk).(h).(y
).(ke.n).no they were burning it

While heavy syllables attract stress, we also nd a reverse phenomenon of quan-

tity being adapted to prominence. Stressed vowels tend to lengthen, increasing
syllable weight, as in Hixkaryana (8a; see also section 4.3). Conversely, vowels
in unstressed syllables tend to shorten or reduce, decreasing syllable weight, as
in English tom~ atmic (8b):

(8) Stress contrasts supported by lengthening and reduction

a. (ne.m).(ko.t).no it fell
b. (E.tm) ~ .(tL.mk)

Mutually reinforcing relations of prominence and quantity are highly typical for
stress systems. (Although stress may be attracted by high-toned vowels, quant-
itative processes supporting tonal contrasts are rare in tone and pitch accent
languages.)

4.2.2 An alphabet of prosodic categories

In order to represent stress, metrical theory rst assumes a set of universal pro-
sodic categories in a hierarchical relation: the prosodic hierarchy (Selkirk 1980,
McCarthy and Prince 1986).

(9) Prosodic hierarchy

PrWd Prosodic Word

Ft Foot

Syllable

Mora

Every prosodic category in the hierarchy has as its head an element of the next-
lower level category. More precisely, every PrWd contains a (main-stressed) foot,
every foot contains a (stressed) syllable, while every syllable contains a mora, a
unit of quantity.

Second, to represent distinctions of vowel length and syllable weight, a theory of

quantity is required. Following Van der Hulst (1984), Hyman (1985), and Hayes
146
 4.2 Word stress: general background

(1989), we will assume moraic theory. According to this theory, a syllables quan-
tity is a function of its number of weight-bearing units, or moras. Universally,
short vowels are represented by one mora, while long vowels have two. The mora
is symbolized by the Greek letter :

(10) a. Light syllables (one mora) b. Heavy syllables (two moras)

t a t a n t a t a n
These diagrams represent weight, omitting the division into subsyllabic constitu-
ents onsetnucleuscoda.1 CV syllables are universally light (monomoraic), while
CVV syllables are universally heavy (bimoraic). The weight of a CVC syllable
depends on whether or not its coda consonant is moraic. This varies from lan-
guage to language, and is determined by the ranking of Weight-by-Position
(Hayes 1989, Sherer 1994):

(11) Weight-by-Position
Coda consonants are moraic.

For languages treating CVC as light, we assume that Weight-by-Position ranks

below a constraint militating against bimoraic syllables.2
Third, a theory of rhythmic units or feet assumes the following universal invent-
ory (McCarthy and Prince 1986, Hayes 1987, 1995, Kager 1993):

(12) a. Syllabic trochee (quantity-insensitive): ()

b. Moraic trochee (quantity-sensitive): (LL) (H)
c. Iamb (quantity-sensitive): (LL) (H) (LH)

In the notation of quantitative foot shapes, we will use L for a light syllable and
H for a heavy syllable. The head of the foot (stressed syllable) will be marked
by bold-face. It is a topic of debate whether the foot inventory should be enforced
by violable constraints, or hard-wired in the universal alphabet of prosodic
representations.3 In section 4.4, we will nd out to what extent this seemingly
1
Not taking a position on the issue of whether moraic theory should incorporate a subsyllabic
constituent rhyme (see for an overview Steriade 1988b), we will maintain these labels as
convenient shorthand notations. For reasons irrelevant here, we assume non-moraic consonants
(onsets, and coda consonants in light syllables) to adjoin below moras.
2
Weight-by-Position was originally conceived of as a parameter, since languages differ in
analysing CVC syllables as light or heavy.
3
For example, unheaded feet have been proposed by Crowhurst and Hewitt (1995), double-
headed feet by Bye (1996), and non-moraic syllables by Kager (1989) and Kirchner (1991).

147
Metrical structure and parallelism

arbitrary collection of feet is deducible from the interaction of constraints, such

as those governing quantity-sensitivity, foot size, etc.

4.3 Case-study: rhythmic lengthening in Hixkaryana

With this brief survey in mind, we will now delve into a genuine metrical system,
that of Hixkaryana, and try to spot metrical forces at work. Not surprisingly,
we will nd erce competition between forces, or, stated in OT terminology, con-
icting universal constraints. Hixkaryana (Derbyshire 1979) is a language of the
Cariban family, and is spoken in the Amazon (northern Brazil). It has no distinct-
ive vowel length all surface length reects rhythmic organization, or so-called
rhythmic lengthening. To see what is rhythmic about lengthening, consider the
following examples, in which foot structure has been marked.

(13) Words consisting of open syllables only

a. (to.r).no small bird
b. (a.
).wo.wo wind
c. (ne.m).(ko.t).no it fell

The generalization that governs the lengthening pattern is the following: in words
of three or more syllables, every even-numbered non-nal open syllable is
stressed, and its vowel is lengthened. This alternating length pattern can be char-
acterized as iambic: it involves a grouping of syllables into weakstrong rhyth-
mic units, of which the strong syllable is lengthened (Hayes 1995). In (13) this
rhythmic unit, the canonical iamb (LH), occurs once in words of three and four
syllables, and twice in ve-syllable words.
Surprisingly, in a disyllabic word consisting of two open syllables, the rst
syllable (rather than the second) is lengthened:

(14) Initial lengthening in disyllabic word

(kw).ja red and green macaw

Divergence from iambic rhythm is not a totally idiosyncratic fact, though, as it is

related to a more general observation, that nal syllables in Hixkaryana never undergo
vowel lengthening, and are never stressed. Whatever factor blocks nal stress and
lengthening may also push the rhythmic peak backward onto the rst syllable.

Turning now from open syllables to closed syllables, we nd that these are always
stressed, but that their vowels never lengthen. Even though closed syllables fail
to undergo lengthening themselves, they do affect the lengthening pattern of open
syllables, serving as reference points in counting syllables. For example, in words
that have a closed syllable in initial position, the lengthening pattern is similar to
that of (13), but with the difference that it is shifted one syllable to the right:
148
 4.3 Case-study: rhythmic lengthening in Hixkaryana

(15) Words beginning with a closed syllable

a. (k).(ma.t).r branch
b. (th).(ku.r j).ho.na to Tohkurye
c. (th).(ku.r j).(ho.n).(ha.).ha nally to Tohkurye
In such words length occurs on odd-numbered syllables (third, fth, seventh, etc.),
rather than on even-numbered ones (second, fourth, etc.). More generally, the
lengthening pattern in a sequence of open syllables starts over again at each
closed syllable:
(16) Words with multiple closed syllables
a. (nk).(h).(y
).(ke.n).no they were burning it
b. (kha.n).(nGh).no I taught you
c. (m.h).(na.nGh).no you taught him
In section 4.3.1 we will develop a rule-based analysis, after Hayes (1995). Then
we will turn to an OT analysis in section 4.3.2, and point out its descriptive and
conceptual advantages.

4.3.1 A rule-based analysis

A rule-based analysis of the Hixkaryana pattern would proceed in the following
ve steps (after Hayes 1995):
(17) Step 1: Syllabify (open syllables are light, closed syllables are
heavy).
Step 2: Mark the nal syllable of each word as extrametrical.
Step 3: Assign iambs {(LH), (LL), (H)} iteratively from left to
right.
Step 4: When the entire metrical domain is a single light syllable,
assign (L) to it.
Step 5: Lengthen the vowel of each strong open syllable.
In Step 2, extrametricality is the property of being invisible to rules of foot
construction. This analysis is illustrated by the derivations in (18):
(18) /a
owowo/ /kwa.ja/ /tohkur jehona/
Step 1: a.
o.wo.wo kwa.ja toh.ku.r je.ho.na
Step 2: a.
o.wo.<wo> kwa.<ja> toh.ku.r je.ho.<na>
Step 3: (a.
).wo.<wo> --- (th).(ku.r j).ho.<na>
Step 4: --- (kw).<ja> ---
Step 5: (a.
).wo.<wo> (kw).<ja> (th).(ku.r j).ho.<na>
Even though this analysis correctly captures the lengthening pattern, it has a
number of important conceptual drawbacks.
149
Metrical structure and parallelism

First, the rule that assigns iambs (Step 3) is not intrinsically connected to the
rule that lengthens open syllables in strong positions of feet (Step 5), although
both rules aspire toward a single output target: the canonical iamb (LH). This
generalization is missed.
Second, this analysis relies on an intermediate stage in the derivation of disyl-
labic words in which a degenerate foot (L) is temporarily allowed (Step 4). This
foot is repaired at the surface by lengthening its vowel into (H). This (L) foot is
abstract since the output pattern contains only (H) and (LH). Again, a general-
ization is missed.
Third, the analysis fails to explain why a degenerate foot is assigned in disyl-
labic words, whereas the prenal syllable of the longer words in (18) remains un-
parsed. That is, Step 4 is just an ad hoc incarnation of culminativity it only applies
in disyllabic words to satisfy the imperative that every word must have a foot.

4.3.2 An OT analysis
Instead we will nd it more fruitful to look at Hixkarayana length and stress
as a system of conicting preferences. We will nd out how these conicts are
resolved by comparing the actual outputs with potential outputs, which might
have arisen if the priorities between preferences had been different. That is, log-
ically possible but suboptimal outputs serve to support ranking arguments. Rank-
ing arguments will turn out to have a general form: for a pair of an actual form
and a suboptimal form, a constraint violated by the suboptimal form (but satised
in the actual form) dominates a constraint of which the violation pattern is the
reverse. (A more complete and formal treatment of what constitutes a ranking
argument in OT will be presented in chapter 7.)
In each step, we will proceed as follows. First an actual form will be juxtaposed
with a potential form, one which might have been, but is not, realized. We will
then identify two conicting preferences, relating these to the cross-linguistically
recurrent properties of stress systems of section 4.2. Preferences will then be
translated into universal metrical constraints, which will be ranked one by one on
the basis of the available empirical evidence. Finally, we will integrate the sub-
ranking thus found into the (current) total ranking for the metrical pattern of
Hixkaryana.
The metrical constraints to be introduced in the analysis of Hixkaryana will be
discussed in more detail in section 4.4. (There proper credit will be given to their
inventors, and all constraints will be motivated in more detail, paying attention to
their function and format.)

What conicting preferences hide behind the length pattern of Hixkaryana? A

rst conict that we identify is that between a preference for the canonical iamb
150
 4.3 Case-study: rhythmic lengthening in Hixkaryana

(LH), versus a preference for nal syllables to be unstressed. A disyllabic word

of two open syllables can (potentially, at least) be metried as a canonical iamb,
by lengthening the vowel of its second syllable. However, this potential form is
not what we actually nd. Vowel length is absent from the nal syllable, and
instead length occurs initially. We can write this as a statement of relative met-
rical harmony in (19), where the symbol
is to be read as is a better
metrication than:

(19) (kw).ja
(kwa.j)

It is apparently more important to leave the nal syllable unfooted than it is for
every foot to be a canonical (LH) iamb. Both preferences can be translated
directly into metrical constraints. The rst states that nal syllables must not be
metried:

(20) NonFinality
No foot is nal in PrWd.

The second constraint denes the best rhythmic shape of the iamb, a foot
having weakstrong prominence and a quantitative make-up light-heavy (Hayes
1995). Uneven-Iamb says that (LH) is a better iamb than (LL) or (H):4

(21) Uneven-Iamb
(LH)
(LL), (H)

The harmony statement in (19) translates into a ranking argument for these
constraints:

(22) NonFinality
Uneven-Iamb

Tableau (23) supports this ranking. Each candidate violates one constraint while
satisfying the other. The optimal candidate is (23a) since it satises the highest-
ranking constraint of the two, NonFinality:

(23) Input: /kwaja/ NonFinality Uneven-Iamb

a. + (kw).ja *
b. (kwa.j) *!

We will now extend this simple ranking (using the same basic method) until we
arrive at a complete characterization of the Hixkaryana lengthening pattern.

4
In section 4.4.5 we will decompose requirements on prominence (iambs vs. trochees) and
quantitative structure (LL vs. LH, H vs. LH, etc.) that are combined in Uneven-Iamb.

151
Metrical structure and parallelism

Let us consider another potential way out of the dilemma of conicting con-
straints. It seems that both the Scylla of NonFinality and the Charybdis of
Uneven-Iamb can be avoided by an output that has no feet at all. This unparsed
output vacuously satises both Uneven-Iamb and NonFinality, simply because
it has no foot to be evaluated. Then why would Hixkaryana still prefer the input
/kwaja/ to surface with a foot, even when this foot fails to match the best form?

(24) (kw).ja
kwa.ja

The unparsed output violates the imperative that grammatical words must have
prosody the culminative property:

(25) GrWd=PrWd
A grammatical word must be a prosodic word.

From this constraint it indirectly follows that grammatical words must have min-
imally one foot. Recall from section 4.2.2 that every prosodic category (PrWd,
foot, and syllable) takes as its head a category of the next-lower category. There-
fore, every PrWd must have a foot as its head. Requiring PrWd-status of a
grammatical word thus amounts to requiring that it must have at least a foot.
The preference for grammatical words to be PrWds (hence to have feet) is
stronger than the dispreference against bad iambs:

(26) GrWd=PrWd
Uneven-Iamb

A tableau illustrates this ranking:

(27) Input: /kwaja/ GrWd=PrWd Uneven-Iamb

a. + (kw).ja *
b. kwaj *!

Culminativity of stress outranks foot well-formedness. This is precisely the gen-

eralization left uncaptured by the rule-based analysis.

A second piece of evidence showing that nal syllables are strongly preferred to
remain unfooted is the length pattern of words with four open syllables. Here the
second syllable is lengthened, while the fourth syllable remains short. Apparently
the left hand parse, with a single canonical (LH) iamb, is preferred to the right
hand parse, which has two:

(28) (a.
).wo.wo
(a.
).(wo.w)

152
 4.3 Case-study: rhythmic lengthening in Hixkaryana

This shows that it is more important to leave the nal syllable unfooted than it is
for all syllables to be parsed by feet. Or, translating the latter preference into a
constraint:
(29) Parse-Syl
Feet are parsed by feet.
The resulting ranking of these two constraints is:
(30) NonFinality
Parse-Syl

(31) Input: /a
owowo/ NonFinality Parse-Syl

a. + (a.
).wo.wo **
b. (a.
).(wo.w) *!

This ranking is conrmed by disyllabic words see the harmonic comparison in

(19).

We have found that NonFinality dominates both Uneven-Iamb and Parse-Syl,

but is it also possible to establish a ranking between both dominated constraints?
We start by asking under what conditions Uneven-Iamb and Parse-Syl could
possibly conict. Relevant contexts are those in which a more complete metricat-
ion would produce non-canonical feet.
Three-syllable forms, for example, cannot be exhaustively parsed without vio-
lating Uneven-Iamb. We therefore compare the pair of forms (32). Does this
harmonic ranking tell us that Uneven-Iamb dominates Parse-Syl?
(32) (to.r).no
(to.r).(n)
Unfortunately, this pair tells us nothing about the ranking we are interested in, as
there is an independent explanation for the suboptimality of the righthand form:
it fatally violates NonFinality. All we can infer from (32) is a conrmation of
the ranking NonFinality
Parse-Syl. Still, we have learnt something from this
failure: we must compare forms that do not involve violations of NonFinality.
We start looking for other evidence bearing on the ranking of Uneven-Iamb
and Parse-Syl. Such evidence may come from a comparison of four-syllable
forms.
(33) (a.
).wo.wo
(a.
).(w).wo

().(
o.w).wo
A single-foot parse is preferred to a two-foot parse even if this respects NonFi-
nality. As indicated, a two-foot parse might have either a (LH) iamb plus a (H)
heavy foot (with the second and third syllable long), or (H) plus (LH) (with the
153
Metrical structure and parallelism

rst and third syllable long). Both two-foot parses are rejected, though, in favour
of a single-foot parse. This shows that it is more important for every foot to be a
canonical (LH) iamb, than it is for all syllables to be parsed by feet. Or, in terms
of constraint ranking:
(34) Uneven-Iamb
Parse-Syl
This ranking is supported by a tableau:

(35) Input: /a
owowo/ Uneven-Iamb Parse-Syl

a. + (a.
).wo.wo **
b. (a.
).(w).wo *! *
c. ().(
o.w).wo *! *

This is an important result showing that ranking arguments are consistent with
transitivity. To see this, we must recapitulate some results. Firstly, NonFinality
dominates Uneven-Iamb (see tableau 23). Secondly, Uneven-Iamb dominates
Parse-Syl (see tableau 35). Thirdly, NonFinality dominates Parse-Syl (tab-
leau 31), conrming transitivity of strict domination. We now arrive at the fol-
lowing hierarchy:
(36) GrWd=PrWd, NonFinality
Uneven-Iamb
Parse-Syl
Tableau (37) summarizes two results booked so far. It proves Uneven-Iamb

Parse-Syl, as well as NonFinality
Parse-Syl (as already established by
tableau 31):

(37) Input: /a
owowo/ NonFinality Uneven- Parse-Syl

Iamb

a. + (a.
).wo.wo **
b. (a.
).(w).wo *! *
c. (a.
).(wo.w) *!

The optimal candidate (37a) violates Parse-Syl more severely than any of its
competitors (37b c). However, these competitors fatally violate either NonFi-
nality (37c) or Uneven-Iamb (37b), both of which dominate Parse-Syl.

What other preferences govern the length pattern of Hixkaryana? For example, is
there any evidence for quantity-sensitivity apart from a preference for the canon-
ical iamb? Such evidence indeed exists, and it comes from two sources.
154
 4.3 Case-study: rhythmic lengthening in Hixkaryana

The rst nding is that it is more important to stress heavy syllables than for
every foot to be a (LH) iamb. Evidence comes from the pair in (38). The actual
form stresses a heavy syllable, but has a non-canonical foot (H). The suboptimal
form has one canonical foot, but fails to stress the heavy syllable.

(38) (th).(ku.r j).ho.na
toh.(ku.r j).ho.na

We conclude that foot-form has less priority than quantity-sensitivity.

The suboptimal status of the righthand form in (38) is due to a constraint
enforcing quantity-sensitivity, the close relation between syllable weight and
prominence. This is the Weight-to-Stress-Principle (henceforth WSP):

(39) WSP
Heavy syllables are stressed.

This constraint is violated by any heavy syllable that is not prominent, either
within a foot or outside a foot. The harmonic relation in (38) translates into a
ranking argument:

(40) WSP
Uneven-Iamb

A related piece of evidence for quantity-sensitivity comes from a comparison

made in (41). Both the actual (lefthand) form and the potential (righthand)
form contain one foot in initial position that is not a canonical iamb, (H) and
(HH), respectively. Although the righthand form might seem preferable for its
more complete parsing, it is suboptimal. The origin of its failure is its quantity-
disrespecting initial foot:

(41) (th).(ku.r j).ho.na
(toh.k).(r je.h).na

This evidence tells us that it is more important to stress heavy syllables than it is
for all syllables to be parsed by feet. We now infer the ranking

(42) WSP
Parse-Syl

WSP is actually undominated in Hixkaryana, for both kinds of heavy syllables,

CVC and CVV, are stressed without exception. We thus rank WSP right at the top
of the hierarchy, together with GrWd=PrWd and NonFinality:5

(43) GrWd=PrWd, NonFinality, WSP
Uneven-Iamb
Parse-Syl

5
NonFinality and WSP potentially conict in a PrWd ending in a closed syllable. If the closed
syllable is stressed, then WSP
NonFinality; if it is unstressed, then NonFinality
WSP.
Unfortunately this ranking cannot be established since Hixkaryana lacks nal closed syllables.
Speculatively, the lack of nal CVC might be construed as support for the undominated status
of NonFinality and WSP: violation of both constraints is avoided at the expense of nal
consonant deletion, violating Max-C. However, no alternations occur.

155
Metrical structure and parallelism

Tableau (44) lays out the most complex interaction of metrical constraints that
we have seen thus far, involving the full hierarchy in (43) except GrWd=PrWd.
(44)
Input: /tohkur jehona/ NonFinality WSP Uneven- Parse-Syl
Iamb

a. + (th).(ku.r j).ho.na * **
b. (th).(ku.r j).(h).na **! *
c. (th).(ku.r j).(ho.n) *! *
d. (toh.k).(r je.h).na *! * *
j
e. toh.(ku.r ).ho.na *! ***

The tableau contains the quantity-disrespecting candidates (44de), presented

earlier in (38) and (41), which fatally violate WSP. Candidate (44b), in comparison
to optimal (44a), has an additional parsed syllable, causing an additional (and
fatal) violation of Uneven-Iamb. Such an additional violation of Uneven-Iamb
is avoided by candidate (44c), but at a high price: a fatal violation of Non-
Finality. This once again corroborates that NonFinality
Parse-Syl, as we
had already established in tableau (31).

Continuing our search for priorities, we nd that it is more important to avoid

monomoraic (L) feet, than to avoid vowel length in underlyingly short vowels
(where the input has none). The evidence consists of the comparison of both
forms in (45), where a disyllabic form that has a (H) foot with a lengthened vowel
is preferred to a potential form involving no lengthening, but a monomoraic (L)
foot.

(45) (kw).ja
(kw).ja

This demonstrates that Dep--IO (the anti-lengthening constraint) is relatively

weak:

(46) Dep--IO
Output moras have input correspondents.

What the harmonic comparison (45) directly tells us is that Dep--IO is dominated
by Ft-Bin, which we will learn to know as the anti-degenerate-foot constraint:

(47) Ft-Bin
Feet are binary under moraic or syllabic analysis.
156
 4.3 Case-study: rhythmic lengthening in Hixkaryana

Ft-Bin is undominated in Hixkaryana: no forms contain light stressed syllables,

and hence no forms contain (L) feet. Because of this, the ranking Ft-Bin
Dep--
IO is not particularly informative about the precise position of Dep--IO in the
constraint hierarchy.
Fortunately, a second source of evidence bearing on this ranking can be
identied. In fact, every word that has a lengthened vowel in a (LH) iamb tells
us that the avoidance of vowel length (Dep--IO) is subordinated to the prefer-
ence for every foot to be a canonical (LH) iamb (Uneven-Iamb). So much is
shown, for example, by a comparison of an actual trisyllabic form in (48) to a
potential form that omits vowel length:

(48) (to.r).no
(to.r).no

This means that we can now add Ft-Bin and Dep--IO to the hierarchy, and
locate the latter as accurately as possible in a position directly dominated by
Uneven-Iamb:

(49) GrWd=PrWd, NonFinality, Ft-Bin, WSP
Uneven-Iamb

Parse-Syl, Dep--IO

What remain to be claried are the factors that determine the position of feet
with respect to the edges of the PrWd. We have already seen that words of four
open syllables fall into a metrication that has only one canonical iamb (see 33).
Still, this single foot can be placed in two different positions, even respecting
NonFinality in both cases. Comparison of the actual four-syllable form with a
potential form shows that the foot is preferrably initial in the word. (This is
Hixkaryanas form of demarcative stress):

(50) (a.
).wo.wo
a.(
o.w).wo

This preference for feet to be initial in the PrWd is translated into a maximally
general constraint:

(51) All-Ft-Left
Every foot stands at the left edge of the PrWd.

Violations of this constraint will amount to the number of syllables between a

foot and the left edge of the word. When a word contains multiple feet, the
violations for all feet will be added up. This is an example of a gradient
constraint.
How is All-Ft-Left ranked with respect to other constraints? To nd out we
must identify forms in which it is actually violated. There happen to be plenty of
forms in which All-Ft-Left is violated: every PrWd having more than one foot
157
Metrical structure and parallelism

constitutes a violation, as only one foot per PrWd can be strictly initial. We thus
nd that it is more important for all syllables to be parsed by feet than it is for
every foot to stand at the left edge:
(52) (ne.m).(ko.t).no
(ne.m).ko.to.no
Accordingly, we nd the important ranking:
(53) Parse-Syl
All-Ft-Left
The resulting integrated hierarchy of all constraints for Hixkaryana is in (54):6
(54) GrWd=PrWd, NonFinality, Ft-Bin, WSP


Uneven-Iamb


Parse-Syl


All-Ft-Left, Dep--IO
Note that we now predict that WSP must dominate All-Ft-Left, by transitiv-
ity. This prediction is conrmed by the comparison in (55):
(55) (kha.n).(nGh).no
(kha.n).nph.no
The following display summarizes all rankings that we have found, and for
each ranking, the evidence on which it is based.
(56)
Its more important: than it is:
to leave the nal syllable unfooted for every foot to be a (LH) iamb
NonFinality
Uneven-Iamb
(kw).ja
(kwa.j)
for every word to contain a foot for every foot to be a (LH) iamb
GrWd=PrWd
Uneven-Iamb
(kw).ja
kwa.ja
to avoid (L) feet to avoid vowel length
Ft-Bin
Dep--IO
(kw).ja
(kw).ja
for every foot to be a (LH) iamb to avoid vowel length
Uneven-Iamb
Dep--IO
(to.r).no
(to.r).no
6
This analysis has been inuenced by analyses of iambic patterns in Axininca Campa (McCa-
rthy and Prince 1993b, Hung 1994) and Carib (Van der Hulst and Visch 1992, Kenstowicz
1995).

158
 4.3 Case-study: rhythmic lengthening in Hixkaryana

to stress heavy syllables for every foot to be a (LH) iamb

WSP
Uneven-Iamb
(th).(ku.r j).ho.na
toh.(ku.r j).ho.na
to stress heavy syllables to parse all syllables by feet
WSP
Parse-Syl
(th).(ku.r j).ho.na
(toh.k).(r je.h).na
for every foot to be a (LH) iamb to parse all syllables by feet
Uneven-Iamb
Parse-Syl
(a.
).wo.wo
(a.
).(w).wo
to leave the nal syllable unfooted to parse all syllables by feet
NonFinality
Parse-Syl
(a.
).wo.wo
(a.
).(wo.w)
to parse all syllables by feet for every foot to be at the left edge
Parse-Syl
All-Ft-Left
(ne.m).(ko.t).no
(ne.m).ko.to.no

To wind up the analysis of Hixkaryana, we present extensive tableaux of four key

forms (57 60).

(57)
Input: /kwaja/ Non- GrWd= Ft- WSP Un- Parse- All- Dep
Finality PrWd Bin even- Syl Ft- -
Iamb Left -IO

a. + (kw).ja * * *
b. (kw).ja *! * *
w
c. k a.ja *! **
d. (kwa.j) *! *

(58)
Input: /torono/ Non- GrWd= Ft- WSP Un- Parse- All- DEP
Finality PrWd Bin even- Syl Ft- -
Iamb Left -IO

a. + (to.r).no * *

b. (to.r).no *! *

c. (to.r).(n) *! * ** **

159
Metrical structure and parallelism

(59)
Input: /a
owowo/ Non- GrWd= Ft- WSP Un- Parse- All- Dep
Finality PrWd Bin even- Syl Ft- -
Iamb Left -IO

a. + (a.
).wo.wo ** *

b. a.(
o.w).wo ** *! *

c. (a.
).(w).wo *! * ** **

d. (a.
).wo.wo *! **

e. (a.
).(wo.w) *! ** **

(60)
Input: /tohkur jehona / Non- GrWd= Ft- WSP Un- Parse- All- Dep
Finality PrWd Bin even- Syl Ft- -
Iamb Left -IO

a. + (th).(ku.r j).ho.na * ** * *
j
b. (th).ku.(r e.h).na * ** **! *

c. (toh.k).(r je.h).na *! * * ** **

d. toh.(ku.r j).ho.na *! *** * *

j
e. (th).(ku.r ).(ho.n) *! * *, *** **

4.3.3 Conclusions
This analysis of Hixkaryana illustrates important ingredients of the analysis of
word stress patterns: basic metrical constraints and their interactions in producing
a rhythmic pattern. Note in particular the following four rankings:
(61) Binary rhythm starting at the left word edge
Ft-Bin
Parse-Syl
All-Ft-Left
(62) Metrical inertness of the nal syllable
NonFinality
Parse-Syl
(63) Subminimal lengthening
GrWd=PrWd, Ft-Bin
Dep--IO
(64) Iambic lengthening
Uneven-Iamb
Dep--IO
These ranking schemata will be placed in broader typological perspective in sec-
tion 4.45.
160
 4.4 A set of metrical constraints

Returning to a basic theme of this book, the interaction of well-formedness and

faithfulness, we have seen another example of this in the analysis of Hixkaryana:
rhythmic vowel lengthening (at the expense of Dep--IO, the anti-lengthening
constraint) is forced by foot-form constraints (Ft-Bin and Uneven-Iamb). This
interaction is interesting since it is parallel, in the following sense: evaluation
of foot form takes place in a single hierarchy together with the evaluation of input
faithfulness. There is a major difference between the OT analysis and the deriva-
tional analysis in this respect. In the derivational analysis vowel lengthening is
a way to repair ill-formed feet created by previous metrication rules. This
division of labour is due to the strict distinction in rule-based theory between
rules of metrical parsing and rules of quantitative adjustment. The former con-
struct feet, while the latter carry out the quantitative changes required to render
the feet perfect. In OT no such serial interaction of metrical parsing and quan-
titative changes is possible, as intermediate levels of derivation are lacking. All
aspects of foot well-formedness (their distribution and quantitative make-up) are
evaluated at one level: the output. Consequently the OT analysis integrates
aspects of foot well-formedness in a more generalizing way.

4.4 A set of metrical constraints

In this section we will present a set of metrical constraints. We will also place the
metrical constraints that were introduced in the analysis of Hixkaryana in a
broader cross-linguistic perspective, by relating them to our discussion of com-
mon preferences in stress systems in section 4.2. In the course of the exposition,
additional constraints will be introduced, which are (as the earlier-mentioned
constraints) identical in name and effect to constraints in standard OT literature
(particularly: Prince and Smolensky 1993: 2832, 3866, McCarthy and Prince
1993a, b: 150 6), unless specic mention is made of a deviation.

4.4.1 Rhythm: FT-BIN, PARSE-SYL, and ALL-FT-X

As a powerful cross-linguistic preference in stress languages, we have identied
rhythm, or regular alternation of strong and weak syllables. In metrical phonology,
such alternating patterns are interpreted as the grouping of syllables into rhythmic
units or feet.
The fundamentally rhythmic requirement that feet be binary is expressed in
Ft-Bin (Prince 1980, Kager 1989, Prince and Smolensky 1993):

(65) Ft-Bin
Feet are binary under moraic or syllabic analysis.

A foot must contain either two moras, as in (H) or (LL), or two syllables, as in
(). A key function of Ft-Bin is to exclude degenerate feet (L), which contain
161
Metrical structure and parallelism

a single light syllable. Many languages have an absolute ban on degenerate feet,
due to undominated Ft-Bin. The usual diagnostic of undominated Ft-Bin is the
absolute enforcement of the minimal word.
Enforcing rhythmic binarity, Ft-Bin also plays a key role in rhythmic alterna-
tion. Note that Ft-Bin by itself does not sufce to guarantee the binary alterna-
tion of weak and strong syllables. This requires something more all syllables
must be parsed by feet (Hayes 1980, Halle and Vergnaud 1987, Prince and Smo-
lensky 1993):
(66) Parse-Syl
Syllables are parsed by feet.
Syllables violating this constraint (any which are not parsed by a foot) will be
assumed to be metried as immediate daughters of PrWd. This assumption is
known as weak layering (It and Mester 1992). The Hixkaryana example
[(to.r).no], represented hierarchically, has a weakly-layered nal syllable, one
that is unparsed by a foot:
(67) PrWd

Ft

to.ro.no
Only the nal syllable, the one that is unparsed by a foot, violates Parse-Syl.
In combination, high-ranking Ft-Bin and Parse-Syl have the effect of enforc-
ing binary alternation: a parse of the word into multiple binary feet. When a word
can be fully parsed by binary feet without any violation of Ft-Bin or Parse-
Syl, the foot distribution is uniquely determined by both constraints. Then an
exhaustive parsing arises, with all feet packed tight together. This is the case, for
example, in words with an even number of light syllables; see (68a). But in words
with an odd number of light syllables, one syllable must remain that cannot be
parsed by binary feet. Accordingly, there are multiple ways in which a maximally
dense distribution of feet can be achieved; see (68b).
(68) a. ()()() Exhaustive parsing: feet packed tightly
b.i ()()()
b.ii ()()()
Non-exhaustive parsing: multiple options
b.iii ()()()
b.iv ()()()
Here the constraint pair Ft-Bin or Parse-Syl fails to uniquely determine foot
distribution. To produce a genuine rhythmic distribution of feet, one oriented
162
 4.4 A set of metrical constraints

toward a specic edge of the word, a foot alignment constraint must pick out one
of the parsings in (68b.iiv). Such a constraint is stated below after McCarthy and
Prince (1993a):

(69) All-Ft-Left
Align (Ft, Left, PrWd, Left)
Every foot stands at the left edge of the PrWd.

This is an alignment constraint of the general format Align (Cat1, Edge1, Cat2,
Edge2) (see chapter 3), requiring that the left edge of every foot coincide with
the left edge of a PrWd. This is satised only in the case of a single foot standing
at the absolute left edge of the word: any additional foot will incur a violation,
since two feet cannot both stand at the left edge. If undominated, All-Ft-Left
enforces a single foot per PrWd a non-iterative pattern.

But then how can All-Ft-Left resolve ties among candidates containing multi-
ple feet, such as those in (68)? The answer is simple and straightforward. Each
foot that is not in absolute initial position incurs some violation of All-Ft-Left.
If we count the distance between a foot and the left word edge in syllables, then
each foot will incur a violation of All-Ft-Left that equals this number. In words
containing multiple feet, we simply add up violations for individual feet to arrive
at the total violation of All-Ft-Left. The candidate incurring the smallest total
violation wins. This is an example of a gradient constraint.

A left-oriented distribution of feet is achieved under the ranking All-Ft-Left

All-Ft-Right (its mirror image). In tableau (70), commas separate the viola-
tions per foot, counting foot by foot from the relevant word edge. Exclamation
marks are placed at the point at which the number of violations becomes fatal:

(70) Left-oriented rhythm All-Ft-Left All-Ft-Right

a. + ()()() **, **** *, ***, *****

b. ()()() **, *****! ***, *****
c. ()()() ***, ****!* **, *****
d. ()()() *, ***, ***!** **, ****

Reversing the ranking of both constraints will have the effect of a right-oriented
rhythm:
163
Metrical structure and parallelism

(71) Right-oriented rhythm All-Ft-Right All-Ft-Left

a. ()()() *, ***, **!*** **, ****

b. ()()() ***, ***!** **, *****
c. ()()() **, ****!* ***, *****
d. + ()()() **, **** *, ***, *****

The highest-ranked foot alignment constraint determines the directionality of foot

parsing. Non-directional candidates such as (70b, c) and (71b, c) can never be
optimal, regardless of the ranking of foot alignment constraints.7 This captures the
fact that rhythmic patterns are directional, as a wave starting at one edge and
owing towards the other edge.

Both rhythmic types are attested in a number of languages. Left-oriented rhythm

is exemplied by the Australian language Pintupi (Hansen and Hansen 1969,
Hayes 1995), which stresses the initial syllable and following odd-numbered syl-
lables, except the nal. Right-oriented rhythm is exemplied by the Amazonian
language Warao (Osborn 1966, Hayes 1980), which stresses the penult and pre-
ceding alternating syllables.

(72) a. Pintupi: Ft-Bin
Parse-Syl
All-Ft-Left

(tj.a).ja many
(m.a).(w.na) through (from) behind
(p.i).(k.la).tju we (sat) on the hill
j j
(t .mu).(lm.pa).(t .ku) our relation
(tj.i).(r.u).(lm.pa).tju the re for our benet ared up
b. Warao: Ft-Bin
Parse-Syl
All-Ft-Right
ko.(r.nu) drink it!
yi.(w.ra).(n.e) he nished it
(y.pu).(r.ki).(t.ne).(h.se) verily to climb
e.(n.ho).(r.a).(h.ku).(t.i) the one who causes him to eat

In both languages, degenerate feet are avoided at the expense of violations of

Parse-Syl. This signals an important interaction between Ft-Bin and Parse-Syl
which will return in various forms in the remainder of this chapter:

7
In section 4.4.3 we will see that such patterns arise due to interactions with word alignment
constraints.

164
 4.4 A set of metrical constraints

(73) Maximally binary metrication

Ft-Bin
Parse-Syl
All-Ft-X
A tableau of the Pintupi example [(p.i).(k.la).tju] serves to illustrate the activ-
ity of all three constraints in this ranking:

(74) Input: /puikalatju/ Ft-Bin Parse-Syl All-Ft-Left

a. + (p.i).(k.la).tju * **
b. (p.i).ka.(l.tju) * ***!
c. pu.(.ka).(l.tju) * *, **!*
d. (p.i).ka.la.tju **!*
e. (p.i).(k.la).(tj) * **, ****

Undominated Ft-Bin rules out degenerate feet (as in 74e), at the expense of
violations of Parse-Syl (74ad). But since Parse-Syl is minimally violated, any
metrications having multiple unparsed syllables (74d) are ruled out. All remain-
ing metrications have multiple feet (74ac), at the expense of multiple violations
of All-Ft-Left. Minimal violation of All-Ft-Left results in an optimal output
whose collective feet stand as close as possible to the left word edge this is
(74a).
Related to rhythmic distribution of stresses, two more preferences have been
observed to be strong in many languages. First, adjacent stressed syllables
(clashes) are often avoided:
(75) *Clash
No stressed syllables are adjacent.
*Clash has its roots in pre-OT work (Liberman 1975, Liberman and Prince
1977, Prince 1983, Hammond 1984, Selkirk 1984). See section 4.4.5 for further
discussion.
Second, nal syllables are metrically inert, in a double sense. For Hixkaryana,
we have seen that nal syllables must be unfooted the PrWd must not end in a
foot. But for many other languages, a somewhat weaker requirement holds that
the nal syllable must be unstressed the PrWd must not end in a stressed
syllable. Both requirements are put in a single statement in Prince and Smolen-
skys version of NonFinality:
(76) NonFinality
No prosodic head is nal in PrWd.

165
Metrical structure and parallelism

This is the OT counterpart of extrametricality in rule-based theory. It is not

completely clear whether both interpretations of NonFinality (unparsed/
unstressed nal syllable) should be merged in a single constraint.8

4.4.2 Culminativity: GRWD=PRWD and word minima

As was explained in section 4.3, the culminative property is enforced by a con-
straint requiring that every grammatical word must be a prosodic word as well:

(77) GrWd=PrWd
A grammatical word must be a PrWd.

The prosodic hierarchy guarantees that every PrWd dominates at least one foot
hence the combined result of GrWd=PrWd and the prosodic hierarchy is that
every grammatical word has at least one stressed syllable the culminative prop-
erty. (Most languages make a distinction in prominence among multiple feet in a
word. The issue of how to capture this distinction by constraints, and how to
locate the strongest foot primary stress in a word, will be dealt with in section
4.4.3.)
Related to GrWd=PrWd is the notion of word minimum the requirement that
a word have minimally two moras or syllables. A word minimum can be enforced
in various ways, for example by subminimal lengthening, making a monomoraic
word bimoraic, or by the epenthesis of a vowel into a monosyllabic word, making
it disyllabic. The example below is taken from Mohawk (Piggott 1995):

(78) /k-tat-s/ k.tats I offer it

This effect is due to GrWd=PrWd in combination with Ft-Bin. The chain of

reasoning is as follows. By GrWd=PrWd, a grammatical word must equal a
prosodic word. By the prosodic hierarchy, a PrWd must dominate a foot. Finally,
by Ft-Bin, a foot must have two moras or syllables. Epenthesis (or augmenta-
tion of the subminimal form) is enforced when the anti-epenthesis constraint
Dep-IO is dominated by GrWd=PrWd and Ft-Bin:

(79) Word minimum enforced by epenthesis

GrWd=PrWd, Ft-Bin
Dep-IO

This interaction is illustrated for the Mohawk example below:

8
*Clash and NonFinality were collapsed by Hung (1994) into a single constraint Rhythm
every strong syllable must be followed by a weak syllable. See Elenbaas (1996, 1999) for
empirical arguments against this merging.

166
 4.4 A set of metrical constraints

(80) Input: /k-tat-s/ GrWd=PrWd Ft-Bin Dep-IO

a. + (k.tats) *
b. (ktts) *!
c. ktats *!

The next section deals with metrical sensitivity to word edges.

4.4.3 Demarcative stress: ALIGN-HD and ALIGN-WD

Another strong cross-linguistic preference (mentioned earlier in section 4.2) is for
stresses to fall at edges of the domain: the so-called demarcative property of
stress. In OT, edge-marking is due to alignment constraints requiring cooccur-
rence of edges of categories, for example of the prosodic categories PrWd and Ft.
We have already seen two constraints to that effect: All-Ft-Left and All-Ft-
Right. Here we will extend the family of metrical alignment constraints with four
members, coming in two pairs.

The rst pair of metrical alignment constraints, Leftmost and Rightmost, align
the strongest foot (or head foot) with a specied edge of the word. (These
constraints are named after Edgemost from Prince and Smolensky 1993):

(81) a. Leftmost
Align (Hd-Ft, Left, PrWd, Left)
The head foot is leftmost in PrWd.
b. Rightmost
Align (Hd-Ft, Right, PrWd, Right)
The head foot is rightmost in PrWd.

The function of peak-aligning constraints is dual. First, they serve to locate the
primary stress on a foot standing at the specied edge of the word. Thus far we
have had no constraints making prominence distinctions among feet; since most
languages make such distinctions, (81a, b) are well-supported.
The second function of Leftmost/Rightmost is perhaps less obvious it
appears in the analysis of so-called bidirectional stress systems. Such systems are
characterized by a rhythmic pattern that is oriented toward two word edges, rather
than toward a single edge. For example, the Australian language Garawa (Furby
1974, Hayes 1980, McCarthy and Prince 1993a) stresses the initial syllable and
alternating syllables preceding the penult. Feet are oriented toward the right edge
(as is evident from words of seven or more syllables), but the primary stress is
always initial. See (82a).
167
Metrical structure and parallelism

(82) a. Garawa: Ft-Bin, Leftmost
Parse-Syl
All-Ft-Right

(pn.ja).a white
(k.ma).a.(.i) wrist
(y.ka).(l.ka).(lm.pa) loose
(n.ki).i.(k.rim).(p.yi) fought with boomerangs
b. Piro: Ft-Bin, Rightmost
Parse-Syl
All-Ft-Left
ru.(:.:a) he observes taboo
(s.lwa).ye.(hk.kna) they visit each other
(p.i).(h.ma).(tl.na) they say they stalk it
(r.slu).(n.ti).ni.(tk.na) their voices already changed
The mirror-image system of Garawa is exemplied by Piro (Matteson 1965,
Hayes 1995), which stresses the penult and alternate syllables following the initial
syllable. See (82b).
The analysis of bidirectional stress patterns involves domination of peak align-
ment over foot alignment at the opposite edge. Thus, Leftmost dominates All-
Ft-Right in the analysis of Garawa, while Rightmost dominates All-Ft-Left
in the analysis of Piro. The basis of the rankings is the normal interaction pro-
ducing binary alternation in the analyses of Pintupi/Warao, that is, Ft-Bin

Parse-Syl
All-Ft-X. This ranking is now extended by inclusion of undomin-
ated peak alignment constraints, as in the rankings given in (82). The tableau of
a seven-syllable example of Garawa is presented below:
(83)
Input: /ankiikirimpayi/ Ft- Leftmost Parse- All-Ft-
Bin Syl Right

a. + (n.ki).i.(k.rim).(p.yi) * **, *****

b. (n.ki).(.ki).rim.(p.yi) * ***, *****!
c. (n.ki).(.ki).(rm.pa).yi * *, ***, ****!*
d. (n.ki).i.ki.rim.(p.yi) **!* *****
e. (n.ki).i.ki.rim.pa.yi **!*** *****
f. an.(k.i).(k.rim).(p.yi) *! * **, ****
g. (n.ki).(.ki).(rm.pa).(y) *! *, ***, *****

By undominated Leftmost, every word has an initial main stress foot (excluding
83f). By Ft-Bin, all feet are binary (excluding 83g). Among the remaining can-
didates, Parse-Syl picks out candidates that have a single unparsed syllable
(excluding 83d e). The survivors (83ac) differ only in the distribution of their
168
 4.4 A set of metrical constraints

three feet (of which the initial one is xed by Leftmost). Minimal violation of
All-Ft-Right draws all feet as close as possible to the right edge, selecting
(83a). In sum, main stress is placed at the left edge due to Leftmost, while all
other stresses are distributed directionally due to All-Ft-Right.
The second pair of demarcative constraints is stated in (84):
(84) a. Align-Wd-Left
Align (PrWd, Left, Ft, Left)
Every PrWd begins with a foot.
b. Align-Wd-Right
Align (PrWd, Right, Ft, Right)
Every PrWd ends in a foot.
These constraints are functional antagonists of All-Ft-Left and All-Ft-Right.
Instead of making a requirement about feet (in terms of word edges), they make
a requirement about word edges (in terms of feet). Accordingly, they are violated
when no foot is present at the specied edge of the word. (Recall the discussion
of asymmetric alignment in chapter 3, section 3.4.1.)
The analytic function of these constraints resides in complex bidirectional sys-
tems, of which an example, Indonesian (Cohn 1989), is given below:
(85) Indonesian: Ft-Bin, Rightmost
Align-Wd-L
Parse-Syl

All-Ft-R
a. bi.(c.ra) speak
b. (kn.ti).nu.(.si) continuation
c. (.ro).(d.na).(m.ka) aerodynamics
d. (.me).ri.(k.ni).(s.si) Americanization
In Indonesian primary stress falls on the penultimate syllable, while words of
minimally four syllables long have an initial secondary stress. Words of six or
more syllables display an alternating stress pattern starting at the right edge.
Ignoring the right-aligned position of primary stress, this foot distribution looks
much like that of Garawa (82a), with one major difference: three-syllable words
have a right-aligned foot, rather than a left-aligned foot. Upon comparing the
Indonesian pattern to that of Piro (82b), we nd that everything is identical except
words of seven syllables long. These have secondary stress on the third syllable
in Piro, but on the fourth syllable in Indonesian.
(86)
3 syllables 5 syllables 7 syllables
Indonesian bi.(c.ra) (kn.ti).nu.(.si) (.me).ri.(k.ni).(s.si)
Garawa (pn.ja).a (k.ma).a.(.i) (n.ki).i.(k.rim).(p.yi)
Piro ru.(:.:a) (s.lwa).ye.(hk.kna) (r.slu).(n.ti).ni.(tk.na)
169
Metrical structure and parallelism

In sum, the Indonesian rhythmic pattern is a hybrid of the Garawa and Piro
patterns. It has undominated Rightmost, as in Piro, but the active foot-alignment
constraint is All-Ft-Right, as in Garawa. Accordingly, ranked in between
undominated Rightmost and All-Ft-Right in the Indonesian grammar must be
Align-Wd-Left:

(87) a. Rightmost
Align-Wd-Left

bi.(c.ra)
(b.ca).ra
b. Align-Wd-Left
All-Ft-Right
(kn.ti).nu.(.si)
kon.(t.nu).(.si)
(.me).ri.(k.ni).(s.si)
(.me).(r.ka).ni.(s.si)

A tableau of a three-syllable word shows why the metrical grammar of Indo-

nesian is distinct from that of Garawa: Rightmost dominates Align-Wd-Left.

(88)
Input: / bicara/ Ft-Bin Rightmost Align-Wd- Parse-Syl All-Ft-
Left Right

a. + bi.(c.ra) * *
b. (b.ca).ra *! * *
c. (b).(c.ra) *! **

The tableau of a seven-syllable word shows that Indonesian and Piro have
distinct metrical grammars: alternating stresses run from the right edge in Indo-
nesian, rather than from the left edge, as in Piro.

(89)
Input: /amerikanisasi / Ft- Right- Align- Parse- All-Ft-Right
Bin most Wd-Left Syl

a. + (.me).ri.(k.ni).(s.si) * **, *****

b. (.me).(r.ka).ni.(s.si) * ***, *****!

c. (.me).ri.ka.ni.(s.si) **!* *****

d. a.(m.ri).(k.ni).(s.si) *! * **, ****

e. a.me.ri.ka.ni.(s.si) *!**** *****

f. (.me).(r.ka).(n.sa).si *! * *, ***, *****

g. (.me).(r.ka).(n.sa).(s) *! *, ***, *****

170
 4.4 A set of metrical constraints

4.4.4 A factorial typology of rhythmic and demarcative constraints

We summarize the discussion of rhythmic patterns of section 4.4.1 and section
4.4.3 in a factorial typology. The set of constraints under reranking consists of
Parse-Syl and the metrical alignment constraints All-Ft-X, X-Most, and
Align-Wd-X, with X taking values left or right.
First, when All-Ft-X is undominated, non-iterative systems arise, since only
one foot can stand at the absolute edge of the word. This single-sided non-
iterative pattern can be modied into a double-sided non-iterative pattern by
ranking Align-Wd-Y above All-Ft-X (with X, Y referring to opposite edges).

(90) Non-iterative binary systems

a. All-Ft-X
Parse-Syl (single-sided non
iterative)
b. Align-Wd-Y
All-Ft-X
Parse-Syl (double-sided non-
iterative)

The second major group of metrical systems is characterized by the presence of

alternating rhythm. The distribution of feet is such that all stand maximally close
to an edge of the word, either the left or right edge. This maximal rhythmic
organization is guaranteed by Parse-Syl, which dominates All-Ft-X. This is
achieved by the following ranking:

(91) Ft-Bin
Parse-Syl
All-Ft-X

A factorial typology of iterative systems with strictly binary rhythm arises on the
basis of this ranking schema when we vary the rankings of All-Ft-X with respect
to the PrWd-alignment constraints. (Ft-Bin remains undominated throughout the
factorial typology.)

(92) Iterative binary systems

a. Ft-Bin, X-most
Parse-Syl
All-Ft-X
unidirectional (Pintupi/Warao)
b. Ft-Bin, X-most
Parse-Syl
All-Ft-Y
bidirectional (simple) (Piro/Garawa)
c. Ft-Bin, X-most
Parse-Syl
Align-Wd-Y
All-Ft-X
bidirectional (complex) (Indonesian)

Observe that other rankings of these constraints, or of different values for X and
Y, will result in a duplication of patterns that have already been covered.

4.4.5 Quantity-sensitivity and foot form: WSP and RHTYPE

The nal cross-linguistic preference to be discussed in this section is quantity-
sensitivity, the matching of syllable weight and prominence. Perhaps the most
171
Metrical structure and parallelism

important constraint in this respect is the Weight-to-Stress Principle (Prince 1983,

Prince and Smolensky 1993):

(93) WSP
Heavy syllables are stressed.

WSP relates syllable weight and metrical prominence. Or, as argued by Prince
(1983), this amounts to the natural requirement that intrinsic sonority prominence
of heavy syllables be registered as stress prominence. This relation between
sonority and prominent positions in prosodic structure can also be seen in syl-
labication segments of the highest sonority are the optimal syllable peaks
(Prince and Smolensky 1993).9

Two additional foot-form constraints determine the rhythmic type of feet (that is,
iambic or trochaic):10

(94) a. RhType=I
Feet have nal prominence.
b. RhType=T
Feet have initial prominence.

These constraints simply state the position of the head in a foot, without requiring
that this matches its quantitative make-up. Partly this function is carried out by
WSP, but this is not the end of the story, as we will now see.
Tabulating the quantitative shapes of iambs and trochees that violate neither
Ft-Bin nor WSP, we arrive at the following:

(95) Satisfy WSP & Ft-Bin Violate WSP Violate Ft-Bin

Iambs: (H), (LL), (LH) (HL), (HH) (L), (LLL), etc.
Trochees: (H), (LL), (HL) (LH), (HH) (L), (LLL), etc.

The combined ltering functions of WSP and Ft-Bin still underdetermine the
choice of quantitative make-up of iambs and trochees. For example, consider two
logically possible iambic and trochaic parses of the following sequences:

(96) Iambs . . . (LL)(LL)(H) . . . or . . . L)(LL)(LH) . . .

Trochees . . . (H)(LL)(LL) . . . or . . . (HL)(LL)(L . . .

9
Related constraints require that stressed syllables be heavy (SWP, Riad 1992) or that within
a domain the heaviest syllable be stressed (Pk-Prom, Prince and Smolensky 1993).
10
For arguments that RhType=I should be eliminated as a constraint, see van de Vijver (1998).

172
 4.4 A set of metrical constraints

Extensive cross-linguistic research into stress systems (Hayes 1995) reveals that
the best quantitative shapes of disyllabic iambs and trochees are (LH) and (LL),
respectively.

(97) Quantitative asymmetry between iambs and trochees (Hayes 1995):

Iambs: (LH)
(LL) Uneven iambs are preferred to even
iambs.
Trochees: (LL)
(HL) Even trochees are preferred to uneven
trochees.

We have already observed a preference for uneven iambs (LH) to even iambs
(LL) in the length pattern of Hixkaryana. Hayes (1995) supports the uneven iamb
with cases of iambic lengthening from many languages. There is ample cross-
linguistic evidence for the reverse preference for quantitative evenness in trochaic
systems. For example, English and Fijian shorten a long vowel precisely in those
contexts where the resulting light syllable is footed with a following light syllable
in an even (balanced) trochee (LL). The example below is from English (Myers
1987, Prince 1990):

(98) a. cone [(kn)] (H)

b. con-ic [(kL.nk)] (LL)
[(k).nk] (H)L

Ideally, a single explanation would exist for the markedness of both uneven (HL)
trochees and even (LL) iambs. Various attempts at rhythmic explanations have
been made in the pre-OT literature, which will not be reviewed here (Hayes 1985,
1995, Prince 1990). Let us instead consider an approach developed in Kager
(1993, 1995), based on the idea that a heavy syllable is rhythmically composed
of two microbeats, one on each composing mora (Prince 1983). In (99) the
microbeats of a heavy syllable are represented in a grid-notation (where * rep-
resents a strong beat, and . a weak beat):

(99) * . micro-level strongweak rhythmic contour of a heavy syllable

[]

This micro-level strongweak contour on heavy syllables is grounded in sonority

factors: in a CVC syllable, for example, the vocalic peak is intrinsically of higher
sonority, and hence of greater prominence, than the consonantal coda. (More gen-
erally, the sonority peak of the syllable is situated as early as possible in the nucleus.)
We can now represent iambs and trochees of different quantitative composi-
tions in the following way, to bring out their rhythmic contours at the (moraic)
micro-level:
173
Metrical structure and parallelism

(100) a. ending in strongweak b. not ending in strongweak

(i) * . (ii) * . *. .
[] [][] [][]

(H) (L L) (H L)
(101) a. ending in strongweak b. not ending in strongweak
(i) * . (ii) . *. . *
[] [][] [][]

(H) (L H) (L L)
Note that at this micro-rhythmic level, the cross-linguistically preferred feet {(H),
(LL), (LH)} end in a strongweak contour, while dispreferred {(HL), (LL)} do
not end in such a contour. The preference for feet of the rst type is stated in the
following constraint:

(102) Rh-Contour
A foot must end in a strongweak contour at the moraic level.

We will assume that this constraint underlies the quantitative asymmetry in the
foot inventory. Languages having this constraint in a high-ranked position of the
grammar employ strategies to avoid (HL) trochees or (LL) iambs, depending on
interactions with specic faithfulness constraints. Languages having it in a low-
ranked position, specically below faithfulness constraints preserving input quan-
tity, will allow these anti-rhythmical feet, which gives the impression of quantity-
insensitive stress.

Let us now turn to some other issues in the theory of quantity-sensitivity. The
rst relates to the often made assumption that quantity-sensitivity is a property
of languages as a whole (and of prosodic systems in particular), which can be
switched on or off. In the pre-OT literature, this was encoded as a parameter
of foot construction (Hayes 1980), or a choice of foot type from the universal
inventory (moraic trochee versus syllabic trochee, Hayes 1995). If the OT
approach is generally correct, then this holistic view cannot be correct, for the
following reasons.
First, it is held that grammars are rankings of all universal constraints, rather
than rankings of specic sets of constraints selected on a language-specic basis.
This has an important consequence: all constraints are potentially active in every
grammar. The presence of a quantity contrast11 in a language leaves open the
11
For example, a vowel length contrast reects the domination of faithfulness (Max--IO) over
markedness (No-Long-V, Rosenthall 1994).

174
 4.4 A set of metrical constraints

ranking of metrical constraints which refer to this quantity contrast (such as WSP
or Rh-Contour). It is then predicted that constraint re-rankings produce various
degrees of quantity-sensitivity. Typological studies of trochaic languages with
quantity contrasts (Kager 1992a, b) conrm this scattered picture: there is a range
from fully quantity-sensitive systems, in which every heavy syllable is stressed,
to systems which assign more importance to other factors, such as binary rhythm,
at the expense of stress on heavy syllables. (But no language completely disrespects
its quantity-contrast in its metrical system, and there is always some aspect to
which quantity does matter. For example, most languages have bimoraic word
minima.)
The second reason for which quantity-(in)sensitivity cannot be a global prop-
erty of languages is that the pre-OT foot inventory is decomposed into rhythmic
and quantitative constraints, such as FtBin, WSP, Rh-Contour, and RhType.
Under this approach, no direct counterpart of quantity-insensitive feet remains, and
quantity-disrespecting effects which were previously attributed to the syllabic
trochee must now be reattributed to rhythmic constraints. As pointed out by
Kager (1992a, 1995), one of the strongest rhythmic factors to produce quantity-
disrespecting stress patterns is clash avoidance. When No-Clash takes pre-
cedence over WSP, some heavy syllables must give up their prominence in order
to avoid a clash with a neighbouring stressed syllable. This is the case in Estonian,
Finnish, and other trochaic languages. (See Alber 1997 for an analysis along these
lines.)

Another important issue in quantity-sensitivity is its relationship with processes

of epenthesis (or lengthening) and deletion (or shortening). It has been argued in
the metrical literature that such processes may achieve metrical targets, for exam-
ple ideally shaped feet (Dresher and Lahiri 1991, Hayes 1995). This observation
has led to the hypothesis that phonologies are metrically coherent: foot well-
formedness is initially asserted in metrical parsing, and is reasserted by the quan-
titative changes operating on the output of parsing. Such metrical coherence is
entirely expected under OT, since this theory no longer makes a distinction of
labour between initial metrical parsing and quantitative processes: both are
evaluated at a single level, that of the output. Accordingly cases that are put for-
ward as examples of metrical coherence are straightforwardly analysed under OT.
Very interesting examples showing the relevance of foot shape and foot posi-
tion to quantitative changes come from two Polynesian languages, Boumaa Fijian
and Tongan. Both languages avoid a quantitative sequence of a heavy plus light
syllable at the end of a word (HL#) and both strive toward a situation in which a
bimoraic trochee is right-aligned in the word. To achieve this goal, Boumaa Fijian
(Dixon 1988, Prince 1990, Hayes 1995) employs trochaic shortening, while Tongan
175
Metrical structure and parallelism

(Churchward 1953, Mester 1991, Prince and Smolensky 1993, Hayes 1995)
breaks the long vowel:

(103) Two reactions to a sequence . . . HL#

a. Fijian: trochaic shortening
/siii/ s.i to exceed
/siii-ta/ si..ta to exceed-trans
b. Tongan: vowel breaking
/ huu/ hu to go in
/ huu-/ hu. to open ofcially

The grammars of both languages have undominated foot shape constraints Ft-
Bin and Rh-Contour, as well as undominated Rightmost. But they differ min-
imally in the reranking of Onset and Max--IO. Fijian gives up input quantity
to attain the perfect bimoraic foot at the right edge (violating Max--IO):

(104)
Input: /siii/ Ft-Bin Rightmost Rh-Contour Onset Max--IO

a. + (s.i) *
b. si.(.i) *!
c. (si.i) *!
d. (si).i *!
e. (si).() *!

Tongan selects another strategy to accomplish the same conguration: it breaks

the long vowel into two short vowels, respecting input quantity, but violating
Onset:

(105)
Input: / huu-/ Ft-Bin Rightmost Rh-Contour Max--IO Onset

a. (h.) *!
b. + hu.(.) *
c. (hu.) *!
d. (hu). *!
e. (hu).(f) *!

176
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

We nd another example of functionally related processes (see chapters 2 and 3).

Both vowel shortening (Fijian) and vowel breaking (Tongan) serve to achieve a
single output goal: the perfect bimoraic foot at the right word edge.
The next section supports these results by a case-study of Southeastern Tepe-
huan, a language that exhibits an interesting interaction of foot shape and quan-
titative processes.

4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

This section deals with interactions between well-formedness constraints and
faithfulness constraints, focussing on rhythmically induced vowel deletion in
Southeastern Tepehuan.

4.5.1 Accent and syncope: the patterns

Southeastern Tepehuan (E. Willett 1982, T. Willett 1991, Rice 1992, Kager 1997a)
is an Uto-Aztecan language spoken southeast of Durango in Mexico. It has an
interesting process of vowel deletion (syncope) in prosodically dened contexts:
vowel deletion is rhythmic, in the sense that alternating vowels are affected.
Surprisingly, there is no phonetic evidence for a corresponding rhythmic pattern
that might condition syncope. Lacking this pattern, we must answer two ques-
tions: (i) what prosodic factors trigger syncope? (ii) what blocks the application
of syncope to adjacent syllables? As we will see, an interaction of prosodic well-
formedness constraints with segmental faithfulness constraints captures the
syncope pattern, without reference to an abstract layer of syncope feet or an
intermediary level of derivation.

Before we can look into the contexts of syncope, we must discuss the place of
the accent. Accent falls on the initial stem syllable when it is heavy (long-
vowelled or closed), see (106ac). It falls on the second stem syllable if this is
heavy while the rst syllable is light; see (106df). (No stem begins with a sequ-
ence of two light syllables: this output gap is due to syncope.) The accent pattern
points to a single iamb (H) or (LH) at the left edge:
(106) #(H) #(LH)
a. (vo).hi bear d. (ta.ka).rui chicken
b. (kr).va goat e. (sa.pk) story
c. (tat).pM eas f. (ta.pGM) ea
As mentioned, Southeastern Tepehuan has a vowel alternation12 that is due to
syncope of vowels in even-numbered open syllables (107c), including long vowels

12
Two other vowel alternations will not be discussed here: vowel shortening and apocope. See
Kager (1997a) for an integrated analysis of syncope, apocope, and shortening.

177
Metrical structure and parallelism

(107d). Syncope applies in the domain of the stem, which may include a reduplic-
ant prex (107b d).

(107) a. /tMrovi/ (tGr).vi rope

b. /tGM-tarovi/ (tGMt).ro.pi ropes
c. /maa-matuidja/ (mam).tu.dja will teach
d. /gaa-gaaga/ (ga).ga he will look around for it

Examples in (108) illustrate the blocking of syncope in heavy syllables following

a light rst syllable. In (108ab) this is shown for the context #CVCVV, and in
(108cd) for the context #CVCV(V)C.

(108) a. /takaarui/ (ta.ka).rui chicken

b. /va-vaium/ (va.pi).um metals
c. /ka-karva/ (ka.kr).va goats
d. /ha-haannu/ (ha.han).nu clothes

The central question is: what is the relation between metrical structure and syn-
cope? The alternating nature of syncope suggests a foot-based analysis. Yet there
is no secondary stress pattern supporting this hypothesis.

4.5.2 An OT analysis
Our analysis of syncope in Southeastern Tepehuan starts from the following
observation. Syncope obviously occurs at the expense of input vowels, hence
there must be some output target that is achieved at the expense of faithful-
ness. Below we will argue that this target is exhaustive parsing of syllables
by feet. By deleting vowels, the number of syllables will be reduced as well,
and thereby the number of violators of Parse-Syl. To those who have carefully
read through the previous sections, syncope may sound a rather absurd strategy
to achieve an exhaustive metrical parsing. Did we not conclude from analyses of
Hixkaryana, Pintupi, Warao, Garawa, and other languages, that exhaustivity is
attained by multiple (iterative) footing, hence rhythmic alternation due to a rank-
ing Parse-Syl
All-Ft-L? Why would languages which opt for this ranking
need vowel deletion to achieve exhaustivity? This question is equal to asking
under what conditions vowel deletion is triggered in Southeastern Tepehuan.
What is the difference between the strategies, iterative footing and syncope, in
terms of constraint rankings? It resides in the priorities of three factors: faithful-
ness to input vowels (Max-V-IO), proximity of feet to edges, and exhaustivity.
Core rankings are below:

178
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

(109) Two strategies toward exhaustive metrication

a. Iterative footing (Pintupi): Parse-Syl, Max-V-IO

All-Ft-L
b. Syncope (Southeastern Tepehuan): All-Ft-L
Parse-Syl

Max-V-IO

A preliminary answer to the question What causes syncope in Southeastern Tehe-

huan? is this: the language gives more priority to its metrication than to its
input vowels. It allows only one foot per word, strictly at the left edge. This foot
should parse as many syllables as possible but it must not grow larger than the
disyllabic iamb. Any syllables that are not contained by it violate Parse-Syl. A
most effective way of reducing the number of such syllables is deleting their
vowels, without which they must perish. Yet there is a powerful constraint oppos-
ing across-the-board vowel deletion: all outputs must be syllabied in keeping
with rigid phonotactics which rule out complex onsets and codas. In sum, while
syncope is set in motion by prosodic well-formedness constraints at the metrical
level, it is simultaneously checked by prosodic well-formedness constraints at the
syllabic level.
But what is the role of segmental faithfulness in this all-prosodic conict? As
we will see, it plays a key role: while vowels are easily deleted, consonants are
not. But since consonants need to be parsed into well-formed syllables, a neces-
sary minimum of vowels must remain. Paradoxically, faithfulness to consonants
is the key factor in the preservation of vowels by an indirect link the canonical
syllable.
This scenario is deeply parallel in the following sense. Interaction occurs
between well-formedness constraints whose targets are at different levels of the
prosodic hierarchy the foot and the syllable. Metrical structure mounts pressure
to minimize syllable number, yet syllable well-formedness militates against this.
At the level of the syllable, consonants (immune to deletion due to high-ranking
faithfulness) require vowels to be able to surface in well-formed syllables. There-
fore segmental faithfulness to consonants, indirectly, saves a certain number of
vowels from deletion.

Let us now consider how this syncope scenario is encoded in constraint interac-
tion. We start by observing that ranking (109b) has two subrankings, All-Ft-L

Parse-Syl and Parse-Syl
Max-V-IO. Both will be motivated below.
First, Southeastern Tepehuan has non-iterative footing. This implies that it
has the reverse ranking (at least, as compared to Pintupi) of All-Ft-L and
Parse-Syl:

179
Metrical structure and parallelism

(110) All-Ft-L
Parse-Syl

a. /CVVCVCVCVC/ (CVVC).CV.CVC
(CVV).CV.CV.CVC
/tMM-tarovi/ (tGMt).ro.pi (tGM).(tM.r).(p)
b. /CVVCVCVCVCVC/ (CVVC).CVC.CVC
(CVV).CV.CV.CV.CVC
/maa-matuidja/ (mam).tu.dja (ma).(ma.t).(i.dj)

This key ranking is illustrated in a tableau:

(111) Input: /tMM-tMrovi/ All-Ft-L Parse-Syl

a. + (tGMt).ro.pi **
b. (tGM).(tM.r).(p) *!, ***

Second, the subranking Parse-Syl
Max-V-IO expresses that syncope serves

to improve exhaustive footing. This is demonstrated schematically in CV-skeletal
forms, as well as in genuine forms, in (112). Observe that the optimal outputs are
closer to exhaustive parsing than their competitors, since they contain a smaller
number of unparsed syllables.

(112) Parse-Syl
Max-V-IO

a. /CVVCVCVCVC/ (CVVC).CV.CVC
(CVV).CV.CV.CVC
/tMM-tarovi/ (tGMt).ro.pi (tGM).tM.ro.pi
b. /CVVCVCVCVCVC/ (CVVC).CVC.CVC
(CVV).CV.CV.CV.CVC
/maa-matuid ja/ (mam).tu.d ja (ma).ma.tu.i. d ja

(113) Input: /tMM-tMrovi/ Parse-Syl Max-V-IO

a. + (tGMt).ro.pi ** *
b. (tGM).tM.ro.p ***!

Observe also that the optimal forms of (112) have not yet completely achieved
the target, as they still contain two unparsed syllables each. How can exhaustivity
form the pressure behind syncope if it cannot be achieved completely? The
answer is typical of OT: there is yet another factor which precludes the full
achievement of the target, and this has priority. To nd out what this factor is, all
we have to do is take exhaustive parsing to its ultimate logical conclusion, and
see what happens.
Note that optimal forms of (112) cannot be compressed any further by the
deletion of additional vowels without fatally violating rigid syllable phonotactics
of Southeastern Tepehuan, requiring syllables to have non-branching margins.
This is due to undominated *Complex.
180
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

(114) *Complex
Parse-Syl

a. /CVVCVCVCVC/ (CVVC).CV.CVC
(CVVC).CCVC
/tMM-tarovi/ (tGMt).ro.pi (tGMt).rpi
b. /CVVCVCVCVCVC/ (CVVC).CVC.CVC
(CVVC).CCVCC
/maa-matuidja/ (mam).tu.dja (mam).
adj

Righthand forms in (114) fatally violate undominated *Complex, even though

they have one fewer unparsed syllable than the optimal forms. A tableau shows
this for (114a):

(115) Input: /tMM-tMrovi/ *Complex Parse-Syl

a. + (tGMt).ro.pi **
b. (tGMt).rpi *! *
c. (tGMt).rop *! *

Next observe that, unlike vowels, consonants are never deleted under pressure
of exhaustivity. For example, the stem /karva/ might, in principle, surface as a
single (LH) iamb without violation of Parse-Syl, if only one of its medial
consonants were deleted: e.g. *(ka.v) or *(ka.r), rather than (kr).va. This
never occurs outputs must preserve the consonants of their inputs. Therefore
Max-C-IO must dominate Parse-Syl:

(116) Input: /karva/ Max-C-IO Parse-Syl

a. + (kr).va *
b. (ka.v) *!
c. (ka.r) *!

Summing up results of the discussion so far, we arrive at the preliminary ranking

in (117):

(117) *Complex, All-Ft-L, Max-C-IO
Parse-Syl
Max-V-IO

This is the basic analysis: modications will later be added, in a more detailed
discussion. But it is not too early to claim an important result. This is that we
have an explanation for the alternating or rhythmic character of syncope. Syncope
of vowels in adjacent syllables would necessarily lead to consonant clusters that
cannot be parsed by canonical syllables, while any deletion of consonants is
blocked (see again 11516 for evidence).
181
Metrical structure and parallelism

Another immediate result, again typical of OT, is this: if the target of syncope
is exhaustive parsing of syllables into feet, then we predict that no deletion takes
place when nothing can be gained by it when it would yield no progress in terms
of syllable parsing. This is exactly what we nd. In (118) syncope fails to apply,
simply because it would not move the output any closer to the target, exhaustivity.
Everything else being equal, it is always better to be faithful to input segments:

(118) /takaarui/ (ta.ka).rui
*(tk).rui

Thus Max-V-IO may be dominated, but that does not mean that its role is
neglectable:

(119) Input: /takaarui/ *Complex All-Ft-L Parse- Max-

Syl V-IO

a. + (ta.ka).rui *
b. (tk).rui * *!
c. (ta.ka).(ri) *!

Thus far we have developed the idea that exhaustivity of metrical parsing and
left foot alignment have priority over realizing underlying vowels in the output.
Forces that drive vowel deletion are effectively counterbalanced by phonotactics:
syllables must not have branching margins. Having completed the core of the
analysis, we now move on to some residual matters.

First consider the fact that there is another potential method by which outputs
may improve their completeness of metrication. By shortening the initial long
vowel in /tMM-tMrovi/, a canonical (LH) iamb would be created in the proper
place, producing *[(tM.tGr).pi]. This output would fare better in terms of exhaust-
ivity than the actual output [(tGMt).ro.pi]. What blocks initial shortening here is
the templatic requirement that the reduplicative prex must equal a heavy syllable
(McCarthy and Prince 1986, 1993b):13

(120) Red=
The reduplicant equals a heavy syllable.

13
The prosodic size of the reduplicative prex (a heavy syllable or a light syllable) is based on a
stem-specic choice: compare the examples in (107b d) with those in (108bd). Reduplica-
tive allomorphy is apparently not governed by phonological or morphological factors. See
Kager (1997a) for details.

182
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

The relevant ranking is that in (121), and illustration of it is offered in tableau

(122):

(121) Red=
Parse-Syl

/CVVCVCVCVC/ (CVVC).CV.CVC
(CV.CVC).CVC
/tMM-tarovi/ (tGMt).ro.pi (tM.tGr).pi

(122) Input: /tMM-tMrovi/ Red= Parse-Syl

a. + (tGM-t).ro.pi **
b. (tM-tGr).pi *! *

Second, we must pay attention to yet another aspect of prosodic well-formed-

ness: foot shape. Southeastern Tepehuan iambs are (H) and (LH), but never (LL).
Avoidance of (LL) is attributed to Rh-Contour (102), requiring iambs to end in
falling sonority contours (or heavy syllables (CVC, CVV, or CVVC) in the case
of Southeastern Tepehuan). This is the second force (besides Parse-Syl) driving
syncope. It explains why syncope affects all inputs beginning with #CVCV:

(123) /tMrovi/ (tGr).vi
(tM.r).vi

Rh-Contour is undominated, here at the expense of the anti-vowel-deletion

constraint Max-V-IO. Interestingly, the royal way to satisfy Rh-Contour, by
lengthening a vowel to create the canonical (LH) iamb, is blocked:

(124) /tMrovi/ (tGr).vi
(tM.r).vi

This signals that the constraint militating against vowel lengthening, Dep--IO,
dominates the anti-deletion constraint Max-V-IO. In sum, we have the following
ranking:

(125) Weight increase of stressed syllables at the expense of input vowels

Rh-Contour, Dep--IO
Max-V-IO

This ranking is illustreated by the forms of (1234) in tableau (126):

(126) Input: /tMrovi/ Rh-Contour Dep--IO Max-V-IO

a. + (tGr).vi *
b. (tM.r).vi *!
c. (tM.r).vi *!

183
Metrical structure and parallelism

Finally we must consider in more detail the quantitative shape of the foot, and
try to integrate foot shape constraints into the current ranking. So far we have
assumed that the single foot standing initial in the stem is a binary iamb with its
strong syllable heavy, (H) or (LH). This selection of iambs is due to a set of four
undominated constraints stated in (127), which will be united here into a cover
constraint Ft-Form (for expository reasons).

(127) Ft-Form (cover constraint)

a. Ft-Bin Feet are binary under moraic or syllabic analysis.
b. RhType=I Feet are right-headed.
c. Rh-Contour Feet must end in a strongweak contour at the
moraic level.
d. WSP-Ft Heavy syllables within the foot are prominent.

Note that WSP-Ft serves to exclude the anti-rhythmic iamb (HH) (see Kager
1997a).
Adding up all the subrankings motivated so far, we arrive at:

(128) Rhythmic vowel deletion in Southeastern Tepehuan

*Complex, Ft-Form, All-Ft-Left, Red=, Max-C-IO
Dep--
IO
Parse-Syl
Max-V-IO

The rhythmicity of the syncope pattern follows from this ranking there is no
need to set up an intermediate representation of iterative trochees, as in the rule-
based analysis (see section 4.5.3).
Tableaux (129 30) illustrate the full constraint ranking in action.

(129)
Input: /takaarui/ *Complex Ft- All- Red Max- Dep- Parse- Max-
Form Ft-L = C-IO -IO Syl V-IO

a. + (ta.ka).rui *

b. (tk).rui * *!

c. (ta.kar) *! *

d. (ta.ka).(ri) *!*

e. (ta.kaa.ri) *!

f. (tak.ri) *! *

g. (ta.kar) *! *

184
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

(130)
Input: /tMM-tMrovi/ *Complex Ft- All- Red Max- Dep- Parse- Max-
Form Ft-L = C-IO -IO Syl V-IO

a. + (tGMt).ro.pi ** *

b. (tGM).t .ro.pi ***!

c. (tGMt).pi *! * **

d. (tM.tGr).pi *! * *

e. (tGMt).(ro.p) *! *

f. (tMM.tGr).pi *! * *

g. (tGMt).rop *! * **

4.5.3 Alternative: a rule-based analysis

Analysing the Southeastern Tepehuan syncope pattern, derivational theory has
one option available that is lacking from OT: setting up an intermediate level of
the derivation that contains an exhaustive parsing into binary feet, from which
syncope may be predicted. In such an analysis, syncope would become a natural
reduction rule: weakening unstressed syllables all the way up to the point of
deletion. For principled reasons, such an analysis is unavailable in OT: well-
formedness constraints refer to the surface level only, and cannot have access to
any other derivational level.
A derivational analysis based on intermediate levels is given below (after Rice
1992):

(131) a. Syllabify: Assign syllable structure.

b. Assign feet: Build left-headed quantity-sensitive feet from left
to right.
c. End Rule Left: Promote the leftmost foot to main stress.
d. Syncopate: Delete vowels in non-head syllables (and
resyllabify).

