Control and Complementation

Author(s): Joan Bresnan

Source: Linguistic Inquiry , Summer, 1982, Vol. 13, No. 3 (Summer, 1982), pp. 343-434
Published by: The MIT Press

Stable URL: http://www.jstor.com/stable/4178286

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to Linguistic
Inquiry

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
Joan Bresnan Control and Complementation

This article develops a theory of control and complementation based on the

lexical-functional theory of grammar, LFG, presented in Bresnan (1982a). A formal
theory of grammar, such as the theory of LFG, is not itself a substantive linguistic
theory. Rather, it is a language for precisely expressing descriptive rules and universal
postulates of grammar. The choice of such a formal theory of grammar is extremely
important. If the formal theory contains the appropriate concepts and representations,
then linguistic principles and grammatical descriptions expressed within it will imme-
diately generalize along the right dimensions, simplifying both descriptive rules and
theoretical postulates. An inappropriate formal theory will require a host of auxiliary
concepts and definitions, and may obscure the underlying regularities that optimal gram-
mars must express. Having no formal theory at all will lead to vague and inconsistent
formulations at both the theoretical and the descriptive levels. Despite its importance,
however, a formal theory of grammar is only one step in the construction of a substantive
linguistic theory of universal grammar. The present work adds to the formal theory of
LFG a set of substantive postulates for a universal theory of control and complemen-
tation. Not all of the relevant phenomena could be treated here, but the central concepts
of function, category, syntactic encoding of function in categorial structure, government,
and functional and anaphoric control are discussed, and comparisons are made with
alternative theories including the government and binding theory of Chomsky (1981).
The basic assumptions of the theory of control are articulated in the initial sections:
first, grammatical functions are universal primitives of syntax, not derived from phrase
structure representations or from semantic notions; second, grammatical functions are
lexically encoded in predicate argument structures in varying ways; third, constituent
structure categories are universally decomposed into features and types, the features
being definable in terms of the primitives SUBJ, OBJ; and fourth, grammatical functions
are syntactically encoded directly in surface representations of phrase structure, accord-
ing to structural configurations or morphological features. In lexical-functional theory,
the grammatical relations of a sentence-that is, the set of associations between its

This article is part of a work published by the MIT Press (J. Bresnan, ed., The Mental Representation
of Grammatical Relations) and appears here with their kind permission.
I wish to thank K. P. Mohanan, Per-Kristian Halvorsen, Ronald Kaplan, and Marilyn Ford for detailed
comments on earlier drafts of this study. I am also grateful to Jane Grimshaw, Lorraine Levin, Malka Rap-
paport, and Jane Simpson for suggesting several improvements. This study is based upon work supported by
the National Science Foundation under Grant No. BNS 80-14730 to MIT and also by the Cognitive and
Instructional Sciences Group of the Xerox Palo Alto Research Center.

Linguistic Inquiry, Volume 13, Number 3, Summer 1982

0024-3892/82/030343-92 $02.50/0
?1982 by The Massachusetts Institute of Technology 343

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 344 JOAN BRESNAN

surface syntactic constituents and

resented by a functional structure (f
of constituency relations (c-struct
a universal format, while constituency relations vary radically across languages and
even across constructions within a single language. In this article it will be argued that
the syntactic conditions on government and control are properly stated in terms of
f-structure representations, not c-structure representations. Evidence will be presented
to choose between alternative theories of government and control, and a number of
major generalizations will be derived from the present theory.
This study presupposes previous work in LFG, including Kaplan and Bresnan
(1982), Grimshaw (1982), and Bresnan (1982b); these, together with related work, are
collected in Bresnan (1982a).

1. Grammatical Functions

In the lexical-functional theory, grammatical functions are universal, syntactically pr

itive elements of grammar. Although grammatical functions are basic concepts in
ditional grammars as well as in recent studies of a wide variety of languages (Perlmu
(to appear)), some researchers have objected to taking grammatical functions as prim
itives at the level of syntactic representation.
The theory of universal grammar, like other theories, can be axiomatized as a set
of basic concepts and a set of postulates from which the explanatory results of the theory
are deducible as theorems; all grammatical concepts of the theory can be defined in
terms of the basic set of concepts, which are taken to be primitive, or undefined. In
general, there are many alternative sets of primitives in terms of which theories can be
axiomatized. However, Chomsky (1981) asserts that we are not free to choose our
primitives in constructing an explanatory theory of universal grammar; in order to explain
the fact of language acquisition, we must require that the primitives of our formal
grammatical theory correspond directly to categories of prelinguistic experience. This
he calls the criterion of epistemological priority. He then suggests that the primitive
concepts in terms of which constituent structure has been defined may be "epistemo-
logically prior" to the primitives in terms of which grammatical relations have been
defined. This is the basis of Chomsky's (1981, 10) conclusion that "we should . . . be
wary of hypotheses that appear to assign to grammatical relations too much of an
independent status in the functioning of rule systems."
However, Chomsky's suggestion that the primitives of categorial or constituent
structure are epistemologically prior to the primitives of relational or functional struc-
tures is unsupported; no evidence or argumentation is offered for this view. Indeed, this
view is implausible, for if it is assumed that the primitives of constituent structure are
epistemologically prior, and that grammatical relations are derived from them as defined
concepts, then it is difficult to explain the invariance of grammatical relations across
the radically varying constituent structures of different languages. If constituency prim-

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 345

itives were epistemologically prior, the appearance of universal grammatical relations

in languages which lack sufficient constituent structure to support configurational def-
initions (see Perlmutter (to appear), Hale (1979), Mohanan (1982a)) would be completely
mysterious.
A more fundamental problem is the criterion of epistemological priority itself. Pinker
(1982) and Pinker and Lebeaux (1981) have shown that there is a natural inductive basis
for the elementary concepts of category and function that are assumed in the
lexical-functional theory of syntax; for example, the notion of SUBJ iS grounded in the
concept of animate agent or actor. However, this inductive basis is supplemented by
a distributional learning algorithm that can infer the existence of nonagent subjects. As
a result, according to the Pinker-Lebeaux model, the mature grammar eventually ac-
quired by the language learner is not represented in terms of formal primitives that
correspond directly to prelinguistic notions such as animate agent or actor. Pinker and
Lebeaux argue that this model can not only explain the fact of language acquisition,
which is the goal of learnability models, but also account for the course of language
acquisition; that is, the model provides a consistent and coherent account of the evidence
from studies of child language. In short, since well-motivated language acquisition pro-
cedures can induce classes of formal grammars whose primitive concepts bear no direct
correspondence to prelinguistic categories, the criterion of epistemological priority is
simply irrelevant to the goal of constructing an explanatory linguistic theory.
The doctrine that theories of higher cognitive processes should be formulated in
terms of primitives that correspond directly to categories of primary experience derives
from empiricist theories of knowledge that have long since been rejected in the psy-
chology of language and perception, and which in fact have never been accepted in
generative linguistic theory. The evidence suggests, on the contrary, that the human
mind is innately endowed with powerful conceptual structures and operations that are
only abstractly and indirectly related to "primary experience" (Miller and Johnson-
Laird (1976)); elsewhere, Chomsky has also championed this view (Chomsky (1969)).
The fact is that many elementary linguistic concepts play an important theoretical role
in rule systems and yet cannot be reduced to primary experience as required by the
criterion of epistemological priority. For example, stress is a basic concept of Chomsky
and Halle's (1968) theory of English phonology, although phonetic research had already
shown that stress has no reliable acoustical correlate in amplitude, duration, or funda-
mental frequency of sound (Lehiste (1970)). What is required of an explanatory theory
of grammar is not that it be axiomatizable in terms of such "epistemologically prior"
primitives, but rather that its knowledge representations can be naturally embedded in
substantive theories of the processes of language acquisition and use. The theory of
syntax adopted here has already begun to meet this requirement in the work of Pinker
(1982), Pinker and Lebeaux (1981), Ford (1982), and Ford, Bresnan, and Kaplan (1982).
A different form of objection to considering grammatical functions as primitives
could be based on constructive elimination, by constructing from constituency primitives
alone a syntactic theory which is capable of explaining the universal properties of gram-

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 346 JOAN BRESNAN

matical relations. Research in re

studies in Bresnan (1982a) and els
(1980), Roberts (1981), Simpson and Bresnan (in preparation)), shows that the invariance
of grammatical relations across languages and across constructions with radically varying
constituency relations is not explained by purely structure-based theories. In fact, to
express certain universals, recent versions of transformational grammar now make ex-
plicit or tacit appeal to grammatical relations that cannot be eliminated through defini-
tions in terms of independently motivated constituent structure configurations (see
Chomsky (1980a, 1981); Mohanan (1981a)).
These objections to considering grammatical functions as primitives at the level of
syntactic representation can therefore be rejected. The question remains, however,
whether grammatical functions can be reduced to nonsyntactic notions at other levels
of linguistic representation. The important work of Dowty (to appear) bears on this
question. Dowty proposes a definition of the grammatical functions in terms of their
universal roles in the semantic composition of sentences (assuming the framework of
Montague (1973)). For example, the object is the argument that combines with a verb
meaning to produce an intransitive verb phrase meaning, and the subject is the argument
that combines with an intransitive verb phrase meaning to produce a sentence meaning.
A successful definition in these terms would be a welcome advance in grammatical
theory. However, while it is possible to identify semantic notions that can be put into
correspondence with syntactic concepts, what must be shown in order to claim that the
syntactic concepts can be eliminated is that this correspondence is explanatory rather
than arbitrary. In other words, the syntactic properties of grammatical functions-such
as their roles in reflexivization, causativization, passivization, and control-should not
be arbitrarily associated with the proposed semantic definitions, but should be shown
to follow from the appropriate semantic characterizations. Unless this is done, syn-
tactically primitive grammatical functions will not have been eliminated by these defi-
nitions, but only renamed. For example, in terms of the semantic composition of sen-
tences, the subject plays the same semantic role with both transitive and intransitive
verbs. Yet causativization in Malayalam (as in many languages) affects the subjects of
transitive and intransitive verbs in syntactically distinct ways: the subject of an intran-
sitive base verb behaves syntactically like the object of the causativized form of the
verb, but the subject of a transitive base verb behaves syntactically like an oblique
instrumental of the causativized verb (Mohanan (1981a)). Since the semantic properties
of the causativized subjects are the same regardless of whether they function syntac-
tically as objects or as oblique instrumentals, there is no principled semantic distinction
between these two grammatical functions. The proposed semantic definitions of gram-
matical functions do not explain the syntactic properties of objects and obliques (although
one could of course stipulate arbitrary differences in the semantic composition of cau-
sativized transitives and intransitives in order to match the syntactic differences). We
therefore conclude that the definitions do not eliminate grammatical functions, but
merely rename them in semantic terms. (See Mohanan (1981a) for a similar argument

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 347

with respect to recent attempts to define grammatical functions in terms of Chomsky's

government and binding theory.)
Because grammatical functions in the present theory play an important role both
in lexical representations and in semantically interpreted syntactic representations, it
is of course possible that lexical or sentential semantics (or both) may ultimately provide
true nonsyntactic definitions of these concepts, just as many of the elementary features
of phonology are definable in terms of phonetic theory. But the elimination of functions
as primitives is no more necessary to an explanatory theory of grammar than is the
elimination of phonological features.
The grammatical functions assumed in this study are classified in figure 1. The
functions XCOMP and xADJ(unct) are distinguished as open grammatical functions; all
others are closed. (Slightly different notations are used elsewhere in Bresnan (1982a);
for example, Mohanan (chapter 8) uses the symbol coMP for xCoMP, and Kaplan and
Bresnan (chapter 4) and Bresnan (chapter 1) use the symbols vcoMP and scoMP for xcoMP
and CoMp, respectively.) The distinction between open and closed functions plays a role
in the theory of control (section 8). The subcategorizable functions in figure 1 correspond
to the governable functions of Kaplan and Bresnan (chapter 4 of Bresnan (1982a)); these
are the only functions to which lexical items can make reference.
Not all aspects of this classification are universal; in particular, the subcategoriz-
ability of TOPIC and FOCUs appears to be a typological property that distinguishes "'sub-
ject-oriented" from "topic-oriented" languages. In addition to these functions, there is
assumed to be a universal set of functional features, including CASE, TENSE, NUMBER,

Grammatical functions

Subcategorizable Nonsubcategorizable

ADJ( UNCT)

Semantically Semantically XADJ(UNCT)

unrestricted restricted

SUBJ OBL0

OBJ COMP

OBJ2 XCOMP

FOCUS*

TOPIC*

*The subcategorizability of these functions is a parameter that distin-

guishes "subject-oriented" from "topic-oriented" languages.

Figure 1
Classification of grammatical functions assumed in this theory

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
348 JOAN BRESNAN

GENDER, PERSON, FINITE, PRED, and SPEC. In the formal representation of functional
structures (Kaplan and Bresnan (1982)), function and feature names are formally distin-
guished by the types of their values.
The role of grammatical functions in the lexical-functional theory of syntax is to
provide the mapping between surface categorial structure and semantic predicate ar-
gument structure. This is done by assigning the grammatical functions semantic roles
in the lexicon and syntactic realizations in the categorial component of the syntax. A
fundamental constraint on all references to functions and functional features, whether
by lexical or by syntactic rules and representations, is the principle offunctional locality:
designators in lexical and grammatical schemata can specify no more than two function
applications (Kaplan and Bresnan (1982)). This principle means that function designators
can refer only to functions and features or to immediate subfunctions and features of
the f-structure to which they apply.

2. Lexical Encoding of Grammatical Functions

In the lexical-functional theory of syntax, lexical items subcategorize for functions, not
constituent structure categories, and lexical items exert their selectional restrictions on
a subset of their subcategorized functions. Lexical subcategorization for function rather
than structure explains the fact that when a function is freely realizable in a set of
different phrase structure positions, the subcategorization restrictions of any lexical item
that subcategorizes that function are satisfied by the entire set of positions (Grimshaw
(1981; 1982), Montalbetti (1981)). Moreover, it provides an invariant theory of subcate-
gorization for configurational and nonconfigurational grammars alike (section 4).
The predicate argument structure of a lexical item lists the arguments for which
there are selectional restrictions; the grammatical function assignment lists the syn-
tactically subcategorized functions (without repetitions) and may include a null symbol
';' the lexical form pairs arguments with functions. This study employs the notati
given in (1), which differ slightly from those in Kaplan and Bresnan (1982) and Bresnan
(1980; 1982b). (The notation given here provides a simple formal interpretation of the
coherence condition; see section 5.) (1) gives notation for representing properties of the
lexical item seem as it is used in sentences like John seems sick to Mary.

(1) a. predicate argument structure: SEEM(l ,2)

b. grammatical function assignment: {(XCOMP), (OBLGO), (SUBJ)}
c. lexical form: 'SEEM-TO((XcOMP)(OBLGO))(SUBJ)'

In this usage seem has a dyadic predicate argument structure, whose first argument
denotes a state of affairs (e.g. John's being sick) and whose second argument denotes
a perceiver of that state of affairs (e.g. Mary). (lb) gives the functions that are subcat-

' It is important to note that f is not a grammatical function, but a lexical symbol indicating that an
argument is semantically filled in the lexicon and is not assigned any function (Bresnan (1980), Halvorsen
(1981)). See examples (8) and (9) below.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 349

egorized by seem. In (lc) the function XCO

and the function (OBLGO) is assigned to
indicates that seem subcategorizes three
only on the XCOMP and oBLGO function
thematic" function, upon which seem im
to the XCOMP is determined by the theo
angled brackets in (la,c) is called the argu
icate's semantic arguments, which give r
(1982b)).
Although example (1) shows a verb that does not impose selectional restrictions on
its subject, it is well established that lexical items may also impose semantic selectional
restrictions on subjects. In (2) and (3), for example, admire selects an animate subject
while frighten selects an animate object:

(2) a. Mary admires John.

b. #lSincerity admires John.
(3) a. Mary frightens John.
b. #Mary frightens sincerity.

It has been argued that certain semantic asymmetries between subjects and non
subjects justify the suppression of the "logical" subject argument from lexical predica
argument structure (Marantz (1981)). (The term logical subject involves a confusion of
distinct levels of representation, as pointed out in Bresnan (1982b). In what follows, t
term logical subject argument will be used to refer to the thematic argument of a predicat
that has SUBJ assigned to it in the unmarked active lexical form.) According to this vi
the selectional restrictions in (2) would be attributed, not to the verb admire, but to
meaning of the verb phrase admires John. Nevertheless, many selectional restrictio
imposed on the subject ultimately reduce to properties of the lexical item, for subj
selection occurs when the only content of the verb phrase is the lexical predicate itse
as in (4).

(4) a. Mary will relax.

b. #Sincerity will relax.

The question is not, then, whether subject selection is a property of lexical items, but
whether this property is represented by the same predicate argument structure that
represents nonsubject selection.
Two rationales have been given for an asymmetrical representation of the logical
subject argument and other arguments in predicate argument structure (Marantz (1981,
50-51)). First, the "choice of object (or other argument of a verb) affects the semanti
role of the logical subject while choice of logical subject does not affect the semantic
role of the object," and second, there exist object idioms but no subject idioms with
free arguments. (Arguments similar to Marantz's have been attributed to Chomsky in
lectures.) That is, while many idioms express properties which are noncompositionally

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 350 JOAN BRESNAN

derived by combining a verb and an object (these are the so-called "object idioms"),
there are claimed to be no idioms that express properties which are noncompositionally
derived by combining a verb and a subject (these would be "subject idioms"). Examples
like (5a-d) are given by Marantz to illustrate these points.

(5) a. kill an insect

b. kill a conversation
c. kill a bottle (i.e. empty it)
d. kill an audience (i.e. wow them)

The wide range of predicates expressed by the examples in (5), including literal, figur-
ative, and idiomatic senses of kill, depends upon the choice of object of the verb.
As we will see, neither of these considerations is well-founded factually, but it must
also be recognized that the logic of the argument that is based upon them is faulty. Note
that both of these considerations are based upon the compositional semantics of sen-
tences. The assumption is that if the subject is always the last argument to be semantically
composed with the predicate, one can explain the generalization that the choice of
nonsubject arguments does not depend on the choice of the subject argument. By sup-
pressing the subject argument from the verb's predicate argument structure, one prevents
the subject from being combined directly with the verb before the verb and its nonsubject
arguments have been assembled into a predicate; the subject can then be semantically
composed only with a completely formed predicate. But the issue of whether or not the
subject argument has a special role in the semantic composition of the sentence is
logically independent of the issue of whether or not a subject argument position should
appear in lexical predicate argument structure. For example, in Dowty's (to appear)
theory, the subject is always the last argument to be semantically composed with the
predicate; yet the lexical function that expresses the meaning of a transitive verb in his
theory contains variables for both the subject and the object arguments. In short, one
could capture the subject/nonsubject generalizations without affecting the lexical rep-
resentation of predicate argument structure, simply by giving the subject a distinguished
role as final argument in the semantic composition of the sentence.
We will not take this approach, however, because the subject/nonsubject general-
ization itself is factually ill-founded, as mentioned above. First, there are in fact subject
idioms with free nonsubject arguments: for example, The cat's got x's tongue ('x can't
speak'), What's eating x? ('What is making x so irritable?'), Time's up (for x) ('The time
(for x) has expired'), x's goose is cooked ('x is in trouble and there is no way out').
Second, there are clear cases in which the semantic choice of a nonsubject argument
does depend upon choice of the subject. Consider, for example, (6a-c).

(6) a. The ceiling caved in on John.

b. The wall caved in on John.
c. The roof caved in on John.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 351

Example (6c) has a figurative or metaphorical sense ('Everything went wrong for John')
as well as the literal sense that part of the structure of a house collapsed on John; but
examples (6a,b) are unambiguously literal. Thus, the choice of the roof as subject of
caved in gives rise to a special meaning. Under this special meaning, one can choose
as the object of on an abstract noun phrase which cannot occur with the literal meaning:

(7) a. #The ceiling caved in on John's dreams.

b. #The wall caved in on John's dreams.
c. The roof caved in on John's dreams.

How do we know that it is not the object of on which gives rise to the special meaning
and thereby determines the choice of the subject the roof? When we omit the on-phrase
Altogether, we find that The roof caved in still has the figurative sense 'Everything went
wrong' while The ceiling caved in and The wall caved in lack it. It is not difficult to find
other examples which support the same conclusion (e.g. A truck hit John vs. An idea
hit John, pointed out by K. P. Mohanan).
In conclusion, there is no justification for suppressing the subject argument from
predicate argument structure. Moreover, theories of lexical representation which omit
the subject argument from predicate argument structure so as to give it the special status
of final argument in the semantic composition of the sentence are incompatible with the
evidence of subject idioms and the subject-determined selection of nonsubject functions.
By separating the predicate argument structure from the grammatical function as-
signment, as in (la,b), we open the possibility of finding principles which will enable us
to predict the possible function-argument correspondences and so derive, or at least
narrowly constrain, the set of lexical forms. Current research in lexical representation
suggests that there are universal constraints on the function-argument pairings in lexical
forms. One constraint which is discussed elsewhere in Bresnan (1982a) is function-argument
biuniqueness: in each lexical form, the predicate arguments and the functions they are
paired with must be in one-to-one correspondence (Bresnan (1980), Grimshaw (1982)).
Recent research suggests that there are also semantic constraints on function-argument
pairings: there are semantically restricted functions that can only be paired with argu-
ments of specified semantic types (see Rappaport (1980)). The specified types of argu-
ments may include the thematic relations AG(ent), TH(eme), ExP(eriencer), so(urce),
Go(al), Loc(ation), DIR(ection), BEN(eficiary), INsTR(umental), as well as MNR (manner),
MEANS (a secondary TH), CAUSEE (a secondary AG), PART, PATH, QUANT(ity), and
PROP(Ositional). (Nothing rests on the details of the classification of argument types; see
Jackendoff (1976) and Amritavalli (1980) for some discussion.) To enforce the semantic
constraints on function-argument pairings, it is convenient to assume that there is a
labeling function from the argument types to the predicate arguments of each predicate
argument structure. As illustrated in (8), a single argument may bear more than one label
(Jackendoff (1972)).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 352 JOAN BRESNAN

(8) a. SEEM-TO (PROP, GOAL)

EXP

b. BUY (AG, TH, SO)

GO

C. SELL (AG, TH, GO)

SO

A semantically restricted function, then, can be paired only with an argument one of
whose labels matches its semantic type. In particular, the oblique functions can be paired
only with an argument type whose index they carry: for example, OBLAG must be paired
with an AG argument; OBLGO, with a GO; and OBLLOC, with a LOC. The open and closed
complement functions XCOMP and COMP can be paired only with what may be broadly
referred to as "propositional" arguments (those labeled PROP). COMP and XCOMP differ
in their syntactic encoding properties (section 4) and in their control properties (sections
8, 9). The null symbol 4) is not a function; it appears only in argument positions, where
it signifies that the argument is lexically interpreted and no function is assigned (footnote
1). For example, in (9) 4) indicates that the AG iS semantically filled in the mediopassive
lexical form of read, and in (10) it indicates that the TH iS filled in the intransitivized
lexical form of the same verb:

(9) a. The novel reads easily.

b. READ ( ) (SUBJ))
AG TH

(10) a. John reads frequently.

b. READ ((SUBJ) 4 )
AG TH

We can formulate these restrictions as follows. Let G be a semantically restricted func-

tion and R its associated argument type. A designator of G is any functional designator
in the sense of Kaplan and Bresnan (1982) which mentions the function G: for example,
( T SUBJ) is a designator of SUBJ. Now we will require that in every lexical entry, every
lexical designator of G must be assigned to an argument of type R, either by appearing
in an argument list at an argument of that type or by being equated with another lexical
designator that does so. (The latter condition allows an extraposed COMP function, for
example, to be equated with a SUBJ.) Note that these restrictions specify the only pos-
sibilities for lexical items to refer to semantically restricted functions; this will have
consequences for the theory of control (section 8), for it restricts the set of lexically
induced functional controllers ((20)) and the set of lexically induced functional anaphors
((35)). Because the open functions have broader syntactic properties, their designators
may also appear in certain other contexts (see sections 4, 8, 9).
Unlike the semantically restricted functions, the semantically unrestricted functions
SUBJ, OBJ, OBJ2 may be paired with any argument type or remain unpaired with an
argument (as in the case of "nonlogical" subjects). There are nevertheless important
constraints on their assignment to lexical forms. In particular, there appears to be a

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 353

hierarchy for the assignment of SUBJ, O

the unmarked case, OBJ2 is assigned on
has been (cf. Bresnan (1980) and Rappaport (1980)).
Lexical rules can alter the assignment of functions to predicate argument structures
and only lexical rules can do so, given the principle of direct syntactic encoding (Bresnan
(1982b)). Lexical rules will not be motivated in this study; see elsewhere in Bresnan
(1982a) and Roberts (1981), Rappaport (1980), Neidle (in preparation).

3. Categories

Constituent structure categories are decomposed into a type, or level of structure, a

afeature matrix of universal categorial features as in Bresnan (1975; 1976a). The feat
and types of major categories assumed in this study are shown in figures 2 and 3,
respectively, with their common notations. As we will see (section 4), the categorial
features "predicative" and "transitive" of figure 2 can be defined in terms of the
functional primitives SUBJ and OBJ. Hence, we can eliminate primitive categorial features
from our theory altogether. For simplicity in representing c-structure trees in what
follows, the convention will be adopted that when a nonsentential category of type 1 is
exhaustively dominated by a category of type 2 or exhaustively dominates a category
of type 0, the type 1 category will be suppressed.
The categories of type 0 are called lexical categories; the categories of types 1 and
2 are called projections; projections of the highest type (i.e. type 2 in figure 3) are called
maximal projections; and lexical categories and their projections are all considered major
categories. S and S are considered to be major categories which are projections of no
lexical category (Hornstein (1977), McCloskey (1979)). There are also minor categories
of null or degenerate type, including DET and COMP (Bresnan (1975; 1976a)) and
certain uses of P(reposition) (see section 4). Not all languages have instantiations of all
categories. For example, Warlpiri does not instantiate the category A; instead, the
category N functions as both A and N (Nash (1980), Simpson (in preparation)). Nor do
all languages distinguish all types in every category; for example, many VSO languages
do not instantiate V' or V". In short, the instantiation of both types and features of
categories is a source of variation among languages.

"predicative" "transitive"

V + + verbal
P + + pre- or postpositional
N ? - nominal
A + - adjectival
S - sentential

Figure 2
Feature matrices of major categories

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
354 JOAN BRESNAN

type 0 1 2

category V V' V" (VP)

P P' P" (PP)

N N' N" (NP)

A A' A" (AP)

S s

Figure 3
Major category types

C-structure rules are context-free rewriting rules (or recursive transition networks)
defined over the vocabulary of major and minor categories. The set of possible rules is
narrowly constrained, as discussed in the next section. Natural classes of categories and
rule schemata can be designated by the X-bar (X-prime) notation as in Bresnan (1975;
1976a). For example, X = {V, P} and X" = {VP, PP, NP, AP}.
[+t] [+p]

4. Syntactic Encoding of Grammatical Functions

Grammatical functions are assigned to the particular c-structure rules and inflectional
features of each language. The possible syntactic encodings of functions into structure
are highly constrained. In configurational encoding a basic form of c-structure rule is,
for any categorial feature matrix X, Xn + l> Cl . . Xn . . . Cm, where n - 0 and Ci is
either a minor category or a maximal projection. (Falk (1980a) suggests that this basic
rule form is further decomposed into separate specifications of dominance and prece-
dence relations.) For this rule form, the basic principle of configurational encoding is
to associate a function-assigning equation ( T G) = I with each Ci if and only if C,
is a maximal projection, and to associate the equation T = I elsewhere. A major
category bearing the equation T = I is called the head. Hence, according to the basic
principle of configurational encoding, all and only the maximal projections of categories
can bear functions, every phrase has a unique head, and the functional features of a
phrase are identified with those of its head. It follows immediately that in predominantly
configurational languages, only maximal projections of categories will appear to be
subcategorized for (section 1); this is a widely recognized characteristic of subcatego-
rization.
The X-bar theory must be elaborated to permit other rule forms such as those for
coordinate structures and for exocentric constructions (Bloomfield (1933)). The syntax
and semantics of coordination in the present theory are discussed in Andrews (1981),
Peterson (1981a), and Halvorsen (1981). Among the exocentric constructions of English
are the sentence (e.g. John walks slowly) and the gerund phrase (e.g. John's walking

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 355

slowly); in these the head is VP, a max

is not a projection of V. Examples are given in (11) and (12). (Here I am adapting the
analysis of the English S proposed by Falk (1980b).) In the gerund construction of rule
(12), the case-marking constraints are omitted.