This analysis is illustrated in (132) by the derivation of [(mam).tu.d ja]:

(132) UR /maa-matuid ja/

Syllabify maa.ma.tu.i.d ja
Assign feet (ma.ma).(t.i).(d j)
End Rule Left (ma.ma).(t.i).(d j)
Syncopate (mam).(t).(d j)
PR [(mam).tu.d ja]
185
Metrical structure and parallelism

Although the analysis derives most forms correctly, it runs into conceptual and
empirical problems that cannot be solved without additional machinery, including
extra rules and even output constraints.
First, the complete lack of surface secondary stress must be accounted for by
a rule of stress conation (Halle and Vergnaud 1987), eliminating secondary
stress feet. This rule obscures the conditioning environment on which syncope
is based, rendering the process opaque. From the viewpoint of the language
learner, the abstractness of this analysis, as well as the extrinsic rule ordering on
which it is based, is problematic. (See chapter 9 for discussion of opacity in OT.)
Second, iterative uneven trochees do not sufce to predict the syncope and
accent pattern. Problematic are words beginning with a sequence #CVVCVV.
Recall that syncope deletes the long vowel in the second syllable, as in /gaa-
gaaga/ [ga.ga]. Quantity-sensitive footing produces a monosyllabic foot
on the heavy second syllable, which must be rst deleted in clash, before syncope
applies, see (133b). Now compare these #CVVCVV cases with those beginning
with a sequence #CVCVV, as in [ta.ka.rui] (133a). These surface with accent
on the second syllable. The quantity-sensitive trochaic parsing results in the inter-
mediary stage [(t).(ka).(ri)], in which the rst foot, rather than the second, is
deleted. The foot deletion rule should be quantity-sensitive in order to make the
proper distinction: if both syllables are equally heavy (#HH, as in 133b), it deletes
the righthand foot, but if the lefthand syllable is lighter (#LH, as in 133a), it
deletes the lefthand foot:

(133) UR a. /takaarui/ b. /gaa-gaaga/

Syllabify ta.kaa.rui gaa.gaa.ga
Assign feet (t).(ka).(ri) (ga).(ga).(g)
Delete foot ta.(ka).(ri) (ga).gaa.(g)
End Rule L ta.(ka).(ri) (ga).gaa.(g)
Syncopate --- (gag).(g)
Conate (ta.ka).rui (gag).ga
PR [(ta.ka).rui] [(ga).ga]

The quantity-sensitive destressing rule that compares the relative weight of the
rst two syllables is a powerful device. Still worse, this analysis misses the gen-
eralization that destressing and syncope conspire toward outputs that begin with
iambs: (H) or (LH).
Third, of the four trochees that the analysis sets up to dene the context of
syncope {(LL), (HL), (H), (L)}, three never surface (only (H) does). This greatly
contributes to the abstractness of the analysis. (Moreover, two of these abstract
trochees, degenerate (L) and uneven (HL), are cross-linguistically marginal; see
Hayes 1995.) In spite of the used trochees, surface stress contours in the stem,
186
 4.5 Case-study: rhythmic syncope in Southeastern Tepehuan

#H and #LH, are easily interpreted as iambic, given the fact that foot constituency
is inaudible.
Fourth, the rule-based analysis must be extended by phonotactic output condi-
tions on syncope that block it when it would create a sequence that cannot be
syllabied. For example, the initial light syllable in a word beginning with
#CVCVV (cf. [ta.ka.rui]) must not undergo syncope, even though destressing
places it in a metrically weak position. Another context in which the rule-based
analysis cannot do without an output constraint is that of an initial closed syllable
that is followed by a light syllable, again metrically weak (cf. [vt.vi.rak]):
(134) UR a. /takaarui/ b. /vatvirak/
Syllabify ta.kaa.rui vat.vi.rak
Assign feet (t).(ka).(ri) (vt.vi).(rk)
End Rule L ta.(ka).(ri) (vt.vi).(rk)
Syncopate blocked blocked
Conate (ta.ka).rui (vt.vi).rak
PR [(ta.ka).rui] [(vt.vi).rak]
Clearly this blocking reects an output constraint on syncope: it must not create
syllables with branching onsets or codas (cf. *Complex). Output constraints on
syncope defeat the derivational nature of rule-based theory. But a mixed theory
that has both rules and output constraints might still be acceptable if the rules
were motivated independently of the constraints so as to satisfy output targets,
which is not the case here, as the previous three objections have shown. Without
such independent evidence for separate rules and constraints, the OT analysis
remains conceptually superior.

Let us summarize the results, and make a nal comparison between the rule-
based analysis and the OT analysis. Below I have summed up ve contexts in
which syncope in Southeastern Tepehuan is blocked, and the explanations for
blocking by both analyses:

(135) Rhythmicity: no syncope of both vowels in adjacent syllables.

Rule-based analysis: There is an abstract intermediate level in the
derivation at which words are parsed exhaustively by binary feet.
Targets of syncope are syllables in weak positions of binary feet.
OT analysis: Deletion of vowels in adjacent syllables would inevit-
ably create clusters of three or more consonants. These long clusters
cannot be syllabied without fatally violating undominated syllable
well-formedness constraints, in particular *Complex. (There is no
remedy in further simplifying clusters by deletion of consonants: this
escape route is effectively shut off by undominated Max-C-IO.)
187
Metrical structure and parallelism

(136) Initial blocking: no syncope of a vowel in an initial syllable.

Rule-based analysis: Syncope feet are trochaic, hence the initial syl-
lable is rhythmically strong. But when the initial syllable is rhythm-
ically weak (as in #CVCVV . . . ) syncope must be blocked by an
output constraint against complex onsets.
OT analysis: Vowels in initial syllables are immune from syncope
since their deletion (in #CVCV or #CVCVV) would produce an out-
put #CCV(V), again in violation of *Complex.

(137) Quantity-sensitivity I: no syncope of a vowel in a closed syllable.

Rule-based analysis: Syncope feet are quantity-sensitive, and closed
syllables are heavy. Hence closed syllables are heads of feet.
OT analysis: Again, the deletion of a vowel in a closed syllable would
result in a fatal violation of *Complex.

(138) Quantity-sensitivity II: no syncope of a light syllable following a

closed syllable.
Rule-based analysis: Blocked by an output constraint against com-
plex margins.
OT analysis: Again, the deletion of a vowel in a closed syllable would
result in a fatal violation of *Complex.

(139) Quantity-sensitivity III: no syncope of long vowel following initial

light syllable.
Rule-based analysis: Quantity-sensitive destressing repairs the (inter-
mediary) trochaic parse #(L)(H) into #L(H).
OT analysis: Nothing is gained by syncope in terms of exhaustivity.
Since there is no trigger, there will be no effect faithfulness forbids.

In sum, we nd that a rule-based analysis uses excessive machinery to achieve

effects that an OT analysis attributes to a single interaction. Let me hasten to add
that the rule-based analysis reviewed here (essentially Rice 1992) might perhaps
be improved in ways which reduce the machinery. The reason to dedicate so much
discussion to this analysis was not pointed against this specic analysis, but rather
to point out various general problems that are inherent to derivational theory.

4.5.4 Parallel interactions between prosodic levels

The analysis of Southeastern Tepehuan is based on parallel interaction of
metrical well-formedness and faithfulness. In earlier sections we have observed
Parse-Syl active in its most obvious role, of enforcing exhaustive parsing in lan-
guages such as Pintupi. However, as the analysis of Southeastern Tepehuan shows,
vowel deletion is yet another strategy to achieve the same goal of exhaustivity:
188
 4.6 Conclusions

when a vowel is deleted, its syllable goes, and with it a potential violator of
Parse-Syl. In evaluating this deletion strategy towards exhaustivity, it is helpful
to make explicit the truth conditions of Parse-Syl:

(140) Dual interpretation of Parse-Syl

a. IF X forms a syllable, THEN X must be parsed by a foot.
(bottom-up)
b. IF X is not parsed by a foot, THEN X must not form a syllable.
(top-down)

It is not too difcult to imagine under which conditions the deletion strategy may
become optimal. Even though vowel deletion involves the violation of a seg-
mental faithfulness constraint (Max-V-IO), there may be special reasons favour-
ing it over the standard way of achieving exhaustivity (the creation of multiple
feet). Of course, these special reasons have to do with the costs of multiple feet
in terms of metrical well-formedness violations. A major candidate for a con-
straint resisting the assignment of multiple feet is All-Ft-X.
Perhaps the clearest contribution of OT to the eld of prosody resides in captur-
ing parallel interactions of prosodic well-formedness constraints and faithfulness
constraints, allowing for a highly generalizing theory of quantitative phenomena
(vowel shortening, vowel lengthening, gemination, syncope, word minima, etc.).
Parallelism establishes complete communication between all levels of the prosodic
hierarchy. Consequently, changes at lower prosodic levels have effects at higher
levels, and vice versa. In rule-based theory it is not a priori impossible to describe
such chains of effects between prosodic levels. In fact, quantitative rules may
refer to foot structure, while foot assignment rules may refer to quantity. But such
descriptive means leave totally unexpressed the organicity of metrical systems.
The achievement of OT is its elegance of expression: no direction of processes
(upward or downward through the prosodic hierarchy) needs to be stipulated, as
effects of changes at different levels are weighed against each other. If some
action downstairs happens to be too costly in terms of the constraint hierarchy,
then an alternative action in the upstairs region will automatically occur.

4.6 Conclusions
The typology of metrical systems has been a highly fruitful branch of pre-OT
phonology. Theoretical attractiveness of this eld resides mainly in two factors:
on the one hand, the relative autonomy of metrical principles from segmental
phonology; on the other hand, a great deal of cross-linguistic variation in stress
systems. In rule-based theory, this invited an approach based on metrical prin-
ciples and parameters, resulting in a typology that had the dual qualities of com-
pactness and symmetry (though perhaps it was overly symmetric, as Hayes 1995
189
Metrical structure and parallelism

emphasized). In comparison, OT offers a number of additional advantages over

rule-based theory in the eld of metrical phonology.
Meanwhile we have also learned a bit more about the architecture of OT itself.
We have seen new episodes in the ongoing battle between faithfulness and well-
formedness. This conict was at the heart of the success in handling parallel
interactions of rhythm and syncope in Southeastern Tepehuan: input vowels are
sacriced to achieve better parsing. The key observation was that input vowels are
never given up for free: there must be some pressure (due to a well-formedness
constraint) to syncopate, and when vowels are deleted, their number always (pre-
cisely) sufces to restore the balance. This is the economy principle do-only-
when-necessary, a true hallmark of OT.

SUGGESTIONS FOR FUR THER READING

Metrical phonology
Halle, Morris and Jean-Roger Vergnaud (1987) An essay on stress. Cambridge,
Mass.: MIT Press.
Hayes, Bruce (1995) Metrical stress theory: principles and case studies. Univer-
sity of Chicago Press.
(1980) A metrical theory of stress rules. PhD dissertation, MIT, Cambridge,
Mass. (Published 1985, New York: Garland Press.)
Prince, Alan (1983) Relating to the grid. Linguistic Inquiry 14. 19 100.
Selkirk, Elisabeth O. (1984) Phonology and syntax: the relation between sound
and structure. Cambridge, Mass.: MIT Press.

Metrical studies in OT
Cohn, Abigail and John McCarthy (1994) Alignment and parallelism in Indone-
sian phonology. Ms., Cornell University and University of Massachusetts,
Amherst. [ROA-25, http://ruccs.rutgers.edu/roa.html.]
Inkelas, Sharon (forthcoming) Exceptional stress-attracting sufxes in Turkish:
representations vs. the grammar. To appear in Kager, Van der Hulst, and
Zonneveld. [ROA-39, http://ruccs.rutgers.edu/roa.html]
Kager, Ren (1994) Ternary rhythm in alignment theory. Ms., Utrecht University.
[ROA-35, http://ruccs.rutgers.edu/roa.html]
Kenstowicz, Michael (1995a) Cyclic vs. non-cyclic constraint evaluation. Phono-
logy 12. 397436.
Pater, Joe (1995) On the nonuniformity of weight-to-stress and stress preservation
effects in English. Ms., McGill University. [ROA-107, http://ruccs.rutgers.edu/
roa.html]
190
 4.64.6
Conclusions
Exercises

EXERCISES

1 Wargamay
Consider the stress pattern of the Australian language Wargamay (Dixon 1981):
(i) bda dog
(ii)
a
ara dilly bag
(iii)
C"awlu freshwater jewsh
(iv) "ur
aymDri Niagara Vale-from
a. State the generalizations governing the position of primary and sec-
ondary stress.
b. What ranking of metrical constraints accounts for this pattern? Sup-
port your analysis by a tableau of Zurway-mri.

2 Manam
Manam stress (Lichtenberk 1983) is exemplied by the set of syllabied words
below:
(i) p.tu stone
(ii) si..ba bush
(iii) ta.np.wa chief
(iv) m.la.b ying fox
(v) m.be.i sacred ute
(vi) ta.np.wa.t.na real chief
a. State the generalizations governing the position of primary and sec-
ondary stress.
b. What ranking of metrical constraints accounts for this pattern? Sup-
port your analysis by tableaux of si.f.ba, m.la.bf, m.be.xi, and
ta.np.wa.t.na.

3 Murinbata
The stress pattern of Murinbata (Street and Mollinjin 1981) is exemplied below:
(i) b march y
(ii) mme I/he/she/ said/did to her
(iii) lmal shoulder
(iv) wlmoma blue-tongue lizard
(v) phJrwJrKn (season just before the dry when grass dries, seeds
fall, etc.)
Street and Mollinjin claim that all non-nal stresses are primary, whereas nal
stresses are secondary. For the purposes of this problem, you can ignore distinc-
tions among stresses.
191
Metrical structure and parallelism

a. State the generalizations governing the positions of stresses.

b. What ranking of metrical constraints accounts for this pattern? Sup-
port your analysis by a tableau of phcrcwcrbydn.
c. Discuss the consequences of this analysis for the interpretation of
Ft-Bin. Hint: is the degenerate foot the only type of non-binary foot?
d. Discuss any differences in foot-alignment between Murinbata and
Pintupi (section 4.4.1), taking into account the similarities between
the stress patterns of these languages.

4 Warlpiri
In bare stems, Warlpiri (Nash 1980) has a stress pattern typical of Australian
languages:

(i) wti man

(ii) wtiya tree
(iii) mnaka spinifex plain

The special interest of the pattern resides in its sensitivity to morphological

structure.

(iv) wti-ka man-loc

(v) wti-k-u man-loc-erg
(vi) ypa-u-u person-for example-erg
(vii) wtiy-a tree-loc
(viii) wtiya--u tree-loc-erg
(ix) ypaa-u fathers mother-elat
(x) mnaka-a spinifex-loc
(xi) mnaka--u spinifex-loc-erg

a. State the generalizations governing the stress pattern of bare stems

and sufxed words.
b. What constraint ranking accounts for the pattern of bare stems? Sup-
port your analysis by tableaux of wtiya and mnafkea.
c. Now consider sufxed stems, which require an extension of the basic
analysis. State a constraint aligning morphemes with feet. Integrate
this in the current ranking.
d. Support your analysis by tableaux of wtiy-ea, wtiya-e-eu, and
ypa-efu-eu.
e. Nash (1980) provides no examples of the type trisyllabic stem
plus trisyllabic afx. What would your analysis predict for such
examples? How does this issue bear on the ranking of the alignment
constraint?
192
 4.64.6
Conclusions
Exercises

5 A factorial typology of stress systems

a. Construct a factorial typology of the following ve metrical con-

straints: Parse-Syl, Leftmost, Rightmost, All-Ft-L, All-Ft-R.
Assume that Ft-Bin and RhType=T are undominated in all rankings.
Base the typology on four-syllable and ve-syllable words composed
of light syllables. Identify any rankings that produce identical stress
systems.
b. How many different stress systems does your typology predict?
Characterize each of the predicted systems in prose, making sure that
your description is sufciently general, as well as precise enough to
capture all distinctions.
c. Instantiate the factorial typology with actual stress systems (dis-
cussed in this chapter, or introduced in preceding exercises). Do you
nd any gaps in the typology? Try to ll these by consulting metrical
literature referred to in Suggestions for further reading.

193
 5
Correspondence in reduplication

5.1 Introduction
A central idea of OT is that the optimal output form arises from competition
of markedness constraints and faithfulness constraints. Markedness constraints
require that output forms match certain segmental or prosodic targets. Faithful-
ness constraints require that outputs be identical to their lexical inputs, each
militating against some type of change segment deletion, segment insertion,
or featural changes. Both constraint types are inherently in conict: markedness
constraints trigger changes, while faithfulness constraints, by their very nature,
oppose changes. Moreover, faithfulness constraints state their requirements about
inputoutput relations in terms of correspondence.

This central idea will be extended in this chapter to a phenomenon that has been
a hotly debated topic in pre-OT phonology (e.g. Wilbur 1973, Marantz 1982,
Broselow and McCarthy 1983, McCarthy and Prince 1986, Shaw 1987, Uhrbach
1987, Steriade 1988a), and continues to be a focus of OT research: reduplication.
From a purely morphological point of view, reduplication is simply a kind of
afxation, both in its morpho-syntactic contribution (it forms morphological
categories, such as plural), and in its linear position with respect to the stem
(preceding it, as a prex, or following it, as a sufx). But from a phonological
viewpoint, the special property of reduplication is that the reduplicative afx is
not fully specied for segmental content. Its segmental content is copied from the
stem that undergoes reduplication. Reduplication is therefore by its very nature a
phenomenon involving phonological identity between the reduplicant and the
base to which it adjoins. (Both terms are used informally here, and will be
dened in section 5.2.)
Segmental and prosodic identity of the reduplicant and the base is obvious in
the case of total reduplication, which involves copying of a complete word. The
following examples of total reduplication in Indonesian (Cohn 1989) show that it
is impossible to tell apart the reduplicant from the base:
194
 5.1 Introduction

(1) Indonesian plural reduplication (total reduplication, copies entire

word)
a. wanita wanita-wanita women
b. ma9arakat ma9arakat-ma9arakat societies

But other languages have reduplication processes that copy only part of the
segments of the base. This phenomenon is known as partial reduplication. For
example, reduplication in the Wakashan language Nootka (Stonham 1990) copies
the rst sequence of consonant plus vowel (CV or CVV) of the base, and prexes
this material to the base:

(2) Nootka reduplication (copies initial CV(V) sequence, equalling open

syllable)
a. 0ims-i# 0i-0ims-i# hunting bear
b. was-0i$ wa-was-0i$ naming where . . .

Observe that the reduplicant prexes [ 0i-] in (2a) and [wa-] in (2b), both open
syllables, do not match similar open syllables in the base, but copy segments that
stand in closed syllables: [0im] and [was] in the base.
The size of the reduplicant in partial reduplication varies between languages.
This is clear from a comparison of the Nootka pattern (2) with reduplication in
the Australian language Diyari (Austin 1981, Poser 1989, McCarthy and Prince
1994a):

(3) Diyari reduplication (copies initial foot, minus coda of second syllable)
a. wi.2a wi.2a -wi.2a woman
b. ku.ku.a ku.ku -ku.ku.ato jump
c. tjil.par.ku tj il.pa -t jil.par.ku bird species

Here the reduplicant is a disyllabic unit, specically a binary foot, although sur-
prisingly it is not simply a copy of the rst two syllables of the base. The restric-
tion holds that the second syllable of the reduplicant is open, regardless of
whether the second syllable in the base is open or closed. This is shown in (3c),
where the reduplicant [ tj il.pa -] fails to match the rst two syllables of the base,
[tjil.par], the second of which is closed.

5.1.1 Shape invariance

The reduplication patterns that we have seen thus far illustrate two observa-
tions that will become of central importance in this chapter. The rst observation
is that the reduplicant tends to have an invariant prosodic shape that has no one-
to-one relation with a prosodic unit in the base. In both Nootka and Diyari, the
195
Correspondence in reduplication

reduplicants can be described in terms of prosodic units (syllable and foot). How-
ever, in neither case does the reduplicant exactly match an analogous prosodic
unit in the base.1

(4) Reduplication does not copy a prosodic constituent

a. Nootka (): wa-was-0i$ (not *was-was-0i$)
b. Diyari (Ft): tj il.pa -t jil.par.ku(not * tj il.pa -t jil.par.ku)

For this reason reduplication cannot be simply constituent copying, but it

involves a prosodic shape invariant in the reduplicant, also known as a reduplicat-
ive template. This observation of shape invariance was rst stated by Moravcsik
(1978), and it has become the basis of the template-and-association theory of
reduplication (e.g. Marantz 1982, Clements 1985a, McCarthy and Prince 1986).
The OT approach to reduplication is, to a certain extent, a continuation of
templatic theory in the sense that constraints may govern the prosodic size of
a reduplicant. But it diverges from templatic theory by decomposing templatic
requirements into independent prosodic constraints and alignment constraints, as
we will see in section 5.3.

5.1.2 Unmarkedness of the reduplicant

The rst observation naturally leads us to a second one, which can also be
illustrated by reduplications in Nootka and Diyari. Reduplicants tend to have
unmarked phonological structures, as compared to the phonotactic options gen-
erally allowed in the language. In Nootka, unmarked prosodic structure emerges
in the reduplicant prex, an open syllable (2). This is interesting since Nootka
does not generally ban codas. Nevertheless we know that open syllables are
universally less marked than closed syllables, from various kinds of evidence
discussed in chapter 3, in particular cross-linguistic patterns of epenthesis. (The
unmarked status of open syllables is captured in OT by the constraint No-Coda.)
The reduplicant prex in Diyari (3) matches a disyllabic foot, another unmarked
prosodic unit, as we know from stress patterns and word minima, as discussed in
chapter 4. (The constraint Ft-Bin captures the unmarked status.) Incidentally, the

1
One reduplication pattern is known which involves copying of an entire prosodic unit. In the
Australian language Yidi (Dixon 1977, McCarthy and Prince 1986) plurals arise by prexing
an exact copy of the rst foot of the base, that is, including its syllabication:
(i) Yidi plural reduplication (copies initial two syllables, or foot)
a. mulari mula-mula.ri initiated men
b. gindalba gindal -gindal.ba lizard species ( plural)
Note that the reduplicant in (a) ends in /a/, matching this segments position at the end of a
foot in the base. And in (b), the reduplicant ends in /l/, matching this segments position in a
closed syllable in the base.

196
 5.1 Introduction

Diyari reduplicant ends in an open syllable, again an unmarked prosodic unit,

even though closed syllables occur in Diyari stems.2
The observation that reduplicants tend to have unmarked phonological struc-
tures (that is, as compared to the general phonotactic possibilities of the language)
was stated explicitly by Steriade (1988a: 80):

Prosodic templates frequently eliminate certain unmarked options from

their syllabic structures. Although the language may in general allow com-
plex onsets or consonantal nuclei, the template might specically revoke
these options.
We observe the following range of syllabic simplication in prosodic tem-
plates: onset simplication (Tagalog ta-trabaho, bo-bloaut from McCarthy
and Prince 1986: 16); coda simplication (Sanskrit intensive kan-i-krand
[ . . . ]); coda drop (cf. the French hypocoristics zabe, mini from (i)zabel,
(do)minik); elimination of consonantal nuclei (Sanskrit perfect va-v,ma
[ . . . ]).

This observation will become a focal point of attention in this chapter. The ques-
tion is, what explains it? Steriade, working in a derivational framework, proposed
that marked syllable properties are literally cancelled from the reduplicant. This
proposal assumes that a reduplicant, in a specic stage of its derivation, is a full
copy of its base, including its syllable structure. For example, a derivation of the
Nootka examples (left undiscussed by Steriade) would involve the following two
steps:

(5) a. Copying of the base

(total copy, including
syllabication) was-0i$ - was-0i$

b. Elimination of marked
structure disallowed by
the template (=syllable, wa - was-0i$
with codas disallowed)
Steriades proposal maintained the formal distinction between a morphological
operation (afxation, copying) and its construction-specic phonology: the can-
cellation of marked properties. This strict division of labour between morphology
and phonology was indeed assumed by most generative linguists in the 1970s and
1980s (although there were some exceptions; see for example Anderson 1975).
But a fundamental question remains unanswered by Steriades full-copy-plus-
cancellation theory: why is it that cancellations of prosodically marked structure

2
The interesting question of what causes the asymmetry between the rst and second syllable
in the reduplicant in (4b) (in terms of markedness) will be addressed in section 5.4.

197
Correspondence in reduplication

fail to affect the base? We might state the unmarkedness as a template-specic

fact, but that only raises another question: what makes reduplication templates so
vulnerable to simplication? A priori, the universal markedness principles treat
template and base alike, since both are built of the same stuff.
As McCarthy and Prince note, the key to the problem resides in the very fact
that reduplication is, by nature, a phenomenon which is dependent in its identity
upon another morpheme. Since the reduplicant is not burdened with lexical con-
trasts, its phonological form naturally drifts towards the unmarked. There is a
straightforward expression of this idea in OT, which involves ranking of marked-
ness constraints between input faithfulness constraints and copying constraints
(reduplicative identity constraints). In the base, the effects of markedness
remain hidden behind other factors, in particular the preservation of contrast
(faithfulness). But unmarked structures (open syllables) manifest themselves in
contexts where input faithfulness has no grip, such as the reduplicant, which has
no specication of input segments. Any language, given the chance, will develop
unmarked structures in contexts where the inuence of faithfulness is absent. This
emergence of the unmarked (McCarthy and Prince 1994a) provides a major argu-
ment for the OT assumption that languages are subject to universal markedness
constraints (see section 5.3).

5.1.3 Identity of reduplicant and base

A third important observation on reduplication should be added to those discussed
in the preceding sections: reduplicants tend to preserve phonological identity with
the base. The pressure towards identity may even happen at the cost of transpar-
ency in the application of the phonology, giving rise to so-called overapplications
or underapplications of phonology. Overapplication refers to a situation in which
the reduplicant undergoes some change, even though it fails to satisfy the struc-
tural condition of this change. Underapplication is the reverse situation, in which
a reduplicant fails to undergo some change, even though it meets the context in
which this change normally takes place. Both effects simultaneously occur in
Javanese, as the examples show (Dudas 1976, Kiparsky 1986, Steriade 1988a):

(6) underlying surface reduplicated

a.i /donga/ do.ng do.ng-do.ng prayer
a.ii /donga-ne/ do.nga.ne do.nga -do.nga.ne (*do.ng-do.nga.ne)
b.i /abur/ a.br a.b.r-a.br (*a.bu.r-a.br)
b.ii /abur-e/ a.bu.re a.bu.r-a.bu.re ight

In (6a.ii), a process of word-nal a-rounding /a/ [] underapplies in the

reduplicant, even though its nal vowel matches the context of application. In
(6b.i), another process of closed syllable laxing /u/ [] overapplies to the
198
 5.1 Introduction

reduplicant, even though it fails to meet the structural requirement of a closed

syllable. In both cases, the reduplicant maintains segmental identity with the base.
The underapplication and overapplication of phonology in reduplication
was rst observed by Wilbur (1973: 58), and stated in the form of the Identity
Constraint:

(7) The Identity Constraint

There is a tendency to preserve the identity of Ro [reduplicant] and
Rr [base] in reduplicated forms.

This is not a constraint in the sense of OT: observe the phrasing tendency.
However, it provides the basis of a constraint-based theory of overapplication and
underapplication effects in reduplication, which we will learn about in section 5.4
of this chapter. The basic idea has already been hinted at in the preceding discus-
sion: maximal identity of reduplicant and base increases the transparency of mor-
phological reduplication, but at the cost of a loss of transparency in the phonology
(Anderson 1975). Grammars dene language-specic ways to reconcil both con-
icting forces.
We may summarize the discussion so far by the following three observations:

(8) Three cross-linguistic tendencies in reduplication

a. Shape invariance
Reduplication tends to be dened in prosodic units independent
of the base. (That is, reduplicants do not copy prosodic constitu-
ency from the base.)
b. Unmarkedness
Reduplicants tend to have phonologically unmarked structures
vis--vis the phonotactics of the language. (That is, reduplicants
display a subset of the phonotactic options generally allowed in
the language.)
c. Identity
Reduplication tends to preserve full phonological identity with
the base. (That is, reduplication may involve overapplication and
underapplication of phonology to preserve such identity.)3

Is there an inconsistency between shape invariance and prosodic unmarked-

ness? If a reduplicant template equals a marked prosodic unit (for example, a
heavy syllable), could this not lead to increased markedness in the reduplicant,
as compared to the base? There is no inconsistency since, as (8b) states, the
3
As far as is currently known, the effects of overapplication and underapplication are limited to
segmental phonology, with a remarkable lack of copying of aspects of syllable structure or
foot structure.

199
Correspondence in reduplication

proper basis for comparison is not the actual base, but rather the overall prosodic
possibilities of the language.
For example, reduplication in Agta (Healey 1960, Marantz 1982) copies
enough segments from the initial portion of the base to form a heavy syllable:

(9) Agta reduplication (copies enough material to form a heavy syllable)

a. ba.ri body bar-ba.ri-k kid-in my whole body
b. tak.ki leg tak-tak.ki legs

Form (9a) may be said to involve literally an increase of markedness in the

reduplicant, as compared to the base. While the base [ba.ri] has only open sylla-
bles, the reduplicant is a closed syllable [bar-], combining in it segments that are
copied from the bases initial open syllable, plus the onset of its second syllable.
However, Agta is a language which happens to allow for closed syllables. (In
terms of ranking, it has Max-IO
No-Coda.) Therefore, even if the reduplicant
is prosodically marked as compared to the particular base with which it com-
bines, this markedness still falls within the margins of the overall phonology of
the language.
In reduplicative morphology, as in other kinds of morphology, conicting
forces are at work; naturally, different languages select different types of recon-
ciliations of these conicts. The very goal of optimality-based approaches to
reduplication is to factor out the complex patterns of reduplication into interac-
tions of general principles of structural well-formedness and correspondence, that
is, to constraint interactions.

5.1.4 The Basic Model

In this chapter we will use the three observations in (8) as our point of depart-
ure in a discussion of the Correspondence Theory of reduplication (McCarthy
and Prince 1994a, b, 1995a, forthcoming). This theory claims that reduplication
patterns arise by interactions of three constraint types: (a) well-formedness con-
straints, encoding markedness principles; (b) faithfulness constraints, requiring
lexical forms and surface forms to be identical; and (c) basereduplicant-identity
constraints, requiring identity between the reduplicant and its base. By reranking
these three types of constraints, Correspondence Theory aims to explain the broad
typological differences and similarities among patterns of reduplication in the
worlds languages, as well as the specic patterns of individual languages. The
following issue will be raised time and again in our discussion of reduplication
patterns throughout this chapter:

What interaction of well-formedness, faithfulness, and basereduplicant

identity underlies the pattern of reduplication in some language L?

200
 5.2 Reduplicative identity: the constraints

We will address this issue in the context of what McCarthy and Prince (1995a,
forthcoming) call the Basic Model. This model, as depicted in (10), has an input
and an output level. The input of reduplication consists of a segmentally empty
reduplicative afx, which is abbreviated as Afred or red, plus the stem to which
the afx adjoins. Input faithfulness constraints require that the stems input
specications be preserved in the output the base of the basereduplicant
combination. Basereduplicant identity constraints require that both parts of this
output basereduplicant combination be identical in some respect.

(10) Basic Model

Input: /Afred + Stem /

IO-Faithfulness
Output: R B
BR-Identity

Not shown in this model are the two remaining constraint types which come into
play in reduplication. Well-formedness constraints require that the output (base
reduplicant) meet certain unmarked structures. Alignment constraints require that
base and reduplicant be glued together along the edges of specic prosodic
constituents.

5.1.5 Organization of this chapter

This chapter is organized as follows. In section 5.2, we will introduce the notions
and constraints which are central in the Correspondence Theory of reduplication.
Each constraint will be presented in the context of a simple analysis of some
reduplication pattern. We will also try to explain the observation that reduplicants
tend to be unmarked (as compared to the phonotactic options in the language):
the emergence of the unmarked in the reduplicant. Section 5.3 will develop the
notion of template in OT, explaining unmarked prosodic shapes by constraint
interactions. Section 5.4 discusses alignment in inxing reduplication, building
the rst argument for parallelism in the prosodymorphology interface. Results
of sections 5.3 4 will be summarized in section 5.5. Section 5.6 addresses over-
application and underapplication in reduplicative morphology, focussing on
typology and developing yet another argument for parallelism. Finally, section 5.7
contains an overview of Correspondence Theory as developed in this chapter.

5.2 Reduplicative identity: the constraints

5.2.1 The notions reduplicant, base, and correspondence

Before we actually discuss the constraints of reduplicative identity, we must rst
dene the notions of reduplicant and base in a more precise way than we have
201
Correspondence in reduplication

done so far. First and foremost, both notions refer to strings of output segments,
rather than to input strings. (In fact, it will be essential that there is no such thing
as the phonological input representation of a reduplicative afx.) Reference to the
output shapes of reduplicant and base is at the heart of the Correspondence
Theory of reduplication.
The denitions of reduplicant and base that are given by McCarthy and
Prince (1994b) are paraphrased in (11):

(11) The reduplicant is the string of segments that is the phonological

realization of some reduplicative morpheme red, which is phonolog-
ically empty.
The base is the output string of segments to which the reduplicant
is attached, more specically:
for reduplicative prexes, it is the following string of segments
for reduplicative sufxes, the preceding string of segments

Note that the reduplicant need not be identical to a unique morpheme, for ex-
ample a root. That reduplication may copy more than just a single morpheme is
shown by examples from the Bantu language Kinande (Mutaka and Hyman 1990,
McCarthy and Prince 1995b):

(12) Kinande noun reduplication (copies nal two syllables of base)

a. ku-gulu- gulu real leg
b. m-twe-m-twe real head
c. mw-na-mw-na real child

This is a case of sufxing reduplication, with a disyllabic reduplicant, which has

been underlined. Following the denition in (11), the base contains all material
preceding the reduplicant. Observe that in all examples in (12) the base contains
both prexal and root material. More specically, prexes are ku- and mu-/mw-,
roots gulu-, twe-, and ana-. Observe that the disyllabic reduplicant may consist
entirely of segments copied from the root, as in (12a). This occurs when the
root provides enough segments to ll the disyllabic reduplicant. But examples
(12bc) show that the reduplicant copies additional segments from the prex, if
the roots segmental material does not by itself sufce to ll out both syllables
of the reduplicant. In sum, we nd that the base may span a string of segments
that exceeds a single morpheme.
The one-to-one relationship between morphemes and reduplicants may also be
broken in the opposite direction, when a part of the stem serves as the reduplic-
ants base. In some languages the reduplicant afxes to a base-string that is
smaller than the stem. This phenomenon is known as internal reduplication. In
the examples below, note that the reduplicant adjoins as a prex to the bimoraic
202
 5.2 Reduplicative identity: the constraints

stress foot in the Polynesian language Samoan (Marsack 1962, Broselow and
McCarthy 1983).

(13) Samoan plural reduplication (copies initial CV of stress foot before

the foot)
a.i nfo no-nfo sit
a.ii ta ta-ta strike
b.i a.lfa a-lo-lfa love
b.ii fa.nu fa-na-nu be born, give birth

In examples (13a), reduplication may easily give the appearance of a prexation

to the stem, since the stem happens to be the size of a bimoraic foot. But exam-
ples (13b) show that this naive hypothesis is incorrect, since here the reduplicant
shows up in the middle of a stem. According to the denition in (11), the base of
an internal reduplication is the string of segments following the reduplicative
prex. (The analysis of the interesting phenomenon of internal reduplication
will be deferred to section 5.4.)

With the preliminary denitions of reduplicant and base, we can now move on
to another notion that requires denition: correspondence in reduplication. In
previous chapters we have seen correspondence relationships between input and
output elements, to which we attributed the term faithfulness. In fact, this notion
of correspondence is all we need to capture the identity relation between the base
and the reduplicant, if only we generalize it a bit. We will see that faithfulness
(or input-to-output correspondence) and reduplicative identity (or base-to-
reduplicant correspondence) are just two aspects of a general relation dened
between elements in two representations. Pairs of representations can be either an
output form and an input form, or a reduplicant and its base.
McCarthy and Prince (1995a: 262) dene this generalized notion of corre-
spondence as follows:4

(14) Correspondence. Given two strings S1 and S2, correspondence is a

relation from the elements of S1 to those of S2.
Correspondents. Elements S1 and S2 are referred to as cor-
respondents of one another when .

From here on, we must be somewhat more specic about what it means for
two elements to be correspondents. Following McCarthy and Prince (1995a) we
make the simplifying assumption that elements standing in correspondence are

4
A precursor of the notion correspondence is proposed in Bybee (1988).

203
Correspondence in reduplication

segments, although it would be a straightforward matter to generalize the notion

to other elements, such as moras, syllables, or even distinctive features. Moreover,
we will assume that correspondence relations are supplied by Gen; accordingly,
any pair of segments (when a member of the pairs of strings dened above) can
be correspondents. Like all other aspects of representations, correspondence rela-
tions are evaluated by Eval using ranked constraints. Crucially we do not make
the assumption that corresponding segments must be segmentally identical. This
will allow us to account for featural changes between input and output separately
from insertion and deletion.
Taking the Samoan example /red + nofo/, we assume that Gen produces output
can-didates supplied with all logically possible correspondence relations. Con-
sider a subset of output candidates, in which subscript indexes indicate corre-
spondence relations:

(15) /red + nofo/ Comments

a. n1 o2 n1 2 f o (f, o in base lack correspondents, others
identical)
b. n1 o2 n 2 f1 o (n, o in base lack correspondents, n1, f1
distinct)
c. n1 o2 f3 o4 n1 2 f3 o4 (all segments have identical
correspondents in the same linear order:
perfect correspondence)
d. f1 o2 f3 o4 n1 2 f3 o4 (all segments have correspondents in the
same linear order, although f1, n1 are
distinct)
e. o4 f3 o2 n1 n1 2 f3 o4 (all segments have identical
correspondents but in the reverse linear
order)
f. n1 o2 f o n1 2 f o (f, o -twice- lack correspondents, others
identical)

These candidates (and many others) are then submitted to Eval for evaluation by
violable correspondence constraints, interacting with faithfulness constraints and
well-formedness constraints. The forms and functions of basereduplicant corres-
pondence constraints are a topic to which we now turn.
We will extend the set of faithfulness constraints known from previous chap-
ters: Max-IO, Dep-IO, Ident-IO, Linearity-IO, and Anchoring-IO, to the
basereduplicant relation. The rst step is a generalization of the IO-faithfulness
constraints to other kinds of strings, including base and reduplicant. According to
denition (14), correspondence is a relation between elements in two strings: S1
204
 5.2 Reduplicative identity: the constraints

and S2. Applying this generalization to the rst three IO-faithfulness constraints
mentioned above, we arrive at the following set:

(16) Generalized correspondence constraints

a. Maximality
Every element of S1 has a correspondent in S2.
b. Dependence
Every element of S2 has a correspondent in S1.
c. Identity[F]
Let be a segment in S1 and be a correspondent of in S2.
If is [F], then is [F].

To arrive at actual basereduplicant identity constraints, we must substitute S1 and

S2 by the specic strings relevant to reduplication, reduplicant (R) and base (B).
This step will be taken in the next section for Maximality, Dependence, and
Identity.

5.2.2 MAX, DEP, and IDENT

5.2.2.1 Violation of MAX-BR in Nootka

In input faithfulness, Max-IO is the anti-deletion constraint. In reduplication the
notion of deletion is not accurate, since no segment in a reduplicant has an input
correspondent. The reduplicative instantiation of Maximality, Max-BR, requires
that every element of the base must reoccur in the reduplicant. This is the total
reduplication constraint:

(17) Max-BR
Every element of B has a correspondent in R.
(No partial reduplication)

If Max-BR is undominated, total reduplication will indeed result, as in Indone-

sian (1). But if it is dominated by a well-formedness constraint militating against
any kind of elements that might result from total reduplication, then partial redu-
plica-tion will be the outcome, as in Nootka (2), Diyari (3), and Samoan (13).
Let us take a closer look at the Nootka example (2a), discussed earlier as a case
of unmarked syllable structure in the reduplicant. Recall that the reduplicant in
[0i-0imsi#] is an open syllable, omitting the coda consonant m (and remaining
segments of the base). If m were copied, it would syllabify as a coda in the
reduplicant, *[ 0im.0imsi#]. Clearly, Nootka avoids this marked syllable struc-
ture in the reduplicant. But then why does it allow the same marked structures in
the base? McCarthy and Prince (1994a) show that the answer is surprisingly
simple. The rst step in the argument is that the total avoidance of codas in the
205
Correspondence in reduplication

reduplicant must be due to the well-formedness constraint No-Coda, which out-

ranks the reduplicative identity constraint Max-BR:

(18) Partial reduplication due to avoidance of codas

No-Coda
Max-BR

This ranking is illustrated by tableau (19). The total reduplication candidate (19d)
is ruled out by No-Coda, since a better candidate (19a) is available which has
fewer violations of this constraint:

(19) Input: /Red-0ims-i#/ No-Coda Max-BR

a. + 0i-0im.si# ** ****
b. 0im-0im.si# ** *! ***
c. 0im.si-0im.si# ** *! *
d. 0im.si#-0im.si# ** *!*

Violations of No-Coda are marked independently for every coda in every candi-
date, that is, regardless of whether it stands in the reduplicant or the base. Observe
that the base has two closed syllables, hence incurs two violations of No-Coda by
itself. (We will nd out shortly why these violations are unavoidable.) But No-
Coda disallows further violations in the reduplicant, sharply limiting its size to
an open syllable at the expense of violations of Max-BR. In the optimal candidate
(19a) Max-BR is violated four times, once for each base segment missing from
the reduplicant.
Of course, this cannot be the complete story, because the consonants [m] and
[#] that are banned from the reduplicant freely occur as codas in the base. Base
segments are different from reduplicant segments in being subject to input faith-
fulness. Their deletion is penalized by Max-IO. Hence, No-Coda must, in its
turn, be dominated by Max-IO:

(20) Emergence of the unmarked in Nootka reduplication

Schema Faithfulness
Well-formedness
Reduplicative
identity
Instantiation Max-IO
No-Coda
Max-BR

The sandwiching of well-formedness between both kinds of correspondence

constraints is what makes the base more faithful to input segments than the
reduplicant.
To prove our point, let us inspect tableau (21). This includes all four candidates
of the previous tableau, plus one critical new candidate (21e), representing a
206
 5.2 Reduplicative identity: the constraints

reduction of the base back to a monosyllabic format, matching the prosodic shape
of the reduplicant:

(21) Input: /Red-0ims-i#/ Max-IO No-Coda Max-BR

a. + 0i-0im.si# ** ****
b. 0im-0im.si# ** *! ***
c. 0im.si-0im.si# ** *! *
d. 0im.si#-0im.si# ** *!*
e. 0i-0i *!***

Candidate (21e) seems to have much in its favour. It is composed of two open
syllables (satisfying No-Coda). Both of its base segments have correspondents in
the reduplicant (satisfying Max-BR). However, it fatally violates faithfulness by
omitting four segments in the base which are present in the input incurring four
marks for Max-IO. Crucially, Max-IO does not militate against the omission of
input segments in the reduplicant. This is because the reduplicant has no input
segments of its own, to which it might potentially be faithful. This is the essence
of reduplication: a reduplicative afx has no input segments, hence escapes from
the gravity forces of faithfulness.
This argument shows that unmarked syllable structure (open syllables) sud-
denly and unexpectedly come to the surface in contexts where input-to-output
faithfulness has no grip. This emergence of the unmarked is a major argument
for OTs assumption that languages are all subject to the same set of universal
markedness constraints. Effects of markedness may remain hidden behind other
factors, in particular preservation of contrast (faithfulness). But any language,
given the chance, will develop unmarked structures in contexts where inuence
of contrast-preserving factors is absent. In reduplication, such cases are easily
dismissed as imperfect copying. This qualication is correct from the viewpoint
that what counts towards perfection is faithfulness. But it seems equally correct
to say that reduplication of the Nootka type involves a kind of unmarked perfec-
tion that is otherwise lacking from the surface phonology of the language.
This analysis exemplies the general line of reasoning to be pursued in the rest
of this chapter. For every pattern of reduplication, we will attempt to reduce it to
interacting constraints of well-formedness, faithfulness, and reduplicative identity.

5.2.2.2 Violation of DEP-BR in Makassarese

Next we turn to the logical counterpart of Max-BR, the constraint Dep-BR.
Recall from chapter 2 that Dep-IO is the anti-epenthesis constraint, militating
207
Correspondence in reduplication

against any segments in the output which have no correspondents in the input.
Analogously, Dep-BR requires that segments in the reduplicant must have corres-
pondents in the base:

(22) Dep-BR
Every element of R has a correspondent in B.

If undominated, the reduplicant is composed solely of segments that occur in the

base. In fact Dep-BR was satised in every example of reduplication that we have
seen thus far. What could a violation look like?
An example of a violation of Dep-BR occurs in Makassarese (Aronoff et al.
1987, McCarthy and Prince 1994a). Reduplication copies the rst two syllables
of a (trisyllabic or longer) base, but also adds an epenthetic consonant // at the
end of the reduplicant:

(23) Makassarese reduplication (copies rst two syllables, adding epen-

thetic //)
a. kalu|-kalu.arak ant
b. mana|-mana.ra sort of tower
c. bala|-bala.o toy rat
d. bai|-bai.ne many women

Observe that epenthesis is blocked at the end of the base, pointing to a strict
enforcement of Dep-IO. We will not present the complete analysis by McCarthy
and Prince, since it is fairly complex, and we refer the interested reader to the
original work.

5.2.2.3 Violation of IDENT-BR in Akan

Finally, the reduplicative instantiation of Identity is Ident-BR. This constraint
requires featural identity between correspondents in base and reduplicant.

(24) Ident-BR[F]
Let be a segment in B, and be a correspondent of in R.
If is [F], then is [F].

Thus far we have only seen cases where this constraint was actually satised.
However, it would not be a true OT constraint if it were unviolable. Violations of
Ident-BR are due to differences in feature values of corresponding segments
in reduplicant and base. As we expect, featural deviations in the reduplicant (as
compared to the base) tend to reduce the segmental markedness in the reduplic-
ant. For example, a feature contrast occurring in the base may be disallowed in
the reduplicant. This gives the appearance of neutralization of segmental contrasts
in the reduplicant, as compared to the base.
208
 5.2 Reduplicative identity: the constraints

An example of a reduction of featural markedness in the reduplicant, discussed

by McCarthy and Prince (1995a), is reduplication in Akan (Schachter and From-
kin 1968):

(25) Reduced segmental markedness in Akan reduplication

a. si-si stand f. bu-bu() bend
b. f-f vomit g. s-s() carry on the head
c. si-se say h. su-so seize
d. s-s resemble i. s-s light
e. s-sa cure

The vowel in the reduplicative CV-prex is a copy of the base vowel in every
respect, except that it is always high. Raising of mid vowels has a striking parallel
in the cross-linguistically common phenomenon of vowel reduction occurring in
unstressed syllables. Reduction amounts to the loss of featural contrasts in met-
rically weak positions, hence a replacement of marked segments by unmarked
segments.5
McCarthy and Prince (1995a) analyse the loss of vowel height contrasts in the
Akan reduplicant as another emergence of the unmarked. Crucially vowels in the
base are exempt from the neutralization of height contrasts taking place in redu-
plicants. Hence the Ident-IO constraint for [high] must dominate the context-free
markedness constraint *[high], which bans [high] vowels:

(26) Faithfulness
Well-formedness

Ident-IO(high)
*[high]

This is a straightforward example of a preservation of contrast due to the domin-

ation of a faithfulness constraint over a markedness (well-formedness) constraint,
of a kind that we have studied in detail in chapter 1. The ranking is illustrated in
the following tableau:

(27) Input: /so/ Ident-IO(high) *[high]

a. + so *
b. su *!

Next observe that a reduction of the vowel height contrast does occur in the
reduplicant. This is another emergence of the unmarked: the reduplicant contains
5
An analysis by Marantz (1982) was based on a prespecication of the feature [ +high] in the
reduplicant, overriding the copied melody. McCarthy and Prince (1995a) criticize this analysis
for prespecifying an unmarked feature value, [+high]. Underspecication Theory (Kiparsky
1985, Archangeli 1988) predicts that only marked values can be lexically specied.

209
Correspondence in reduplication

segments of lesser markedness than the corresponding base segments. This effect
arises due to the lack of an input correspondent for a vowel in a reduplicant,
which is thus exempt from faithfulness requirements. For unmarked vowels to
arise in the reduplicant, the markedness constraint *[high] must actually dom-
inate the reduplicative identity constraint Ident-BR(high), requiring identity
between the base and reduplicant for the feature [high]. This analysis involves the
type of ranking that we saw earlier for Nootka.

(28) Emergence of the unmarked in Akan reduplication

Schema Faithfulness
Well-formedness
Reduplicative
identity
Instantiation Ident-IO(high)
*[high]
Ident-
BR(high)

We have now arrived at the rst part of an explanation for the different behaviours
of the reduplicant and base with respect to marked segments.
The second part is to prove that no reduction of markedness can apply to the
base without also applying to the reduplicant. For this purpose, let us turn to
tableau (29):

(29) Input: /red-so/ Ident-IO(high) *[high] Ident-BR(high)

a. + su-so * *
b. so-so **!
c. su-su *!
d. so-su *! * *

Optimal (29a) preserves the input value [high] in the stem vowel, due to undom-
inated Ident-IO(high), thus violating *[high]. Minimal violation of *[high]
blocks a second occurrence of [high] in the reduplicant, even though this would
yield complete identity between the reduplicant and the base (29b).
Next consider candidate (29d), whose reduplicant vowel copies the inputs
value for [high], while its base vowel is unfaithful to this value. This pattern of
vowel height is precisely the reverse of that in the optimal candidate, where the
unmarked value emerges in the reduplicant. Could this reversed pattern of reduc-
tion in base and reduplicant arise in any language? The correct prediction is that
it cannot, since no constraint ranking can derive it. Candidate (29c) is identical
in its pattern of violation marks to candidate (29d), except for *[high] and
Ident-BR(high), which evaluate (29d) as worse than (29c). Therefore (29d) cannot
be optimal under any possible ranking of these constraints, since in the end it will
210
 5.2 Reduplicative identity: the constraints

always fall victim to *[high] or Ident-BR(high), regardless of where these are

ranked, rendering it suboptimal to (29c). No reduction of markedness is possible
in the base without an accompanying reduction of markedness in the reduplicant,
which is a signicant result.

5.2.3 Why input faithfulness and reduplicative identity cannot be unied

So far we have seen reduplicative counterparts of the constraint triplet Max-IO,
Dep-IO, and Ident-IO. We have learned why the reduplicant can have a reduction
of markedness as compared to the base, and why things could not have been
reversed.
One major question has not yet been answered: if input-to-output faithfulness
and base-to-reduplicant identity are both due to correspondence constraints, then
would it not be possible to unify both types of constraints into a single general-
ized set? One glance at any reduplication example discussed above tells us that
this conclusion cannot hold.
Consider the case of Nootka discussed earlier. If No-Coda were to dominate
both Max-BR and Max-IO, then all words in the language, reduplicated and non-
reduplicated words alike, would be strictly open-syllabled. Of course we know
that such languages are attested (CV-languages, as have been discussed in chap-
ter 3). But Nootka happens not to be among these languages since it allows codas,
for example in bases of the reduplicative examples that we saw. It follows that
Nootka maintains dominance of faithfulness (Max-IO) over the well-formedness
constraint No-Coda. Suppose that we were to take the bold step of collapsing
both Max constraints into a single constraint, then the inevitable result for Nootka
would be disastrous. We would end up with either total reduplication (30b), or a
simplication of the base into an open syllable (31c), depending on ranking:

(30) Input: /red-0ims-i#/ Max-IO/BR No-Coda

a. 0i-0im.si# *!*** **
b. I 0im.si#-0im.si# ****
c. 0i-0i *!***

(31) Input: /red-0ims-i#/ No-Coda Max-IO/BR

a. 0i-0im.si# *!* ****

b. 0im.si#-0im.si# *!***
c. I 0i-0i ****

211
Correspondence in reduplication

In sum, we must distinguish input-to-output faithfulness and base-to-reduplicant

identity for empirical reasons: base and reduplicant may differ in their phonotac-
tic options, which implies that both types of constraints are ranked in different
positions in the hierarchy.6
In the next section we will introduce two more BR-correspondence constraints.

5.2.4 ANCHORING and CONTIGUITY

So far all reduplicative examples have been of the following general type: starting
at an edge of the base (the left edge in the case of a reduplicative prex, the right
edge in the case of a reduplicative sufx), a contiguous substring of base seg-
ments is copied to the reduplicant:

(32) Anchoring and contiguity satised by various reduplications

a. prexing Diyari: tj ilpa -t jilparku
b. sufxing Kinande: ku-gulu- gulu
c. inxing Samoan: a-lo-lfa

Anchoring and contiguous copying are highlighted in correspondence diagrams

in (33). In each diagram, reduplicant and base are represented parallelly, as in
IO-correspondence:
(33) a. tj i l p a - R b. -g u l u R c. l o- R

tj i l p a r ku B ku g u l u B l ofa B

In (33a), the leftmost segment in the reduplicative prex [tjilpa] corresponds with
the leftmost segment in the base [tjilparku]. The corresponding segments [tjilpa]
form a contiguous string of both the base and the reduplicant. In (33b), the
rightmost segment in the reduplicant [-gulu] corresponds with the rightmost base
segment, [kugulu]. The corresponding segments [gulu] form a contiguous string
of both the base and the reduplicant. In (33c), the leftmost segment in the redu-
plicant [lo-] corresponds with the leftmost segment of the base. (Recall that the
base is dened as the segment string to which the reduplicant is adjoined. For
reduplicative prexes, it is the following string of segments: here the foot [lofa]
to which the reduplicant is prexed.) The corresponding segments [lo] are a
contiguous string of both base and reduplicant.
In all three cases, correspondence holds between contiguous strings of seg-
ments which are anchored at an edge. However, this is not a logical necessity in
reduplication. As we will see, partial reduplication need not respect contiguity

6
A more comprehensive exploration of the interactions between both types of correspondence
constraints and well-formedness constraints will take place in section 5.6.

212
 5.2 Reduplicative identity: the constraints

and anchoring, which we will state as violable constraints in the following two
subsections.

5.2.4.1 Violation of ANCHORING in Sanskrit

The rst constraint, Anchoring, requires correspondence between two major
landmarks in the base and reduplicant: segments standing at edges. Depending on
whether the reduplicant is a prex or a sufx, this requirement must hold at the
left or right edge. (A prex-initial segment must correspond with the initial seg-
ment of the base, while a sufx-nal segment must correspond with the nal
segment of the base.) This requirement is stated in a general way in (34):7

(34) Anchoring-BR
Correspondence preserves alignment in the following sense: the left
(right) peripheral element of R corresponds to the left (right) periph-
eral element of B, if R is to the left (right) of B.

Convince yourself that all reduplications that we have seen so far satisfy this
constraint: see again the diagrams in (33).
But Anchoring-BR is actually a violable constraint, as can be illustrated by
the perfective prexing reduplication pattern of Sanskrit (Whitney 1889, Steriade
1988a). In all examples below the reduplicant simplies the complex onset of
the base, dropping the initial consonant of a cluster [sk] or [st]:

(35) Sanskrit perfective reduplication, full grade (copies rst CV, drop-
ping /sC/)
a. ka-skand-a leap
b. ta-stambh-a prop

Onset simplication is yet another illustration of reduction of syllabic markedness

in the reduplicant (Steriade 1988a). This goes at the expense of Max-BR; also,
what is more, of Anchoring-BR. The violation of Anchoring-BR in [ka-skand-
a] is shown in (36). The leftmost segment of the base, the [s], fails to correspond
to the leftmost segment of the reduplicant, the [k]:

(36) Correspondence diagram for violation of Anchoring-BR

k a- Reduplicant

s k a n d -a Base

7
Recall that in chapter 3 we saw its input-faithfulness version, Anchoring-IO, active in the
blocking of consonant deletion at the edge of a stem in Diola-Fogny.

213
Correspondence in reduplication

As Steriade observes, there is a change towards less marked structure in the

reduplicant. A complex onset in the base ([sk]) is simplied into a simple onset
([k]) in the reduplicant. Lack of complex onsets in Sanskrit reduplicants is due
to *Complex, a syllabic well-formedness constraint introduced in chapter 3.
This emergence of the unmarked is due to a ranking of *Complex above both
Max-BR and Anchoring-BR. But *Complex must itself be dominated by Max-
IO, for complex onsets freely surface in the base. The preliminary ranking is:

(37) Emergence of the unmarked in Sanskrit reduplication

Schema Faithfulness
Well-formedness
Reduplicative
identity
Instantiation Max-IO
*Complex
Max-BR,
Anchoring-BR

5.2.4.2 Violation of CONTIGUITY in Sanskrit

The Sanskrit perfect reduplication pattern also serves to clarify the notion of
contiguity. Although Anchoring-BR is violated in the examples in (35), these
examples still preserve the contiguity of corresponding elements in the base and
the reduplicant. As shown in (36), the two segments that stand in correspondence,
[ka], form a contiguous substring of both base and reduplicant. This property is
stated in constraint format in (38):

(38) Contiguity-BR
The portion of the base standing in correspondence forms a contig-
uous string, as does the correspondent portion of the reduplicant.

Contiguity-BR requires the reduplicant to be a contiguous substring of the

base, leaving no medial gap. In Diyari prexing reduplication (33a), this protects
reduplicant-medial coda consonants, though not reduplicant-nal ones: tj il.pa -
tjil.par.ku, * tj i.pa -tjil.par.ku.
Like previous basereduplicant identity constraints, Contiguity-BR is viol-
able. To see this we must consider another set of examples of Sanskrit perfective
reduplication (Whitney 1889, as cited in Steriade 1988a: 120):

(39) Sanskrit perfective reduplication, full grade (copies rst CV, drop-
ping C2)
a. pa -prath-a spread
b. ma-mna-u note
c. sa-swar sound
d. da-dhwans-a scatter

214
 5.2 Reduplicative identity: the constraints

Violation of Contiguity-BR is illustrated in (40). The corresponding string of

elements in reduplicant and base, [pa], is not a contiguous substring of the base,
since in the base [r] interferes. The segment /r/ in the base is skipped:

(40) Correspondence diagram for violation of Contiguity-BR

p a- Reduplicant

p r a th -a Base

As in the previous example of Sanskrit of section 5.2.4.1, we nd that the redu-

plicant is of lesser markedness than its base, since it contains no complex onset.
Again onset simplication is due to *Complex. This points to an interaction of
faithfulness, well-formedness, and reduplicative identity constraints that is very
similar to what we saw earlier in (37):

(41) Emergence of the unmarked in Sanskrit reduplication

Schema Faithfulness
Well-formedness
Reduplicative
identity
Instantiation Max-IO
*Complex
Max-BR,
Contiguity-BR

A nal aspect of the Sanskrit pattern merits consideration. Reduplication cop-

ies the second consonant in a base beginning with /s/ + plosive ([ ka-skand-a],
35a), whereas it copies the initial C in all other /CC/-initial bases ([pa -prath-a],
39a). Gnanadesikan (1995) observes that the choice depends on sonority. The
consonant of lowest sonority is the one to surface in the reduplicant. Plosives rank
lower on the sonority scale than fricatives:

(42) Sonority scale

plosivesfricativesnasalsliquidsvocoids

The preference for low-sonority onset segments coincides with the preference for
abrupt rises of sonority at the beginning of the syllable (Clements 1990).8 We will
not present an analysis of sonority-governed onset simplication, and refer the
reader to Gnanadesikan (1995).

5.2.5 The emergence of the unmarked

An observation that has been made time and again in previous sections is that
reduplicants eliminate segmental or prosodic markedness, whereas the same

8
We also observe the logically complementary requirement of that made by HNuc (Prince and
Smolensky 1993), that syllable nuclei are maximally sonorous.

215
Correspondence in reduplication

markedness is tolerated in any non-reduplicant in the language. The interaction

of faithfulness, well-formedness, and reduplicative identity constraints that pro-
duces an unmarked reduplicant is referred to by McCarthy and Prince (1994a) as
an emergence of the unmarked. It is stated in (43):

(43) Ranking schema for emergence of the unmarked

IO-faithfulness
Well-formedness
BR-identity

We have already seen three examples of this ranking schema. The rst was redu-
plication in Nootka (20), which produces an open syllable, even though the lan-
guage allows closed syllables in other morphemes than reduplicants. The second
was Akan reduplication (28), which produces high reduplicant vowels, even
where the base contains non-high vowels. Finally, Sanskrit reduplication (37 41)
produces single onsets where the base contains a complex onset. All three rank-
ings (repeated below) are instantiations of the schema (43):

(44) a. Nootka: Max-IO
No-Coda
Max-BR

b. Akan: Ident-IO(high)
*[high]
Ident-BR(high)
c. Sanskrit: Max-IO
*Complex
Max-BR

This completes our discussion of reduplicative identity constraints. In the next

section we will look into another core ingredient of analyses of reduplication: the
constraints that determine the prosodic shape invariant.

5.3 From classical templates to generalized templates

5.3.1 The classical template in Agta

We will now address the rst observation made in the introduction of this chapter.
That is, reduplicants tend to have prosodic shape invariants which are not neces-
sarily related to a prosodic unit in the base. To learn about the contributions of
OT in this respect, we must rst focus on aspects of the ancestor theory of
templates, and then compare these aspects with the OT viewpoint. We will outline
the classical template as it played a role in pre-OT prosodic morphology, dis-
cussing the major conceptual changes with respect to templates that took place as
a result of OT.
A CVC shape invariant occurs in the following examples of reduplication in
Agta (Healey 1960, Marantz 1982, McCarthy and Prince 1986), partly repeated
from (9):

(45) Agta reduplication (copies enough material to form a heavy syllable)

a. ba.ri body bar-ba.ri-k kid-in my whole body
b. tak.ki leg tak-tak.ki legs
216
 5.3 From classical templates to generalized templates

The generalization is that the reduplicant has the size of a heavy syllable, regard-
less of whether the segments which ll it are taken from open or closed syllables
in the base. To account for such shape invariance, Marantz (1982) assumed a
notion of template, a morpheme having the dening property of being melodic-
ally empty. In his template-and-association theory, all prosodic properties den-
ing shape-invariance of reduplicants are stated in the template. The Marantzian
template is a sequence of skeletal segments, for example CVC- in the case of
Agta, to which the melodic content of the base is associated from a fully copied
string. Unassociated melodic elements are deleted:

(46) C V C + C V C V copy C V C + C V C V assoc. C V C + C V C V

b a r i ba r i ba r i b a r (i) b a r i

CV-templates were replaced by prosodic templates by McCarthy and Prince

(1986, 1990), whose Prosodic Morphology Hypothesis says that templates are
dened in terms of authentic units of prosody: mora (), syllable (), foot (F),
prosodic word (W), and so on (1990: 209). In Agta, the reduplicative template
is a heavy syllable . McCarthy and Prince argued that all other aspects of the
reduplicants make-up (the presence of an onset, obligatory coda to satisfy sylla-
ble weight, etc.) follow from universal principles as well as language-specic
principles of well-formedness. For example, since coda consonants make weight
in Agta, the second mora in the template is satised by associating a consonant
to it:

(47) + copy + associate +

ba r i ba r i b a r i b a r (i) b a r i

This classical theory of reduplicative templates includes other prosodic cat-

egories, such as the core syllable (Sanskrit), the disyllabic foot (Diyari), and
PrWd (Indonesian).
It is a straightforward matter to translate this derivational analysis of Agta into
ranked OT constraints; all it takes is dening a templatic constraint:

(48) Red=
The reduplicant is a heavy syllable.

and ranking this above the competing constraints No-Coda and Max-BR (the
OT analogue of full-copying-and-association). A simple tableau illustrates this
ranking:
217
Correspondence in reduplication

(49) Input: /red + bari/ Red= No-Coda Max-BR

a. + bar-ba.ri * *
b. ba-ba.ri *! **
c. ba.ri-ba.ri *!

In early OT, McCarthy and Prince (1993b: 82) construed templates as con-
straints on the prosodymorphology interface, alignment constraints asserting the
coincidence of edges of morphological and prosodic constituents:
The place to look for generalization of the notion of template, we propose,
is in the family of constraints on the prosody/morphology interface, such
as Align. The idea is that the Reduplicant must be in a particular align-
ment with prosodic structure. The strictest such alignments will amount to
classical templates.

The classical template is thus translated into a constraint schema (McCarthy and
Prince 1993b):
(50) Constraint schema for classical templates
Mcat=Pcat
where Mcat Morphological Category Prex, Sufx, Red,
Root, Stem, LexWd, etc.
and Pcat Prosodic Category Mora, Syllable (type), Foot
(type), PrWd (type), etc.
However, recent developments in OT have surpassed this simple translation of
classical templates into constraints. This new approach to templates in OT will
now be illustrated by an analysis of Diyari reduplication, which involves a sharp
reduction of the role of templates, and a correspondingly increased role for gen-
eral prosodic constraints.

5.3.2 The Generalized Template in Diyari

McCarthy and Prince (1986) assumed a monolithic notion of the template, as a
single unit combining in it all invariant shape properties of the reduplicant, rang-
ing from its status as a prex or sufx to its prosodic size. McCarthy and Prince
(1994a, 1994b) argued against this view of templates, and proposed to eliminate
it by interactions among independently motivated constraints. This approach is
known as Generalized Template theory. Its key idea is that templatic specicat-
ion is minimal, consisting only of a statement to the effect that the reduplicant
equals an afx or stem. Any shape characteristics of reduplicants are derived
from interactions of prosodic well-formedness constraints and constraints on
218
 5.3 From classical templates to generalized templates

reduplicative identity.9 Such a decomposition of the classical template into inter-

actions of constraints is a highly natural result in OT.
We will illustrate this approach by an analysis of reduplication in Diyari, after
McCarthy and Prince (1994a, forthcoming). The Diyari pattern was introduced in
section 5.1, where we already observed the disyllabicity of the reduplicant, a
shape invariance which we linked with the category foot. Feet are primarily units
of stress, and quite appropriately, properties of the Diyari stress system will
explain the disyllabicity of the reduplicant. We must therefore begin by studying
Diyari stress in some detail.
Diyari lacks a distinction of syllable weight: all its syllables count as light.
Hence, feet are binary under syllabic analysis, or in metrical terms, feet are
disyllabic trochees. The stress pattern of (unreduplicated) stems is illustrated in
(51). Primary stress is initial, and a secondary stress falls on the third syllable of
a four-syllable stem:

(51) a. k=a man

b. t jClparku bird species
c. wC2apDna old woman

From chapter 4, we know that this iterative stress pattern is due to the constraint
ranking in (52). This in fact equals the ranking for Pintupi (see again tableau 74
in chapter 4):

(52) Ft-Bin
Parse-Syl
All-Ft-Left

A second observation directly relevant to the disyllabic size of the reduplicant is

that this equals the minimal prosodic word of Diyari: all stems are minimally
disyllabic.10
Our hypothesis that the reduplicant is a PrWd, which was based on its disyl-
labic size, is conrmed by stress patterns. The examples below show that a pri-
mary stress falls on both the base and the reduplicant:

(53) a. wC2a wC2a -wC2a woman

b. tjClparku tj Clpa -t jClparku bird species
c. wClapDna wCla -wClapDna old woman

Since each primary stress heads one PrWd, we must conclude that the reduplicant
equals a PrWd itself.

9
This step was anticipated by Steriade (1988a), and more explicitly by Shaw (1991), who
claims that Template = sum of constraints on prosodic weight and constraints on markedness
on its structure (pp. vii1).
10
Poser (1989) observes that the only exception is the monosyllabic function word ya and.

219
Correspondence in reduplication

In classical template theory (McCarthy and Prince 1986) this pattern involves
a reduplicative template that directly encodes the prosodic size of the reduplicant
as a foot, which equals a minimal PrWd. In OT, setting up a templatic constraint
analogous to the classical template is certainly possible, for example Red=Ft.
But a far more interesting approach is possible, which makes the metrical phono-
logy of the language do most of the work, and also eliminates the template in its
classical sense of a prosodic shape invariant associated with a morphological
category. This new approach is known as Generalized Template theory (McCarthy
and Prince 1994b). As stated earlier, its key idea is that the shape invariance
of reduplicative templates is the net result of interaction between prosodic well-
formedness constraints and constraints of reduplicative identity, an approach
which radically simplies the specication of the reduplicative morpheme as
either stem or afx. Recall the important observation made in section 5.1.2,
that reduplicants tend to have unmarked prosodic structures. This is explained by
making well-formedness constraints directly responsible for determining shapes
of reduplicants. Accordingly, we must strip templates of any template-specic
prosodic requirements, reducing them to a bare minimum and specifying them
maximally as the morphological categories afx or stem, differing from their
non-reduplicative counterparts only in having no segmental input specication.
Their prosodic shapes will then emerge from constraint interaction.

In the case of Diyari, all we need to state specically for the reduplicant is that
it equals a stem:

(54) Red==Stem
The reduplicant is a stem.

This is the only requirement specic to reduplication. It sets in motion a chain of

effects, dening the prosodic shape of the reduplicant by interactions of prosodic
constraints that are not specic to reduplication, but independently motivated in
the Diyari stress system.
The reduplicants stem status implies its PrWd status, mediated by undomin-
ated Stem=PrWd, a constraint requiring that a stem must equal a PrWd:

(55) Stem==PrWd
A stem equals a PrWd.

This is nothing but the minimal stem requirement of Diyari, independently motiv-
ated for non-reduplicative categories. Naturally the reduplicant-as-a-stem is auto-
matically subject to any prosodic limitation that may hold for PrWd. Crucially,
PrWd must be minimally a foot in size, due to the prosodic hierarchy, in which

220
 5.3 From classical templates to generalized templates

every PrWd is headed by a foot. This single-foot minimum translates as a disyl-

labic minimum, due to undominated Ft-Bin. (Diyari has no contrast of syllable
weight, hence feet are strictly disyllabic.)

We have now tracked the chain of effects which follow from a single requirement
that the reduplicant is a stem, resulting in the prediction that the reduplicant is
minimally disyllabic. However, the Diyari reduplicant is not only minimally disyl-
labic, but rather exactly disyllabic. Earlier theories of reduplication took this exact
size to be an indication of its templatic nature, assuming a morphological tem-
plate of the exact size of a metrical category, here a foot (Poser 1989). However,
a more interesting approach is possible, as McCarthy and Prince (1994a, 1995a)
demonstrate. The exact limitation of the reduplicant to a single foot follows from
interactions of metrical constraints which are independently motivated for Diyari.
Observe that the single-foot reduplicant is in most ways metrically perfect
since its single foot is aligned at both edges, while it has no unparsed syllables.
Consider three candidate reduplicants for old woman (53c), of two, three, and
four syllables:

(56) a. [(wC.la)]-
b. [(wC.la).pi.]-
c. [(wC.la).( pD.na )]-

The trisyllabic reduplicant (56b) is less harmonic than a disyllabic one (56a) in
terms of Parse-Syl. In a trisyllabic reduplicant, one syllable must remain
unparsed since feet are strictly binary, while a disyllabic reduplicant can be
exhaustively parsed by a single foot. Moreover, a four-syllable reduplicant (56c)
is less harmonic than a disyllabic one. This is because to parse four syllables it
takes two feet, one of which is necessarily misaligned at the left edge, violating
All-Ft-Left. We nd that a disyllabic reduplicant approximates the metrically
ideal PrWd, one exhaustively parsed by a single, perfectly aligned foot:

(57) a. The reduplicants foot is disyllabic (by Ft-Bin)

b. The reduplicants syllables are exhaustively parsed (by Parse-Syl)
c. The reduplicants foot is left-aligned, hence single-footed (by
All-Ft-L)

This is an emergence of the unmarked in the Diyari reduplicant. Its ideal disyl-
labic size, favoured by the sum of metrical constraints, is not attested in normal
non-reduplicative stems, which exceed disyllabicity in the case of trisyllabic or
quadrisyllabic stems.

221
Correspondence in reduplication

Next we turn to the issue of what constraint interaction limits perfect disyllabicity
to reduplicants, while leaving other stems free to surpass it. Non-reduplicant
stems fail to undergo segment deletion in order to attain a perfect disyllabic size:
IO-faithfulness takes precedence over metrical well-formedness. But for the redu-
plicant, priorities are reversed. Metrical well-formedness takes precedence over
reduplicative identity, and as a result reduplication is partial, approaching the
ideal metrical size. The schematic interaction is:

(58) Emergence of the unmarked in Diyari reduplication

Schema Faithfulness
Well-formedness
Reduplicative
identity
Instantiation Max-IO
Metrical constraints
Max-BR

The full hierarchy involves a ranking of all metrical well-formedness con-

straints. Some of these are in fact undominated (Stem=PrWd, Ft-Bin), while
others (Parse-Syl, All-Ft-Left) are dominated by segmental faithfulness
(Max-IO). At the bottom of the hierarchy we nd Max-BR, enforcing reduplic-
ative identity.

(59) Diyari disyllabic reduplication

Max-IO, Red=Stem, Stem=PrWd, Ft-Bin
Parse-Syl
All-
Ft-Left
Max-BR

This analysis is illustrated by the tableaux (60 1). We leave out any candidates
violating undominated Red=Stem, focussing on the effects of the genuinely pro-
sodic constraints. Tableau (60) shows reduplication of a trisyllabic stem.

(60)
Input: /red + tjilparku/ Max- Stem= Ft- Parse- All-Ft- Max-BR
IO PrWd Bin Syl Left

a. + [( tj Cl.pa )]-[(tjCl.par).ku] * r, k, u
j j
b. [( t Cl.par ).ku]-[(t Cl.par).ku] **!
j j
c. [( t C )]-[(t Cl.par).ku] *! * l,p,a,r,k,u
d. [( tj C -tjil).(pr.ku)] *! l,p,a,r,k,u
e. [( tj Cl.pa )]-[(tjCl.pa)] r!,k,u

Candidate (60e) deletes input segments, fatally violating Max-IO. Next, candid-
ate (60d) represents an attempt at prosodically integrating a monosyllabic redu-
plicant with the base at the expense of Stem=PrWd. Remaining candidates differ
in the number of syllables in the reduplicant. A monosyllabic reduplicant (60c)
222
 5.4 From circumscription to alignment

fails on undominated Ft-Bin, while a trisyllabic total reduplicant (60b) exces-

sively violates Parse-Syl. Therefore disyllabic reduplication (60a) is optimal
even though it has three violations of Max-BR.11
Tableau (61) is yet another illustration of the same point with a subtle differ-
ence: an attempt at total reduplication (candidate 61b) now falls victim to All-
Ft-Left, instead of Parse-Syl (as in the previous tableau).

(61)
Input: /red + wilapina/ Max- Stem= Ft- Parse- All-Ft- Max-BR
IO PrWd Bin Syl Left

a. + [(wC.la)]-[(wC.la).(pD.na)] p, i, n, a

b. [(wC.la).( pD.na )]-[(wC.la).(pD.na)] !

c. [(wC.la). pi .]-[(wC.la).(pD.na)] *! n, a

d. [(wC)]-[(wC.la).(pD.na)] *! l,a,p,i,n,a

e. [(wC.la)]-[(wC.la)] p!i,n,a

This analysis of Diyari has shown that unmarked prosody in the shape and
size of the reduplicant is indeed due to universal prosodic markedness constraints.
But it takes violable constraints to reach this conclusion: the same universal
markedness constraints which govern the reduplicant are massively violated in the
non-reduplicative forms of the language. This is not an inconsistency, but simply
reects the differences in strength and scope of two sets of correspondence con-
straints: inputoutput and basereduplicant.
In the next section, we will add a nal ingredient to the analysis of reduplica-
tion: the constraints determining the position of the reduplicant.

5.4 From circumscription to alignment

5.4.1 Inxing reduplication and prosodic circumscription

The template-and-association theory of reduplication (Marantz 1982, McCarthy
and Prince 1986) involves a strict separation between the morphological operation
(the afxation of a prosodic template) and the phonological operation (the asso-
ciation of the segments of the base, subject to language-specic and universal
principles). However, this separation of morphology and phonology causes a
number of descriptive problems that turn out to undermine the very fundamentals
of the theory.
11
Observe that yet another candidate, [( tj Cl.par )]-[(t jCl.par).ku], with the reduplicant ending in
a closed syllable, is not given in (60). As we have shown in section 5.2.2.1 for Nootka, this
candidate will be excluded by the ranking No-Coda
Max-BR.