(11) S NP VP
( SUBJ) = I T = I
(12) NP NP 's VP
(T SUBJ) = t =1

Because it is the projection of no lexical category, S is an exocentric category in all

languages. S can be headed by VP (as in English and other SOV languages), by V (as
in VSO languages), by nominals (as in Warlpiri (Simpson (in preparation))), and by
adjectival phrases (as in Russian (Neidle (in preparation))).
Exocentric rules fall under the rule form Xn _ Cl ... Cm, where n > 0 and Ci is
either a minor category (e.g. 's) or a maximal projection of a lexical category. The
principle of function assignment for rules of this form is to associate the equation
T = I with a single major category (the head) and to any minor categories, and to
associate a function assignment equation ( T G) = I with all other maximal projections.
Consequently, for both exocentric and endocentric categories, the head is a major cat-
egory annotated with j = l, while all other major categories (and only these) are
annotated with ( G) = I. Thus, it remains true that in configurational encoding, only
the maximal projections of categories can bear functions, every phrase has a unique head,
and the functional features of a phrase are identified with those of its head.
In nonconfigurational encoding, the basic form of c-structure rule is C -* X*, where
C is a major nonlexical category and X is a lexical or nonlexical category. The basic
principle of nonconfigurational encoding is to associate pairs of function-assigning and
feature-assigning equations of the form given in (13a) with arbitrary X:

(13) a. f(IF)= V
|( G) = I J
b. I = I
In (13a), G is a function selected by the value v of the feature F. For example, taking
F to be CASE and v to be NOM or ACC, (I CASE) = NOM could be associated with ( T SUB
= i, and (j CASE) = ACC with ( T OBJ) = I. These pairs of schemata are arbitraril
associated with categories in c-structure rules; see Mohanan (1982a) for a detailed ex-
ample. The schema T = I in (13b) is also associated arbitrarily with categories in
c-structure rules. (The association with T = I could also depend upon some inflectiona
feature.) The head of C is defined to be any major category which is annotated with
T = I and which has a PRED. It follows from the consistency condition (section 5 and
Kaplan and Bresnan (1982)) that the head is unique.
In nonconfigurational encoding as in configurational encoding, every phrase has a
unique head, and the features of a phrase are identified with those of its head. However,

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 356 JOAN BRESNAN

in nonconfigurational encoding, functions need not be assigned to maximal projections;

instead, they may be assigned to submaximal projections, or even to single lexical
categories, as in the example given in figure 4. This example is based on the analysis
of Warlpiri given in Simpson and Bresnan (in preparation). Although alternative assign-
ments of functions to categories are possible in figure 4, the conditions on well-formed-
ness of f-structures (section 5) eliminate "incorrect" assignments, admitting only those
that yield consistent, coherent, and complete f-structures.
In sum, the fundamental difference between configurational and nonconfigurational
syntactic encoding lies in the surface realizations of grammatical functions. In config-
urational encoding, functions are identified by the category and by the order of maximal
constituents within the immediately dominating phrase, while in nonconfigurational en-
coding, functions are identified by the case and other inflectional features of unordered,
possibly submaximal, constituents. It follows from this theory that in predominantly
nonconfigurational languages, only the case marking of constituents will appear to be
subcategorized for, and not their phrase structure configurational properties.
Languages may employ both types of encoding, as observed by Mohanan (1981a).
In English, the OBJ and OBJ2 are configurationally encoded. Rule (14) illustrates this.

(14) V'-> V ( NP 0 ( NP j PP*

I = t \([ tOBJ)= 1 J OBJ2)= I J ( OBLO) =
(I PCASE) = OBL0 J

In (14), the function names OBJ and OBJ2 are assigned, respectively, to the first and secon
NPs immediately dominated by VP. However, the OBLH function names are assigned on
the basis of the features of the unordered PPs. The pair of equation schemata assigned
to PP* in (14) actually abbreviates the disjunction of a finite set of pairs, one for each
value of 0: {( t OBLAG) = |, (4. PCASE) = OBLAG} or . . . or {( ' OBLGO) = i ,
(4. PCASE) = OBLGO}. The PCASE features are borne by prepositions, which can serve
as case markers in English (Bresnan (1979)). For example, the preposition to carries the
lexical information ( T PCASE) = OBJGo and the preposition by carries the equation
( T PCASE) = OBLAG. Since the value of the PCASE feature is always identical with the
function name of the PP in rule (14), we can abbreviate the disjunction of these pairs
of equations by substituting equals for equals. Thus, we replace the symbol OBLH in
( T OBLO) = by the designator ( 4 PCASE), whose value is OBL0: the result is the single
equation (4. (4. PCASE)) = 4 that appears in rule (15).

(15) V' V NP NP PP*

T = 4 ( OBJ) = 4 ( 4 OB
We see, then, that despite its superficial dissimilarity from the schemata in (14), the
English PP schema in (15), which is discussed in Kaplan and Bresnan (1982), is an
instance of nonconfigurational encoding. (It should be remembered that the rules just
discussed are not the only sources of PP in English; there are PP adjuncts and comple-
ments as well (Bresnan (1979; 1980)). The introduction of various types of PPs under

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 357

c-structure
S

N AUX V N N

P3

kurdu-ngku ka wajili-pi-nyi maliki-0 wita-ngku

child-erg asp running-do-nonpast dog-abs small-erg
'The small child is chasing the dog.'

syntactic encoding of functions

I(tSUBJ)= 1 (TOBJ)= 1
l (CASE)=ERG ,et.(CASE)=ABS

f-structure

SUBJ CASE ERG'I

PRED CHILDE

ADJ SUBJ PRED PRO']

PRED 'SMALL ((SUBJ))'

ADJ L CASE ERG i
PRED &CHASE((SUBJ)(OBJ))'

TENSE NONPAST

ASPECT PRES-IMPERF

OBJ CASE ABS

L PRED DOG']

Any order of constituents is possible (so long as AUX is seco

grammatical relations of the sentence remain invariant. The f
are syntactically encoded by case. (Simpson and Bresnan 1982)

Figure 4
Nonconfigurational encoding in Warlpiri

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 358 JOAN BRESNAN

different nodes, such as S, VP (V"), and V', may give rise to some ordering constraints
among prepositional phrases.)
A second example of mixed configurational and nonconfigurational encoding in one
grammar is the rule S -* X AUX X* of Warlpiri, where the minor category AUX has
a fixed position in the configuration of S (Hale (1979)); in addition, there is evidence
that Warlpiri has optional phrasal constituency in nominals as well as infinitival phrase
constituents (Nash (1980), Simpson and Bresnan (in preparation), Simpson (in prepa-
ration)).
Fundamental to X-bar theory are the principles of syntactic encoding which relate
possible functions to categorial features. For example, it was pointed out in Bresnan
(1975; 1976a) that the "verbal" ([ + V]) categories V and P take direct object NPs while
the "nominal" ([ + N]) categories A and N do not. Subsequently, Jackendoff (1977) made
the important observation that the categories of natural language are better defined in
terms of their relations to the subject and object functions than in terms of the categorial
features [? N], [? V] of Chomsky (1970). This insight has a very natural expression in
lexical-functional theory. The "intransitive" (i.e. [-t]) categories are those which do
not permit ( T OBJ) = I to be annotated to any symbol within the phrase structure rules
that expand the intransitive category. This means that adjectival phrases and noun
phrases may not contain phrasal direct objects, while verb phrases and pre- or post-
positional phrases may.2 The "predicative" (i.e. [+ p]) categories are those which do
not permit ( T SUBJ) = I to be annotated to any symbol within the phrase structu
rules that expand the predicative category. This means that VP, AP, and predicative
NP and PP cannot contain phrasal subjects, while S and nonpredicative NP and PP may.
(We have already seen that S and NP may contain phrasal subjects; a possible example
in English of a PP which contains a phrasal subject is the absolutive PP construction
involving with: With John impossible to talk to, Mary left the room, With John away,
Mary was happy.) Thus, we see that we can eliminate primitive categorial features from
our theory altogether.
Not only can the basic phrase structure categories of natural language be defined
in terms of the primitive functions SUBJ and OBJ, but the function of a category is a fairly
good predictor of what are likely to be the categorial and case features of the category.
For example, knowing that a category C is a SUBJ, OBJ, or COMP, one can predict that
C is likely to be (respectively) a nominative NP, an accusative NP, or an S, although
S can also have the SUBJ function, and PP, the OBJ function (Grimshaw (1981)), and
although nominative NPs can be objects and accusative NPs, subjects (Andrews (1982)),
Mohanan (1982a)). Where function and category diverge from their typical correlations,
it is the function that correctly predicts the direction in which linguistic rules generalize
in such domains as subcategorization, agreement, and control (Grimshaw (1981), Neidle
(1982; in preparation), Mohanan (1982a), Simpson (in preparation), Simpson and Bresnan

2 It is clear that As and Ns may have oblique NP objects (Maling (to appear)). In Russian, adjectives
appear to have NP objects; whether or not they must be analyzed as direct objects deserves further investi-
gation.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 359

(in preparation)). To express the typical, or unmarked, relations between function and
category, however, we postulate substantive constraints on the pairing of functions with
categorial and case features. For example, we may say that NOM case is the unmarked
feature encoding the SUBJ function. Similarly, we may postulate substantive constraints
on the pairing of functions with categories. In particular, the predicative/nonpredicative
categorial feature is correlated with the open/closed function distinction.
There are two logically possible ways of pairing the predicative/nonpredicative
categories with the open/closed functions: we could say that the predicative categories
must have closed functions, and the nonpredicative categories, open functions; or the
opposite. It turns out that other properties of our theory dictate which correlation is
correct: a category is predicative if and only if it can be assigned an open function.3
This principle of association is motivated by very general conditions on f-structures;
only by the function-category associations asserted in this principle can it be ensured
in all cases that the f-structures of categories will be consistent and complete (section
5) with respect to their subject functions. For predicative categories lack structural
subjects; the assignment of an open function to them induces functional control, which
obligatorily provides a subject from a category-external source (section 8.2). If these
categories were assigned closed functions, a subject could not always be ensured. In
contrast, nonpredicative categories may have structural subjects; the assignment of a
closed function to them precludes functional control and thereby prevents the assignment
of a subject from an external source, which could create inconsistent f-structures. As
we will see below (sections 9.2, 9.6), this principle of function-category association has
a number of explanatory consequences.
We digress briefly to point out that the analysis of PPs given here differs in two
ways from the analysis presented in Kaplan and Bresnan (1982). First, wherever possible,
the PCASE values used here are drawn from a universal set of feature values {OBLGO,
OBLAG, OBLINST, etc.} rather than from the English-particular set {TO, BY, WITH, etc.}.
This is motivated by the need for a universal format for representing grammatical re-
lations (see Bresnan (1981), Pinker (1982), and Halvorsen (1981)). The choice of which
prepositions mark which oblique functions appears to depend upon the meaning of the
prepositions in their predicative uses. In This road is to London. the preposition to
designates a semantic relation between the SUBJ (this road) and the prepositional OBJ
(London), in which the object is the goal of some motion. Similarly, in This work is by
a famous sculptor, the preposition by designates a semantic relation between the SUBJ
(this work) and the prepositional OBJ (a famous sculptor), in which the object is the agent
of some action. We can therefore assume that each case-marking preposition is selected
from the set of semantically related predicative prepositions. That is, the preposition
that serves as a formal marker of the OBLH function is chosen from the set of prepositions
whose lexical forms associate their prepositional OBJ with the argument type 0. This

3 Fassi Fehri (1981) has proposed the further restriction that only S can be assigned the closed COMP
function.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 360 JOAN BRESNAN

assumption supports the "semantic bootstrapping" theory regarding acquisition of gram-

mars described in Pinker (1982) and also accounts for the fact that the case-marking uses
of prepositions appear to be related to their predicative uses cross-linguistically. There
do, however, appear to be some idiosyncratically marked oblique objects. To account
for these, we will extend our notation to OBLFORM, where FORM is a specific preposition
name. (For example, in John relies on Mary, we could say that on Mary is an OBLON.)
Future research may reveal a semantic motivation for these apparently idiosyncratic
markings.
The second difference between this analysis of PP and the analysis of Kaplan and
Bresnan (1982) is that the oblique PP is treated here as an exocentric constituent in
which the NP actually functions as the head of the PP and the P is a minor category that
contributes only the PCASE feature. Thus, it is assumed that in structures of the form
[pp P NP], either P is the head and NP is an OBJ (yielding the endocentric, predicative
PP) or else P is a minor category and NP is the head (yielding the exocentric, oblique
PP). The motivation for this is to eliminate compound names (e.g. TO OBJ in Kaplan and
Bresnan (1982)) for oblique objects. Andrews (1982) has shown that Icelandic verbs which
subcategorize for oblique prepositional objects may also govern the case marking of the
NP within the PP; if the oblique prepositional object had a compound name, the case-
marking restriction on this NP would violate the functional locality principle (for ex-
ample, no verb could specify "( T TO OBJ CASE) = (ACC)").'
In conclusion, this theory of syntactic encoding has three main results. First, it
provides a characterization of the syntactic notion head of a phrase which holds uni-
versally for endocentric and exocentric constituents and for configurational and non-
configurational structure types. Second, it unifies the theories of lexical subcategori-
zation for configurational and nonconfigurational languages. And third, it identifies the
basic parameter of syntactic variation for configurational and nonconfigurational lan-
guages (see Simpson and Bresnan (in preparation)).

5. Representation of Grammatical Relations

The lexical and syntactic encodings of grammatical functions determine a mapping, or

a set of associations, between the word and phrase configurations of a language and its
semantic predicate argument structures. The grammar of the language assigns to each
sentence a set of these associations between surface form and predicate argument struc-
tures; these are the grammatical relations of the sentence.
Grammatical relations are formally represented by f-structures, which are pairs of

4 Andrews (1982) presents an analysis of PPs that differs from the one assumed here. Given our revision,
case marking by a preposition can still be distinguished in f-structure from case marking by a nominal inflection,
in that the former is expressed by PCASE and the latter by CASE. An example of the need for this distinction
is the use of prepositional and nominal instrumentals in Malayalam (Mohanan (1982a)); to express OBLINsTR,
the instrumental case and the instrumental postposition are used in free variation; to express the OBLAG in
passives, only the instrumental case is used; and to express the causativized SUBJ, only the postposition is
used.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 361

fnames and fvalues. (This differs from the terminology used in Kaplan and Bresnan
(1982), where fnames and fvalues are referred to as attributes and values.)