223
Correspondence in reduplication

One problem involves internal reduplication: in many languages the reduplic-

ant is an afx to a prosodically circumscribed part of the stem. For example,
Samoan (Marsack 1962) prexes the reduplicant to the stress foot:

(62) Samoan (reduplicant is prexed to the stress foot)

a. fa.nu fa-na-nu be born, give birth
b. a.lfa a-lo-lfa love

Timugon Murut (Prentice 1971) prexes the reduplicant to a part of the stem that
is the residue of subtracting its initial vowel (or initial onsetless syllable):

(63) Timugon Murut inxing reduplication: prexation skips over stem-

initial onsetless syllable
a.i bulud bu-bulud hill/ridge
a.ii tulu tu-tulu point at
a.iii dondo do-dondo one
b.i ulampoy u-la-lampoy (no gloss)
b.ii indimo in-di-dimo ves times
b.iii ompod om- po -pod atter

The fundamental analytic problem is this: in order to apply the morphological

operation (the prexation of a template), some prosodic operation must rst take
place, establishing the base of prexation its precise position internal to the
stem. Development of the theory led to a skipping mechanism termed prosodic
circumscription (McCarthy and Prince 1990, 1995b). Circumscription places a
cut in the stem at a prosodically dened position, either before or after a prosodic
unit (or a segment, as in Timugon Murut), thus splitting the domain into two
parts. Reduplication takes as its base one part (positively circumscribing the
foot in Samoan, while negatively circumscribing the initial vowel in Timugon
Murut), and subsequently both parts are reunited, producing the inxation:

(64) Circumscription analyses of Samoan and Timugon Murut inxing

reduplication
Stem fanu ulampoy
Circumscribe {fa} {nu}Ft {u}V {lampoy}
Reduplicate () {fa} {na-nu}Ft {u}V {la-lampoy}
Reunite fa-na-nu u-la-lampoy

Technically speaking, prosodic circumscription saves the day. It produces a gen-

eralizing and ingenious analysis of internal reduplication, but that is where it
ends. The analysis still fails to answer the fundamental question of why a mor-
phological operation should apply to the output of a prosodic operation. Prosodic
circumscription has no purpose but to serve the morphology with a base of
224
 5.4 From circumscription to alignment

reduplication.12 Moreover, circumscription leaves the relation between the posi-

tion of the inx and its status as a reduplicant unexplained. Consider, for example,
the fact that in Timugon Murut inxation it is an initial vowel which is skipped
rather than a consonant, a fact which should be related to the perfect CV shape
of the reduplicant (it has an onset and lacks a coda). What we do not nd, and
will probably not nd in any language, is the reverse pattern in which an initial
consonant, but not an initial vowel, is skipped (*b-u-ulud, *u-ulampoy). The point
is clear: circumscription helps to full the requirements of prosody yet this is
totally unclear from the analysis: circumscription is principally unable to look
ahead to the output of reduplication in order to locate the position of the cut.
McCarthy and Prince (1993a) observe another major defect of the circumscrip-
tion analysis: it leaves unexpressed the cross-linguistic generalization that it is
a reduplicative inx that skips over the initial vowel. The reduplicant is totally
dependent on its base for its segmental content. Therefore an onset consonant
must be copied from the base, which is exactly what circumscription accom-
plishes: it makes the reduplicant sit before a CV sequence in the base. But, in
principle, the same circumscription analysis might have been employed for the
inxation of a (hypothetical) non-reduplicative afx, for example, -ka-:

(65) Hypothetical inxation

a. bulud ka-bulud
b. ulampoy u-ka-lampoy

However, such patterns are remarkably absent from the worlds languages. This
can only be due to the fact that a fully specied afx is not segmentally depend-
ent on its stem, and hence need not immediately precede a consonant in the stem.
But circumscription theory is unable to make this connection, since it totally
separates the circumscription operation from the morphological operation that it
potentiates.

A circumscriptional analysis cannot explain why, in all known cases, it is

always a reduplicative inx that skips over the initial onsetless syllable.
(McCarthy and Prince 1993a: 126)

These problems turn out to be serious enough to warrant a new approach, one
in which morphology and prosody apply in parallel and in which alignment
plays a central role. Such an approach is highly compatible with the organization
of the grammar in OT, where prosodic constraints and morphological constraints
are part of the same hierarchy. This approach will now be illustrated for inxing
reduplication in Timugon Murut.
12
Actually, circumscription saw various other ingenious applications in prosodic morphology
(see McCarthy and Prince 1990, 1995b), which also are subject to the same general criticisms.

225
Correspondence in reduplication

5.4.2 Inxing reduplication and alignment

Let us now consider the analysis of Timugon Murut inxing reduplication in OT.
As we noted earlier, the interest of this pattern is that it is partly prexing,
and partly inxing. The reduplicant is a light syllable prex located strictly at the
beginning of the word if the base begins with a consonant (66a). But if the word
begins with a vowel, the prex becomes an inx, skipping over an initial onset-
less syllable (66b):

(66) Timugon Murut (repeated from 63)

a. bulud bu-bulud hill/ridge
b. u.lampoy u-la-lampoy (no gloss)

This initial-vowel-skipping reduplication pattern is fairly common cross-

linguistically.
McCarthy and Princes analysis captures the following idea. The reduplicative
inx sits initially in the word, except when prexation would produce an onset-
less reduplicant syllable, as in the case of a vowel-initial stem (67):

(67) u-la-lampoy
*u-ulampoy (rejected due to extra violation of Onset)

The basic constraint interaction is analogous to that of Tagalog -um-inxation,

discussed in chapter 3, section 3.4.2. It involves domination of Onset13 over an
edgemost constraint requiring that the reduplicant be left-aligned in the word:

(68) Align-Red-L
Align the left edge of the reduplicant with the left edge of the PrWd.

This interaction is illustrated by tableau (69), which also presents a third candid-
ate (69c), showing the activity of Align-Red-L in keeping the inx leftmost in
the word:

(69) Input: /red, ulampoy/ Onset Align-Red-L

a. + u-la-lampoy * u
b. u-ulampoy **!
c. ulam-po-poy * ul!am

The reduplicative prex becomes an inx because of a high-ranking prosodic

constraint. Even if the prex cannot be a genuine prex, it still locates itself as

13
Undominated input faithfulness constraints, Dep-IO and Max-IO, block consonant epenthesis
and vowel deletion as strategies to avoid onset-less syllables.

226
 5.4 From circumscription to alignment

close as possible to the left edge, which is what makes (69a) more harmonic than
(69c).
This nishes the core of the analysis. Its success (as compared to circumscrip-
tion) is due to two factors. First, the fact that prosody and morphology operate
in parallel, rather than serially. The position of the inx (morphology) depends
on the relative prosodic well-formedness of the output. The second key feature of
this analysis is the violability of constraints. The requirement that the reduplicant
is a prex (Align-Red-L) is violable precisely to the extent necessary to avoid
violation of prosodic well-formedness (Onset).

There are some residual issues, to which we now turn. Consider another logically
possible strategy to avoid excessive violation of Onset, while keeping the redu-
plicative afx strictly initial in the word. This strategy is to copy yet another
segment of the base, producing the candidate *ul-ulampoy. It may appear that this
is a genuine improvement over u-la-lampoy: both have one onset-less syllable, but
the former is a perfect prex, while the latter is a near miss. Then what excludes
*ul-ulampoy? Observe that the single improvement of *ul-ulampoy over *u-
ulampoy resides in its lesser violation of Onset, due to its additional onset con-
sonant [l]. But in order to make this [l] a genuine onset, it must syllabify into
the base, thereby blurring the edge between reduplicant and base.

(70) u-la-lampoy
*u.l-u.lampoy (rejected due to violation of red =)

This violates the undominated templatic alignment constraint Red=:

(71) Red=
Align both edges of the reduplicant with the edges of a syllable.

This is an example of a templatic alignment constraint.14 It actually fulls two

functions. It enforces the alignment of the reduplicant with a syllable, and it also
restricts the size of the reduplicant to a single syllable. We assume gradient
evaluation for Red=, with one violation for each segment exceeding the strict
syllable-size of the reduplicant.
The interaction is shown in tableau (72), where violations of both constraints
are marked by the offending segments:

(72) Input: /red, ulampoy/ Red= Align-Red-L

a. + u-la-lampoy u
b. u.l-u.lampoy l!

14
We leave open the consequences of adopting this constraint for Generalized Template theory.

227
Correspondence in reduplication

As noted above, apart from its alignment function, Red= is also responsible for
the size limitation of the reduplicant to one syllable. Therefore it must dominate
Max-BR, ruling out output forms such as *u- lam.poy -lam.poy.
Let us now turn to the interaction between IO-faithfulness constraints and
syllabic well-formedness constraints. Timugon Murut disallows consonant epen-
thesis and vowel deletion as strategies to repair onset-less or closed syllables.
From this we conclude that faithfulness (Dep-IO, Max-IO) dominates prosodic
well-formedness (Onset, No-Coda):

(73) Input: /ulampoy/ Dep-IO Max-IO Onset No-Coda

a. + u.lam.poy * **
b. lam.poy *! **
c. u.la.po *!* *
d. Tu.lam.poy *! **
e. u.la.mA.po.yA *!* *

It thus appears that reduplicative inxation in Timugon Murut is yet another

emergence of the unmarked: markedness constraints that are normally invisible
in the phonology of the language suddenly jump into full activity in situations
where input faithfulness is not at stake.15
There is independent evidence from reduplication for the activity of No-Coda,
a constraint whose effects are normally suppressed by faithfulness. The inx is
always of the shape CV, even though a closed syllable would no doubt be prefer-
able on Max-BR:

(74) u-la-lampoy
*u-lam-lampoy (rejected due to violation of No-

Coda)

Therefore No-Coda must dominate Max-BR.

Let us now put all pieces together, and consider the complete analysis. We have
evidence for three undominated constraints: Red=, Dep-IO, Max-IO. These
dominate a pair of syllable well-formedness constraints: Onset and No-Coda,
which in their turn dominate two reduplicative constraints: Align-Red-L and
Max-BR:

15
In fact, one input faithfulness constraint is violated under inxation: Contiguity-IO.

228
 5.5 Classical versus OT-based prosodic morphology: conclusions

(75) Red=, Dep-IO, Max-IO
Onset, No-Coda
Align-Red-L,

Max-BR

This total ranking is illustrated by the tableau of [u-la-lam.poy]:

(76)
Input: /red, ulampoy/ Red= Dep- Max- Onset No- Align- Max-
IO IO Coda Red-L BR

a. + u-la-lam.poy * ** u mpoy

b. u-lam-lam.poy * ***! u poy

c. u-u.lam.poy **! ** lampoy

d. ul.-u.lam.poy **! *** ampoy

e. la-lam.poy *! ** mpoy

f. Tu-Tu.lam.poy *! ** lampoy

g. u.l-u.lam.poy l! * ** ampoy

h. u- lam.poy -lam.poy p!oy * **** u

Finally, reconsider the serial analysis of this pattern presented in section 5.4.1,
which was based on the circumscription of the initial vowel and prexation of the
reduplicant to the residue of the base. We ended section 5.4.1 by reecting on the
observation, due to McCarthy and Prince (1993a), that a circumscriptional ana-
lysis is principally unable to explain why it is always a reduplicative inx that
skips over the initial onset-less syllable.
Now observe that an OT analyis does make this prediction, for nothing would
be gained (over simple prexation) by the inxation of a segmentally specied
prex /ka-/. Both of the output candidates forms below have one violation of the
constraint Onset:

(77) ka-u.lam.poy
u-ka-lam.poy

Hence the choice is automatically made in favour of the optimally aligned

prexation candidate. (We leave it to the reader to work out the details in a
tableau.)

5.5 Classical versus OT-based prosodic morphology: conclusions

The following overview sums up the major differences, discussed in the preceding
sections, between classical and OT-based prosodic morphology:

229
Correspondence in reduplication

(78) Classical prosodic morphology OT-based prosodic

morphology
a. Templates have xed prosodic shapes. Templates emerge from
interactions.
b. Prosodic circumscription of domains. Violable alignment
constraints.
c. Serialism: morphology prosody. Parallelism: prosody

morphology.

The modications of the classical theory of prosodic morphology are integrated

into the revised version of the principles of prosodic morphology:

(79) New prosodic morphology (McCarthy and Prince 1993b: 138)

a. Prosodic Morphology Hypothesis
Templates are constraints on the prosody/morphology inter-
face, asserting the coincidence of morphological and prosodic
constituents.
b. Template Satisfaction Condition
Templatic constraints may be undominated, in which case they
are satised fully, or they may be dominated, in which case they
are violated minimally, in accordance with general principles of
Optimality Theory.
c. Ranking Schema
P
M

In the following sections, we will discuss further evidence that lends support to
the OT-based theory of reduplication. We will rst turn to another central theme
of research in reduplication theory, overapplication and underapplication.

5.6 Overapplication and underapplication in reduplication

5.6.1 Introduction
As we saw in previous sections, the basic idea behind the OT model of redup-
lication is that reduplicative identity constraints compete with well-formedness
constraints; within the same form, well-formedness constraints compete with
input-faithfulness constraints. Both two-way interactions are summed up below:

(80) a. Well-formedness BaseReduplicant Identity

b. Well-formedness InputOutput Faithfulness
230
 5.6 Overapplication and underapplication in reduplication

The most complex interaction combining all these ingredients that we have seen
thus far is the emergence of the unmarked:

(81) IO-Faithfulness
Well-formedness
BR-Identity

But this interaction does not exhaust the possibilities of the Correspondence
Theory of reduplication. Simple reranking predicts that reduplicative identity may
be promoted to a position in the hierarchy where it produces a more complete
identity between the base and reduplicant than we have seen until here. Particu-
larly, high-ranked basereduplicant identity may cause the transfer of context-
sensitive phonological properties of the base to the reduplicant. Such effects are
known as overapplication and underapplication, and in this section we will
discuss their analysis by McCarthy and Prince (1995a, forthcoming).
The basis of the OT model of overapplication and underapplication resides in
an observation by Wilbur (1973: 58):

(82) The Identity Constraint

There is a tendency to preserve the identity of Ro and Rr in reduplic-
ated forms.

The function of overapplication and underapplication is to make the relation

between the base and the reduplicant more transparent (see Anderson 1975 for
an earlier statement of this idea). Before we turn to actual cases of overapplication
and underapplication, we rst consider a case of normal application.

5.6.2 Normal application in Washo

An example of normal application of phonology to the base and reduplicant
comes from reduplication in Washo (Jacobsen 1964), which is discussed in
Wilbur (1973):

(83) Normal application of coda devoicing in Washo

a. /red + wis-i/ wis-wi.si its squeaking
b.i /red + wed-i/ wet-we.di its quacking
b.ii /red + ba
-i/ bak-ba.
i hes smoking
b.iii /red + 9ub-i/ 9up -9u.bi hes crying gently

The reduplicant prex, a heavy syllable, copies the rst CVC string from the
base. Note that if the base contains a voiced obstruent, this appears as voiceless
in the reduplicant. This is due to the familiar process of coda devoicing, which
applies generally in Washo. In contrast, the intervocalic obstruent in the base
reects its input status, unhampered by coda devoicing. The result is featural non-
identity for [voice] between the corresponding segments in the reduplicant and
the base.
231
Correspondence in reduplication

The Washo pattern of normal application follows naturally from the interac-
tion of the following three constraints:

(84) *Voiced-Coda
No voiced coda.

(85) Ident-IO(voice)
Let be a segment in I, and be a correspondent of in O.
If is [voice], then is [voice].

(86) Ident-BR(voice)
Let be a segment in B, and be a correspondent of in R.
If is [voice], then is [voice].

Because Washo has normal coda devoicing outside reduplication, *Voiced-

Coda must dominate Ident-IO(voice). But in its turn, Ident-IO(voice) must
dominate Ident-BR(voice). Loss of identity between the reduplicants coda con-
sonant and its correspondent, the intervocalic base consonant, is less important
than the preservation of the input value of [voice]. This can be inferred from the
following harmonic relation:

(87) wet-we.di
wet-we.ti Hence Ident-IO(voice)
Ident-BR(voice)

Basereduplicant identity is sacriced for the preservation of input voice. In

principle, the outcome might have been different (with devoicing overapply-
ing to the consonant in the base, to render it identical to the coda consonant in
the reduplicant). (Actually, we will come across such reduplicant-to-base back-
copying cases later in this chapter.)
Adding these rankings up, we nd:

(88) Normal application of coda devoicing in Washo reduplication

*Voiced-Coda
Ident-IO(voice)
Ident-BR(voice)

The following tableau illustrates normal application of coda devoicing in Washo:

(89) Input: /red + wed-i/ *Voiced-Coda Ident-IO Ident-BR

(voice) (voice)

a. + wet-we.di *
b. wet-we.ti *!
c. wed-we.di *!

232
 5.6 Overapplication and underapplication in reduplication

In sum, the normal application of phonology in the reduplicant is due to the

dominated position of a basereduplicant identity constraint below both input-
faithfulness and well-formedness.

(90) Ranking schema for normal application

Well-formedness
IO-Faithfulness
BR-Identity

Let us now turn to overapplication and nd out about the ranking schema under-
lying this phenomenon.

5.6.3 Overapplication in Malay

Wilbur denes overapplication as follows: the reduplicant undergoes some pho-
nological process even though it fails to meet the structural condition. That is, a
rule applies in the wrong environment.
A spectacular example of overapplication of phonology in reduplication comes
from Malay (Onn 1976, Kenstowicz 1981, McCarthy and Prince 1995a, forth-
coming). Malay has a complementary (allophonic) distribution of oral and nasal
vocoids (vowels and semivowels). When a word contains a nasal consonant, all
vocoids that follow this nasal are also nasalized. Vocoids that are not in post-nasal
position are always oral, with one exception: in reduplications nasal harmony
overapplies, as shown in the vocoids that are underlined in (91).

(91) Overapplication in Malay nasal harmony16

Stem Reduplicated
a.i hamO germ a.ii hmO-hmO germs
b.i waP fragrant b.ii SP-SP (intensied)
c.i an reverie c.ii n-n ambition
d.i aQn wind d.ii Qn-Qn unconrmed news

Nasality in the underlined vocoids cannot have normal spreading as its source,
as nasal harmony applies exclusively rightwards. This a clear example of a pho-
nological process overapplying to establish complete featural identity between
the reduplicant and the base. Observe that reduplicant and base are featurally
identical; for this reason it is impossible to identify the initial or nal elements
as either the reduplicant or base. However, the identication of the reduplicant
is irrelevant to the argument: all that matters is that these are reduplicated forms,
while nasal harmony overapplies exclusively in this context.

16
Nasality is transcribed in glides by McCarthy and Prince (1995a), but not by Onn (1976), the
source.

233
Correspondence in reduplication

Which constraints are involved? Let us rst take a look at the mechanism that
is responsible for the complementary (allophonic) distribution of oral and nasal
vocoids: nasal vocoids occur in post-nasal position, and oral vocoids elsewhere.
This we recognize as a typical case of complete dominance of markedness con-
straints with respect to [nasal] over IO-faithfulness: see chapter 1, section 1.5.2.
Two markedness constraints govern the distribution of [nasal]. The rst is context-
sensitive, and it excludes a sequence of a nasal and an oral vocoid, see (92).

(92) *NVORAL
*[nas] % [nas, vocalic]

This is undominated in Malay: no oral vocoids ever follow a nasal in the same word.
The second constraint active here is a context-free markedness constraint that
excludes nasal vocoids:

(93) *VNASAL
No nasal vocoids.

Obviously, *Vnasal is dominated by *NVoral in Malay, since nasal vocoids do occur

but only in post-nasal contexts, precisely where they are required by *NVoral.
Recall from chapter 1 that a complementary distribution points to the domin-
ance of markedness constraints over IO-faithfulness constraints. Here the relevant
faithfulness constraint is:

(94) Ident-IO(nasal)
Let be a segment in I, and be a correspondent of in O.
If is [nasal], then is [nasal].

When this is dominated by both markedness constraints governing nasality, the

outcome will be that input values of [nasal] in vocoids are totally irrelevant to
their output values. That is, whatever the lexical specication of [nasal] in a
vocoid may be, its surface value only reects markedness constraints. (Naturally
the language learner is never tempted to posit lexical values of [nasal] that are
irrecoverable from the surface evidence; this is Lexicon Optimization or Stam-
pean occultation; see again chapter 1.)
These remarks lead us to the following ranking:

(95) Nasal harmony in Malay

*NVoral
*Vnasal
Ident-IO(nasal)

Two tableaux of the same non-reduplicated form [waP] (differing in input

specication for the feature [nasal] in vocoids) serve to demonstrate the irrelev-
ance of the input. First consider tableau (96), which takes as its input a form
/waP/ that is actually identical to the surface form:
234
 5.6 Overapplication and underapplication in reduplication

(96) Input: /waP/ *NVoral *Vnasal Ident-IO(nasal)

a. + waP *
b. SP **!* **
c. wa *! *

The outcome would have been precisely the same if one or both vocoids had had
different input values for [nasal], for example the reverse pattern /S/:

(97) Input: /S/ *NVoral *Vnasal Ident-IO(nasal)

a. + waP * ***
b. SP **!* *
c. wa *! **

This analysis of the nasal harmony pattern serves as a necessary preparation

for the analysis of overapplication in reduplication, which is our genuine interest
in Malay. Recall that reduplicant and base are always featurally identical, even
though the context of nasalization is not met by the reduplicant. All we need to
do to capture this identity of base and reduplicant is rank Ident-BR(nasal) at the
top of the hierarchy, next to *NVoral:

(98) Overapplication of nasal harmony in Malay

Ident-BR(nasal), *NVoral
*Vnasal
Ident-IO(nasal)

This ranking captures the basic nature of overapplication: a well-formedness con-

straint at the top of the hierarchy, *NVoral, is amplied by a reduplicative iden-
tity constraint, Ident-BR(nasal), which carries its effects over to the reduplicant.
The overapplication of nasal harmony occurs at the expense of additional viola-
tions of *Vnasal:

(99) XP R-SPB
waP R-SPB

The nasalization of the initial vocoids [S] in the base is due to their post-nasal
position (with the triggering nasal as part of the reduplicant), that is, to normal
nasal harmony. The same holds for the last vowel in the reduplicant. But things
are different for the initial vocoids [S] of the reduplicant: being in word-initial
position, these vocoids are outside the scope of normal nasal harmony. They
do undergo nasalization, however, in order to retain featural identity with their
nasalized correspondents [S] in the base.
235
Correspondence in reduplication

Let us now turn to the constraint-based evaluation of some output candidates

for the reduplicated form of /waP/. The primary contestors of the optimal can-
didate (100a) are [ waP -waP] (100b), which preserves reduplicative identity by
underapplying nasal harmony, and [ waP -SP] (100c), which would be the
output of normal application:

(100) Input: /red + waP/ Ident-BR *NVoral *Vnasal Ident-IO

(nasal) (nasal)

a. + SP -SP ****** **
b. waP -waP *!* **
c. waP -SP *!* **** **

First, candidate (100b) preserves reduplicative identity at the expense of *NVoral,

which is directly fatal. This is the underapplication candidate: an attempt to
establish identity by overriding the application of nasal harmony where it should
have applied. However, the undominated position of the constraint enforcing nasal
harmony (*NVoral) precludes this attempt. (It may seem that underapplication
effects can be simply derived from the promotion of IO-faithfulness over a pho-
nological constraint; however, matters are not so simple, as we will see below.)
Second, the normal application candidate (100c) is doomed because it estab-
lishes normalcy at the expense of basereduplicant identity. This is fatal, due to
undominated Ident-BR(nasal).
Recall that due to the complete featural identity of the reduplicant and base, it
is actually impossible to tell the reduplicant and base apart. So far we have been
assuming that reduplication is prexation, but in fact it might just as well be
sufxation. In what respect would the analysis become different if the reduplicant
were sufxed to the base? The answer is: surprisingly little. This is due not so
much to the markedness constraints, whose evaluation is blind to notions such as
base and reduplicant (and registers violations to both equally alike), but to the
correspondence constraints.
First, observe that evaluations by Ident-BR(nasal) are unaffected by the dif-
ferent analysis into reduplicant + base: its violations remain the same, regardless
of what strings are taken as reduplicant and base. Second, in principle at least, the
identication of the base is relevant to Ident-IO(nasal), since only base segments
are subject to faithfulness. But we have just seen that this constraint is ranked low
enough never to be of relevance to the nasal harmony pattern. Hence the sufxa-
tion tableau in (101) is identical to the prexation tableau (100) in all relevant
aspects. The single (irrelevant) difference is the evaluation by Ident-IO(nasal).
236
 5.6 Overapplication and underapplication in reduplication

(101) Input: /waP + red/ Ident-BR *NVoral *VNASAL Ident-IO

(nasal) (nasal)

a. + SP- SP ****** **
b. waP- waP *!* **
c. waP- SP *!* ****

Note that if prexation were actually sufxation, then Malay nasal harmony
would be an overapplication of phonology to the base, rather than to the redup-
licant. The fact that this effect easily falls out of constraint interaction initially
seems a liability of the parallel OT model, rather than a strength. Does not this
parallel model predict languages in which a phonological property is back-
copied from the reduplicant onto the base? The answer is simple: this is an
empirical issue. There is no a priori reason to assume that back-copying does not
occur in the languages of the world. (It is actually attested, as we will see below
for Southern Paiute.)17

Let us now draw conclusions from the discussion of Malay. First, we have found
strong evidence for the role of basereduplicant identity, hence for the notion of
output-to-output correspondence, which is deeply parallel. Perhaps this argu-
ment gains in depth if we compare the analysis with a serial analysis. McCarthy
and Prince (1995a, forthcoming) point out that the Malay pattern poses severe
problems for a serial theory, in which the surface identity of reduplicant and base
is due to a copy rule. This copies the segmental content of the base onto the
reduplicant. Now consider an analysis of [ SP -SP] in which the reduplicant
is a prex.18 The nasalization of the initial segments [ S] in the reduplicant can
only be understood as being due to the copy rule, as these segments are not in
the proper context for nasalization. (They do not follow a nasal consonant.) The
string [S] must therefore be copied from the nasalized base segments [S],
which, at that point in the derivation, must already have been nasalized. From this
it follows that the copy rule must apply after the rule of nasalization. This is what
it means for a rule to overapply in serial theory: the effects of a phonological
rule in the base are copied onto the reduplicant. But there is a snag: nasalization
of base segments [S] is itself triggered by a preceding nasal but the base itself

17
A precursor of the correspondence analysis of back-copying effects can be found in Russell
(1993).
18
The problems posed to serial theory by a sufxing analysis of Malay reduplication are even
more severe, as this involves back-copying onto the base. However, since there is no way of
knowing which analysis ( prexing or sufxing) is correct, it is only fair to base an argument
on the prex analysis. In section 5.6.5 we will look into a clear case of back-copying.

237
Correspondence in reduplication

contains none! The trigger of nasalization in the base can only be the preceding
nasal segment [], which is part of the reduplicant. There is a paradox: by its very
nature, a reduplicant segment is a result of the copy rule. Hence the copy rule
must also precede nasalization. This mutual dependence of reduplicant and base
is shown by arrows in the diagram in (102):

(102) - - - - - - - -> [] in reduplicant induces nasalization of initial base

segments
SP + SP
<- - - - - - - - nasalized portion [ S] of base copied to reduplicant

This two-sided information ow requires a complex rule ordering in which the

copy rule applies twice in the derivation: once before nasalization, and once
after it:

(103) Underlying form /red + wa/

Copy wa + wa
Nasalization waP + SP
Copy SP + SP
Surface form [SP + SP]

The basic claim of derivational theory is that rules apply in a strict linear order.
Hence, a reapplication of a rule in a single derivation (within a single domain)
should be allowed under extremely limited circumstances only. More specically,
it has been proposed (see Myers 1991) that some rules reapply whenever their
structural description is met.19 If the Malay copy rule belongs to this class of
persistent rules, then what exactly does it mean for it to apply whenever its
structural description is met? McCarthy and Prince argue that the only reason-
able interpretation is: whenever base and reduplicant are non-identical, which
is, of course, a condition on the surface identity of the base and reduplicant, or
an output-to-output correspondence constraint in disguise.

The second result of the preceding section is a (preliminary) ranking schema for
overapplication. The core property of overapplication, as compared to normal
application (90), is this: a reduplicative identity constraint is promoted to an
undominated position:

(104) Ranking schema for overapplication

BR-Identity, Well-formedness
IO-Faithfulness

19
The standard example of a persistent rule is resyllabication, which reapplies after every
rule that affects syllable structure.

238
 5.6 Overapplication and underapplication in reduplication

Let us now evaluate this schema in a broader typological context. As has been
emphasized at various points in this book, a highly suitable way to assess overall
typological predictions of the theory is to construct a factorial typology by per-
mutation of constraint rankings. The central claim of the Basic Model of Corres-
pondence Theory is that reduplication patterns depend on interactions of three
constraint types: well-formedness, BR-identity, and IO-faithfulness. We have now
seen the following three ranking schemata (repeated from 43, 90, and 104 above):

(105) a. Ranking schema for emergence of the unmarked

IO-Faithfulness
Well-formedness
BR-Identity
b. Ranking schema for normal application
Well-formedness
IO-Faithfulness
BR-Identity
c. Ranking schema for overapplication
BR-Identity, Well-formedness
IO-Faithfulness

What we see is a progressively lowered position of IO-faithfulness in the hier-

archy from (105a) to (105c). To complete this factorial typology, we will now
add the case of total domination of well-formedness by IO-faithfulness and BR-
identity, which is stated in (105d) as total non-application:

(105) d. Ranking schema for total non-application

IO-Faithfulness, BR-Identity
Well-formedness

This ranking prohibits any changes potentially triggered by well-formedness in

either the base or reduplicant. The base must be faithful to its input due to IO-
faithfulness
well-formedness, while the reduplicant copies this unmodied
input shape due to BR-identity
well-formedness. This case is hardly interest-
ing, as it involves no phonology at work at all, hence we did not discuss it
previously. Yet, it is instantiated in every case in which lexical contrasts are
preserved in the base, as well as copied to the reduplicant.

Let us now return to the issue of underapplication. So far, McCarthy and Princes
Basic Model has been able to accommodate normal application and overapplica-
tion. It may seem that the factorial typology has exhausted all logical possibilities,
among which we did not identify underapplication. This conclusion comes as a
mild shock: how does the Basic Model deal with underapplication? This issue
will occupy us in the next section.

5.6.4 Underapplication in Japanese

Wilbur (1973) denes underapplication as the non-application of a phonological
process in the reduplicant even though this meets the structural condition. That
is, a rule fails to apply in the right environment.
239
Correspondence in reduplication

Before looking into an actual case of underapplication, let us see if we can

dene its general properties in terms of the Basic Model. First, some process
must be blocked in a context in which it normally applies. The fact that the pro-
cess normally applies in non-reduplication contexts implies that well-formedness
outranks IO-faithfulness. (Under the reverse ranking IO-faithfulness would simply
always block the process.) Second, the fact that underapplication is reduplication-
specic makes it an identity effect: its analysis will therefore involve high-ranked
basereduplicant identity. To our surprise, we nd that the combined rankings
result, once again, in the ranking schema of overapplication, stated in (104). To
distinguish overapplication from underapplication, we need something extra. This
something extra is another constraint that has the effect of blocking the change,
and must be an antagonist of the well-formedness constraint that triggers the
change:

(106) Ranking schema for underapplication

BR-Identity, Blocker-constraint
Trigger-constraint

IO-Faithfulness

Are we on the right track? In fact, this general argument seems to produce noth-
ing but a fatal inconsistency: since how can a change ever apply, if the constraint
that triggers it is itself dominated by a constraint blocking the change?
The answer takes into account the fact that the relevant segments appear twice
in a basereduplicant combination: once in the base, and once in the reduplic-
ant. Moreover, the phonological contexts in which these correspondents occur are
typically different. For example, let us assume that one segment (of a pair of
correspondents) is in a blocking context, while its counterpart is in a trigger-
ing context. By top-ranking BR-identity, the correspondents in reduplicant and
base must be identical. In such a situation, whichever segment wins can only
depend on the pushpull of the blocker- and trigger-constraints. Therefore, in
underapplication there is an undominated constraint blocking the change in either
the base, or the reduplicant. If this argument sounds too abstract, we will now
make it more concrete step by step in the analysis of an actual underapplication
case.

An example of underapplication due to an undominated blocker-constraint is the

pattern of reduplicated mimetics in Japanese (Mester and It 1989, McCarthy
and Prince 1995a, forthcoming). Japanese has a general phonological process by
which [
] alternates with [], depending on position in the word. In fact these
segments are allophones, since they are in complementary distribution. What is
relevant here is that [
] occurs word-initially, while [] occurs medially:

240
 5.6 Overapplication and underapplication in reduplication

(107) Complementary distribution of [

], [] in Japanese
#
VV
a.i
eta clogs b.i kai key
a.ii
iri duty b.ii oyuu to swim
a.iii
ai-koku foreign country b.iii koku-ai abroad

McCarthy and Prince (1995a, forthcoming) argue that the alternations are trig-
gered by two phonotactic well-formedness constraints. The rst rules out a voiced
velar oral stop:

(108) PostVcls
No voiced velar oral stops.

This registers as a markedness constraint the effects of a property of articulation:

voicing is difcult to maintain in an oral velar stop, due to the limited size of the
oral cavity (Ohala 1983, Vance 1987). In order to maintain voicing, air will be let
out of the nasal cavity, with as a result a velar nasal. Indeed languages occur that
have voiced oral stops [b] and [d], but not [
] (an example is Dutch). This is the
trigger-constraint, producing the impetus for nasalization of [
] into [] to
occur. PostVcls is essentially a context-free markedness constraint (in spite of
its name, suggesting some post-vocalic context).
The second markedness constraint involved in the allophonic pattern rules out
word-initial velar nasals:

(109) *[
No word-initial velar nasals.

The fact that [

] occurs word-initially in Japanese, shows that *[ dominates
PostVcls. Moreover, IO-faithfulness with respect to the feature [nasal] is totally
dominated by these phonotactic constraints, since the pattern is totally predictable
(allophonic).

(110) Allophonic variation of [

] and [] in Japanese
Blocker-constraint
Trigger-constraint
IO-Faithfulness
*[
PostVcls
Ident-IO(nasal)

This is just another instantiation of the allophonic ranking schema of chapter 1.

But it also instantiates a part of the ranking schema (106), containing the con-
guration of a blocker-constraint, a trigger-constraint, and IO-faithfulness. (We
will turn to identity effects in reduplication shortly.) This interaction is illustrated
by two tableaux:

241
Correspondence in reduplication

(111) Input: /
ara/ *[ PostVcls Ident-IO(nasal)

a. +
ara *
b. ara *! *

(112) Input: /ka

i/ *[ PostVcls Ident-IO(nasal)

a. ka
i *!
b. + kai *

In these tableaux, both velars are attributed to an input oral stop [

]. But in fact,
there is no evidence for this, and the presence of [
] or [] in the input is totally
irrelevant. (The tableaux would remain essentially the same if the inputs had been
/ara/ and /kai/, with only some irrelevant changes of violation marks in the
IO-faithfulness column.)
Let us now turn to reduplication. Voiced velar stop nasalization underapplies
in reduplicated mimetics, where both the base and the reduplicant have the voiced
stop [
], rather than the nasal []. Consider the examples in (113):

(113) Underapplication of voiced velar stop nasalization in reduplicated

mimetics
a.
ara-
ara rattle (*
ara-ara, *ara-ara)
b.
eji-
eji centipede (*
eji-eji, *eji-eji)
c.
era-
era laughing (*
era-era, *era-era)

With normal application of voiced velar stop nasalization, we would have

expected *[
ara-ara]. With overapplication, we would have expected *[ara-
ara]. Since there is no evidence for oral or nasal stops in the input, pattern (113)
may equally well be considered an overapplication of initial velar stop denasal-
ization, or an underapplication of velar stop nasalization. This does not affect the
argument, though, because the pattern involves all ingredients that are character-
istic of the underapplication schema (106). In sum, we have an interaction of the
following four constraint types:

(114) a. BR-Identity: Ident-BR(nasal)

b. Blocker-constraint: *[
c. Trigger-constraint: PostVcls
d. IO-faithfulness: Ident-IO(nasal)

We have just seen a partial ranking in (110) of all constraints except BR-
identity. Arguably it is BR-identity that makes voiced velar stop nasalization
242
 5.6 Overapplication and underapplication in reduplication

underapply in medial position in reduplicated mimetics. In (115), the sources of

both occurrences of [
] in [
ara-
ara] are attributed to their respective sources.

(115) [
] due to Ident-BR(nasal)
PostVcls

ara-
ara

[
] due to *[
PostVcls

Accordingly we add the factor of BR-identity to the analysis, in the form of

undominated Ident-BR(nasal). The four constraints in (114) are then ranked
according to the schema of underapplication (106):

(116) Ranking for underapplication of voiced velar stop nasalization in

Japanese
BR-Identity, Blocker-
Trigger-
IO-Faithfulness
constraint constraint
Ident-BR(nasal), *[
PostVcls
Ident-IO(nasal)

This ranking is illustrated by the following tableau:

(117) Input: /
ara + red/ Ident-BR *[ PostVcls Ident-IO
(nasal) (nasal)

a. +
ara-
ara **
b. ara- ara *! *
c.
ara- ara *! *

The normal application candidate (117c) fatally violates Ident-BR(nasal), while

the overapplication candidate (117b) suffers from a word-initial [], fatally
violating *[. Satisfaction of both high-ranked constraints happens at the expense
of a double violation of PostVcls in the optimal candidate (117a), hence this is
an underapplication.

Now that we have identied pairs of well-formedness constraints to be relevant

to underapplication, we must return to overapplication, and see how this nding
relates to this case. In fact, the situation is completely symmetrical. The difference
between an overapplication and an underapplication resides in the relative posi-
tion of the blocker-constraint and the trigger-constraint. The resulting ranking
schemata are given in (118):
243
Correspondence in reduplication

(118) a. Ranking schema for underapplication

BR-Identity, Blocker-constraint
Trigger-constraint

IO-Faithfulness
b. Ranking schema for overapplication
BR-Identity, Trigger-constraint
Blocker-constraint

IO-Faithfulness

To show how the expanded overapplication schema (118b) subsumes the case of
Malay, we repeat it below (from 98):

(119) Overapplication of Malay nasalization in the expanded schema

BR-Identity, Trigger-
Blocker-
IO-Faithfulness
constraint constraint
Ident-BR(nasal), *NVoral
*Vnasal
Ident-IO(nasal)

The ranking schemata (118) summarize the results of our discussion of over-
application and underapplication. As we have seen in the case of Japanese, it is
sometimes not clear whether the special behaviour of reduplicants with respect
to phonological processes is to be referred to as overapplication or underappli-
cation. Whenever the phonology involved has an allophonic character, both terms
seem to t equally well. However, it is not so much the terminology that matters,
but the insights expressed by the analysis: the special behaviour of reduplication
is a kind of identity effect, and the correspondence model takes this identity effect
to be the heart of the matter.
But the ultimate (and most dramatic) consequence of the output correspond-
ence approach to reduplication is yet to be discussed: the case of back-copying
of reduplicant phonology into the base. To this we will now turn.

5.6.5 The case for parallelism: base copies reduplicant in

Southern Paiute
In all cases of reduplicative identity that were discussed until here, the over-
application or underapplication occurred in the reduplicant. However, nothing
in the model predicts that this should be the case. If phonology may overapply
or underapply because of BR-identity, then why not in the base? The clearest
evidence for the parallel OT model of reduplication is that of BR-identity induc-
ing an overapplication or underapplication in the base. This is what we will call
reduplicative back-copying:

(120) Reduplicative back-copying: the overapplication or underapplica-

tion of some process in the base, under pressure to preserve identity
with the reduplicant.
244
 5.6 Overapplication and underapplication in reduplication

The logic of the situation is identical to that of overapplication and underapplica-

tion in the reduplicant, which we have discussed until here: it is the preservation
of identity of base and reduplicant that overrules a normal application of some
process. What makes a process overapply or underapply to the base, rather
than the reduplicant, depends on the relative ranking of the blocker and trigger
well-formedness constraints governing each of the correspondents segments in
the base and the reduplicant.

Back-copying effects from reduplicant to base seem to occur (upon preliminary

analysis, at least) in several languages, as McCarthy and Prince (1995a, forthcom-
ing) show. We focus on a case from Southern Paiute (Sapir 1930). This language
has an allophonic distribution of the labio-velar glide [w] and the labialized nasal
[w], which is contextually conditioned essentially like the alternation of [
] and
[] in Japanese. That is, the glide [w] occurs in the onset of a word, but it
alternates with the nasal [w] in medial positions.

(121) Alternations of [w], [w] in Southern Paiute

#w V\w V
a.i waai to shout a.ii tV]-waai to give a good shout
b.i waiTa- to have a council b.ii na]v -WaiTap] council (of chiefs)

The analysis of the allophonic pattern resembles that of Japanese: again two
phonotactic constraints dominate an IO-faithfulness constraint. One phonotactic
constraint blocks the nasalization of /w/ in word-initial positions, and is stated as:

(122) *[\w
No word-initial labio-velar nasals.

Presumably this constraint can be simplied to the anti-initial-velar-nasal con-

straint *[ that is undominated in Japanese and English. Then the shared labiality
of both alternants, [w] and [w], must be due to undominated Ident-IO(labial),
preserving input labiality.20
The second phonotactic constraint required in this pattern is:21

(123) *VwV
No intervocalic labio-velar glides.

This constraint plays the role of trigger-constraint for intervocalic nasalization

of /w/. It must therefore dominate Ident-IO(nasal).
20
The fact that input /w/ alternates with [w], rather than another labial consonant such as [b]
or [m], may be attributed to faithfulness constraints as well.
21
*VwV may be reducible to a more general ban on post-vocalic glides, a kind of minimal
sonority distance effect. See Rosenthall (1994) for discussion of a similar effect in Lenakel.

245
Correspondence in reduplication

Next, observe that reduplication deviates from the alternation pattern of (121),
in the sense that roots beginning with [w] unexpectedly retain this glide in an
intervocalic position after a reduplicant.

(124) Underapplication of medial nasalization in the base in Southern

Paiute
a.i wVV- vulva a.ii wV-wVxVA- vulvas (obj.)
b.i wai- several enter b.ii wa-waxpia all entered

This is a case of underapplication: a phonological process is blocked in a context

where it would normally apply, due to the inhibiting effect of BR-identity. The
analysis is not too difcult, and essentially mirrors the analysis of Japanese (116):
(125) Underapplication of medial nasalization in the base in Southern
Paiute
BR-Identity, Blocker-
Trigger-
IO-Faithfulness
constraint constraint
Ident-BR(nasal) *[w
*VwV
Ident-IO(nasal)
With undominated BR-identity, the initial consonants of reduplicant and base
must both be glides [w], or both be nasal [w]. The initial consonants of reduplic-
ant and base are in different phonological contexts, each of which falls in the
scope of different phonotactic constraints, *[w and *VwV. Hence, the outcome
(double [w] or double [w]) depends on the relative ranking of these constraints.
Obviously *[w must take precedence over *VwV since the double [w] pattern is
what surfaces.
Tableau (125) illustrates this ranking:

(126) Input: /red + wVV-A/ Ident-BR *[w *VwV Ident-IO

(nasal) (nasal)

a. + wV-wVxVA- *
b. w V-w VxVA- *!
w
c. wV- VxVA- *!

This tableau is almost identical to that of underapplication in Japanese reduplicat-

ive mimetics (117). The single difference is that in Southern Paiute effects of
reduplicative identity unambiguously appear on the base, whereas in Japanese the
identication of the reduplicant and base was unclear reduplication being total.
Underapplication of a phonological process to the base is an important nding,
and we should rule out any alternative analyses. Crucially, underapplication of

246
 5.6 Overapplication and underapplication in reduplication

medial nasalization depends on identity, rather than on some obscure morpholog-

ical property of reduplication. The additional data below show that intervocalic
nasalization is perfectly productive in reduplication, as long as both the reduplic-
ant and base are in intervocalic position. Such a double intervocalic context is
easily created: when a vowel-nal prex precedes the reduplicant, labio-velar
nasals do occur in both reduplicant and base:

(127) Normal application of medial nasalization in the base in Southern

Paiute
a.i wVn- to stand
a.ii ya-w-w nT- while standing and holding

Note that this is precisely the result that is expected under the analysis given
above: since none of the corresponding consonants in reduplicant and base are in
word-initial position, the constraint *[w is never active, so that the next-lower
(trigger-)constraint *VwV determines the outcome.
If this argument is indeed valid, then these data constitute a severe problem for
a derivational theory, which is based on copying. The copy rule always copies a
string of segments of the base, upon the reduplicant. Copying of segments of the
reduplicant into the base is a complete mystery under this theory.
To wind up this section, let us now generalize the ndings for Southern Paiute
to general ranking schemata for back-copying. For the case of underapplication
to the base, the schema is (128):

(128) Ranking schema for back-copying underapplication

BR-Identity, Blocker-constraint
Trigger-constraint

IO-Faithfulness
where the triggering context is met by the base (rather than the
reduplicant).

The overapplication case arises by reversing the ranking of the trigger and
blocker:

(129) Ranking schema for back-copying overapplication

BR-Identity, Trigger-constraint
Blocker-constraint

IO-Faithfulness
where the blocking context is met by the base (rather than the
reduplicant).

This concludes the section on overapplication and underapplication. The nal

section of this chapter provides an overview of Correspondence Theory, to the
point where we have seen it now.

247
Correspondence in reduplication

5.7 Summary of Correspondence Theory

Inputoutput faithfulness and basereduplicant identity are notions of a general
theory of correspondence which includes constraints requiring identity between
elements in pairs of representations. Thus far in this book we have applied Cor-
respondence Theory to the pairs of representations in (130):

(130) Scope of Correspondence Theory (so far in this book)

a. Input and output -IO faithfulness
b. Base and reduplicant -BR reduplicative identity

The notion of correspondence is dened such that it covers both situations

(McCarthy and Prince 1995a, forthcoming):

(131) Correspondence
Given two strings S1 and S2, related to one another as input output,
basereduplicant, etc., correspondence is a relation from the ele-
ments of S1 to those of S2. Elements S1 and S2 are referred
to as correspondents of one another when .

Given this general denition of correspondence, the correspondence constraints

(Max, Dep, etc.) must be relativized for different values of S1 and S2. For example,
we saw in section 5.2.3 that it is necessary for empirical reasons to distinguish Max-
IO from Max-BR. In fact, Maximality now becomes a family of constraints,
whose members are distinguished by their values of S1 and S2 (InputOutput,
BaseReduplicant, etc.). Before discussing the actual members of these constraint
families, we have to repeat some basic assumptions of Correspondence Theory.
The output-generating component Gen supplies pairs of strings S1 and S2, as
well as any correspondence relationships holding between the elements of these
strings. That is, correspondence is not a relationship that is established by con-
straints, but rather one that is evaluated by constraints. Constraints evaluating the
Gen-supplied correspondence relationships are crucially violable, giving rise to
optimal candidates displaying imperfect correspondence relations. These imper-
fections appear as deletion, epenthesis, feature change, metathesis, partial
reduplication, skipping, etc.

Following McCarthy and Prince (1994b, forthcoming), I will introduce the

notions of domain of and range of . Dened formally, these are stated as
in (132):

(132) Domain(): for a relation AB, x Domain() iff x A and

y B such that xy
Range(): y Range() iff y B and x A such that xy
248
 5.7 Summary of Correspondence Theory

In these denitions, a structure Si is encoded as a set of elements. The relation

on elements of structures (S1, S2) is dened as a subset, or any subset, of S1 S2.
Paraphrasing the denitions, we say that x (an element of A) is in the domain of
if and only if there is some y (an element of B), such that x and y are related
by . Accordingly, y (an element of B) is in the range of if and only if there
is some x (an element of A), such that x and y are related by .
For example, the rst correspondence constraint family, Maximality, has as
its domain the set of elements in S1 (either the input or the base), and as its range
the set of elements standing in correspondence, which is a subset of S2 (either the
output or the reduplicant):

(133) Maximality
Every element of S1 has a correspondent in S2.
Domain() = S1
Members: S1 S2 Effect:
Max-IO Input Output No deletion of segments
Max-BR Base Reduplicant Total reduplication

Max-IO was introduced in chapter 1 as the anti-deletion constraint. In this

chapter we have seen another important member of the Maximality family,
Max-BR, the constraint enforcing total reduplication. (Chapter 6 will extend the
Maximality family to still other S1S2 pairs, such as Base and Truncated
form.)

The second family of constraints, Dependence, we have come to know by its

most prominent member Dep-IO, the anti-epenthesis constraint.

(134) Dependence
Every element in S2 has a correspondent in S1.
Range() = S2
Members: S1 S2 Effect:
Dep-IO Input Output No epenthesis of segments
Dep-BR Base Reduplicant No base-alien segments in the
reduplicant

Dep-IO is the anti-epenthesis constraint (which was motivated at length in

chapter 3). Its counterpart in basereduplicant identity is Dep-BR, the constraint
militating against base-alien segments in the reduplicant.

Third, the presence or absence of a corresponding element must be distinguished

from its featural identity, which is evaluated by yet another constraint family:
249
Correspondence in reduplication

(135) Identity[F]
Correspondent segments have identical values for feature [F].
If and is [F], then is [F]
Members: S1 S2 Effect:
Ident-IO[F] Input Output No featural changes
Ident-BR[F] Base Reduplicant No featural discrepancy
between R and B
By varying features for [F], different members of the Identity family arise.
Research in progress presents evidence for certain extensions of Identity. In its
current formulation, Identity is symmetrical for S1 and S2. For a pair of corres-
ponding segments, values of [F] are simply compared, without checking whether
discrepancies are due to deletion or insertion of [F]. Pater (forthcoming), as
reviewed in chapter 2, argues for an asymmetrical format of Identity on the
basis of segmental fusions, distinguishing Ident-IO[F] from Ident-OI[F].
Another extension of Identity distinguishes positive and negative values of the
same feature, Ident[+F] and Ident[F] (McCarthy and Prince forthcoming).
Finally, Identity is segment-based in the following sense: featural identity is
evaluated indirectly through corresponding segments, instead of directly through
corresponding features. A strictly segmental view of Identity may have to be
abandoned in favour of a featural view, to account for autosegmental phenomena,
such as oating features and featural stability. This would involve a distinction
between Max[F] and Dep[F] (see Zoll 1996).

Fourth, the constraint militating against medial epenthesis and /or deletion:
(136) (I-, O-) Contiguity
The portion of S1 standing in correspondence forms a contiguous
string, as does the correspondent portion of S1.
I-Contig (No Skipping) Domain() is a single contiguous string
in S1
O-Contig (No Intrusion) Range() is a single contiguous string
in S2
Members: S1 S2 Effect:
Contig-IO Input Output No medial epenthesis or deletion of
segments
Contig-BR Base Reduplicant No medial intrusion or skipping in
reduplicant
McCarthy and Prince (forthcoming) distinguish two forms of this constraint, I-
Contig and O-Contig, differing as to which string (S1 or S2) is taken as basic
with respect to the other. Accordingly violations of Contig are of two types,
250
 5.7 Summary of Correspondence Theory

called skipping and intrusion. In the basereduplicant domain, Contig-BR is

respected if the reduplicant copies a contiguous substring of the base, for example
in the reduplication patterns (32). It is violated if the reduplicant contains a
medial gap as compared to the base (skipping, cf. Sanskrit, section 5.2.4.2), while
in the case of intrusion, the reduplicant contains a medial segment not present
in the base. Likewise, Contig-IO demands that inputoutput mappings involve
contiguous substrings. It is violated by medial deletion (skipping) and medial
epenthesis (intrusion). Contig-IO is motivated by languages which prohibit
morpheme-internal epenthesis (e.g. Chukchee; Kenstowicz 1994b, Spencer 1994)
and by languages which restrict deletion to segments in morpheme-peripheral
positions (Diyari; McCarthy and Prince 1994a).

Fifth, Linearity, the anti-metathesis constraint:

(137) Linearity
S1 is consistent with the precedence structure of S2, and vice versa.
Let , S1 and , S2
If and , then
< iff ( < )
Members: S1 S2 Effect:
Linear-IO Input Output No metathesis
Linear-BR Base Reduplicant No segment reversal in
reduplicantbase
We have seen the faithfulness version of this constraint Linear-IO active in
chapter 2 in Paters analysis of coalescence in Indonesian. However, McCarthy
and Prince attribute the no coalescence function of Linearity to a separate
faithfulness constraint, Uniformity (no element of S2 has multiple correspond-
ents in S1). See also McCarthy (1995a) and Hume (1995).

Sixth, Anchoring, the constraint which is responsible for edge-in mapping:

(138) {Right, Left}Anchoring
Any element at the designated periphery of S1 has a correspondent
at the designated periphery of S2.
Let Edge(X, {L, R}) = the element standing at the Edge = L,
R of X.
Right-Anchor. If = Edge(S1, R) and = Edge(S2, R) then
.
Left-Anchor. Likewise, mutatis mutandis.
Members: S1 S2 Effect:
Anchor-IO Input Output No epenthesis or deletion at edges
Anchor-BR Base Reduplicant Edge-in association
251
Correspondence in reduplication

In chapter 3, we saw evidence for Anchor-IO, blocking epenthesis at morpheme

edges in Lenakel. Its counterpart in reduplication, Anchor-BR, enforces edge-in
association. It requires that the initial (or nal) segment of the reduplicant stand
in correspondence with the initial (or nal) segment in the base. A specication
of the edge (L/R) in Anchoring serves to distinguish prexing and sufxing
reduplication. (The former ranks L-Anchor
R-Anchor, the latter has the
reverse ranking.) McCarthy and Prince (forthcoming) argue that anchoring
should subsume Generalized Alignment. Their unication builds on the idea that
edges of constituents can also be matched by the correspondence of segments
standing at edges, using the Anchoring format. Correspondence of peripheral
segments in constituents (for example, Stem and PrWd) within a single output
form implies that correspondence becomes a reexive relation. McCarthy and
Prince give a hypothetical case (b.ta) where the left edge of the foot and the
head syllable align because b and its correspondent (which is, reexively, b) are
initial in both.

SUGGESTIONS FOR FUR THER READING

Pre-OT studies of prosodic morphology

Broselow, Ellen and John McCarthy (1983) A theory of internal reduplication.
The Linguistic Review 3. 2589.
Clements, George N. (1985a) The problem of transfer in nonlinear morphology.
Cornell Working Papers in Linguistics 7. 3873.
Marantz, Alec (1982) Re reduplication. Linguistic Inquiry 13. 43583.
McCarthy, John (1981) A prosodic theory of non-concatenative morphology.
Linguistic Inquiry 12. 373 418.
McCarthy, John and Alan Prince (1986) Prosodic morphology. Ms., University of
Massachusetts, Amherst and Brandeis University.
Steriade, Donca (1988a) Reduplication and syllable transfer in Sanskrit and else-
where. Phonology 5. 73155.

Correspondence Theory

Reduplication
Blevins, Juliette (1996) Mokilese reduplication. Linguistic Inquiry 27. 52330.
Downing, Laura J. (forthcoming) Verbal reduplication in three Bantu languages.
To appear in Kager, van der Hulst, and Zonneveld.
Klein, Thomas B. (1997) Output constraints and prosodic correspondence in
Chamorro reduplication. Linguistic Inquiry 28. 70715.
252
 5.7 Summary of Correspondence
5.7 Further Reading
Theory

McCarthy, John and Alan Prince (1994a) The emergence of the unmarked:
Optimality in Prosodic Morphology. In M. Gonzlez (ed.), Proceedings of
the North-East Linguistics Society 24. 33379.
(forthcoming) Faithfulness and identity in prosodic morphology. To appear in
Kager, van der Hulst, and Zonneveld.
Myers, Scott and Troi Carleton (1996) Tonal transfer in Chichewa. Phonology 13.
3972.
Urbanczyk, Suzanne (1996) Morphological templates in reduplication. In K.
Kusumoto (ed.), Proceedings of the North-East Linguistics Society 26. 425
40.
(forthcoming) Double reduplications in parallel. To appear in Kager, van der
Hulst, and Zonneveld. [ROA-73, http://ruccs.rutgers.edu/roa.html]

Coalescence
Lamontagne, Greg and Keren Rice (1995) A correspondence account of coales-
cence. In Beckman, Walsh Dickey, and Urbanczyk. 21123.
Pater, Joe (forthcoming) Austronesian nasal substitution and other Nt effects.
To appear in Kager, van der Hulst, and Zonneveld. [ROA-92,
http://ruccs.rutgers.edu/roa.html]

Metathesis
Hume, Elizabeth (1995) Beyond linear order: prosodic constraints and C/V
metathesis. Proceedings of the Formal Linguistics Society of the Midwest 6.
Bloomington: Indiana University Linguistics Club.
McCarthy, John (1995a) Extensions of faithfulness: Rotuman revisited. Ms.,
University of Massachusetts, Amherst. [ROA-110, http://ruccs.rutgers.edu/
roa.html]

Featural identity
Alderete, John, Jill Beckman, Laura Benua, Amalia Gnanadesikan, John McCar-
thy, and Suzanne Urbanczyk (1997) Reduplication with xed segmentism.
Ms., University of Massachusetts, Amherst. [ROA-226, http://ruccs.rutgers.edu/
roa.html]
Gnanadesikan, Amalia (1995) Markedness and faithfulness constraints in
child phonology. Ms., University of Massachusetts, Amherst. [ROA-67,
http://ruccs.rutgers.edu/roa.html]
Lombardi, Linda (1995b) Why Place and Voice are different: constraint interac-
tions and feature faithfulness in Optimality Theory. Ms., University of Mary-
land, College Park. [ROA-105, http://ruccs.rutgers.edu/roa.html]
253
Correspondence in reduplication

McCarthy, John (1997) Process specic constraints in Optimality Theory. Lin-

guistic Inquiry 28. 23151.
Myers, Scott (1997a) OCP effects in Optimality Theory. Natural Language and
Linguistic Theory 15. 84792.

Miscellaneous
Bat-El, Outi (1996) Selecting the best of the worst: the grammar of Hebrew
blends. Phonology 13. 283328.
It, Junko, Yoshihisa Kitagawa, and R. Armin Mester (1996) Prosodic faithfulness
and correspondence: evidence from a Japanese argot. Journal of East Asian
Linguistics 5. 21794.

EXERCISES

1 Reduplication in Axininca Campa

Consider a sample of reduplicated forms of Axininca Campa (Payne 1981, Spring
1990, McCarthy and Prince 1995b) below.

Sufxing reduplication in Axininca Campa

(i.a) kawosi kawosi-kawosi bathe
(i.b) koma koma-koma paddle
(ii.a) osampi osampi- sampi ask
(ii.b) osakina osakina- sakina write

Observe that the reduplicant sufx copies all segments of consonant-initial roots
(i). But roots beginning with a vowel (ii) are not copied entirely here the
reduplicant omits the initial vowel. The question then is, what constraint interac-
tion accounts for this pattern.

a. Identify the well-formedness and BR-identity constraints that are

actively involved in shaping the reduplicant, and rank them.
b. Consult chapter 3 (section 3.3.4.1) for an analysis of consonant epen-
thesis in Axininca Campa, involving Max-IO and Dep-IO. Then inte-
grate the analysis of epenthesis with the reduplication analysis.
c. Explain how your ranking correctly excludes *sampi- sampi and
*tosampi- tosampi (with epenthetic t).
d. Next, consider another ill-formed candidate, *osampi- tosampi , with
epenthesis in the reduplicant, but not in the base. Integrate the
responsible constraint in your ranking.

254
 5.7 Summary of Correspondence
5.7 Exercises
Theory

e. What excludes the inxational candidate *osamp- osamp -i? Inte-

grate the responsible constraint in your ranking.
f. Present all results in a tableau of osampi- sampi .

2 Reduplication in Oykangand
The prexing reduplication pattern of Oykangand (Sommer 1981, McCarthy and
Prince 1986) is exemplied below:

Oykangand prexing reduplication

(i.a) eder ed-eder rain
(i.b) i
un i
-i
un go
(ii) al
al al
-al
al straight

Assume that the reduplicants nal consonant syllabies as the onset for the base-
initial vowel (e.d-e.der, i.w -i.wun, al.w -al.wal). This syllabication poses the ana-
lytic challenge.

a. Identify the constraints determining the reduplicants size and syl-

labication, and rank them.
b. Explain how your ranking excludes * al.wal -al.wal and *a.l-al.wal.
c. What excludes the inxational candidates *e-der-der and *i- wun -
wun? Integrate the responsible constraint in your ranking.
d. Present all results in a tableau.

3 Inxing reduplication in Pangasinan

Pangasinan has various reduplication patterns (Benton 1971, McCarthy and
Prince 1986), among which are the following:

Pangasinan inxing reduplication

(i.a) sakey saksakey one
(i.b) talo taltalora three
(ii.a) apat apatpatira four
(ii.b) anem anemnemira ve

This is inxing reduplication, very similar to the pattern of Timugon Murut (dis-
cussed in section 5.4.2). The patterns of Pangasinan and Timugon Murut differ
minimally, though.

a. Accurately describe this difference in prose.

b. Starting from the constraint ranking of Timugon Murut (section
5.4.2), can you change this minimally so that the Pangasinan pattern
is obtained?

255
Correspondence in reduplication

c. Explain how your analysis excludes *a- pa -pa.ti.ra, * ap -a.pa.ti.ra,

and * a.p -a.pa.ti.ra.
d. Present all results in a tableau.

4 Inxing reduplication in Mangarayi

Mangarayi has inxing reduplication in plurals (Merlan 1982, McCarthy and
Prince 1986, 1993a), as exemplied below:

Mangarayi inxing reduplication

(i.a)
abuji
ababuji old person
(i.b) yira
yirira
father
(ii.a) jim
an jim
im
an knowledgeable person
(ii.b) wa
ij wa
a
ij child

The special feature of this inxing pattern is the rst consonant of the root, which
always immediately precedes the reduplicant.

a. Which constraint interaction accounts for the position of the reduplic-

ant immediately after the rst consonant of the root?
b. Rank the constraints determining the reduplicants size, and integrate
this ranking with your earlier results. Hint: reconsider your analysis
of Oykangand (see exercise 2).
c. Explain how your analysis excludes * jim -jim.wan, *j-im-im.wan, and
*j-i.m-im.wan.
d. Present all results in a tableau.

5 A factorial typology of reduplicative systems

a. Construct a factorial typology of the ve constraints Align-Red-L,

Max-BR, No-Coda, Onset, and red=. Identify any rankings that
produce identical reduplicative systems.
b. How many reduplicative systems does your typology predict? Char-
acterize each of the predicted systems in prose, making sure that your
description is sufciently general as well as precise enough to cap-
ture all distinctions.
c. Instantiate the factorial typology with reduplication systems (dis-
cussed in this chapter or introduced in preceding exercises). Do you
nd any gaps in the typology?

256
 6
Output-to-output correspondence

6.1 Introduction
In this chapter we will consider similarities in the shapes of morphologically
related words which are not due to common inputs. We will pursue the idea that
such similarities involve the notion of output-to-output correspondence, the
maximization of phonological identity between morphologically related output
forms.
A theoretical precursor of OO-correspondence, the notion of paradigm uni-
formity, enjoyed a long tradition in pre-generative linguistics (see for example
Kury(owicz 1949). This notion played a modest role in generative phonology (but
see Kiparsky 1982a), where similarities between morphologically related forms
were attributed to derivational means, in particular the phonological cycle.
Recently, paradigm uniformity has been revived in OT by Benua (1995), Flem-
ming (1995), McCarthy (1995a), Burzio (1996), Kenstowicz (1996), Steriade
(1996), and others. Disregarding the technical differences between these pro-
posals (referred to as either paradigm uniformity, uniform exponence, base-
identity, or OO-correspondence), we will subsume all under the general notion
of OO-correspondence.
OO-correspondence elaborates on the notion of reduplicative identity, dis-
cussed in chapter 5. The reduplicant, part of a surface form, is subject to cor-
respondence constraints requiring identity with its base, which is part of the
same surface form. Interactions between reduplicantbase-identity constraints and
markedness constraints turned out to be a variation on the major type of con-
straint interaction in OT: that between faithfulness and markedness.
For example, we discovered a deep similarity between constraints militating
against the deletion of input segments, and those enforcing total reduplication.
Both constraints require that elements in one string (input or reduplicants base)
match elements in another string (output or reduplicant). These (and other) sim-
ilarities were captured by a small set of generalized correspondence constraints:
Maximality, Dependence, Identity, Linearity, Contiguity, and Anchoring.
We also learned that correspondence offers an explanation of underapplication
257
Output-to-output correspondence

and overapplication in reduplicative morphology, based on interactions of

markedness constraints and correspondence constraints: IO-faithfulness and BR-
identity. This analysis produced new evidence for parallelism, based on the
observation that identity effects may back-copy to the base, as in the case of
Malay vowel nasalization.
In this chapter, Correspondence Theory will be extended to fresh empirical
domains, again involving correspondence between outputs, in particular to mor-
phological truncation (Benua 1995, McCarthy 1995a) and stem-based afxation
(Burzio 1994, 1996, Benua 1995, Kenstowicz 1996, Steriade 1996). A new and
highly interesting aspect of these phenomena, as compared to reduplication, is
that they involve correspondence between separate output forms: words. In redu-
plication, BR-correspondence is dened as a relation between pairs of repres-
entations that are co-present in a single output form. But co-presence no longer
holds for truncation, nor for stem-based afxation, where each of the representa-
tions involved is a stem, a free-standing element of a morphological paradigm.
Pairs of forms to be related by OO-correspondence can be either a truncated form
and the stem on which it is based, or an afxed form and the non-afxed stem
on which it is based.

(1) Extending the scope of Correspondence Theory

a. Lexicalsurface (IO) Chapters 12
b. Basereduplicant (BR) Chapter 5
c. Basetruncated stem (BT) This chapter, section 6.2
d. Stemafxed stem (BA) This chapter, sections 6.3 4

The generalization that covers truncation and stem-based afxation is that in both
cases the derived form (the truncated form, or the afxed form) copies a phono-
logical property of its base (the non-truncated output form, or the stem in a
stem-based afxation). In both cases, the base is a free-standing form an output.
This extension of Correspondence Theory to relations between free-standing
output forms is the topic of this chapter. We will learn that OO-identity effects
involve both the underapplication and overapplication of phonology to a morpho-
logically derived (truncated or afxed) form. As we concluded earlier for redupli-
cation, these effects result from a small number of ranking schemata: interactions
of IO-faithfulness, OO-identity, and markedness constraints.

This chapter is organized as follows. First, in section 6.2, we will look into
identity effects in truncation, on the basis of case-studies of English nicknames
and Icelandic truncation. Section 6.3 will extend Correspondence Theory to rela-
tions between an afxed word and the stem on which it is based, considering
English data. Finally, section 6.4 will show how OO-correspondence serves to
258
 6.2 Identity effects in truncation

reanalyse data that were analysed by the phonological cycle in derivational theory.
Here we will focus on examples from Palestinian Arabic, keeping an eye on the
advantages of the correspondence model over a serial (cyclic) analysis.

6.2 Identity effects in truncation

6.2.1 Introduction
Truncation is a general term for any morphological category that is derived by a
systematic phonological shortening of a basic form. A cross-linguistically com-
mon type of truncation is the formation of hypocoristics by shortening a name
(nicknames). It has been observed for various languages that truncations preserve
phonological properties of their base, even though these properties need not be
contextually motivated in the truncated form (Anderson 1975, Benua 1995,
McCarthy 1995a). We will discuss Benuas idea that maximization of phonolog-
ical identity between a truncated form and its base involves OO-correspondence,
a notion that is modelled after base-reduplicant identity.
This section is organized as follows. Section 6.2.2 will outline the analysis of
morphological truncation in Correspondence Theory. Discussion will focus on
examples from American English, analysed by Benua (1995). In section 6.2.3 this
analysis will be generalized into a proposal of a Basic Model, analogous to that
for reduplication in chapter 5. Finally, section 6.2.4 will discuss truncation in
Icelandic, again following the analysis by Benua (1995).

6.2.2 Truncation in American English

A set of examples motivating OO-correspondence in truncation is given in (2).
All examples in the bottom row are truncations of names (into nicknames) in
American English:
(2) a.i Larry [l.ri] b.i Harry [h.ri] c.i Sarah [s.r]
a.ii Lar [lr] b.ii Har [hr] c.ii Sar [sr]
Kahn (1976) and others have observed that the vowels in the truncated forms are
exempt from certain phonological regularities that govern underived words. In
particular, American English lacks the low front vowel [] in syllables closed by
/r/, realizing this vowel by its low back allophone [], as in car (3a). In contrast,
[] freely occurs in syllables closed by other consonants than /r/ (3b), as well as
syllables in which /r/ is not tautosyllabic (3c).
(3) __r] __C] __] rV
a.i car [kr] b.i kat [kt] c.i carry [k.ri]
a.ii mar [mr] b.ii map [mp] c.ii marry [m.ri]
a.iii lark [lrk] b.iii last [lst] c.ii Larry [l.ri]
259
Output-to-output correspondence

Non-truncated words in American English never contain [] in a /r/-closed syl-

lable: *[r]. Nevertheless, all truncated forms in (2) have precisely this sequence,
for example Lar [lr]. Apparently truncated forms copy the vowel of their base,
in spite of the allophonic pattern (which would predict the truncated form *[lr]).

6.2.2.1 An OO-correspondence analysis

Benua (1995) argues that the vowel in the truncated form Lar [lr] is related to
the vowel in its base Larry [l.ri] by a correspondence relation. This relation is
marked in (4) by a line connecting the vowels in the truncated form and the base:
(4) Larry [l.ri] Base

Lar [lr] Truncated form

The requirement that the corresponding vowels in the base and truncated form are
identical in their feature content produces a kind of underapplication of the
allophonic distribution in the truncated form. The realization of the low vowel as
[] before tautosyllabic /r/ fails to occur, as that would obliterate the identity
between the truncated form and its base. This is fully analogous to the under-
application effects in reduplication in chapter 5. This analogy is strikingly sup-
ported by the constraint interaction underlying the pattern, as we will see.
First, the allophonic distribution of front [] and back [] instantiates a schema
for allophonic variation seen in chapter 1:

(5) Schema for allophonic variation

Contextual markedness
Context-free markedness
Faithfulness

This ranking schema is instantiated by the following three constraints:1

(6) a. *r]
No // before tautosyllabic /r/.
b. *Back-low
Low vowels are front.
c. Ident-IO(back)
Let be a segment in the input, and be a correspondent of
in the output. If is [back], then is [back].

1
Constraints (6ab), in spite of their idiosyncratic appearances, are in fact phonetically
grounded. First, *r] captures the well-known centralization of vowels before [r], here rela-
tivized to the domain of the syllable. Second, *Back-low is not mentioned by Benua (1995),
who gives only a partial analysis of the allophonic pattern. *Back-low functions to support the
lexical contrast in low vowels along the frontback dimension, which is otherwise based on
length (//~// ). This makes sense as a language-specic instantiation of a no-perceptual-
confusion principle.

260
 6.2 Identity effects in truncation

The irrelevance of the input is shown in the following tableaux of [kr] and
[l.ri]:

(7) Input: /kr/ *r] *Back-low Ident-IO(back)

a. [kr] *!
b. + [kr] * *

(8) Input: /lri/ *r] *Back-low Ident-IO(back)

a. + [l.ri] *
b. [l.ri] *!

To state underapplication as a constraint interaction, we need a constraint

Ident-BT(back), requiring featural identity of backness between the base and its
truncated form:

(9) Ident-BT(back)
Let be a segment in the base, and be a correspondent of in the
truncated form. If is [back], then is [back].

The truncated form [lr] shows that Ident-BT(back) dominates *r], and by
transitivity, all constraints dominated by it:

(10) Underapplication of backness adjustment in truncated forms

Ident-BT(back)
*r]
*Back-low
Ident-IO(back)

This interaction is illustrated by the tableau of the truncated form [lr]:

(11)
Input: /lri/ Ident-Bt(back) *r] *Back-low Ident-IO(back)
Base: [l.ri]

a. [lr] *! *
b. + [lr] *

In sum, an allophonic process which adjusts the backness of vowels in /r/-closed

syllables is blocked in the truncated form, thereby maximizing phonological iden-
tity with its base. This blocking is highly similar to underapplication effects in
reduplication. Both increase the shape identity between surface forms, either
reduplicantbase or truncated form and base.
261
Output-to-output correspondence

This resemblance between reduplication and truncation has been stated in

terms of interactions between identity and markedness in (12):

(12) Underapplication in reduplication and truncation

a. In reduplication: BR-Identity
Markedness
IO-Faithfulness
b. In truncation: BT-Identity
Markedness
IO-Faithfulness

Overapplication, the opposite situation of underapplication, is also attested in the

truncation patterns of the worlds languages. We will consider an example of
overapplication in a different type of phenomenon, that of stem-based afxation,
in section 6.3.

6.2.2.2 A derivational analysis

A derivational analysis of the underapplication effect in truncation faces severe
problems. To derive truncated forms, a problematic kind of ordering has to be
called on. Rule ordering must be such that a morphological rule (truncation)
follows an allophonic rule (specifying backness in low vowels, with reference to
tautosyllabic /r/).

(13) /lri/ Lexical form

l.ri Syllabication, Backness Adjustment / __ r]
lr Truncation
[lr] Surface form

If the ordering were reverse (i.e. feeding), then Backness Adjustment would
have a chance to apply to the output of truncation, which is a closed syllable,
incorrectly predicting *[lr]. This ordering of allophonic rules before morpholog-
ical rules is problematic to the standard version of derivational theory according
to which all morphological rules are in the lexicon, and all allophonic rules in the
post-lexical component (Lexical Phonology, Kiparsky 1982b). But even if it were
assumed that Backness Adjustment is a lexical rule (that is, in spite of its allo-
phonic character), the derivational analysis would still face a second problem.
That is, it cannot explain why Backness Adjustment fails to reapply to the output
of truncation. Note that truncation is followed by the automatic resyllabication
of /r/ into the preceding syllable. At that point in the derivation, an automatic
reapplication of Backness Adjustment must be blocked. However, this blocking
must be stipulated by brute-force means.

6.2.3 Truncation: the Basic Model

In this subsection we will extract the general properties of Benuas analysis of
truncation in American English into a Basic Model of OO-correspondence, which
262
 6.2 Identity effects in truncation

is similar to that of reduplication. The correspondence model of truncation is

presented below:

(14) Basic Model of morphological truncation

BT-Identity
Base Truncated form

IO-Faithfulness
Input

The truncated form (T) is a stem, a free-standing form, hence an output. This is
related to a non-truncated form, itself a stem and free-standing form, which
Benua (1995) refers to as the base (B). This base, like any output form, has its own
input (I). Correspondence between elements in the input I and base B (an output)
is evaluated by IO-faithfulness constraints, in a standard way. Correspondence
between the truncated form T and its base B is evaluated by BT-identity con-
straints, militating against any dissimilarities between both.
This Basic Model, when applied to the truncated form [lr], which we dis-
cussed in the previous section, indicates the following correspondence relations
holding between the base, its input, and the truncated form.

(15) Correspondence relations in truncated form [lr]

BT-Identity
B [lri] T [lr]

IO-Faithfulness
I /lAri/

BT-identity constraints comparing the identity of morphologically related forms

presuppose the notion of correspondence, a relation holding between pairs of
segments in two strings, whose denition (McCarthy and Prince 1995a) is
repeated below from chapter 5:

(16) Correspondence
Given two strings S1 and S2, correspondence is a relation from the
elements of S1 to those of S2. Segments (an element of S1) and
(an element of S2) are referred to as correspondents of one another
when .

This notion of correspondence is general enough to include relations be-

tween two separate output forms. The strings S1 and S2 are instantiated by two
free-standing, morphologically related outputs. This is OO-correspondence
(Benua 1995, McCarthy 1995a, Burzio 1996). Here we consider the case of
263
Output-to-output correspondence

morphological truncation, where OO-correspondence takes the form of a relation

between segments in a truncated form and its base.2
The constraints enforcing BT-identity are highly similar to BR-identity con-
straints (relating base and reduplicant) introduced in chapter 5. Max-BT requires,
for example, that every element in the base must have a correspondent in the
truncated form.

(17) Max-BT
Every element in B has a correspondent in T.
Dep-BT requires that every element in the truncated form have a correspondent
in the base.
(18) Dep-BT
Every element in T has a correspondent in B.
Ident-BT[F] requires that correspondents in the base and the truncated form have
identical values for feature [F].
(19) Ident-BT[F]
Let be a segment in the base, and be a correspondent of in the
truncated form. If is [F], then is [F].

Other correspondence constraints (Anchoring-BT, Linearity-BT, Contiguity-

BT) may eventually be necessary, but the above set of three sufces for the cases
discussed here.

Finally, we need to address the constraints that trigger the truncation as such.
These constraints are counterparts of the templatic constraints in reduplication.
Benua (1995) has little to say about this issue, but we may assume an analysis
along the lines of It (1990). The basic idea is that truncated forms are morpho-
logical stems, and accordingly must full all relevant prosodic requirements for
stems. Perhaps the most important requirement is that a stem must equal a PrWd:
(20) Stem=PrWd
A stem equals a PrWd.
From this requirement the minimum size of the truncated form follows, typically
a binary (bimoraic or bisyllabic) foot. PrWd-status entails foot-status by the
prosodic hierarchy, as we saw in chapter 5, section 5.3.2.
The maximum size of the truncated form can then be modelled by morpho-
prosodic alignment constraints. For example, a cross-linguistically common size
2
The term base was chosen so as to reect the strong similarities with other types of OO-
correspondence, for example reduplication and truncation.

264
 6.2 Identity effects in truncation

of truncated forms is a heavy syllable (Mester 1990). This equals the smallest
possible foot, hence the minimal PrWd. Analogously to the constraint Red=
from chapter 5, this can be stated as follows:

(21) Trunc=
A truncated form equals a syllable.

An alternative, potentially more interesting, analysis would be analogous to the

emergence of the unmarked template shape (McCarthy and Prince 1994a; see the
analysis of disyllabic reduplication in Diyari in chapter 5). If we assume that
truncation involves a segmentally empty morpheme trunc (analogous to red
in reduplication), the prosodically unmarked single-foot size of the truncated
form will follow from the ranking: Markedness
Max-BT.
With these remarks in our minds, we now turn to the analysis of Icelandic
truncation.

6.2.4 Truncation in Icelandic

6.2.4.1 Introducing the pattern

Icelandic deverbal action nouns are truncations of the innitive, which arise by
the deletion of its nal vowel (rnason 1980: 52, Anderson 1982: 6, Kiparsky
1984: 156). Examples are given below:

(22) Innitive Deverbal action noun

a.i klifra climb klifr climbing
a.ii kumra bleat kumr bleating
a.iii grenja cry grenj crying
b.i stra sip str sipping
b.ii puukra conceal puukr concealment
b.iii kjkra wail kjkr wailing
b.iv siifra lament siifr lamentation

It can be argued that deverbal action nouns are derived from innitives by the
truncation of -a (rather than the other way around, deriving innitives by the
sufxation of -a). Deverbal action nouns have two idiosyncratic phonological
properties that set them apart from other forms. First, they may end in a cluster
of consonants that has a rising sonority, such as [tr], [kr], [fr], [mr], or [nj],
whereas word-nal clusters in Icelandic are generally of falling sonority, for
example bjrn bear, folald young foal. The second idiosyncratic property of
deverbal action nouns is that they may contain a long vowel standing before a
consonant cluster, whereas nal consonant clusters are generally preceded by
short vowels; see again bjrn.
265
Output-to-output correspondence

The generalization is that both idiosyncratic properties of deverbal action

nouns (that is, nal clusters of rising sonority and long vowels before clusters)
preserve phonological identity between the truncated form and the innitive
(Anderson 1975, Benua 1995). In the innitival base, consonant clusters of rising
sonority incur no violation of phonotactics, as they stand before a vowel, forming
a licit onset. And long vowels in innitives are syllable-nal, a position in which
length is phonotactically allowed. (Stressed open syllables always contain long
vowels.)

(23) Identity of length and consonant cluster in base and truncated form
s.tra Innitive: to sip
str Deverbal action noun: (the act of) sipping

Vowel length and consonant cluster originate in the innitive, and both are pre-
served in the related deverbal action noun. Deverbal action nouns thus preserve
the phonological shape of the related innitive, although they fail to provide the
prosodic context which normally licenses these shape characteristics. Or to state
it differently, deverbal action nouns copy innitival phonology, but thereby vio-
late otherwise generally respected constraints on the word-level phonology of
Icelandic.
Benua (1995) argues that identity effects are due to OO-correspondence con-
straints which require complete phonological identity between a truncated form
and its base. These identity constraints dominate markedness constraints (which
are responsible for ruling out nal clusters of rising sonority, and long vowels
before consonant clusters). Before we can discuss the precise interactions of
identity constraints and markedness constraints, we must rst nd out how Ice-
landic normally avoids violations of markedness constraints. That is, what pro-
duces the stringent phonotactic restrictions on non-truncated forms?

6.2.4.2 Sonority restrictions

Let us rst discuss the restriction that words should not end in a sequence of
consonants of rising sonority. Inputs ending in consonant clusters of rising sonor-
ity can be repaired in two ways, of which we will discuss one, epenthesis.3 Final
clusters of the form consonant-plus-/r/ (in shorthand notation /Cr/), are broken up
by epenthesis of [u] (Anderson 1974, 1975):

(24) Epenthesis: normal application

a. /tek-r/ te.kur take (pres.ind.)
b. /hest-r/ hes.tur horse (nom.sg.)

3
A second repair strategy is the deletion of a word-nal glide /j/ after a consonant, for example
/ bylj/ [byl] snowstorm (nom.sg.). See for details Benua (1995).

266
 6.2 Identity effects in truncation

Avoidance of nal clusters of rising sonority is cross-linguistically common. It is

known as the Sonority Sequencing Principle (Clements 1990). This general-
ization can be stated as a markedness constraint:4
(25) Son-Seq
Complex onsets rise in sonority, and complex codas fall in sonority.
Son-Seq dominates the anti-epenthesis constraint Dep-IO. For if the ranking
were reverse, no epenthesis would take place. Epenthesis is selected as a repair
strategy for /Cr/ clusters, over another logically possible strategy, the deletion of
a consonant in the nal cluster. The choice of (vowel) epenthesis over (consonant)
deletion shows that Max-IO dominates Dep-IO. This is a familiar interaction of
which we have seen examples in chapter 3. In tableau (26) this interaction is
shown by evaluations of three candidates for the input /tek-r/:

(26) /tek-r/ Son-Seq Max-IO Dep-IO

a. + te.kur *
b. tek *!
c. tekr *!

We assume that the position of the epenthetic vowel inside the stem is due to
high-ranking Anchoring-IO, which rules out the candidate [te.kru].

6.2.4.3 Quantitative restrictions

The second phonotactic restriction that appears to be disrespected by deverbal
action nouns is the ban against long vowels in words of the type CVCC. Let us
take a closer look at the distribution of quantity in Icelandic. Stressed vowels are
always long in open syllables, and short in non-nal closed syllables:
(27) Stressed long vowel in open Stressed short vowel in closed
syllable syllable
a.i h.fu head b.i har.ur hard
a.ii aa.kur eld b.ii el.ska love
a.iii faa.ra ride b.iii kal.la call
In monosyllables, stressed vowels are always long before single consonants (28b),
but they are short before consonant clusters (28c).
4
The situation is somewhat more complex, since consonant clusters of rising sonority occur:
vopn weapon, kukl witchcraft, vatn water (all examples cited from Booij 1986). These ex-
amples suggest that the prohibition against codas of rising sonority falls apart into several con-
straints, with *Cr] dominating other *CC] constraints in which the second consonant is not /r/.

267
Output-to-output correspondence

(28) Length in Icelandic monosyllables

long in CVV long in CVVC short in CVCC
a.i skoo shoe b.i haas hoarse c.i bjrn bear
a.ii buu homestead b.ii ljoos light c.ii haft have
a.iii /t/ tea b.iii skiip ship c.iii skips ships

This pattern can be understood as an interaction of two constraints. (The analysis

developed from here on is the authors.) The rst is Stress-to-Weight (Myers
1987, Riad 1992):

(29) Stress-to-Weight
If stressed, then heavy.

The second constraint trims syllabic overweight by militating against trimoraic

syllables:

(30) *3
No trimoraic syllables.

In combination, high-ranking Stress-to-Weight and *3 have the effect that

all stressed syllables are bimoraic. In principle, there are two ways in which
a syllable can be bimoraic. The rst is by vowel lengthening. This occurs in
open syllables (where no coda consonant is available), and in syllables closed
by a single consonant. The second way in which weight may arise is by render-
ing a coda consonant moraic. This is what we will assume for syllables closed
by two consonants (CVCC). The analytic problem is this: why do syllables
closed by a single consonant (CVVC) behave like open syllables (CVV) for
vowel lengthening?
The fact that vowels are long before a single word coda shows that nal con-
sonants are non-moraic. (If the nal consonant in [haas] hoarse were moraic,
then there would be no need for vowel length.) Non-moraicity of nal consonants
is due to the constraint in (31) (Lorentz 1996, Borowsky and Harvey 1997):5

(31) *Final-C-
The nal consonant is weightless.

This is a kind of NonFinality constraint (see chapter 4) at the level of the mora.
For words ending in a consonant cluster, it only affects the nal consonant.

5
Actually Icelandic has word-nal geminate consonants, e.g. [vis] certain, but these are never
preceded by long vowels (behaving like consonant clusters in this respect). Therefore *Final-
C- is dominated by a faithfulness constraint preserving input quantity in consonants. See
Spaelti (1994) and Lorentz (1996) for analyses of similar phenomena in Norwegian and Swiss
German.

268
 6.2 Identity effects in truncation

Next consider the question of how to block *CVVCC syllables, that is, a long
vowel before a nal cluster. We explain this blocking by the ban on trimoraic syl-
lables, on the assumption that the rst consonant in the cluster is moraic. This mora
renders vowel length impossible (for trimoraic syllables are ruled out by *3).
Moraic status of coda consonants is due to the following constraint (Hayes 1989):

(32) Weight-by-Position
Coda consonants are moraic.

Evidently Weight-by-Position must be dominated by *Final-C-, since the vowel

length in CVVC-monosyllables shows that the nal consonant is weightless.

(33) *Final-C-
Weight-by-Position

[haas]
[has]

In contrast, the nal consonant in CVCC-monosyllables remains weightless due

to *Final-C-, while the prenal one is obligatorily moraic. Consider the follow-
ing two possibilities for satisfying syllable bimoraicity in a monosyllable ending
in a cluster:

(34) Syllable weight satised by moraic coda rather than long vowel
[bjrn]
[bjrn]
The actual form [bjrn] has a short vowel, from which we infer that the initial
consonant in the cluster is moraic. The suboptimal form [bjrn] has a long
vowel, but none of the coda consonants is moraic. We thus nd a preference for
a moraic coda (Weight-by-Position) to a long vowel.
Since vowel length is an input property, the neutralization of length before a
cluster in CVCC syllables must happen at the expense of a violation of a faith-
fulness constraint militating against changes in input length. This faithfulness con-
straint is Wt-Ident-IO, stated below in the correspondence format (McCarthy
1995a):

(35) Wt-Ident-IO
If Domain(),
if is monomoraic, then () is monomoraic. (no lengthening)
if is bimoraic, then () is bimoraic. (no shortening)
Wt-Ident-IO requires identical quantity of output segments and their input cor-
respondents. This requirement is made in both directions, so that the constraint
militates against both the addition of quantity (lengthening) and loss of quantity
(shortening).
Since surface vowel length entirely depends on stress and syllable structure
in Icelandic, Wt-Ident-IO must be dominated by the length-inducing constraint
269
Output-to-output correspondence

(Stress-to-Weight), as well as by the length-inhibiting constraints (*3,

Weight-by-Position). Since we already know that *Final-C-
Weight-by-
Position, we arrive at:

(36) Stress-to-Weight, *Final-C-, *3
Weight-by-Position

Wt-Ident-IO

This ranking correctly predicts that underlying vowel length is irrelevant to the
outcome. Regardless of whether an input vowel is long or short, it will be long if
and only if it stands in a stressed open syllable. The tableaux (379) demonstrate
the irrelevance of input length by reversing it (as compared to the output).
The rst tableau is simple: the requirement that the stressed syllable be heavy
means that the vowel of /sko/ is lengthened, regardless of the violation of faith-
fulness to length.

(37)
Input: /sko/ Stress-to- *Final- *3 Weight-by- Wt-Ident-
Weight C- Position IO

a. + skoo *
b. sko *!

The second tableau, of [haas], is slightly more complicated. It shows that the
vowel of a CVC input lengthens to satisfy Stress-to-Weight. Strategies to retain
input shortness of the vowel fail, specically that of making the coda moraic,
which fails on *Final-C-:

(38)
Input: /has/ Stress-to- *Final- *3 Weight-by- Wt-Ident-
Weight C- Position IO

a. + haas * *
b. has *!
c. has *! *

Weight-by-Position is violated in the optimal candidate to avoid violation of

*Final-C-.
The third tableau, of [bjrn], adds the factor of a nal consonant cluster.
Strategies to preserve the hypothetically long input vowel all fail. First, preserving
vowel length in a syllable with a moraic coda (39c) simply fails on *3. Second,
270
 6.2 Identity effects in truncation

preserving length by keeping both coda consonants weightless (39b) fails on non-
minimal violation of Weight-by-Position:

(39)
Input: /bjrn / Stress-to- *Final- *3 Weight-by- Wt-Ident-
Weight C- Position IO

a. + bjrn * *
b. bjrn **!
c. bjrn *! *

This tableau thus demonstrates the ranking Weight-by-Position
Wt-Ident-IO.

In sum, input vowel length is completely overruled by quantitative well-
formedness of the syllable. The latter requirements can now be written as a cover
constraint:

(40) Quant-Form (cover term for constraint interaction)

Stress-to-Weight, *Final-C-, *3
Weight-by-Position

With Son-Seq (from section 6.2.4.2) and Quant-Form available as the con-
straints spelling out the phonotactic laws in the phonology of Icelandic, we can
now return to the original problem of accounting for the violations of these
phonotactics in truncated forms.

6.2.4.4 OO-correspondence in truncation

The analysis of truncation requires three constraints which are specic to trunca-
tion. First, a templatic constraint requiring that truncated forms be monosyllabic:

(41) Trunc=
A truncated form equals a syllable.

This constraint is undominated, as far as the examples can tell us. (As mentioned
above, the effect may also be due to an emergence of the unmarked, but we will
not pursue this idea.)
Second, a constraint requiring identical quantity between correspondents in the
base and the truncated form:

(42) Wt-Ident-BT
If Domain(),
if is monomoraic, then ( ) is monomoraic. (no lengthening)
if is bimoraic, then ( ) is bimoraic. (no shortening)
271
Output-to-output correspondence

This is of course the output-to-output variant of Wt-Ident-IO, introduced in (35).

However, as we will soon nd out, this identity constraint is crucially ranked
higher in the grammar than the faithfulness constraint.
Third, a constraint militating against the loss of base segments in the
truncated form. This anti-truncation constraint is stated as a member of the
Max-family:
(43) Max-BT
Every element of B has a correspondent in T.
(No truncation)
The crucial interactions of these constraints have been spelled out below. Mono-
syllabicity of truncation has priority over preserving segments of the base in
the truncated form (44a). In its turn, segmental base identity has priority over
phonotactic requirements (with respect to sonority and quantity), giving the over/
underapplication effects (44b). Finally, identity of vowel length (in base and
truncated form) has priority over quantitative requirements (44c):
(44) a. Trunc=
Max-BT
str
s.tra
b. Max-BT
Son-Form, Quant-Form
str
st
c. Wt-Ident-BT
Quant-Form
str
str
Integrating all this into a total hierarchy, we arrive at:
(45) Icelandic truncation
Trunc =, Wt-Ident-BT
Max-BT
Son-Seq, Quant-Form

Wt-Ident-IO
The tableau of [str] is straightforward now.
(46)
Input: /str-a/ Trunc= Wt- Max- Son- Quant- Wt-
Base: [s.tra] Ident-BT BT Seq Form Ident-IO

a. + str * * * *
b. st **! *
c. str *! * *
d. s.tra *! *
e. s.tur *! * *

272
 6.3 Identity effects in stem-based afxation

With any disyllabic candidates (46de) eliminated by Trunc=, the choice is

among three candidates: [str] (46c), [st] (46b), and [str] (46a). The rst of
these has a short vowel, fatally violating undominated Wt-Ident-BT. Among the
remaining candidates (46ab), the optimal one is (46a), since this minimally
violates Max-BT. This happens at the expense of the violation of both phono-
tactic constraints (Son-Seq, Quant-Form).

6.2.4.5 A derivational analysis

Kiparsky (1984: 156) proposes that truncation is due to a late phonological rule
of /a/-deletion, rather than to a (subtractive) morphological process. As shown in
the derivation below, this /a/-deletion rule applies after syllabication and vowel
lengthening:
(47) UR /str-a/
Syllabication s.tra
Vowel lengthening s.tra
/a/-deletion str
PR [str]
By stating truncation as a phonological rule, the generalization is saved that all
morphology is additive not subtractive. But another generalization is missed:
only truncation produces output forms which violate the (otherwise completely
respected) phonotactics of Icelandic, including sonority-based and quantitative
restrictions. As Anderson (1975) observes, such violations serve to preserve max-
imal phonological identity between the truncated form and innitive. The OT
analysis captures this insight, while the derivational analysis ignores it.

6.3 Identity effects in stem-based afxation

This section will discuss a second source of evidence for OO-correspondence:
the phonology of stem-based afxations. Empirically, we will focus on identity
phenomena in stems and afxed forms in American English.

6.3.1 Identity and stem-based afxation in English

A second argument for OO-correspondence presented by Benua (1995) again
involves a set of vowel alternations in American English. Here we consider dia-
lects spoken in the PhiladelphiaNew York region. These dialects have a tense
allophone [E] of the low front vowel [] in closed syllables before certain con-
sonants, among which /s/, for example pass [pEs], but passive [p.sv] (48a).6
Note that -tensing does not apply in open syllables before /s/; see (48b c):

6
The sets of tautosyllabic consonants that trigger -tensing vary slightly between the dialects,
but this is irrelevant for the following argument. See Ferguson (1975), Dunlop (1987).

273
Output-to-output correspondence

(48) -tensing: normal application

a.i pass [pEs] b.i passive [p.sv] c.i acid [.sd]
a.ii mass [mEs] b.ii massive [m.sv] c.ii tacit [t.st]
a.iii class [klEs] b.iii classic [kl.sk] c.iii lasso [l.so]

The words in (48b) are morphologically related to those of (48a) by afxation of

adjectival /-iv/ and /-k /, both so-called Class 1 afxes. The words in (48c) are
simplex stems.

Next observe that in the same dialects, words which are sufxed by Class 2
sufxes (-ing, -able, and -y) display special behaviour with respect to -tensing
(Dunlop 1987). The stems of (48a) preserve their tense allophone [E] under Class
2 sufxation:

(49) -tensing: overapplication in Class 2 afxed forms

a.i pass [pEs] b.i class [klEs] c.i mass [mEs]
a.ii passing [pE.s] b.ii classy [klE.si] c.ii massable [mE.s.bg]

Preservation of [E] in the afxed form can be considered an overapplication of

-tensing. -tensing overapplies in the sense that the vowels in open syllables of
passing, classy, and massable surface as tense, even though they fail to meet the
context of tensing. The afxed form copies the vowel quality of its stem, in
which tensing applies normally.

The key difference between Class 1 and Class 2 afxes is that only the latter
require a stem, a free-standing lexical item, as their base (Kiparsky 1985, Inkelas
1989, Borowsky 1993). It has often been observed (starting with Siegel 1974)
that Class 1 afxations behave in many phonological aspects as morphologic-
ally underived words. For example, compare the identical vocalisms of pass-ive
(48b.i) and acid (48c.i). The base of a Class 1 afx need not be a free-standing
stem, but it can be a morpheme which does not occur independently: a root. (For
example, the root aggress- occurs in aggress-ive and aggress-ion; the root pacif-
in pacif-ic and pacif-y.) This difference between Class 1 and Class 2 afxation is
crucial to the observed differences in allophonic distribution between both types
of word formation.

6.3.2 Stem-based afxation: the Basic Model

Benua (1995) argues that phonological identity effects in Class 2 afxation are
due to OO-correspondence. Overapplication in passing is due to its relatedness
to its base, pass [pEs]. In a diagram, this is portrayed as follows (B abbreviates
base, and A afxed form):
274
 6.3 Identity effects in stem-based afxation

(50) Basic Model of stem-based afxation

BA-Identity
[pEs] [pE.s]

IO-Faithfulness
/ps/(cf. passive)
Benua (1995: 51) argues more generally that Class 2 afxation is derived through
an O/O correspondence with the unafxed word. This OO-correspondence rela-
tion is analogous to that seen in truncated forms in the previous subsection.7
Two constraint interactions are relevant. First, the allophonic distribution of
tense and lax low vowels, once more, points to the allophonic ranking schema of
chapter 1:
(51) Schema for allophonic variation
Contextual markedness
Context-free markedness
Faithfulness
The contextual markedness constraint is -Tensing, militating against [] in
closed syllables. It outranks a context-free markedness constraint *Tense-low,
requiring low vowels to be lax.
(52) a. -Tensing
*C]
No [] in closed syllables8
b. *Tense-low
Low vowels are lax.
c. Ident-IO(tense)
Let be a segment in the input, and be a correspondent of in
the output.
If is [tense], then is [tense].
This ranking is supported by the following tableaux of [pEs] and [.sd], which
show that input values of [tense] are overruled:

(53) Input: /ps/ -Tensing *Tense-low Ident-IO(tense)

a. [ps] *!
b. + [pEs] * *

7
Benua (1997) argues that Class 1 afxation is subject to its own set of OO-identity constraints.
These constraints are ranked below those for Class 2 afxation: OO2-Identity
OO1-Identity.
8
Benua refers to -Tensing as a descriptive constraint, since it must be specic to the
[]~[E] alternation, and not force tensing of other vowels. In contrast, *Tense-low is
grounded to the extent that it reects the cross-linguistically common tendency to avoid tens-
ing in low vowels.

275
Output-to-output correspondence

(54) Input: /Esd/ -Tensing *Tense-low Ident-IO(tense)

a. + [.sd] *
b. [E.sd] *!

The second interaction relevant to the afxed form [pE.s] relates to BA-
identity. The relevant BA-identity constraint requires that the tenseness of a vowel
in the base and its correspondent in the afxed form should be identical.

(55) Ident-BA(tense)
Let be a segment in the base, and be a correspondent of in the
afxed form. If is [tense], then is [tense].

This constraint outranks *Tense-low, as is supported by the selection of [pE.s]

rather than [p.s].

(56) Ident-BA(tense)
*Tense-low

[pE.s]
[p.s]

In sum, the overapplication follows from the ranking in (57).

(57) Overapplication of allophonic phonology in Class 2 afxed forms

Ident-BA(tense), -Tensing
*Tense-low
Ident-IO(tense)

The interaction of these constraints is shown in a tableau of the afxed form

[pE.s]:

(58)
Input: /ps/ Ident- -Tensing *Tense-low Ident-IO(tense)
Base: [pEs] BA(tense)

a. [p.s] *!
b. + [pE.s] * *

Upon comparing this tableau with that of truncation [lr] in (11), it will be clear
that, once more, we witness a domination of a markedness constraint over an OO-
identity constraint. The analysis of stem-based afxation is fully analogous to that
of truncation, and both show that Correspondence Theory elegantly captures rela-
tions between output forms.
Again, a derivational analysis of the pattern is possible, but it encounters prob-
lems that are analogous to those pointed out in section 6.2.4.5 with respect to
truncation. The general idea of a derivational analysis is ordering the allophonic
276
 6.4 The cycle versus base-identity

rule which adjusts tenseness in closed syllables before the morphological rule of
Class 2 afxation.

(59) /ps/ Lexical form

pEs Tenseness Adjustment
pE.s Class 2 afxation, resyllabication
[pE.s] Surface form

This analysis is subject to the same criticisms as the derivational analysis of

truncation that was discussed earlier. By ordering an allophonic rule (of Tenseness
Adjustment in closed syllables) before a morphological rule (of /-/ afxation),
it violates standard assumptions of Lexical Phonology (Kiparsky 1985). That is,
allophonic rules should not apply in the lexical component, since they are not
structure-preserving.9

6.4 The cycle versus base-identity

6.4.1 Introduction
Proponents of serial theory have argued for derivational levels which do not
coincide with the input, nor with the output. An important argument for interme-
diate derivational levels was based on phonological properties carried over from
morphologically simplex forms to complex forms. Such transderivational transfer
was modelled as the transformational cycle (Chomsky and Halle 1968), a mode
of rule application in morphologically complex words in which rules apply in an
inside-out fashion, from smaller to larger morphological domains. Cyclic rule
application is intrinsically derivational, as it implies intermediate levels between
the input and output at which phonological generalizations are captured.
Like derivational theory, OT searches for explanation in maximizing general-
izations at some level of the grammar. The output is the privileged level at which
signicant linguistic generalizations are captured. OT grammars map underlying
representations (lexical inputs) into surface forms (outputs) without intermediate
levels, drastically reducing derivations to one-step mappings. The crucial differ-
ence between both theories resides in how interactions between generalizations

9
Most dialects of English offer other examples of underapplication and overapplication in words
derived by Class 2 afxes. Among these are the following two processes (Borowsky 1993):
Simplication of nal /mn / Syllabicity of /r/
hymn /hmn / hinder /hndr/
hymn-al [hmnl] hinder-ance [hndrns]
hymn# [hm] hinder# [hndl]
hymn#ing [hm] hinder#ing [hndl]
Final nasal cluster simplication overapplies in [hm], and /r/-syllabication in [hndl].

277
Output-to-output correspondence

take place: either by linear precedence of rules (derivational theory) or by hierar-

chical ranking of constraints (OT).

Here we will compare two theoretical means of capturing transderivational iden-

tity: the cycle and base-identity. To evaluate predictions made by derivational
theory and OT, we will analyse interactions of stress and vowel deletion in Pales-
tinian Arabic. This language has a set of sufxes which, like Class 2 sufxes in
English, exhibit special phonological behaviour. Words derived by these sufxes
display an underapplication of a vowel deletion process that normally applies to
words of analogous segmental make-up and stress pattern. A cyclic analysis of
these underapplication effects (Brame 1974) will be compared with an analysis
using base-identity (Kenstowicz 1996, Steriade 1996, Kager forthcoming).
Discussion is organized as follows. In section 6.4.2, we will compare deriva-
tional theory and OT with respect to transderivational relations, on the basis of
data from Palestinian Arabic involving underapplication of vowel deletion with
respect to the surface stress pattern. An analysis will be proposed featuring a new
kind of base-identity constraint, requiring identity between vowels in the afxed
form and stressed correspondents in the base. We will argue for a notion of para-
digmatic relatedness, predicting the morphological relations under which two
forms display phonological identity effects. Section 6.4.3 will compare this base-
identity analysis with a cyclic analysis in derivational theory. Section 6.4.4 dis-
cusses similar identity effects in epenthesis in Palestinian Arabic. Here we will
focus on the possibility, predicted by OT, that an output form reects the effects
of both IO-faithfulness constraints and BA-identity constraints. This will support
parallel evaluation of output forms, in the sense that both the base and lexical
input are accessible simultaneously. Section 6.4.5 will contain conclusions.

6.4.2 Underapplication of syncope in Palestinian Arabic

6.4.2.1 The cyclic analysis

A famous example of cyclic rule application comes from the interaction of stress
and vowel deletion in Palestinian Arabic, as analysed by Brame (1974). The stress
rule interacts with a rule of i-Syncope, deleting unstressed /i / in open non-nal
syllables. This is stated below:

(60) i-Syncope
i
/ __ CV
stress

Verbal forms inected for subject (person, number, and gender) illustrate the
application of i-Syncope:
278
 6.4 The cycle versus base-identity

(61) Normal application of i-Syncope in verbal forms with subject

sufxes
a. /him/ to understand (verb stem)
b.i /him/ f him he understood
b.ii /him-na/ fhm-na we understood
b.iii /him-u/ f hm-u they understood

Observe that i-Syncope preserves stressed vowels, hence it must be ordered after
stress. The stress rule of Palestinian Arabic (which we will not state here in rule
notation) places stress on a heavy penultimate syllable, otherwise on the antepe-
nult. Derivations of these forms are presented in (62):

(62) /him/ /him-na/ /him-u/

Stress fhim hm-na fhim-u
i-Syncope --- fhm-na fhm-u
[fhim] [fhm-na] [fhm-u]

So far, we have not seen any cases of cyclic rule application. The actual examples
on which Brame (1974) rested his case for transderivational preservation of stress
are verbal forms containing accusative sufxes. Accusative sufxes express mor-
phological features of the object. More precisely, verbal forms may be inected
for person, number, and gender of the object by a sufx that is added to a verb
form inected for subject. Observe that bold-face [i] in the forms in (63c) fails
to delete, even though it stands in the context of i-Syncope, an open unstressed
non-nal syllable.

(63) Normal application and underapplication of i-Syncope in accusatives

a. /him/ fhim he understood
b.i /him-ak/ f hm-ak he understood you m.
b.ii /him-ik/ fhm-ik he understood you f.
b.iii /him-u/ fhm-u he understood him
c.i /him-ni/ fihm-ni *fhm-ni he understood me
c.ii /him-ha/ fihm-ha *fhm-ha he understood her
c.iii /him-na/ fihm-na *fhm-na he understood us

Brame observes that the accusative forms in which [i] fails to delete are all
based on a free form [fhim] (63a) in which this [i] is stressed. The generalization
is that i-Syncope underapplies in any afxed form that has a morphological base
form in which [i] is stressed. Or to state it differently, any [i] having a stressed
correspondent in the base is protected from deletion.
The correctness of this generalization is shown by possessives, nouns which
contain a sufx indicating person and number of the possessor. Observe that
279
Output-to-output correspondence

i-Syncope is blocked in the bold-face vowels of these forms. Like accusatives,

possessives have morphologically related nouns in which the corresponding [i] is
stressed (Kenstowicz and Abdul-Karim 1980):

(64) Underapplication of i-Syncope in possessives

a. / birak / brak pools
b. / birak-u / brak-u his pools
c. / birak-na/ birk-na *brk-na our pools

Again, underapplication of i-Syncope correlates with the presence of a stressed

[i] in a base form. Unstressed [i] in [birk-na] is left unaffected by syncope when
this vowel is stressed in the base [brak]. And precisely as we saw before for the
accusative forms, the possessive forms are formed by the addition of a (posses-
sive) sufx to a free form (or base), which is inected itself.
Underapplication of i-Syncope cannot be due to an actual (secondary) stress on
the initial syllable of accusatives and possessives, e.g. [fhm-na] and [brk-na].
Kenstowicz and Abdul-Karim (1980) found that for speakers of Palestinian Ara-
bic the analogous form that is based on a CaCaC verb, e.g. [7arb-na], is ambig-
uous between we hit and he hit us. If there is no phonetic difference between
the two forms, then certainly there cannot be a secondary stress on the [i] that is
protected from syncope in the accusatives and possessives. This then makes the
property which is responsible for the blocking of i-Syncope in accusatives and
possessives abstract to a certain degree. Of course, the abstractness of the prop-
erty which blocks i-Syncope was precisely what raised Brames interest. Stress
is not present in lexical representations, nor is it present at the surface on the
vowel which is mysteriously protected from i-Syncope. Hence it must be the case
that i-Syncope applies to a level of representation intermediate between the
lexical representation and the surface representation. This, in turn, implies that
phonological theory should allow such intermediate levels of representation.
Thus, what Brame argued for is that derivations are an essential part of phono-
logical theory.
In derivational theory the notion of relatedness between forms is characteris-
tically modelled in derivational terms, that is, cyclically. Under cyclic application,
an ordered set of phonological rules (R1Rn) rst apply to the minimal morpho-
logical domain, and then to successively larger domains. Brame (1974) assumes
that accusatives and possessives have an additional internal layer of morpholog-
ical structure, which triggers a cyclic application of the stress rule.
The cyclically assigned stress on the rst syllable of [f him] is carried over
to the second cycle [[fhm]-na], even though it becomes subordinated to the
new main stress. The secondary stress protects the initial syllables vowel against
280
 6.4 The cycle versus base-identity

post-cyclic i-Syncope. Finally, initial secondary stress is erased by a post-cyclic

rule of Destressing:

(65) Input [him-na]Subj [him-u]Subj [[him]na]Acc [[him]u]Acc

Cycle1
Stress hm-na fhim-u fhim fhim
Cycle2
Stress --- --- fhm-na (vacuous)
Post-cyclic
i-Syncope fhm-na f hm-u blocked fhm-u
Destressing n.a. n.a. hm-na n.a.
Output fhmna fhmu fihmna f hmu

This analysis was regarded as strong evidence for extrinsic rule ordering, in the
sense that a phonological property (stress) that is acquired in the course of the
derivation blocks a rule that is sensitive to its presence (i-Syncope), even though
it is absent from the surface form due to a subsequent rule that deletes it
(Destressing). This is achieved by linearly ordered rules that are blind to under-
lying representations (no globality), and have access only to the representation
that arises at the point in the derivation at which they apply.

6.4.2.2 The base-identity analysis

How can underapplication of i-Syncope be analysed in OT, a theory which (in its
strictest form) avoids derivations, hence lacks the intermediary level of represen-
tation which Brame argued for? An OT analysis based on base-identity was pro-
posed by Kenstowicz (1996), Steriade (1996), and Kager (forthcoming). We will
discuss it below.
Let us return to the initial generalization made by Brame: the unstressed vowel
[i] in an accusative or possessive is protected from deletion if its counterpart
in the base word is stressed. Blocking of i-Syncope is a kind of base-identity,
keeping morphologically related forms maximally similar in their phonological
form. Accordingly the central idea of the OT analysis is that underapplication of
i-Syncope is due to OO-correspondence. Let us now develop such an analysis.10

From here on, we will use the notion of base in a specic sense, making explicit
two criteria. First, the base is a free-standing output form of the language, that is,
a word. Second, the base is compositionally related to its derived counterpart (the
afxed form or truncated form). That is, the base contains a proper subset of the
grammatical (semantic, morphological) features of the derived form.

10
The analysis presented below draws on Kager (forthcoming).

281
Output-to-output correspondence

(66) Denition of base

a. The base is a free-standing output form a word.
b. The base contains a subset of the grammatical features of the
derived form.

Note that the morphological relation between a truncated form and its base is
included in this denition. Both occur as free-standing independent words. Trun-
cation also matches the second criterion. Consider, for example, the morpho-
semantic status of a truncated form in Icelandic, the deverbal action noun. This
properly contains the morpho-semantic features of the innitive, which is its base.
In fact, the only thing that is subtractive about truncation is phonological: the
truncated form is a segmental reduction of the base. From a morpho-semantic
viewpoint, truncation is an addition, precisely as overt afxation. (Note that the
very terminology truncation correctly suggests that the truncated form results
from a morphological operation on the base.)

This denition of base is precise enough to capture the distinction among

subject and accusative forms in Palestinian Arabic which is relevant for the
(under)application of i-Syncope.
Subject forms such as [fhm-na] we understood (61b.ii) have no base since
no form occurs which matches both criteria for base-hood. In particular, the
verb stem /him/ to understand, which would be appropriate in a compositional
sense, lacks inection features. Hence the verb stem is no word, thereby failing
the rst criterion. Nor can the word [fhim] he understood serve as the base for
[fhm-na] we understood. These words are not compositionally related to one
another, because of a conict of inectional features (1sg. versus 2pl.), thereby
failing the second criterion.
Accusative forms, in contrast, have bases by both criteria. For example, the
form [hm-na] he understood us has as its base a free-standing form [f him]
he understood, of which it contains all morpho-semantic features. More gener-
ally, every accusative form has precisely one subject form which is its base.

(67) Form Gloss Base Gloss

a.i [fhm-na] we understood (none)
a.ii [f h.mu] they understood (none)
b.i [fihm-na] he understood us [fhim] he understood
b.ii [f hm-u] he understood him [fhim] he understood

By the same criteria, possessive forms have bases, for example [birk-na] our
pools has as its base the word [brak] pools, which contains a subset of its
morpho-semantic features.
282
 6.4 The cycle versus base-identity

Keeping these remarks on the notion of base in our minds, we can now return
to the original problem, of how to characterize underapplication of i-Syncope
in object forms and possessives. As we have seen, the generalization is that
i-Syncope underapplies to vowels which are stressed in the base:

(68) a. [fhim] he understood [fihm-na] he understood him

b. [brak] pools [birk-na] our pools

From here on, graphic means will serve to indicate the correspondence relations
between an afxed form and (on the one hand) its input and (on the other hand)
its base. These relations will be marked by vertical lines between correspondents
at three levels (input, output, and base). In (69a), bold-face [i] indicates the
underapplication of i-Syncope in the output:

(69) a. he understood us b. we understood

/f i h i m -n a/ /f i h i m -n a/ Input
IO-correspondence
[f i h m -n a] [f h m -n a] Output
OO-correspondence
[f h i m] Base
he understood
The correspondence-based perspective of this pattern is that syncope under-
applies in the accusative and possessive because the relevant vowels have
stressed correspondents in the base. More precisely, syncope is blocked by an
OO-correspondence constraint requiring that vowels which are stressed in the
base must have correspondents in the afxed form:

(70) HeadMax-BA
Every segment in the bases prosodic head has a correspondent in the
afxed form.

This is a head-identity constraint, which involves an extension of correspondence

to stress properties of segments. Such an extension was proposed by McCarthy
(1995a) and Alderete (1995).
Vowel deletion (i-Syncope) is triggered by a markedness constraint disallow-
ing [i] in open syllables.11

11
This constraint is modelled after an analogous constraint No [a] in Orgun (1995). An inter-
esting question is whether No [i] can be decomposed into general constraints which, taken
together, produce its effects. Presumably this is possible, if we assume a constraint against
monomoraic syllables (argued for by Broselow 1992: 32), and rank this in between faithful-
ness constraints for specic vowels, for example Ident-IO[a]
*
Ident-IO[i]. We will
not pursue this issue here, but maintain the formulation of No [i] as it is.

283
Output-to-output correspondence

(71) No [i]
/i/ is not allowed in light syllables.
Note that no reference is made to stress, an aspect that will be conrmed in
section 6.4.4, where we discuss the interaction of syncope, stress, and epenthesis.
No [i] clearly outranks the faithfulness constraint requiring that input segments
have correspondents in the output (McCarthy and Prince 1995a):
(72) Max-IO
Every segment in the input has a correspondent in the output.
This ranking is motivated by the fact that deletion occurs in forms such as
[fhmna] we understood, which satisfy No [i] at the expense of Max-IO.

The underapplication of i-Syncope requires HeadMax-BA
No [i] since dele-

tion is blocked wherever base-identity is active. The constraint ranking that now
arises is (73):
(73) Underapplication of i-Syncope
HeadMax-BA
No [i]
Max-IO
The analysis is illustrated by tableaux (745). Only candidates that satisfy the
canonical stress patterns of Palestinian Arabic are considered; the issue of how
stress interacts with the processes of syncope and epenthesis will be taken up in
section 6.4.4 below.
Tableau (74) illustrates the underapplication of i-Syncope, while tableau (75)
shows how syncope occurs in a form that lacks a base:

(74) Input: /him-na/ HeadMax-BA No [i] Max-IO

Base: [f.him]

a. + [.hm.na] *
b. [fhm.na] *! *

(75) Input: /him-na/ HeadMax-BA No [i] Max-IO

Base: none

a. [.hm.na] *!
b. + [fhm.na] *

This shows that we are on the right track: the basic observation that i-Syncope
underapplies in accusatives with consonant-initial sufxes has now been
accounted for.
284
 6.4 The cycle versus base-identity

Next let us consider forms that have vowel-initial sufxes, the accusative
[fh.mu] he understood him (with its base [f.him] he understood), and the
subject form [fh.mu] they understood (lacking a base). The constraint ranking
correctly predicts that i-Syncope applies in both forms, resulting in a pair of
homophonous outputs. In the accusative form, the second occurrence of [i] in the
stem is freely deleted since its counterpart in the base is not stressed. Therefore
HeadMax-BA is not violated in candidate (76b), and the outcome is left to the
markedness constraint No [i]:

(76) Input: /him-u/ HeadMax-BA No [i] Max-IO

Base: [f.him]

a. [f .hi.mu] *!*
b. + [fh.mu] *
c. [fh.mu] *! * *

In the subject form, HeadMax-BA is simply irrelevant, since there is no base in

this case:

(77) Input: /him-u/ HeadMax-BA No [i] Max-IO

Base: none

a. [f .hi.mu] *!*
b. + [fh.mu] *
c. [fh.mu] *! *

In sum, syncope is not blocked by the presence of a counterpart vowel in the base,
but only by the presence of a stressed counterpart vowel in the base.

6.4.3 The cycle versus base-identity

From the above discussion a number of general conclusions can be drawn. First,
the notion base in OO-correspondence is rmly linked to compositionality.
Second, the assumption that the base is a free-standing output form correctly
predicts that transderivational relations involve output forms. This prediction does
not follow from derivational theory, since what counts as a cycle is inherently
unconstrained by the criterion of free-standing form. Any layer of morpho-
logical structure, regardless of its relation with free-standing words, is predicted
potentially to display cyclicity. In order to distinguish between subject morpho-
logy and object morphology in Palestinian Arabic verbs, Brame had to make the
285
Output-to-output correspondence

(arbitrary) assumption that only the latter type invokes cyclicity. However, if
matters had been reverse (with subject morphology invoking cyclicity), this would
have been equally easy to express on the cyclic theory.12
Let us make a more detailed comparison between the base-identity analysis and
the cyclic derivational analysis, based on the subject form [fhmu] they under-
stood. The cyclic analysis (in derivation 65) imposes a condition on i-Syncope
restricting its application to unstressed vowels. This stress condition crucially
protects the vowel in the initial syllable of /him-u/ against i-Syncope, in order
to rule out the incorrect form *[fhmu].
In the base-identity analysis, no need arises for such a stress condition, and
actually it would be impossible to state anything like it. It is impossible to refer
to lack of stress in the target vowel of syncope, as this vowel does not appear in
the output. Nor could blocking of initial syncope be attributed to base-identity,
since subject forms generally lack a base. Instead, the base-identity analysis
blocks candidate (77c) *[fh.mu] by the same constraint that triggers i-syncope in
the rst place: No [i]. If the goal of i-Syncope is the avoidance of light syllables
containing [i], then some progress toward this goal must be made in the output,
due to minimal violation. (Compare optimal [fh.mu] in 77b, which has no open
syllables with [i], with suboptimal *[fh.mu] in 77c, where one open syllable with
[i] is still present.)

Nevertheless, the base-identity analysis does still require some reference to stress
to account for underapplication effects, as in the accusative form [hmna] (74a).
As we have seen, reference to stress is made by HeadMax-BA, a base-identity
constraint requiring that stressed vowels be preserved in morphologically related
forms. The difference in approach between the analyses can be stated as follows:
the OT analysis rationalizes the reference to stress as a case of paradigm uniformity,
involving OO-correspondence. The derivational analysis instead takes the para-
digmatic dimension in blocking to be coincidental, and views blocking as due to
the presence of an abstract stress at an intermediate level of derivation. Which
view is more appropriate?
Arguably, the abstractness of the cyclic analysis, in spite of its striking appear-
ance, constitutes a drawback in terms of learnability. In general, the language
learner is presented with extra difculties in learning a rule whose effects are
obscured at the surface level. The stress that is present at the intermediate level
of representation (where it blocks i-Syncope) is no longer present at the surface
level, and hence has to be inferred or reconstructed by a morphological

12
This objection does not hold for the prosodic theory of Lexical Phonology that was developed
by Inkelas (1989), which links the notion of cyclic category to its prosodic independence.

286
 6.4 The cycle versus base-identity

analysis of the relevant word. That is, the abstract stress in [fihmna] must
be related to the fact that it actually appears in [fhim]. Now observe that assum-
ing this morphological reconstruction process is tantamount to establishing a
paradigmatic link between a derived form and its base, precisely as in the OT
analysis. The only difference is that an OT analysis turns this reference to
morphologically related forms into a principle of grammar, taking the form of
OO-correspondence. But this step pays off, since it eliminates abstractness. As
a general strategy, eliminating abstractness is attractive in terms of learnability.

However, it is fair to point out a liability of the OT analysis, one which deriva-
tional theory avoids. The derivational analysis assumes that phonological surface
specications of every output form are derived on the basis of information present
in lexical representations, plus its internal morphological structure. In a way, every
complex form is self-sufcient, which gives a restrictive local theory of trans-
derivational relations. This localism no longer holds under OO-correspondence,
where surface specication of a morphologically complex word becomes depend-
ent upon surface properties of other words. The important question is: to what
extent is Correspondence Theory able to impose restrictions on logically possible
paradigmatic relations?13 Does the elimination of abstractness compensate for
this increase in globality? At present, it is very difcult to answer this question.
However, whatever loss of restrictiveness may occur as a result of the increase in
globality must be compensated for by constraining the notion of paradigmatic
relation. The denition of base, as presented here, is an attempt at constraining
the notion of OO-correspondence.

6.4.4 Epenthesis and opacity of metrical structure

We now turn to opacity of metrical structure. We will rst discuss the process of
epenthesis as a source of metrical opacity, and its interactions with base-identity
and faithfulness. Then we will discuss the metrical constraints proper, ranking
them with respect to the constraints responsible for epenthesis and syncope.
Palestinian Arabic has a process of i-Epenthesis, which inserts [i] between the
rst and second consonant in a sequence of three consonants, or between two
consonants at the end of the word:

(78) i-Epenthesis
i / C ___ C
!C #
@# $

13
Various answers to this question were given by Benua (1995), Buckley (1995), Flemming
(1995), Orgun (1994, 1995, 1996), Kenstowicz (1996), Steriade (1996), Kager (forthcoming),
and others.

287
Output-to-output correspondence

This process is the source of opaque stress in possessive examples such as those
in (79a.iii, 79b.iii) below. Epenthetic vowels appear in bold-face:
(79) Epenthesis and opaque stress in Palestinian Arabic
a.i /hm/ f.him understanding
a.ii /hm-u/ f h.mu his understanding
a.iii /hm-na/ f.him.na *.hm.na our understanding
b.i /akl/ .kil food
b.ii /akl-u/ k.lu his food
b.iii /akl-ha/ .kil.ha *a.kl.ha her food
Stress is opaque in the sense that it fails to accord with surface syllabication.
Palestinian generally stresses the penultimate syllable if it is heavy, and otherwise
the antepenult. (Here we disregard nal stress on superheavy syllables.) Never-
theless, forms (79a.iii b.iii) are stressed on their antepenults, in spite of their
closed (heavy) penults.

In a derivational analysis, this interaction involves ordering i-Epenthesis after

stress. The derivation in (80) shows how i-Epenthesis comes too late to deter-
mine surface stress:
(80) /hm/ /hm-u/ /hm-na/
Stress Cycle 1 fhm fhm fhm
Cycle 2 fhm fhm-u fhm-na
i-Epenthesis fhim --- fhim-na
[fhim] [fhmu] [fhimna]
Again the question arises as to how OT captures the opacity. Let us rst see what
constraint interaction produces i-Epenthesis.
Epenthesis is triggered by a combination of two high-ranked constraints on
syllable well-formedness (81ab), which we will combine into a cover constraint
Syll-Form:
(81) Syll-Form
a. *Complex (cf. Prince and Smolensky 1993)
No complex syllable margins.
b. Son-Seq (cf. Clements 1990)
Complex onsets rise in sonority, and complex codas fall in sonority.

Avoidance of syllabic ill-formedness takes priority over avoidance of epenthetic

vowels. Or, to state it in terms of constraint ranking: *Complex
Dep-IO.14
14
Perhaps more appropriately, *CCC, since at the word beginning complex onsets are allowed,
and actually created by the deletion of input vowels (cf. [fhm-na]).

288
 6.4 The cycle versus base-identity

Violations of *Complex can be avoided in two different ways, that is, by epen-
thesis between the rst and the second consonant, or after the second consonant.

(82) a. C_CC f.him.na

b. CC_C *f h.mi.na

In a derivational analysis, the locus of epenthesis is stated in the rule. More

sophisticated derivational analyses would predict the locus from directional syl-
labication (see It 1989). Neither of these options is available in an OT analysis.
There are no rules in OT, let alone directional rules.
Once again, base-identity is involved. If we consider the optimal output form
[f.him.na], we nd that this maximally resembles its base, in the sense that both
vowels appearing in the base [f.him] are preserved in the afxed form. The fact
that one of the vowels (underlined in 83a) is epenthetic with respect to the input
is simply irrelevant for base-identity:

(83) a. /f i h m -n a/ b. /f i h m -n a/ Input

[f h i m -n a] [f h m i -n a] Output

[f h i m] [f h i m] Base

Correspondence between base and output is broken in (83b), where the epenthetic
vowel [i] lacks a correspondent in the base. This points to a base-identity con-
straint of the Max type, requiring that base segments have correspondents in the
output (the afxed form):

(84) Max-BA
Every segment in the base has a correspondent in the afxed form.

Max-BA is more general than HeadMax-BA as it does not refer to stress in the
base vowel. Both constraints also differ in their positions in the ranking, crucially
with respect to No [i]. Max-BA outranks No [i], whereas HeadMax-BA ranks
below it, as the following syncope data show:

(85) a. HeadMax-BA
No [i] /him-na/ he understood us

[hm-na]
[fhm-na]
b. No [i]
Max-BA /him-u/ he understood him
[fhm-u]
[fhim-u]

The next question is: if the (bold-face) vowel in [f.him.na] our understanding
copies the epenthetic vowel in its base [f.him] understanding, then what pre-
dicts the epenthesis site in the base? We must consider two candidates, one
289
Output-to-output correspondence

with stem-internal epenthesis, [f.him], the other with stem-external epenthesis,

[fh.mi]. The former is selected by a Generalized Alignment constraint (chapter
3, after McCarthy and Prince 1993a):

(86) Align-R
]GrWd = ]

This constraint is apparently undominated in Palestinian Arabic.

Let us briey consider the featural content of the epenthetic vowel, which
depends on featural markedness (chapter 3). Epenthetic [i] is apparently less
marked than any other vowel. This result follows from the context-free marked-
ness constraint *[high]:

(87) high]
*[

In the tableaux below we will not include this constraint, and tactitly assume it
by including only [i] as an epenthetic vowel.15

After this preliminary analysis, let us now localize the source of metrical opac-
ity. As we have seen, opacity resides in the lack of accordance between stress
and syllabication in surface forms. Palestinian Arabic generally assigns stress to
any heavy penultimate syllable. Then why is epenthetic [i] in [f.him.na] and
[.kil.ha] unstressed even though it stands in a closed penult, a syllable that
normally attracts stress? The rejection of stress by epenthetic vowels is in fact a
cross-linguistically common phenomenon (Piggott 1995, Alderete 1995). To cap-
ture it, Alderete (1995) proposes a head dependence constraint requiring that
stressed vowels must have input correspondents:

(88) HeadDep-IO
Every vowel in the outputs prosodic head has a correspondent in the
input.

Epenthetic [i] in [f.him.na] lacks a correspondent in the input, as illustrated by

the diagram in (89a). Therefore stressing it would violate HeadDep-IO. In con-
trast, the stressed vowel in [ba.kr.na] our cattle (89b), with a base [b.kar]
cattle, has an input correspondent and hence stressing it does not violate
HeadDep-IO.

15
See Abu-Salim (1980) for data showing that [u] patterns much like [i] in syncope and
epenthesis.

290
 6.4 The cycle versus base-identity

(89) a. /f i h m -n a/ b. /b a k a r -n a/ Input

[f h i m -n a] [b a k r -n a] Output

[f h i m] [b k a r] Base

Indirectly, form (89b) rules out an alternative hypothesis about opaque stress in
[f.him.na], according to which a vowel that is stressed in the output must have a
stressed correspondent in the base. This hypothesis is ruled out since the stressed
output vowel of form (89b) has no stressed correspondent in the base.

Metrical opacity can now be modelled as the domination of an identity con-

straint over a metrical well-formedness constraint. More precisely, HeadDep-IO
dominates the well-formedness constraint which is responsible for stress on heavy
(penultimate) syllables (the Weight-to-Stress Principle, or WSP, Prince and Smo-
lensky 1993):

(90) Metrical opacity in Palestinian Arabic

HeadDep-IO
WSP

The reverse ranking would produce penultimate stress, e.g. *[hm-na], rather
than [fhim-na].
The analysis of vowelzero alternations in Palestinian is illustrated by three
tableaux of minimally contrasting output forms: [hm-na] he understood us
in (91), [fhm-na] we understood in (92), and [fhim-na] our understanding
in (93).

(91)
Input: /him-na/ Syll- HeadMax No [i] HeadDep- Max-BA WSP
Base: [f.him] Form -BA IO

a. + [.hm.na] *
b. [f .him.na] * *!
c. [f h.mi.na] * *!
d. [fhm.na] *! *
e. [f hm.na] *! *

291
Output-to-output correspondence

(92)
Input: /him-na/ Syll- HeadMax No [i] HeadDep- Max-BA WSP
Base: none Form -BA IO

a. [.hm.na] *!
b. [f .him.na] *! *
c. [f h.mi.na] *!
d. + [fhm.na]
e. [f hm.na] *!

The nal tableau is the crucial one: it shows the opacity effect with respect to
epenthetic [i]:

(93)
Input: /him-na/ Syll- HeadMax No [i] HeadDep- Max-BA WSP
Base: [f.him] Form -BA IO

a. [.hm.na] * *!
b. + [f.him.na] * *
c. [f h.mi.na] * *!
d. [fhm.na] *! *
e. [f hm.na] *! *

Finally we are in a position to substantiate the claim that the interaction of

epenthesis and stress in Palestinian Arabic involves the parallel evaluation of
IO-faithfulness and OO-identity. Such parallelism is demonstrated by the act-
ivity, within the same hierarchy, of both constraint types. Note that within a
single output form [f.him.na] (93b) the bold-face vowel is treated as epenthetic
with respect to the input (which is why it cannot surface as stressed due to
HeadDep-IO), while it is, paradoxically, treated as non-epenthetic with respect to
the base (which is why it must be retained due to Max-BA).

6.4.5 Conclusions
The OT analysis of Palestinian Arabic stress and vowelzero alternations leads to
a number of conclusions with respect to the morphologyphonology interface. On
rst inspection, the opacities in these patterns seem to be problematic to OT, and
292
 6.5 Output-to-output correspondence: conclusions

to support intermediate levels of representation and the derivational mechan-

ism of the cycle. However, as we now see, an analysis which is based on OO-
correspondence is equally adequate in descriptive terms, and it has the additional
conceptual advantage of eliminating abstractness. A denition of base was pro-
posed, as a free-standing form compositionally related to the complex word.
Finally, the evaluation of OO-correspondence and IO-correspondence in output
forms turns out to be a parallel process.

6.5 Output-to-output correspondence: conclusions

The goal of this chapter was to discuss extensions of Correspondence Theory
to two novel empirical domains, morphological truncation and stem-based
afxation. It may be too early for an evaluation of the overall merits of the base-
identity approach, but certainly the rst results are encouraging.
Brames famous cyclic analysis of vowelzero alternations and stress in Pales-
tinian Arabic has stood as a major example of intermediate levels in phonology
for a long period. The correspondence analysis maintains its essential insight
that morphologically complex forms depend on less complex forms in their pho-
nological shape, but it eliminates the serial (or transderivational) nature of the
dependence relation. At the same time it eliminates the abstractness involved in
the serial analysis. Brames analysis was abstract in the sense that a blocking of
syncope was attributed to a pre-surface stress pattern in the complex form rather
than to the stress pattern in the base form, where it is overt.
The consequences of this base-identity approach for other empirical domains
remain to be determined by careful analysis. For example, to what extent can the
notion of possible paradigmatic relation be constrained in Correspondence The-
ory? Another focus of attention of future OT research will be on phenomena that
have been argued to require intermediate levels of derivation in serial theory.
Among these are a set of phenomena that were analysed by so-called counter-
feeding and counterbleeding rules orders in serial theory. That is, OT must
develop a theory of opacity. We will return to this important issue in chapter 9.

SUGGESTIONS FOR FUR THER READING

Identity in truncation
Benua, Laura (1995) Identity effects in morphological truncation. In Beckman,
Walsh Dickey, and Urbanczyk. 77136. [ROA-74, http://ruccs.rutgers.edu /
roa.html]
Colina, Sonia (1996) Spanish truncation processes: the emergence of the
unmarked. Linguistics 34. 1199 218.
293
Output-to-output correspondence

Identity in afxation
Burzio, Luigi (forthcoming) Cycles, regularization, and correspondence. To
appear in J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimal-
ity Theory: syntax, phonology, and acquisition. Oxford University Press.
It, Junko and R. Armin Mester (1997) Correspondence and compositionality: the
Ga-gyo variation in Japanese phonology. In Roca. 419 62.
Kager, Ren (forthcoming) Surface opacity of metrical structure in Optimality
Theory. To appear in Hermans and van Oostendorp.
Kenstowicz, Michael (1996) Base-identity and uniform exponence: alternatives to
cyclicity. In J. Durand and B. Laks (eds.), Current trends in phonology: models
and methods. CNRS, Paris X, and University of Salford: University of Sal-
ford Publications. 36393. [ROA 103, http://ruccs.rutgers.edu /roa.html]

The phonologymorphology interface

Inkelas, Sharon (1996) Dominant afxes and the phonologymorphology inter-
face. In U. Kleinhenz (ed.), Interfaces in phonology. Studia Grammatica 41.
Berlin: Akademie Verlag. 12854.

EXERCISES

1 Spanish
Spanish (Harris 1983) neutralizes place of articulation (to alveolars) in coda
consonants. Neutralization is illustrated by the following alternations, where input
palatals /, / in coda position surface as alveolars [l, n], respectively:

/done-a/ donea lass /desde-ar/ desdear to disdain

/done-a-s/ doneas lass, pl. /desde-os-o/ desdeoso disdainful
/done/ donel lad, sg. /desde/ desden disdain, sg.

Neutralization of place of articulation overapplies in the following plurals:

/done-es/ doneles lad, pl. /desde-es/ desdenes disdain, pl.

a. Rank the constraints pertaining to coda neutralization of place features.

b. Next analyse the overapplication effect, proposing an OO-corres-
pondence constraint and ranking it with respect to the constraints you
found in (a).
c. Discuss any assumptions that you have found necessary with respect
to the notion base, then explain the contrasts [donea] ~ [don-
eles] and [desdeoso] ~ [desdenes].
d. Present tableaux of [donea], [donel], and [doneles].
294
 6.5 Output-to-output correspondence:
5.7conclusions
Exercises

2 Belfast English
In Belfast English (Harris 1990, Borowsky 1993) alveolar non-continuants [t, d,
n, l] are dentalized to [4, 1, 3, 2], respectively, before rhotics [r], [r]:

fat [ft] late [let] train [4ren] matter [m4r]

wide [wad] dine [dan] drain [1ren] spider [spa1r]
ll [fl] kill [kl] ladder [l1r] pillar [p2r]

However, dentalization fails to apply before the comparative and agentive sufxes
/-r/, even though these are phonetically indistinguishable from the endings in
spider etc.:

[4] matter [t] fatter

[1] spider [d] wider
[2] pillar [l] ller

a. Propose an analysis of the allophonic variation between alveolars and

dentals, setting up a contextual markedness constraint. (Assume that
dentals are [+distrib], while alveolars are [distrib].) To show the
allophonic nature of the distribution, present two tableaux of train,
one for each possible input value of [distrib] (/tren/ and /4ren).
b. Next analyse the underapplication of dentalization, introducing a new
OO-correspondence constraint pertaining to [distrib], and ranking it
with respect to the constraints of (a).
c. Present tableaux of fat, matter, and fatter, for both possible input
values of [distrib].

3 Factorial typology of OO-correspondence

a. Compute the factorial typology of IO-correspondence, OO-corres-

pondence, context-free, and contextual markedness constraints. Make
sure that you cover normal application, overapplication, and under-
application. (Follow the methodology of chapter 5, section 5.6.)
b. Compare your factorial typology with the one presented in chapter 5
on reduplication. Do you nd any signicant differences? In partic-
ular, does the factorial typology of OO-correspondence contain the
analogue of reduplicative back-copying? (Chapter 5, section 5.6.5.)
c. Construct an imaginary case of back-copying to the base in stem-
based afxation, and try to instantiate it by an actual case in a real
language. Comment on your ndings addressing any consequences
for the notion base, addressing any consequences for the notion
base.

295
 7
Learning OT grammars

7.1 Introduction
Generative theory assumes that the grammars of individual languages are all vari-
ations on a single theme, that of Universal Grammar. By hypothesis, UG is innate
rather than acquired, and hence it denes the range of possibilities within which
natural languages fall. From this viewpoint the study of rst language acquisition
becomes of crucial relevance to the study of natural language, as it may offer a
window onto properties of UG, hence into the human language faculty. The
central issue in rst language acquisition is this: how do principles of UG deter-
mine the acquisition process?
In previous studies, this issue of language acquisition has been approached from
a variety of viewpoints and methods. Some researchers study data from childrens
language, while others consider the preconditions that are necessary to make a
grammar learnable.
One eld of investigation is based on the collection of data from childrens
speech, in the form of corpora of spontaneous speech, or elicited through experi-
mental methods. Researchers study these data in close connection with assump-
tions about UG, and attempt to establish the relevant properties of UG which
guide the child during her acquisition of the target grammar (Nouveau 1994,
Demuth 1995, Gnanadesikan 1995, Goad 1997).
The second eld of investigation, which we will actually focus on in this
chapter, is learnability. Here the central issue is what formal and substantive
properties of universal grammar make it possible that grammars of individual
languages can actually be learned. OT makes a number of assumptions about
UG which are very different from other models, particularly Principles-and-
Parameters Theory (Wexler and Culicover 1980, Hyams 1986, Dresher and Kaye
1990), which denes the learners task as determining values of a set of universally
available binary choices, each corresponding to an inviolate property of the target
grammar. In contrast, OT assumes that UG denes a set of universal and violable

296
 7.2 Learning constraint rankings

constraints, as well as principles by which constraints interact. Individual lan-

guages differ along the dimension of constraint ranking (as well as in their lexi-
cons). If grammars are essentially rankings of universal constraints, then
acquiring a language must involve the acquisition of a language-specic hierarchy
of universal constraints. The language learner has exclusive access to the output
forms of the target language, and is faced with the task of extracting the informa-
tion from these outputs that is necessary to rank all constraints in a way predicting
the outputs. What strategies might the language learner use to bring this task to
a successful end? What guarantee is there that the child will eventually arrive at
the target grammar by a process of ranking and reranking constraints? These and
other questions will be central in this chapter.

This chapter is organized as follows. In section 7.2 we focus on the issue of the
learnability of OT grammars. Assuming that languages differ mainly by their
constraint rankings, can it be proved that it is actually possible for the language
learner to infer the correct ranking on the basis of output data? We will discuss
results of a learning algorithm developed by Tesar and Smolensky (1993), which
lead to a positive answer to the above question. The algorithm will be seen in
action in section 7.3. In section 7.4 we will discuss aspects of language acquisi-
tion which are currently not included in this model, and we will review some
attempts to include these. Finally, section 7.5 will sketch an approach to adapt the
algorithm to the complications involved in learning phonological alternations.

7.2 Learning constraint rankings

7.2.1 Dening the problem

The specic question which we will address in this section is this: given the
availability of a set of surface forms of the target language, and a set of universal
constraints, is it possible for the language learner to discover the correct con-
straint ranking of the target language? If so, what strategies does the learner use
to discover the proper ranking? We will discuss an approach to the issue of
learnability of OT grammars which was developed by Tesar and Smolensky
(1993). The learning algorithm serves as a model of the language learner, and it
is faced with the task of constructing a hierarchy of constraints for a language on
the basis of output forms of this language. Results of the algorithm are positive,
in the sense that it is indeed possible to deduce constraint rankings on the basis
of a given input plus an output. This result is of key importance to the assumption
that grammars are essentially language-specic hierarchies of constraints.

297
Learning OT grammars

7.2.2 The constraint demotion algorithm: basic ideas

An algorithm developed by Tesar and Smolensky demonstrates that it is possible
to deduce rankings of constraints on the basis of output forms plus a set of
universal constraints. The algorithm assumes that the input (the form from which
the attested output derives) is given. Moreover it assumes that the output is a
linguistically structured representation, rather than a raw phonetic form. Both
assumptions represent idealizations of a real-life learning situation (in which
inputs have to be inferred by the learner, and outputs are raw acoustic forms). In
a more sophisticated learning algorithm, these idealizations will have to be aban-
doned. We will address various possibilities of how this might be achieved in
sections 7.4 7.5.
The Tesar and Smolensky algorithm is based on an idea which we have, implic-
itly, made use of several times before in this book. That is, information about
constraint ranking can be extracted from the violation (rather than satisfaction)
of constraints in the optimal candidate. The key idea is that constraints which are
violated in the optimal output must be dominated by some other constraint. To
identify the relevant dominating constraint(s), the algorithm compares the attested
output to various other (suboptimal) candidates. For each pair consisting of the
attested output and a suboptimal candidate, it lists constraints which are violated
by one member of the pair, but not by the other, and vice versa. From this
crossed comparison, the algorithm then deduces the constraint hierarchy which
pairs the attested form as the optimal output to a given input. In this way, maximal
use is made of information provided by the single source of positive evidence:
output forms.
The central principle by which the algorithm applies is constraint demotion.
The initial state of the algorithm is one in which all constraints are unranked with
respect to one another. That is, all constraints are undominated. See (1):

(1) Initial state of the constraint hierarchy

{C1, C2, C3 . . . Cn}

Starting from this initial unranked state, the algorithm gradually develops a hier-
archy by reranking constraints. Reranking is only allowed if there is positive
evidence in the form of a constraint violation in the optimal output. Reranking
always involves the demotion (rather than promotion) of a constraint below
another constraint. Demotion is minimal in the sense that a constraint is demoted
to a position immediately below the highest-ranking constraint that induces its
violation in the optimal output. See (2), in which constraints are indicated by the
symbol C plus a subscript:

298
 7.2 Learning constraint rankings

(2) Demote C2 below C4 Demote C3 below C4 Demote C1 below C3

{C1 . . . C3, C4 . . . Cn} {C1 . . . C4 . . . Cn} {C4 . . . Cn}




{C2} {C2, C3} {C2, C3}


{C1}

Sets of constraints included in pairs of curly brackets { . . . } are unranked with

respect to one another. Such a set of constraints is called a stratum. A stratum
may also consist of a single constraint, as illustrated by stratum {C2} in the
hierarchy which arises after the demotion of this constraint. A hierarchy contain-
ing (one or more) strata is called a stratied hierarchy.

(3) A stratied hierarchy

{C1 . . . Cn}


{Cn+1 . . . Cp}


...
{Cq+1 . . . Cr}

It is important to bear in mind that the stratied hierarchies which arise in the
course of progress of the algorithm are hypothetical. At each moment during the
learning process, stratied hierarchies represent the current knowledge which
the learner at that moment has accumulated about constraint interactions under-
lying a given output form. This knowledge is dynamic, as it continuously changes
while the learning algorithm processes information from output forms. After all
information has been absorbed from some output form in the form of rankings,
the current hierarchy may already fully match the hierarchy of the target gram-
mar. But it is equally possible that the hierarchy is still incomplete, meaning that
one or more constraints have not yet been assigned to their proper positions.
Crucial evidence required to complete the hierarchy may reside in new output
forms, which the algorithm has not yet considered. In this sense, the learner can
never be sure that the acquisition process has terminated after any output form.
However, any new output form the learner encounters may only serve to rene
the current hierarchy, rather than totally redene it. This is because (by hypothe-
sis) all output forms of the target language consistently reect a single constraint
hierarchy.
We now turn to a demonstration of Tesar and Smolenskys algorithm in action.

299
Learning OT grammars

7.3 Learning the Pintupi grammar of stress

7.3.1 The target grammar

The learning problem to be solved is this: to deduce the ranking of constraints
correctly predicting the stress pattern of the Australian language Pintupi (Hansen
and Hansen 1969; Hayes 1995). Recall from chapter 4 (section 4.4.1) that Pintupi
words have initial primary stress, and secondary stresses on alternating odd-
numbered syllables. Final syllables are always unstressed. An example of a ve-
syllable word, together with its metrical parsing, is given in (4):
(4) (p.i).(kB.l).tju we (sat) on the hill
The Pintupi pattern is due to an interaction of universal and violable metrical
constraints, among which are the following six (from chapter 4, section 4.4.1 and
section 4.4.3):
(5) a. Ft-Bin
Feet are binary under moraic or syllabic analysis.
b. Parse-Syl
Syllables are parsed by feet.
c. Leftmost
Align (Hd-Ft, L, PrWd, L) The head foot is initial in PrWd.
d. Rightmost
Align (Hd-Ft, R, PrWd, R) The head foot is nal in PrWd.
e. All-Ft-Left
Align (Ft, L, PrWd, L) Feet are initial in PrWd.
f. All-Ft-Right
Align (Ft, R, PrWd, R) Feet are nal in PrWd.
The grammar of Pintupi ranks these constraints in the following way (the
notation identies four strata in the hierarchy):
(6) A fragment of the grammar of Pintupi
{Ft-Bin, Leftmost}


{Parse-Syl, Rightmost}


{All-Ft-Left}


{All-Ft-Right}
For completeness this ranking is illustrated by a tableau for the input /puikltju/
in (7):
300
 7.3 Learning the Pintupi grammar of stress

(7)
Input: /puikltju/ Ft- Left- Parse- Right- All-Ft-L All-Ft-R
Bin most Syl most

a. + (p.i).(kB.l).tju * *** ** *, ***

b. (p.i).k.(lB.tju) * *** ***! ***

j
c. (p.i).k.l.t u **!* *** ***

d. pu.i.k.(lA.tju) *!** *** ***

j
e. (p.i).(kB.l).(t ) *! *** **, **** *, ***

As always, violations of foot alignment constraints are marked in syllables.

7.3.2 Dening the task of the learning algorithm

Thus far we have tacitly assumed that the hierarchy in (6), as any hierarchy which
we have posited so far, is deducible from output forms. But is this implicit
assumption correct? May we indeed assume that output forms sufce to charac-
terize uniquely the grammar which generates them? If this turns out not to be the
case, then the grammar of Pintupi, or of any other language, would not be learn-
able. This would pose an unsurmountable problem for the linguistic theory at
stake, since any theory which posits unlearnable grammars cannot be a correct
theory of the human language capacity.
Therefore what we must do is put the theory under discussion, Optimality
Theory, to the test, and nd out whether an OT grammar is actually learnable. To
this end, we will use a model of the language learner. This model will reect cer-
tain assumptions about the organization of the grammar, specically the assump-
tion that the grammar of any natural language consists of a strict hierarchy of
universal and violable constraints (among other aspects, such as a lexicon).
The model takes the shape of a learning algorithm, which is nothing but a
series of instructions which, if blindly followed, will lead to a certain goal. What
we want from the learning algorithm is stated in general terms in (8):

(8) Task of the learning algorithm

To deduce the constraint hierarchy under which the given surface
form emerges as the optimal output of the given input form.

The specic task of the algorithm in the Pintupi learning situation is to deduce
the ranking of constraints in (5) which correctly predicts the given output form
[(p.i).(kB.l).tju] on the basis of a given input /puikltju/. To this end, the
learning algorithm is provided with precisely this information:
301
Learning OT grammars

(9) a. UG, including the six universal constraints in (5)

b. An output form [(p.i).(kB.l).t ju]
c. An input form /puiklt ju/

7.3.3 The initial state, and Mark-Data Pairs

What the algorithm does rst is initializing the hierarchy. Recall that the algo-
rithm initially expects that all constraints are unranked. This is equal to the
assumption that all constraints are undominated, sharing a single stratum in the
hierarchy:

(10) Initial state of the constraint hierarchy

{All-Ft-L, All-Ft-R, Ft-Bin, Leftmost, Parse-Syl, Rightmost}

Out of this initial single mother stratum, the learning algorithm will develop a
complete hierarchy by a consecutive demotion of constraints into new lower
strata. Demotion will take place on the basis of information which is deduced
from a comparison of constraint violations in the optimal form and those in
suboptimal candidates. Such comparison of candidates requires some crucial pre-
processing of information, to which we now turn.
As stated earlier, we assume that the learning algorithm is fed with the surface
form [(p.i).(kB.l).t ju], including its metrical structure, and together with the
information that its input is the form /puiklt ju/ we (sat) on the hill.

(11) Input: /puiklt ju/ we (sat) on the hill

Output: [(p.i).(kB.l).t ju]

The task for the learner (in its model form, the algorithm) is to deduce the
grammar which leads to this particular pairing of input and output. Deducing
the grammar means: ranking universal constraints into a hierarchy under which
[(p.i).(kB.l).tju] is the most harmonic output form of /puiklt ju/, out of all
possible output forms. This implies that all possible output forms of /puiklt ju/
are somehow available to the learner, together with violation marks for specic
constraints which render them suboptimal. But are suboptimal forms available? It
has been observed by generative linguists time and again that a child learning her
native language has access to positive evidence only. That is, the ungrammatical-
ity of forms is never directly offered as data to the learner.1
In OT terminology, the child has direct access to optimal outputs of the gram-
mar, while suboptimal output forms (negative evidence) are not part of the
childs linguistic environment. For the sake of the argument, let us reduce the data
available to the learner of Pintupi (who is faced with the task of constructing the

1
Except in cases where the child is explicitly corrected.

302
 7.3 Learning the Pintupi grammar of stress

metrical grammar in 6) to only a single piece of positive evidence, that is, the
output form (11). Consequently the learner is deprived of any suboptimal outputs
of /puiklt ju/, such as candidates (7be) in the above tableau. Thus far in this
book, we have built arguments for constraint rankings on the comparison of
suboptimal candidate outputs with the actual optimal output. This is the only
method that produces valid ranking arguments. But if (as we now assume) the
learner has no access to suboptimal forms, then how could the relevant rankings
ever be inferred? It seems that the learner has got stuck before she has even
started. How to break the spell?
The learners strategy is to make maximal use of information which is implicit
in the surface form, but which is nevertheless reconstructable. If [(p.i).(kB.l).t ju]
is the optimal output of the input /puiklt ju/, then any output candidate must
be suboptimal, by denition. The learner may freely posit literally any output can-
didate as a competitor of [(p.i).(kB.l).t ju], and still be sure that this hypothetical
competitor is less harmonic. The algorithm capitalizes on this, and generates a set
of suboptimal candidates,2 which are the competitors for [(p.i).(kB.l).t ju], for
example the following set:

(12) [(p.i).k.(lB.t ju)], [(p.i).k.l.t ju], [pu.i.k.(lA.t ju)],

[(p.i).(kA.l).(t j)]

She is then automatically served with information packages of the following kind:

(13) subopt  opt

ba [(p.i).k.(lB.t ju)]  [(p.i).(kB.l).t ju]
ca [(p.i).k.l.t ju]  [(p.i).(kB.l).t ju]
da [pu.i.k.(lB.t ju)]  [(p.i).(kB.l).t ju]
ea [(p.i).(kB.l).(t j)]  [(p.i).(kB.l).t ju]

Each chunk of information consists of a harmonic ranking of two candidates, one

of which is a suboptimal output form (subopt), and the other the optimal output
form (opt).3

2
There is a procedure for automatically selecting informative suboptimal candidates, which is
called Error-Driven Constraint Demotion (Tesar 1995). See also note 7.
3
Given Freedom of Analysis, Gen will supply the learner with an innite number of such pairs.
Clearly not all such pairs are informative to the learner, due to the presence of a large number
of intrinsically suboptimal candidates (see again chapter 1, section 1.7.5). The issue of how
to deal with such noise in the data will not be discussed here; in any event, no negative effects
arise for the learner by considering intrinsically suboptimal candidates. Moreover, the algo-
rithm does not know beforehand how many mark-data pairs must be considered in order to
arrive at the target grammar, nor is it necessary to know this. The ideal learner will consider
all available information, after which she will positively have acquired the target grammar, as
will be shown below.

303
Learning OT grammars

This is the point at which constraint evaluation comes into play. OT is based
on the central claim that the observed surface form is the most harmonic output
(of all possible output candidates), given some specic ranking of violable con-
straints. The optimal output may violate constraints and in fact it always will.
However, the very fact that the observed output is optimal implies that no other
output is possible which is more harmonic, at least not under the same constraint
ranking. The learner can simply be certain that any constraint violation in the
observed output is forced. That is, if the optimal output violates a constraint C1,
then any suboptimal output which respects C1 must itself violate some higher-
ranking constraint C2 which is unviolated in the optimal output. (For if C1 were
not violated in the optimal form, then this form would have been suboptimal
itself.) The key to the problem of establishing the unknown constraint ranking is
precisely this property of OT grammars. The learner must identify the constraint
violations in the observed output (which is optimal by denition), matching these
against violations of other constraints in suboptimal outputs. From this informa-
tion, she deduces the target constraint hierarchy.

The learning algorithm starts by constructing an overview of constraint violations

for each pair in the list (13). The result is tabulated in (14). Each row represents
the marks of one pair in (13). For each candidate pair, a cell in the column loser-
marks lists all violation marks of the suboptimal candidate, while a cell in the
column winner-marks lists all violation marks of the optimal candidate. Viola-
tions are indicated by an asterisk before the name of the violated constraint.
Candidates inducing multiple violations of the same constraint are doubly marked
for this constraint. (As in tableau 7 of [(p.i).(kB.l).t ju].)

(14) Mark-data Pairs

subopt  opt loser-marks winner-marks

ba [(p.i).k.(lB.t ju)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *All-Ft-R, *All-Ft-R, *All-Ft-R,
*All-Ft-R, *All-Ft-R, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *R-most, *Parse-Syl, *R-most,
*R-most, *R-most} *R-most, *R-most}

ca [(p.i).k.l.t ju] {*All-Ft-R, *All-Ft-R, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-R, *Parse-Syl, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *Parse- *All-Ft-R, *All-Ft-R,
Syl, *R-most, *R-most, *Parse-Syl, *R-most,
*R-most} *R-most, *R-most}

304
 7.3 Learning the Pintupi grammar of stress

subopt  opt loser-marks winner-marks

da [pu.i.k.(lA.t ju)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *L-most, *All-Ft-R, *All-Ft-R,
*L-most, *L-most, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *Parse- *Parse-Syl, *R-most,
Syl, *Parse-Syl} *R-most, *R-most}

ea [(p.i).(kB.l).(t j)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *All-Ft-L, *All-Ft-R, *All-Ft-R,
*All-Ft-L, *All-Ft-L, *All-Ft-R, *All-Ft-R,
*All-Ft-R, *All-Ft-R, *Parse-Syl, *R-most,
*All-Ft-R, *All-Ft-R, *R-most, *R-most}
*Ft-Bin, *R-most,
*R-most, *R-most}

Observe that every pair of (13) contains the optimal output [(p.i).(kB.l).t ju].
This is a logical necessity, since the optimal candidate is the single one which is
presented to the learner. Since every harmonic comparison involves the optimal
output, every cell winner-marks will contain identical sets of violation marks. In
contrast, the contents of loser-marks differ from row to row.