(16) fname, fvalue1

fnamen fvaluenj

The uniqueness condition on f-structures (also called consistency) requires that every
fname have a unique value. An fname is a symbol denoting one of a universal set of
features and functions; an fvalue is a symbol (denoting one of a universal set of feature
values), a semantic form, an f-structure, or a set of fvalues. Semantic forms differ from
symbols in that they are uniquely instantiated when their designators are instantiated
(Kaplan and Bresnan (1982)). Consequently, two otherwise identical f-structures with
separately instantiated semantic forms are formally distinct. For linguistic motivation
for this property, see Grimshaw (1982) and Montalbetti (1981). There are two kinds of
semantic forms: those with argument lists-called lexicalforms-and those without.
Certain f-structures are distinguished as clause nuclei; these contain an fname PRED
whose fvalue is a lexical form (i.e. a predicate argument structure paired with a gram-
matical function assignment). For example, the f-structure shown in figure 5 has two
clause nuclei, the f-structure labeled f, and its subsidiary f-structure labeled f2. (The
f-structure of figure 4 contains three clause nuclei.)
The clause nucleus is the domain of lexical subcategorization in the sense that it
makes locally available to each lexical form the grammatical functions that are subcat-
egorized by that form. Lexical subcategorization is enforced by the requirement that
every f-structure be both coherent and complete. An f-structure is locally coherent if
and only if all of the subcategorizable functions that it contains are subcategorized by
its PRED; an f-structure is then (globally) coherent if and only if it and all of its subsidiary
f-structures are locally coherent. Similarly, an f-structure is locally complete if and only
if it contains values for all of the functions subcategorized by its PRED; and an f-structure
is then (globally) complete if and only if it and all of its subsidiary f-structures are locally
complete. This treatment of coherence and completeness permits a very simple formal
interpretation: to be locally complete and coherent, an f-structure must contain sub-
categorizable functions GI, .. . , Gn if and only if it contains a PRED whose value is a
lexical form with the grammatical function assignment {G1, . . . ,
The level of f-structure differs crucially from representations of constituent structure

5 Note that in this formulation of the coherence condition, local is interpreted to mean 'in the same
f-structure'. There is an alternative formulation of coherence, based on the functional locality principle (section
1), which interprets local to mean 'in the f-structure that immediately contains or is contained in this
f-structure': an f-structure is locally coherent if and only if the values of all of the subcategorizable functions
that it contains are subcategorized by a PRED; by the functional locality principle, only a local PRED (in the
second sense) can subcategorize these function values (cf. Kaplan and Bresnan (1982)). The two formulations
have empirically distinguishable consequences; for example, the formulation adopted in the text above is
inconsistent with the use of compound function names such as "TO OBJ". Further research is required to
determine which interpretation of coherence is optimal. Note also that, to give a uniform definition of head,
we have adopted the instantiation procedure of Kaplan and Bresnan (1982, footnote 6).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
a. c-structure c. f-structure

S fl SUBJ GEND MAS

NUM S

NP VP PERS 3
CASE N

LPRED P
V AP PP
TENSE PRES

PRED SEEM-TOK(XCOM
A P NP
OBLGO FPCASE OBL

CASE A

he seems sick to her GEND FEM

NUM

PERS 3

PRED P

XCOMP f2 SUB
_ andfa sPRED SICK

f, and f2 are cla

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
b. c-structure annotated with functional schemata
S

(TSUBJ) = =
NP VP

T = l (IXCOMP)=t (T(IPCAS
V AP PP

|=4 T=i (ICA

A P

(1GEND)=MASC (1PRED)= (1PRED)= (1PCASE)=OBLGO (

(TNUM)=SG 'SEEM-TO((XCOMP)(OBLGO))(SUBJ)' 'SICK ((SUBJ)) t

(1PERS)=3 (1TENSE)= PRES sick (IP
(TCASE) =NOM (TSUBJ NUM)=SG (TCAS
(TPRED)= PRO' (TSUBJ PERS)=3 (TPRE
he (IXCOMP SUBJ)=(TSUBJ
seems

Figure 5
Clause nuclei

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 364 JOAN BRESNAN

in that it represents grammatical relations universally, locally with respect to lexical

subcategorization, and compositionally with respect to semantic interpretation (Bresnan
et al. (in preparation), Simpson and Bresnan (in preparation)). See Halvorsen (1981) for
a compositional procedure for the semantic interpretation of f-structures, which accounts
for quantification, adverb scope, control, the distribution of intensional objects, and the
interpretation of idiom chunks and grammatical expletives.

6. Representation of Constituency

The representation of grammatical relations by f-structures rather than by c-structures

permits a massive simplification in the representation of the constituency relations of
natural language. These can now be represented by a context-free phrase structure
grammar (or a recursive transition network), defining a single level of c-structure rep-
resentation over a small set of universal features and types.6 The form of c-structure
rules is largely predictable from the theory of syntactic encoding, and the categorial
features are defined in terms of the primitives SUBJ and OBJ (section 4). Since there is
only a single level of c-structure to represent the syntactic constituency relations among
the words of a sentence, syntactic operations that restructure the dominance or pre-
cedence relations among the words and their affixes are eliminated. Consequently, only
fully inflected and morphologically complete words are lexically inserted into phrase
structures, as Selkirk (1981), Lieber (1980), Lapointe (1980), Nash (1980), and Mohanan
(1982c) have argued on independent grounds. A further consequence is that any "re-
structuring" rules that alter word structure (as proposed by Chomsky (1977a), Hornstein
and Weinberg (1981)) must be presyntactic rules of morphology, as justified by Bresnan
(1982b), Rothstein (1981), Peterson (1981b). In addition, it follows that there are no
structure-dependent operations that map c-structures into c-structures, although there
are structure-dependent operations that map c-structures into f-structures, and there
may be restricted kinds of operations that map c-structure rules into c-structure rules
in a function-preserving fashion (ct. Kaplan and Bresnan (1982), Bresnan (1982b), Gazdar
(in press)). Thus, syntactic transformations are eliminated.
Finally, the use of null c-structure can be virtually eliminated, restricting both the
expressive and the generative power of grammars (Kaplan and Bresnan (1982)). The
formal theory of grammatical representation permits a nonterminal node to exhaustively
dominate the empty terminal string e only in cases of constituent control, the long-
distance structure-dependent relation that characterizes "wh-movement" and similar
constructions, as illustrated in figure 6.

6 By themselves, context-free grammars are not sufficient to represent the constituency relations of
natural language. Bresnan et al. (in preparation) show that there is no context-free phrase structure grammar
that strongly generates just the set of empirically motivated c-structure representations for Dutch. However,
the correct representation of constituency relations can be strongly generated by letting the lexical subcate-
gorization conditions on f-structures filter the associated c-structures.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 365

7. Government

Government refers to the ability of lexical items to determine the features of other
constituents. The term agreement is often used in preference to government when an
inflectional morpheme of a word and an inflectional morpheme of a governed constituent
have mutually constraining features. Prepositions govern the case of their objects (An-
drews (1982)); verbs and adjectives govern the case, person, or number marking of their
subjects, objects, and oblique objects (Andrews (1982), Mohanan (1982a,b), Maling (to
appear), Craig (1977), Roberts (1981), Robertson (1976)); complementizers and verbs
may govern the finiteness, mood, and constituent control features of their clauses
(Zaenen (1981), McCloskey (1979)).
The fundamental problem of government is to explain the observed structural re-
lations between governing morphemes and the constituents they govern. The theory of
syntax developed in the preceding sections already provides a solution to this problem.
Inflectional affixes are lexical items which must be combined with other lexical items
prior to lexical insertion (Selkirk (1981), Lieber (1980), Bresnan (1982b), Mohanan
(1982c)). Syntactic information may be encoded in any lexical item in the form of func-
tional features; inflectional affixation sums the feature set of the inflectional morpheme
and the feature set of the lexical category that it inflects (Bresnan (1982b)). After lexical
insertion, the features of a word become the features of the phrase of which the word
is either the head or a minor category (as defined in section 4). Thus, referring to figure
6b, we see that the features of the preposition become features of the PP that dominates
it, the features of the determiner become features of the NP, and the features of the
verb become features of the S.
Consider in particular how a verb bearing the present tense inflectional affix -s
governs (i.e. agrees with) features of the subject. As in Kaplan and Bresnan (1982) and
in Bresnan (1982b), -s encodes the information that the SUBJ person is 3 and the SUBJ
number is singular. When -s is affixed to seem, this information becomes part of the
feature set of seems. When seems is lexically inserted into V in figure 6b, the feature
set is propagated up the tree: by the T = t equations, the features of the (f-structure
of the) V become features of the (f-structure of the) VP, and the features of (the
f-structure of) VP become features of (the f-structure of) S and ultimately of S. The
uniqueness principle guarantees that any subject that the clause contains will have the
features required by the verbal affix and vice versa. Now suppose, hypothetically, that
a morpheme similar to -s had been affixed to whom, the prepositional object of to. At
lexical insertion, the features of this hypothetical whom-s bearing the information about
the number and person of the SUBJ would again be propagated up the tree-but only as
far as the PP dominating whom. Because this PP is neither a head nor a minor category
in S or VP, its features will not propagate to S. Hence, the syntactic information bor
by the affix will remain within the PP without affecting the SUBJ of the clause.
Our theory of syntax thus provides an explanation for the fact that governing mor-

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 c. f-structure

SPEC THE'
a. c-structure

NP PRED PERSON
ADJUNCT TOPIC R

A ~~~~~~~~~~~~~~CASE NO
PCASE OB
NP S CAK AC C
PRED 'P
DET N PP S SUBJ PRED 'PRO

NUM SG
P NP NP VP K PERS 3
GEND M

V AP P TENSE PRES
PRED SEE
A XCOMP [SUBJ SC((U)

the person to wm he sGPRED SICK (

the person to whom he seems Nick e oBLGo

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
b. c-structure annotated with functional schemata
NP

T14 (TADJUNCT)= 1
NP S

T = l 1 =l( 1TOPIC) = T 1
DET N 1S 4pp S
(1R) =C +
PP

T = , (,CASE) =ACC (1SUBJ) T=4

P T=1 NP VP
NP

T=1 (TxCOMP)=
V AP

(1SPEC)= (TPRED)= (TPCASE)= (TPRED)='PRO' (TPRED)='PRO' (TPRED)= (IP

THE' 'PERSON' OBLGO (TCASE)=ACC (TCASE)=NOM 'SEEM-TO...'
the person to (1R)= + (IGEND)=MASC sick
whom (TNUM)=SG seems
(TPERS) = 3
he

Figure 6
Constituent control

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 368 JOAN BRESNAN

phemes appear either in the heads (or heads (of heads ... )) or in minor categories of
the phrase whose constituents they govern. There is no need to stipulate what the
domains of government are, as much recent work does (Chomsky (1981)): the possible
governing relations follow from the lexical theory of inflectional affixation (section 6)
and the theory of syntactic encoding (section 4)-both of which are independently
motivated subtheories.
This theory has a further consequence of interest. Because government relations
follow from the functional relations of categories and not from a set of stipulated struc-
tural configurations, the same kinds of government relations can appear in languages
with very different c-structure configurations as well as in different c-structures within
the same language. Figure 7 illustrates several different configurations in which verbal
affixes govern (i.e. agree with) the person and number or case features of the SUBJ (and
OBJ): number and person agreement in English (figure 7a); case and person agreement
in the Mayan language Jacaltec (figure 7b) (see Craig (1977, 108)); animacy and person
agreement in the Athapaskan language Navajo (figure 7c). In this last example, the same
affix -yi- determines the number and person of the SUBJ and OBJ (Roberts (1981)). Figure
8 illustrates two configurations in Dutch in which the verb glimlachen 'to smile' governs
the oblique PP object, determining the choice of the preposition naar 'at'; for justification
of these structures, see Evers (1975) and Bresnan et al. (in preparation). In both structures
8a,b, V3 governs the preposition, but this similarity is not captured at the level of
c-structure representation; rather, it is expressed in the functional relation between the
verb glimlachen and its oblique object, which holds regardless of the very different
structural configurations that express it. This result follows from the uniqueness prin-
ciple: the functions in each functional structure must be unique; hence, the XCOMP within
the XCOMP in figure 8b, though it is described in discontinuous c-structure positions
(namely, at both VP3 and V3), forms a single functional unit containing both the OBL0
and the PRED of glimlachen. See Bresnan et al. (in preparation) for details.
We see, then, that a word may govern the functions in the f-structure that imme-
diately contains the word's features. No set of c-structure categories provides an in-
variant universal characterization of this domain of government. However, given the
syntactic encoding of functions into the c-structure rules of a particular language or
language type, one can predict in which c-structure configurations government relations
will hold in that language or type.
These results-the explanation of the distribution of governing morphemes, the
uniform cross-configurational characterization of government relations, and the pre-
dictability of the actual structural configurations in which government relations can
appear-depend on the assumption that verbs govern their subjects. But this assumption
is in fact a necessary consequence of the theory of syntax adopted here. Recall from
the theory of lexical encoding that lexical items exert selectional restrictions on a subset
of their subcategorized functions (section 2), and that verbs impose selectional restric-
tions on their subjects. Thus, verbs are subcategorized for their subjects. But then it
follows from the completeness and coherence conditions (section 5) that f-structures

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 369

a. English (numnber and personi agreemiienit)

S

(TSUBJ)=J4 T =
NP VP

T=4, (TOBJ)=J
V NP

he like-s them
I

b. Jacaltec (case and person agreement)

S

T=41 (41SUBJ)=4. (TOBJ)=4.

V NP NP

x-o-s-watx'e naj te' iiah

asp-A3-E3-make cl/he cl/the house
L I I I I
'He made the house.'

c. Navajo (animacy and person agreement)

S

(TSUBJ)=J (ToBJ)=J T1=4.

NP NP V

dzaane'ez yi-o-ztat
horse mule S3-S3-kicked
03
SANIM

'The horse kicked the mule.'