7.3.4 Mark Cancellation

The table of mark-data pairs in (14) provides the raw information that the
algorithm processes in order to arrive at the target grammar. However, from this
table to the target is still a long way to go. It will be clear that in its present form,
the table cannot provide the basis for ranking constraints, since for each mark-
data pair, identical violation marks occur in both the cells loser-marks and
winner-marks. The algorithm must rst purify this information by cancelling
out marks from the table which have no information value.
The rst step to take is eliminating from this table any marks which are shared
by winner and loser. Given the logic of OT, shared violations can never result in
harmony differences between two candidates, hence cannot provide information
about constraint ranking. As we already mentioned, the raw information in table
(14) is puried by the cancellation of violation marks which have no informa-
tion value. Before we discuss how the algorithm performs this function, let us
take a look at the new mark-data pairs table (15), where cancelled marks have
been struck out.

305
Learning OT grammars

(15) Mark-data Pairs after Cancellation

subopt  opt loser-marks winner-marks

ba [(p.i).k.(lB.t ju)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *All-Ft-R, *All-Ft-R, *All-Ft-R,
*All-Ft-R, *All-Ft-R, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *R-most, *Parse-Syl, *R-most,
*R-most, *R-most} *R-most, *R-most}

ca [(p.i).k.l.t ju] {*All-Ft-R, *All-Ft-R, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-R, *Parse-Syl, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *Parse-Syl, *All-Ft-R, *All-Ft-R,
*R-most, *R-most, *Parse-Syl, *R-most,
*R-most} *R-most, *R-most}

da [pu.i.k.(lB.t ju)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *L-most, *All-Ft-R, *All-Ft-R,
*L-most, *L-most, *All-Ft-R, *All-Ft-R,
*Parse-Syl, *Parse-Syl, *Parse-Syl, *R-most,
*Parse-Syl} *R-most, *R-most}

ea [(p.i).(kB.l).(t j)] {*All-Ft-L, *All-Ft-L, {*All-Ft-L, *All-Ft-L,

 [(p.i).(kB.l).t ju] *All-Ft-L, *All-Ft-L, *All-Ft-R, *All-Ft-R,
*All-Ft-L, *All-Ft-L, *All-Ft-R, *All-Ft-R,
*All-Ft-R, *All-Ft-R, *Parse-Syl, *R-most,
*All-Ft-R, *All-Ft-R, *R-most, *R-most}
*Ft-Bin, *R-most,
*R-most, *R-most}
How does this cancellation take place? First we should realize that the absolute
number of violations of some constraint is never relevant to a candidates well-
formedness. Instead, what is relevant is the difference in violations of a constraint
with other candidates. For constraints which can have multiple violations per can-
didate, what really matters is which of the candidates violates the constraint worst
(rather than absolute numbers of marks). In order to establish a difference in
violations of a single constraint between two candidates, mark cancellation is done
on an item-by-item basis. See, for example, the rst pair in table (15), in which two
marks *All-Ft-L are cancelled, so that one mark remains in the cell loser-marks.
This insight is captured by the rst two steps of the algorithm Mark
Cancellation:
(16) Mark Cancellation
For each pair (loser-marks, winner-marks) in mark-data:
a. For each occurrence of a mark *C in both loser-marks and winner-
marks in the same pair, remove that occurrence of *C from both.
306
 7.3 Learning the Pintupi grammar of stress

b. If, as a result, no winner-marks remain, remove the pair from

mark-data.
c. If, after the preceding steps, a row of the mark-data table contains
multiple tokens of the same type of mark, duplicates are elimin-
ated, leaving at most one token of each type.

Step (16c) achieves a further purication of the table by removing multiple

tokens of the same mark, which may still occur after cancellation. As stated
earlier, all that matters to the learner is which of the candidates in the pair has
the largest violation of a constraint; it is therefore irrelevant to know the exact
difference in degree of violation.
Table (17) represents the result of the elimination of cancelled marks (steps
16bc).

(17) Mark-data Pairs after Mark Cancellation

subopt  opt loser-marks winner-marks

ba [(p.i).k.(lB.tju)] {*All-Ft-L} {*All-Ft-R}

 [(p.i).(kB.l).tju]
ca [(p.i).k.l.tju] {*Parse-Syl} {*All-Ft-L,
 [(p.i).(kB.l).tju] *All-Ft-R}
da [pu.i.k.(lA.tju)] {*All-Ft-L, *Leftmost, {*All-Ft-R,
 [(p.i).(kB.l).tju] *Parse-Syl} *Rightmost}
ea [(p.i).(kB.l).(tj)] {*All-Ft-L, *Ft-Bin} {*Parse-Syl}
 [(p.i).(kB.l).tju]
After this preprocessing by Mark Cancellation, the information generated from
the output form [(p.i).(kB.l).t ju] is now ready to be fed into the core of the
learning algorithm, which deduces the actual constraint hierarchy, and which is
called Recursive Ranking by Tesar and Smolensky.

Before we look into the actual statement of Recursive Ranking, we must rst
expand the algorithms work-slate beyond the mark-data pairs table in (17). From
now on the algorithm will register rerankings in the constraint hierarchy, which
will acquire more and more structure until it reaches a point at which no addi-
tional rerankings are motivated on the basis of positive evidence available. At that
point, the learning process will have terminated, and hopefully we will then nd
that the algorithm has established a constraint hierarchy which matches the frag-
ment of the grammar of Pintupi. The current state of the constraint hierarchy is
still identical to its initial state. Recall that this initial state is not a genuine

307
Learning OT grammars

hierarchy, but includes the entire set of universal constraints without any ranking
dened among them. All constraints are lumped together in a single stratum.

(18) H0 = {All-Ft-L, All-Ft-R, Ft-Bin, Leftmost, Parse-Syl,

Rightmost}

Consecutive states of the hierarchy which result from a demotion of constraints

from this initial hierarchy will be notated as H1, H2 . . . Hn, etc.

7.3.5 Recursive Ranking: basic ideas

The basic ideas underlying Recursive Ranking have already been sketched in
section 7.2. Now it is time to discuss these ideas in more detail. The rst idea is
that only constraint violations (not satisfactions) are relevant to constraint rank-
ing. Constraints which are satised in the optimal output are less informative
since constraint satisfaction need not reect anything about some constraints
position in the hierarchy. Consider the fact that in all four-syllable words in
Pintupi, for example [(mA.).(wB.n)] through from behind, Parse-Syl is
satised. Yet, we may not conclude from this that Parse-Syl is undominated. And
indeed, output forms with an odd number of syllables ([(p.i).(kB.l).t ju]) show
that Parse-Syl is dominated. We may conclude that no solid ranking arguments
can be based on constraint satisfaction. However, crucial information on ranking
can be deduced from constraint violations. This is because we know for sure that
a constraint which is violated in the optimal output must be dominated by some
other constraint.4 For if this constraint were not violated in the optimal output,
then the optimal output would not be optimal. By denition, the optimal output
cannot be improved, under any given constraint hierarchy.
This important idea is captured in the algorithm by constraint demotion, trig-
gered by positive evidence from mark-data pairs. Generally speaking, the algo-
rithm demotes constraints assessing winner-marks below constraints assessing
loser-marks. The rationale behind this is the following. The single reason why the
optimal output is actually optimal is that its overall evaluation by the constraint
hierarchy is superior to any of its suboptimal competitors. Therefore, when the
winner fares worse on some constraint than the loser, then it must be the case
that this constraint Cwinner is dominated by another constraint Closer on which the
loser fares worse than the winner. Given the positive fact that the winner
is the winner, and the loser is the loser, the constraints can only be ranked
as below:
4
To be more precise and technical, a top-ranked constraint can be violated by the optimal output
as long it is violated at least as many times by every other output. Thanks to Bruce Tesar for
pointing this out. (Such a situation may occur, in principle at least, if Gen places restrictions
on all output candidates.)

308
 7.3 Learning the Pintupi grammar of stress

(19) Closer Cwinner

opt (winner) *
subopt (loser) *!

If the constraints were ranked reversely, then the winner would become the
loser, and the loser would become the winner. See tableau (20):

(20) Cwinner Closer

opt (no-longer-the-winner) *!
subopt (no-longer-the-loser) *

Therefore, given mark-data as in (21), in which both cells loser-marks and

winner-marks contain one constraint each (Closer and Cwinner):

(21) Mark-data
subopt  opt loser-marks winner-marks

loser  winner {Closer} {Cwinner}

the algorithm can legitimately deduce the ranking in (22):

(22) {Closer}


{Cwinner}

The algorithm will achieve this by demoting Cwinner to a stratum below Closer.

A question promptly arises: what happens when more than one constraint occurs
in loser-marks? More specically, to what stratum should a constraint Cwinner be
demoted when the constraints in loser-marks occupy different strata in the hier-
archy? Should all constraints come to dominate Cwinner, as depicted in (23a)? Or
should one of the constraints in loser-marks be selected as dominating Cwinner, for
example the highest-ranked among them, as in (23b)?

(23) a. {Closer-1} b. {Closer-1}

{Closer-2 , Closer-3} {Cwinner}



{Cwinner} {Closer-2 , Closer-3}

309
Learning OT grammars

This issue is not void of empirical content since the rankings in (23a) and (23b)
may give rise to different optimal outputs. Is there a hard-and-fast rule which will
always produce a correct result? To this issue we will now turn.

The second general idea underlying Recursive Ranking is that all constraints are
in the highest possible stratum in the hierarchy which is compatible with the
evidence from mark-data (which has been considered so far). In the initial state
of the hierarchy, this idea is seen in a most exemplary way: no evidence from
mark-data has been considered in that stage, so that all constraints are granted the
status of being undominated. Demotions of a constraint are always triggered by
positive evidence from mark-data, and even then, no constraint is demoted farther
down than can be positively motivated by the marks pattern. This implies that a
constraint assessing a winner-mark (Cwinner) is demoted to the highest possible
stratum in the hierarchy which is still compatible with mark-data. That is, Cwinner
is demoted to a stratum immediately below the highest-ranking constraint Closer in
the current hierarchy. Or to state it differently, we opt for demotion strategy (23b),
rather than (23a). We safely infer that this constraint cannot be ranked higher in
the hierarchy than it should be; for if it were higher-ranked, then it would not be
violated in the optimal output. (See again tableaux 1920.) Each constraint is
given the benet of the doubt: the algorithm considers it to be as high in the
hierarchy as can be, keeping open the possibility that new evidence may eventu-
ally turn up triggering additional demotions of the constraint. This was referred
to as a conservative demotion strategy earlier in this section. The full advantages
of this strategy become apparent only when we consider alternatives.
Suppose that constraints assessing winner-marks (Cwinner) were instead demoted
to a stratum below the lowest-ranked constraint in loser-marks. Chances are fair
that a demoted constraint falls too deep, so that it must later be rehabilitated by
demotions of other constraints which it had passed on its way down. But when
these new demotions also occur at random intervals, a never-ending jumping
over of constraints may start. Consequently the acquisition process will never
converge into the stable state of the target grammar.
This is illustrated schematically in (24), in the form of a cyclic demotion
process of three stages. Observe that the result of the nal stage (Hn+3) is identical
to the input to the initial stage (Hn). Demoted constraints appear in bold-face,
while . . . indicates a site from which a constraint has been demoted. This loop
is eternal since the learner will never reach the stable state in which no more
demotions can be justied by the data. The alternative strategy of minimal demo-
tion, which is part of the Tesar and Smolensky algorithm, serves to guarantee that
the learning process converges into a stable ranking. We will now see that this
indeed works.
310
 7.3 Learning the Pintupi grammar of stress

(24) Stage 1 Stage 2 Stage 3

Cwinner: C3, C6 Cwinner: C4 Cwinner: C5, C6
Closer: C2, C4, C5 Closer: C1, C5 Closer: C4
Hn Hn+1 Hn+2 Hn+3 (=Hn )
{C1, C2} {C1, C2} {C1, C2} {C1, C2}





{C3, C4} { . . . C4} { . .. }




{C5, C6} {C5 . . . } {C5 . . . } { ...}




{C6 , C3} {C6, C3, C4} { . . . C3, C4}


{C5, C6}

Consider the part of Tesar and Smolenskys algorithm which captures this key
idea, Constraint Demotion:

(25) Constraint Demotion (repeat this step until no demotions occur)

For each pair loser-marks  winner-marks in mark-data
i. nd the mark *Closer in loser-marks which is highest-ranked in H
ii. for each *Cwinner in winner-marks
if Closer does not dominate Cwinner in H, then demote constraint
Cwinner to the stratum H immediately below that of Closer (creat-
ing such a stratum if it does not already exist).

Observe that Constraint Demotion considers the pairs in mark-data one by one,
extracting information on constraint rankings from their distribution of marks.
Constraint Demotion is recursive.5 That is, it is repeated until no further demo-
tions occur. In our example this stable state of the hierarchy will arise after nine
steps, when the complete hierarchy of six metrical constraints of the target
grammar of Pintupi will have been deduced. Crucially, the order in which mark-
data pairs are examined by Constraint Demotion makes no difference to the nal
result, although it may be necessary to examine some mark-data pairs more than
once. In our test-case below we will actually select the least efcient way to reach
the target grammar, to emphasize the fact that the correct grammar emerges
regardless of the order in which mark-data pairs are examined.

5
We will present the recursive version of Tesar and Smolenskys constraint demotion algorithm,
entirely for expository reasons. It should be mentioned that there is also a non-recursive
version of the algorithm, also presented in Tesar and Smolensky (1993), which produces the
same results in a somewhat different fashion. This version is actually assumed in most of the
literature.

311
Learning OT grammars

7.3.5.1 Recursive Ranking in action

The rst mark-data pair in table (17) which Constraint Demotion considers is the
one in the top row, that is, b  a or [(p.i).k.(lB.t ju)]  [(p.i).(kB.l).t ju]:

(26) Mark-data (step 1)

subopt  opt loser-marks winner-marks

ba [(p.i).k.(lB.tju)] {*All-Ft-L} {*All-Ft-R}

 [(p.i).(kB.l).tju]

This pair embodies the generalization that Pintupi prefers a left-oriented distribu-
tion of feet to a right-oriented one. Constraint Demotion is able to deduce the
correct ranking:

The mark *All-Ft-L is the highest-ranked in loser-marks in the current hier-

archy, H0. (This is trivial, since it is the only mark in loser-marks.)
The constraint assessing the winner-mark (All-Ft-R) is not dominated by
Closer (All-Ft-L). (Again this is trivial, since the initial hierarchy H0 is totally
unranked.) Therefore All-Ft-R is demoted to a position immediately below
All-Ft-L, thereby creating a new stratum in the hierarchy:

(27) State of the constraint hierarchy after step 1

H1 = {All-Ft-L, Ft-Bin, Parse-Syl, Leftmost, Rightmost}


{All-Ft-R}

The second pair which Constraint Demotion considers is c  a in mark-data,

that is: [(p.i).k.l.t ju]  [(p.i).(kB.l).t ju]:

(28) Mark-data (step 2)

subopt  opt loser-marks winner-marks

ca [(p.i).k.l.tju] {*Parse-Syl} {*All-Ft-L,

 [(p.i).(kB.l).tju] *All-Ft-R}

This pair embodies the generalization about Pintupi that metrication of syllables
in feet takes priority over the (left edge and right edge) alignment of feet. Con-
straint Demotion deduces the corresponding rankings as follows:

*Parse-Syl is the highest-ranked loser-mark in the current hierarchy, H1.

(Trivially so, since it is the only constraint assessing loser-marks.)
None of the constraints in winner-marks (All-Ft-L, All-Ft-R) are domin-
ated by Closer (Parse-Syl). The rst, All-Ft-L, shares a stratum with Parse-

312
 7.3 Learning the Pintupi grammar of stress

Syl, while All-Ft-R is already dominated by Parse-Syl. Therefore only

All-Ft-L needs to be demoted to the stratum immediately below Parse-Syl.
It comes to share this stratum with All-Ft-R, which already occupies it:

(29) State of the constraint hierarchy after step 2

H2 = {Ft-Bin, Parse-Syl, Leftmost, Rightmost}


{All-Ft-R, All-Ft-L}

Observe that no new stratum is created for All-Ft-L. This is placed in the
stratum already occupied by All-Ft-R, in accordance with the conservative
demotion strategy: no constraint is demoted further down than is justied by the
evidence from mark-data.

The third mark-data pair to be considered by Constraint Demotion is that of the

third row in table (17), d  a, or [pu.i.k.(lA.t ju)]  [(p.i).(kB.l).t ju]:

(30) Mark-data (step 3)

subopt  opt loser-marks winner-marks

da [pu.i.k.(lA.t ju)] {*All-Ft-L, {*All-Ft-R,

 [(p.i).(kB.l).t ju] *Leftmost, *Rightmost}
*Parse-Syl}

This mark-data pair embodies the generalization that left-edge primary stress is
preferred to right-edge primary stress. Constraint Demotion processes the infor-
mation as follows:

*Leftmost and *Parse-Syl are the highest-ranked loser-marks (Closer) in the

current hierarchy, H2. Both are in the top stratum. (Observe that the third
loser-mark, *All-Ft-L, occupies a dominated stratum in H1 as a result of its
demotion in step 2.)
Only one constraint assessing winner-marks (Rightmost) is not already
dominated by Closer (Parse-Syl, Leftmost). Therefore only Rightmost is
demoted to the stratum below Closer, which was already occupied by All-Ft-
L and All-Ft-R:

(31) State of the constraint hierarchy after step 3

H3 = {Ft-Bin, Parse-Syl, Leftmost}


{All-Ft-R, All-Ft-L, Rightmost}

313
Learning OT grammars

The fourth mark-data pair to be considered is the one in the bottom row of table
(17), e  a or [(p.i).(kB.l).(t j)]  [(p.i).(kB.l).t ju]:

(32) Mark-data (step 4)

subopt  opt loser-marks winner-marks

ea [(p.i).(kB.l).(t j)] {*All-Ft-L, {*Parse-Syl}

 [(p.i).(kB.l).t ju] *Ft-Bin}

This harmonic ranking reects the generalization that degenerate feet are no licit
strategy to improve exhaustivity of parsing. This information is processed as
follows:

The mark *Ft-Bin is the highest-ranked loser-mark in the current hierarchy,

H3. (It is in the top stratum, while *All-Ft-L is in the bottom stratum.)
The single constraint assessing a winner-mark (Parse-Syl) is not dominated
by Closer (Ft-Bin). Hence Parse-Syl is demoted to a stratum immediately
below Ft-Bin. Again no new stratum needs to be created:

(33) State of the constraint hierarchy after step 4

H4 = {Ft-Bin, Leftmost}


{All-Ft-R, All-Ft-L, Rightmost, Parse-Syl}

In four steps, the algorithm has produced a broad division into two strata. The top
stratum contains both undominated constraints, while the bottom stratum contains
all dominated constraints. However, this hierarchy must be further rened in
order to achieve the status of target grammar. In particular, renements must
be achieved in the mutual rankings of the dominated constraints in the bottom
stratum.
The algorithm does not know that it has not yet nished its reranking opera-
tion, and it steadily goes on considering mark-data pairs. Let us assume that it
works through the pairs once more in the same order. The rst pair b  a (which
we will not fully repeat) contains a loser-mark {*All-Ft-L} and a winner-mark
{*All-Ft-R}:

The mark *All-Ft-L is the highest-ranked in loser-marks in the current hier-

archy, H4. (This is trivial, since it is the only mark in loser-marks.)
The constraint assessing the winner-mark (All-Ft-R) is not dominated by
Closer (All-Ft-L). (Both are in the same bottom stratum of H4.) Therefore
All-Ft-R is demoted to a position immediately below All-Ft-L, creating a
new stratum in the hierarchy:

314
 7.3 Learning the Pintupi grammar of stress

(34) State of the constraint hierarchy after step 5

H5 = {Ft-Bin, Leftmost}


{All-Ft-L, Rightmost, Parse-Syl}


{All-Ft-R}

Next the algorithm reconsiders the second mark-data pair in (17), c  a, which
contains the loser-mark {*Parse-Syl} and the winner-marks {*All-Ft-L, *All-
Ft-R}:

The mark *Parse-Syl is the highest-ranked in loser-marks in H5.

One constraint in winner-marks (All-Ft-L) is not already dominated by Closer
(Parse-Syl). Only All-Ft-L is demoted to the stratum immediately below
Closer. It comes to share this stratum with All-Ft-R, which already occupied it:

(35) State of the constraint hierarchy after step 6

H6 = {Ft-Bin, Leftmost}


{Rightmost, Parse-Syl}


{All-Ft-R, All-Ft-L}

In two more steps following the rst round of four, Constraint Demotion has
created the stratum {Rightmost, Parse-Syl} below the topmost stratum, which
is in fact part of the target grammar. The only step to take place is a demotion of
All-Ft-R below All-Ft-L, which requires a third (and nal) round through the
mark-data pairs (it will eventually be triggered by the rst pair b  a). But rst
we will see two vacuous steps.
Next, Constraint Demotion reconsiders the pair d  a, which contains loser-
marks {*All-Ft-L, *Leftmost, *Parse-Syl} and winner-marks {*All-Ft-R,
*Rightmost}:

*Leftmost is the highest-ranked loser-mark in H6,

while both constraints in winner-marks (All-Ft-R, Rightmost) are already
dominated by Closer (Leftmost). No demotion takes place, and the hierarchy
remains as it is:

(36) State of the constraint hierarchy after step 7

H7 = H6

315
Learning OT grammars

A similar story holds for the pair e  a, which contains the loser-marks {*All-
Ft-L, *Ft-Bin}, and the winner-marks {*Parse-Syl}:

*Ft-Bin is the highest-ranked loser-mark in H7 (= H6),

while the constraint in winner-marks (Parse-Syl) is already dominated by
Closer (Ft-Bin). This entails that no demotion takes place.

(37) State of the constraint hierarchy after step 8

H8 = H7 = H6

What both previous steps have in common is that in both cases, loser-marks con-
tain a mark which is assessed by an undominated constraint. In the pair d  a
this is *Leftmost, and in the pair e  a, this is *Ft-Bin. These marks are
identied as Closer, regardless of the content of winner-marks in these pairs. Con-
straint Demotion says that a constraint can be demoted only if it is not already
dominated by Closer. But in fact, no winner-marks in these pairs are assessed by
constraints which are not already dominated by the top stratum of constraints.

Finally, a third round is needed to nish the hierarchy. Once again, the initial pair
b  a is considered. It contains a loser-mark {*All-Ft-L} and a winner-mark
{*All-Ft-R}:

The mark *All-Ft-L is the highest-ranked in loser-marks in the current hier-

archy, H8.
The constraint assessing the winner-mark (All-Ft-R) is not dominated by
Closer (All-Ft-L). (Both are in the same bottom stratum of H8.) Therefore
All-Ft-R is demoted to a position immediately below All-Ft-L, creating a
new stratum in the hierarchy:

(38) State of the constraint hierarchy after step 9

H9 = {Ft-Bin, Leftmost}


{Rightmost, Parse-Syl}


{All-Ft-L}


{All-Ft-R}

This hierarchy precisely matches the target grammar (6). The learning algorithm
has been successful, having accomplished what it should. Interestingly, this hier-
archy cannot be changed by any further rounds of reconsiderations of the mark-
data pairs in table (17). It is a stable ranking, crucially so. Let us see how this
316
 7.3 Learning the Pintupi grammar of stress

conclusion follows. We have already seen that the pairs d  a and e  a cannot
lead to further rerankings of the hierarchy. Their loser-marks contain a mark
which is assessed by an undominated constraint, either Ft-Bin or Leftmost. But
in the nal hierarchy, H9, no constraint occurs which is not already dominated by
either of these constraints. Therefore no constraint can undergo a further demo-
tion which is triggered by either Ft-Bin or Leftmost.
Similar reasoning holds for the two remaining pairs b  a and c  a. Evidently
the former cannot lead to further demotion, since it was itself the trigger of the
nal demotion in step 9, which led to the current hierarchy. The latter pair c  a
contains the loser-mark {*Parse-Syl} and the winner-marks {*All-Ft-L, *All-
Ft-R}. Again the constraint in loser-mark is already dominating both of the
constraints in winner-marks. Therefore this is insufcient to trigger a demotion.
This result is important, as it shows that the learning algorithm, on the basis of
a single output form, converges to a state in which all information from this
output form has been put to its maximal use. All information has been extracted,
and no more can be extracted from the output form under consideration. Regard-
less of how many times the mark-data pairs are reconsidered by the algorithm, no
further changes in the hierarchy will result. The learner is eager, but she knows
where to stop.6
It may occur (and typically does occur) that a single output form is insufcient
to rank completely all constraints. This we have in fact just seen in the Pintupi
ve-syllable example, from which no mutual rankings of constraints within strata
{Ft-Bin, Leftmost} and {Rightmost, Parse-Syl} could be deduced. In fact,
the language may offer no evidence for a complete ranking. But it is equally
possible that when new output forms are fed into the algorithm, new rerankings
will be motivated. This issue will be taken up in sections 7.4 7.5.

7.3.5.2 Robustness of Recursive Ranking

We must now consider what happens when Recursive Ranking processes the
mark-data pairs in a different order from the one we assumed in the previous
section. Will the algorithm still converge be able to arrive at the target

6
Bruce Tesar ( p.c.) points out that there is a theoretical limit (an upper bound) on the number
of informative mark-data pairs necessary to determine a correct constraint hierarchy for a
language. An informative mark-data pair is one which contains some information about the
ranking that is not contained in the pairs already observed. The upper bound is on the order
of the square of the number of constraints. More precisely, if N is the number of constraints,
the upper bound on the necessary number of informative mark-data pairs is N(N-1)/2. This is
just a theoretical limit; in practice, the actual necessary number is much, much lower than this,
and in fact it would be quite challenging to construct an optimality theoretic system that
actually could require that many pairs.

317
Learning OT grammars

hierarchy? A positive answer to this question is crucial, since no inherent property

of the mark-data pairs would force one order of processing over another. If the
learner can get stuck just by considering the data in a clumsy order, something
is wrong. In that case the learner will arrive at the target grammar only by chance:
by coincidentally picking a smart order. In that case, we can no longer assume
that OT grammars are learnable.
But fortunately, the attainment of the target grammar is independent of the
order in which mark-data pairs are considered by Constraint Demotion. The only
side-effect of selecting one order over another is the number of intermediate
stages that must be passed on the way towards the goal. There are efcient and
less efcient orders, so to speak. We will not consider a formal proof of this
conclusion, but as a way of illustrating the relative length of the trajectory, we
will consider another order of processing mark-data pairs.
In fact, the order of mark-data pairs which we considered in the previous
section was the least efcient of all possible orders. The quickest way to reach the
goal considers the mark-data pairs in precisely the reverse order:

(39) Mark-data pairs considered in reverse order of previous section

subopt  opt loser-marks winner-marks

ea {*All-Ft-L, *Ft-Bin} {*Parse-Syl}

da {*All-Ft-L, *Leftmost, {*All-Ft-R,
*Parse-Syl} *Rightmost}
ca {*Parse-Syl} {*All-Ft-L,
*All-Ft-R}
ba {*All-Ft-L} {*All-Ft-R}

Let us go through the essential steps once again, now following the new order.
Again, the initial hierarchy has a single stratum only, containing all constraints.
First, consideration of mark-data pair e  a demotes Parse-Syl to a new
stratum. Second, consideration of mark-data pair d  a demotes All-Ft-R and
Rightmost into the stratum already occupied by Parse-Syl. No new stratum is
created because demotion is minimal (both of the highest-ranking constraints
assessing loser-marks, All-Ft-L and Leftmost, already dominate this stratum).
Third, the mark-data pair c  a produces a demotion of All-Ft-L and All-Ft-
R into a new stratum, immediately below that of Parse-Syl. Fourth, the mark-
data pair b  a leads to a demotion of All-Ft-R to a new stratum, below that of
All-Ft-L. And that is all! In (40) are the hierarchies which arise after each of
these four steps.
318
 7.3 Learning the Pintupi grammar of stress

(40) Demotion steps when considering mark-data pairs in reverse order

Steps Hierarchies after steps 14

H0 {All-Ft-L, All-Ft-R, Ft-Bin, L-most,

Parse-Syl, R-most}
H1 e  a {All-Ft-L, All-Ft-R, Ft-Bin, L-most, R-most}

{Parse-Syl}
H2 d  a {All-Ft-L, Ft-Bin, L-most}

{Parse-Syl, All-Ft-R, R-most}
H3 c  a {Ft-Bin, L-most}

{Parse-Syl, R-most}

{All-Ft-L, All-Ft-R}
H4 b  a {Ft-Bin, L-most}

{Parse-Syl, R-most}

{All-Ft-L}

{All-Ft-R}

Here it takes four steps, rather than nine, to arrive at the target grammar. We nd
a rather large variation in length of the trajectory between the initial state and the
target grammar, depending on the order in which mark-data pairs are considered.
But this variation does not affect the outcome: in both cases the algorithm con-
verges into the target grammar.7

7.3.5.3 Why demotion must be minimal

Another question to be addressed is that of why demotion is minimal, rather than
maximal. Recall that by minimal demotion we understand that a constraint assess-
ing winner-marks is demoted to a stratum directly below the highest-ranking

7
Yet another issue is that of the generation of informative suboptimal candidates. As mentioned
in note 2, an algorithm was developed by Tesar (1995) to accomplish this task. Error-Driven
Constraint Demotion works by interweaving learning and parsing. The learner uses her current
constraint hierarchy, and computes the description for the underlying form which is best. If it
matches the correct description (the winner), then learning is done for that winner. If it does
not match the winner, then it will make a good loser, so a mark-data pair is made combining
the winner ( pre-determined) with the currently optimal description as the loser. Constraint
demotion can then be applied with that mark-data pair, yielding a new constraint hierarchy.
The learner then applies parsing using the new constraint hierarchy, to see if the winner is
optimal yet. If not, then the constraint which is best under the new constraint hierarchy
becomes the next loser. In this way, learning is stopped for a given winner once a constraint
hierarchy is reached which makes that winner actually optimal.

319
Learning OT grammars

constraint assessing loser-marks. We have seen the success of the minimal demo-
tion strategy in the Pintupi case study. But it is useful to understand that minimal
demotion is crucial to the success of the algorithm. To understand this, we must
consider the alternative strategy of maximal demotion.
By maximal demotion we might understand the following: a constraint assess-
ing winner-marks is demoted to a stratum below the lowest-ranking constraint that
assesses loser-marks. The failure of maximal demotion can be demonstrated in
either of two ways. First, it can be shown that the algorithm continues performing
constraint rerankings even when the target grammar has been reached. Of course
the algorithm does not know that it has reached the target grammar. Therefore
it will continue considering mark-data pairs, and performing rerankings whenever
the mark-data pairs motivate these. It is only when rerankings are no longer justi-
ed that we say that the target grammar has been acquired. Now suppose that
the algorithm would, by chance, arrive at the target grammar. Another demotion
would occur if another mark-data pair (say, e  a) were taken into consideration:

(41) Target grammar ea Non-target grammar

{Ft-Bin, L-most}
{Ft-Bin, L-most}

{Parse-Syl, R-most}
{R-most}

{All-Ft-L}
{All-Ft-L}

{All-Ft-R} {Parse-Syl, All-Ft-R}

Parse-Syl is demoted below the lowest-ranking constraint assessing loser-marks,

which is All-Ft-L in this case. This means that an algorithm using maximal
demotion is unable to converge on the target grammar, since it continues chang-
ing it once it has reached it.
But in fact, there is even no guarantee that such an algorithm will ever arrive
at the target grammar. This is the second argument to the effect that maximal
demotion is a non-effective strategy. Table (42) compares sequences of states
through which minimal and maximal demotion go when using the order of mark-
data pairs in (39).

(42) Non-convergence of algorithm using maximal demotion

Steps Minimal demotion Maximal demotion

H0 {All-Ft-L, All-Ft-R, Ft-Bin, (identical to minimal demotion)

L-most, Parse-Syl, R-most}

H1 e  a {All-Ft-L, All-Ft-R, Ft-Bin, (identical to minimal demotion)

L-most, R-most}

{Parse-Syl}

320
 7.4 The learning algorithm: discussion

Steps Minimal demotion Maximal demotion

H2 d  a {All-Ft-L, Ft-Bin, L-most}
{All-Ft-L, Ft-Bin, L-most}

{Parse-Syl, All-Ft-R, R-most} {Parse-Syl}

{All-Ft-R, R-most}

H3 c  a {Ft-Bin, L-most}
{Ft-Bin, L-most}

{Parse-Syl, R-most}
{Parse-Syl}

{All-Ft-L, All-Ft-R} {All-Ft-L, All-Ft-R, R-most}

H4 b  a {Ft-Bin, L-most}
{Ft-Bin, L-most}

{Parse-Syl, R-most}
{Parse-Syl}

{All-Ft-L}
{All-Ft-L, R-most}

{All-Ft-R} {All-Ft-R}

H5 e  a {Ft-Bin, L-most}

(no demotion justied ) {All-Ft-L, R-most}

{Parse-Syl, All-Ft-R}

H6 d  a {Ft-Bin, L-most}

(no demotion justied ) {All-Ft-L}

{Parse-Syl}
{All-Ft-R, R-most}

H7 c  a (no demotion justied ) (=H3)

H8 b  a (no demotion justied ) (=H4)

H9 e  a (no demotion justied ) (=H5)

H10 d  a (no demotion justied ) (=H6)

Minimal demotion arrives at the target grammar in four steps. But maximal
demotion has not reached the target grammar even after ten steps, and it has
actually started reiterating previous stages from H7 onwards. It has run into an
eternal loop. Such an oscillation of demotions offers no hope that this algorithm
will ever acquire the target grammar. In sum, we have found that constraint
demotion must be minimal in order for effective learning to become possible at all.

7.4 The learning algorithm: discussion

Tesar and Smolenskys learning algorithm has been shown to be successful in the
complex task of deducing a hierarchy of constraints, that is, an OT grammar. Of
course we have only considered a very simple case, involving six constraints
whose interactions could be construed from the examination of a single example.
But in fact it can be shown that the algorithm will successfully construct a hierar-
chy of constraints of any complexity, as long as the input data are consistent.
321
Learning OT grammars

Therefore, if the model can learn the constraint hierarchy of an OT grammar, then
it means that the problem is logically solvable. Consequently we may assume that
the language-learning child is also able to perform the same task. The learnability
of constraint rankings is a crucial ingredient of the explanation of learnability of
grammars. It is only an ingredient, however, since much more is at stake in
learning a grammar than constraint ranking. To this issue we now turn.

Tesar and Smolensky make a number of simplifying assumptions about the nature
of the acquisition task, which will (ultimately) have to be reconsidered in order
to arrive at a more complete model of the language learner. Let us discuss these
assumptions.
First, Tesar and Smolenskys algorithm works on the assumption that output
forms are available to the learner in the form of (properly structured) linguistic
representations, rather than in a raw phonetic form. Under real-life conditions
in which acquisition takes place, we cannot make this assumption, simply because
linguistic units such as syllables, feet, and distinctive features are not readily
available from the phonetic forms, and must therefore be inferred by the learner.
Underdetermination of the linguistic representation in the phonetic form raises
the problem of how the learner gets around it. This problem has been addressed
recently by Tesar (1996), who has developed an extension of the learning algo-
rithm having the ability to reach progress along two dimensions, by iterative
testing of hypotheses about the outputs representation and the constraint ranking.
Tesar (1996: 1, 7) states the general idea as follows:
the learner can go back and forth between estimating the hidden structure
and estimating the constraint ranking, eventually converging on the correct
forms for both . . . The goal is to nd a ranking such that the interpretation of
each overt form matches the corresponding optimal description, both of
which are determined with respect to the current hypothesis ranking. Chang-
ing the hypothesis ranking may change the optimal description assigned to
the underlying form of the prior interpretation. But, it may also change the
interpretation assigned to the overt form by interpretative parsing.
Tesar demonstrates the feasibility of this approach in an algorithm that is faced
with the task of constructing a ranking of metrical constraints on the basis of
outputs that have no overt metrical structure, but only raw stress markings.

Second, Tesar and Smolenskys learning algorithm assumes that the learner has
already mastered the lexicon, including the correct underlying representations for
each morpheme. Therefore every output form considered by the algorithm will be
paired with its correct input form. But in real-life language acquisition, under-
lying forms are hypothetical and have to be inferred from combined analytic
assumptions about the output and the constraint hierarchy. Of course this problem
322
 7.4 The learning algorithm: discussion

is not unique to OT in fact every theory of phonology which assumes that

contextual alternants of a morpheme derive from a single underlying form is
faced with it. Note that with the addition of underlying forms to the logical
problem of acquisition, in fact three variables must be continuously monitored:
(a) the underlying form, (b) its output representation, and (c) the constraint hier-
archy relating both levels of representation:

(43) Interdependence of three factors in language acquisition

Output representations

Constraint hierarchy Underlying representations

Perhaps a solution to this problem will be possible along the lines of the extended
model of Tesar (1996) sketched above, and we will make an attempt at this in
section 7.5. This model will have to be extended so as to allow the iterative testing
of hypotheses for all three variables.

Third, Tesar and Smolensky assume that the hierarchy of the target language has
the property of total ranking. The assumption that all constraints are ranked with
respect to the others guarantees that each input corresponds with a single output.
To see this, consider the fact that no two output forms that have distinct violation
marks can both be optimal. Hence if two output forms have different violation
marks for conicting constraints, then these forms must have different harmonic
values, since both constraints are, by denition, ranked with respect to each other.
In close relation to this issue, Prince and Smolensky (1993: 51) remark that the
total ranking hypothesis is no a priori assumption of OT, but an empirical matter:
It is entirely conceivable that the grammar should recognize nonranking of
pairs of constraints, but this opens up the possibility of crucial nonranking
(neither can dominate the other; both rankings are allowed), for which we
have not yet found evidence.

There is potential empirical evidence that bears on this issue. The assumption that
an input has unique output may be problematic in the sense that languages are
known to exhibit widely variability of output forms, even under grammatical con-
trol. Some phenomena to display grammar-governed variability in many languages
are stress patterns, phonological phrase structure, and syntactic phenomena such
as linear order of adjunct phrases (scrambling). See chapter 9 (section 9.4) for
more discussion of variation in OT.

Finally, it is assumed that among the data presented to the learner no incorrect
data (ungrammatical outputs) occur. Again this assumption does not match real-
life conditions under which language acquisition takes place. Tesar and Smolensky
323
Learning OT grammars

show that the learning algorithm is capable of signalling such mists, and make
suggestions about the possible role of this alarm function in acquisition.

7.5 Learning alternations and input representations

7.5.1 The nature of the problem

The major question arises whether the algorithm can be extended such that it will
be able to handle alternations in the shape of morphemes. The key aspect of
alternations is that two or more output forms share a single input form the
lexical or underlying representation. Usually the input form is identical to one of
the surface alternants, but there is no label on the data to tell the learner which
one this is. The learners task is therefore more complex than was assumed in the
previous section, where the input was an unmetried word. We will consider the
question of whether we can generalize the iterative learning strategy to the learn-
ing of underlying forms.
The alternation problem will be tackled using a relatively simple set of data
from Dutch, involving syllable-nal devoicing. As we learned in chapter 1, this
language has a voicing contrast in obstruents which is neutralized in syllable-
nal position, where all obstruents are voiceless. No alternations of voicing in
obstruents are apparent in stems such as pet, voet, and bloot (44), in which
underlying forms contain voiceless obstruents:

(44) a.i pet /pt/ [pt] cap

a.ii petten /pt-n/ [p.tn] caps
b.i voet /vut/ [vut] foot
b.ii voeten /vut-n/ [vu.tn] feet
c.i bloot /blot/ [blot] naked (predicative)
c.ii blote /blot-/ [blo.t] naked (attributive)

But in stems whose nal obstruent is underlyingly voiced, coda devoicing pro-
duces voice alternations, as in the singular and plural of bed, hoed, and breed
(45):

(45) a.i bed /bd/ [bt] bed

a.ii bedden /bd-n/ [b.dn] beds
b.i hoed /hud/ [hut] hat
b.ii hoeden /hud-n/ [hu.dn] hats
c.i breed /bred/ [bret] wide (predicative)
c.ii brede /bred-/ [bre.d] wide (attributive)

The learning algorithm must nd the constraint ranking (involving an interaction

of well-formedness and faithfulness constraints) on the basis of assumptions
324
 7.5 Learning alternations and input representations

about the underlying forms (deducing those of 445), and a set of universal con-
straints. In this set of universal constraints are included the two well-formedness
constraints (46a b) and the faithfulness constraint (46c):

(46) a. *Voiced-Coda
Coda obstruents are voiceless.
b. Inter-V-Voice
Intervocalic consonants are voiced.
c. Ident-IO(voice)
Let be a segment in the input, and be a correspondent of in
the output. If is [voice], then is [voice].

Devoicing of coda consonants and intervocalic voicing of consonants are both

processes which are frequently attested in the languages of the world. Also, both
are natural states of affairs from an articulatory point of view. Being markedness
statements, they are captured by the well-formedness constraints (46ab).
The grammar of Dutch ranks these constraints in the following way (the nota-
tion identies three strata in the hierarchy):

(47) A fragment of the grammar of Dutch

{*Voiced-Coda}


{Ident-IO(voice)}


{Inter-V-Voice}

The task facing the language learner is to nd the proper constraint hierarchy
(47), and also to infer the correct underlying forms which are mapped to the
attested outputs. A priori, the learner may state two hypotheses about the alterna-
tion [bt]~[b.dn] in (45), which involve different markedness constraints. The
rst (correct) hypothesis is that the stem bed has an underlying voiced obstruent
/bd/, which is devoiced at the end of the syllable in the singular. Syllable-nal
devoicing involves domination of the markedness constraint *Voiced-Coda over
the featural faithfulness constraint Ident-IO(voice), as we have seen in chap-
ter 1. The alternative hypothesis is that the stem bed has an underlying voiceless
obstruent /bt/, which is voiced in intervocalic position in the plural. Intervocalic
voicing involves domination of the markedness constraint Inter-V-Voice over
Ident-IO(voice). This hypothesis is demonstrably incorrect, as no constraint hier-
archy that involves this ranking is consistent with the data. But let us rst discuss
the mapping of the correct underlying forms to the attested outputs.
Under the correct hypothesis, the underlying forms /pt/ and /bd/ are mapped
onto surface forms by the ranked constraints of (47), as illustrated by the tableaux
325
Learning OT grammars

(489). The evidence for the ranking *Voiced-Coda
Ident-IO(voice) comes

from the singular form [bt] in tableau (49), where voiced input /d / corresponds
with voiceless /t / in the output.

(48) Input: /pt/ *Voiced-Coda Ident-IO(voice) Inter-V-Voice

a. + pt
b. pd *! *

(49) Input: / bd/ *Voiced-Coda Ident-IO(voice) Inter-V-Voice

a. + bt *
b. bd *!

None of these tableaux contain evidence for the ranking of Inter-V-Voice, since
all output candidates which are considered satisfy this constraint. (There is no
intervocalic context.)

For evidence for the ranking Ident-IO(voice)
Inter-V-Voice we must turn to

the tableau of the plural form [p.tn] caps, which is faithful to the underlying
voiceless /t/ in spite of the pressure to voice this in intervocalic position, due to
Inter-V-Voice:

(50) Input: /pt-n / *Voiced-Coda Ident-IO(voice) Inter-V-Voice

a. + p.tn *

b. p.dn *!

(51) Input: / bd-n / *Voiced-Coda Ident-IO(voice) Inter-V-Voice

a. b.tn *! *

b. + b.dn

Note that tableau (51) of [b.dn] beds gives no information about ranking
whatsoever.

Now consider the (incorrect) alternative hypothesis, under which the voiced [d]
in [b.dn] derives from an input /t/ in /bt-n/ by intervocalic voicing. The
grammar which performs this mapping must crucially rank Inter-V-Voice
326
 7.5 Learning alternations and input representations

above Ident-IO(voice), since intervocalic voicing violates featural faithfulness

to [voice]. Under this hypothesis there is no reason to demote *Voiced-Coda
from its undominated position because it is unviolated in the examples at hand
(actually, as it is in all output forms of the language). Therefore we consider the
following hypothetical ranking:

(52) Hypothetical (and incorrect) fragment of the grammar of Dutch

*Voiced-Coda, Inter-V-Voice
Ident-IO(voice)

This ranking actually derives all the forms correctly except one: it predicts
*[p.dn] caps rather than [p.tn]. (The incorrect winner is indicated by a L
symbol.)

(53) Input: /pt-n/ *Voiced-Coda Inter-V-Voice Ident-IO(voice)

a. p.tn *!

b. L p.dn *

The intervocalic voicing contrast has vanished due to the demotion of Ident-
IO(voice).

The argument for the correct grammar (47) over the hypothetical grammar in (52)
is based on the voicing alternation in bedbedden, and lack thereof in petpetten.
This makes a perfectly straightforward argument. However, we must ask whether
the choice in favour of grammar (47) is equally transparent to the learner, who
must determine both the underlying forms and the constraint rankings, on the
basis of only surface forms and a set of universal constraints. What makes a child
learning Dutch eventually select this correct combination of grammar and lex-
icon, rather than one in which the alternation in bedbedden is due to intervocalic
voicing of an underlying /t/? Actually, we will show that the correct grammar and
underlying forms are learnable by the Tesar and Smolensky learning algorithm, if
only this algorithm is extended by iterative learning in the sense of Tesar (1996).
That is, the learner can go back and forth between estimating the underlying
forms and estimating the constraint ranking, and eventually converges on the
correct forms for both. We will argue that Lexicon Optimization plays an import-
ant role in this learning procedure.

7.5.2 Learning the hierarchy on the basis of correct inputs

We will rst prove that the correct constraint ranking is learnable simply on the
basis of the correct underlying forms, by the standard Tesar and Smolensky algo-
rithm. That is, we temporarily put aside the problem of how to determine the
327
Learning OT grammars

underlying forms, and return to this in due course. The initial state of the learner
is the universal set of unranked constraints contained in a single stratum, of which
(54) forms a subset:

(54) { . . . *Voiced-Coda, Ident-IO(voice), Inter-V-Voice . . . }

Assume that the learner is rst confronted with the surface form [bt], equipped
with the correct underlying form / bd/. The mark-data pairs of the optimal can-
didate [bt] and its major competitor [bd] are given below:

(55) Mark-data pairs

subopt  opt loser-marks winner-marks

bd  bt {**Voiced-Coda} {*Ident-IO(voice)}

No cancellation is required. Form here on Constraint Demotion applies as fol-

lows. First it is made sure that the constraint which induces the loser-mark,
*Voiced-Coda, does not dominate the constraint which induces the winner-mark,
Ident-IO(voice). This is trivial, since the hierarchy is still in its initial unranked
state. Next the constraint which induces the winner-mark is demoted to a stratum
immediately below that of the constraint which induces the loser-mark:

(56) { . . . *Voiced-Coda, Inter-V-Voice . . . }
{Ident-IO(voice)}

Observe that this happens to be the incorrect hypothetical grammar of (52). What
makes the child reject this? It will be clear that no valuable information can be
gathered from the surface forms [pt] and [b.dn], since these violate none of
the three constraints in (56). (See tableaux 48 and 51.) And since there is no
interaction of constraints in these cases, no new rankings can be established.
But sooner or later, the Dutch child comes across the crucial surface form
[p.tn], which will induce a second round of the constraint demotion algorithm.

(57) Mark-data pairs

subopt  opt loser-marks winner-marks

p.dn  p.tn {*Ident-IO(voice)} {*Inter-V-Voice}

Again, no cancellation of marks occurs. It is made sure that the constraint which
induces the loser-mark, Ident-IO(voice), does not dominate the constraint induc-
ing the winner-mark, Inter-V-Voice. In accordance with the algorithm the latter
constraint is demoted to a stratum immediately below that of the former con-
straint:

(58) { . . . *Voiced-Coda . . . }
{Ident-IO(voice)}
{Inter-V-Voice}

328
 7.5 Learning alternations and input representations

which is the correct mini-grammar of the voicing alternations in (44 5). We leave
it as an exercise to the reader to establish that the learning algorithm would have
yielded the same result when surface forms were presented in the opposite order,
with [p.tn] preceding [bd].

7.5.3 Learning the hierarchy without given inputs

Next we provide the child with a severe handicap, which is nevertheless universal
among language learners: no underlying forms are innate, and hence the learner
has to infer them solely on the basis of surface forms. What we do assume is that
the learner is able to do sufcient morphological analysis on surface forms, which
results in a segmentation of the stem and plural sufx /-n /. (Morphological
analysis is a matter of hypothesis testing itself. In a real-life, full-blown model it
will have to be integrated with the rest of the acquisition, but we will leave this
aspect out of consideration here.)
The same child who has successfully ranked the three constraints in (58) must
now forget all the results of this laborious task, as well as the underlying forms,
and is reinstalled with a constraint ranking in its initial maiden state. The question
is: will this naturally handicapped child still be able to deduce the correct
constraint ranking, as she achieved earlier while knowing all of the underlying
forms? Moreover, will this child be able to deduce the correct underlying forms?
In what follows we will see that the answers to these questions are positive, if we
assume that the child has the innate disposition to go back and forth between
hypotheses about constraint rankings and underlying forms.
A trap which the child must avoid is setting up an analysis of bed as /bt/,
where the surface form [bt] is an exact copy of the input. In connection, a second
trap will have to be avoided, which is to analyse the form [b.dn] as the result
of intervocalic voicing of underlying /t/. When the child falls into such traps, will
she ever be able to proceed on the way to the constraint ranking (47)? Can the
learning algorithm get stuck after making an incorrect assumption about under-
lying forms? Surprisingly, we will see that the child is a robust learner. Falling
into the rst trap setting up an underlying form /bt/ is not fatal, since this
analysis has no consequences for constraint ranking whatsoever. Perhaps even
more surprisingly, setting up an underlying form /bt-n / is avoided by the
learner, since an alternative underlying form /bd-n/ is available which will
always be preferred, due to Lexicon Optimization.
We will assume that the language learner may cancel out earlier hypotheses
about underlying forms, but not hypotheses about ranking. That is, the correct
ranking will be established incrementally, as in the Tesar and Smolensky model.
The starting point is a completely at initial state in which constraints are all part
of a single stratum. Into this initial state more and more hierarchical structure is
329
Learning OT grammars

introduced, in a succession of stages of stratied hierarchies of increasing com-

plexity. This process terminates in a state in which no further rankings can be
established on the basis of any new evidence, which is the target ranking.8 The
property of the learning algorithm that constraint rankings established on the
basis of positive evidence need not be undone qualies it as sufciently robust
for the requirements of real-life language learning.

7.5.3.1 Setting the trap

Let us assume that the Dutch child is confronted with the worst-case scenario, in
which both of the poisonous forms [bt] and [b.dn], by accident, are pre-
sented to her prior to the forms [pt] and [p.tn]. That is, the child will start
analysing the voicing alternation on the basis of forms which are fully consistent
with both hypotheses, and before having seen the crucial form [p.tn]. Still
worse, we will assume that [bt] is presented before [b.dn] (this is certainly a
realistic scenario, since the singular is likely to be more frequently used than the
plural). That is, the trap is set: the child will be lured into setting up an analysis
of bed as / bt /, in which the surface form [ bt] is an exact copy of the assumed
underlying form. Since this analysis involves no violation of any of the three
constraints, it is selected by Lexicon Optimization.
The mark-data pair in (59) shows that no information about constraint ranking
can be extracted, since no constraints interact.

(59) Mark-data pairs

input subopt  opt loser-marks winner-marks

/ bt / bd  bt {* *Voiced-Coda,
*Ident-IO(voice)}

Actually this mark-data pair is never considered by the core algorithm since it is
eliminated by provision (16b) of Marks Cancellation even before Constraint
Demotion applies. The constraint ranking therefore remains in its initial state.
Let us assume that the learner is now confronted with the plural form [b.dn],
and (after judicious morphological analysis) attributes this form to the morpheme
/ bt/ which occurs in the singular [bt]. Starting from this incorrect lexical
form /bt /, the following mark-data pairs will be constructed:

8
It is an entirely different question whether the child actually proceeds incrementally on the
way to the target constraint ranking. Evidence presently available strongly suggests that chil-
drens grammars may have different constraint rankings from adult grammars. See the next
section.

330
 7.5 Learning alternations and input representations

(60) Mark-data pairs

input subopt  opt loser-marks winner-marks

/ bt-n / b.tn  b.dn {*Inter-V- {*Ident-IO(voice)}

Voice}
The winner-mark is assessed by Ident-IO(voice). Therefore Constraint Demotion
demotes this constraint directly below the highest-ranking constraint which
assesses the loser-mark, here Inter-V-Voice. The result is a new hierarchy:
(61) { . . . *Voiced-Coda, Inter-V-Voice . . . }


{Ident-IO(voice)}
In fact, this conclusion is correct, even though it is based on incorrect premisses.
That is, the voicing alternation in the pair [bt] ~ [b.dn] necessarily involves
some violation of faithfulness in either the singular or the plural form. The algo-
rithm incorrectly assumes that this violation occurs in the plural, rather than the
singular; but from this it draws a correct conclusion about the hierarchy. The
hierarchy (61) encodes precisely this observation: that faithfulness for [voice] is
violated. (The learner would have arrived at the same conclusion when she had
initially embarked on the correct hypothesis about the input form, / bd /.)

Next, the child encounters the singular form [pt], which she considers to be on
a par with the singular [bt], currently attributed to the lexical form / bt /. The
child simply registers [pt] under the lexical form /pt/,
(62) IO-pairings after learners exposure to output set {[bt], [b.dn],
[pt]}
a. /pt / [pt]
b. / bt / [bt] ~ /bt-n/ [ b.dn]
At this stage, the learner is completely unaware of its plural [p.tn], which will
induce an important constraint demotion in a later stage of learning. In fact, if the
child were asked to produce the plural of [pt] in this acquisitional stage, she
would have no choice but to reply: [p.dn]!, since this is literally what the
ranking (61) predicts that it should be. The model predicts overgeneralization of
phonology in alternations, which is a widely observed characteristic of childrens
grammars. See section 7.5.4 for Dutch language acquisition data which corrobor-
ate this prediction.

The nal form which the child is confronted with is [p.tn]. This comes as a
shock to the child, as this is the rst surface form she encounters in which
*Inter-V-Voice is violated. A mark-data pair is immediately drawn up:
331
Learning OT grammars

(63) Mark-data pairs

input subopt  opt loser-marks winner-marks

/pt-n / p.dn  p.tn {*Ident-IO(voice)} {*Inter-V-

Voice}

This presents ranking information which conicts with that of the form [b.dn]
beds in (60). However, since [p.tn] is the current form under analysis, the
learner proceeds to track mechanically its consequences for the hierarchy. Let us
carefully consider the steps taken by the Constraint Demotion algorithm.

First, it identies the mark *Ident-IO(voice) in loser-marks as the highest-

ranked in the current hierarchy. (This is trivial, since there is only one mark
in loser-marks.)
Second, it establishes that the constraint assessing the winner-mark (Inter-V-
Voice) is not dominated by the constraint assessing the loser-mark (Ident-
IO(voice)).
Third, it demotes Inter-V-Voice to a position immediately below Ident-
IO(voice).

This produces the correct target ranking in (64):

(64) { . . . *Voiced-Coda . . . }


{Ident-IO(voice)}


{Inter-V-Voice}

The child has actually succeeded in learning the correct target ranking. This
achievement is remarkable, given the fact that it is based on incorrect inferences
about underlying forms. But may we also conclude that the child has now suc-
cessfully acquired the fragment of the grammar which derives the voicing pattern?
Of course we may not, since the underlying forms of the pair [bt] ~ [bdn]
are still misrepresented as /bt / ~ / bt-n/. Once the ranking (64) has been
established, the child will make the incorrect prediction *[b.tn] for the plural
form see tableau (65):

(65) Input: / bt-n / *Voiced-Coda Ident-IO(voice) Inter-V-Voice

a. L b.tn *

b. b.dn *!

332
 7.5 Learning alternations and input representations

In fact, similar types of overgeneralizations are attested in childrens language,

which is to the credit of the model. But eventually they become extinct, suggest-
ing that underlying forms are subject to restructuring. How this restructuring
takes place is what we will consider in the remainder of this section.

7.5.3.2 Restructuring input forms

Once the conict between the acquired ranking (64) and the surface form [bdn]
has been signalled (the nature of this signal will be discussed below), the child
has two options. The rst is to adjust the hierarchy to the observed output [bdn]
while preserving its current input / bt/. The second option is to restructure the
input from / bt / into / bd/, so that the unchanged hierarchy correctly maps it
onto the observed output. We will see that the second strategy, restructuring the
input, is the only one leading to success.

The rst strategy, going through another constraint demotion, may in fact produce
a short-term success with respect to the problematic form, but will not converge
into a stable ranking. The mark-data pairs of table (60) reactivate Constraint
Demotion in the hierarchy (64). Concretely, Ident-IO(voice) will be demoted
below Inter-V-Voice:

(66) {*Voiced-Coda}
{ . . . }
{Inter-V-Voice}
{Ident-IO(voice)}

This hierarchy accounts for [bdn], but of course we run into a loop now, as we
have been in this stage before, at a point prior to the introduction of the form
[ptn]. The learner will go through the entire demotion process once more,
exchanging positions between Inter-V-Voice and Ident-IO(voice), without ever
converging into the target grammar. Note that these stages crucially include the
mark-data pairs (60) and the ranking (61). In sum, the rst option of repairing the
hierarchy by Constraint Demotion produces an eternal loop.
Of all things, eternal loops must be avoided most. Hence the learner must
capitalize on any information warning her that she is entering a loop. Actually, a
signal is given to the learner right at this point: a stratum is vacated. To vacate
a stratum means: to demote a constraint from a stratum in which it was the single
constraint. An empty stratum has been indicated in (66) by { . . . }.
Stated generally, a loop occurs whenever the learner is presented with incon-
sistent information with respect to the ranking of two constraints.9 That is, a pair
of constraints C1 and C2 must be ranked C1
C2 for one set of forms, and as C2

C1 for another set of forms. If the learner encounters both sets of forms, an
9
The same idea that vacating of a stratum points to an inconsistency in the underlying forms
was used by Broihier (1995) to solve the problem of learning grammars with tied constraints.

333
Learning OT grammars

eternal process of ranking and reranking starts that never terminates in a stable
hierarchy. The diagnostic for inconsistency in the data with respect to ranking is
the vacation of a stratum a consistent target ranking never gives rise to a loop.
Since the learner may assume that the target ranking is consistent with the data,
any sign of an eternal loop must be due to another cause an incorrect assump-
tion about input forms.
What causes an incorrect assumption about the input? Clearly the error must
reside in an alternating morpheme, whose alternants are specied as [+F] and
[F], respectively, in different contexts. By Lexicon Optimization, the learner
will be led to analyse the rst occurring alternant of a morpheme to reect its
input value. Hence, if the learner happens to encounter the alternant [F] rst,
she will set up the incorrect input value [F], even when the correct input value
is [+F].
This erroneous assumption on inputs causes the (apparent) inconsistency in
ranking, as follows. The value of the feature [F] in a non-alternating morpheme
always equals the value of [F] in one of the alternants of an alternating mor-
pheme. This is simply due to the neutralizing nature of the processes causing
alternations. (For example, word-nal [voice] occurs in both the alternating form
[bt] and in the non-alternating form [pt].) Alternating and non-alternating mor-
phemes have identical specications for [ F] in a specic context, as is shown
schematically below:

(67) Neutralization Elsewhere

context context
Alternating morpheme . . . [F] . . . . . . [+F] . . .
Non-alternating morpheme . . . [F] . . . . . . [F] . . .

The problem occurs when the learner attributes both types of morphemes to the
same input value, [F]. This causes an apparent inconsistency in the ranking of
faithfulness constraints and markedness constraints in the elsewhere context
(here: bedden versus petten). The loop signals that such a situation occurs.

Signalling an inconsistency in rankings is one thing, but to repair it is another. To

avoid eternal loops, we propose that the learning algorithm blocks Constraint
Demotion if this would result in an empty stratum. This is stated below as the
Anti-Loop Provision:

(68) Anti-Loop Provision

If Constraint Demotion produces an empty stratum:
a. cancel demotion
b. adjust the input form in accordance with Lexicon Optimization
c. adjust the mark-data pairs in accordance with the new input form
334
 7.5 Learning alternations and input representations

Once this alarm bell goes, the learner adjusts her underlying representation of the
observed output form in accordance with Lexicon Optimization. Restructuring of
the current input / bt-n/ into the new input /bd-n/ is forced as the single
method of reaching compatibility with the current hierarchy (64). This is achieved
by provision (68b):

(69) Adjustment of underlying form for [b.dn]:

/ bt-n / / bd-n/

Moreover, the mark-data pair for [bdn] is changed as a consequence of provi-

sion (68c). Since the input now equals the output, no violation of faithfulness
occurs any longer in the optimal form. Hence the mark-data pair becomes unin-
formative for Constraint Demotion:

(70) Adjusted mark-data pair for [bdn]

input subopt  opt loser-marks winner-marks

/ bd-n / b.tn  b.dn {*Inter-V-Voice,

*Ident-IO(voice)}

Only consideration of new forms might, potentially, lead to additional demo-

tions. Let us see what occurs when the algorithm reconsiders the singular form
[bt]. Uniqueness of underlying forms implies that the adjustment of the input
form of the plural / bt-n/ into /bd-n/ is accompanied by a change of the input
of the singular of / bt / into /bd/. At the same time, the mark-data pair for [bt]
is changed as a consequence of provision (68c):

(71) Adjusted mark-data pairs for [bt]

input subopt  opt loser-marks winner-marks

/bd / bd  bt {* *Voiced-Coda} {*Ident-IO(voice)}

At this point, when the learner is confronted with the output form [bt], the
hierarchy (64) still holds. The algorithm

First identies the mark * *Voiced-Coda in loser-marks as the highest-ranked

in the current hierarchy, (64).
Second, it establishes that the constraint assessing the winner-mark
(Ident-IO(voice)) is dominated by the constraint assessing the loser-mark
(*Voiced-Coda). Therefore no demotion is motivated.

Finally, it remains to be demonstrated that no adjustments of the current hierarchy

(64) can be triggered by non-alternating forms such as [pt]~[p.tn]. Their input
335
Learning OT grammars

forms are unchanged, since the Anti-Loop Provision is only activated by forms
which pose the threat of vacating a stratum. It is easy to show that such a threat
is not posed by [pt] or [p.tn], though. First, mark-data pairs for [pt] are
trivially uninformative. Second, mark-data pairs of [p.tn] in (63), if anything,
might motivate a demotion of Inter-V-Voice below Ident-IO(voice). However,
this demotion is inapplicable, as Ident-IO(voice) already dominates Inter-V-
Voice in the current ranking (64). In sum, non-alternating forms are incapable of
triggering any demotion, hence the Anti-Loop Provision is never activated, auto-
matically preserving the inputs.
Therefore we may safely conclude that the learning algorithm has converged
into the target grammar of Dutch: the hierarchy (64). But even more importantly,
the learner has succeeded in establishing the correct set of underlying forms for
alternating morphemes. That is, the learner has overcome the handicap that under-
lying forms are not pre-given, but have to be inferred from output forms on the
basis of hypothetical constraint rankings.

It remains to be seen whether the Anti-Loop Provision is sufciently general to

deal with more complex alternations. Alternations involving multiple features
seem to require a more sophisticated version of input adjustment (provision 68b),
limiting adjustment to the feature [F] at stake, while preserving input specicat-
ions of other features. The feature [F] must be inferred from the constraints whose
ranking is at stake (here, constraints referring to [voice]). Moreover, the problem
of how to deal with dependencies between alternating features must be solved.
We will leave these issues open for future research.

7.5.4 Intervocalic voicing in child language

The prediction that Dutch children go through a stage of overgeneralization of
intervocalic voicing (discussed in section 7.5.3.1) is corroborated. Actual ex-
amples, as produced by Emma, a three-year old, are presented below:10

(72) Emmas voicing pattern (3;03;3)

Emma adult form
a.i krij[t] ~ krij[d]en krij[t]en chalks
a.ii boo[t] ~ bo[d]en bo[t]en boats
a.iii groo[t] ~ gro[d]e gro[t]e big(in.)
(varies with gro[t]e)

10
I wish to thank Martin Everaert for sharing these data with me.

336
 7.5 Learning alternations and input representations

b.i kaar[t] ~ kaar[d]en kaar[t]en cards

b.ii taar[t] ~ taar[d]en taar[t]en cakes
c. olifan[t] ~ olifan[t]en olifan[t]en elephants
d. kas[t] ~ kas[t]en kas[t]en cupboards

During this stage, Emma voices obstruents in intervocalic positions where adult
forms have voiceless ones (72a). The context of voicing is actually more general
than intervocalic. That is, Emma voices obstruents in between approximants and
vowels, as (72b) shows, but not after a nasal (72c) or an obstruent (72d).
Only two months later, the pattern has extended into prevocalic obstruents
following nasals (73c), generalizing the process to post-sonorant obstruents:

(73) Emmas voicing pattern (3;43;5)

Emma adult form
a.i pe[t] ~ pe[d]en pe[t]en caps
a.ii poo[t] ~ po[d]en po[t]en paws
(varies with po[t]en)
a.iii voe[t] ~ voe[d]en voe[t]en feet
b.i bul[t] ~ bul[d]en bul[t]en lumps
b.ii taar[t] ~ taar[d]ies taar[t]jes small cakes
c.i mun[t] ~ mun[d]en mun[t]en coins
c.ii achterkan[t] ~ achterkan[d]en achterkan[t]en back sides

In both acquisitional stages, Emma displays variation in intervocalic voicing,

showing that the ranking of Inter-V-Voice with respect to Ident-IO(voice) has
not yet stabilized. Note also that intervocalic obstruent clusters are never voiced
in any acquisitional stage.

7.5.5 Discussion
For expository reasons, the alternation problem discussed has been a simple one,
involving only a single alternation. Yet our success in solving this problem sug-
gests that the Tesar and Smolensky algorithm is extendable to handle more com-
plex alternations. The central idea that we have explored here is that of building
inertness of the hierarchy into the learning algorithm, prohibiting the demotion
of a constraint which forms a stratum on its own. This presents a warning bell
to the learner that current assumptions about input forms may be incorrect, as a
result of which lexical restructuring takes place. Such adjustment of input forms
by the Anti-Loop Provision is always concrete, in the sense that it copies pho-
nological structure of the output (here, a feature value for [voice]), in accordance
with Lexicon Optimization. In chapter 9 (section 9.6) we will re-evaluate the need
for underlying forms in OT, sketching an alternative in terms of allomorphs.
337
Learning OT grammars

SUGGESTIONS FOR FUR THER READING

Parameters
Dresher, Elan and Jonathan Kaye (1990) A computational learning model for
metrical phonology. Cognition 34. 13795.

Connectionist models
Daelemans, Walter, Steven Gillis, and Gert Durieux (1994) The acquisition of
stress: a data-oriented approach. Computational Linguistics 20. 42151.
Gupta, Prahlad and David Touretzky (1994) Connectionist models and linguistic
theory: investigations of stress systems in language. Cognitive Science 18.
150.

Optimality Theory (see also references to OT in computational phonology

in chapter 1)
Hale, Mark and Charles Reiss (1997) Grammar optimization: the simultaneous
acquisition of constraint ranking and a lexicon. Ms., Concordia University,
Montral. [ROA-231, http://ruccs.rutgers.edu/roa.html]
Pulleyblank, Douglas and William J. Turkel (1997) Gradient retreat. In Roca.
153 93.
Smolensky, Paul (1996) On the comprehension/production dilemma in child lan-
guage. Linguistic Inquiry 27. 720 31.
Tesar, Bruce (1996) An iterative strategy for learning metrical stress in Optimality
Theory. In E. Hughes, M. Hughes, and A. Greenhill (eds.), The proceedings
of the 21st annual Boston University Conference on Language Development,
November 1996. 61526. [ROA-177, http://ruccs.rutgers.edu /roa.html]
Tesar, Bruce and Paul Smolensky (1993) The learnability of Optimality Theory:
an algorithm and some basic complexity results. Technical Report CU-CS-
678 93, Computer Science Department, University of Colorado, Boulder.
[ROA-2, http://ruccs.rutgers.edu/roa.html]
Tesar, Bruce and Paul Smolensky (1998) Learnability in Optimality Theory. Lin-
guistic Inquiry 29. 22968.

Constraint ranking software

Hayes, Bruce (1996b) Constraint ranking software. Ms., University of California,
Los Angeles. [Manual of computer program ranking OT constraints, on line
available from http://www.humnet.ucla.edu/humnet/linguistics/people/hayes/
otsoft/otsoft.htm]
338
 7.5 Learning alternations and input representations
7.5 Exercises

Child phonology
Demuth, Katherine (1995) Markedness and the development of phonological
structure. In J. Beckman (ed.), Proceedings of the 25th North East Linguistic
Society, II: Papers from the workshops on language acquisitions and lan-
guage change. 13 25.
Gnanadesikan, Amalia (1995) Markedness and faithfulness constraints in
child phonology. Ms., University of Massachusetts, Amherst. [ROA-67,
http://ruccs.rutgers.edu/roa.html]
Goad, Heather (1997) Consonant harmony in child language. In S. J. Hannahs and
M. Young-Scholten (eds.), Focus on phonological acquisition. Amsterdam:
John Benjamins. 11342.

Syntactic acquisition
Broihier, Kevin (1995) Optimality theoretic rankings with tied constraints: Slavic
relatives, resumptive pronouns and learnability. Ms., Department of Brain
and Cognitive Sciences, MIT. [ROA-46, http://ruccs.rutgers.edu /roa.html]

EXERCISES

1 Why the learner must use stratied hierarchies

Here we will consider the question of why learning should use stratied hierar-
chies. Apart from the appeal of beginning with all of the constraints in a single
stratum, there is another reason. This is closely related to the argument for min-
imal demotion made in section 7.3.5.3. The purpose of this problem is to pinpoint
the argument for stratied hierarchies.11 We will assume a hypothetical language
which has the following target ranking of three constraints:

(i) Target ranking (goal): Top
Mid
Low

Suppose we have the example pairs in (ii), b  a and c  a.

(ii) subopt  opt loser-marks winner-marks

ba {*Mid} {*Low}

ca {*Top, *Low} {*Mid}

The goal is to nd a ranking of the three constraints such that the winner is more
harmonic than the loser, for both pairs. The question is, why not create a new
stratum containing the demoted constraint Mid by itself?
11
Thanks to Bruce Tesar (p.c.) for suggestions which have inspired this problem.

339
Learning OT grammars

a. First show that constraint demotion with stratied hierarchies pro-

duces the correct result from an initial state {Top, Mid, Low} with
all three constraints unranked. (Recall that demotion is minimal.)
Construct a table as in (iii) which keeps track of the learners
progress toward the target grammar.

(iii) Demotion steps using stratied hierarchies, from initial state {Low,
Mid, Top}:
Steps Hierarchies after steps 13
H0 {Low, Mid, Top}
H1 b  a
H2 c  a
H3 b  a

b. Now we change our assumptions, such that demotion can produce

stratied hierarchies. Suppose we apply demotion, but always create
a new stratum. The starting hierarchy is Low
Mid
Top, the
reverse of the target. Construct a new table which keeps track of the
learners progress. Does this strategy always converge on the correct
grammar?

2 Learning input forms under positional faithfulness

Read chapter 9, section 9.5 on positional faithfulness. Then propose an analysis
of Dutch devoicing using only the following three constraints:

VOP Obstruents are voiceless.

Ident-IO(voice) Output segments preserve values of [voice]
for input correspondents.
Ident-IO(Ons, voice) Output segments in onset position preserve
values of [voice] for input correspondents.

Apply the ranking algorithm, including the Anti-Loop Provision, to test the learn-
ability of underlying forms. (Make sure that you present the data to the algorithm
in an order that will trigger the incorrect initial input form / bet /, as in the text of
section 7.5.3.1.) Do you experience any new difculties in learning input forms?
If so, can you propose a remedy?

340
 8
Extensions to syntax

8.1 Introduction
In this chapter we will look into results of OT in the domain of syntax. It is
important to emphasize once again that OT is not a theory of phonology proper,
but rather a theory of grammar. Therefore the basic idea underlying OT, that of
hierarchically ranked constraints which are minimally violated, is, in principle,
equally well applicable to non-phonological phenomena. The present chapter should
be read as an attempt at such an extension of OT into syntax, an attempt mainly
based on a recent paper on English auxiliaries by Grimshaw (1997). We will
discuss the syntactic counterparts of constraint types which we have seen active
in previous chapters, and will elaborate the OT architecture of faithfulness and
well-formedness constraints. Besides dening constraints in the domain of syn-
tax, we will also address the nature of the input in syntax, and the denition of
the function of Gen. These are non-trivial matters, that merit discussion in their
own right. Finally we will develop the factorial typology of the core syntactic
constraints of Grimshaws paper.

8.2 OT and syntax

How could OT contribute to a better understanding of syntactic phenomena? Let
us rst point out some dimensions along which recent Minimalist syntactic
theory and OT-based syntax differ, as well as some aspects in which they are sur-
prisingly similar. (See Radford 1997 for an overview of minimalism in syntax.)
First consider the major notion of parameter. In the Principles-and-Parameters
Theory of Chomsky (1981b) and much work thereafter, it is assumed that gram-
matical principles are universal, in the sense of being inviolable. On this view,
differences between grammars stem from differences in parameter settings. Para-
meters constitute a subset of the universal and inviolable principles of the theory
of grammar, among which they have a special status, offering a (typically binary)
choice of options among which languages are free to choose. Parameters serve the
descriptive goal of allowing for cross-linguistic diversity (linguistic typology)

341
Extensions to syntax

without having to give up on universality.1 Of course, a similar typological goal

is on the agenda of OT, but it is elaborated quite differently. Like parametric
theory, OT adheres to the notion of a core of universal principles, but shapes these
as a set of universal and violable constraints. Typologies are obtained by reranking
the universal constraints in language-specic hierarchies, rather than relativizing
a subset of universal principles to languages, as in parametric theory. Superci-
ally, then, both theories reach the same goal by largely parallel means: languages
are free to escape from universal uniformity either by switching off parameters,
or by domination of violable constraints. Constraint reranking and parameter
setting may thus be thought of as two a priori equivalent ways of reaching lin-
guistic typologies. When a parametric theorist would assert that the absence of
some grammatical effect in a language L is due to the negative value of para-
meter P, then an OT-theorist might equally well argue that in this language L the
constraint P is crucially dominated by some other constraint, suppressing the
effects of constraint P. However, this is an oversimplied and misguided view of
the relation between parameters and OT, for the following reasons.
The paramount difference between a parametric approach and OT is that in OT,
constraints are not switched off, but only dominated by other constraints.
Therefore it is predicted that the effects of some constraint may show up even
in a language in which it is dominated. Given the chance, even a dominated
constraint will make its presence felt, and break into activity. The canonical
examples of such situations that we have seen in this book are cases of the
emergence of the unmarked. In contrast, a parameter, once it has been switched
off, can never thereafter leave its mark on the grammatical patterns of the lan-
guage. Consider a language which selects the value negative for the parameter
onsets are obligatory (as we may infer from the fact that onset-less syllables occur
in the surface patterns of this language). On a parametric view it is predicted that
such a language lacks processes avoiding hiatus, or other selectional means to
bring about syllables with onsets, rather than onset-less syllables. Yet we know by
now, after discussion in previous chapters, that this prediction is simply false. In
OT the (correct) prediction is made that the relevant constraint Onset may con-
tinue to be active in a phonology even when the language allows for onset-less
syllables. (Onset-less syllables show that Onset is dominated by faithfulness
constraints, obscuring its effects in most contexts.) There is a subtle yet robust
difference between parametric theory and OT in this respect. The question then
becomes: do similar effects arise in syntax as well? As we will nd out in this
chapter, there is indeed strong evidence for the activity of dominated constraints.

1
Or as Grimshaw (1997: 405) puts it, parameters serve to accommodate language variation in
a system of inviolable constraints.

342
 8.2 OT and syntax

A stronger kind of parallelism between standard syntax and OT, and one that will
also play an important role in the discussion below, is that of economy. This
notion stems from recent work in Minimalist syntactic theory (Chomsky 1993,
1995). At the heart of this theory is the assumption that syntactic derivations and
representations must not exceed what is minimally necessary to meet general
well-formedness conditions. For example, no movement must take place without
some general structural requirement forcing it,2 and no projection must be present
without an overt licensing of this amount of structure.
This notion of economy of derivation and representation is highly similar to the
do only when necessary principle of OT.3 That is, outputs will be identical to
inputs (due to faithfulness constraints) except when divergence between them is
forced by a high-ranking well-formedness constraint. But even then, the diver-
gence between input and output will be kept at a bare minimum, that is, to the
extent that is required to meet well-formedness constraints. Such inertness fol-
lows from interactions of faithfulness and well-formedness constraints, as we
have seen on many occasions in earlier chapters of this book, in which we were
dealing with phonological phenomena. In fact, inertness eventually follows from
the general OT idea that constraints are violable, but that violation must be
minimal. Minimalism in syntax is based on theoretical assumptions that are
sufciently parallel to OT notions to warrant an attempt at assimilation.4 The
advantage of modelling economy in OT is, of course, that it is a soft system
by nature, rather than a hard system such as Minimalist syntax. Recently syn-
tacticians have indeed tried to explain economy using OT assumptions, and we
will see some results of this ongoing enterprise below. We will base our discus-
sion on syntactic phenomena of Wh-movement and inversion of subject and aux-
iliary verb in English, as have been analysed by Grimshaw (1997). We will also
sketch some of the typological consequences of an OT analysis of Wh-movement.
From here on, this chapter is organized as follows. Directly below, in section
8.2.1, we will sketch the theoretical assumptions on which OT-based syntax is
founded, developing these as variants on more general assumptions underlying
OT. Here issues will be taken up such as: What denes syntactic inputs?, What
is the function of Gen in syntax?, and What is the inventory of syntactic con-
straints?. This amount of theoretical background will enable us to outline in
2
Compare the ban on unmotivated movement in pre-minimalist syntax.
3
This similarity was pointed out by Prince and Smolensky (1993). See chapter 3 (section 3.3.3),
for a discussion of do only when necessary in phonology.
4
One difference between Chomskys notion of economy and its OT counterpart of minimal
violation is the type of evaluation. Chomsky assumes that economy is evaluated locally, rather
than globally (as in parallel OT). What are compared are not complete derivations, but only the
possibilities arising at specic points in a derivation. Hence a choice made at a particular point
in a derivation cannot be inuenced by any negative consequences that it may have later on.