Figure 7
Verbal inflections governing (agreeing with) properties of SUBJ

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 370 JOAN BRESNAN

which contain the PREDS of verbs must also contain SUBJS (unless, of course, a particular
verb in question is subjectless, as in cases of "spontaneous demotion" (Comrie (1977),
Sridhar (1979))). Therefore, verbs govern their subjects.
By the same reasoning, we can establish the more general consequence that lexical
items govern all of their subcategorized functions. Since we have assumed that the only
functions that can be referred to in lexical entries are the subcategorizable functions
(section 1), it follows that the subcategorizable functions and the governable functions
are one and the same. This is an extremely strong substantive constraint on government
which further research must test and may perhaps modify.
To summarize, we see that several major results follow from the theory of syntax
proposed here: first, that governing morphemes universally appear either in the heads
(or heads (of heads . . . )) or in minor categories of the phrases whose constituents they

a. c-structure 1
S

(tSUBJ) =T=1
NP VP1

Jan (tXCOMP)= lT
VP2 VI

ADV (IXCOMP)=1 1T=

VP3 V2

ADV (TOBLNAAR)=Jt T= I
PP V3

1=1 T=1
P NP

best ongezien naarl Marie glimlachen willen zou

'very' 'without *an 'smile' 'want' 'would'
being seen' (to)

Figure 8
Verb governing oblique object in Dutch

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 371

b. c-structure 2
S

(TSUBJ)=j T=
NP VP1

Jan (1XcOMP)=k T=I

VP2 V

ADV (1xcoMP)=I T= (XCOMP)=I

VP3 V1 V

ADV (TOBLNAAR)=4 = (XcOMP)=j

PP V2 V3

P NP

best ongezien naarl Marie zou willen glimlachen

'very' 'without *an 'would' 'want' 'smile'
being seen' 'at'

c. f-structure

SUBJ L"Jan'

PRED 'WOULD((SUBJ)(XCOMP) )'

XCOMP SUBJ

PRED WANT((SUBJ)(XCOMP) )'

XCOMP SUBJ

PRED SMILE((SUBJ)(OBL0))'

OBLNAAR ["at Marie"]

ADJ ["without being seen"]
ADJ ["very"]

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 372 JOAN BRESNAN

govern; second, that similar government relations are instantiated in configurationally

dissimilar structures; and third, that the types of structural configurations which in-
stantiate government relations in a particular language (type) are predictable from the
syntactic encoding of functions in that language (type).
This theory contrasts with what may be called configurational theories of govern-
ment, which stipulate a set of possible governing relations in terms of phrase structure
configurations. For example, it is assumed in the theory of government and binding
(Chomsky (1981)) that government is a structural relation between a major type 0 cat-
egory and an NP which is dominated by a projection of that category and that V and
INFL (the category of the tense and subject-agreement morpheme) are distinct major
type 0 categories having separate syntactic projections (VP and S, respectively). Verbs
govern their objects, but not their subjects, which lie outside of the VP (i.e. outside of
the projections of V). However, subjects are governed by the finite inflectional mor-
pheme in INFL (or by certain superordinate verbs and prepositions that are exceptionally
permitted to govern the subjects of a subordinate clause).
While it is easy to stipulate sets of possible government relations in terms of various
c-structure configurations, this does not solve the fundamental problem of government,
which is to explain the observed structural relations between governing morphemes and
the constituents they govern. Why, for example, are the case, number, and person of
the subject of a sentence recorded in the affix of its verb and not, say, in the affix of
a prepositional object of the verb? Why is it that verbs and complementizers govern the
finiteness of their clauses, while the determiners of direct objects do not? Why do the
same kinds of government relations appear in unrelated languages with radically different
constituent structures? Our theory provides principled answers to these questions.

8. Control

Control refers to a relation of referential dependence between an unexpressed subject

(the controlled element) and an expressed or unexpressed constituent (the controller);
the referential properties of the controlled element, including possibly the property of
having no reference at all, are determined by those of the controller. The term controlled
clause is also applied to the clause whose subject is controlled. To illustrate, two control
relations appear in examples (17a,b).

(17) a. At the moment, the goal of the police is to try to prevent a riot.
b. At the moment, the goal is to try to prevent a riot.

In (17a), the unexpressed subject of try is controlled by the police, and the unexpressed
subject of prevent is controlled by the (unexpressed) subject of try. In (17b), the unex-
pressed SUBJ of try lacks an antecedent, but this is often viewed as a degenerate control
relation, called arbitrary control.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 373

8.1. Theoretical Considerations

Theories differ according to what are considered the major generalizations to be der
and what are taken as basic assumptions. For example, Chomsky's (1981) theory of
control is designed to obtain the following generalizations as theorems.

(18) Only subjects are controlled.

(19) Only nonfinite clauses have controlled subjects.

To derive these results, it is assumed in Chomsky's theory that government is a structur

relation, as summarized in section 7. It is further stipulated that all NPs must be governed,
except for the controlled NP ('PRO'), which must not be governed. Verbs govern their
objects, but not their subjects, which lie outside of the VP. However, subjects are
governed by the finite inflectional morpheme in INFL (or by certain superordinate verbs
and prepositions that are exceptionally permitted to govern the subjects of a subordinate
clause). Therefore, the controlled NP 'PRO' can occur only in the ungoverned position
of subject of a nonfinite clause. The main idea of this theory is illustrated in figure 9.
Because phrase structure configurations are used to express government and control
relations, this can be referred to as a configurational theory of control.
It is evident that these assumptions are incompatible with the lexical-functional
theory of syntax. In this theory, government is a functional, not a phrase-structural,
relation (section 7); verbs must govern their subjects (section 7); bound inflectional
morphemes must be affixed prior to lexical insertion into phrasal structure (section 6);
and null structure appears only in cases of constituent control (section 6), that is, only
in cases of unbounded, structure-dependent syntactic binding relations (cf. Bresnan and
Grimshaw (1978), Kaplan and Bresnan (1982)).
We have already seen that these properties of the theory adopted here lead to
several explanatory results in the theories of subcategorization, inflection, and govern-
ment (section 7). The same results do not follow from the assumptions of the configu-
rational theory of control without further stipulations. Since in that theory only the
inflectional affix of the verb and not the verb itself governs the subject, and since
government is a structural relation that associates each governing category with a distinct
phrasal domain, it follows that the phrase structure representation in terms of which
government relations are stated will be distinct from the representation of the surface
constituency relations of the morphemes in a sentence. Therefore, a separate component
of "affix hopping" or "morphological spelling" or "realization" rules must be postulated
to map between these two phrase structure representations. As in figure 9, one can
stipulate that such rules must place the morpheme dominated by INFL on the morpheme
dominated by V, but any other stipulation is equally possible. Thus, in the configurational
theory of control, it is only an accident that the inflectional affixes that govern the
subject in figures 7a-c are verbal affixes.
Similar considerations show that the invariance of government relations across such

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 374 JOAN BRESNAN

a. S

NP INFL VP
[+Tns]
_+Agr]

*PRo V NP

gov't I gov't

affix hopping

b. S

NP INFL VP
[-Tns]

PRQ, V NV

gov't

Figure, 9
A configurational theory of control (Chomsky 1981)

configurationally different structures as those in figures 7 and 8 does not follow from
the configurational theory without a0ditional stipulations. Indeed, if government wer
a structural relation between phrasal nodes, and if such structural relations determined
the distribution of 'PRO', one vyoul4 predict that in 1a1nguages that lack VPs, subjects
could not be controlled (because ij such languages verbs would always govern their
subjects). However, research on Irish, a VSO language (McCloskey (1980; 1981)), and
Malayalam, a verb-final language haying "flat" clausal structure (Mohanan (1982a,b)),
falsifies this prediction. To maintain the configurational theory, one could draw the
procrustean conclusion that there is an underlying phrasal structure in terms of which
government is invariantly defined for all languages and that it resembles English surface
structure prior to affix hopping as in figure 9 (Chomsky (1981)). With this move, however,
the claim that government is a phrase-structural relation is emptied of its most interesting
and falsifiable content. The stipulation of a further level of abstract phrase structure

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 375

representation complicates the theory with the addition of yet another component of
realization rules to map the English-like abstract phrase structure onto the internally
motivated phrase structures of each language.
The contrasting properties of the lexical-functional theory of syntax have strong
independent justification, in addition to the above explanatory results. First, careful
research on the syntax of Irish (McCloskey (1980; 1981)) and Malayalam (Mohanan
(1982a,b)), taking into account the various options for configurational theories of control,
has yielded clear evidence that verbs govern their subjects; and in Icelandic (Andrews
(1982)) and in Russian (Neidle (1982)), controlled subjects show evidence of case marking
governed by the verb or by the nonfiniteness of the clause. Second, the research of
Selkirk (1981), Nash (1980), Lapointe (1980), Lieber (1980), and Mohanan (1982c) sup-
ports the more constrained lexicalist theory of inflectional morphology over the trans-
formationalist theory. Third, the constraints on null structure are supported by research
into null anaphora in Malayalam (Mohanan (1981b)), Navajo (Hale (1979)), and English
(Grimshaw (1979), Bresnan (1980), Levin (1982)).
It would therefore be a mistake to suppose that the configurational theory of control,
in deriving generalizations (18) and (19) from other assumptions, is more "explanatory"
than other theories. In fact, we have just seen that its assumptions lead to the loss of
major generalizations which follow instead from the theory described here.
An important further difference between the configurational theory of control and
the one presented here is the treatment of the controlled element. As a structuralist
theory, the configurational theory of control cannot make direct reference to grammatical
functions such as SUBJ. Instead, it takes the property of not being governed to be a
universal defining property of the controlled element. Thus, to derive generalization
(19), the configurational theory first stipulates that 'PRO' cannot be governed and then
that the NP immediately dominated by nonfinite S (i.e. the subject position of a nonfinite
clause) is an NP position that is not governed. The latter property is just as much a
stipulation as the former, because there is no reason in principle why the type 0 category
of the nonfinite S-the [ - agr] or [ - tns] INFL in figure 9b-should not be just as much
a governor as the [+ agr, + tns] INFL in figure 9a; but if it were, 'PRO' could never
appear in subject position. Similarly, there is no reason in principle why verbs are always
taken to govern their objects. The category V could bear a syntactic diacritic feature
exactly as the category INFL does, which would make it an optional governor of its
object; but if it were, 'PRO' could appear in nonsubject positions.
In contrast, the theory of control adopted here can make direct reference to gram-
matical functions. Instead of stipulating that not being governed is the defining property
of the controlled element 'PRO', our theory assumes that the property of being a SUBJ
is a defining property of the functionally controlled element. (The qualification "func-
tionally controlled" is explained below.) Similarly, instead of stipulating that the only
domain in which the controlled element can appear is nonfinite S, our theory assumes
that only the open functions XCOMP and XADJ can denote functionally controlled clauses.
We turn now to the precise expression of these assumptions within the formal system

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 376 JOAN BRESNAN

of grammatical representation. We will then see that a number of major generalizations

follow from the present theory of control relations.
Recall that control has been characterized as a relation of referential dependence
between an unexpressed subject and an expressed or unexpressed controller. The char-
acterization of control relations in terms of referential dependence captures their se-
mantic similarity, but masks an important underlying grammatical difference. Where the
referential dependence is accompanied by the complete identity of all functional features
of the controller and the controlled element, we have functional control (also called
grammatical control elsewhere in Bresnan (1982a)). Where the referential dependence
implies no identity of grammatical features, we have anaphoric control. That is, func-
tional control entails identity of f-structures of the controller and controlled element,
while anaphoric control entails mere "identity of reference" (i.e. only referential de-
pendence). The next two sections articulate the theory of functional and anaphoric
control. The reader is referred to figure 14 for a schematic summary of the theory.

8.2. Functional Control

In functional control relations, the controlled element is the SUBJ function and the
controlled clauses are designated by the open grammatical functions xcoMP and XADJ.
Thus, the term controlled clause refers to a clause at the level of f-structure: a clause
nucleus (section 5). The control relation is expressed by a control equation, a functional
schema which equates the f-structures of the controller and the controlled element.
Functional control relations are either lexically induced or constructionally induced,
depending on whether the control equation is part of a lexical entry or a c-structure rule
annotation. The range of possible controllers depends upon whether the functional con-
trol relation is lexically or constructionally induced.
In lexically induced functional control, the control equation is part of a lexical entry.
The control relation is defined in terms of functions that are subcategorized by the lexical
item that induces functional control. Thus, the controlled clause (that is, the function
whose subject is controlled) is the xcoMP (the predicative or open complement). The
controller is specified by a control equation of the form ( T G) = ( t XCOMP SUBJ
is added to the lexical entry of the item. From the severe constraints on the lexical
encoding of semantically restricted functions (stated in section 2), it follows that these
functions cannot be lexically induced functional controllers. Hence, only SUBJ, OBJ, and
OBJ2 are possible functional controllers, or values of G, in the above control equation.
The unmarked choice of controller is predictable from the following universal rule:
(20) Lexical Rule of Functional Control
Let L be a lexical form and FL its grammatical function assignment. If XCOMP
E FL, add to the lexical entry of L:
( OBJ2) = ( XCOMP SUBJ) if OBJ2 E FL;
otherwise:
( T OBJ) = ( T XCOMP SUBJ) if OBJ E FL;
otherwise:
( T SUBJ) = ( T XCOMP SUBJ)

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 377

That is, the XCOMP of a lexical form is functionally controlled by the OBJ2 if there is one,
otherwise by the OBJ if there is one, otherwise by the SUBJ.
This rule of unmarked lexical control is interpreted as a redundancy rule; that is,
the rule obligatorily expands an eligible lexical entry which lacks a control equation,
but it blocks if the otherwise eligible lexical entry already has a control equation. For
example, rule (20) specifies that the controllers of the predicative complements of appear
and regard in (21) and (22) are the SUBJ and OBJ, respectively, but the rule is blocked
by the lexically marked control equation for strike in (23).

(21) a. John seems sick to Mary.

b. ( PRED) = SEEM((XCOMP)(OBL0))(SUBJ)'
(22) a. John regards Mary as friendly.
b. ( T PRED) = 'REGARD((SUBJ)(OBJ)(XCOMP))'
(23) a. John strikes Mary as friendly.
b. ( PRED) = STRIKE((SUBJ)(OBJ)(XCOMP))'

( SUBJ) = (T XCOMP SUBJ)

An example of lexically induced functional control by OBJ2 is given in (24).

(24) Tom will serve you the fish raw.

The adjective in (24) is a state predicate, an optional predicative complement to a verb,

which describes a state of one of the verb's arguments at the time of the action denoted
by the verb. There are strong lexical restrictions on the introduction of a state predicate
as complement to a verb. For example, the predicate must describe an objective state
(cf. *Tom will serve you the fish tasty), the verb's argument must be a THEME (cf. *Tom
will give the fish a sauce raw), and the verb must denote an action rather than a state
(cf. *I envy Tom the fish raw). Within these lexical restrictions, any of SUBJ, OBJ, or OBJ2
are possible functional controllers:

(25) a. The package arrived unopened.

b. John will serve the fish to you raw.
c. I sent you the letter sealed.

Our theory predicts that only SUBJ, OBJ, and OBJ2 are possible controllers in cases
of lexically induced functional control; hence, the ungrammaticality of examples like
(26a,b)-noted by Williams (1980)-is explained.

(26) a. I presented it to John dead.

b. *1 presented John with it dead.

The verb present in (26) allows its THEME argument to be expressed either as OBJ or as
OBL0; however, as predicted by our theory, only the object can control the state com-
plement. These state complements are a special type of xcOMP; they differ from XADJS
to be described directly.
In constructionally induced functional control, the control equation is part of a
c-structure rule annotation. The controlled clause is the XADJ (the predicative or open

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 378 JOAN BRESNAN

adjunct), and the controller is specified by a functional schema of the form ( T G)

= ( 4 SUBJ), which is added to the functional annotations of the adjunct. Because the
control equation is syntactically, rather than lexically, specified, it is not constrained
by the restrictions on lexical encoding of functions; this means that a wider range of
controllers is available in principle. The set of possible controller functions F appears
to be a parameter of variation across languages. In Malayalam, F = {SUBJ} (Mohanan
(1981b)), while in Russian (Neidle (1982)) and in English, f = {SUBJ, OBJ, OBJ2, OBLe}.
The rule of constructional control is given in (27).

(27) Constructional Rule of Functional Control

If ( T XADJ) = I is a syntactically encoded functional annotation, conjo

it the disjunction of the schemata {( T G) = ( 1 SUBJ) I G E F}.
Rule (27) is interpreted as a syntactic redundancy rule: if a predicative adjunct lacks a
controller, rule (27) obligatorily specifies the set of possible controllers; but if an adjunct
construction is marked for a particular controller, the rule's application is blocked.
To illustrate the effect of rule (27), let us consider examples of predicative adjuncts
in English. (This discussion disregards other kinds of adjuncts, such as relative or
appositional adjuncts.) In (28), either the SUBJ Mary or the OBJ John can be the controller
of drunk as usual, although there may be a slight preference to interpret Mary as con-
troller:

(28) Mary passed John in the hall yesterday drunk as usual.

The following example brings out the interpretation in which the OBJ John is the con-
troller, because we assume that Mary would probably not describe herself as "drunk
as usual":

(29) Mary said she passed John in the hall yesterday drunk as usual.

Note that in contrast to the cases of lexically controlled state complements, the control
of these predicative adjuncts is not restricted to one thematic argument of the verb.
Moreover, in (30) we see that an OHL0 can control the adjunct, for it is possible for drunk
as usual to be predicated of the OI3LAG Mary as well as the SUBJ John:

(30) John was passed by Mary in the hall yesterday drunk as usual.

Again, a reportive example such as (31) brings out the interpretation in which the subject
is not the controller, for once again we assume that John would probably not describe
himself as "drunk as usual":

(31) John said he was passed by Mary in the hall yesterday drunk as usual.

Furthermore, these adjuncts modify a much wider range of lexical predicates than the
complements do and can describe nonobjective states. For example, in contrast to *John
will serve you the fish tasty, we find Mary caught a glimpse of the fish lying on the
table, tasty and fragrant with herbs and even John will serve you the fish, tasty and

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 379

fragrant with herbs. The distinctio

previously discussed corresponds to Halliday's (1967) distinction between "conditional
attributes" and "depictive attributes". Halliday observes that multiple conditional at-
tributes can occur in a simple sentence, while only one depictive attribute may. This is
exactly what the analysis in terms of XCOMPS and XADJS predicts (Bresnan (1980), Kaplan
and Bresnan (1982)).
While functionally controlled XADJS generally have a range of possible controllers,
including OBL0 as we have just seen, there is one construction in English in which
functional control of the XADJ is restricted. This is the clause-initial position of the
adjectival phrase adjuncts shown in (32) and (33).

(32) a. Sure of winning, Mary entered the competition yesterday.

b. #Sure of winning, the competition was entered by Mary yesterday.
(33) a. #Sure of winning, the competition excited Mary yesterday.
b. Sure of winning, Mary was excited by the competition yesterday.

In all of (32a,b) and (33a,b), sure of winning is controlled by the SUBI of the sentence;
where the SUBJ denotes an inanimate entity, the result is semantically anomalous (sig-
naled by #). Thus, we will assume that suBJ-control is a marked property of the clause-
initial XADJ construction shown in (34). ((34) is one instance of a more general rule
introducing S-initial adjuncts of various categories.)

(34) S f AP NP VP
| ( T XADJ)- | T SUBJ) =
\T(1 SUBJ) = ( I SUBJ)JJ

Since the functional controller of the adjunct has been marked in the clause-initial
construction given in (34), the rule of constructionally induced functional control (27)
is blocked from applying there.
Because functional control relations have been formalized by equations identifying
the controlled SUBJ with the controller, the f-structures of controller and controlled will
be "merged" into one and the same f-structure; as a result, the controlled SUBJ will have
all of the grammatical features (CASE, PERSON, NUMBER, etc.) of the controller. The
referential dependence of the controlled SUBJ upon the controller follows from the se-
mantic interpretation of the functional control relation by "quantifying in" (Halvorsen
(1981)).

8.3. Anaphoric Control

Anaphoric control relations arise from the presence of a functional anaphor ('PRO') which
is not expressed in c-structure. The functional anaphor is created by an optional func-
tional schema of the form ( G PRED) = 'PRO' for any function G. The possible occur-
rences of these schemata are quite limited by independent principles. 'PRO' is a semantic
form, and all semantic forms are introduced in the lexicon (Halvorsen (1981)). Hence,
the optional schema (( l G PRED) = 'PRO') must belong to some lexical entry. That cannot

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 380 JOAN BRESNAN

be the entry for a null NP, for our theory does not permit the use of null c-structure to
represent local grammatical relations (section 6). Therefore, since (( T G PRED) = 'PRO')
lacks its own lexical entry as a null category, it must be introduced as part of the lexical
entry of a lexical form that governs G (section 1) and G must be a subcategorizable
function. Moreover, for any lexical entry L to which an equation (( T G PRED) - 'PRO
is added, the completeness and coherence conditions restrict G to the particular functions
subcategorized by the lexical form of L. The constraints on lexical encoding of functions
(section 2) further restrict G to be one of the set of semantically unrestricted functions
{SUBJ, OBJ, OBJ2}. Finally, a language-particular parameter may restrict G to a subset A
of the set of semantically unrestricted functions. We will further assume that only
[? FIN] lexical items permit the functional anaphor, and that the value of the feature
[SlFIN] is a language-particular parameter. These properties are incorporated in rule (35).

(35) Rule of Functional Anaphora

For all lexical entries L, for all G E A, assign the optional pair of equations
{(( T G PRED) = 'PRO'), (T FIN) =c a} to L.

By fixing the parameters (x = - and A = {SUBJ}, we derive the rule of functional

anaphora for English. Thus, the functional anaphor 'PRO' arises in English only as the
subject of a nonfinite verb (infinitive or gerund).
An anaphor is a grammatical element which may be assigned an antecedent by the
rules of sentence grammar. Semantically, anaphors are either coreferential with or ref-
erentially dependent upon their antecedents. 'PRO' iS an anaphor similar (but not
identical) in its interpretation to the definite pronouns her, he, they, etc. The so-called
control relation between 'PRO' and its controller is actually an anaphoric relation between
'PRO' and its antecedent; we designate this relation anaphoric control to distinguish it
from functional control. Note that the term anaphor as used here applies both to those
pronouns that are obligatorily assigned antecedents within the sentence, such as the
reflexive pronoun herself, and to those pronouns that are only optionally assigned an-
tecedents within the sentence, such as the definite pronoun her. The former can be
distinguished as bound anaphors.
In contrast, in the government and binding theory, anaphors are distinguished from
pronominals in that anaphors "have no capacity for 'inherent reference"' (Chomsky
(1981, 188)). The semantic content of this notion of "inherent reference" is unclear, but
what appears to be meant by it is that anaphors lack the independent capacity to refer
to specific extrasentential referents. For example, in We thought Mary should have the
operation. *However, herself is the only one that can make that decision, the anaphor
herself cannot refer by itself to the extrasentential referent Mary; but in We thought
Mary should have the operation. However, she herself is the only one that can make
that decision, the pronominal she can refer to Mary. Anaphors and pronominals are
distinguished by the "binding conditions" of the government and binding theory, which
assert (A) that an anaphor must be bound within its minimal governing category (NP
or S), and (B) that a pronominal must be free within its minimal governing category

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 381

(Chomsky (1981, 188ff)). Thus, herself in

her in Mary voted for her is a pronominal. As Chomsky (1981, 191) writes, 'PRO' iS
considered to be both a pronominal and an anaphor:

Consider next pronominals without a phonetic matrix, i.e., PRO. Note that PRO is like overt
pronouns in that it never has an antecedent within its clause or NP. PRO also resembles
anaphors in that it has no intrinsic referential content but is either assigned reference by an
antecedent or is indefinite in interpretation, lacking specific reference. It is reasonable, then,
to regard PRO as a pronominal anaphor. If so, it is subject to both the binding conditions
(A) and (B). Then PRO is bound and free in its governing category, a contradiction if PRO
has a governing category. Therefore PRO has no governing category and is therefore un-
governed. We therefore derive the principle (20) [that PRO is ungoverned], which . . . is the
essential property of PRO.

Thus, the motivation for the claim that 'PRO' is an anaphor is that when it does not
derive its reference from an antecedent within the sentence, it lacks definite reference,
as in It is unclear what to do. (This is the so-called arbitrary control.) The claim that
'PRO' is an anaphor (as well as a pronominal) in turn motivates the crucial assumption
that it is ungoverned, from which follow its most essential properties for the government
and binding theory. In particular, it follows that 'PRO' can be the subject of a nonfinite
clause (in which INFL does not govern the subject), but not an object or a subject of
a finite clause (in which INFL does govern the subject).
However, the assumption that 'PRO' is an anaphor in the sense of the government
and binding theory presents a serious problem: unlike anaphors, but like pronominals,
'PRO' does have the capacity to refer independently to specific extrasentential referents.
In English, many examples arise in the context of indirect discourse, as (36) illustrates.

(36) a. Mary was happy and excited. To have involved herself in the group was
a risky action. But it was proving that she could change her life.
b. Tom felt sheepish. Pinching those elephants was foolish. He shouldn't
have done it.
c. She sighed and looked around the empty room. It was unclear what to do
with herself now that Molly was gone.

In these examples, the PRO subjects of the infinitives are respectively understood as
referring to Mary, Tom, and the referent of she. Examples of definite reference of 'PRO'
also occur in other discourse contexts such as (37).

(37) Frankly, I'm worried about Mary. What has she gotten herself into? Don't
get me wrong: I think it was fine to join the group. But getting herself pho-
tographed with those starving wolves was dangerous.

Here, too, the PRO subjects of the infinitive and gerund refer specifically to Mary. Like
expressed pronouns, 'PRO' can also be understood as referring to a specific referent
which is presented in the nonverbal context of an utterance. Consider a situation in
which two observers have witnessed a young man commit suicide by leaping from a

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 382 JOAN BRESNAN

fortieth-floor window. One obser

was a terrible mistake", speaking specifically of the young man and meaning by this
that his killing himself was a terrible mistake.
In general, the reference of reflexive pronouns is determined by their antecedents.
In the above examples, the reflexive pronouns refer to specific individuals, and their
antecedents are 'PRO'. Furthermore, in these examples, 'PRO' does not derive its ref-
erence from an antecedent within the sentence. Thus, we are led to the conclusion that
'PRO' does have the capacity to refer independently to specific extrasentential referents.
(This argument is due to P.-K. Halvorsen.) Thompson (1973) argues convincingly that
the interpretation of 'PRO' as definite or generic is predictable from temporal or aspectual
and contextual properties of the sentence in which it occurs; in contexts of specific
temporal reference such as (36) and (37), 'PRO' may have a definite interpretation, but
in contexts of nonspecific temporal reference such as (38), it has the generic interpre-
tation.

(38) a. To involve oneself in that kind of group is risky.

b. Pinching elephants is foolish.

The fact that 'PRO' can have definite reference in the previous examples shows that
it cannot be an anaphor in the sense of the government and binding theory. However,
if 'PRO' is not an anaphor in this sense, then there is no motivation for the stipulation
that it must be ungoverned; 'PRO' should occur in ungoverned positions, such as object
position, like other nonanaphors in English. Even if we assumed that 'PRO' iS ambiguous,
being either [ + anaphoric, + pronominal] or [ - anaphoric, + pronominal], the govern-
ment and binding theory would be unable to account for the fact that 'PRO' has the same
syntactic distribution in English (namely, subject of nonfinite clause) when it is an
anaphor as when it is a nonanaphor. Thus, "the essential property" of 'PRO' for the
government and binding theory-its being ungoverned-remains unexplained.7
Note that the government and binding theory cannot adopt our definition of anaphors
as grammatical elements that may be assigned antecedents by the rules of sentence
grammar. That definition would admit pronouns as anaphors, because pronouns can,
and sometimes must, be assigned sentential antecedents (as in Louise craned her neck);
by Chomsky's reasoning given above, it would then follow (contrary to fact) that pro-
nouns could occur only in ungoverned positions. The alternative of defining anaphors
as grammatical elements that must be assigned sentential antecedents would also fail,
because 'PRO' need not have such an antecedent; for example, it lacks an antecedent in
It is unclear what to do and I think that killing himself was a terrible mistake.
In the theory proposed here, 'PRO' is not a bound anaphor, but a pronominal element.
It must be distinguished from the expressed definite pronouns, however, because it has
special restrictions on its anaphoric relations (to be discussed below). Let us therefore

7The implications of this analysis of 'PRO' for the government and binding theory were pointed out by
K. P. Mohanan (personal communication).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 383

assume that there is some feature-call it

separates 'PRO' from other pronouns. Th
therefore resemble (39a,b).

(39) a. LPRED PRO

_U + J
b. PRED PRO'

GEND FEM

NUM SG

PERS 3

CASE NOM J

Given that 'PRO' is an anaphor which need not be bound, how c

contrast between (40a) and (40b)?

(40) a. Mary wished to vote.

b. Mary wished for her to vote.

The government and binding theory attributes the contrast to the analysis of 'PRO' a
a bound anaphor and her as a pronominal. Assuming that the complementizer for ex-
ceptionally governs her in (40b), the minimal governing category of the pronominal he
is extended to include both the matrix clause and the complement clause. The minimal
governing category now includes Mary; thus, her must be free in this domain and disjoin
from Mary. In contrast, the 'PRO', as a (bound) anaphor, must be bound within the
sentence or assigned the arb (indefinite) interpretation. Presumably, lexical properties
of wish force the bound interpretation.
The present theory provides a natural alternative to the above analysis. Note that
in these examples, the [ + u] anaphor ('PRO') is bound to the SUBJ Mary, while the [ - u
anaphor (her) cannot be bound to the SUBJ. This is reminiscent of the phenomenon
known in many languages as obviation, by which certain pronouns exclude coreferenc
with specified types of antecedents (Hale (1978)). We therefore formulate the Obviation
Principle given in (41). This formulation should not be regarded as a maximally genera
statement of obviation across languages, but rather as an alternative hypothesis to th
claim of the government and binding theory (Chomsky (1981)) that (40b) is an instanc
of disjoint reference (i.e. of a pronoun being free within its minimal governing category
In fact, (41) can be derived from a general theory of pronominal obviation, but this topi
is beyond the scope of our present concerns. See Simpson and Bresnan (in preparation)
and the references cited there.