343
Extensions to syntax

section 8.3 an OT-analysis of Wh-movement and inversion of subject and auxil-

iary verb in English. Typological consequences of this analysis will be outlined
in section 8.4, on the basis of a factorial typology. Finally section 8.5 will present
conclusions and general perspectives.

8.2.1 OT-based syntax: general assumptions

Let us once more state the general assumptions of OT, which will help us to
conceive of their implementation in syntactic theory:

Universality, violability: Constraints are universal and can be violated, but

violation is minimal.
Strict domination: Constraints are hierarchically ranked in a grammar, such
that any higher-ranking constraint takes absolute priority over any lower-
ranking constraint.
Freedom of analysis: Any amount of structure may be posited on the input
by Gen, while keeping within the representational vocabulary of linguistic theory.
Harmony: Candidate outputs are ranked by the grammar according to their
relative harmony with the constraint hierarchy. The optimal output candidate
is the one that incurs the minimal violation of higher-ranking constraints, pos-
sibly at the expense of any number of violations of lower-ranking constraints.

Accordingly, to construct an OT framework for the grammatical domain of syn-

tax, we must at least answer the following questions:

What is the nature of the input?

What is the specic representational vocabulary of syntactic theory within
which Gen must keep?
What is the constraint inventory of syntactic theory?

8.2.2 Dening the input of syntax

Let us rst consider the nature of the input of a syntactic component of an OT
grammar. Grimshaws general idea is that syntactic inputs are dened in terms of
lexical heads and their argument structure, but are otherwise void of syntactic
structure. That is, no syntactic projections are given in the input the assignment
of such structures is the responsibility of Gen, to be discussed in section 8.2.3.

Accordingly, Grimshaw denes the input for the type of syntactic constructions
that she considers in her paper verbal extended projections as follows:

(1) Input
a lexical head plus its argument structure
an assignment of lexical heads to its arguments
344
 8.2 OT and syntax

a specication of the associated tense and semantically meaningful

auxiliaries

For example, consider the input of the sentence What did Mary say? At the level
of the input, this is dened by the lexical head say, which is a predicate taking
two arguments, plus an assignment of two lexical heads (Mary, what) to these
arguments.

(2) say (x, y)

x = Mary
y = what
tense = past

Note that no semantically empty auxiliaries (do, did) are present in the input. The
presence of such elements in the output, and the constraint interactions by which
they appear, will be our main concern in this chapter.
Grimshaw makes two further assumptions about the input that are implicitly
guided by the principle of Containment (Prince and Smolensky 1993),5 which was
briey discussed in chapter 3 (section 3.3.1). This principle requires that no
element be literally removed from the input. This may be stated for the domain
of syntax as follows:

(3) Containment
competing candidates are evaluated as analyses of the same lexical
material
competing candidates to be generated for a single input must be
semantically equivalent

To state it differently, each analysis of the input competes with other analyses of
the same input, and all these analyses must have non-distinct semantic representa-
tions. In fact, we are already committing ourselves to a denition of Gen in syntax,
an issue which will be given special attention in the next subsection.

8.2.3 Dening Gen for syntax

We have already encountered some limitations on candidate analyses generated
by Gen in the form of the Containment-based principles (3). In this subsection
we will consider the residue of structural restrictions on Gen as it is dened for
syntactic purposes, and also spell out the types of syntactic operations that are
allowed to Gen.

5
Although Grimshaw does not explicitly mention Containment, her approach is very much
similar to that of Prince and Smolensky (1993) in this respect.

345
Extensions to syntax

Grimshaw denes Gen as a function generating all possible analyses of an

input within the universal structural requirements of X theory. That is, candidate
analyses that are submitted for evaluation by the constraint hierarchy must all be
proper X structures this may be considered as an inviolable constraint, or
alternatively, as a property of Gen. Grimshaw in fact assumes the following rather
minimal version of X theory:

(4) X theory (respected under analysis by Gen): each node must be a

good projection of a lower one, if a lower one is present.

Note that this X theory does not require that some head be present in every
projection: such a requirement will in fact be made by a violable constraint,
which is referred to by Grimshaw as Obligatory Heads. (We will nd out why
this is a violable constraint later on when we come to discuss patterns of sub-
ordinate interrogatives.)

The notion of extended projection plays a central role in Grimshaws paper, as it

restricts the nature of the candidate set (only extended projections are fed into
Eval ). It is dened as follows (after Grimshaw 1991):

(5) Extended projection: a unit consisting of a lexical head and its

projection plus all functional projections erected over the lexical
projection.

By this denition IP and CP are both extended projections of V (the lexical head
which has VP as its minimal projection). The schematic structure in (6) contains
a verb in its own projection VP, upon which a hierarchy of extended projections
are built (here consisting of a Inectional Phrase and a Complementizer Phrase).

(6) CP

XP C

C IP

XP I

I VP

XP V

V ...
346
 8.2 OT and syntax

In sum, every candidate analysis of an input must be an extended projection

conforming to the minimal X theory in (4).

Let us now turn to the types of syntactic operations that Gen may perform on
the input, while keeping within the previously dened limits. These include the
following:

(7) Operations allowed under analysis by Gen

introducing (extended) projections conforming to X theory
introducing functional heads as they do not appear in the input, due
to their lack of semantic content (e.g. the complementizer that and
do-support in English)
introducing empty elements (traces, etc.), as well as their coindexa-
tions with other elements
moving lexical elements

Note that, in keeping with Containment (3), no elements may be literally removed
from the input, and no semantically meaningful elements may be introduced.
Let us now consider a number of candidate analyses of the input of the sen-
tence What did Mary say? as it was dened in (2) above:

(8) Various candidate analyses for {say (x, y), x = Mary, y = what,
tense = past)
a. [IP Mary [VP said what]]
b. [CP what [IP Mary [VP said t]]]
c. [CP what saidi [IP Mary [VP ei t]]]
d. [CP what didi [IP Mary ei [VP say t]]]

These outputs are all extended projections, conforming to the X schema dened
in (45). Evidently these analyses do not all correspond to grammatical sentences
of English: in fact only one does: (8d). This involves Wh-movement, do-support,
and subjectauxiliary inversion. Yet all of the analyses (8ac) correspond to gram-
matical sentences in languages other than English. Analysis (8a), for example, has
no movement of any kind, as in Chinese. Analysis (8b) involves Wh-movement,
but it lacks both subjectverb inversion and do-support (or its analogue), the set
of properties found in Czech and Polish. Finally, analysis (8c) has both Wh-
movement and subjectlexical verb inversion, but it lacks do-support, as in Dutch.
The idea of the OT approach to syntax is that the diversity of syntactic struc-
tures across languages reects differences in the rankings of universal and viol-
able constraints. Each of the analyses in (8) violates one or more constraints, but
each can be considered as maximally harmonic (optimal ) with respect to some
347
Extensions to syntax

ranking of constraints that is specic to some language. But in order to be able

to develop such an analysis, we must of course rst dene the constraints. This
will be the topic of the next subsection.

8.2.4 Dening the syntactic constraint inventory

The nal (and in many respects crucial) component of an OT syntax that we must
dene is Eval, the inventory of universal and violable constraints to evaluate
candidate outputs. Grimshaws paper contains a set of violable constraints, all of
which are evaluated at the level of surface structure. In this section we will con-
sider a number of constraints that play an important role in Grimshaws analysis
of extended verbal projections.

Exactly as in phonology, well-formedness constraints make general requirements

of structural well-formedness of the output. Such requirements involve speciers,
heads of projections, government, or other aspects of syntactic structure. An essen-
tial function of syntactic well-formedness constraints is to trigger movement of
elements. Now consider a rst example of a syntactic well-formedness constraint,
Op-Spec:

(9) Operator in Specier (Op-Spec)

Syntactic operators must be in specier position.

The primary function of this constraint is to force the movement of Wh-elements

(which are syntactic operators) into the specier position of some extended pro-
jection, typically the specier position of a Complementizer Phrase (or Spec-of-
CP), as below.6

(10) CP

Spec-of-CP ...

Wh
Spec-of-CP will be the single available position for a Wh-element to move into
when the specier position of the next-lower projection, Spec-of-IP, is occupied
by the subject. Wh-movement into Spec-of-CP produces a chain of effects in
English syntax, which are known as subjectauxiliary inversion and do-support.

6
Later we will go into the question of what blocks movement of a Wh-phrase to other specier
positions, such as Spec-of-VP or Spec-of-IP. Furthermore, Grimshaw assumes that the option
of base-generating the Wh-phrase in Spec-of-CP is ruled out by the Theta Criterion, which may
be interpreted as a part of Gen, or as a constraint dominating Stay.

348
 8.2 OT and syntax

That is, the creation of Spec-of-CP implies the presence of a CP, due to inviolate
X requirements. This CP, in its turn, is an extended projection subject to struc-
tural constraints that, indirectly, trigger the effects mentioned above.

Consider, for example, the following structural constraint, whose function should
be clear from its statement:

(11) Obligatory Heads (Ob-Hd)

A projection has a head.

For verbal projections, Verb Phrase and its extended projections Inectional
Phrase and Complementizer Phrase, this implies that each must have a verb as
its head. Satisfaction of this constraint may require the movement or insertion of
a verb to become the head of an extended verbal projection. For example, Ob-Hd
may trigger subjectauxiliary inversion, the movement of an auxiliary to become
the head of CP (Den Besten 1983).7
The representation below shows the output of subjectauxiliary inversion in the
sentence What will Mary say?:

(12) CP

whati C

C IP

willj Mary I

I VP

ej V

V ti
say

The auxiliary verb willj undergoes head-movement to CP, from its original posi-
tion in IP, as has been indicated by its trace ei. Crucially, both the auxiliary verb
and its trace function as proper heads in their respective extended projections, IP
and CP.

7
Of course Ob-Hd can also be satised by a complementizer (that, whether, etc.) in the head
of CP.

349
Extensions to syntax

Naturally, Gen is allowed to generate competitor candidates to (12). For example

consider a schematic candidate (13a) in which the auxiliary will is not moved into
CP (but remains the head of IP), and yet another candidate (13b) in which will is
generated as the head of CP, without movement from IP:

(13) a. CP b. CP

C C

C IP C IP

I will I

I VP I VP

will

Each of these analyses violates Ob-Hd, however, since each has one projection
that is not headed: CP in (13a) and IP in (13b).

Apart from Op-Spec and Ob-Hd, a number of other well-formedness constraints

are proposed by Grimshaw (1997). We will not consider these, however, but turn
to the question of what are the costs, stated in terms of constraint violations, of
satisfying Op-Spec and Ob-Hd. That is, we now turn to the economy, or faith-
fulness, part of the constraint inventory.

To dene faithfulness constraints in syntax requires some degree of elaboration.8

The notion of faithfulness, as we have used it throughout this book, involves
evaluation of the degree of identity between the input and output. In phonology,
the input is dened as a phonological representation itself, hence it is possible
literally to compare the input and output as to their degree of identity. This is, of
course, the very essence of correspondence constraints: evaluating the degree of
similarity between the input and output for various aspects of phonological rep-
resentation. A priori, it seems entirely logical to conceive of an anti-movement
constraint as the syntactic counterpart of (phonological) Linearity as both mil-
itate against differences in the order of elements in the input and the output.
Along the same lines, a syntactic constraint do not insert is a plausible counter-
part of Dependence.
8
See Keer and Bakovic (1997) on syntactic faithfulness. See Bresnan (forthcoming) on syntac-
tic correspondence, as well as a reanalysis of the results of Grimshaw (1997) in the framework
of Lexical-Functional Grammar.

350
 8.2 OT and syntax

However, the input of the syntax as Grimshaw denes it (see again section
8.2.2) is, strictly speaking, not a (hierarchical) syntactic representation itself. It
only contains lexical heads and their argument structure, but not extended syntac-
tic projections, nor functional heads. This renders a correspondence approach to
economy, in principle at least, problematic.
Nevertheless, we may still conceive of syntactic faithfulness constraints in a
less correspondence-like fashion. First consider the general anti-movement con-
straint Stay:

(14) Economy of Movement (Stay)

Trace is not allowed.

The function of this constraint is to militate against any movement per se, or in
the case of gradient violation, to select the shortest movement, the one that has
the minimal number of intermediate steps. What we refer to as a movement is
actually construed as a chain of an element plus its traces as this occurs at surface
structure, the level at which Stay is evaluated. For example, the chain of a Wh-
element and its trace ti in (15a) incurs one violation of Stay, while the chain of
an Aux-element and its two traces ei in (15b) incurs two violations:

(15) a. Whi . . . ti
b. Auxi . . . ei . . . ei

Next consider the question of why Stay penalizes traces, rather than divergences
of word order between the input and output. (A correspondence approach to
economy of movement might set up constraints measuring the distance between
some element at surface structure and its input correspondent.) One problem,
already pointed out above, is that the input is not a syntactic representation, so
that the position of input elements cannot be established. Furthermore, under a
denition of the input as a lexical head plus its argument structure, it would
become impossible to penalize any movement of elements that are not present at
the level of the input.9 A third reason why such a correspondence approach
to economy may not be feasible is the following. What matters to economy of
movement is the number of landing sites of an element between its base position
and its position at surface structure, rather than its distance.10 Since every trace
9
A correspondence approach to economy of movement which would circumvent this problem
would involve setting up an intermediate syntactic level of deep structure that is distinct
from Grimshaws input: it is dened as a phrase-marker in which every element occupies its
base-generated position. This level would allow the representation of syntactic elements that
have no counterparts in the input (or lexical conceptual structure).
10
Ackema and Neeleman (forthcoming) propose an evaluation of Stay that is based on the
number of nodes intervening between positions in a movement chain. It remains an empirical
issue which type of evaluation is most adequate.

351
Extensions to syntax

in a chain incurs a violation of Stay, surface structure chains such as (15) are
indispensable in measuring violations.

Stay subsumes all types of movement, including Wh-movement and movement

of lexical and functional heads. A more specic ban against movements is made
by the next constraint, No-Lex-Mvt:

(16) No Movement of a Lexical Head (No-Lex-Mvt)

A lexical head cannot move.

The effect of this constraint is that a lexical head (for example, a lexical verb)
must stay in the projection that it heads (e.g. the VP). The conguration (17)
presents the typical case of violation of No-Lex-Mvt:

(17) . . . Vi . . . [VP . . . ei . . .]

Note that No-Lex-Mvt is a special case of the general anti-movement constraint

Stay. That is, every violation of No-Lex-Mvt implies a violation of Stay, but
not conversely. This subset relation means that independent effects of No-Lex-
Mvt may occur only in a language which ranks No-Lex-Mvt above Stay (while
separating both constraints by some intervening constraint). As we will see below,
this is the case in English.

Finally, consider the constraint that militates against the suppression of semantic
roles of lexical elements, Full-Int:

(18) Full Interpretation (Full-Int)

Lexical conceptual structure is parsed.

The key function of this constraint is to ban semantically empty auxiliary verbs,
as in do-support in English. In a sentence such as What did Mary say? the verb
do is semantically empty, functioning merely as an auxiliary for the lexical verb
say. It will be assumed that do in such sentences is lacking from the input, and
inserted from the lexicon into candidate analyses by Gen. Since, as was argued
above, Gen is semantically neutral, any insertion of a lexical element must occur
at the expense of its lexical conceptual structure. Observe that the lexical verb do
is a theta-marker (hence, an argument-taker) in a sentence such as Mary did the
vocal parts. This can be represented as below:

(19) do (x, y)

Under do-support, this lexical conceptual structure of the verb do is not parsed,
hence the sentence What did Mary say? incurs a violation of Full-Int. In
352
 8.3 The structure of extended verbal projections in English

sum, Full-Int expresses a faithfulness requirement with respect to lexical ele-

ments, which is violated if the output suppresses their theta-marking and argument-
taking properties.
Since violation of constraints is minimal, we expect that lexical verbs are used in
a semantically empty way only when this is absolutely necessary. This idea is the
base of Grimshaws OT analysis of do-support in English, to which we now turn.

8.3 The structure of extended verbal projections in English

Empirically, Grimshaw (1997) focusses on patterns of Wh-movement, subject
verb inversion, and do-support in English. All three phenomena strongly interact,
and we will see that this interaction centres on extended verbal projections:
syntactic structures on top of the VP, which include the Inectional Phrase (IP)
and Complementizer Phrase (CP). (See again representations (6), (12), and (13)
for illustration.) The specic issue that we will address is this: can economy of
derivation and representation, as this is manifest in the domain of extended verbal
projections, be reduced to the interaction of syntactic constraints? Our answer
will be positive; we will in fact argue that the analysis proposed by Grimshaw
(1997) is naturally integrated with general principles of OT, and that it is, point
by point, analogous to OT analyses of phonological phenomena involving eco-
nomy, such as epenthesis.
Directly below (in section 8.3.1) we will discuss Wh-movement and subject
auxiliary inversion, and we will turn to do-support in section 8.3.2.

8.3.1 Wh-movement and subjectauxiliary inversion

We begin our exploration by considering a declarative sentence containing the
auxiliary verb will (marking future tense). In declarative sentences such as (20a) the
inversion of the subject and auxiliary verb is not allowed (compare the ungram-
maticality of 20b):

(20) a. Mary will say too much.

b. *Will Mary say too much.

In Grimshaws analysis of a simple declarative sentence such as (20a), the verb

phrase (or VP) contains the lexical verb say and its object too much, while the
subject Mary and the auxiliary verb will occupy positions outside the VP, in the
Inectional Phrase (IP):

(21) [IP Mary will [VP say too much]]

Each projection, the lexical projection VP and its extended projection IP, is
headed by an appropriate head. The VP is headed by the lexical verb (say), while
the IP is headed by the auxiliary verb (will). Since no heads are missing from this
353
Extensions to syntax

structure, there is no need for an additional CP plus inversion of the subject and
auxiliary verb, as in (22):

(22) [CP willi [IP Mary ei [VP say too much]]]

Structure (22) has unnecessary complexities in comparison to (21). It has an

additional extended projection CP, and it also has an additional movement to ll
the head of CP. Such avoidance of unnecessary structure and movement is what
Chomskys Minimalist programme refers to as economy. In OT, economy is
subsumed under the do only when necessary principle. That is, inversion in the
absence of any structural necessity leads to a fatal violation of the anti-movement
constraint Stay, as in (23b):11

(23) {say (x, y), x = Mary, y = too much, Op- Ob- Stay
tense = future, auxiliary = will} Spec Hd

a. + [IP Mary will [VP say too much]]

b. [CP willi [IP Mary ei [VP say too much]]] *!
c. [CP e [IP Mary will [VP say too much]]] *!

Candidate (23c) is a failed attempt at improving over candidate (23b). It maintains

CP, but leaves the auxiliary will in head position in IP. Although avoiding a
violation of Stay, this attempt fails by its violation of Ob-Hd, requiring that every
projection must have a head (which is violated by the CP in 23c). In sum, no
inversion takes place in declarative sentences because of lack of necessity both
Op-Spec and Ob-Hd are satised without movement.
We make the additional observation that the economy of the optimal output
(23a) is independent of the constraint ranking. It is easy to see that no ranking is
possible under which less-economical candidates (23b c) could ever defeat

11
Under the hypothesis that the subject is base-generated in the VP (Zagona 1982), an additional
candidate merits consideration: [IP will [VP Mary say too much]]. This would even be optimal,
since, as compared to (23a), it avoids movement of the subject into Spec-of-IP. Grimshaw
rules this candidate out by the constraint Subject a clause must have a subject, in which
subject is dened as the highest A(rgument) specier in a clause. If Subject dominates
Stay, the subject is driven out of VP into IP, which would otherwise remain subject-less:

{say (x, y), x = Mary, y = too much, Op-Spec Ob-Hd Subject Stay
tense = future, auxiliary = will}

a. + [IP Mary will [VP t say too much]] *

b. [IP will [VP Mary say too much]] *!

354
 8.3 The structure of extended verbal projections in English

(23a), which is awless with respect to all three constraints. A more general predic-
tion is that no language has subjectauxiliary inversion in simple declaratives.
Next consider interrogatives, where Wh-movement is obligatory, in combina-
tion with inversion, as in (24a):

(24) a. What will Mary say?

b. *What Mary will say?
c. *Will Mary say what?
d. *Mary will say what?

Wh-movement without inversion (24b) causes ungrammaticality, and so does lack

of Wh-movement (24c d). Combining this observation on interrogatives with
those on declaratives (see 212), we arrive at the following generalization:12

(25) Generalization
Subjectauxiliary inversion occurs if and only if Wh-movement
applies.

Observe the biconditional nature of (25): subjectauxiliary inversion is obligatory

if Wh-movement applies (hence *What Mary will say?); conversely inversion is
impossible elsewhere (hence *Will Mary say too much). What causes this corre-
lation of Wh-movement and inversion? We can only answer this question by rst
addressing two related issues. First, what makes Wh-movement obligatory? Sec-
ond, what functional projections are involved in Wh-movement and inversion?
Wh-movement itself is obligatory due to high-ranking Op-Spec (9) requiring
that syntactic operators occupy a specier position. Wh-phrases are operators
(Rizzi 1991), and hence must move into some specier position. But to which
specier position can a Wh-phrase in sentences such as (24) move? The subject
already occupies the specier position in IP, hence this position is not available.
Therefore an additional extended projection is created outside the IP, which is the
CP. At this point, we must consider the following pair of structures, both of which
have Wh-movement to CP:

(26) a. *[CP what [IP Mary will [VP say t ]]]

b. [CP what willi [IP Mary ei [VP say t ]]]

In (26a) the Wh-phrase occupies the specier position of CP, as is required by

Op-Spec. However, the head of CP is missing, which violates Ob-Hd, the con-
straint requiring that every projection must be headed. By moving the auxiliary
verb into CP, as in (26b), this defect is repaired, since the auxiliary verb provides

12
Inversion is triggered also by yes/no interrogatives, as in Will Mary say anything? See Grim-
shaw (1997) for analysis of this construction.

355
Extensions to syntax

a proper head for CP. At the same time, its trace ei in IP still functions as the
head of IP. Crucially, both lexical heads and traces can serve as heads of a
projection.

We thus arrive at the following scenario: Wh-movement serves to place the Wh-
element, an operator, in specier position of CP (as forced by Op-Spec); con-
sequently the auxiliary verb must also move into CP to provide a head for this
extended projection (due to Ob-Hd). Both movements naturally take their toll as
they occur at the expense of violations of Stay, the general anti-movement con-
straint. This motivates the following ranking:

(27) Inversion only under Wh-movement

Op-Spec, Ob-Hd
Stay

Op-Spec must dominate Stay in order for Wh-movement to take place; with the
reverse ranking movement would be blocked by Stay. And Ob-Hd must also
dominate Stay in order for head movement of the auxiliary verb to take place.
This ranking is shown to be appropriate in the following tableau:

(28) {say (x, y), x = Mary, y = what, Op- Ob- Stay

tense = future, auxiliary = will} Spec Hd

a. [IP Mary will [VP say what]] *!

b. [CP e [IP Mary will [VP say what]]] *! *
c. [CP what e [IP Mary will [VP say t ]]] *! *
d. + [CP what willi [IP Mary ei [VP say t ]]] **
e. [CP willi [IP Mary ei [VP say what]]] *! *

The optimal candidate (28d) is the only one which satises both Op-Spec and Ob-
Hd. Its representational perfection is purchased at the expense of two violations
of economy of movement (Stay), one for the trace t left by Wh-movement, and
another for the trace ei left by head movement. But (28d) is optimal precisely for
the reason that it cannot be improved over by other candidates. To improve (28d)
could only mean to avoid violations of Stay, hence to reduce movements. This,
however, cannot be achieved without violating higher-ranked constraints, Op-
Spec or Ob-Hd. For example, no improvement is made over (28d) by candidates
(28ab), which totally avoid movements, but consequently run into fatal viola-
tions of Op-Spec (28a), or even worse, of both Op-Spec and Ob-Hd (28b).
Next, candidate (28c) has the apparent virtue of avoiding head movement to CP.
356
 8.3 The structure of extended verbal projections in English

However, this leaves CP unheaded, fatally violating Ob-Hd. Finally, candidate

(28e) atly avoids Wh-movement, fatally violating Op-Spec.13,14

Let us now see how the property of economy follows from the OT analysis. For
the sake of the argument, we distinguish derivational and representational eco-
nomy. Derivational economy (no unnecessary movement) is enforced directly
by Stay. Head movement of the auxiliary occurs whenever necessary (that is,
in interrogative sentences) to meet the top-ranking structural constraint Ob-Hd
even though movement violates Stay. Naturally any movement of the auxiliary
that yields no gains with respect to Ob-Hd will be blocked by Stay (that is,
in declarative sentences). Representational economy (no empty projections) is
achieved without a constraint explicitly militating against unnecessary projec-
tions. Instead, exclusion of superuous extended projections such as CP is
achieved indirectly, by cooperation of two constraints, Ob-Hd and Stay. To posit
a projection is to accept responsibility for its structural well-formedness. But
doing so may cause problems of derivational economy: an extended projection
left unheaded is ruled out by Ob-Hd. Alternatively, an extended projection sup-
plied with a proper head by head movement will be ruled out by Stay (except
when it yields an improvement with respect to Ob-Hd).

13
Again it may be interesting to consider the consequences of base-generating the subject in VP.
The candidate left out of consideration in tableau (28) is [CP what will [VP Mary say t ]], which
lacks an IP altogether. The subject Mary occupies Spec-of-VP, and the Wh-element is moved
to Spec-of-CP. This structure is more economical than optimal (28d = b below) since it has
no movement of auxiliary to IP. See the following tableau:

{say (x, y), x = Mary, y = what, Op-Spec Ob-Hd Stay

tense = future, auxiliary = will}

a. + [CP what will [VP Mary say t ]] *

b. [CP what willi [VP Mary ei [VP t say t ]]] **!*

c. [CP what e [IP Mary will [VP t say t ]]] *! **

Grimshaw does not discuss this option, presumably because she assumes that auxiliaries
always head the IP, either in situ, or as a trace. Moreover, this additional candidate does not
affect the main result that inversion is obligatory in this case, due to the ranking Op-Spec,
Ob-Hd
Stay. That is, the non-reversing candidate (c) is ruled out in the same way as it is
in (28c).
14
Peter Ackema (p.c.) points out that this analysis requires modications in order to account for
double Wh-constructions. In languages like English, only one Wh-element may move into CP,
while all others remain in situ (Who has given what to whom yesterday?). Here Grimshaws
analysis, which has undominated Ob-Hd and Op-Spec, predicts that every Wh-element will
be moved to a new CP, with the auxiliary verb moving upward through the head position of
each CP (*Who has what to whom given yesterday?).

357
Extensions to syntax

This analysis is unifying in the sense that both types of economy are, either
directly or indirectly, reduced to a single economy constraint, Stay. Economy
requires no separate constraints for derivation and representation. Stay directly
enforces derivational economy, while it indirectly enforces representational eco-
nomy, assisted by Ob-Hd.
In sum, given the pair Ob-Hd and Stay, both movement and extended pro-
jections are kept at a bare minimum. No movement will occur, and no extended
projection will be constructed, unless it is necessary. This warrants the conclusion
that economy is enforced by every grammar containing Ob-Hd and Stay hence
that economy is universal.

8.3.2 Do-support

8.3.2.1 The pattern

We now turn to the phenomenon of do-support as this occurs in interrogative
sentences, and to its interaction with Wh-movement and inversion. The generaliza-
tion to be captured by the analysis is simple: do-support is possible only when
it is necessary (Chomsky 1957, 1991). We will see that OT offers a straightfor-
ward explanation of this generalization. Because of space limitations we cannot
address the complete distribution of the auxiliary do here, and we refer the reader
to Grimshaw (1997) for further discussion.
The key observation which we will focus on in this section, and which will
serve to highlight Grimshaws analysis, is that do occurs in simple interrogative
sentences where it is the single auxiliary verb. Or to state it in negative contexts,
do cannot occur in positive declarative sentences, nor can it occur in any sentence
that has another auxiliary verb. We thus nd a complementary distribution of do
and other auxiliaries.

Firstly, do is obligatorily present in simple interrogative sentences:15

(29) a. What did Mary say?

b. *What Mary said?

Observe that do occupies a position typical for the auxiliary in interrogatives,

immediately after the Wh-element (compare What will Mary say? in 24a and
30a). This is, of course, a key to the explanation of the complementary distribu-
tion of do and other auxiliary verbs: do functions as obligatory head of the CP,
precisely like the auxiliary will in (26b). Filling the head position of the CP is

15
A third logical possibility, that of inversion involving the lexical verb (cf. What said Mary?),
will be discussed in section 8.3.2.2.

358
 8.3 The structure of extended verbal projections in English

what is meant by when it is necessary in the generalization about do-support

stated above.

Secondly, do cannot cooccur with other auxiliary verbs in interrogatives:

(30) a. What will Mary say?

b. *What does Mary will say?
c. *What will Mary do say?

This complementary distribution of do and other auxiliary verbs in interrogatives

points to the conclusion that do functions as a kind of place-holder for the aux-
iliary (as it appears in sentences such as 30a). When another auxiliary verb is
already in place, do is no longer necessary, and therefore it will be left out for
reasons of economy.

Thirdly, that do-support is impossible in positive declarative sentences is shown

by (31a) and its ungrammatical counterpart in (31b):

(31) a. Mary said much.

b. *Mary did say much.

Positive declarative sentences are structured in a way that renders addition of the
auxiliary do superuous. As expected under economy, superuity entails pro-
hibition. (In section 8.3.2.2 we will return to the structure of such sentences.)

Fourthly, as we now expect, the occurrence of auxiliary do is ruled out in declara-

tive sentences that already contain another auxiliary verb, such as will:

(32) a. Mary will say much.

b. *Mary does will say much.
c. *Mary will do say much.

Again, there is a lack of necessity for do-support: (32a) already contains an

auxiliary verb in the proper position. Again, this distribution is economical: lack
of necessity for do goes hand in hand with an actual ban on do-support.

Fifthly, auxiliary do cannot cooccur with itself, even in interrogatives:

(33) a. What did Mary say?

b. *What did Mary do say?

A single occurrence of auxiliary do sufces, any additional occurrences are

ruled out. This is yet another demonstration of economy: no more occurrences of
do-support take place than are strictly necessary.
359
Extensions to syntax

The generalization about do-support can now be stated in a maximally simple

way:

(34) Generalization about do-support (Chomsky 1957, 1991)

The auxiliary do is possible only when it is necessary.

This reminds us of generalizations found earlier in this book, all having the gen-
eral form do only when necessary. Then do referred to some change imposed
on the input, for example an epenthesis or deletion of a segment. Here, do
literally refers to do-support.
Grimshaw goes on subsume generalization (34) under the standard OT account
for generalizations of the do only when necessary type: a domination of well-
formedness over faithfulness.

8.3.2.2 The analysis

As argued above, the auxiliary do is a semantically empty verb, one which only
serves the syntactic function of head of extended projections which, without it,
would be fatally unheaded. The semantic emptiness of auxiliary do implies the
suppression of the lexicalconceptual properties of its source, the lexical verb do,
which are roughly similar to that of to act. (Hence the double occurrence of do
in Did you do your homework?: the rst do is an auxiliary, the second a lexical
verb.) The suppression of lexicalconceptual properties of do constitutes a viola-
tion of Full-Int (section 8.2.4), requiring that lexicalconceptual structure be
parsed. This is the price to be paid for using do as an auxiliary verb, which is
forced by higher-ranking constraints requiring the presence of some verbal head;
specically Ob-Hd.
Grimshaw points out that it is not at all a coincidence that English uses do as
the semantically empty verb: among the lexical verbs of English, do is (appar-
ently) the verb which has the smallest lexical conceptual structure of its own.
Using do as a semantically empty verb causes the minimal suppression of lexical
conceptual structure, which explains why other verbs are not used in a semantic-
ally empty way. In fact the prediction is broader than English: every language
having an equivalent of do-support will pick the lexical verb of minimal semantic
content. Of course it remains to be seen whether a theory of semantic lightness
can be developed which picks out the minimal verb for any language.
Note that Grimshaws analysis of do-support bears a resemblance to the ana-
lysis of epenthesis presented in chapter 3. Epenthetic segments in phonology are
considered to be lacking from the input since they belong to no input morpheme,
and lack inherent feature content. In phonology epenthesis follows from interac-
tions of well-formedness constraints (making requirements of syllable structure)
and faithfulness constraints (specically Dep, militating against output elements
360
 8.3 The structure of extended verbal projections in English

not present in the input). Again the parallelism extends to the syntactic analysis:
do-support is triggered by the well-formedness constraint (Ob-Hd), at the expense
of violations of the syntactic faithfulness constraint Full-Int.

Once we have spotted the main idea behind the analysis, its execution becomes a
fairly simple matter. Evidently Full-Int must be dominated by Ob-Hd, since
violations of Full-Int (in the form of do-support) are tolerated to avoid viola-
tions of Ob-Hd.
(35) Ob-Hd
Full-Int
Turning now to an actual example, this ranking is supported by the comparison
of two candidates for an interrogative, one with do-support (36a), another without
it (36b):
(36) a. What did Mary say?
b. *What Mary said?

Thus far we have not seen any evidence bearing on the ranking of Full-Int with
respect to Stay. Lacking this evidence, we arrive at a partial hierarchy in (37),
which inserts the ranking Ob-Hd
Full-Int (35) into the earlier hierarchy
Op-Spec, Ob-Hd
Stay (27):
(37) do-support only when necessary
Op-Spec, Ob-Hd
Full-Int, Stay

In tableau (38), both candidates (36ab) are represented as (38a b). The former
candidate is optimal, as it has no violations of either highest-ranking constraint
Op-Spec and Ob-Hd, while all other candidates violate at least one of these. In
(38a) the auxiliary verb did heads CP, while its trace ei properly heads IP. Since
all its extended projections are headed, (38a) fully satises Ob-Hd. This happens
at the expense of violations of lower-ranked Full-Int (due to do-support) and
Stay (due to Wh-movement and head movement of did to CP).16
16
If subjects are base-generated in VP, a possibility mentioned in notes 1113, the optimal can-
didate would be [CP what did [VP Mary say t ]], leading to the same word order. Movement of
Mary out of VP is not necessary to provide a subject for the clause (cf. Subject in note 11),
hence this would violate Stay without reason. Avoidance of IP brings yet another advantage:
the auxiliary need not move to become head-of-CP, but can be simply inserted in that position:

{say (x, y), x = Mary, y = what, Op-Spec Ob-Hd Full-Int Stay

tense = past}

a. + [CP what did [VP Mary say t]] * *

b. [CP what didi [IP Mary ei [VP t say t]]] * **!*

361
Extensions to syntax

(38)
{say (x, y), x = Mary, y = what, Op- Ob-Hd Full- Stay
tense = past} Spec Int

a. + [CP what didi [IP Mary ei [VP say t ]]] * **

b. [CP what e [VP Mary said t ]] *! *
c. [CP what e [IP Mary did [VP say t ]]] *! * *
d. [CP what e [IP Mary e [VP said t ]]] *!* *
e. [IP Mary [VP said what]] *!

Its direct competitor (38b) omits do-support, thereby avoiding a violation of

Full-Int, as well as an additional violation of Stay due to head movement of
do. But at a fatal cost: by leaving the CP unheaded, it violates Ob-Hd. Note that
the subject of (38b) occupies Spec-of-VP, a possibility that we did not consider
until here (but see notes 11, 13, and 16). The reason for this previous neglect was
that VP-internal subjects in earlier examples did not affect the result. But from
here on, we will include this option in candidate sets. It will be optimal in some
examples to be discussed below (starting with tableau 42).
Turning to the remaining competitors (38ce), we nd that none can improve
on (38a) by avoiding violations of Full-Int or Stay, without incurring fatal
violations of Op-Spec or Ob-Hd. Candidate (38c) has do-support, thus providing
a head for IP, but no head movement, leaving CP unheaded and causing a fatal
violation of Ob-Hd. Next, candidate (38d) not only lacks a head of CP but also
leaves IP unheaded, causing two fatal violations of Ob-Hd. Finally, candidate
(38e) represents an attempt to avoid violations of Ob-Hd by not creating extended
projections, neither IP nor CP. This attempt, of course, runs into Op-Spec, requir-
ing that Wh-elements occupy a specier position.
The attentive reader may have noticed that one obvious output candidate is
missing from tableau (38). This involves inversion on the lexical verb:

(39) a. *What said Mary?

b. [CP what saidi [VP Mary ei t ]]

Apparently the lexical verb is not allowed to move (due to No-Lex-Mvt, intro-
duced in section 8.2.4). We postpone discussion of this construction, including its
typological status, to section 8.4, and now continue our survey of the distribution
of the auxiliary do in English.
362
 8.3 The structure of extended verbal projections in English

As noted earlier, the auxiliary do stands in complementary distribution with other

auxiliaries, such as will. Do-support causes ungrammaticality in interrogatives
containing another auxiliary. See again the set of sentences below, repeated from
(30):

(40) a. What will Mary say?

b. *What does Mary will say?
c. *What will Mary do say?

This complementary distribution is readily captured by the constraint hierarchy.

Recall that the difference between do and other auxiliaries (such as can, will,
may) is that the latter have semantic and functional content, and therefore must
occur in the input. In contrast, do is semantically empty, hence it is not part of
the input. The generalization is that whenever an input auxiliary is available as
the head of CP, there is no need for inserted auxiliary do. Since epenthesis of an
auxiliary amounts to a violation of Full-Int, do-support is avoided whenever the
input contains an auxiliary. Similarly, epenthesis in phonology is avoided when-
ever possible, avoiding unnecessary violation of Dep-IO.
The tableau of What will Mary say? contains the three candidates of (40),
plus two additional candidates (41de), which are ruled out by undominated
constraints.

(41)
{say (x, y), x = Mary, y = what, Op- Ob- Full- Stay
tense = future, auxiliary = will} Spec Hd Int

a. + [CP what willi [IP Mary ei [VP say t ]]] **

b. [CP what does [IP Mary will [VP say t ]]] *! *
c. [CP what willi [IP Mary ei [XP do [VP say t ]]]] *! **
d. [CP what e [IP Mary will [VP say t ]]] *! *
e. [IP Mary will [VP say what]] *!

This tableau adduces evidence that Full-Int dominates Stay.17 This resides in
the fact that candidate (41a) has more violations of Stay than (41b), but is
nevertheless optimal as it has no violations of Full-Int.

The next tableau evaluates candidates for simple declarative sentences that lack
an auxiliary (Mary said little). Economy (no do-support) follows from the same

17
Grimshaw (1997) does not discuss this ranking.

363
Extensions to syntax

constraint ranking as before. Interestingly, the ranking now predicts an analysis

under which the full clause, including its subject and predicate, is a single VP:

(42) {say (x, y), x = Mary, y = little, Op- Ob- Full- Stay
tense = past} Spec Hd Int

a. + [VP Mary said little]

b. [IP Mary did [VP say little]] *!
c. [IP e [VP Mary said little]] *!

Consider how this result follows. A matrix clause that lacks auxiliaries in the
input is, in a way, self-supporting. Specically it has no need for IP, an extended
projection canonically headed by auxiliaries, nor is there any need for CP, due to
the lack of Wh-elements. As we have seen often before, when extended projec-
tions are unnecessary, syntactic faithfulness rules them out. The economy mech-
anism is essentially the same as before: introduction of extended projections
triggers the well-formedness constraints for projections. Here Ob-Hd requires that
IP be headed, thereby excluding candidate (42c). Moreover, no head of IP can be
supplied by movement of an auxiliary as no auxiliary occurs in the input. (The
option of moving the lexical verb will be discussed in section 8.4.2.) The single
remaining option to provide a head for IP is do-support, as in (42b). But this
option crashes into Full-Int. In sum, all attempts at postulating an IP have
resulted only in losses, and no gains, as compared to the structurally simpler
candidate (42a). Again, representational complexity never wins unless it brings a
bonus in terms of lesser violations.

Finally, the observation that do-support never cooccurs with itself gives additional
evidence that Full-Int dominates Stay.

(43)
{say (x, y), x = Mary, y = what, Op- Ob- Full- Stay
tense = past} Spec Hd Int

a. + [CP what didi [IP Mary ei [VP say t ]]] * **

b. [CP what did [IP Mary do [VP say t ]]] **! *
c. [CP what e [IP Mary did [VP say t ]]] *!
d. [VP Mary said what] *!

364
 8.3 The structure of extended verbal projections in English

It is more economical to insert an auxiliary in IP and to move it to CP, than to

introduce auxiliaries in both extended projections. (But see note 16.)

All examples of Wh-movement discussed thus far involved Wh-elements which

are objects to the lexical verb. In such cases we always nd inversion, in com-
bination with do-support when the input lacks an auxiliary. Observe that, as
compared to Wh-objects, Wh-subjects behave differently, in that do-support is no
longer required:

(44) a. Who said that?

b. *Who did say that?

In fact, this observation readily follows from the analysis as it stands now. The
difference between Wh-objects and Wh-subjects is that the latter can satisfy Op-
Spec by staying in Spec-of-VP, a legitimate specier position for a syntactic
operator. In contrast a Wh-object must move out of VP to nd a proper specier
position, in particular when Spec-of-VP is already occupied by the subject or its
trace.
The following tableau of sentence (44a) shows that without structural necessity
for Wh-subjects to move, their movement is blocked:

(45) {say (x, y), x = who, y = that, Op- Ob- Full- Stay
tense = past} Spec Hd Int

a. + [VP who said that]]]

b. [IP who did [VP t say that]]] *! *
c. [IP who e [VP t said that]]] *! *

This tableau of an interrogative is strikingly similar to that of the simple declar-

ative in (42).18

18
The fronted position of the subject Wh-element in Who will say that? is due to the constraint
Subject, which was previously mentioned in notes 11 and 16. Structure (b) with the auxiliary
in I, and the Wh-element in Spec-of-VP, would violate Subject since this constraint requires
that the highest A(rgument) specier in a clause have a subject, in this case Spec-of-IP. Hence
Wh-movement becomes obligatory:

{say (x, y), x = who, y = that, Op-Spec Ob-Hd Subject Full-Int Stay
tense = future, auxiliary = will}

a. + [IP who will [VP t say that]] *

b. [IP will [VP who say that]] *!

365
Extensions to syntax

This concludes our discussion of Grimshaws analysis of the interactions of

Wh-movement, inversion, and do-support. In her paper, Grimshaw discusses a
larger array of constructions than we are able to do here, including negative
sentences and subordinate clauses. The interested reader is referred to the paper.
Before we evaluate the analysis in terms of its adequacy, we will rst address its
typological consequences, on the basis of constraint reranking, in section 8.4.

8.4 Typological consequences

Below a factorial typology will be sketched of Wh-movement, inversion and
do-support, as predicted on the basis of Grimshaws constraint inventory. In
chapter 2 we discussed the following assumptions underlying the OT approach
to typology. Constraints are universal while their ranking is language-specic.
Since language variation stems from reranking the set of universal constraints,
the grammatical patterns of languages emerge as variations on universal themes.
While typological variation due to constraint reranking is considerable, it is not
unlimited: certain grammatical patterns (combinations of properties) are cross-
linguistically lacking. Such typological gaps are due to intrinsic properties of
constraint interaction, in particular to strict domination and minimal violation. OT
predicts that typologies are inherently asymmetrical, as opposed to a theory which
assumes a set of logically independent parameters, thereby predicting free com-
bination of their values into symmetrical typologies.

8.4.1 The general typology

This reranking approach to language typology will be applied in this section to a
small set of syntactic constraints. This set consist of the three constraints which
constitute the core of Grimshaws analysis of Wh-movement and inversion: Op-
Spec, Ob-Hd, and Stay. Yet this small set is large enough to illustrate the general
idea and merits of the reranking approach.
Restricting rerankings to a set of three core constraints {Stay, Op-Spec, Ob-
Hd}, we nd six logically possible rankings:

(46) Factorial typology of Wh-movement and inversion

a. Stay
Op-Spec
Ob-Hd
b. Stay
Ob-Hd
Op-Spec
c. Ob-Hd
Stay
Op-Spec
d. Op-Spec
Stay
Ob-Hd
e. Op-Spec
Ob-Hd
Stay
f. Ob-Hd
Op-Spec
Stay

366
 8.4 Typological consequences

Working our way through this typology from top to bottom, we rst arrive at the
rankings in (46a b). Here Stay dominates both well-formedness constraints Op-
Spec and Ob-Hd. When violations of Stay are avoided at the expense of viola-
tions of well-formedness, a grammar arises lacking Wh-movement as well as
inversion. Languages of this type are attested, for example Chinese (Ackema and
Neeleman forthcoming).
Next, ranking (46c), where Stay still dominates Op-Spec, produces a similar
lack of Wh-movement; nor is there movement of auxiliaries into head positions
of extended projections due to Wh-movement. Therefore ranking (46c) leads to
the same inert picture as both rankings in (46ab).
We now move on to ranking (46d), which merits some discussion. Here Op-
Spec dominates Stay, hence Wh-movement is forced. But Stay, in its turn, dom-
inates Ob-Hd, so that inversion cannot be used to ll the head position of
extended functional projections which are the result of Wh-movement. In sum,
this represents a language that has Wh-movement, but not inversion. Such a
language is French (Ackema and Neeleman forthcoming).
Finally, the rankings (46ef) are apparently equivalent since both have Op-
Spec and Ob-Hd on top of Stay. As in English, both Wh-movement and inversion
occur in violation of Stay, to satisfy both top-ranking constraints. However, as
Grimshaw points out, there is a potential empirical difference between the rank-
ings (46e) and (46f) which arises when a fourth constraint comes into play.

8.4.2 Variations on subjectverb inversion

We continue our typology discussion by returning to English one more time, to
make good on a promise. Recall that we observed a perfectly obvious alternative
output candidate for tableau (38) of What did Mary say? the sentence resulting
from inversion on the lexical verb said:

(47) a. *What said Mary?

b. [CP what saidi [VP Mary ei t ]]

We must now consider the cause of its ungrammaticality. This structure seems to
have a lot in its favour. Wh-movement satises Op-Spec, and the projection thus
created (a CP) satises Ob-Hd due to the lexical verb, which has moved out of
the VP. On top of all this, structure (47b) avoids do-support. Then why should it
be ruled out? The answer resides in the undominated constraint No-Lex-Mvt that
was introduced in (16) of section 8.2.4, but has been left out of consideration thus
far because it was not crucial to any other examples which were previously dis-
cussed. No-Lex-Mvt blocks head-movement of a lexical verb out of VP. We may
thus extend tableau (38) with the new candidate (47b = 48e) in the following way:

367
Extensions to syntax

(48)
{say (x, y), x = Mary, y = what, No-Lex- Op- Ob- Full- Stay
tense = past} Mvt Spec Hd Int

a. + [CP what didi [IP Mary ei [VP say t ]]] * **

b. [CP what e [VP Mary said t ]] *! *

c. [CP what e [IP Mary did [VP say t ]]] *! * *

d. [CP what e [IP Mary e [VP said t ]]] *!* *

e. [CP what saidi [VP Mary ei t ]] *! **

When movement of the lexical verb to the head position of CP is blocked, while
this head position must be lled by some verb, then there is nothing better than
to insert a form of do. English thus prefers violations of Full-Int to violations
of No-Lex-Mvt.

Now we return to the typological aspects of this analysis. Surely one would expect
that languages occur in which No-Lex-Mvt takes less priority than it does in
English, and even ranks below Full-Int, so that subjectverb inversion involves
the lexical verb. Such a language is Dutch:19

(49) a. Wat ziet Marie?

b. [CP wat zieti [VP Marie t ei]]

Wh-movement is supported by the movement of the lexical verb to head position

of CP, thus violating No-Lex-Mvt. The following tableau contains the exact
Dutch analogues of candidates occurring in the English tableau (48):

(50)
{say (x, y), x = Marie, y = wat, Op- Ob- Full- Stay No-Lex-
tense = past} Spec Hd Int Mvt

a. [CP wat doeti [IP Marie ei [VP t zien]]] *! **

b. [CP wat e [VP Marie t zien]] *! *
c. [CP wat e [IP Marie doet [VP t zien]]] *! * *
d. [CP wat e [IP Marie e [VP t ziet]]] *!* *
e. + [CP wat zieti [VP Marie t ei]] ** *

19
The verb moves from a VP-nal position, since Dutch has a basic Subject ObjectVerb order.

368
 8.5 Conclusions

Grimshaw points out that languages of this type (which allow movement of lex-
ical verbs) are incompatible with a semantically empty auxiliary: there is always
a better option than (the analogue of) do-support, which is moving the lexical
verb. This conclusion follows naturally from the analysis developed, in which
minimal violation plays a crucial role.

Importantly, the typology does not predict the following situation: do-support in
a sentence containing a Wh-element, but left unaccompanied by Wh-movement.
No possible ranking of constraints will produce this (logically possible) situation,
as will be clear from the following argument. First, for Wh-movement to be
prohibited, it must be the case that Stay dominates Op-Spec. Second, for do-
support to occur, it must be the case that Ob-Hd dominates Full-Int. That is,
do-support is triggered by the requirement that (extended) projections be properly
headed. But an extended projection never occurs without structural necessity, in
particular Wh-movement. But when Wh-movement is absolutely prohibited, this
situation simply cannot arise.

8.5 Conclusions
In this chapter we have demonstrated that OT is a theory of grammar, rather than
a theory of phonology proper, since it is applicable to linguistic phenomena
outside phonology. The empirical domain was syntactic in nature: we considered
interactions between the patterns of Wh-movement, subjectauxiliary inversion,
and do-support in English. It was demonstrated that inversion and do-support
have economical distributions: they occur if and only if the construction has
properties which make inversion and/or do-support necessary. Necessity, stated in
terms of independently motivated terms, was then translated into well-formedness
constraints (such as Op-Spec and Ob-Hd); while syntactic effects (movement,
inversion, do-support, etc.) were translated into faithfulness constraints (such as
Stay, Full-Int, and No-Lex-Mvt). This set of syntactic constraints, in combina-
tion with general principles of constraint interaction (such as strict domination and
minimal violation), sufce to explain the observed syntactic patterns, as well as
the economy of movements and representations.
The observed economy follows as the natural result of constraint interaction.
Movement necessarily involves domination of a syntactic well-formedness con-
straint over a syntactic faithfulness constraint (Stay). Since any movement incurs
a violation of Stay, movement is kept to the bare minimum necessary to obey
higher-ranking well-formedness constraints. Any movement in excess of this min-
imum will serve no purpose in satisfying the higher-ranking constraints, and
therefore incurs an unnecessary violation of Stay. This explanation requires no
369
Extensions to syntax

stipulation that movement be minimal, but simply follows from general OT prin-
ciples: that constraints are violable, but violation must be minimal.
Similarly, the explanation for economy of representation is based on the same
principle of minimal violation. Any (extended) projection postulated in the ana-
lysis of an input is subject to the general well-formedness constraints pertaining
to syntactic structure. To satisfy well-formedness constraints, some movement or
other effect (for example, do-support) may be required. But this, in its turn, incurs
violations of (faithfulness) constraints militating against these effects. The explana-
tion for economy of representation is thus reducible to economy of movement.
In the nal section we have shown that this OT theory of the interactions
of Wh-movement, subjectverb inversion, and do-support produces a factorial
typology which is adequate in terms of what is currently known about syntactic
typology.20

SUGGESTIONS FOR FUR THER READING

Minimalist syntax
Chomsky, Noam (1993) A minimalist program for linguistic theory. In K. Hale
and S. J. Keyser (eds.), The view from Building 20. Cambridge, Mass.: MIT
Press. 4158.
(1995) The minimalist program. Cambridge, Mass.: MIT Press.
Radford, Andrew (1997) Syntax: a minimalist introduction. Cambridge University
Press.

OT syntax
Ackema and Neeleman (forthcoming) Optimal questions. To appear in Natural
Language and Linguistic Theory 16:3. [A prenal version of the paper can
be found in ROA-69, http://ruccs.rutgers.edu/roa.html]
Barbosa, Pilar, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky
(eds.) (1998). Is the best good enough? Optimality and competition in syntax.
Cambridge, Mass.: MIT Press.
Bresnan, Joan (forthcoming) Optimal syntax. To appear in J. Dekkers, F. van der
Leeuw, and J. van de Weijer (eds.), Optimality Theory: phonology, syntax,
and acquisition. Oxford University Press.
Grimshaw, Jane (1997) Projection, heads, and optimality. Linguistic Inquiry 28.
373 422.

20
Other works pursuing issues of syntactic typology in OT are Ackema and Neeleman (forth-
coming) and Legendre, Raymond, and Smolensky (1993).

370
 8.5 8.5
Further
Conclusions
reading

Keer, Edward and Eric Bakovic (1997) Have Faith in syntax. Proceedings
of the West Coast Conference on Formal Linguistics 16. [ROA-200,
http://ruccs.rutgers.edu/roa.html]
Legendre, Graldine, William Raymond, and Paul Smolensky (1993) An Optimality-
Theoretic typology of case and grammatical voice systems. Proceedings of
the Berkeley Linguistics Society 19. 46478.
Mller, Gereon (1997) Partial Wh-movement and Optimality Theory. The Lin-
guistic Review 14. 249 306.
Pesetsky, David (forthcoming) Some Optimality principles of sentence pronun-
ciation. To appear in Barbosa et al.

Syntaxphonology interaction
Golston, Chris (1995) Syntax outranks phonology: evidence from ancient Greek.
Phonology 12. 343 68.

371
 9
Residual issues

9.1 Introduction
The previous eight chapters of this book have offered an overview of Optimality
Theory, with an emphasis on areas in which the theory has proved successful.
This chapter will add an overview of issues that have not yet been successfully
resolved, and developments and modications in the theory that are currently
taking place. The issue that is perhaps most urgently in need of a solution is
opacity, to be discussed in section 9.2. Remaining sections each deal with a major
theme in ongoing research. First, absolute ungrammaticality will be addressed in
section 9.3. Section 9.4 will evaluate strategies to deal with optionality and free
variation. Next, section 9.5 will discuss positional faithfulness in relation to func-
tional approaches to phonology. Finally, section 9.6 will re-evaluate the role of
underlying forms in OT, linking this issue with phonologically driven allomorphy.
Conclusions and a perspective on future developments in OT will be presented in
section 9.7.

9.2 Opacity

9.2.1 Introduction
Opacity refers to the phenomenon that output forms are shaped by generalizations
that are not surface-true. Opaque generalizations lurk at a level deeper than the
output, which becomes apparent by peeling off effects overlaid by other surface-
true generalizations. Opacity is predicted by any theory allowing non-surface
levels of description (the input, or any level mediating between input and output).
However, opacity presents a potential problem for surface-oriented OT, a theory
disallowing reference to preoutput levels by well-formedness constraints. (Of
course, the fact that a generalization is not surface-true is, by itself, not immedi-
ately problematic for OT. Output forms are always in violation of some con-
straints.) Before we look into strategies for dealing with opacity in OT, we will
rst discuss its varieties. That is, opaque generalizations are either non-surface-
apparent or non-surface-true, following McCarthys (1998) terminology.
372
 9.2 Opacity

We call a generalization non-surface-apparent if it takes effect at a level concealed

at the surface. A set of forms undergo a process, although they fail to match its
structural description at the surface. This kind of interaction occurs in Turkish
(Zimmer and Abbott 1978, Kenstowicz and Kisseberth 1979, Sezer 1981, Inkelas
and Orgun 1995).

(1) Opacity of vowel epenthesis in Turkish

a. Vowel epenthesis
/ b-m/ b.m my head
/jel-m/ je.lim my wind
b. Velar deletion
/jk-I/ .j. his foot
/inek-I/ i.ne.i his cow
c. Interaction
/jk-m/ .j.m my foot
/inek-m/ i.ne.im my cow

Vowel epenthesis is motivated by forms such as [ b.m], where a consonant

cluster at the end of a word is broken up by a vowel (whose quality harmonizes
with the stem vowel by a process that is irrelevant here). Next, velar deletion is
apparent from [.j.] where /k/ is deleted in intervocalic position. Finally, con-
sider the interaction of both processes in [.j.m]. While velar deletion applies
transparently (the output contains both vowels that trigger it), epenthesis does
not, since only one of the triggering consonants, namely /m/, surfaces. The output
is opaque with respect to epenthesis, because its context of application is not
recoverable at the surface level. The effect is an overapplication of epenthesis.
(Compare similar overapplications in reduplication and stem-based morphology
chapters 5 and 6.)
This is not an isolated example: many similar cases have been reported (for a
wide variety of other languages), most of which involve well-motivated processes
whose application is otherwise completely general. Our second example is from
Tunica (Haas 1940, Kenstowicz and Kisseberth 1979).

(2) Opacity of vowel harmony in Tunica

a. Vowel harmony
/p-ki / pki she looks
/pC-ki / pCki she emerges
b. Syncope
/ hCpu-uhki/ hCpuhki he dances
/nAi-uhki/ nAuhki he leads (someone)
373
Residual issues

c. Interaction
/ hCpu-ki/ hCpki she dances
/nAi-ki/ nAki she leads (someone)

Vowel harmony copies the colour (backness and rounding) of the rightmost stem
vowel onto the vowel of the 3 sg. f. sufx /-ki/.1 Syncope deletes unstressed
vowels standing immediately before //. The output form [hCpki] she dances
is opaque with respect to vowel harmony, since the stem vowel /u / that triggers
the back and rounded quality of the sufx vowel cannot be recovered from the
surface. (It is indirectly recoverable, of course, precisely because of the sufx
vowels quality.)
A generalization is non-surface-true if it has cases of non-application at the
surface that are controlled by a non-surface level. This is the logically opposite
situation of the previous case. A set of forms fail to undergo a process even
though their surface forms match its structural description. An example comes
from Isthmus Nahuat (Law 1958, Kenstowicz and Kisseberth 1979). A process of
apocope (optionally) deletes a word-nal unstressed vowel. A second process
devoices approximants /l, w, j/ at the end of a word:2

(3) Opacity of approximant devoicing in Isthmus Nahuat

a. Apocope
tmi~tm it ends
b. Approximant devoicing
tjo6 shelled corn
c. Interaction
ikkCli~ikkCl put it in it

The effect is an underapplication of approximant devoicing, in terms of its

output form.
It should now be clear that what distinguishes opaque cases from exceptions
is their systematicity. Opaque generalizations become transparent once we take
into account a level preceding the output. Usually this level is the input, or
underlying form. Taking into account the input level, vowel epenthesis in Turkish,
vowel harmony in Tunica, and approximant devoicing in Isthmus Nahuat are
perfectly transparent. Incidentally, however, opacity involves a level mediating
between the input and the output, as we will see later.

1
Glottal stop // is transparent to harmony.
2
Syllable-nal approximants may also be devoiced, under conditions that are irrelevant here.

374
 9.2 Opacity

Opacity has received much attention in rule-based serial phonology since it offers
a major argument for this theory. Therefore let us rst take a look at opacity from
a serial viewpoint. For a rule to apply, all that matters is whether its structural
context is satised at the point of the derivation at which the rule applies.3 For
this reason serial theory easily captures generalizations that are not surface-true
or not surface-apparent, thus predicting opacity on principled grounds. The two
types of rule interactions that produce opacity are known as counterbleeding and
counterfeeding, respectively (Kiparsky 1973).
Counterbleeding arises when a rules structural context is potentially destroyed
or removed by the application of a prior rule, but the ordering is such that both
rules apply. The second rule, which might have destroyed the context of applica-
tion for the rst rule, applies too late to actually do so. In the Tunica example,
harmony rst copies rounding and backness of the rightmost stem vowel onto the
sufx vowel. This triggering vowel is then deleted by syncope:

(4) Counterbleeding (Tunica and Turkish)

a. / hCpu-ki/ b. /jk-m/
hCpuki Vowel harmony .j.km Vowel epenthesis
hCpki Syncope .j.m Intervocalic k-deletion
[hCpki] [.j.m]

In both examples, a rule (Syncope, Intervocalic k-deletion) deletes the trigger of

another rule (Vowel harmony, Vowel epenthesis), but both deletions apply too
late, missing their chance to bleed the earlier rule. Therefore, in both examples,
both rules apply.
Counterfeeding represents the logically opposite situation. The structural con-
text of a rule is potentially satised due to the application of a prior rule, but the
ordering is such that only one rule applies. The second rule, which might have
created the context of application for the rst rule, applies too early to actually
feed it. Such a counterfeeding order produces opacity of approximant devoicing
in Isthmus Nahuat:

(5) Counterfeeding (Isthmus Nahuat)

/ikkCli /
--- Devoicing (non-applicable)
ikkCl Apocope
[ikkCl]

3
At the same time, this output-blindness is a serious disadvantage of serial theory, as we have
seen in chapter 2, since it frustrates an account of conspiracies.

375
Residual issues

Apocope misses its chance to establish the context for devoicing, which does not
apply.4
Interactions such as those in (45) seem to motivate strongly abstract inter-
mediate levels of representation between input and output. OT recognizes no such
intermediate levels, in line with standard assumptions stated earlier in this book.
In standard OT, the mapping from input (Lexical Representation) to output
(Phonetic Representation) is direct. Moreover, well-formedness constraints state
requirements on output forms only, ignoring the input. This assumption is another
cornerstone of OT, essential to explanation of segment inventories, conspiracies,
and the Duplication Problem (see chapters 12).

Let us rst consider the problem in more detail, focussing on the Turkish case.
The two processes involved, both of which can be independently motivated for
Turkish, are stated as constraint interactions below:

(6) a. Intervocalic k-deletion

*VkV (No intervocalic k)
Max-IO
b. Vowel epenthesis
*Complex, Max-IO
Dep-IO

*VkV and *Complex never conict, and cannot be ranked with respect to one
another. The rankings (6a) and (6b) are integrated in tableau (7), considering
all logically possible combinations of k-deletion and epenthesis, for the input
/jk-m/:

(7) Input: /jk-m/ *Complex *VkV Max-IO Dep-IO

a. .jkm *!
b. .j.km *! *
c. .j.m * *!
d. L .jm *

Candidate [.j.m] (7c) is the actual output, but it is not the most harmonic
candidate. The transparent form [.jm] (7d) has one fewer violation of Dep-IO,
obeying the OT adage do-only-when-necessary: why epenthesize if there is no
4
Kiparsky (1973) proposes the hypothesis that grammars avoid orderings of rules that produce
opacity. That is, the ordering of rules will tend toward a relation that maximizes the transpar-
ency of the rules involved. The cost of opaque interactions clearly resides in the abstractness
they involve, which increases learnability difculties. On the other hand, opacity may also be
functionally motivated as contributing to learnability since it makes lexical representations
recoverable from output representations (Kisseberth 1973, Kaye 1974).

376
 9.2 Opacity

need for it? Note that the case for [.jm] is robust enough to be independent of
constraint ranking. A comparison of the violation patterns of both candidates
shows that violation marks incurred by (7d) form a proper subset of those
incurred by (7c). This nding destroys all hope that the problem is solvable
without the help of additional constraints (or even new theoretical machinery).

The underapplication type of opacity, as occurs in Isthmus Nahuat, is equally

problematic. Each of the interacting processes, apocope (3a) and devoicing (3b),
involves domination of a well-formedness constraint over a faithfulness constraint:

(8) a. Apocope: Final-C (Stem ends in C)
Max-IO

b. Devoicing: *Voiced-Coda
*Ident-IO(voice)

But no matter how we integrate the subrankings into a total ranking, the opaque
candidate (9b) will never become optimal. Instead the transparent candidate (9c)
is predicted:

(9) Input: /ikkCli/ Final-C Max-IO *Voiced- Ident-IO

Coda (voice)

a. ikkCli *!
b. ikkCl * *!
c. L ikkC6 * *

Opaque (9b) and transparent (9c) only differ in their violation marks for *Voiced-
Coda and Ident-IO(voice). However, the mutual ranking of these constraints is
forced, since it is independently motivated by approximant devoicing in words
such as [tjo6] shelled corn (3b). The incorrect prediction crucially depends on
constraint ranking (rather than being independent of it, as in overapplication).
This is a clear ranking paradox.

Opacity appears to be a direct empirical refutation of the surface-based evalua-

tion of well-formedness constraints in OT. Since opacity is OTs Achilless heel,
researchers have attempted to nd solutions for it which maximally preserve the
theorys advantages. This section will present an overview of these attempts,
reecting on their relative merits. OT approaches to opacity can be put under
various headings, all discussed below:

Two-level well-formedness (section 9.2.2)

Intermediate levels (section 9.2.3)
OO-correspondence (section 9.2.4)
377
Residual issues

Sympathy (section 9.2.5)

Local Conjunction (section 9.2.6)

9.2.2 Two-level well-formedness

One approach to opacity allows reference to the input by well-formedness con-
straints, in a more or less direct way. By giving up the assumption that well-
formedness constraints have access to output forms only, this approach preserves
direct mapping of input to output, avoiding intermediate levels. Simultaneous
reference to input and output by well-formedness constraints originates in pre-OT
work by Koskenniemi (1983) and Karttunen (1993). We will refer to this idea as
two-level well-formedness. Note that Correspondence Theory is a two-level model
in a more restricted way, since only faithfulness constraints (and not well-formed-
ness constraints) can refer to the input and output simultaneously.
A conceptually related OT model is the Parse/Fill theory of Prince and
Smolensky (1993), also known as Containment Theory, briey discussed in chap-
ter 3 as a precursor of Correspondence Theory. This model is named after its key
assumption that no element may be literally removed from the input, which is thus
contained in all candidate outputs. Deletion of an element is represented by leav-
ing it prosodically unparsed, resulting in the lack of phonetic interpretation. Since
deleted input elements are contained in the output representation, they can be
phonologically active.5 In Correspondence Theory, reference to the input by well-
formedness constraints has been implemented by Cole and Kisseberth (1995),
McCarthy (1995b), Orgun (1995), and Archangeli and Suzuki (1997).

To illustrate the basic idea, let us consider the Tunica example. Opacity of vowel
harmony is accounted for if harmony is triggered by input vowels, but takes effect
in the output. This can be achieved by a well-formedness constraint that refers to
the input and output simultaneously, for example:6

(10) Harmony-IO
If input V1 . . . V2 then V1 and V2 agree in backness and rounding.

output V2

5
In its original form, Containment Theory was never used as a framework for analysing opacity.
Yet the idea that input features are abstractly represented in the output has been elaborated by
Cole and Kisseberth (1995) in Optimal Domains Theory. The innovative aspect of this theory
is its assumption that input features occur in the output in special constituents, or F(eature)-
domains, which largely coincide with (autosegmental) harmony domains. Relations between
F-domains and input features are enforced by violable faithfulness constraints.
6
A complete analysis should account for the joint reference to backness and rounding (colour),
and also explain the directionality of harmony: rightward, or more likely, stem to afx.

378
 9.2 Opacity

Harmony-IO is a two-level well-formedness constraint, stating a requirement of

featural agreement between segments at different levels of representation, input
and output. It can be read as follows:

(11) If V1 . . . V2 is a sequence of vowels in the input, and V2 is the corres-

pondent of V2 in the output, then V1 and V2 agree in backness and
rounding.

Featural harmony between V1 and V2 can be said to be transmitted by the corre-

spondence relation between V2 and V2. Note that its dependence on correspond-
ence does not give Harmony-IO the status of an IO-faithfulness constraint. The
output vowel is required to resemble another vowels input, rather than its own.
This is a well-formedness target, not a faithfulness target.7
Two-level well-formedness constraints have been extended to express restric-
tions on the cooccurrence of different features. This is exemplied by a constraint
proposed by Archangeli and Suzuki (1997):

(12) Lowering-IO
Any output correspondent of an input long vowel must be [high].

Lowering has the format of a correspondence constraint, but with one crucial
difference: correspondents in the input and output are not required to agree with
respect to the same feature (as in Ident-IO[F]). Like (12), Lowering-IO is not
an IO-faithfulness constraint, since feature-cooccurrence is required for different
features (here, length and height), and for unrelated values. In Archangeli and
Suzukis analysis of Yokuts, Lowering triggers a featural change from the input
value of [high]. Therefore, it functions as a well-formedness constraint, although
it has the IO-format of standard correspondence constraints. A theory allowing
for constraints such as (10) and (12) essentially gives up the distinction between
well-formedness constraints and faithfulness constraints.

Let us now evaluate two-level well-formedness models. The rst question to be

asked is: Do such models account for opacity?, and the second question: If so,
at what theoretical cost? We will spot some serious empirical and theoretical
problems.
Empirically, two-level well-formedness models fail to account for types of
opacity that are not controlled by the input, nor by the output. An opaque gener-
alization may hold at an intermediate level, for example with respect to prosodic

7
Nevertheless, Harmony-IO interacts with IO-faithfulness constraints in a standard way it
dominates Ident-IO(round) and Ident-IO( back).

379
Residual issues

structure. (So far we have not seen such cases.) For example, an opaque gen-
eralization referring to prosody cannot be analysed by reference to the input,
assuming that prosody is not present in input forms.
A well-studied case is compensatory lengthening (CL), a situation where a
vowel is lengthened to compensate for the quantity of a deleted consonant
(Hayes 1989). An example comes from Oromo (Lloret 1988, Sprouse 1997):

(13) a.i /fe-na/ feena we wish

a.ii /fe-sisa/ feesisa I make wish
b.i /fe-a/ fea I wish
b.ii /fe-aa/ feaa wish for self

A serial derivation shows the origin of the long stem vowel in [feena] we wish.
A mora is assigned to the coda consonant closing the rst syllable (by Weight-
by-Position, chapter 4). This consonant is then deleted before a consonant-initial
sufx, but its mora is preserved in the form of an extra element of length on the
preceding vowel. Opacity is explained by a counterbleeding rule order (Weight-
by-Position Pre-consonantal deletion):

(14)

/f e-na/ f e na f e na f e na

As Sprouse (1997) points out, the mora that triggers the lengthening cannot be
part of the input. If faithfulness to input moras were to trigger CL, and the
consonant // be specied as moraic in the input, then CL should occur in any
context in which // cannot hold on to its mora. This is incorrect; see *[feea]
(13b), where // surfaces as an onset. Instead CL is restricted to the context of
Weight-by-Position. Moreover, if input consonants were contrastively specied as
moraic (so that weight of CVC syllables would lexically vary), then certain stems
would undergo CL, while other stems would simply undergo deletion. Such con-
trasts are apparently not found in Oromo. Finally, as McCarthy (1998) points out,
input specication of moras runs into a general problem: it is at odds with Rich-
ness of the Base, a cornerstone of OT. In sum, the length-triggering mora cannot
be part of the input, ruling out a two-level analysis of CL. The best a two-level
analysis can achieve is to encode the context of Weight-by-Position into a two-
level constraint:

(15) -Projection
Project an output for every input C followed by a C.
380
 9.2 Opacity

Of course, this is a bad constraint. Sprouse (1997: 2) notes it is more of a rule,

combining a well-formedness constraint (how syllables should be organized) with
its repair ( project a mora) [ . . . ] even when the well-formedness constraint cant
be evaluated directly since the mora and segment that projects it are never present
at the same level. Finally, -Projection loses the generalization that a mora is
projected for every coda consonant. It stipulates the opaque pattern, rather than
explaining it.
The theoretical status of both earlier two-level constraints (10) and (12) is
equally problematic. Both function as rules, combining a structural condition (the
input structure) and a repair. A theory allowing for two-level well-formedness
constraints may stipulate any type of relation between the input and output, being
equivalent in this respect to rule-based theory (Lakoff 1993). This power under-
mines standard OTs solutions to problems inherent to rule-based serialism, in
particular conspiracies and the Duplication Problem. Recall from chapter 2 that
standard OT improves over rule-based theory by separating structural conditions
(triggers) and structural changes (repairs), thus predicting a range of repair strat-
egies in response to a single well-formedness requirement. However, reference to
the input by well-formedness constraints implies a crucial deviation from stand-
ard OT. Triggers and repairs become once more entangled, hence the explanation
of conspiracies and the Duplication Problem is fatally undermined. As Sprouse
(1997) observes, blurring the borderline between well-formedness constraints and
faithfulness constraints produces negative consequences for factorial typology as
well.
Finally, we may ask to what extent two-level well-formedness constraints can
still be considered to be grounded. Each of the constraints Harmony-IO and
Lowering-IO, interpreted as a single-level output constraint, makes a phonetic-
ally natural requirement. (Harmony reduces the number of articulatory gestures,
and lowering supports the quantity contrast by matching it with the longer intrin-
sic duration of non-high vowels.) However, phonetic grounding does not carry
over to the two-level versions of the constraints since the goals are not manifest
in the output. This mismatch is, of course, inherent to opacity. But explanation of
opacity will be lost if it is hard-wired into the constraints, rather than reduced
to interactions of grounded constraints.

9.2.3 Intermediate levels

Yet another approach to opacity is based on intermediate levels between input and
output. The idea behind this move should be clear: the success of serial theory in
accounting for opacity was based precisely on pre-surface levels. Although this
gives up the standard OT assumption of direct mapping, there is also a potential
pay off: preserving the standard constraint format. This offers a reverse picture as
381
Residual issues

compared to two-level well-formedness discussed in section 9.2.2, which gave up

constraint format while preserving direct mapping.
Let us assume an OT grammar which is internally organized into serially
ordered strata (Stratum 1n, with n 2), each containing both functions Gen
and Eval. The input to each stratum is dened by the output of the previous one,
except that the input to Stratum 1 is identical to the lexical representation. Within
each stratum, the input output mapping is direct. Each stratum contains its own
constraint ranking:

(16) Multi-stratal evaluation

Input

Stratum 1 Gen1 Eval1

...

Stratum n Genn Evaln

Output

Multi-stratal models have their origins in derivational Lexical Phonology (Kiparsky

1982b, Mohanan 1982, Booij and Rubach 1987), where they accounted for a range
of phenomena, such as word domain effects, afx ordering, structure preserva-
tion, and cyclicity. With the emergence of constraint-based theories, direct map-
ping became the default assumption. Nevertheless, constraint-based multi-stratal
models were proposed, by Goldsmith (1993a, Harmonic Phonology) and in OT
by, among others, McCarthy and Prince (1993b), Inkelas and Orgun (1995), and
Sprouse (1997).

In multi-stratal OT the generation of output forms essentially follows a derivation,

as in serial theory. This is illustrated below for Turkish, a case of overapplying
opacity, involving counterbleeding in serial theory. This grammar contains two
strata. Stratum 1 takes the lexical representation /jk-m/ as its input, and produces
an intermediate output [.j.km], where epenthesis has applied but intervocalic
/k/ is preserved.

(17) Overapplication in multi-stratal evaluation (Turkish)

Stratum 1: I1 /jk-m/
*Complex, Max-IO
Dep-IO, *VkV
O1 [.j.km]
382
 9.2 Opacity

Stratum 2: I2 /.j.km/
*Complex, *VkV
Dep-IO, Max-IO
O2 [.j.m]

The output of S1, [.j.km], is due to the ranking *Complex, Max-IO
Dep-
IO, *VkV:

(18) Input: /jk-m/ *Complex Max-IO Dep-IO *VkV

a. .jkm *!
b. + .j.km * *
c. .j.m *! *
d. .jm *!

This S1 output is then fed as an input into S2. Ranking at S2 is such that intervo-
calic /k/ is deleted (*VkV
Max-IO), giving an S2 output [.j.m], which is
the surface form.

(19) Input: /.j.km/ *Complex *VkV Dep-IO Max-IO

a. .jkm *! *
b. .j.km *!
c. + .j.m *
d. .jm **!

In sum, both strata have minimally different rankings, involving only reranking
of a well-formedness constraint and a faithfulness constraint.8
Let us now briey discuss a case of underapplication, which is captured
in serial theory by counterfeeding. Approximant devoicing in Isthmus Nahuat is
opaque, as shown earlier in this chapter. Here intermediate outputs are [ ikkili]
and [tajo6], respectively:

(20) Underapplication in multi-stratal evaluation (Isthmus Nahuat)

Stratum 1: I1 /ikkili/ /tajol/ Max-IO
Final-C
*Voiced-Coda
Ident-IO(voice)
O1 [ikkili] [tajo6]
8
It is difcult to establish the ranking of Dep-IO at S2. We may assume it to be undominated.
It marks no violation for epenthetic [], because this vowel occurs in the input of Stratum 2.