(41) Obviation Principle

If P is the pronominal SUBJ of an obviative clause C, and A is a potent

antecedent of P and is the SUBJ of the minimal clause nucleus that proper
contains C, P is or is not bound to A according to whether P is + or - u
respectively.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 384 JOAN BRESNAN

In English, an obviative clause is any clause that can be marked by the complementizer
for.8 Since the infinitival complement to wish can be marked by the complementizerfor,
it is an obviative clause, and given our analysis of pronominals, principle (41) immediately
explains the contrast between (40a) and (40b).
Our hypothesis that a type of pronominal obviation is involved in examples like
(40a,b) makes a strikingly different prediction from the government and binding theory
of pronominals and anaphors. In our theory, obviation is a functional relation which
holds between matrix subjects and complement subject pronominals. In the government
and binding theory, by contrast, the property of being free within a minimal governing
category is a structural relation which prohibits all NPs within the matrix clause-
subject and nonsubject NPs alike-from binding the pronominal subject of the for-to
complement. The government and binding theory therefore predicts no difference be-
tween matrix subject and nonsubject antecedents of a complement subject pronominal;
the complement subject should be disjoint from both. The Obviation Principle in our
theory predicts that disjoint reference applies only with subject and not with nonsubject
antecedents in the matrix. The evidence given below crucially chooses between the two
theories. (Examples (42d) and (43c) contain nonobviative clauses; they are included as
minimal contrasts.)

(42) a. For her to lose after all that effort would really surprise Louise.
[her = Louise, possibly]
b. It would really surprise Louise for her to lose after all that effort.
[her = Louise, possibly]
c. Louise would really be surprised for her to lose after all that effort.
[her $A Louise]
d. Louise would really be surprised if she lost after all that effort.
[she = Louise, possibly]
(43) a. Louise signaled to Ted for him to follow her.
[him = Ted and her = Louise, possibly]
b. Louise signaled to Ted for her to follow him.
[her $ Louise and him = Ted, possibly]
c. Louise signaled to Ted that she would follow him.
[she = Louise and him = Ted, possibly]

8 We conjecture that the subjunctive complementizer in Romance languages is also an obviative

marker. In other words, we propose that examples like Marie veut qu'elle parte, wherein elle must be disjoint
from Marie, are instances of subject-subject obviation parallel to Mary would like for her to leave in English.
Since the subjunctive clause of French is finite, the [ + u] 'PRO' does not arise in it, but the [ - u] pronouns are
nevertheless subject to the principle. Simpson and Bresnan (in preparation) show that complementizers can
be obviative markers in Warlpiri.
It is possible that the Obviation Principle (41) should be reformulated to apply only to [- u] pronouns,
for the obligatory assignment of an antecedent to the [ + u] 'PRO' in examples like (40a) could be attributed to
lexical properties of the matrix verb, which could induce "thematic binding" of [+ u] 'PRO' (Kisala (in prep-
aration)). Until further research decides this question, the present formulation of the Obviation Principle will
be tentatively maintained.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 385

Examples (42) and (43) show that, contrary to the predictions of the government and
binding theory, the complement subject pronoun is disjoint only from the matrix subject,
and not from nonsubject arguments of the matrix. (It might be supposed that the extra-
posed clause in (42b) lies outside the minimal governing category of Louise, as would
be the case if the extraposed clause were dominated by S rather than by VP. However,
there is evidence that the extraposed clause is dominated by VP. Note that the extraposed
clause in It would really surprise her for Louise to lose after all that effort must be
c-commanded9 by the direct object in order to account for the noncoreference of her an
Louise (assuming Reinhart's (1976) formulation of the noncoreference condition); but
this example is exactly parallel in structure to (42b). Hence, the extraposed clause in
(42b) does lie within the same minimal governing category as the matrix object.)
The present theory also predicts the contrast in the following examples.

(44) a. To be able to shave oneself at five years of age would really surprise one's
father.
b. It would really surprise one's father to be able to shave oneself at five
years of age.
c. One's father would really be surprised to be able to shave *oneself/himself
at five years of age.

In examples (44a,b), the matrix subject is not a possible antecedent for the complement
'PRO', so binding is not obligatory by the Obviation Principle. By contrast, the matrix
subject in (44c) is a possible antecedent, and the binding of 'PRO' is obligatory, as is
shown by the fact that the reflexive pronoun must agree with one's father.'0
It must be noted that the Obviation Principle (41) does not provide the only con-
ditions under which anaphoric control of 'PRO' by a matrix argument may be obligatory.
It has frequently been observed that the semantic or thematic structure of the matrix
predicate can induce control of a complement 'PRO'. This appears to account for the
examples in (45), which indicate that 'PRO' is obligatorily bound to the GOAL argument
of signal.

(45) a. Louise signaled to Ted to shave himself.

b. *Louise signaled to Ted to shave oneself.
c. *Louise signaled to Ted to shave herself.

The Obviation Principle is inapplicable to (45) even though Louise is a subject, because
Louise is thematically excluded as a possible antecedent A of the complement 'PRO'.
'PRO' is further distinguished from expressed definite pronouns by a universal con-
dition on anaphoric control which is due to K. P. Mohanan (1981b). We formulate this
condition as (46).

9 A c-structure node N1 c-commands a c-structure node N2 if N1 does not dominate N2 and every node
that dominates N, dominates N2.
10 If, however, we adopt the alternative formulation of the Obviation Principle suggested in footnote 8,
the contrast between (42a,b) and (42c) would have to be attributed to lexical differences between the verb
surprise and the deverbal adjective surprised.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
386 JOAN BRESNAN

(46) Universal Condition on Anaphoric Control

If A is a grammatically assigned antecedent of P, where the value of P is [PRED

'PRO', u + ], then A must f-command P.

F-command is a relation on f-structures defined as follows:

(47) F-command

For any occurrences of the functions a, P in an f-structure F, a f-co

1B if and only if a does not contain P and every f-structure of F tha
ax contains 1P.
To illustrate the effect of the Universal Condition on Anaphoric Control, figure 10 shows
the f-structure for the sentence People who know John often discuss working too hard.
We see in this figure that the SUBJ atI f-commands the SUBJ 1, while OBJ a2 does not
f-command SUBJ . Hence, it follows from the Universal Condition on Anaphoric Control
that (a 1, P) but not (aX2, 1) is a possible control relation. This explains why the PRO subject
of working in (48a) can be interpreted as people but not as John; (48b) shows that the
universal condition applies only to the unexpressed 'PRO' and not to definite personal
pronouns.

(48) a. People who know John often discuss working too hard.
b. People who know John often discuss his working too hard.

Similarly, the Universal Condition on Anaphoric Control explains why, in (49a),

Mr. Jones does not anaphorically control the PRO subject of contradicting, even though
Mr. Jones can be a grammatically assigned antecedent of the pronoun his in (49b).

(49) a. *Contradicting himself will demonstrate that Mr. Jones is a liar.

b. His contradicting himself will demonstrate that Mr. Jones is a liar.

Example (49a) is ill-formed when the PRO subject of contradicting is taken to be Mr.
Jones. The reason is that in the f-structure for (49a), shown in figure 11, the SUBJ Mr.
Jones does not f-command 'PRO'. In contrast, Mr. Jones can anaphorically control 'PRO'
in (50).

(50) Contradicting himself will discredit Mr. Jones.

In (50), the object of the verb discredit does f-command the 'PRO' complement subject;
see figure 12.
These examples also show very clearly that f-command, and not c-command, is the
determining factor in the universal condition (46) on anaphoric control relations, for in
neither of the examples Contradicting himself will discredit Mr. Jones and *Contradicting
himself will demonstrate that Mr. Jones is a liar does Mr. Jones c-command the com-
plement subject position (see figures 11 and 12). Similarly, in Improving himself seems
important to John, the OBLGo John f-commands, but fails to c-command, the SUBJ of
improving himself; see figure 13.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 387

c-structure

NP VP

NP S V NP

NP S VP

NP VP V QP

V NP

people who e know John discuss working too hard

f-structure

SUBJa, PRED PEOPLE'

ADJUNCT TOP PRED PRO

(relative)|

SUBJ

PRED KNOW((SUBJ)(OBJ))'

OBJoV2 ,[PRED JOHN']

PRED DISCUSS((SUBJ)(OBJ) '

OBJ SUBJg PRED PRO J

Lu +
PRED WORK((SUBJ))'

ADJUNCT L"too hard"

SUBJai f-commands SUBJp

OBJa2 does not f-command SUBJ8

Figure 10
The F-Command Condition on anaphoric control

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
388 JOAN BRESNAN

c-structure

NP VP

V VP

VP V S

V NP COMP S

NP VP

V NP

contradicting himself will demonstrate that Mr. Jones is a liar

f-structure

SUBJ SUBJJ PRED PRO'J

PRED CONTRADICT ((SUBJ)(OBJ))'

OBJ gPRED PRO'

_ L~~(reflexive) _
PRED WILL((XCOMP))(SUBJ)'

XCOMP SUBJ

PRED DEMONSTRATE ((SUBJ)(COMP))'

COMP SUBJa EPRED MR. JONES']

PRED BE ((XCOMP))'

XCOMP SUBJ

_ _ _ ~~~~~~PRED LIAR((SUBJ)) _

SUBJa does not f-command SUBJfl

Figure 11
F-Command Condition on anaphoric control

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 389

c-structure

NPa VP

V VP

VP V NPa

V NP

contradicting himself will discredit Mr. Jones

NPa does not c-command NPp

f-structure

SUBJ SUBJ LPED PRO J

PRED CONTRADICT((SUBJ)(OBJ) )'

OBJ PRED PRO

OBJ (reflexive) R j
PRED 'WILL((XCOMP))(SUBJ)/

XCOMP SUBJ

PRED DISCREDIT ((SUBJ)(OBJ))'

L OBJa EPRED MR. JONES']

OBJa f-commands SUBJp

Figure 12
F-command vs. c-command

Because 'PRO' can refer to discourse antecedents, as we have seen, the possibility
arises that both 'PRO' and another referential phrase in the sentence may refer to the
same antecedent in discourse. When this happens, 'PRO' is not anaphorically controlled
by the other phrase (because the other phrase has not been grammatically assigned as
the antecedent to 'PRO'). In just these cases, apparent counterexamples to the F-Com-
mand Condition may appear. As an initial example of this situation, consider (51).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 390 JOAN BRESNAN

c-structure

NP,p VP

VP V AP

V NP A PP

P NPa

improving himself seems important to John

NPa does not c-command NP,

f-structure

SUBJ SUBJ LPRED PRO j

PRED IMPROVE ((SUBJ)(OBJ))'

OBJ rPRED PRO

O(reflexive) j j

PRED SEEM ((XCOMP) )(SUBJ)/

XCOMP SUBJ

PRED IMPORTANT-TO ((SUBJ)(O

OBLGOa PRED JOHN_

OBLGoa f-commands SUBJp

Figure 13
F-command vs. c-command

(51) Contradicting himself will demonstrate that he is a liar.

Example (51) is parallel in structure to (49a), yet he and 'PRO' can corefer, where Mr.
Jones and 'PRO' could not. The reason is that he, more easily than Mr. Jones, can be
taken as referring to an implied discourse antecedent; he could even have 'PRO' as its
antecedent. However, there are certain conditions under which a nonpronominal noun

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 391

phrase like Mr. Jones could be coreferen

noun phrases can refer to discourse antecedents much more easily than can stressed
noun phrases; unstressed phrases are usually taken as designating old information, and
stressed phrases, new information. The contrast between (52a) and (52b) illustrates this
phenomenon.

(52) a. John was unhappy. ??The Nobel Prize had prevented the discovery of a
vaccine by John.
b. John was unhappy. The Nobel Prize had prevented John's discovery of
a vaccine.

In these examples, we see that the possessive John's in (52b) can refer to a discourse
antecedent much more easily than can the prepositional object by John in (52a). The
prepositional object, unlike the genitive prenominal NP, is in a stressed position. On the
basis of this difference, we therefore expect the contrast in (53).

(53) a. Winning the Nobel Prize had prevented John's discovery of a vaccine.
b. ??Winning the Nobel Prize had prevented the discovery of a vaccine by
John.

In (53a), but much less so in (53b), John and the PRO subject of winning may be co-
referential with an implied discourse antecedent. Hence, in (53b) 'PRO' can easily refer to
John only if 'PRO' is assigned John as an antecedent within the sentence; but this is ruled
out by the Universal Condition on Anaphoric Control. John is properly contained within
the object of the main clause, and so clearly fails to f-command the complement subject.
Therefore, in (53b) it is very difficult to interpret the PRO subject of winning as John.
We see, then, that the analysis of 'PRO' as an unexpressed pronoun which is subject
to the Obviation Principle (41) and the Universal Condition on Anaphoric Control (46)
better expresses the generalizations that govern its behavior than the theory that it is
an ungoverned pronominal anaphor. It is of interest that nothing in our theory of ana-
phoric control predicts that 'PRO' cannot occur in so-called governed positions, such as
subject of a finite sentence or object or second object of a verb. In this respect, the
theory presented here differs crucially from Chomsky's (1981) government and binding
theory, which was designed to derive that generalization as a theorem. Our theory
predicts that 'PRO' may have any of the semantically unrestricted functions SUBJ, OBJ,
OBJ2. It would therefore be of crucial theoretical interest to know whether there are
languages in which this occurs. Let us now suppose that such a language is found. Note
that a possible response for the government and binding theory would be that the ele-
ments in question are not true instances of 'PRO'. Of course, in bringing evidence to bear
on the choice between alternative theories of control of 'PRO', one must have clear criteria
for identifying an instance of 'PRO' other than its conformity to one of the theories in
question. What, then, are the independent criteria for identifying 'PRO'? Taking into
account the above discussion, we can propose the following properties as collectively
criterial for 'PRO':

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
392 JOAN BRESNAN

(a) 'PRO' is an unexpressed pronoun.

(b) 'PRO' may have either definite or indefinite reference, depending upon the
context.
(C) 'PRO' is not assigned an antecedent within its minimal clause.
(d) 'PRO' may be obligatorily bound to certain thematically or grammatically spec-
ified antecedents.
(e) An antecedent grammatically assigned to 'PRO' must be superordinate to 'PRO'
in the clause structure representing grammatical relations (in our theory, the
antecedent must f-command 'PRO').
(f) 'PRO' has a restricted set of grammatical functions (either {SUBJ} alone as in
the government and binding theory, or a subset of {SUBJ, OBJ, OBJ2} as in the
lexical-functional theory).