383
Residual issues

Stratum 2: I2 /ikkili/ /tajo6/ Final-C
Max-IO

Ident-IO(voice)
*Voiced-Coda
O2 [ikkil] [tajo6]

At Stratum 1, apocope is blocked (Max-IO
Final-C), while nal approximants

become devoiced (*Voiced-Coda
Ident-IO(voice)). This is shown in two
tableaux:

(21) Input: /ikkili/ Max-IO *Voiced- Final-C Ident-IO

Coda (voice)

a. + ikkili *
b. ikkil *! *
c. ikki6 *! *

(22) Input: /tajol/ Max-IO *Voiced- Final-C Ident-IO

Coda (voice)

a. tajol *!
b. + tajo6 *

The reverse rankings hold at Stratum 2, where apocope applies (Final-C
Max-
IO), but devoicing is blocked (Ident-IO(voice)
*Voiced-Coda):

(23) Input: /ikkili/ Final-C Ident-IO Max-IO *Voiced-

(voice) Coda

a. ikkili *!
b. + ikkil * *
c. ikki6 *! *

(24) Input: /tajo6/ Final-C Ident-IO Max-IO *Voiced-

(voice) Coda

a. tajol *! *
b. + tajo6

384
 9.2 Opacity

Consequently each form undergoes only one process, either devoicing or apo-
cope, but not both. ( Note that values for [voice] as these occur in the output of
Stratum 1 are preserved at Stratum 2, regardless of their values in the original
lexical input.)
In contrast to two-level well-formedness theory, an intermediate-levels the-
ory of opacity holds the promise of accounting for types of opacity that are not
controlled by the input. Compensatory Lengthening in Oromo (13), for example,
may be accounted for by an intermediate level at which Weight-by-Position has
applied, but consonant deletion is still blocked. CL becomes an identity effect
preserving quantity occurring at intermediate levels. In sum, multi-stratal evalu-
ation has the advantage of a broad empirical coverage of opacity effects, being
equivalent to serial theory in this respect.
But does it preserve OTs advantages over serial theories, which are largely
based on direct mapping and evaluation of output forms? If OT is not to be
reduced to a notational variant of serial theory, one major question should be
answered: Can we have the best of both worlds? That is, can intermediate levels
be restricted in a way that captures opacity, but also preserves OTs advantages?
Can the number of strata be restricted to a universal maximum? At present, it is
not clear what the ultimate answers to these questions will be, due to a lack of
experience with multi-stratal models. We foresee two problems, however.
Firstly, the problem of nding independent motivation for strata. Motivation
must be independent in the sense that it goes beyond the desire to capture opaque
interactions. If such (presumably morphological) evidence cannot be produced,
then multi-stratal OT is reduced to a variant of serial theory, differing in having
two mechanisms (constraint interaction and derivation) instead of one (derivation).
Considerations of generality of explanation would then favour the serial model.
Secondly, multi-stratal evaluation implies that multiple rankings are active
within a grammar. This predicts that stratal rankings (hypothetically, at least) can
differ as widely as those of different languages. But cases under discussion show
that such radically different rankings for strata do not occur. On the contrary,
rankings differ only in minimal ways, typically by the reranking of a pair of
constraints (a well-formedness constraint and a faithfulness constraint), or per-
haps two pairs, but not by massive rerankings.
Thirdly, multi-stratal models are not learnable by Tesar and Smolenskys algo-
rithm, discussed in chapter 8, and it is doubtful whether they are learnable at
all, due to the large increase in complexity. However, future research will have to
answer this question.

385
Residual issues

9.2.4 OO-correspondence
A third approach to opacity is based on OO-correspondence (McCarthy 1995a;
Benua 1995, chapter 6). Opacity involves underapplication and overapplication
effects, as noted in section 9.2.1. Hence, it seems logical to extend the apparatus
of OO-identity constraints to opacity. If successful, this approach would have the
advantage of preserving two core principles of standard OT: direct mapping and
strictly output-based well-formedness constraints.
An identity-based approach to opacity requires the following pair of conditions.
First, for each opaque output, there must be an output form compositionally
related to it. Second, this base must be transparent with respect to the generali-
zation. Both conditions hold for the analysis of i-Syncope in Palestinian Arabic
(chapter 6, section 6.4).
i-Syncope normally deletes unstressed /i/ in open syllables (25b). It under-
applies in (25c), where unstressed /i/ corresponds with a stressed vowel in the base.

(25) Regular application and underapplication of i-Syncope in Palestinian

Arabic
a. /him/ fhim he understood
b. /fihim-na/ fhmna we understood
c. /[f him]-na/ fihmna *fhmna he understood us

Two-level well-formedness fails here since the abstract stress blocking application
of i-Syncope (rendering it opaque in 25c) is not present in the input.
However, an OO-correspondence analysis works, because the blocking stress
is recoverable from the base. A BA-identity constraint HeadMax-BA refers to it:

(26) HeadMax-BA
Every segment in the bases prosodic head has a correspondent in the
afxed form.

OO-correspondence reanalyses any cases of opacity accounted for by cyclic appli-

cation in serial theory. Yet it fails wherever opacity is controlled by an abstract
intermediate level, one that is not recoverable from a morphologically related
output form.
For example, vowel epenthesis in Turkish (/jk-m/ [.j.m], section 9.2.1)
cannot be analysed by OO-correspondence. Some form should exist (in the para-
digm of /jk/) that has a transparent epenthetic [] following the stem. But no
such output form can exist, given the phonology of Turkish. If epenthesis applied
transparently in any form related to /jk/, then /k/ would have to resist deletion
in an intervocalic context. However, this is precisely what *VkV
Max-IO
excludes. We must conclude that OO-correspondence is insufciently general to
account for all types of opacity.
386
 9.2 Opacity

9.2.5 Sympathy
The most recent attack on opacity is due to McCarthy (1998), and it is called
Sympathy. Its core feature is an extension of the correspondence relation to
pairs of candidate forms. Faithfulness constraints require the output form to re-
semble another candidate of Gen, the sympathetic form, which is analogous to the
abstract intermediate representation in serial theory. Sympathy preserves the key
features of standard OT: direct mapping of inputs to outputs and the output-based
format of well-formedness constraints. However, it implies a vast increase in the
correspondence relations involved in selecting the optimal output.

To acquire an intuitive notion of Sympathy, we return to tableau (7) of Turkish

/jk-m/ (repeated below as 27). Recall that the transparent candidate (27d) is
incorrectly selected as optimal over opaque (27c). This result is independent of
constraint ranking, as violation marks incurred by the former are a proper subset
of those incurred by the latter:

(27) Input: /jk-m/ *Complex *VkV Max-IO Dep-IO

a. .jkm *!
b. .j.km *! *
c. .j.m * *!
d. L .jm *

For [.j.m] to become more harmonic than *[.jm], some constraint C

must outrank Dep-IO, one which [.j.m] obeys but *[.jm] violates. The
unknown constraint is probably not a well-formedness constraint, since [.j.m]
can boast no advantages over *[.jm] in terms of its featural or syllabic compo-
sition. (Its extra [] only works against it in terms of well-formedness, causing
hiatus, and an onset-less syllable.) Therefore constraint C must be a faithfulness
constraint.
To what form can [.j.m] be faithful which [.jm] is unfaithful to? It can-
not be the input /jk-m/ for obvious reasons (IO-faithfulness fails, which is
precisely the problem). Nor is any transparent output form available for OO-
correspondence (as we saw in section 9.2.4). McCarthy opts for a third logical
possibility: the opaque candidate output is faithful to a form that is neither an
input nor an output, but a candidate output itself. That is, he proposes a new type
of correspondence relation dened on pairs of output candidates: this relation is
called Sympathy.
387
Residual issues

Let us rst consider an intuitive (and informal) way of picking the sympathetic
candidate to which [.j.m] is faithful. Focussing on the candidate pair [.j.m]
and [.jm], we nd that whatever advantage of faithfulness the former may have,
this must reside in its epenthetic []. We deduce that the yet-to-be-identied sym-
pathetic candidate shares this vowel with [.j.m]. Such a candidate form is
[.j.km] (27b). This happens to be the form in tableau (27) that is otherwise
maximally faithful to the input this observation will be taken up directly below.
This sympathetic form will be indicated by the ower {.
The {-candidate [.j.km] matches the intermediate form in the serial ana-
lysis (4), arising after vowel epenthesis (the input to intervocalic k-deletion). In
the OT analysis, [.j.km] is more faithful to the input /jk-m/ than opaque
[.j.m], taking an intermediate position between both. Thus the serial analysis
and OT analysis share an insight: both set up an abstract intermediate form
connecting the opaque output to the input.9

The next step is setting up a correspondence constraint demanding faithfulness to

the {-candidate [.j.km]. We need a constraint militating against deletion of
segments, hence a member of the Maximality family. This constraint, Max-{O,
is ranked above Dep-IO, the anti-epenthesis constraint, violation of which would
have been fatal to (28c) (as indicated by ) if Max-{O had not been there:

(28)
Input: /jk-m/ *Complex *VkV Max-{O Max-IO Dep-IO
{-Candidate: .j.km

a. .jkm *! *
b. { .j.km *! *
c. + .j.m * * *
d. .jm **! *

The opaque candidate (28c), which would normally lose to the transparent
candidate (28d) because of its extra violation of Dep-IO, now becomes optimal
due to its maximal faithfulness to the {-candidate. Note that one violation of
Max-{O is incurred by each segment of the {-candidate that lacks a correspond-
ent in the output. Therefore, opaque (28c) is unfaithful to the {-candidate
by one segment (deleted /k/), whereas transparent (28d) is unfaithful by two

9
McCarthy points out that such parallelism is not always exact, particularly in cases involving
multiply opaque generalizations.

388
 9.2 Opacity

segments (deleted /k / and //), with the additional violation being fatal. Finally,
the {-candidate itself fatally violates *VkV.10

Next we turn to a case of overapplying opacity that is not controlled by the input
(which is problematic to reference-to-the-input models). Again, the {-candidate
matches the intermediate representation in a serial analysis, which arises after
Weight-by-Position and is the input to consonant deletion. This is [fe.na]
(29b). The opaque output (29c) is faithful to it by satisfying Max--{O, a con-
straint demanding that each mora of the {-candidate have a correspondent in the
output.
(29)
Input: /fe-na/ *Weight-by- Coda- Max-- Max-IO Dep-
{-Candidate: fe.na Position Cond {O -IO

a. fe.na *! * *
b. { fe.na *! *
c. + fee.na * *
d. fe.na *! *

Finally, consider the case of underapplying opacity in Isthmus Nahuat. In

a serial (counterfeeding) derivation (5) of [ikkil] (from /ikkili /), an inter-
mediate representation occurs in which devoicing applies, but apocope has not yet
applied. Here /ikkili / fails to undergo devoicing since its nal vowel is still
present. Analogously, the {-candidate in tableau (30) happens to be identical to
the input:
(30)
Input: /ikkili / Final-C Ident-{O Max- *Voiced- Ident-IO
{-Candidate: ikkili (voice) IO Coda (voice)

a. { ikkili *!
b. + ikkil * *
c. ikki6 *! * *

10
Another plausible output candidate, [j.km], is equally faithful to the sympathetic candidate.
The difference is that an input vowel is deleted in [j.km] but an input consonant in [.j.m]
(the actual output). We assume that this contrast correlates with a distinction between Max-
C and (higher-ranking) Max-V. (See the analysis of Southeastern Tepehuan in chapter 4.) This
assumption is shared by McCarthy (1998), who does not discuss the Turkish case, though.

389
Residual issues

Candidate (30a), perfectly faithful to the input and to itself (in its quality of {-
candidate), is nevertheless rejected for its fatal violation of Final-C. Opaque
(30b) is optimal due to Ident-{O(voice)
*Voiced-Coda, ruling out trans-
parent (30c).
Observe that this tableau contains two faithfulness constraints of the format
Ident(voice), a low-ranking one evaluating IO pairs (Ident-IO(voice)), and a
top-ranking one evaluating {-O pairs (Ident-{O(voice)). Since the {-candidate
happens to be identical to the input, both faithfulness constraints produce the
same evaluation marks, at different positions in the hierarchy. We will return to
this observation shortly below, in connection with the issue of how the choice of
the {-candidate is made.

The {-candidate is not a priori given, and it should be determinable on the basis
of positive evidence. How can this aim be achieved? In all cases of opacity
discussed thus far, the {-candidate satises an IO-faithfulness constraint that is
violated in the opaque form. In all three cases, this is Max-IO:

(31) {-candidate opaque form faithfulness violation

a. Turkish .j.km .j.m /k/ (*Max-IO)
b. Oromo fe.na fee.na // (*Max-IO)
c. Nahuat ikkili ikkil /i/ (*Max-IO)

We noted in section 9.2.1 that opaque generalizations are overlaid by other gen-
eralizations of the language. Now we see how the {-candidate is related to the
opaque form by peeling off a structural change (here, undoing a segment
deletion). The relation between the opaque form and the {-candidate is such that
the latter obeys an IO-faithfulness constraint that is violated by the former. This
special IO-faithfulness constraint is called the selector.
McCarthy claims that the choice of the {-candidate is uniquely determined by
the selector, given the (independently motivated) constraint ranking of the lan-
guage. That is, the {-candidate is the most harmonic candidate of all candidates
that satisfy the selector. For this purpose, we may think of candidate space as
divided into two mutually exclusive subsets, one of candidates that satisfy the
selector, and another of candidates violating it. Within the former subset, the {-
candidate is the optimal candidate, given the constraint hierarchy of the language.

For Turkish this method is successful, given a number of independently motivated

observations about constraint ranking. *Complex and Anchoring-IO (militating
against nal epenthesis) appear to be never violated in Turkish, hence they are
undominated. But *VkV is actually violable in paradigms of monosyllabic roots
(for example, ok arrow, ok-u his arrow, ok-um my arrow). That is, no mono-
390
 9.2 Opacity

syllabic root ending in /k/ loses /k/ in intervocalic contexts. Therefore *VkV must
be crucially dominated by some constraint enforcing root-minimality, whose
exact nature is irrelevant to our discussion.
The dominated position of *VkV is conrmed by the selection of the {-
candidate [.j.km]. Tableau (32) shows four candidates which obey the selector
constraint Max-IO (omitting candidates that violate this constraint, which do not
qualify by denition). The independently motivated constraint ranking indeed
selects (32c) as the {-candidate:

(32) Subtableau for selection of the {-candidate

Input: /jk-m/ *Complex Anchoring-IO *VkV Max-IO Dep-IO

a. .jkm *! 3
b. .jk.m *! 3 *
c. { .j.km * 3 *
d. .j.k.m *! * 3 **

This is not the tableau selecting the actual output form but only the part that is
relevant to the selection of the {-candidate. In a fully edged tableau, the choice
of the {-candidate is actually made in parallel with the choice of the optimal
output form. That is, (32) is a subtableau of (28). Note that tableau (32), unlike
(28), does not include Max-{O, the Sympathy constraint. McCarthy argues that
the Sympathy constraint must indeed always be invisible to the selection of the
{-candidate, to avoid circularity in selection.
The choice of the selector is critical, since this determines the {-candidate,
hence indirectly the choice of the opaque output. But how is the selector itself
being chosen? The selector is always an IO-faithfulness constraint, violable in
actual output forms of the language (crucially so, in the opaque output itself).
Since most faithfulness constraints of a language are undominated, this narrows
down its choice. But the bottom line is that any positive evidence regarding the
choice of the selector can only reside in the opaque form itself. The learner must
infer the choice of the selector from the opaque form. This may seem an unsatis-
factory conclusion which has a whiff of circularity. However, an analogous conclu-
sion holds for a serial theory of opacity: any evidence pertaining to rule ordering
(bleeding or counterbleeding, feeding or counterfeeding) can only come from the
opaque form itself.

Let us now briey discuss this proposal, starting with its virtues. First, Sympathy
preserves two cornerstones of standard OT direct mapping and the output-based
391
Residual issues

format of well-formedness constraints. All interactions are parallel rather than

serial, involving ranked constraints rather than ordered levels. Well-formedness
constraints never refer to levels other than the output (either inputs or abstract
intermediate representations). In its empirical scope, Sympathy surpasses all OT
approaches to opacity seen thus far except Intermediate Levels theory. However,
it also has a number of problematic aspects.
First it weakens Correspondence Theory by extending it to candidate-to-
candidate faithfulness. Standard correspondence constraints require identity with
an independently established, concrete, form, either a lexical input (IO-faithful-
ness), or an actual output form (OO-identity). Increasing the scope of correspond-
ence constraints poses a potential threat to OTs restrictiveness. Faithfulness to
abstract forms implies a radical increase in the number of constraint interactions,
hence in computational complexity of the theory. A related issue, one which
cannot be addressed without additional research, is learnability. Do the results
that we discussed in chapter 7 carry over to Sympathy?
Second, it is not clear at all whether the {-candidate is uniquely determinable
by the selector on the basis of only the independently motivated hierarchy of the
language, as McCarthy suggests. Rankings of undominated constraints, which
are not independently motivated by output forms, may become relevant to the
selection of the {-candidate. Note that the set of candidates obeying the selector
always contains candidates violating one or more undominated constraints. (In
the Turkish example, the candidate set includes a candidate [.jkm], violating
*Complex.) What happens if, in this set of selector-obeying candidates, no can-
didates occur obeying all undominated constraints? Such a situation may occur if
the selector imposes a requirement that implies the violation of an otherwise
undominated constraint in the {-candidate. In such a case, the ranking of undom-
inated constraints is required to break the tie. But the use of ranked undominated
constraints contradicts the claim that the independently motivated ranking of the
language sufces to choose the {-candidate. By denition, undominated con-
straints are not rankable on the basis of evidence from transparent outputs, in
which they are never violated.
Third, Sympathy does not offer a general theory of opacity, as it cannot deal
with chain shifts, a kind of counterfeeding opacity that will be discussed in the
next subsection.

9.2.6 Local Conjunction

A nal approach to opacity in OT merits attention, partly because it is com-
plementary to Sympathy, and partly for independent reasons. Under Local Con-
junction, two constraints are conjoined as a single composite constraint which is
violated if and only if both of its components are violated within some domain.
392
 9.2 Opacity

The proposal is due to unpublished work by Smolensky (1993). Consider the

schematic tableau (33), which contains four candidates representing all logically
possible combinations of violations of constraints C1 and C2:

(33) C1 C2 [C1 & C2]

Candidate 1
Candidate 2 *
Candidate 3 *
Candidate 4 * * *

The complex constraint [C1 & C2] (the conjunction of constraints C1 and C2) is
violated if and only if both of its components are violated. Moreover, for a viola-
tion of [C1 & C2] to occur, both separate violations must arise within a single
domain (a segment, morpheme, etc.). Evidently some domain is needed for
conjunction: the severity of output ill-formedness is never increased by combina-
tions of violations in random positions in the output. Finally, a conjoined con-
straint does not replace its components, but it is separately ranked. It is generally
assumed that a conjoined constraint is universally ranked above the component
constraints.

(34) Universal ranking schema: [C1 & C2]
C1, C2

We will rst look at evidence for conjunction from opacity (chain shifts), and then
at its wider applications.

The clearest motivation for Local Conjunction of constraints resides in chain

shifts (Kirchner 1996). A chain shift is a situation in which sounds are promoted
(or demoted) stepwise along some scale in some context. Crucially the chain shift
does not result in neutralization, since each input takes precisely one step. This is
schematized as follows:

(35) A B and B C, but not *A C

That is, A reoccupies the position left vacant by B, which itself occupies C,
etc.
A striking example of a chain shift is vowel raising in Western Basque (De Rijk
1970, Kenstowicz and Kisseberth 1979, Kirchner 1996). Mid vowels and high
vowels that precede another vowel are raised by one degree (mid to high and high
to raised):
393
Residual issues

(36) Indenite Denite

a. Mid to high ei seme bat semie son
ou asto bat astue donkey
b. High to raised i i j erri bet errije village
u uw buru bet buruwe head

Kirchner assumes the following specications for vowels for [low], [high], and
[raised]:

(37) Raised {ij, uw} [low, +high, +raised]

High {i, u} [low, +high, raised]
Mid {e, o} [low, high, raised]
Low {a} [+low, high, raised]

(In the Etxarri dialect, on which Kirchner focusses, low vowels are not raised in
hiatus. The raising of clitic vowels /bat/ [bet] is conditioned by a high vowel
in the stem, and we will ignore it.) First we will see that chain shifts are not
problematic to serial theory, while they do pose problems to standard OT.
In serial theory the analysis is straightforward: it is a case of counterfeed-
ing on the focus. Two counterfeeding rules raise high vowels and mid vowels,
respectively:11

(38) Counterfeeding analysis of chain shift

/seme-a/ /erri-a/
High vowel raising --- errija
Mid vowel raising semia ---
[semie] [errije]

Each of the counterfeeding rules applies only once per derivation.

Chain shifts, a kind of underapplying opacity, cannot be analysed under stand-
ard OT, as we will now show. Raising in hiatus is triggered by a well-formedness
constraint (Kirchner 1996):

(39) Hiatus-Raising
In V1 V2, maximize height of V1.

Kirchner argues that this constraint is evaluated gradiently, and that different
values of V1 incur the following violations marks: low ***, mid **, high *,
and raised none. The IO-faithfulness constraints involved are:

11
A third rule raises the sufx vowel from /-a / to [-e] if it is preceded by a high vowel.

394
 9.2 Opacity

(40) a. Ident-IO(high)
If an input segment is [high], then its output correspondent is
[high].
b. Ident-IO(raised)
If an input segment is [raised], then its output correspondent is
[raised].

For raising to take place at all, Hiatus-Raising must be ranked above both
faithfulness constraints. That is, e i shows that Ident-IO(high) is dominated,
while i ij shows that Ident-IO(raised) is dominated. This is summarized in the
following tableau (which, however, gives only a subset of the output candidates
to be nally considered for /e/):

(41) Hiatus- Ident-IO Ident-IO

Raising (high) (raised)

a.i ee *!*
a.ii + e i * *

b.i ii *!
b.ii + i ij *

Note, however, what happens if we add a third output candidate, [ij], to the tableau
of /e/. This ranking incorrectly predicts that /e/ is raised to [ij], going two steps
rather than one:

(41) Hiatus- Ident-IO Ident-IO

Raising (high) (raised)

c.i ee *!*
c.ii ei *! *
c.iii L e ij * *

The incorrect prediction is due to undominated Hiatus-Raising. No pos-

sible ranking of these three constraints allows i ij, and e i, while disallowing
e ij. Mid vowels simply cannot be prevented from going all the way to
raised.

395
Residual issues

The reasoning behind Kirchners Local Conjunction analysis is this: the change
e ij involves violation of two faithfulness constraints, while each of the indi-
vidual steps i ij and e i involves only one violation. Moreover, the faithful-
ness violations incurred by i ij reoccur separately in the individual steps i
ij and e i. Then what is needed is the conjunction of both faithfulness con-
straints into a composite constraint. This must be ranked above Hiatus-Raising
in order to restrict raising to a one-step process:
(42) a. High vowel raising: Hiatus-Raising
Ident-IO(raised)
b. Mid vowel raising: Hiatus-Raising
Ident-IO(high)
c. Interaction: [Ident-IO(raised) & Ident-IO(high)]

Hiatus-Raising
The composite constraint [Ident-IO(raised) & Ident-IO(high)] is violated if and
only if Ident-IO(high) and Ident-IO(raised) are violated with respect to a given
segment. The domain is equal to the segment for which both faithfulness
constraints are evaluated.
Tableau (43a) illustrates how the fell swoop candidate e ij is eliminated by
the composite constraint:
(43)
[Ident-IO(high) & Hiatus- Ident-IO Ident-IO
Ident-IO(raised)] Raising (high) (raised)

a.i ee **!
a.ii + e i * *
a.iii ei j
*! * *

b.i ie **! *
b.ii ii *!
b.iii + i i j *

But in tableau (43b), the same output i i j is allowed as it involves only a single
step, hence the composite constraint is not violated.

Finally, let us briey discuss why Sympathy theory cannot deal with chain shifts.
To account for opacity of raising in e i, the output must be maximally faithful
to a {-candidate that is [raised] itself. Therefore Ident-IO(raised) must be the
selector, and [i] the {-candidate (which also happens to be the output). Assum-
ing Ident-{O(raised) as the undominated {O-faithfulness constraint, we arrive
at tableau (44a):
396
 9.2 Opacity

(44) Ident-{O Hiatus- Ident-IO Ident-IO

(raised) Raising (high) (raised)

a.i ee *!*
a.ii + { e i * *
a.iii e ij *! * *

b.i ie *!* *
b.ii L { i i *
b.iii ii j
*! *

In (44b), however, we see that the same ranking predicts that [i] is the {-
candidate, hence the (incorrect) output. Reversing the ranking of Ident-
{O(raised) and Hiatus-Raising may account for i ij, but will spoil the correct
outcome for input /e/, predicting e ij. This ranking paradox cannot be resolved.

Local Conjunction has seen wider applications (outside chain shifts) in a range
of phenomena, including dissimilation and Obligatory Contour Principle (OCP)
effects (Alderete 1997, It and Mester 1998) and word stress (Kager 1994, Crow-
hurst and Hewitt 1997). The shared property of all these phenomena is that
multiple violations of basic constraints within some domain are banned, while
a single violation of each of the constraints individually is tolerated. A specic
case is that where multiple violations of one constraint within a domain are
banned: self-conjunction.
Here we will illustrate this approach for OCP effects. It and Mester (1998)
discuss various restrictions on the cooccurrence of segments within Japanese morph-
emes, which ban cooccurring segments sharing the same marked value of some
feature. One of these restrictions is known as Lymans Law:

(45) Lymans Law: Stems must not contain more than one voiced
obstruent.

Lymans Law is exemplied by the minimal pairs (46ac) and systematic gaps
(46d):

(46) a.i kaki persimmon a.ii futa lid

b.i ka
i sign b.ii fuda sign
c.i
aki writing c.ii buta pig
d.i *
a
i (non-existent) d.ii *buda (non-existent)
397
Residual issues

This is arguably a case of positional neutralization voicing is distinctive in

obstruents, but up to a certain limit: maximally one voiced obstruent may occur
in a stem. Moreover, it can be seen as a static kind of dissimilation opposite
values for voicing are preferred to cooccurrences of [+voice].
Earlier autosegmental analyses of Lymans Law assumed the tier-adjacency of
the feature [voice], excluding adjacent occurrences of [+voice]. Such analyses
involved the OCP (At the melodic level, adjacent identical elements are prohib-
ited).12 The fact that [+voice] is targeted by the OCP, rather than [voice], is not
an accidental property of Japanese. Cross-linguistically, dissimilations are trig-
gered by adjacent segments specied for marked values for a given feature. For
this reason, OCP analyses assume the underspecication of unmarked feature
values. (See Steriade 1995a for a criticism of such approaches.)
It and Mester (1998) take a different view, and argue that the markedness
content of Lymans Law should be captured directly, without underspecication.
This involves a constraint banning cooccurrences of marked [+voice] within a morph-
eme. This constraint is an extension of VOP, the basic markedness constraint
banning [+voice] in obstruents (known from chapter 1, section 1.7.5):
(47) Voiced Obstruent Prohibition (VOP)
*[+voice, son]
Since voicing is distinctive in Japanese obstruents (within the limits imposed
by Lymans Law), VOP must be dominated by Ident-IO(voice). However, the
absolute ban on double occurrences of [+voice] in a stem requires the conjunc-
tion of VOP with itself, dominating Ident-IO(voice). It and Mester propose the
following schema for self-conjunction:
(48) Self-conjunction of constraints: [C1 & C1] with C1 = C1
Evaluation of [C1 & C1]: [C1 & C1] is violated in domain if
there is more than one violation of
C1 in domain .
Applying self-conjunction to VOP, this schema produces [VOP & VOP]Stem, or
VOP2:

(49) VOP2
No cooccurrence of voiced obstruency with itself.
In order to full its function of positional neutralizer, VOP2 must be ranked
above the basic sequence Ident-IO(voice)
VOP. Considering a hypothetical
input that has a dual occurrence of [+voice, son], one of these is erased, while
12
The OCP was originally proposed for tonal systems (Leben 1973, Goldsmith 1976), and later
generalized into a wider principle governing dissimilations (McCarthy 1986).

398
 9.2 Opacity

the other is preserved. Which of them is erased cannot be determined here, given
the fact that there are no alternations:

(50) Input: /
a
i/ VOP2 Ident-IO (voice) VOP

a.
a
i *! **
b. + ka
i * *
c. +
aki * *
d. kaki **!

Minimal violation of Ident-IO(voice) guarantees that only one occurrence of

[+voice] is erased (rather than both, as in 50d).
Another advantage of Local Conjunction is that it rationalizes existing context-
ual markedness constraints as conjunctions of two basic constraints (Smolensky
1995, It and Mester 1998). For example, *Voiced-Coda (coda obstruents are
voiceless) is construed as the Local Conjunction of the context-free markedness
constraint VOP (obstruents are voiceless) and the prosodic markedness con-
straint NoCoda (no codas). Violation of *Voiced-Coda implies a combined
violation of VOP and NoCoda in a single segment. Hence, [NoCoda &
VOP]Segment is equivalent to *Voiced-Coda. The example is from Dutch (see
chapter 1):

(51) Input: /bd/ [NoCoda Max-IO Ident-IO NoCoda VOP

& VOP] (voice)

a. bd *! * **
b. + bt * * *
c. pt **! *
d. b *! *

The original constraint is somehow unsatisfactory because it restates two inde-

pendently needed markedness constraints. In cases such as these, Local Conjunc-
tion helps to reduce the set of basic constraints, leading to greater generalization,
and decreasing arbitrariness in the statement of individual constraints.
Should Local Conjunction be granted the status of basic theoretical device of
OT? It certainly has its positive sides. First, it lls the empirical gap left by
Sympathy in the analysis of opacity (chain shifts). Second, in its self-conjunction
399
Residual issues

mode, it offers a highly elegant implementation of the insight that multiple viola-
tions of a constraint are worse than a single violation (Mohanan 1993). This extends
to OCP effects, as we saw above. However, Local Conjunction also has some
negative sides.
Firstly, although banning the worst-of-the-worst is undeniably in the spirit of OT,
conceptual problems do arise. Local Conjunction seems to stipulate what should
come for free in OT, given the architecture of the theory. OT is designed to deal
with forms which have more-than-necessary violations, against the principle of
minimal violation. However, minimal violation by itself, as a standard property
of interactions of basic constraints, should sufce to exclude the worst-of-the-
worst, rather than an extra mechanism of constraint conjunction. Accepting it
means that the theory has two means to rule out excessive violations of faithful-
ness: (a) minimal violation of basic constraints, and (b) Local Conjunction.
A second conceptual problem is that Local Conjunction seems to undermine
strict domination, a core principle of OT. Under strict domination, violation of
higher-ranked constraints cannot be compensated for by satisfaction of lower-
ranked constraints. (See chapter 1.) But under Local Conjunction, two con-
straints A and B, each of which is ranked too low to force the violation of C, can
nevertheless dominate C by joining forces in a conjoined constraint [A & B].
Such a situation, in which two constraints team up against a third constraint,
conicts with the intuition behind strict domination (even though Local Conjunc-
tion and strict domination are not formally inconsistent).
Finally, questions arise with respect to the huge increase of possible constraints
in Gen which is implied by Local Conjunction. For example, can any pair of
constraints be conjoined? This predicts constraints of a typologically doubtful
status, such as [Onset & NoCoda] or [*[+round] & Parse-Syl]. Can faithful-
ness constraints be conjoined with markedness constraints?13 If so, this predicts
even more bizarre constraints, such as [*Nt & Ident-IO(Place)]. Can any
number of constraints be conjoined, or is there a (binary?) upper limit? Without
any upper limit, the number of constraints in Gen becomes innite, with fatal
effects on learnability. But even under maximally restrictive assumptions about
Local Conjunction, a vast increase occurs in the amount of constraint interaction.

9.3 Absolute ungrammaticality

The essence of OT is that grammatical outputs are the best possible compromise
between conicting needs. This implies that absolute well-formedness of output
forms cannot be a criterion for grammaticality. (Some constraints are necessarily
violated in every output.) This approach to well-formedness predicts that for every

13
It and Mester (1998) exclude this, but see section 9.5 below.

400
 9.3 Absolute ungrammaticality

lexical input, some grammatical output exists the optimal output, which minim-
ally violates the conicting constraints in a xed hierarchy. Nevertheless, languages
do not strictly conform to this picture, since cases of absolute ungrammaticality
occur. A specic input has no grammatical output:

(52) Input Grammar

The best-studied examples involve the blocking of word formation processes to

avoid the violation of some phonological well-formedness constraint. How does
OT deal with such cases? Does absolute ungrammaticality motivate a fundamental
revision of the model?
We will discuss an example from stress-based conditions on afxation in Eng-
lish, analysed by Raffelsiefen (1996). Raffelsiefen argues that the verbalizing sufx
-ize may not attach to adjectives ending in a stressed syllable, while it product-
ively attaches to adjectives ending in an unstressed syllable:

(53) Blocking of -ize after bases with nal stress in English

a.i rndom rndom-ze a.ii freign freign-ze
b.i corrpt *corrpt-ze b.ii obscne *obscne-ze

Note that -ize is a stressed sufx (it has secondary stress wherever it appears),
and that its attachment to a stem with nal stress would create a stress clash, a
situation of adjacent stresses. Intuitively, clash avoidance takes priority over
afxation. But how can this be captured in OT, a theory of violable constraints?
If word formation were simply blocked if its optimal output violated any phono-
logical constraint, then no word formation would be possible at all. To avoid
wholesale nullication of morphology, the lexical input must be realized by the
grammar, that is, be mapped onto some output. Instead of blocking the input, the
grammar selects its optimal output analysis, possibly by a repair. In the case in
hand, repair may involve violation of accentual faithfulness (Max-Ft-IO input
stresses occur in the output), or well-formedness (*Clash). In sum, absolute
ungrammaticality poses a challenge to OT. In contrast, a theory of inviolate con-
straints has no problems: an output lter (with nullifying power) will reject any
outputs violating *Clash.

Raffelsiefen employs an analytic resource that was originally proposed by Prince

and Smolensky (1993). The grammar always produces an output for every input.
However, by Freedom of Analysis, Gen supplies one special output, which is
identical to the input: a set of disjoint morphemes lacking any morphological or
phonological cohesion. This Null Parse is left unpronounced, due to its lack of
morpho-phonological word status. It will be optimal, though, if no other candid-
ate is available that is more harmonic:
401
Residual issues

(54) Candidates
C1
C2
Input ... Output

...
Cn
The question then is, how can the Null Parse be optimal? Due to its lack of
phonological and morphological word status, it has one substantial advantage over
all other candidates: it cannot violate any well-formedness constraints pertain-
ing to the word level. However, the Null Parse violates the following constraint
(McCarthy and Prince 1993b):

(55) MParse
Morphemes are parsed into morphological constituents.

The ranking of MParse in the grammar determines the robustness of word

formation. If it is undominated, some (non-null) output will be produced, since
anything is better than the null parse. But any constraint that dominates MParse
has the power of annihilating afxation. This is seen in tableau (56), in which
MParse is dominated by an OO-identity constraint Ident (The stem of the
derived word must be identical to the base), as well as by *Clash (Two adjacent
stresses are prohibited).

(56) Input: / krmpt, -z/ Ident *Clash MParse

a. kI.rp.tz *!
b. k.rmp.tz *!
c. + {krmpt, -z} *

The Null Parse (56c) is optimal since all its competitors violate undominated con-
straints. Hence it is better to have no (tangible) output at all than an ill-formed output.
Orgun and Sprouse (1997: 4) point out empirical problems for MParse theory.
They argue that . . . there are cases of ungrammaticality in which the ungram-
matical candidate could be repaired by violating a constraint independently
known to be violable in other (grammatical) output forms in the language. That
is, the constraint ranking required to make the Null Parse win is demonstrably
inconsistent with the ranking required for non-null forms of the language. If these
objections (and the analyses on which they are based) are valid, then they pose
serious challenges to MParse theory.
402
 9.3 Absolute ungrammaticality

Yet another unsatisfactory aspect of the MParse model is the assumption that
the Null Parse, in spite of its lack of phonetic realization, violates no faithfulness
constraints. This assumption is required to prevent the undesired result that the
Null Parse is rejected for violating the same constraints that exclude its non-null
(ungrammatical) competitors. For example, adding a violation mark for Ident to
the Null Parse (56c), penalizing its lack of phonetically realized stress, would
render it less harmonic than all its competitors. Alternatively, it might be stipu-
lated that identity constraints are only activated for outputs that are phonetically
interpreted, but that seems equally ad hoc.
Alternatives to MParse have been proposed, changing the basic design of the
OT grammar, by introducing inviolate constraint components. Orgun and Sprouse
(1997) argue that a component of inviolate constraints (Control) is serially
ordered after Gen and Eval as an output ltering device. As in standard OT, Eval
selects an optimal candidate, which is submitted to Control. Unlike Eval, Control
has the power of marking as ungrammatical any outputs violating its constraints.
Under this conception the grammar contains two sets of undominated constraints:
those whose violation is avoidable by repairs (these are part of Eval ), and those
whose violation is beyond repair (these are part of Control). In the case in hand,
the output [k.rmp.tz], optimal under Eval, is submitted to Control, which lters
it out, leaving the input /krmpt, -z/ without an output.
However, this proposal has its own specic problems. Firstly, it is essential that
constraints in Control, as output lters blocking afxations, be inviolate in all
outputs of the language. But for English, most researchers would agree that
*Clash is violable even within a Prosodic Word, as forms like Chnse reveal.
Secondly, Control is a rather blunt tool, which is principally unable to account for
different types of blocking within a single language. For example, a language may
have two defective afxations, each of which is blocked in a separate phono-
logical context. Here Control would be of no value because it predicts that the
afxations share gaps in the same contexts. In contrast, MParse theory may deal
with such cases by splitting up MParse into different afx-specic versions. In
sum, absolute ungrammaticality is another challenge which OT has not yet
completely mastered.14

14
Broekhuis (forthcoming) argues that absolute ungrammaticality can be dealt with by a com-
ponent of inviolate constraints preceding Eval. This proposal is comparable to the assumption
that Gen contains universally inviolable constraints (Prince and Smolensky 1993), ltering out
candidates even before they are submitted to Eval. Broekhuis deviates from this idea, however,
in his denition of input. In syntax, as we have seen in chapter 8, the input is less clearly
dened than in phonology. Broekhuis, for example, denes the input to Gen as a complete syn-
tactic representation developed in a component preceding Gen, whose requirements (check-
ing) are inviolable. Whether such a model can deal with absolute ungrammaticality in phonology
remains to be seen, however.

403
Residual issues

9.4 Free variation

Another unresolved issue is variation. We will focus on variation of a specic
type: the case of a single input being mapped onto two outputs, each of which is
grammatical. This is free variation, also known as optionality:
(57) Output1
Input Grammar
Output2

By denition, the distribution of both outputs cannot be under grammatical con-

trol, since that would restore (ranking-based) determinism in the choice of both
variants.
Examples of free variation are abundant in natural languages. Examples below are
from English phonology (vowel reduction, 58a) and syntax (complementizer, 58b):

(58) a. sentim[en]tality ~ sentim[n]tality

b. I know that John will leave ~ I know John will leave

The fact that variation is free does not imply that it is totally unpredictable, but
only that no grammatical principles govern the distribution of variants. Never-
theless, a wide range of extragrammatical factors may affect the choice of one
variant over the other, including sociolinguistic variables (such as gender, age, and
class), and performance variables (such as speech style and tempo). Perhaps the
most important diagnostic of extragrammatical variables is that they affect the choice
of occurrence of one output over another in a stochastic way, rather than determin-
istically. Here we will focus on the consequences of free variation for the grammar.
Why does free variation pose a challenge to OT? An OT grammar is essentially
an input output mapping device. The grammar is deterministic, in the sense that
each input is mapped onto a single output the most harmonic candidate for a
constraint hierarchy. Given a single deterministic competition, how can two can-
didates ever both be optimal? If two output candidates O and O are different in
grammatical terms, then this difference must be relevant to some constraint(s) in
the hierarchy. This implies that O and O do not share the same violations marks,
hence one is more harmonic than the other with respect to the hierarchy. The
question is: how can free variation be reconciled with the deterministic nature of
the grammar?15

15
It has been argued that some cases of free variation indeed derive from inability on the part
of the constraint inventory to distinguish different outputs (Hammond 1994, Smolensky
1996). That is, the constraint inventory simply lacks constraints discriminating between two
outputs O1 and O2. However, even if this approach is feasible for the particular cases for which
it has been proposed, it is clearly not generalizable to all cases of free variation. Most involve
presence or absence of grammatical structure which is visible to some constraint (ultimately
to *Struc no phonological structure).

404
 9.4 Free variation

To be fair, the problem of non-unique outputs is not specic to OT. It reoccurs

in derivational theory, or in fact any theory of grammar that is faced with the
situation (57). A solution at a terminological level is to redene the output of the
grammar for a given input as a set of forms {O1, O2 . . . On}, rather than a unique
form. But of course this does not solve the problem of how to generate sets of
output forms. In derivational rule-based theory, specic rules may be marked as
optional. In a derivation, an optional rule may be applied (in which case some
output results), or be left unapplied (in which case another output results). But
what is the counterpart of optional rules in OT? Contrary to rules, constraints
are not language-specic devices, but elements of universal grammar that are
potentially active in every grammar. Therefore a solution to the free variation
problem can only reside in constraint ranking.
A radical method is splitting up the grammar into multiple constraint hier-
archies, or co-phonologies, each of which selects its own optimal candidate by its
own ranking. Such an approach is reminiscent of multi-stratal evaluation, dis-
cussed in section 9.2.3 in relation to opacity. It is not too difcult to imagine an
extension of this stratal idea to free variation. If we assume that strata can be
organized in parallel (in addition to serially, as in section 9.2.3), an input can be
fed into two parallel co-phonologies, giving two outputs:

(59)
Co-phonology 1 Output1
Input
Co-phonology 2 Output2

In section 9.2.3 we mentioned weaknesses in models of multi-stratal evaluation,

one of which has a direct counterpart in parallel co-phonologies. Splitting a
grammar up into subgrammars (strata or co-phonologies) makes the prediction
that each subgrammar is independent, hence that subgrammars can be radically
different. This prediction is clearly incorrect for free variation, where both outputs
are mostly similar, and differ only in a minor respect. (Co-phonologies are better
suited for lexical variation resulting from strata; see Inkelas and Orgun 1995 and
It and Mester 1995; and exceptionality; see Inkelas, Orgun, and Zoll 1997.)
A less radical alternative is to maintain a single constraint hierarchy, while
giving up the idea of a xed ranking of constraints. Throughout this book we have
assumed that constraints are strictly ranked. Two conicting constraints C1 and C2
are ranked in either of two ways: C1 strictly dominates C2, or C2 strictly dominates
C1. This assumption was upheld even for cases in which C1 and C2 cannot be
ranked with respect to each other due to a lack of interaction (so that no empirical
evidence can determine the ranking). In such cases we ranked C1 and C2 in
the same position (called stratum in chapter 7). But we never considered the
405
Residual issues

possibility that constraints were crucially unranked. Free ranking was observed
as a purely theoretical option by Prince and Smolensky (1993: 51), and has since
been argued to be the OT counterpart of optional rule application (Kiparsky
1993b, Kager 1994, 1997b, Reynolds 1994, Anttila 1995). When two constraints
C1 and C2 are freely ranked, the evaluation procedure branches at that point. In
one branch, C1 is ranked above C2, while in the other branch the ranking is
reversed.

(60) Interpretation of free ranking of constraints C1, C2

Evaluation of the candidate set is split into two subhierarchies, each
of which selects an optimal output. One subhierarchy has C1
C2,
and the other C2
C1.

Note that free ranking preserves strict domination, which holds within each sub-
hierarchy.
A (slightly simplied) example of free ranking is given below for English
vowel reduction. The conict between Reduce (vowels lack quality) and Ident-
IO is resolved in favour of either the latter (61a) or the former (61a):

(61) Ident-IO Reduce

a.i + sentim[en]tality *
a.ii sentim[n]tality *!

Reduce Ident-IO

b.i sentim[en]tality *!
b.ii + sentim[n]tality *

In what way(s) does free ranking differ from parallel co-phonologies? Is what
we call subhierarchy just another name for co-phonology? To some extent the
answer is yes, since free ranking entails evaluation of candidate forms by paral-
lel competitions. There is a substantial difference between the approaches, however.
Subhierarchies differ only in constraints whose ranking is not stipulated by the
grammar. One may think of a grammar with free rankings as underdetermined.
Variable ranking remains consistent with the hard rankings stated in the grammar,
a limitation explaining the observed similarities between variable outputs. Radical
differences between outputs are possible, but this entails a large amount of free
ranking, with a corresponding increase in the number of outputs. In a free ranking
model, two factors are positively correlated: the degree of dissimilarity between
406
 9.5 Positional faithfulness

variable output forms and the number of variable outputs. In contrast, the
co-phonologies approach fails to predict such a correlation, since rankings of
different co-phonologies are intrinsically unrelated. A grammar with two co-
phonologies may select only two variable outputs of radically different shapes.
A nal point in favour of free ranking is that it offers a fairly accurate estimate
of the frequency of occurrence for each output (Anttila 1995).
The free ranking approach to variation looks promising, but a number of poten-
tial problems merit attention in future research. First, it is unclear whether OT
grammars with free ranking are learnable at all. (If so, free ranking is likely to
require major adaptions of Tesar and Smolenskys learning algorithm.) Second,
not necessarily on a negative note, the notion of free ranking may have conse-
quences for the concept of OT grammar which are sharply at odds with Prince
and Smolenskys (1993) principle of strict domination. Fine-tuning of free varia-
tion may be achieved by associating a freely ranked constraint with a numerical
index indicating its relative strength with respect to all other constraints. This may
pave the way to a probabilistic view of constraint interaction, replacing the doc-
trine of strict domination and moving into the direction of connectionism. Such
a development is already apparent in functionally oriented OT (Boersma 1997),
with precursors in Goldsmith and Larson (1990) and Mohanan (1993).

9.5 Positional faithfulness

Languages exhibit asymmetries with respect to contexts in which phonetic con-
trasts can be realized. For example, in chapters 13, we have seen that, generally
speaking, codas display a more restricted set of contrasts than onsets do. We
assumed that such positional neutralization is due to contextual markedness con-
straints (militating against a feature value [F] in specic positions), interacting
with context-free faithfulness constraints. But a priori, what we refer to as coda
neutralization may equally well be described as faithfulness in onsets. Here we
will sketch an alternative view, positional faithfulness, which radically revises the
roles of markedness and faithfulness constraints. Resistance to neutralization is
attributed to constraints that license features in specic positions,16 an interaction
with context-free markedness constraints.

(62) Two possible views of positional neutralization:

a. Context-free faithfulness Positional markedness
b. Positional faithfulness Context-free markedness

16
The notion of licensing originates in pre-OT phonology (Goldsmith 1990, Kaye, Lowenstamm,
and Vergnaud 1990, Steriade 1995a).

407
Residual issues

In most cases it is simply impossible to nd evidence for one view or the other,
due to the fact that the contexts of neutralization and faithfulness are both posit-
ively characterized by mutually exclusive labels (onset, coda, etc.). In current
literature both views have been adopted. Regardless of the issue of whether one
or both views are adequate, two general arguments for positional faithfulness can
be put forward.

Firstly, positional faithfulness is supported by functional considerations. It is well

known that contrasts are best realized in perceptually salient positions (Noote-
boom 1981, Hawkins and Cutler 1988, Ohala 1990, Ohala and Kawasaki 1984).
Salient positions include word-initial consonants, prevocalic (or released) con-
sonants, stressed vowels, and vowels in initial syllables. For example, it may be
assumed that all vowels are subject to the same general forces of reduction (min-
imize articulatory effort), regardless of their position or stressing. Yet stressed
vowels, by their inherent perceptual salience (tonal and durational) are best equip-
ped to realize featural distinctions, hence to resist general reduction forces.
Recent OT studies have implemented perception-based asymmetries in the notion
of positional faithfulness. Phonetic grounding renders faithfulness constraints
sensitive to perceptually prominent positions. More specically, a faithfulness
constraint for a feature [F] referring to a prominent position dominates the gen-
eral faithfulness constraint for [F]. This allows a sandwiching of markedness
constraints in between positional and general faithfulness constraints, producing
positional neutralization:

(63) Ranking schema for positional neutralization

IO-Faithfulness (prominent positions)
Markedness

IO-Faithfulness (general)

Some researchers (Steriade 1995b, Flemming 1995, Jun 1995, Kirchner 1995,
1997) have argued that perceptual principles should be directly stated in gram-
mars, thus allowing reference to gradient and non-contrastive phonetic features.
Others (Selkirk 1995, Lombardi 1995b, Beckman 1997a, b) maintain a strict separa-
tion of phonology and phonetics, in the sense of avoiding reference to gradient
features.

The second (empirical) argument for positional faithfulness is that it captures and
unies a number of phonological patterns in different languages. More specic-
ally, there is an overwhelming typological tendency for neutralizing assimilation
to preserve feature values of segments in salient positions (onsets, initial syl-
lables, root segments, etc.), at the expense of segments in other positions (codas,
408
 9.5 Positional faithfulness

medial syllables, afx segments, etc.). In this section we will look into one of
these cases the licensing of the feature [+round] by initial syllables drawing
on non-OT work by Steriade (1995a) and inspired by the OT analysis of Shona
vowel harmony by Beckman (1997a).
A typologically common restriction regarding [+round] is that it occurs on
vowels in specic positions. For example, several Altaic languages allow rounded
vowels only in the initial syllable of the word. This is a case of contextual neutral-
ization: the rounding contrast is suppressed in segments in non-initial syllables.
However, an analysis of this pattern by a contextual markedness constraint ban-
ning [+round] from non-initial syllables is awkward, as the context of neutraliza-
tion (non-initial syllables) is not a natural class. This defect is avoided by a
positional faithfulness analysis. A constraint Ident-IO(round, [) militates
against the loss of [round] in initial syllables:

(64) Ident-IO(round, [)
An output segment standing in the initial syllable has the same value
for [round] as its input correspondent.

This constraint is (nearly) identical to the general faithfulness constraint Ident-

IO(round) (65), the only difference being that the latter fails to refer to initial
syllables:

(65) Ident-IO(round)
An output segment has the same value for [round] as its input
correspondent.

It is assumed that pairs of faithfulness constraints such as (65 6) are universally

ranked such that the position-specic constraint is ranked above the general,
position-insensitive constraint. (Compare harmony scales discussed in chapter
1, section 1.8.) Hence:

(66) Harmony scale

Ident-IO(round, [)
Ident-IO(round)

This installs the faithfulness part of the interaction. Let us now look into the
markedness part, which is very simple. The faithfulness constraints compete with
a general, context-free markedness constraint *[+round], which penalizes round
vowels. In Altaic, rounding is lost in all vowels except those that are licensed by
an initial syllable. This is expressed in the following ranking:

(67) Licensing of rounding in initial syllables

Ident-IO(round, [)
*[+round]
Ident-IO(round)
409
Residual issues

Let us assume a hypothetical input with rounding in both of its vowels. Candidate
(68c) fatally violates positional faithfulness in its initial syllable. Of the two
remaining candidates, (68b) is optimal since it minimally violates *[+round]:

(68) Input: /u-u/ Ident-IO *[+round] Ident-IO

(round, [) (round)

a. u-o **!
b. + u-i * *
c. i-i *! **

This analysis offers an empirical argument for positional faithfulness: it is able to

capture contextual neutralization of [round], without reference to a non-natural
class non-initial syllables as the context of neutralization.

However, an even better argument can be made for positional faithfulness on the
basis of rounding harmony. In Yokuts (Newman 1944), [+round] is contrastive in
the rst syllable only, as in Altaic, but it also spreads beyond its directly licensed
position, into afx vowels.17

(69) a.i dub-hun leads by the hand

a.ii dub-mu having led by the hand
a.iii dub-nut will be led by the hand
b.i xil-hin tangles
b.ii xil-mi having tangled
b.iii xil-nit will be tangled

Rounding in non-initial syllables is indirectly licensed by the initial syllable. This

pattern is a variation on the direct licensing pattern seen in Altaic (67). A marked-
ness constraint Harmony (vowels agree in their values of [round]) dominates
*[+round]:

(70) Rounding harmony controlled by initial syllable

Ident-IO(round, [), Harmony
*[+round]
Ident-IO(round)

For a hypothetical input /dub-hin/, this ranking is illustrated by the following

tableau:

17
We abstract away from the interaction with vowel height, as well as from opacity (section 9.2).

410
 9.5 Positional faithfulness

(71) Input: /dub-hin/ Ident-IO Harmony *[+round] Ident-IO

(round, [) (round)

a. + dub-hun ** *
b. dub-hin *! *
c. dib-hin *! *

Harmony is satised in both (71a) and (71c) but only the latter violates posi-
tional faithfulness in its initial syllable. This analysis is highly elegant, and it also
captures the relation with the direct licensing of rounding in the Altaic languages
by a simple reranking. (In Altaic, *[+round] dominates Harmony.) Finally, a posi-
tional faithfulness analysis is superior to one based on contextual markedness.
With only a general faithfulness constraint Ident-IO(round) to penalize changes
of [round], the inertness of the initial syllable must be due to contextual marked-
ness. Even if we assume a constraint militating against [+round] in non-initial
syllables (ranking it above *[+round]), its interaction with Harmony is highly
problematic. For example, Harmony
*[+round]
Ident-IO(round) predicts
a harmony pattern *dib-hin, regardless of where we rank the contextual marked-
ness constraint. This empirical argument for positional faithfulness complements
the typological argument that was discussed earlier.

Steriade (1995a: 162) points out extensions of the idea of indirect licensing to
the cases of local assimilation in which onsets spread place or laryngeal features
onto adjacent coda consonants. In many languages, codas cannot carry independ-
ent values for [voice] (as in Dutch, chapter 1) or place of articulation (recall the
discussion of the Coda Condition in chapter 3). In a subset of these languages,
coda neutralization is accompanied by coda-to-onset assimilation, resulting in an
indirect licensing of features by the onset.

(72) Coda Onset

[F] [F]

Steriades observations on indirect licensing in local assimilation carry over to

positional faithfulness in OT, as demonstrated in recent work by Jun (1995),
Lombardi (1995b), and Beckman (1997a, b). Similar onsetcoda asymmetries
have been observed with respect to consonant deletions (see chapter 3, section
3.6). Again, positional faithfulness offers an explanation for such asymmetries
(explaining why coda deletion is favoured over onset deletion).
411
Residual issues

Let us now evaluate the notion positional faithfulness by criteria that have played
a major role earlier in this chapter. These are constraint format and factorial
typology.
First, does positional faithfulness maintain the strict separation between faith-
fulness constraints and markedness constraints? Recall that in section 9.2.2 we
criticized two-level well-formedness for blurring this distinction, and giving con-
straints rule-like power. Positional faithfulness constraints resemble two-level
well-formedness constraints in their reference to prosodic positions of output seg-
ments. Unlike two-level well-formedness constraints, they cannot be violated if
the output segment fails to occupy the relevant position, and so cant trigger
changes. In this respect, compare (73) and (74) stated in their correspondence
formats:

(73) Positional faithfulness

If an output segment S is [round] and S stands in an initial syllable ,
then its input correspondent S must be [round].
(not violated if S is not in an initial syllable)

(74) Two-level well-formedness

If an input segment S is [round], then its output correspondent S
must be [round] and stand in an initial syllable .
(violated if S is not in an initial syllable)

To avoid the problems inherent to two-level well-formedness constraints (includ-

ing their rule-like nature, mentioned in section 9.2.2), constraints of the type
(74) should be disallowed.

A second criterion is factorial typology. We have seen that positional faithfulness

is promising in several respects, but nonetheless it appears that contextual marked-
ness, as discussed in chapter 1, cannot be simply eliminated. For that would
produce the fatally impoverished typology below:

(75) A factorial typology of markedness and faithfulness

a. P-Faithfulness
Markedness
Faithfulness Positional neutralization
b. Faithfulness, P-Faithfulness
Markedness Full contrast
c. Markedness
Faithfulness, P-Faithfulness Lack of variation

What is lacking from this factorial typology, as compared to the one of chapter
1, is the case of allophonic variation. This arises under a complete domination
of faithfulness for a feature [F] by markedness constraints for [F], with contextual

412
 9.6 Underlying Representations versus allomorphy

markedness dominating context-free markedness. Positional faithfulness fails to

reproduce this pattern since total domination of faithfulness obscures any con-
textual effects of positional faithfulness. The conclusion must be that positional
faithfulness cannot replace contextual markedness and that both are necessary.
This nding implies a more complex factorial typology, which we will not present
here, though.

9.6 Underlying Representations versus allomorphy

Classical generative phonology (Chomsky and Halle 1968) is based on three
assumptions: (i) all contextual variants of a morpheme derive from a single Under-
lying Representation, (ii) by rewrite rules (A B / X__Y), (iii) applying in a serial
derivation, based on linear ordering. As we have seen, both the second and third
assumptions have been abandoned by OT, but so far we have still assumed that
the input of an OT grammar is an Underlying Representation (UR) in the classical
sense: a unique lexical form underlying all alternants of a morpheme. The notion
of UR may have changed as a result of Lexicon Optimization, but it still functions
as the input of Gen: all candidate outputs evaluated are analyses of a single UR.
The standard correspondence model, which includes URs, is given below:

(76) Standard model (with Underlying Representations)

OO-Identity
Output Base Afxed form

IO-Faithfulness

Input UR (afxed) UR

Base-identity is a priority of language: it enhances the one-to-one relation between

lexical items, atoms of meaning, and the sound shapes encoding them. We will
call this uniform exponence. An overlap in functions exists between base-identity
and UR: both maximize uniform exponence. In this sense standard Correspond-
ence Theory, which has two means for a single goal, is conceptually redundant.
Recently attempts were made to eliminate UR from phonological theory
(Burzio 1996, forthcoming, Hayes forthcoming). This development can be seen
as yet another extension of the role of output forms in explaining alternations, as
well as a logical continuation of ideas on base-identity (Benua 1995, McCarthy
1995a, Kenstowicz 1996).

On the standard view, the input equals an abstract UR, which is mapped onto the
output, an actual surface representation. UR is abstract as it need not surface in
its input form, but serves to supply phonotactically unpredictable shape aspects

413
Residual issues

of alternants. If we eliminate UR, ignoring alternations for the moment, then

abstractness of the input is no longer necessary, since every input will be identical
to its actual output. On this new UR-less view, the input is simply an intended
surface form, whose relative harmony is to be determined by Eval. If the gram-
mars optimal output matches the input, then the input is grammatical. The role
of the grammar as an inputoutput mapping mechanism changes into a checking
mechanism.
This checking role of the grammar in fact originates in the standard model,
where it forms the cornerstone of the theory of neutralization, allophonic varia-
tion, and contrast. Under Richness of the Base, any imaginable input may be
proposed to Eval, but only a subset of inputs actually surface in their original
shape. However, if no alternations occur in a morphemes shape, the learner will
never postulate an input deviating from the actual observable output form. Due
to Lexicon Optimization, the input simply equals the output unless there is reason
to deviate (chapter 1, section 1.6).

Let us now extend this checking view to cases in which alternations do exist, that
is, cases for which the standard model uses an abstract UR. We start by making
the major assumption that each surface alternant of a morpheme has its own
input, which is simply identical to its output under Lexicon Optimization. (That
is, we temporarily ignore the fact that alternants are systematically related in their
shapes.) The assumption of multiple inputs for each morpheme (allomorphs) thus
allows the grammar to continue to function as a checking device.
In the Dutch example, the input of the singular [bt] bed is simply {bt},
while that of the plural [bdn] is simply {bdn}. (We will use the notation
{ . . . } for inputs to avoid confusion with the standard UR notation / . . . /.) Still,
under Richness of the Base, an alternative singular input {bd} might have been
proposed, but this would not have been able to make it to the surface due to a
constraint interaction (*VoicedCoda
Ident-IO(voice)) neutralizing it into the
output [bt]. Therefore the optimal lexical input for [bt] is {bt}, due to Lexicon
Optimization. The grammar does not impose such a neutralization on the input
{bdn}, which makes it to the surface without change:

(77) Input {bt} Output [bt] Input {bdn} Output [bdn]

Input {bd}

In this UR-less model, the input (lexical shape) simply equals the output (surface
shape). IO-faithfulness maintains its original function of reinforcing parts of the
input, protecting it against the neutralizing forces of markedness.

414
 9.6 Underlying Representations versus allomorphy

We now have a conceptually simple model, which accounts for neutraliza-

tion as well as allophonic variation, but which evidently fails to account for
alternations. On this simple model, shapes of morphologically related outputs
([bt] ~ [bdn]) are not related in any phonological sense, since they do not
share a common input (UR). To deal with alternations, morphologically related
output forms must be subjected to constraints which enforce uniform expon-
ence, limiting the phonological dissimilarity between alternants. This is where
OO-correspondence comes into play: it eliminates the function of UR in cap-
turing phonological shape similarities between morphologically related output
forms.
As in the standard model, identity between base and afxed forms is enforced
by OO-correspondence. However, the key difference from the standard model
resides in a new role of the lexicon. The lexicon no longer supplies a unique UR
for each morpheme, but instead it supplies a set of shape variants of the mor-
pheme, allomorphs, chunks ready for insertion in various morphological contexts
(base or afxed form). The role of IO-faithfulness is revised accordingly: this
checks faithfulness of output allomorphs to their lexical shapes. Finally, well-
formedness constraints preserve their function of evaluating output forms. That
is, we replace model (76) with model (78).

(78) Allomorphic model (without Underlying Representations)

OO-Correspondence
Output Base Afxed form

IO-Faithfulness

Input {Base allomorph, Contextual

allomorph}

Such a model has been advocated in Burzio (1996, forthcoming), elaborating

ideas on word-based morphology as proposed by Aronoff (1976), Bybee (1988,
1995) and others. Two sets of faithfulness constraints are assumed in this model.
OO-correspondence is active in its standard role of checking identity in the net-
work of morphologically related output forms, or the paradigm. IO-faithfulness
checks identity between allomorphs in the lexical input and their output counter-
parts.

Let us come to understand this allomorphic model by an example of an alterna-

tion that we have studied before, the case of Dutch nal devoicing. In a theory
without URs, this alternation depends on allomorphs which have different speci-
cations for the feature [voice]. Here we will abstract away from any other possible
shape variation. Furthermore, to avoid complications, we focus on alternations of

415
Residual issues

voice in very simple paradigms which consist of the singular stem and the plural.
Therefore we assume that the input is a mini-paradigm, a set of allomorph pairs,
each containing one base allomorph (for insertion into the singular stem) and one
contextual allomorph (for insertion into the afxed plural). As usual, Gen supplies
a set of candidate output forms for each input pair, that is, a paradigm consisting
of an actual singular stem and an actual afxed plural.
First we will look at the evaluation of an input paradigm {bt ~ bd-n}, which
the grammar should approve of without changes, producing [bt] ~ [bd-n]. Gen
proposes an innite set of candidate output paradigms, of which four are shown
in (79).

(79) Function of Gen in allomorphic model

Input {bt ~ bd-n} Gen [bt] ~ [bd-n]
[bt] ~ [bt-n]
[bd] ~ [bd-n]
[bd] ~ [bt-n]

All four are submitted to Eval, which selects output paradigm (80a), indeed
identical to the input:

(80) Input: {bt ~ bd-n} *VoicedCoda Ident-IO Ident-OO

(voice) (voice)

a. + bt ~ bd-n *
b. bt ~ bt-n *!
c. bd ~ bd-n *! *
d. bd ~ bt-n *! ** *

Note that faithfulness Ident-IO(voice) dominates uniform exponence Ident-

OO(voice), a constraint which would have preferred the uniform output para-
digm [bt] ~ [bt-n] (80b). IO-faithfulness thus assumes its traditional role of
penalizing shape differences between inputs (allomorphs) and outputs (candidate
paradigms). Here the optimal output paradigm is maximally faithful to both allo-
morphs in the input. However, complete faithfulness to the input cannot always
be maintained, due to high-ranked well-formedness constraints.
For example, *VoicedCoda must dominate Ident-IO(voice) so as to lter out
the input paradigm {bd ~ bd-n}, mapping this input onto [bt] ~ [bd-n]:
416
 9.6 Underlying Representations versus allomorphy

(81) Input: {bd ~ bd-n} *VoicedCoda Ident-IO Ident-OO

(voice) (voice)

a. + bt ~ bd-n * *
b. bt ~ bt-n **!
c. bd ~ bd-n *!
d. bd ~ bt-n *! * *

The grammar (actually Gen, by Freedom of Analysis) produces an allomorph

([bt]) for insertion in the base output if the lexicon fails to supply one. This is
the rst example of a lexical paradigm being overruled by the grammar: an under-
sized paradigm is inated to match the requirements of output well-formedness:

(82) Ination of paradigm

Lexical {bd} [bd] (Output allomorph 1, afxed form,
lexically supplied)
[bt] (Output allomorph 2, base form,
added by grammar)

Simultaneously, we also observe a neutralization of the input paradigms (80) and

(81) into a single output paradigm [bt] ~ [bd-n].

Up to this point, no activity of uniform exponence has been visible. But of course
the grammar must allow for such activity, so as to restrict the difference in shape
between allomorphs in a paradigm. Without constraints enforcing uniform expon-
ence, any pair of input allomorphs would be able to surface (with the only restric-
tion imposed by high-ranked well-formedness constraints).
For example, Dutch allows allomorph pairs such as [bt] ~ [bd-n], whereas
pairs such as *[bt] ~ [pt-n], with voice alternations in word onset position, are
systematically excluded. Such candidate output paradigms are rejected for viola-
tion of an undominated positional faithfulness constraint:

(83) Ident-OO(voice, [C)

Let be a segment in the base, and be a correspondent of in the
afxed form.
If is in word onset position and [voice], then is [voice].
Corresponding segments in the word onset of the base and the
afxed form must agree in voicing.
This constraint penalizes any alternations between allomorphs with respect to
voicing in the word onset. Any input paradigm is neutralized towards a surface
417
Residual issues

paradigm respecting uniform exponence with respect to word onset voicing. Since
this neutralization could occur at the expense of the voicing value of either the
base or the afxed form, a tie-breaking constraint is required. In tableau (84),
this is the low-ranked constraint VOP:

(84)
Input: {bt ~ pt-n} Ident-OO *Voiced Ident-IO Ident-OO VOP
(voice, [C) Coda (voice) (voice)

a. bt ~ pt-n *! * *
b. bt ~ bt-n * *!*
c. + pt ~ pt-n *
d. pt ~ bt-n *! ** * *

This example demonstrates the interaction of the three major types of constraints in
an allomorphy model: OO-identity, IO-faithfulness, and well-formedness.
This is the second example of a lexical paradigm being overruled by the gram-
mar, the case of an oversized paradigm that is shrunk to match the requirements
of uniform exponence.

(85) Shrinking of paradigm

Lexical {pt} [pt] (Output allomorph 1, lexically supplied)

{bt} (Input allomorph 2, rejected by grammar)

In sum, this allomorphic model captures systematic variation in the shapes of

alternants, and it also restricts the variation in shape between alternants. That is,
input paradigms are not simply copied to the output, but undergo the ltering
function of the grammar. This means that certain input paradigms are inated,
adding an allomorph to match the needs of high-ranked well-formedness con-
straints, while other input paradigms are shrunk, removing input variation between
allomorphs that exceeds the limits imposed by uniform exponence.

An extension of this model may be considered under which the distribution of the
allomorphs is also brought under control of the grammar. The Dutch lexicon
supplies the set of two allomorphs {bt ~ bd}, without marking one for insertion
into stem and the other for insertion into afxed form. The distribution of allo-
morphs is fully predictable from well-formedness constraints, primarily *Voiced-
Coda but possibly also Inter-V-Voice, which, in spite of its low ranking in
Dutch, may still exert its inuence as an emergence of the unmarked:
418
 9.6 Underlying Representations versus allomorphy

(86) Input: {bt ~ bd} *VoicedCoda Inter-V-Voice

a. + bt ~ bd-n
b. bd ~ bt-n *! *

Conceptually, this extension of the allomorph model is quite satisfactory, since it

makes the distribution of allomorphs predictable by the same constraints which
also check their shape. It remains to be seen, however, whether all types of allo-
morphy can be subsumed under it. Similar proposals for allomorph distribution
governed by well-formedness constraints have been made by Mester (1994) and
Kager (1996).

It is simply too early to evaluate the results of a theory that seeks to eliminate
URs. However, some major points can be made even on the basis of the simple
model outlined above. As usual, the model has pros and cons.
On the positive side, this model reduces the abstractness of lexical repres-
entation by eliminating the UR. This is a noteworthy result, since reduction of
abstractness entails an increase in cognitive plausibility, reducing the role of the
learner in inferring patterns that are remote from actual surface patterns. While
it has been claimed in the classical generative literature that such abstractness
is just what it takes to shape an explanatory theory of alternations, any reduction
of abstractness that leaves the generalizing power of the theory unaffected is very
welcome. (Of course it has to be determined whether such a model indeed
matches or even exceeds the explanatory power of UR theory, a point that we will
return to below.)
The second point in favour of the allomorphic model is that it presents a
uniform analysis of all types of alternations, productive and non-productive alike.
The lexicon is the place where idiosyncratic properties of a language are listed,
so nothing guarantees that every morpheme will indeed exhibit the full range of
allomorphs as compared to other morphemes. For example, English is known
to have exceptions to a process of trisyllabic laxing (op[e]que ~ op[]city, but
ob[i]se ~ ob[i]sity, Kiparsky 1982b). On an allomorphic model the difference
between both morphemes is simply this: opaque has two allomorphs alternating
in vowel length, while obese has only one, hence no length alternation. There are
no high-ranked well-formedness constraints, nor OO-identity constraints pertain-
ing to vowel length to enforce this allomorphy. Thus complete symmetry is not
expected. Or to state it differently, the grammar allows space for lexical variation,
but the lexicon allots this space in the form of allomorphy. The grammar partly
controls allomorphy, but does not fully predict it. (Richness of the Base was never
intended to ll all the accidental gaps in the lexicon of a language.)
419
Residual issues

These advantages are counterbalanced by a number of potential problems. First,

it may not be wise to lump all kinds of alternations (productive and non-
productive) into a single category, given the psycholinguistic evidence. Research
in speech production gives evidence for a distinction between two kinds of alter-
nations, which can be summarized as on line production versus lexical look up.
The allomorphy model claims that alternations all depend on look up, even
though the grammar denes the notion of possible allomorph (as we saw above).
It is unclear how distinctions of productivity between alternations can be captured
in this model.
Second, related to the previous point, it may be overly optimistic to assume that
a general network of uniform exponence constraints may restrict all types of allo-
morphy in a language. For example, if the drastic shape variation in allomorphs
such as go ~ went is indicative of the general position of OO-correspondence
constraints for consonant and vowel identity in English, then it is no longer
evident that OO-correspondence does any work at all. This poses a difcult
dilemma: allomorphic variation must either be radically restricted by high-ranked
uniform exponence (losing the go ~ went case), or left more or less free (predict-
ing random allomorphic variation all over). A solution may be found in a stratied
model, in which different parts of the lexicon set their own specic balance of
faithfulness and uniform exponence. However, it is far from clear whether this is
feasible.

9.7 Conclusion: future perspectives

At the end of the nal chapter of this book, it is time to sketch some future per-
spectives of OT, partly based on developments sketched in the previous sections.
One development that can be foreseen is that an even greater role will be
assigned to surface forms in explaining alternations, moving away from abstract
underlying forms. This emphasis on surface forms has been a cornerstone of
OT in dealing with phenomena as diverse as contrast, neutralization, allophone
distribution, conspiracies, overapplication and underapplication (in reduplication
and stem-based morphology), and transderivational cyclic dependencies. Extra-
polation of this development may lead to the elimination of underlying repres-
entations in favour of an allomorphic model, as we have seen in section 9.6.
Grammatically adequate and cognitively plausible allomorphic models may be
developed which overcome the problems mentioned there. More on a speculative
note, allomorphy may also offer an alternative account of opacity, a phenomenon
for which current OT has no genuinely surface-based analysis.
A second major development that is likely to continue is a reduction of the role
of representations in favour of constraint interaction. Underspecication, once a
cornerstone of theories of segment inventories and assimilation, has come under
420
 9.7 Conclusion: future perspectives

severe pressure from OT analysts (starting with Smolensky 1993, and elaborated
by Inkelas 1995, It, Mester, and Padgett 1995, and Steriade 1995b) who argue
that constraint interactions offer superior accounts. Similar reductionist arguments
were made against representational assumptions of feature geometry (Padgett
1995, N Chiosin and Padgett 1997) and syllable structure (Steriade 1995b).
An issue closely related to that of representations concerns the boundary be-
tween phonology and phonetics. In work by Steriade, Flemming, Kirchner, Hayes,
and others, it is argued that constraints should be able to refer to much more phon-
etic detail (including non-contrastive features and numerical values of acoustic
parameters) than is allowed on classical generative assumptions, which maintain
a strict separation between phonology and phonetics. (An illustration of this
development is the title of Flemmings 1995 UCLA dissertation: Auditory repres-
entations in phonology.) This blurring of the phonologyphonetics boundary
goes hand in hand with an increased role for functional explanations. For ex-
ample, Jun (1995) and Steriade (1995b), in the footsteps of phoneticians such as
Lindblom, Kohler, and Ohala, argue that speakers make more effort to preserve
features in contexts where they are most salient. The articulatory and perceptual
basis of phonological constraints is emphasized by Archangeli and Pulleyblank
(1994), Gafos (1996), Hayes (1996a), Myers (1997b), and others. No doubt, real
progress can be made by this approach. An increased role for functional explana-
tions in grammatical theory matches well with a major goal of OT, which is to
encode directly markedness in the grammar, an enterprise that has been crucial
to OTs typological achievements. The merging of the phonetic and phonological
components into a single hierarchy of constraints comes tantalizingly close to a
single-step mapping, another authentic goal of OT.
However, freely mixing phonology and phonetics comes at a certain cost. There
is overwhelming support for a level coinciding with the output of phonology and
the input to phonetic interpretation. First, while phonological specications are
categorical (that is on or off), phonetic specication of non-distinctive features
is gradient and based on interpolation (Keating 1988, Cohn 1993a; see Kirchner
1997 for a theory of contrast in a one-step mapping model). Second, there is sur-
prisingly little evidence that grammatical generalizations refer to non-distinctive
feature specications. Third, a relatively abstract (phonetically underspecied)
output level of the phonology is conrmed as a cognitive reality by various kinds
of independent evidence (language games, speech errors, second language acqui-
sition, etc.: Fromkin 1971, Stemberger 1991). Apart from the evidence for levels,
there is yet another potential stumbling block for a direct reductionist approach.
A theory reducing phonology to interactions of raw functional (perceptual and
articulatory) factors has difculties in accounting for symmetry as a property of
phonological systems, since such factors, by denition, interact in a gradient
421
Residual issues

fashion. (But see Hayes 1996a for a theory of learning of symmetrical patterns
on the basis of gradient information.) In sum, a theory which gives up the sep-
aration of phonology and phonetics holds much explanatory potential, but also
faces two fundamental problems in dealing with cognitively plausible levels of
representation.
When numerical values (of phonetic parameters) are incorporated into phono-
logy, it seems only logical to take another step into numerical (gradient) types of
interaction, so that constraints exert inuence according to their relative strength,
expressed in an index. In several constraint-based models (for example, Harmonic
Grammar, Legendre, Miyata, and Smolensky 1990), numerical ranking indeed re-
places the principle of strict domination, partly bridging the gap with connection-
ism (Rumelhart and McLelland 1986, Goldsmith and Larson 1990). It is not clear
at all whether an OT model based on numerical ranking can preserve the advan-
tages of standard OT. For example, the phenomenon of the emergence of the
unmarked depends on strict domination. For discussion see Prince and Smolen-
sky (1993, chapter 10).
In sum, in the foreseeable future phonology may look quite different from what
it is today. It has been suggested that there are some striking resemblances
between OT and Structuralism (with respect to views of allophonic patterns and
contrast, surface patterns, functional considerations, and allomorphy), the domin-
ant framework in the pre-generative era. Is history circular? Perhaps, but if so,
there are important differences from the earlier days. Due to the generative legacy,
there remains a strong emphasis on formal precision in grammatical analysis,
combined with the necessity to restrict the descriptive power of linguistic theory.
Both theoretical priorities are solidly integrated into OT. Moreover, in-depth
phonological analyses of many languages have given us a much better insight
into cross-linguistic tendencies and typology than we had half a century ago.
The emphasis on formal accuracy and explanatory adequacy, coupled to a still
expanding typological basis, provides the background that is necessary to move
successfully forward into areas that were previously considered to be too surface-
oriented for phonologists to deal with. The concept of grammar is strong enough
to survive this continuing enterprise.

SUGGESTIONS FOR FUR THER READING

Opacity and OT
Archangeli, Diana and Keiichiro Suzuki (1997) The Yokuts challenge. In Roca.
197226.

422
 9.7 Conclusion:9.7
future
Further
perspectives
reading

Idsardi, William J. (1997) Phonological derivations and historical changes in

Hebrew spirantization. In Roca. 36792.
McCarthy, John (1998) Sympathy and phonological opacity. Ms., University of
Massachusetts, Amherst. [ROA-252, http://ruccs.rutgers.edu /roa.html]
Orgun, C. Orhan (1995) Correspondence and identity constraints in two-level
Optimality Theory. In J. Camacho, L. Choueiri, and M. Watanabe (eds.),
Proceedings of the Fourteenth West Coast Conference on Formal Linguistics.
Stanford: CSLI. 399413.

Local Conjunction
Alderete, John (1997) Dissimilation as Local Conjunction. In K. Kusumoto (ed.)
Proceedings of the North East Linguistic Society 27. 1731. [ROA-175,
http://ruccs.rutgers.edu/roa.html]
Kirchner, Robert (1996) Synchronic chain shifts in Optimality Theory. Linguistic
Inquiry 27. 34150.

Absolute ungrammaticality
Orgun, C. Orhan and Ronald Sprouse (1997) From M-Parse to Control: deriv-
ing ungrammaticality. Ms., University of California, Berkeley. [ROA-224,
http://ruccs.rutgers.edu/roa.html]
Raffelsiefen, Renate (1996) Gaps in word formation. In U. Kleinhenz (ed.), Inter-
faces in phonology. Studia Grammatica 41. Berlin: Akademie Verlag. 194209.

Free variation
Anttila, Arto (1995) Deriving variation from grammar: a study of Finnish genit-
ives. Ms., Stanford University. [ROA-63, http://ruccs.rutgers.edu/roa.html]
Kager, Ren (1997b) Generalized alignment and morphological parsing. Rivista
di Linguistica 9. 24582.

Positional faithfulness
Beckman, Jill (1997a) Positional faithfulness, positional neutralization and Shona
vowel harmony. Phonology 14. 146.
Lombardi, Linda (1995b) Why Place and Voice are different: constraint inter-
actions and feature faithfulness in Optimality Theory. Ms., University of
Maryland, College Park. [ROA-105, http://ruccs.rutgers.edu/roa.html]

Allomorphy
Burzio, Luigi (forthcoming) Cycles, regularization, and correspondence. To
appear in J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimal-
ity Theory: syntax, phonology, and acquisition. Oxford University Press.

423
Residual issues

Hayes, Bruce (forthcoming) Anticorrespondence in Yidi. To appear in B. Her-

mans and M. van Oostendorp (eds.), The derivational residue in phonology.
Amsterdam: John Benjamins.
Kager, Ren (1996) On afx allomorphy and syllable counting. In U. Kleinhenz
(ed.), Interfaces in phonology. Studia Grammatica 41. Berlin: Akademie Verlag.
15571.

Functional approaches to phonology

Hayes, Bruce (1996a) Phonetically driven phonology: the role of Optimality
Theory and inductive grounding. Ms., University of California, Los Angeles.
[ROA-158, http://ruccs.rutgers.edu/roa.html]
Kirchner, Robert (1997) Contrastiveness and faithfulness. Phonology 14. 83 111.
Myers, Scott (1997b) Expressing phonetic naturalness in phonology. In Roca.
125-12.

424
 REFERENCES

Abu-Salim, Issam M. (1980) Epenthesis and geminate consonants in Palestinian Arabic.

Studies in the Linguistic Sciences 10. 111.
Ackema, Peter and Ad Neeleman (1998) Optimal questions. Natural Language and
Linguistic Theory 16. 443 90.
Alber, Birgit (1997) Quantity sensitivity as the result of constraint interaction. In
G. E. Booij and J. van de Weijer (eds.), Phonology in progress: progress in phonology.
HIL phonology papers III. The Hague: Holland Academic Graphics, 1 45.
Alderete, John (1995) Faithfulness to prosodic heads. Ms., University of Massachusetts,
Amherst. [ROA-94, http://ruccs.rutgers.edu/roa.html]
(1997) Dissimilation as local conjunction. In K. Kusumoto (ed.), Proceedings of the
North East Linguistic Society 27. 1731.
Alderete, John, Jill Beckman, Laura Benua, Amalia Gnanadesikan, John McCarthy, and
Suzanne Urbanczyk (1997) Reduplication with xed segmentism. Ms., University of
Massachusetts, Amherst. [ROA-226, http://ruccs.rutgers.edu/roa.html]
Allison, E. (1979) The phonology of Sibutu Sama: a language of the Southern Philippines.
In C. Edrial-Luzares and A. Hale (eds.), Studies in Philippine Languages 3:2.
Linguistic Society of the Philippines and Summer Institute of Linguistics.
Anderson, Stephen R. (1974) The organization of phonology. New York: Academic Press.
(1975) On the interaction of phonological rules of various types. Journal of Linguistics
11. 39 62.
(1982) Differences in rule type and their structural basis, in H. van der Hulst and
N. Smith (eds.), The structure of phonological representations, part 2. Dordrecht:
Foris. 126.
(1985) Phonology in the twentieth century: theories of rules and theories of
representations. University of Chicago Press.
Anttila, Arto (1995) Deriving variation from grammar: a study of Finnish genitives. Ms.,
Stanford University. [ROA-63, http://ruccs.rutgers.edu/roa.html]
Archangeli, Diana (1988) Aspects of underspecication theory. Phonology 5. 183 207.
(1997) Optimality Theory: an introduction to linguistics in the 1990s. In D. Archangeli
and D. T. Langendoen (eds.), Optimality Theory: an overview. Oxford: Blackwell.
132.

425
References

Archangeli, Diana and Douglas Pulleyblank (1994) Grounded phonology. Cambridge,

Mass.: MIT Press.
Archangeli, Diana and Keiichiro Suzuki (1997) The Yokuts challenge. In Roca. 197226.
rnason, Kristjn (1980) Quantity in historical phonology: Icelandic and related cases.
Cambridge University Press.
Aronoff, Mark (1976) Word formation in generative grammar. Cambridge, Mass.: MIT
Press.
Aronoff, Mark, Azhar Arsyad, Hassan Basri, and Ellen Broselow (1987) Tier conation in
Makassarese reduplication. In A. Bosch, B. Need, and E. Schiller (eds.), Papers from
the Chicago Linguistic Society 23:2. 115.
Austin, Peter (1981) A grammar of Diyari, South Australia. Cambridge University Press.
Bat-El, Outi (1996) Selecting the best of the worst: the grammar of Hebrew blends.
Phonology 13. 283 328.
Beckman, Jill (1997a) Positional faithfulness, positional neutralisation and Shona vowel
harmony. Phonology 14. 1 46.
(1997b) Positional faithfulness. PhD dissertation, University of Massachusetts, Amherst.
[ROA-234, http://ruccs.rutgers.edu/roa.html]
Beckman, Jill, Laura Walsh Dickey, and Suzanne Urbanczyk (eds.) (1995) Papers in
Optimality Theory. University of Massachusetts Occasional Papers in Linguistics 18.
Amherst, Mass.: Graduate Linguistic Student Association.
Benton, Richard (1971) Pangasinan reference grammar. Honolulu: University of Hawaii
Press.
Benua, Laura (1995) Identity effects in morphological truncation. In Beckman, Walsh
Dickey, and Urbanczyk. 77136.
(1997) Transderivational identity: phonological relations between words.
PhD dissertation, University of Massachusetts, Amherst. [ROA-259,
http://ruccs.rutgers.edu/roa.html]
Bird, Steven (1990) Constraint-based phonology. PhD dissertation, University of
Edinburgh.
Blevins, Juliette (1995) The syllable in phonological theory. In Goldsmith. 206 44.
(1996) Mokilese reduplication. Linguistic Inquiry 27. 52330.
(1997) Rules in Optimality Theory: two case studies. In Roca. 227 60.
Boersma, Paul (1997) How we learn variation, optionality, and probability. Ms., University
of Amsterdam. [ROA-221, http://ruccs.rutgers.edu/roa.html]
Booij, Geert E. (1986) Icelandic vowel lengthening and prosodic phonology. In F. Beukema
and A. Hulk (eds.), Linguistics in the Netherlands 1986. Dordrecht: Foris. 9 18.
Booij, Geert E. and Jerzy Rubach (1987) Postcyclic versus postlexical rules in lexical
phonology. Linguistic Inquiry 18. 1 44.
Borgstrm, C. Hj. (1937) The dialect of Barra in the Outer Hebrides. Norsk Tidsskrift for
Sprogvidenskap 8. 71242.
Borowsky, Toni (1993) On the word level. In S. Hargus and E. Kaisse (eds.), Studies in
Lexical Phonology. San Diego: Academic Press. 199 234.

426
 References

Borowsky, Toni and Mark Harvey (1997) Vowel length in Warray and weight identity.
Phonology 14. 16175.
Boxwell, Helen and Maurice Boxwell (1966) Weri phonemes. In S. A. Wurm (ed.), Papers
in New Guinea linguistics, V. Linguistic Circle of Canberra Publications, Series A,
No. 7. Canberra: Australian National University. 7793.
Brame, Michael (1974) The cycle in phonology: stress in Palestinian, Maltese, and
Spanish. Linguistic Inquiry 5. 39 60.
Bresnan, Joan (forthcoming) Optimal syntax. To appear in J. Dekkers, F. van der Leeuw,
and J. van de Weijer (eds.), Optimality Theory: phonology, syntax, and acquisition.
Oxford University Press.
Broekhuis, Hans (forthcoming) Against feature strength. To appear in R. van Bezooijen
and R. Kager (eds.), Linguistics in the Netherlands 1998. Amsterdam: John
Benjamins.
Broihier, Kevin (1995) Optimality theoretic rankings with tied constraints: Slavic relatives,
resumptive pronouns and learnability. Ms., Department of Brain and Cognitive
Sciences, MIT. [ROA-46, http://ruccs.rutgers.edu/roa.html]
Bromberger, Sylvain and Morris Halle (1989) Why phonology is different. Linguistic
Inquiry 20. 5170.
Broselow, Ellen (1992) Parametric variation in Arabic dialect phonology. In E. Broselow,
M. Eid, and J. McCarthy (eds.), Perspectives on Arabic linguistics IV: Papers from the
fourth annual symposium on Arabic linguistics. Amsterdam and Philadelphia: John
Benjamins. 7 45.
(1995) Skeletal positions and moras. In Goldsmith. 175205.
Broselow, Ellen, Su-I Chen, and Marie Huffman (1997) Syllable weight: convergence of
phonology and phonetics. Phonology 14. 47 82.
Broselow, Ellen and John McCarthy (1983) A theory of internal reduplication. The
Linguistic Review 3. 25 89.
Buckley, Eugene (1995) Cyclicity and correspondence. Handout of paper presented at the
Tilburg conference The derivational residue in phonology, Fall 1995. [ROA-93c,
http://ruccs.rutgers.edu/roa.html]
Burzio, Luigi (1994) Principles of English stress. Cambridge University Press.
(1995) The rise of Optimality Theory. Glot International 1:6. 37.
(1996) Surface constraints versus Underlying Representation. In J. Durand and B. Laks
(eds.), Current trends in phonology: models and methods. CNRS, Paris X, and
University of Salford. University of Salford Publications. 97122.
(forthcoming) Cycles, regularization, and correspondence. To appear in J. Dekkers,
F. van der Leeuw, and J. van de Weijer (eds.), Optimality Theory: syntax, phonology,
and acquisition. Oxford University Press.
Bybee, Joan L. (1988) Morphology as lexical organization. In M. Hammond and
M. Noonan (eds.), Theoretical morphology. San Diego: Academic Press. 119 41.
(1995) Regular morphology and the lexicon. Language and Cognitive Processes 10.
425 55.

427
References

Bye, Patrik (1996) Correspondence in the prosodic hierarchy and the grid: case studies in
overlength and level stress. Cand. Philol. thesis, University of Troms.
Chomsky, Noam (1957) Syntactic structures. The Hague: Mouton.
(1965) Aspects of the theory of syntax. Cambridge, Mass.: MIT Press.
(1981a) Markedness and core grammar. In A. Belletti, L. Brandi, and L. Rizzi (eds.),
Theory of markedness in generative grammar. Proceedings of the 1979 GLOW
conference. Pisa: Scuola Normale Superiore di Pisa. 123 46.
(1981b) Lectures on government and binding. Dordrecht: Foris.
(1991) Some notes on economy of derivation and representation. In R. Freidin (ed.),
Principles and parameters in comparative grammar. Cambridge, Mass.: MIT Press.
4158.
(1993) A minimalist program for linguistic theory. In K. Hale and S. J. Keyser (eds.),
The view from Building 20. Cambridge, Mass.: MIT Press. 4158.
(1995) The minimalist program. Cambridge, Mass.: MIT Press.
Chomsky, Noam and Morris Halle (1968) The sound pattern of English. New York: Harper
and Row.
Churchward, C. Maxwell (1953) Tongan grammar. Oxford University Press.
Clements, George N. (1985a) The problem of transfer in nonlinear morphology. Cornell
Working Papers in Linguistics 7. 38 73.
(1985b) The geometry of phonological features. In C. Ewen and E. Kaisse (eds.),
Phonology Yearbook 2. Cambridge University Press. 225 52.
(1986) Syllabication and epenthesis in the Barra dialect of Gaelic. In K. Bogers,
M. Maus, and H. van der Hulst (eds.), The phonological representation of
suprasegmentals. Dordrecht: Foris. 31736.
(1990) The role of the sonority cycle in core syllabication. In J. Kingston and M.
Beckman (eds.), Papers in laboratory phonology I: Between the grammar and physics
of speech. Cambridge University Press. 283 333.
Clements, George N. and Samuel J. Keyser (1983) CV phonology: a generative theory of
the syllable. Cambridge, Mass.: MIT Press.
Cohn, Abigail (1989) Stress in Indonesian and bracketing paradoxes. Natural Language
and Linguistic Theory 7. 167216.
(1993a) Nasalisation in English. Phonology 10. 4381.
(1993b) The status of nasalized continuants. In Huffman and Krakow. 329 68.
Cohn, Abigail and John McCarthy (1994) Alignment and parallelism in Indonesian
phonology. Ms., Cornell University, Ithaca, N. Y. and University of Massachusetts,
Amherst. [ROA-25, http://ruccs.rutgers.edu/roa.html]
Cole, Jennifer and Charles Kisseberth (1995) Restricting multi-level constraint evaluation:
opaque rule interactions in Yawelmani vowel harmony. Ms., University of Illinois,
Urbana-Champaign. [ROA-98, http://ruccs.rutgers.edu/roa.html]
Colina, Sonia (1996) Spanish truncation processes: the emergence of the unmarked.
Linguistics 34. 1199 218.
Crowhurst, Megan and Mark Hewitt (1995) Prosodic overlay and headless feet in Yidi,
Phonology 12. 39 84.

428
 References

(1997) Boolean operations and constraint interaction in Optimality Theory. Ms.,

University of North Carolina, Chapel Hill and Brandeis University. [ROA-229,
http://ruccs.rutgers.edu/roa.html]
Daelemans, Walter, Steven Gillis, and Gert Durieux (1994) The acquisition of stress: a
data-oriented approach. Computational Linguistics 20. 42151.
Davis, Stuart (1995) Emphasis spread in Arabic and grounded phonology. Linguistic
Inquiry 26. 46598.
Demuth, Katharine (1995) Markedness and the development of phonological structure.
In J. Beckman (ed.), Proceedings of the North East Linguistic Society 25:2.
1325.
Derbyshire, Desmond C. (1979) Hixkaryana. Lingua Descriptive Studies 1. Amsterdam:
North-Holland.
Dixon, R. M. W. (1977) A grammar of Yidis. Cambridge University Press.
(1981) Wargamay. In R. M. W. Dixon and B. J. Blake (eds.), Handbook of Australian
languages, II. Amsterdam: John Benjamins. 1144.
(1988) A grammar of Boumaa Fijian. University of Chicago Press.
Downing, Laura J. (forthcoming) Verbal reduplication in three Bantu languages. To appear
in Kager, van der Hulst, and Zonneveld.
Drachman, Gaberell and Angeliki Malikouti-Drachman (1973) Studies in the acquisition
of Greek as a native language, I: Some preliminary ndings on phonology. Ohio State
Working Papers in Linguistics 15. 99 114.
Dresher, Elan and Jonathan Kaye (1990) A computational learning model for metrical
phonology. Cognition 34. 13795.
Dresher, Elan and Aditi Lahiri (1991) The Germanic foot: metrical coherence in Old
English. Linguistic Inquiry 22. 251 86.
Dudas, Karen (1976) The phonology and morphology of Modern Javanese. PhD
dissertation, University of Illinois, Urbana-Champaign.
Dunlop, Elaine (1987) English [] tensing in Lexical Phonology. Ms., University of
Massachusetts, Amherst.
Eisner, Jason (1997) Efcient generation in primitive Optimality Theory. Proceedings of
the 35th Annual Meeting of the Association for Computational Linguistics. [ROA-206,
http://ruccs.rutgers.edu/roa.html]
Elenbaas, Nine (1996) Ternary rhythm in Sentani. In C. Cremers and M. den Dikken (eds.),
Linguistics in the Netherlands 1996. Amsterdam: John Benjamins. 6172.
(1999) A unied account of binary and ternary stress: considerations from Sentani and
Finnish. PhD dissertation, Utrecht University.
Ellison, Mark T. (1994) Phonological derivation in Optimality Theory. Proceedings of the
15th International Conference on Computational Linguistics, II. Kyoto, Japan. 1007
13. [ROA-75, http://ruccs.rutgers.edu/roa.html]
Ferguson, Charles (1975) Short a in Philadelphia English. In M. E. Smith (ed.), Studies
in linguistics: in honor of George L. Trager. The Hague: Mouton. 250 74.
Flemming, Edward (1995) Auditory representations in phonology. PhD dissertation,
University of California, Los Angeles.

429
References

Frank, Robert and Giorgio Satta (1998) Optimality Theory and the generative
complexity of constraint violability. Computational Linguistics 24. 30715.
[ROA-228, http://ruccs.rutgers.edu/roa.html]
French, Koleen M. (1988) Insights into Tagalog: reduplication, inxation, and stress from
nonlinear phonology. Dallas: Summer Institute of Linguistics and University of Texas
at Arlington.
Fromkin, Victoria (1971) The non-anomalous nature of anomalous utterances. Language
47. 2752.
Furby, Christine (1974) Garawa phonology. Canberra: Australian National University.
Gafos, Adamantios (1996) The articulatory basis of locality in phonology. PhD
dissertation, Johns Hopkins University.
Gnanadesikan, Amalia (1995) Markedness and faithfulness constraints in child
phonology. Ms., University of Massachusetts, Amherst. [ROA-67,
http://ruccs.rutgers.edu/roa.html]
Goad, Heather (1997) Consonant harmony in child language. In S. J. Hannahs and
M. Young-Scholten (eds.), Focus on phonological acquisition. Amsterdam: John
Benjamins. 113 42.
Goldsmith, John (1976) Autosegmental phonology. PhD dissertation, MIT. [Published
1979, New York: Garland.]
(1990) Autosegmental and metrical phonology. Oxford: Blackwell.
(1993a) Harmonic phonology. In Goldsmith (1993b). 21 60.
(ed.) (1993b) The last phonological rule: reections on constraints and derivations.
University of Chicago Press.
(ed.) (1995) The handbook of phonological theory. Oxford: Blackwell.
Goldsmith, John and Gary Larson (1990) Local modeling and syllabication. In M.
Ziolkowski, M. Noske, and K. Deaton (eds.), Papers from the Chicago Linguistic
Society 26:2. 129 41.
Golston, Chris (1995) Syntax outranks phonology: evidence from ancient Greek.
Phonology 12. 343 68.
(1996) Direct Optimality Theory: representation as pure markedness. Language 72.
713 48.
Greenberg, Joseph (1978) Some generalizations concerning initial and nal consonant
clusters. In J. Greenberg (ed.), Universals of human language, II: Phonology. Stanford
University Press. 243 80.
Grimshaw, Jane (1991) Extended projection. Ms., Brandeis University, Waltham, Mass.
(1997) Projection, heads, and optimality. Linguistic Inquiry 28. 373 422.
Gupta, Prahlad and David Touretzky (1994) Connectionist models and linguistic theory:
investigations of stress systems in language. Cognitive Science 18. 150.
Haas, Mary (1940) Tunica. Handbook of American Indian languages, IV. Washington, D.C.:
Smithsonian Institute, Bureau of American Ethnography.
Haas, Wim de (1988) A formal theory of vowel coalescence: a case study of Ancient Greek.
PhD dissertation, University of Nijmegen. Dordrecht: Foris.

430
 References

Hale, Mark and Charles Reiss (1997) Grammar optimization: the simultaneous acquisition
of constraint ranking and a lexicon. Ms., Concordia University, Montreal. [ROA-231,
http://ruccs.rutgers.edu/roa.html]
Halle, Morris and George N. Clements (1983) Problem book in phonology. Cambridge,
Mass.: MIT Press.
Halle, Morris and Jean-Roger Vergnaud (1987) An essay on stress. Cambridge, Mass.: MIT
Press.
Hammond, Michael (1984) Constraining metrical theory: a modular theory of rhythm and
destressing. PhD dissertation, University of California, Los Angeles. [Published 1988,
New York: Garland.]
(1994) An OT account of variability in Walmatjarri stress. Ms., University of Arizona.
[ROA-20, http://ruccs.rutgers.edu/roa.html]
(1997) Parsing syllables: modeling OT computationally. Ms., University of Arizona.
[ROA-222, http://ruccs.rutgers.edu/roa.html]
Hansen, K. C. and L. E. Hansen (1969) Pintupi phonology. Oceanic Linguistics 8.
153 70.
Harris, James (1983) Syllable structure and stress in Spanish: a nonlinear analysis.
Cambridge, Mass.: MIT Press.
Harris, John (1990) Derived phonological contrasts. In S. Ramsaran (ed.), Studies in the
pronunciation of English: a commemorative volume in honour of A. C. Gimson.
London: Routledge. 87105.
Hawkins, John and Anne Cutler (1988) Psycholinguistic factors in morphological
asymmetry. In J. Hawkins (ed.), Explaining language universals. Oxford: Blackwell.
280 317.
Hayes, Bruce (1980) A metrical theory of stress rules. PhD dissertation, MIT. [Published
1985, New York: Garland.]
(1985) Iambic and trochaic rhythm in stress rules. In M. Niepokuj, M. VanClay, V.
Nikiforidou, and D. Jeder (eds.), Proceedings of the Berkeley Linguistics Society 11.
429 46.
(1986) Inalterability in CV phonology. Language 62. 32151.
(1987) A revised parametric metrical theory. In J. McDonough and B. Plunket (eds.),
Proceedings of the North East Linguistic Society 17. 274 89.
(1989) Compensatory lengthening in moraic phonology. Linguistic Inquiry 20. 253 306.
(1995) Metrical stress theory: principles and case studies. University of Chicago Press.
(1996a) Phonetically driven phonology: the role of Optimality Theory and inductive
grounding. Ms., University of California, Los Angeles. [ROA-158,
http://ruccs.rutgers.edu/roa.html]
(1996b) Constraint ranking software. Ms., University of California, Los Angeles.
[Manual of computer program ranking OT constraints, on line available from
http://www.humnet.ucla.edu/humnet/linguistics/people/hayes/otsoft/otsoft.htm]
(forthcoming) Anticorrespondence in Yidi. To appear in B. Hermans and M. van
Oostendorp (eds.), The derivational residue in phonology. Amsterdam: John Benjamins.

431
References

Hayes, Bruce and May Abad (1989) Reduplication and syllabication in Ilokano. Lingua
77. 33174.
Hayes, Bruce and Tanya Stivers (1995) A phonetic account of postnasal voicing. Ms.,
University of California, Los Angeles.
Healey, Phyllis M. (1960) An Agta grammar. Manila: Summer Institute of Linguistics.
Hjelmslev, Louis (1935) La Catgorie des Cas I. Acta Jutlandica 7:1. Aarhus. [Reissued
1972, Copenhagen and Munich.]
Hooper [Bybee], Joan L. (1976) An introduction to Natural Generative Phonology.
New York: Academic Press.
Hore, Michael (1981) Syllable length and stress in Nunggubuyu. In B. Waters (ed.),
Australian phonologies: collected papers. Work Papers of SIL-AAB, A5. Darwin:
Summer Institute of Linguistics.
Huffman, Marie K. (1993) Phonetic patterns of nasalization and implications for feature
specication. In Huffman and Krakow. 303 27.
Huffman, Marie K. and Rena A. Krakow (eds.) (1993) Phonetics and phonology V: Nasals,
nasalization, and the velum. San Diego: Academic Press.
Hulst, Harry van der (1984) Syllable structure and stress in Dutch. Dordrecht: Foris.
Hulst, Harry van der and Ellis Visch (1992) Iambic lengthening in Carib. In R. Bok-
Bennema and R. van Hout (eds.), Linguistics in the Netherlands 1992. Amsterdam:
John Benjamins. 113 24.
Hume, Elizabeth (1992) Front vowels, coronal consonants and their interaction in nonlinear
phonology. PhD dissertation, Cornell University, Ithaca, N.Y.
(1995) Beyond linear order: prosodic constraints and C/V metathesis. Proceedings of the
Formal Linguistics Society of the Midwest 6. Bloomington: Indiana University
Linguistics Club.
Hung, Henrietta (1994) The rhythmic and prosodic organization of edge constituents.
PhD dissertation, Brandeis University, Waltham, Mass. [ROA-24,
http://ruccs.rutgers.edu/roa.html]
Hyams, Nina (1986) Language acquisition and the theory of parameters. Dordrecht:
Reidel.
Hyman, Larry (1985) A theory of phonological weight. Dordrecht: Foris.
Idsardi, William J. (1997) Phonological derivations and historical changes in Hebrew
spirantization. In Roca. 367 92.
Inkelas, Sharon (1989) Prosodic constituency in the lexicon. PhD dissertation, Stanford
University.
(1995) The consequences of optimization for underspecication. In J. Beckman (ed.),
Proceedings of the North East Linguistic Society 25. 287302.
(1996) Dominant afxes and the phonologymorphology interface. In U. Kleinhenz
(ed.), Interfaces in phonology. Studia Grammatica 41. Berlin: Akademie Verlag.
128 54.
(forthcoming) Exceptional stress-attracting sufxes in Turkish: representations vs.
the grammar. To appear in Kager, Van der Hulst, and Zonneveld. [ROA-39,
http://ruccs.rutgers.edu/roa.html]

432
 References

Inkelas, Sharon and C. Orhan Orgun (1995) Level ordering and economy in the lexical
phonology of Turkish. Language 71. 763 93.
Inkelas, Sharon, C. Orhan Orgun, and Cheryl Zoll (1997) The implications of lexical
exceptions for the nature of grammar. In Roca. 393 418.
It, Junko (1986) Syllable theory in prosodic phonology. PhD dissertation, University of
Massachusetts, Amherst. [Published 1988, New York: Garland.]
(1989) A prosodic theory of epenthesis. Natural Language and Linguistic Theory 7.
21760.
(1990) Prosodic minimality in Japanese. In M. Ziolkowski, M. Noske, and K. Deaton
(eds.), Papers from Chicago Linguistic Society 26:2. 213 39.
It, Junko, Yoshihisa Kitagawa, and R. Armin Mester (1996) Prosodic faithfulness and
correspondence: evidence from a Japanese argot. Journal of East Asian Linguistics 5.
21794.
It, Junko and R. Armin Mester (1992) Weak layering and word binarity. Ms., University
of California, Santa Cruz. Linguistics Research Center Report No. 92 09.
(1995) The core-periphery structure of the lexicon and constraints on reranking. In
Beckman, Walsh Dickey, and Urbanczyk. 181210.
(1997) Correspondence and compositionality: the Gagyo variation in Japanese
phonology. In Roca. 419 62.
(1998) Markedness and word structure: OCP effects in Japanese. Ms., University of
California, Santa Cruz. [ROA-255, http://ruccs.rutgers.edu/roa.html]
(forthcoming) Realignment. To appear in Kager, Van der Hulst, and Zonneveld.
It, Junko, R. Armin Mester, and Jaye Padgett (1995) Licensing and underspecication in
Optimality Theory. Linguistic Inquiry 26. 571 614.
Jacobsen, William (1964) A grammar of the Washo language. PhD dissertation, University
of California, Berkeley.
Jakobson, Roman (1941) Kindersprache, Aphasie und allgemeine Lautgesetze. Uppsala.
[Reissued in Jakobson 1962. 328 401.]
(1962) Selected writings. The Hague: Mouton.
Jun, Jongho (1995) Perceptual and articulatory factors in place assimilation: an Optimality
Theoretic approach. PhD dissertation, University of California, Los Angeles.
Kager, Ren (1989) A metrical theory of stress and destressing in English and Dutch.
Dordrecht: Foris.
(1992a) Shapes of the generalized trochee. In J. Mead (ed.), Proceedings of the West
Coast Conference on Formal Linguistics 11. 298 311.
(1992b) Are there any truly quantity-insensitive systems? In L. A. Buszard-Welcher,
L. Lee, and W. Weigel (eds.), Proceedings of the Berkeley Linguistics Society 18.
12332.
(1993) Alternatives to the iambic-trochaic law. Natural Language and Linguistic Theory
11. 381 432.
(1994) Ternary rhythm in alignment theory. Ms., Utrecht University. [ROA-35,
http://ruccs.rutgers.edu/roa.html]
(1995) Review of Hayes 1995. Phonology 12. 437 64.

433
References

(1996) On afx allomorphy and syllable counting. In U. Kleinhenz (ed.), Interfaces

in phonology. Studia Grammatica 41. Berlin: Akademie Verlag. 155 71. [ROA-88,
http://ruccs.rutgers.edu/roa.html]
(1997a) Rhythmic vowel deletion in Optimality Theory. In Roca. 463 99.
(1997b) Generalized Alignment and morphological parsing. Rivista di Linguistica 9.
245 82.
(forthcoming) Surface opacity of metrical structure in Optimality Theory. To appear in
B. Hermans and M. van Oostendorp (eds.), The derivational residue in phonology.
Amsterdam: John Benjamins. [ROA-207, http://ruccs.rutgers.edu/roa.html]
Kager, Ren, Harry van der Hulst, and Wim Zonneveld (eds.) (forthcoming) The prosody
morphology interface. Cambridge University Press.
Kahn, Daniel (1976) Syllable-based generalizations in English phonology. PhD
dissertation, MIT. [Published 1980, New York: Garland.]
Karttunen, Lauri (1993) Finite-state constraints. In Goldsmith. 173 94.
(1998) The proper treatment of optimality in computational phonology. Ms., Xerox
Research Centre Europe, Meylan, France. [ROA-258, http://ruccs.rutgers.edu/roa.html]
Kaye, Jonathan (1974) Opacity and recoverability in phonology. Canadian Journal of
Linguistics 19. 134 49.
(1990) Phonology: a cognitive view. Hillsdale, N.J.: LEA.
Kaye, Jonathan and Jean Lowenstamm (1981) Syllable structure and markedness theory.
In A. Belletti, L. Brandi, and L. Rizzi (eds.), Theory of markedness in generative
grammar: proceedings of the 1979 GLOW conference. Pisa: Scuola Normale
Superiore di Pisa. 287315.
Kaye, Jonathan, Jean Lowenstamm, and Jean-Roger Vergnaud (1990) Constituent structure
and government in phonology. Phonology 7. 193231.
Kean, Mary-Louise (1975) The theory of markedness in generative grammar. PhD
dissertation, MIT.
(1981) On a theory of markedness: some general considerations and a case in point.
In A. Belletti, L. Brandi, and L. Rizzi (eds.), Theory of markedness in generative
grammar: proceedings of the 1979 GLOW conference. Pisa: Scuola Normale
Superiore di Pisa. 559 604.
Keating, Patricia (1988) Underspecication in phonetics. Phonology 5. 27597.
Keer, Edward and Eric Bakovic (1997) Have Faith in syntax. Proceedings of the
West Coast Conference on Formal Linguistics 16. [ROA-200,
http://ruccs.rutgers.edu/roa.html]
Kenstowicz, Michael (1981) Functional explanations in generative phonology. In
D. L. Goyvaerts (ed.), Phonology in the 1980s. Ghent: E. Story-Scientia. 431 43.
(1994a) Phonology in generative grammar. Oxford: Blackwell.
(1994b) Syllabication in Chukchee: a constraints-based analysis. In A. Davison,
N. Maier, G. Silva, and W. S. Yan (eds.), Proceedings of the Formal Linguistics
Society of the Midwest 4. Iowa City: Department of Linguistics, University of Iowa.
160 81.
(1995) Cyclic vs. non-cyclic constraint evaluation. Phonology 12. 397 436.

434
 References

(1996) Base-identity and uniform exponence: alternatives to cyclicity. In J. Durand and

B. Laks (eds.), Current trends in phonology: models and methods. CNRS, Paris X,
and University of Salford. University of Salford Publications. 36393. [ROA-103,
http://ruccs.rutgers.edu/roa.html]
Kenstowicz, Michael and Kamal Abdul-Karim (1980) Cyclic stress in Levantine Arabic.
Studies in the Linguistic Sciences 10. 55 76.
Kenstowicz, Michael and Charles Kisseberth (1977) Topics in phonological theory.
New York: Academic Press.
(1979) Generative phonology: description and theory. New York: Academic Press.
Kiparsky, Paul (1973) Abstractness, opacity, and global rules. In O. Fujimura (ed.), Three
dimensions of linguistic theory. Tokyo: Taikusha. 5786.
(1982a) Analogical change as a problem for linguistic theory. In P. Kiparsky,
Explanation in phonology. Dordrecht: Foris. 21736.
(1982b) From cyclic phonology to Lexical Phonology. In H. van der Hulst and
N. Smith (eds.), The structure of phonological representations, part 2. Dordrecht:
Foris. 13176.
(1984) On the lexical phonology of Icelandic. In C. C. Elert, I. Johansson, and
E. Strangert (eds.), Nordic prosody III: Papers from a symposium. Stockholm:
Almqvist and Wiksell. 135 64.
(1985) Some consequences of Lexical Phonology. In C. Ewen and E. Kaisse (eds.),
Phonology Yearbook 2. Cambridge University Press. 85138.
(1986) The phonology of reduplication. Ms., Stanford University.
(1993a) Blocking in nonderived environments. In S. Hargus and E. Kaisse (eds.),
Phonetics and phonology, IV: Studies in Lexical Phonology. San Diego: Academic
Press. 277313.
(1993b) Variable rules. Handout of presentation at the Rutgers Optimality Workshop
(ROW1).
(1994) Remarks on markedness. Paper presented at TREND 2.
Kirchner, Robert (1991) Phonological processes without phonological rules: Yidi
apocope and penultimate lengthening. In T. Sherer (ed.), Proceedings of the North
East Linguistic Society 21. 20316.
(1995) Contrastiveness is an epiphenomenon of constraint ranking. Proceedings of the
Berkeley Linguistics Society 21. 198208.
(1996) Synchronic chain shifts in Optimality Theory. Linguistic Inquiry 27. 34150.
(1997) Contrastiveness and faithfulness. Phonology 14. 83111.
Kisseberth, Charles (1969) On the abstractness of phonology: the evidence from
Yawelmani. Papers in Linguistics 1. 248 82.
(1970) On the functional unity of phonological rules. Linguistic Inquiry 1. 291306.
(1973) Is rule ordering necessary in phonology? In B. Kachru (ed.), Issues in linguistics:
papers in honor of Henry and Rene Kahane. Urbana: University of Illinois Press.
418 41.
Klein, Thomas B. (1997) Output constraints and prosodic correspondence in Chamorro
reduplication. Linguistic Inquiry 28. 70715.

435
References

Koskenniemi, Kimmo (1983) Two-level morphology: a general computational model

for word-form recognition and production. Publication 11, Department of General
Linguistics, University of Helsinki.
Kury(owicz, Jerzy (1949) La nature des procs dits analogiques. Acta Linguistica 5,
1537. Reprinted in E. Hamp, F. Householder, and R. Austerlitz (eds.) (1966),
Readings in linguistics II. University of Chicago Press.
Ladefoged, Peter and Ian Maddieson (1996) The sounds of the worlds languages. Oxford:
Blackwell.
Lakoff, George (1993) Cognitive phonology. In Goldsmith. 173 94.
Lamontagne, Greg and Keren Rice (1995) A correspondence account of coalescence.
In Beckman, Walsh Dickey, and Urbanczyk. 21123.
Lapoliwa, Hans (1981) A generative approach to the phonology of Bahasa Indonesia.
Canberra: Pacic Linguistics D 34.
Law, Howard (1958) Morphological structure of Isthmus Nahuat. International Journal
of American Languages 24. 108 29.
Leben, William (1973) Suprasegmental phonology. PhD dissertation, MIT.
Legendre, Graldine, Y. Miyata, and Paul Smolensky (1990) Harmonic Grammar a
formal multi-level connectionist theory of linguistic well-formedness: theoretical
foundations. Proceedings of the Twelfth Annual Conference of the Cognitive Science
Society. Cambridge, Mass: Lawrence Erlbaum. 388 95.
Legendre, Graldine, William Raymond, and Paul Smolensky (1993) An optimality-
theoretic typology of case and grammatical voice systems. Proceedings of the
Berkeley Linguistics Society 19. 464 78.
Leslau, Wolf (1958) The verb in Harari. University of California Publications in Semitic
Philology 21. Berkeley: University of California Press.
Liberman, Mark (1975) The intonational system of English. PhD dissertation, MIT.
Liberman, Mark and Alan Prince (1977) On stress and linguistic rhythm. Linguistic Inquiry
8. 249 336.
Lichtenberk, Frantisek (1983) A grammar of Manam. Honolulu: University of Hawaii
Press.
Lloret, Maria-Rosa (1988) Gemination and vowel length in Oromo morphophonology.
PhD dissertation, Indiana University.
Lombardi, Linda (1995a) Laryngeal neutralization and alignment. In Beckman, Walsh
Dickey, and Urbanczyk. 225 48.
(1995b) Why Place and Voice are different: constraint interactions and feature
faithfulness in Optimality Theory. Ms., University of Maryland, College Park.
[ROA-105, http://ruccs.rutgers.edu/roa.html]
Lorentz, Ove (1996) Length and correspondence in Scandinavian. In P. Bye, O. Lorentz,
and C. Rice (eds.), Papers from the 2nd Workshop on Comparative Germanic
Phonology. Nordlyd 24. Troms: School of Languages and Literature. 11128.
Lowenstamm, Jean and Jonathan Kaye (1986) Compensatory lengthening in Tiberian
Hebrew. In L. Wetzels and E. Sezer (eds.), Studies in compensatory lengthening.
Dordrecht: Foris. 97146.

436
 References

Lynch, John D. (1974) Lenakel phonology. PhD dissertation, University of Hawaii.

Maddieson, Ian (1984) Patterns of sounds. Cambridge University Press.
Marantz, Alec (1982) Re reduplication. Linguistic Inquiry 13. 43583.
Marsack, C. C. (1962) Teach yourself Samoan. London: Hodder and Stoughton.
Mascar, Joan (1976) Catalan phonology and the phonological cycle. PhD dissertation,
MIT. [Distributed 1978, Bloomington: Indiana University Linguistics Club.]
Matteson, Esther (1965) The Piro (Arawakan) language. University of California
Publications in Linguistics 22. Berkeley and Los Angeles: University of California
Press.
McCarthy, John (1979) Formal problems in Semitic phonology and morphology. PhD
dissertation, MIT. [Published 1985, New York: Garland]
(1981) A prosodic theory of non-concatenative morphology. Linguistic Inquiry 12. 373 418.
(1986) OCP effects: gemination and antigemination. Linguistic Inquiry 17. 207 63.
(1993) A case of surface constraint violation. In C. Paradis and D. LaCharite (eds.),
Constraint-based theories in multilinear phonology, special issue of Canadian Journal
of Linguistics 38. 169 95.
(1995a) Extensions of faithfulness: Rotuman revisited. Ms., University of
Massachusetts, Amherst. [ROA-110, http://ruccs.rutgers.edu/roa.html]
(1995b) Remarks on phonological opacity in Optimality Theory. Ms., University of
Massachusetts, Amherst. To appear in J. Lecarme, J. Lowenstamm, and U. Shlonsky
(eds.), Proceedings of the Second Colloquium on Afro-Asiatic Linguistics.
(1997) Process specic constraints in Optimality Theory. Linguistic Inquiry 28. 23151.
(1998) Sympathy and phonological opacity. Ms., University of Massachusetts, Amherst.
[ROA-252, http://ruccs.rutgers.edu/roa.html]
McCarthy, John and Alan Prince (1986) Prosodic morphology. Ms., University of
Massachusetts, Amherst and Brandeis University, Waltham, Mass. [Annotated version
1996, issued as Technical report no. 32, Rutgers Center for Cognitive Science.]
(1990) Foot and word in prosodic morphology: the Arabic broken plural. Natural
Language and Linguistic Theory 8. 20983.
(1993a) Generalized Alignment. In G. E. Booij and J. van Marle (eds.), Yearbook of
Morphology 1993. Dordrecht: Kluwer. 79153.
(1993b) Prosodic Morphology I: constraint interaction and satisfaction. Ms., University
of Massachusetts, Amherst and Rutgers University. [To appear as Technical report
no. 3, Rutgers University Center for Cognitive Science. Cambridge, Mass.: MIT
Press.]
(1994a) The emergence of the unmarked: optimality in prosodic morphology. In M.
Gonzlez (ed.), Proceedings of the North East Linguistic Society 24. 33379.
(1994b) Two lectures on prosodic morphology. Handouts of two lectures at OTS/HIL
Workshop on Prosodic Morphology, Utrecht University, July 1994. [ROA-59,
http://ruccs.rutgers.edu/roa.html]
(1995a) Faithfulness and reduplicative identity. In Beckman, Walsh Dickey, and
Urbanczyk. 249 384.
(1995b) Prosodic morphology. In Goldsmith (1995). 318 66.

437
References

(forthcoming) Faithfulness and identity in prosodic morphology. To appear in Kager, van

der Hulst, and Zonneveld.
Merchant, Jason (1996) Alignment and fricative assimilation in German. Linguistic Inquiry
27. 709 19.
Merlan, Francesca (1982) Mangarayi. Lingua Descriptive Series 4. Amsterdam: North
Holland.
Mester, R. Armin (1990) Patterns of truncation. Linguistic Inquiry 21. 478 85.
(1991) Some remarks on Tongan stress. Ms., University of California, Santa Cruz.
(1994) The quantitative trochee in Latin. Natural Language and Linguistic Theory 12.
1 61.
Mester, R. Armin and Junko It (1989) Feature predictability and underspecication:
palatal prosody in Japanese mimetics. Language 65. 258 93.
Mills, Roger F. (1975) Proto South Sulawesi and Proto Austronesian phonology.
PhD dissertation, University of Michigan.
Mohanan, Karuvannur P. (1982) Lexical phonology. PhD dissertation, MIT.
(1993) Fields of attraction in phonology. In Goldsmith. 61116.
Moravcsik, Edith (1978) Reduplicative constructions. In J. Greenberg, C. Ferguson, and
E. Moravcsik (eds.), Universals of human language III: Word structure. Stanford
University Press. 297334.
Mller, Gereon (1997) Partial Wh-movement and Optimality Theory. Linguistic Review 14.
249 306.
Mutaka, Ngessimo and Larry Hyman (1990) Syllable and morpheme integrity in Kinande
reduplication. Phonology 7. 73120.
Myers, Scott (1987) Vowel shortening in English, Natural Language and Linguistic Theory
5. 485518.
(1991) Persistent rules. Linguistic Inquiry 22. 315 44.
(1997a) OCP effects in Optimality Theory. Natural Language and Linguistic Theory 15.
84792.
(1997b) Expressing phonetic naturalness in phonology. In Roca. 12552.
Myers, Scott and Troi Carleton (1996) Tonal transfer in Chichewa. Phonology 13. 39 72.
Nash, David (1980) Topics in Warlpiri grammar. PhD dissertation, MIT. [Published 1986,
New York: Garland.]
Newman, Stanley (1944) Yokuts language of California. Viking Fund Publications in
Anthropology 2. New York.
Newton, Brian. 1972 The generative interpretation of dialect: a study of Modern Greek
phonology. Cambridge University Press.
N Chiosin, Mire and Jaye Padgett (1997) Markedness, segment realisation, and locality
in spreading. Report LRC-97-01, Linguistic Research Center, University of
California, Santa Cruz. [ROA-188, http://ruccs.rutgers.edu/roa.html]
Nooteboom, Sieb G. (1981) Lexical retrieval from fragments of spoken words: beginnings
vs. endings. Journal of Phonetics 9. 40724.
Nouveau, Dominique (1994) Language acquisition, metrical theory, and optimality. PhD
dissertation, Utrecht University.

438
 References

Odden, David (1986) On the role of the Obligatory Contour Principle in phonological
theory. Language 62. 35383.
Ohala, John J. (1983) The origin of sound patterns in vocal tract constraints. In
P. F. MacNeilage (ed.), The production of speech. New York: Springer. 189216.
(1990) The phonetics and phonology of aspects of assimilation. In J. Kingston and
M. Beckman (eds.), Papers in laboratory phonology I: Between the grammar and
physics of speech. Cambridge University Press. 25875.
Ohala, John J. and H. Kawasaki (1984) Phonetics and prosodic phonology. In C. Ewen and
E. Kaisse (eds.), Phonology Yearbook 1. Cambridge University Press. 11327.
Ohala, John J. and Manjari Ohala (1993) The phonetics of nasal phonology: theorems and
data. In Huffman and Krakow. 225 49.
Onn, Farid M. (1976) Aspects of Malay phonology and morphology: a generative
approach. PhD dissertation, University of Illinois, Urbana-Champaign.
Orgun, C. Orhan (1994) Monotonic Cyclicity and Optimality Theory. In M. Gonzlez (ed.),
Proceedings of the North East Linguistic Society 24. 46174.
(1995) Correspondence and identity constraints in two-level Optimality Theory. In J.
Camacho, L. Choueiri, and M. Watanabe (eds.), Proceedings of the West Coast
Conference on Formal Linguistics 14. 399 413.
(1996) Sign-Based Morphology and Phonology: with special attention to
Optimality Theory. PhD dissertation, University of California, Berkeley. [ROA-171,
http://ruccs.rutgers.edu/roa.html]
Orgun, C. Orhan and Ronald Sprouse (1997) From M-Parse to Control: deriving
ungrammaticality. Ms., University of California, Berkeley. [ROA-224,
http://ruccs.rutgers.edu/roa.html]
Orr, Carolyn (1962) Ecuador Quichua phonology. In B. Elson (ed.), Studies in Ecuadorian
Indian languages. Norman, Okla.: Summer Institute of Linguistics. 60 77.
Osborn, Henry (1966) Warao I: phonology and morphophonemics. International Journal
of American Linguistics 32. 108 23.
Padgett, Jaye (1995) Feature classes. In Beckman, Walsh Dickey, and Urbanczyk. 385 420.
Paradis, Carole (1988) On constraints and repair strategies. Linguistic Review 6. 7197.
Paradis, Carole and Franois Prunet (eds.) (1991) The special status of coronals: internal
and external evidence. San Diego: Academic Press.
Pater, Joe (1995) On the nonuniformity of weight-to-stress and stress preservation effects
in English. Ms., McGill University. [ROA-107, http://ruccs.rutgers.edu/roa.html]
(1996) *Nt. In K. Kusumoto (ed.), Proceedings of the North East Linguistic Society 26.
22739.
(forthcoming) Austronesian nasal substitution and other Nt effects. To appear in Kager,
Van der Hulst, and Zonneveld. [ROA-160, http://ruccs.rutgers.edu/roa.html]
Payne, David L. (1981) The phonology and morphology of Axininca Campa. Arlington:
Summer Institute of Linguistics Publications in Linguistics.
Pesetsky, David (1998) Some optimality principles of sentence pronunciation. In P.
Barbosa, D. Fox, P. Hagstrom, M. McGinnis, and D. Pesetsky (eds.), Is the best good
enough? Optimality and competition in syntax. Cambridge, Mass.: MIT Press. 33783.

439
References

Piggott, Glyne L. (1993) The geometry of sonorant features. Ms., McGill University,
Montreal.
(1995) Epenthesis and syllable weight. Natural Language and Linguistic Theory 13.
283326.
Popjes, Jack and Jo Popjes (1986) Canela-Krah. In D. C. Derbyshire and G. K. Pullum
(eds.), Handbook of Amazonian languages, I. Berlin: Mouton De Gruyter. 12899.
Poser, William (1989) The metrical foot in Diyari. Phonology 6. 117 48.
Prentice, D. J. (1971) The Murut languages of Sabah. Pacic Linguistics, series C, no. 18.
Canberra: Australian National University.
Prince, Alan (1980) A metrical theory for Estonian quantity. Linguistic Inquiry 11. 511 62.
(1983) Relating to the grid. Linguistic Inquiry 14. 19 100.
(1984) Phonology with tiers. In M. Aronoff and R. Oehrle (eds.), Language sound
structure. Cambridge, Mass.: MIT Press. 234 44.
(1990) Quantitative consequences of rhythmic organization. In M. Ziolkowski, M. Noske,
and K. Deaton (eds.), Papers from the Chicago Linguistic Society 26:2. 35598.
(1994) Optimality: constraint interaction in generative grammar. Handout of lectures at
Utrecht University, January 1994.
Prince, Alan and Paul Smolensky (1993) Optimality Theory: constraint interaction in
generative grammar. Ms., Rutgers University, New Brunswick and University of
Colorado, Boulder. [To appear as Technical report no. 2, Rutgers University Center
for Cognitive Science. Cambridge, Mass.: MIT Press.]
(1997) Optimality: from neural networks to universal grammar. Science 275.
1604 10.
Radford, Andrew (1997) Syntax: a minimalist introduction. Cambridge University Press.
Raffelsiefen, Renate (1996) Gaps in word formation. In U. Kleinhenz (ed.), Interfaces in
phonology. Studia Grammatica 41. Berlin: Akademie Verlag. 194 209.
Rehg, Kenneth L. and Damien G. Sohl (1981) Ponapean reference grammar. Honolulu:
University of Hawaii Press.
Reynolds, William T. (1994) Variation and phonological theory. PhD dissertation,
University of Pennsylvania, Philadelphia.
Riad, Tomas (1992) Structures in Germanic prosody: a diachronic study with special
reference to Nordic languages. PhD dissertation, Stockholm University.
Rice, Curtis (1992) Binarity and ternarity in metrical theory: parametric extensions. PhD
dissertation, University of Texas, Austin.
Rice, Keren (1993) A reexamination of the feature [sonorant]: the status of sonorant
obstruents. Language 69. 308 44.
Rizzi, Luigi (1991) Residual verb second and the Wh Criterion. Technical Reports in
Formal and Computational Linguistics 2. Facult des Lettres, University of Geneva.
Roca, Iggy (1994) Generative phonology. London and New York: Routledge.
(ed.) (1997) Constraints and derivations in phonology. Oxford University Press.
Rosenthall, Sam (1994) Vowel/glide alternation in a theory of constraint interaction.
PhD dissertation, University of Massachusetts, Amherst. [ROA-126,
http://ruccs.rutgers.edu/roa.html]

440
 References

(1997) The distribution of prevocalic vowels. Natural Language and Linguistic Theory
15. 139 80.
Rumelhart, D. E. and J. L. McLelland (1986) On learning the past tenses of English.
In D. E. Rumelhart and J. L. McLelland (eds.), Parallel Distributed Processing:
explorations in the microstructure of cognition. Cambridge, Mass.: MIT Press.
216 71.
Russell, Kevin (1993) Morphemes as constraints: the case of copy-back reduplication. In
K. Beals, G. Cooke, D. Kathman, S. Kita, K.-E. McCullough, and D. Testen (eds.),
Papers from the Chicago Linguistic Society 29. 37589.
Sapir, Edward (1930) Southern Paiute, a Shoshonean language. Proceedings of the
American Academy of Arts and Sciences 65:13.
(1965) A grammar of Diola-Fogny. Cambridge University Press.
Schachter, Paul and Victoria Fromkin (1968) A phonology of Akan: Akuapem, Asante, and
Fante. Los Angeles: University of California.
Scobbie, James (1991) Attribute-value phonology. PhD dissertation, University of
Edinburgh.
Selkirk, Elisabeth O. (1980) The role of prosodic categories in English word stress.
Linguistic Inquiry 11. 563 605.
(1981) Epenthesis and degenerate syllables in Cairene Arabic. In H. Borer and J. Aoun
(eds.), Theoretical issues in the grammar of Semitic languages. Cambridge, Mass.:
Department of Linguistics and Philosophy, MIT.
(1984) Phonology and syntax: the relation between sound and structure. Cambridge,
Mass.: MIT Press.
(1986) On derived domains in sentence phonology. Phonology 3. 371 405.
(1995) The prosodic structure of function words. In Beckman, Walsh Dickey, and
Urbanczyk (eds.). 439 69.
Selkirk, Elisabeth and Tong Shen (1990) Prosodic domains in Shanghai Chinese. In
S. Inkelas and D. Zec (eds.), The phonology syntax connection. University of Chicago
Press. 31338.
Sezer, Engin (1981) The k / alternation in Turkish. In G. N. Clements (ed.),
Harvard studies in phonology. Bloomington: Indiana University Linguistics Club.
354 82.
Shaw, Patricia (1987) Non-conservation of melodic structure in reduplication. In A. Bosch,
B. Need, and E. Schiller (eds.), Papers from the Chicago Linguistic Society 23:2.
291306.
(1991) Ling 253: morphology. Handout of LSA Linguistic Institute 1991.
Sherer, Tim (1994) Prosodic phonotactics. PhD dissertation, University of Massachusetts,
Amherst. [ROA-54, http://ruccs.rutgers.edu/roa.html]
Sherrard, Nicholas (1997) Questions of priorities: an introductory overview of Optimality
Theory in phonology. In Roca. 4389.
Siegel, Dorothy (1974) Topics in English morphology. PhD dissertation, MIT.
Smith, Neilson V. (1973) The acquisition of phonology: a case study. Cambridge University
Press.

441
References

Smolensky, Paul (1993) Harmony, markedness, and phonological activity. Handout to talk
presented at Rutgers Optimality Workshop 1, 23 October 1993, New Brunswick,
N. J. [ROA-87, http://ruccs.rutgers.edu/roa.html]
(1995) On the internal structure of the constraint component Con of UG. Handout
to talk presented at University of California, Los Angeles, 7 April 1995. [ROA-86,
http://ruccs.rutgers.edu/roa.html]
(1996) On the comprehension/production dilemma in child language. Linguistic Inquiry
27. 720 31.
Sommer, Bruce (1981) The shape of Kunjen syllables. In D. L. Goyvaerts (ed.), Phonology
in the 1980s. Ghent: E. Story-Scientia.
Spaelti, Philip (1994) Weak edges and nal geminates in Swiss German. In M. Gonzlez
(ed.), Proceedings of the North East Linguistic Society 24. 57388.
Spencer, Andrew J. (1994) Syllabication in Chukchee. In R. Wiese (ed.), Recent
developments in Lexical Phonology. Theorie des Lexikons. Arbeiten des Sonder-
forschungsbereichs 282, No. 56. Dsseldorf: Heinrich Heine University. 20526.
Spring, Cari (1990) Implications of Axininca Campa for prosodic morphology and
reduplication. PhD dissertation, University of Arizona, Tucson.
Sprouse, Ronald (1997) A case for Enriched Inputs. Handout of a presentation at TREND,
3 May 1997. [ROA-193, http://ruccs.rutgers.edu/roa.html]
Stahlke, Herbert (1976) Segment sequences and segmental fusion. Studies in African
Linguistics 7. 44 63.
Stampe, David (1972) How I spent my summer vacation (a dissertation on Natural
Phonology). PhD dissertation, University of Chicago. [Published 1979, New York:
Garland.]
Steinbergs, Aleksandra (1985) The role of MSCs in OshiKwanyama loan phonology.
Studies in African Linguistics 16. 89 101.
Stemberger, Joseph P. (1991) Radical underspecication in language production.
Phonology 8. 73112.
Steriade, Donca (1982) Greek prosodies and the nature of syllabication. PhD dissertation,
MIT.
(1987) Redundant values. In A. Bosch, B. Need, and E. Schiller (eds.), Papers from the
Chicago Linguistic Society 23:2. 339 62.
(1988a) Reduplication and syllable transfer in Sanskrit and elsewhere. Phonology 5.
73155.
(1988b) Review of G. N. Clements and S. J. Keyser (1983). Language 64. 118 29.
(1993) Closure, release, and nasal contours. In Huffman and Krakow. 40170.
(1995a) Underspecication and markedness. In Goldsmith (1995). 114 74.
(1995b) Positional neutralization. Ms., University of California, Los Angeles.
(1996) Paradigm uniformity and the phoneticsphonology boundary. Paper presented at
the 5th Conference in Laboratory Phonology, Evanston, Illinois, July 1996.
Stonham, John (1990) Current issues in morphological theory. PhD dissertation, Stanford
University.

442
 References

Street, Chester S. and Gregory Panpawa Mollinjin (1981) The phonology of Murinbata. In
B. Waters (ed.), Australian phonologies: collected papers. Work Papers of SIL-AAB,
A5. Darwin: Summer Institute of Linguistics.
Tesar, Bruce (1995) Computational Optimality Theory. PhD dissertation, University of
Colorado, Boulder.
(1996) An iterative strategy for learning metrical stress in Optimality Theory. In
E. Hughes, M. Hughes, and A. Greenhill (eds.), The proceedings of the 21st annual
Boston University Conference on Language Development, November 1996. 61526.
[ROA-177, http://ruccs.rutgers.edu/roa.html]
Tesar, Bruce and Paul Smolensky (1993) The learnability of Optimality Theory: an
algorithm and some basic complexity results. Technical Report CU-CS-678-93,
Computer Science Department, University of Colorado, Boulder. [ROA-2,
http://ruccs.rutgers.edu/roa.html]
(1998) Learnability in Optimality Theory. Linguistic Inquiry 29. 229 68.
Trigo, R. Lorenza (1993) The inherent structure of nasal segments. In Huffman and
Krakow. 369 400.
Trommelen, Mieke and Wim Zonneveld (1979) Inleiding in de Generatieve Fonologie.
Muiderberg: Coutinho.
Trubetzkoy, Nicolai S. (1939) Grundzge der Phonologie. Travaux du Cercle Linguistique
de Prague 7. [English translation: Principles of phonology. Berkeley and Los Angeles:
University of California Press, 1969.]
Uhrbach, Amy (1987) A formal analysis of reduplication and its interaction with
phonological and morphological processes. PhD dissertation, University of Texas,
Austin.
Urbanczyk, Suzanne (1996) Morphological templates in reduplication. In K. Kusumoto
(ed.), Proceedings of the North East Linguistic Society 26. 425 40.
(forthcoming) Double reduplications in parallel. To appear in Kager, van der Hulst, and
Zonneveld. [ROA-73, http://ruccs.rutgers.edu/roa.html]
Vance, Timothy J. (1987) An introduction to Japanese phonology. Albany, N.Y.: SUNY
Press.
Vijver, Ruben van de (1998) The iambic issue: iambs as a result of constraint interaction.
HIL dissertations 37. Leiden: Holland Institute of Generative Linguistics.
Vogel, Irene (1976) Nasals and nasal assimilation patterns, in the acquisition of Chicano
Spanish. Papers and Reports on Child Language Development 15. 20114.
Walther, Markus (1996) ot simple: a construction-kit approach to Optimality
Theory implementation. Ms., Heinrich Heine University, Dusseldorf. [ROA-152,
http://ruccs.rutgers.edu/roa.html]
Wexler, Kenneth and Peter W. Culicover (1980) Formal principles of language acquisition.
Cambridge, Mass.: MIT Press.
Whitney, William D. (1889) Sanskrit grammar. Cambridge, Mass.: Harvard University Press.
Wilbur, Ronnie (1973) The phonology of reduplication. PhD dissertation, University of
Illinois, Urbana-Champaign.

443
References

Willett, Elizabeth (1982) Reduplication and accent in Southeastern Tepehuan. International

Journal of American Linguistics 48. 164 84.
Willett, Thomas L. (1991) A reference grammar of Southeastern Tepehuan. Summer
Institute of Linguistics and University of Texas at Arlington Publications in
Linguistics 100. Arlington: University of Texas.
Yip, Moira (1988) The Obligatory Contour Principle and phonological rules: a loss of
identity. Linguistic Inquiry 19. 65 100.
(1993) Cantonese loanword phonology and Optimality Theory. Journal of East Asian
Linguistics 2. 26191.
Zagona, Karen (1982) Government and proper government of verbal projections. PhD
dissertation, University of Washington, Seattle.
Zimmer, Karl and Barbara Abbott (1978) The k/ alternation in Turkish: some
experimental evidence for its productivity. Journal of Psycholinguistic Research 7.
35 46.
Zoll, Cheryl (1996) Parsing below the segment in a constraint based framework.
PhD dissertation, University of California, Berkeley. [ROA-143,
http://ruccs.rutgers.edu/roa.html]

444
 INDEX OF LANGUAGES

(bold-face numbers refer to analysis sections)

Agta 200, 216 18 Harari 141

Akan 208 11, 216 Hebrew, Tiberian 130
Arabic 93, 94, 130, 144, 145 Hindi 130
Arabic, Palestinian 58, 27892, 386 Hixkaryana 146, 148 60, 161, 162, 173
Axininca Campa 93, 98 9, 101 4, 11012, 137, Hungarian 130
158, 254 5
Icelandic 26573
Basque 130, 3937 Indonesian 5961, 62, 64, 6770, 727, 79, 81,
Burmese 70 84, 86, 87, 16970, 1945, 217, 251

Canela-Krah 97 Japanese 50 1, 93, 97, 1312, 133, 138, 2403,

Carib 158 245, 246, 3979
Cayuvava 94 Javanese 1989
Cheremis, Eastern 145
Chinese 347, 367
Kaingang 81
Chukchee 251
Kikuyu 89
Czech 347
Kinande 202, 212
Konjo 73
Diola Fogny 93, 134 8, 213
Diyari 195 7, 205, 212, 214, 217, 21823,
251 Lenakel 96, 1079, 11217, 1269, 136, 137,
Dutch 14 17, 18, 27, 402, 52, 54, 74, 90, 131, 245, 252
324 7, 330, 331, 336 7, 347, 368, 399, 411,
414 19 Makassarese 207 8
Malay 233 8, 244
English 6, 1718, 28, 2932, 34, 37, 51, 79, 83, Malay, Kelantan 7980, 87
93, 94, 97, 142, 146, 173, 245, 259 62, Manam 191
273 7, 295, 347, 353 65, 367 8, 369, Mandar 81 2, 87
401 3, 404, 406, 419 Mangarayi 256
Estonian 175 Maore 79
Mazateco 94
Fijian 94, 173 Mohawk 130, 166 7
Fijian, Boumaa 94, 97, 1757 Mongolian, Khalka 145
Finnish 97, 175 Murinbata 1912
French 28, 197, 367
Nahuat, Isthmus 374, 375, 377, 3835,
Gaelic, Barra 95 38990
Garawa 1679, 170 Nootka 1957, 205 7, 211, 216
German 268 Norwegian 268
Greek 79, 88 9, 104 Nunggubuyu 146

445
Index of languages
Oromo 3801, 385, 389, 390 Tepehuan, Southeastern 177 88
OshiKwanyama 836, 87 Timugon Murut 2249, 255
Oykangand 95, 255 Toba Batak 81
Tongan 175 7
Paiute, Southern 2447 Totonac 97
Pangasinan 255 6 Tunica 94, 97, 3734, 375, 378
Pintupi 144, 145, 1645, 188, 192, 300 1, 302, Turkish 373, 374, 3767, 3823, 386, 3879,
303, 308, 312 390 1
Piro 168, 170
Polish 347
Venda 79
Ponapean 93, 133 4, 136, 137, 138

Quechua, Puyu Pungo 801, 84, 85, 87 Warao 144, 145, 164
Wargamay 191
Samoan 203, 204, 205, 212, 224 Warlpiri 192
Sanskrit 197, 21315, 216, 217, 251 Washo 2312
Sedang 97 Weri 144
Sibutu Sama 144
Spanish 79, 97, 294
Yidi 145, 196
Swahili 79
Yokuts 58, 97, 104, 379, 41011
Tagalog 121 4, 141, 197
Temiar 93 Zoque 80

446
 INDEX OF SUBJECTS

(bold-face numbers refer to denitions and special discussions)

abstractness 186, 286 7, 419 constraint reranking 35, 298

alignment (see also Index of constraints) free ranking 406
alignment constraint 11115, 11721 strict ranking (see strict domination)
Generalized Alignment 118 transitivity of ranking 21
morpho-prosodic alignment 115, 11819, 122, universal ranking 44, 76, 409
135 constraint satisfation 9
allomorphy 130, 41420 Constraint set (Con) 21
allophonic variation 27 8, 30, 38, 241, 260, 275, constraint stratum 299, 333, 382
412 constraint violation 9, 21
absolute (categorical) violation 120
base 202, 282 gradient violation 1201
base in reduplication 194, 201, 202 minimal violation 25, 66, 105, 108
base in truncation 263 violation marks 21, 24, 1201, 3047
base in stem-based afxation 282 correspondence constraint (see
base-identity 200, 2812 correspondence)
faithfulness constraint (see faithfulness)
candidate (see output candidate) grounded constraint (see grounding)
coalescence 5 960 markedness constraint (see markedness)
compensatory lengthening 380, 389 output constraint 3, 9, 58
computation 25 7 universality of constraints 3, 9, 11, 119
Connectionism 422 violability of constraints 3, 12, 21, 11012,
conspiracy 55 6, 83 4, 381 1223, 342 4
constraint 3, 9 (for specic constraints, see Index well-formedness constraint (see well-
of constraints) formedness)
alignment constraint (see alignment) Containment 99 100, 345, 378
constraint conict 3, 4, 1213, 1516 contrast 5, 10, 29, 39
constraint conjunction (see Local Conjunction) co-phonology 405
constraint demotion 298 9, 311 correspondence 15, 203, 248, 263
blocking of constraint demotion 344 candidate-to-candidate correspondence (see
constraint demotion algorithm 311 Sympathy)
minimal constraint demotion 319 21 correspondence constraint 67, 205, 24952,
constraint family 121, 248 264, 379 81, 412 (see also Index of
constraint hierarchy 13, 21, 2989 constraints)
stratied constraint hierarchy 299 Correspondence Theory 100, 200, 24852,
constraint interaction 225 258, 392
constraint ranking 3, 4, 21, 323, 4056 input-to-output correspondence (see
constraint ranking algorithm (see constraint faithfulness)
demotion, Recursive Ranking) output-to-output correspondence 237, 257 8,
constraint ranking argument 16, 150, 298, 263, 287, 386, 415, 420
303, 3089 cycle 277, 2801, 285 6

447
Index of subjects
deletion 67 language acquisition 1, 296, 3367
consonant deletion 134, 380 learnability 296
vowel deletion (see syncope) of serial grammar 186, 2867
derivation 57 8, 1857, 238, 280 of lexical representations 322, 329, 334
derivational theory 57 8 of OT grammar 2967, 407
derived environment effect 75 of output representations 322
distinctive feature 3, 36, 71, 127, 408 lexical representation 33, 413, 415
do-only-when-necessary (see economy) lexicon 19, 32 4, 415
domination 13 Lexicon Optimization 323, 32930, 334,
strict domination 223, 344, 400, 405 414
Duplication Problem 56, 84, 381 licensing 409 11
linguistic theory 13
economy 21, 74, 104 7, 343, 357 goals of 1
emergence of the unmarked 130, 206 7, Local Conjunction 392400
209 10, 21516, 222
epenthesis 68, 98 100, 124 5 Mark Cancellation 306
consonant epenthesis 105 Mark-Data pairs 304
vowel epenthesis 106, 112, 126, 133, 287 markedness 2 3, 45, 44
subminimal epenthesis 144, 166 markedness constraint 910 (see also Index of
Evaluator (Eval) 19, 205 constraints)
extrametricality 149, 166 context-free markedness constraint 28, 125
extraprosodicity 124 context-sensitive markedness constraint
28, 125
markedness scales 44
factorial typology 35, 171, 239, 366, 412
metathesis 63
faithfulness 5 6, 10, 55, 74, 200, 211, 414
Minimalism 343
faithfulness constraint 10, 14, 200, 350
Morpheme Structure Constraint 20, 56
positional faithfulness 407 13
morphology 3, 19, 115, 194, 274
root faithfulness 76
morphological category (constituent) 118 19
fallacy of perfection 16
morphological constraint 122, 21718
foot 142, 145, 147
morphological edge 75, 115, 135
foot distribution (parsing) 1625, 16770
morphological feature 282
foot well-formedness 161, 172 4
morphological rule (process, operation) 197,
free variation (see optionality)
223 4, 262, 273, 282
Freedom of Analysis 20, 25, 535, 344, 417
morphological truncation (see truncation)
functional motivation 6 7, 10, 421
morphologically related 263, 2812, 415
functionally related processes 56, 87, 139, 177
morphology-phonology interface 25, 118,
122 4, 197, 2235, 262
Generative Grammar 1, 296
Generator (Gen) 19, 20, 545, 204, 303, neutralization 14, 29, 30, 36
3457 complete neutralization 36
grammar 1, 8, 422 positional neutralization 389, 4078
components of 19
structure of 18 21, 382 opacity 3724
grammatical word 118, 152, 166 optimality 1213
grounding of constraints 5, 11, 61, 94, 144, 241, Optimality Theory
381, 408 and goals of linguistic theory 34
basic ideas of 3 4, 8 13
harmony 3, 8 9, 1617, 151, 344 comparison with derivational (rule-based)
harmony scale 44, 409 theory 52 8, 86 8, 1389, 187 8,
237 8, 262, 273, 285 7
identity (see base identity and reduplicant) optionality 4047
implicational universal 93 4 output 3, 8, 1213, 56, 58, 87, 401, 404, 420
inxation 121 4, 2245 nonunique output 405
input 8, 1920, 33, 57 8, 322, 324, 344 5, output candidate 20
378 81, 413 14 candidate set 8, 257
intermediate levels 57 8, 280, 376, 3815 faithfulness to (see Sympathy)

448
 Index of subjects
output constraint (see constraint) reduplicative template (see template)
output-to-output correspondence total reduplication 195, 205
(see correspondence) rewrite rule (see rule)
overapplication Richness of the Base 19, 29, 31, 34, 414
in opacity 373 rule 523
in reduplication 1989, 2338 optional (see optionality)
in stem-based afxation 274 6 rewrite rule 1, 53
rule ordering 1, 57
paradigm 415 18 counterbleeding 375
paradigmatic relation 287 counterfeeding 375
paradigm uniformity 6, 257 structural change of 53
parallelism 25 structural description of 53
in reduplication 2447 rule-based theory 523
of base identity and faithfulness 292 comparison with OT (see Optimality Theory)
of evaluation of candidates 55
of morphology and prosody 122 3, segment inventory 434
225, 227 stem-based afxation 273 4
of prosodic levels 188 9 Basic Model 275
parameter 2, 296, 3412 structural change of rule (see rule)
Persistent Rule Theory 56 structural description of rule (see rule)
phonetics 5, 11, 61, 408, 421 2 Structuralism 2, 422
phonetic grounding (see grounding) syllable 917
phonology 5 6 coda 91, 94, 96, 130
autosegmental phonology 138 9 nucleus 91
Declarative Phonology 12 onset 91, 93, 96
Generative Phonology 2, 52, 413 syllable template 98
Harmonic Phonology 4, 382 syllable weight 147, 155
Lexical Phonology 2, 262, 277, 382 Sympathy 38792
metrical phonology 142, 1467 syncope 177 8, 278
Natural Phonology 2 syntax 2, 341, 3447
positional faithfulness (see faithfulness) syntactic constraint 34852
positional neutralization (see neutralization)
positive evidence 302 tableau 13
Principles-and-Parameters Theory 2, 3, 296, 341 template 217
prosodic circumscription 2245, 229 classical template 21617
prosodic morphology 217, 230 decomposition of 21819
prosody Generalized Template 218, 220
prosodic categories 119, 146 reduplicative template 196, 230
prosodic head 146, 167 truncation template 2645
prosodic hierarchy 146 violability of template 227, 230
prosodic word 111, 136, 146, 220 1 truncation 259, 265
Basic Model 263
quantity-sensitivity 145 6, 1715 truncation template (see template)
typology 1, 11, 34
Recursive Ranking 308, 311 of structural changes 645, 78
robustness of 31719
reduplicant 194, 201, 202 underapplication
constraint on 218, 220, 227 in opacity 374
identity with base 1989 in reduplication 198, 240 4
shape invariance of 195 6, 199 in stem-based afxation 27885
unmarkedness of 196 8, 199 in truncation 2601
reduplication 194 5 Underlying Representation (see lexical
as inxation 203, 212, 2249 representation)
as prexation 195, 202, 212 underspecication 33, 70, 127, 209, 398,
as sufxation 202, 212 4201
Basic Model 201 ungrammaticality 3, 23
partial reduplication 195, 205, 222 absolute ungrammaticality 4003

449
Index of subjects
Universal Grammar 1 well-formedness
universality 12, 18 output-based nature of 878
two-level well-formedness 378 81
violation (see constraint violation) well-formedness constraint 93, 104

450
 INDEX OF CONSTRAINTS

(page numbers refer to constraint denitions)

Correspondence constraints Ident-IO(voice) 14, 40, 232, 325, 340

Anchoring 251 Wt-Ident-IO 269
Anchor-IO 137, 251 Ident-OO(voice) 416
Anchor-BR 213, 251 Ident-OO(voice, [C) 417
Anchor-BT 264 Ident-{O(raised) 396
Contiguity 250 Ident-{O(voice) 389
Contig-BR 214, 250 Linearity 251
Contig-IO 250 Linear-BR 251
Contig-BT 264 Linear-BT 264
Dependence 205, 249 Linear-IO 63, 251
Dep-BR 208, 249 Maximality 205, 249
Dep-BT 264 HeadMax-BA 283, 386
Dep-IO 68, 101, 249 Max-BA 289
Dep--IO 156 Max-BR 205, 249
HeadDep-IO 290 Max-BT 264, 272
Identity[F] 205, 250 Max-IO 67, 102, 249, 284
Ident-BA(tense) 276 Max-C-IO 181
Ident-BR[F] 208, 250 Max-V-IO 178
Ident-BR(high) 210 Max--IO 176
Ident-BR(nasal) 242 Max-{O 388
Ident-BR(voice) 232 Max--{O 389
Ident-BT[F] 264 RootLin-IO 76
Ident-BT(back) 261 Uniformity 251
Wt-Ident-BT 271 Uniformity-IO 63, 77
Ident-IO[F] 250
Ident-IO[a] 283 Featural markedness constraints
Ident-IO[i] 283 *[back] 127
Ident-IO(back) 128, 260 *[cor] 44
Ident-IO(high) 395 *[high] 209, 290
Ident-IO(low) 128 *[+high] 128
Ident-IO(nasal) 29 *[lab] 44
IdentIO(nasal) 72 *[+low] 127
Ident-IO(ObsVce) 70 *[+round] 127, 409
Ident-IO(Ons, voice) 340 *Back-low 260
Ident-IO(Place) 45, 132 *Nt 61
Ident-IO(raised) 395 *NVoral 234
[Ident-IO(raised) & Ident-IO(high)] 396 *Tense-low 275
Ident-IO(round) 128, 409 *VkV 376
Ident-IO(round, [) 409 *Vnasal 28, 234
Ident-IO(tense) 275 *Voiced-Coda 14, 40, 232, 325

451
Index of constraints
*VoralN 28 Uneven-Iamb 151
*VwV 245 Weight-by-Position 147, 269
*r] 260 WSP 155, 172
*[ 241 WSP-Ft 184
*[W 245
-Tensing 275 Various phonological constraints
Coda-Cond 131 *Struc 404
Cor-high 128 Harmony-IO 378
Harmony 410 Fill 99
Hiatus-raising 394 Lowering-IO 379
Inter-V-Voice 325 MParse 402
No [a] 283 Parse 100
No [i] 284 -Projection 380
OCP 56
PostVcls 241 Alignment constraints
Reduce 406 Align-L 111
VOP (Voiced Obstruent Prohibition) Align-Morph-L 115, 136
40, 340, 398 Align-R 113, 290
VOP 2 398 Align-Red-L 226
Align-Stem-R 119
Prosodic markedness constraints Align-Syllable-R 119
*Clash 165 Align-um-L 122
*Complex (cover constraint) 288 Align-Wd-Left 169
*ComplexCod 97 Align-Wd-Right 169
*ComplexOns 97 All-Ft-Left 157, 163, 300
*Final-C- 268 All-Ft-Right 163, 300
*3 268 Edgemost 122, 167
* 283 GrWd=PrWd 152, 166
Final-C 377 Leftmost 167, 300
Ft-Bin 156, 161, 184, 300 Red=Stem 220
Ft-Form (cover constraint) 184 Red= 227
HNuc 215 Red= 182, 217
No-Coda 94, 112 Rightmost 167, 300
[NoCoda & VOP]Segment 399 Stem=PrWd 220, 264
NonFinality 151, 165 Trunc= 265, 271
Onset 93, 110
Parse-Syl 153, 162, 300 Syntactic constraints
Quant-Form (cover constraint) 271 Economy of Movement (Stay) 351
Rh-Contour 174, 184 Full-Interpretation (Full-Int) 352
RhType=I 172, 184 No Movement of a Lexical Head
RhType=T 172 (No-Lex-Mvt) 352
Son-Seq 267, 288 ObligatoryHeads (Ob-Hd) 349
Stress-to-Weight 268 Operator in Specier (Op-Spec) 354
Syll-Form (cover constraint) 288 Subject 354

452