It might be questioned whether (e) should be admitted as a theory-independent property

of 'PRO', since there is no alternative to the F-Command Condition. (Although it has
been suggested in passing by Chomsky (1981) that the controller must be an argument
of the matrix of the controlled complement, this requirement cannot replace the
F-Command Condition (cf. footnote 16 and the discussion of examples (141a-c)).) In fact,
a condition similar to the F-Command Condition was embodied in the earliest generative
analyses of PRO phenomena, which hypothesized a cyclic Equi NP Deletion transfor-
mation. The condition of strict cyclicity limited the application of cyclic rules so that
no rule could analyze material solely contained within previous cycles, a constraint that
had essentially the same effect as (e).
Malayalam provides crucial evidence regarding the existence of governed 'PRO'S as
defined by the above criteria (Mohanan (1981b)). First, Malayalam has unexpressed
pronouns. Second, these pronouns may have either generic or definite reference, de-
pending on context. Third, they cannot have antecedents within their minimal clauses.
Fourth, they may be obligatorily bound to certain functionally or thematically specified
antecedents. An example given by Mohanan (personal communication) is (54).

(54) ellaawarkkum [faawile kufik'k'unnat;] istamaanm

all-dat morning bathe-pres-it liking-be-pres
'Everyone likes bathing in the morning.'

In example (54), the PRO subject of 'bathing' is obligatorily bound to the matrix subject
'everyone'; in contrast, the expressed pronoun subject in (55) is not obligatorily bound
to 'everyone'.

(55) ellaawarkkum [awar raawile kulik'k'unnata] istamaan;)

all-dat they-nom morning bathe-pres-it liking-be-pres
'Everyone likes their bathing in the morning.'

Fifth, the Malayalam unexpressed pronouns, unlike the expressed pronouns, must be
f-commanded by their grammatically assigned antecedents within the sentence. Example
(56), from Mohanan (1981b), illustrates this point.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 393

(56) [[oon meefiye ummawecca] kaafyam] awal/+ afootum paraninilla

John-nom Mary-acc kiss-past thing she-nom/PRo anyone say-past
'Shei/*PRoi did not tell anyone that John had kissed Maryi.'

Sixth, and crucially, these unexpressed pronouns can be SUBJ, OBJ, and OBJ2, but not
OBLH or ADJ. Examples of a PRO OBJ having definite and indefinite interpretations are
given in (57) and (58) (from K. P. Mohanan (personal communication)).

(57) wakkiilanmaar caticcu

lawyer-nom-pl cheat-past
'The lawyers cheated [definite pronoun: him, her, them, me, etc.].'
(58) wakkiilanmaar catik'k'um
lawyer-nom-pl cheat-fut
'Lawyers will cheat one.'

An example illustrating the F-Command Condition on an OBJ 'PRO' is given in (59) (from
K. P. Mohanan (personal communication)).

(59) [joon nulliyappool] meeriyute acchan kafannu

[John-nom pinch-past then] Mary-poss father cry-past
'Maryi's fatherj cried when John pinched him-/*her,.'

For further discussion of the F-Command Condition and evidence that the C-Command
Condition does not hold, see Mohanan (1981b). Finally, an example illustrating that
'PRO' cannot be OBLQ in Malayalam is given in (60) (from Mohanan (1981b)).

(60) raajaawa waal etuttu. addeeham mantriye wadhiccu

king-nom sword-nom take-past he-nom minister-acc kill-past
'The king took the sword. He killed the minister (*with it).'

As indicated in the gloss of (60), it is not possible to interpret the second sentence as
having a definite 'PRO' as instrumental. (Further examples are given in Mohanan (198 lb).)
In Malayalam, the criterial properties of 'PRO' are shared by the unexpressed subjects,
objects, and second objects of finite clauses (Mohanan (1981b and personal communi-
cation)).
Thus, the evidence from Malayalam crucially favors the present theory of anaphoric
control over the government and binding theory of 'PRO' as an ungoverned pronominal
anaphor. For Malayalam, the parameters cx and A of the universal Rule of Functional
Anaphora (35) are fixed as co = + and A = {SUBJ, OBJ, OBJ2}.
Our theory of functional and anaphoric control is summarized in figure 14.

9. Consequences of the Theory of Control

9.1. Interpretation, Function, and Agreement

From this theory of functional and anaphoric control, there follow a number of interest
generalizations. First, from the conditions on functional control (section 8.2), it follows
immediately that anaphoric and arbitrary control of subject arise only in closed functions

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 control

identity of all "identity o

grammatical reference'*
features

functional anaphor

lexically constructionally
induced induced

domain (= controlled lj l
clause): XCOMP XADJ closed funct
controller: SUBJ, OBJ, OBJ2 r* closed functions f-commandin
controlled element: SUBJ SUBJ A* C {SUBJ, OB
representation: lexical equations: c-structure rule annotations: lexical e
(tG)=(TXCOMP SUBJ) for (tG)=(JSUBJ) for G a (tG PRE
G a controller controller (tG U)= + f
con

rules: Lexical Rule of Constructional Rule of Rule of Fu

Functional Control (20) Functional Control (27) Anaph
Prin

*language-particular parameter

Figure 14
The theory of control

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 395

(and, depending on the finiteness paramet

antecedentless interpretation of the contr
COMP clauses (or in main clauses, as in Malayalam); it never occurs in the predicative
complement (xcoMP) or open adjunct (XADJ).
An alternative hypothesis concerning the arb (arbitrary) interpretation of 'PRO' iS
Chomsky's proposal that in the structure ... V ... [s COMP PRO ... ] . . ., where
V strictly governs S, (i) if COMP (lementizer) is not null and V has no controller, the
PRO is assigned the index arb; otherwise (ii) PRO is assigned the index i of a controller
(Chomsky (1980b)). In other words, arbitrary control arises only in complement clauses
with nonnull complementizers, such as It is unclear what to do, It is difficult to leave.
(It is assumed in this theory that a nonnull complementizer for occurs in the latter
example; this complementizer, through its exceptional government of the subject, is
what is supposed to sanction the presence of a lexical NP subject of the infinitive in
examples like It is difficultfor us to leave (cf. section 7)).
The fact that arbitrary control does arise in complement clauses with nonnull com-
plementizers follows as a special case of our theory, for by the principles of syntactic
encoding (section 4) S-complement clauses can only be assigned closed functions such
as SUBJ, OBJ, OBL0, or COMP. In contrast to Chomsky's proposal, however, our theory
predicts that arbitrary control can arise in clauses that lack nonnull complementizers,
so long as these clauses have closed functions. There is evidence from Italian that
confirms our theory. Manzini (1980) has observed that arbitrary control occurs in infin-
itival clauses in Italian, despite the fact that these infinitival clauses lack nonnull com-
plementizers. An example is given in (61).

(61) E' difficile andarsene.

'It is difficult to leave.'

In (61), the infinitive andarsene functions as either the SUBJ of difficile or an extraposed
COMP, which is bound to the SUBJ. In either case, it has a closed function, and given the
absence of an anaphoric controller, arbitrary control arises. The fact that these infinitival
clauses never have nonnull complementizers is irrelevant to the determination of control.
Another consequence of our theory of control is that when a lexical item agrees
with (or governs) grammatical features of its SUBJ, it must also agree with (or govern)
functional controllers of its SUBJ; but it need not agree with anaphoric controllers of it
SUBJ. English has only vestigial systems of case marking and agreement, and so provides
little illustration of this consequence. However, research on Modern Arabic (Fassi Fehri
(1981)), Icelandic (Andrews (1982; to appear)), and Russian (Neidle (1982; in prepara-
tion)) bears out the predictions of the theory. For example, Andrews and Neidle show
that there are predicative adjuncts in both Russian and Icelandic that agree in case with
the subject of the clause in which they occur. In controlled clauses, these adjuncts show
two patterns of case marking: in one pattern, the adjunct necessarily agrees in case with
the controller of the clause; in the other, the adjunct may appear in a case (nominative
in Icelandic, dative in Russian) distinct from that of the controller, the case of the subject

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 396 JOAN BRESNAN

of the controlled clause. It turns out that agreement with the controller is necessary just
where there is independent syntactic evidence of functional control. No special rules
of agreement are required to account for these facts; this is exactly the result that our
theory predicts. Similarly, Fassi Fehri's careful syntactic analysis of Modern Arabic
shows that predicative adjectives exhibit different patterns of agreement depending upon
whether they are anaphorically or functionally controlled. (Note that while agreement
with a functional controller is necessary, agreement with an anaphoric controller is not
prohibited; however, it must be effected by a further rule.)
These results of the theory of control reveal a surprising interrelation among se-
mantic interpretation, syntactic function, and morphological agreement. Theories which
give a uniform syntactic representation to all control relations-for example, theories
which treat the controlled element everywhere as a pronominal anaphor 'PRO'-leave
this deep relation among interpretation, function, and agreement unexplained.

9.2. NP Subjects and Split Antecedents

Our theory implies that lexical NPs cannot appear as the subjects of functionally con-
trollable clauses (sections 4, 8.2). Moreover, given the consistency condition, it follows
that functionally controllable clauses cannot have split antecedents. (A pronoun that
refers to more than one noun phrase is said to have split antecedents; for example, in
Tom told Mary that they should leave, Tom and Mary can be split antecedents of they.)
Functional control by split antecedents is impossible, for the f-structures of each of the
functional controllers would be merged with that of the controlled subject, resulting in
a clash of features. In contrast, anaphorically controllable clauses may have lexical NP
subjects, and (provided that the conditions on anaphoric control permit) they may also
have split antecedents, since nonreflexive anaphors in general allow this. Thus, our
theory predicts the following correlation: if a controlled clause can have a lexical NP
subject, then (subject to conditions on anaphoric control) it can have split antecedents.
In English, both participle phrases and adjective phrases can function as predicative
adjuncts, controlled by some element of the clause:

(62) a. Angry at John, Mary left.

b. Being angry at John, Mary left.

Unlike the adjectival adjuncts (which, as we have seen in section 8.2, are functionally
controlled XADJS), the participial adjuncts have optional NP subjects, as shown by the
contrast between (63a) and (63b).1"

" Jane Simpson has pointed out that Stump (1981) discusses absolutives which might seem to be counter-
examples. However, his examples fall into several classes that appear to have special properties. One class
includes examples which could be analyzed as verbal participles, such as Her hair braided, Mary began to
put on make-up. A second class includes such examples as The children asleep, Bill watched TV; it is well
known that asleep belongs to a marked class of predicates, including awake, aglow, which have mixed adjectival
and nonadjectival properties. For example, in contrast to most adjectives, asleep cannot occur prenominally
and does allow modification by right in standard dialects of English: contrast an angry child, *an asleep child
and *%The child became right angry, The childfell right asleep. A third class consists of true adjectives which

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 397

(63) a. *John angry, Mary left.

b. John being angry, Mary left.

Thus, if the participles in (62b) and (63b) are instances of the same construction, it
cannot be the functionally controlled XADJ. Rather, it must be a closed ADJ having an
internal structure similar to the one shown in (64):

(64) S - NP VP
( SUBJ) = T = I
( k PART) = ing
Given this analysis, our theory then predicts that, unlike the open adjectival ad-
juncts, the closed participial adjuncts will show anaphoric and arbitrary control:

(65) a. *Hatless, it is unnecessary to bow.

b. Having taken off your hat, it is unnecessary to bow.

(Some speakers of English reject examples like (65b); it is unclear whether this is because
they class the participial adjuncts (62b) with the phrasal XADJS (62a) rather than with the
absolute (sentential) ADJS (63b), or because they wish to avoid the stylistic stigma of
"dangling participles". Our theory further predicts that, since the PRO subject of the
participial adjuncts is a nonreflexive anaphor, these adjuncts may have split antecedents
(conditions on anaphoric control permitting):

(66) a. *Mary lost track of John because, angry at each other, he had gone one
way and she the other.
b. Mary lost track of John because, having been angry at each other, he had
gone one way and she the other.

(Again, there are speakers who react to (66) as to (65).)

Note that the converse of this generalization does not hold. It is true that if the
subject position of a clause is functionally controlled, then it cannot be replaced by a
lexical NP. But it is not true that if the subject position of a clause cannot be replaced
by a lexical NP, the subject is functionally controlled. The reason is that other factors
than functional control may prevent lexical expression of the subject of a clause. For
example, the subject of an infinitival S in English can be lexically expressed if and only
if the complementizerfor is present, as illustrated in (67).

(67) a. Louise gestured/said for me to follow her.

b. *Louise gestured/said me to follow her.
c. *Louise gestured/said for to follow her.
d. Louise gestured/said to follow her.

modify body parts or possessions of the subject: an example is Her face scarlet, Mary left the room. These
examples appear to be semantically restricted in ways that require further study: contrast the previous examples
with ??Her lobster scarlet, Mary left the kitchen and ?Her dog sick, Mary left the room. In general, the
existence of exceptions to categorial generalizations, including the generalization that adjectives and nouns
lack direct objects (Maling (to appear)), supports a markedness interpretation of the category definitions.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 398 JOAN BRESNAN

However, since wh-phrases in English exclude the presence of for, the subjects of wh-
infinitival clauses cannot be lexically expressed:

(68) a. Whether to grant equal rights to women is under debate.

b. *Whether men to grant equal rights to women is under debate.

Nevertheless, it is clear that the wh-infinitival clause of (68a) cannot be functionally

controlled, because it bears the closed function SUBJ and has the indefinite interpretation
of arbitrary control. Similarly, there exist lexically governed cases of obligatory ana-
phoric control in nominals (Kisala (in preparation)).

9.3. Functional Restrictions on Controllers

A further generalization of control follows from this theory. Recall that in lexically
induced functional control relations, OBL0 cannot be a controller. Since the oblique
functions are marked by prepositions in English (Bresnan (1979)), it follows that prep-
ositional objects cannot be lexically induced functional controllers in English. The fol-
lowing examples were previously given to illustrate this point.

(69) a. I presented it to John dead.

b. *1 presented John with it dead.

There are apparent counterexamples to this generalization, such as (70a,b):

(70) a. Louise signaled to Ted to follow her.

b. Mary relies on John to dress himself.

However, these are either cases of anaphoric control or cases of Verb-Preposition

Incorporation (V-P Incorporation; see Bresnan (1982b)). For example, the fact that we
find Louise signaled to Ted for him to follow her, in which the subject is a lexically
expressed NP, indicates that the complement of signal must be anaphorically, not func-
tionally, controlled in (70a). As for rely on, it is subject to the rule of V-P Incorporation,
which produces the two lexical forms shown in (71).

(71) a. [rely]v, ( t PRED) = 'RELY-ON((SUBJ)(OBLON))'

b. [rely on]v, ( T PRED) = 'RELY-ON((SUBJ)(OBJ))'

The verb with lexical form (71a) is inserted into the structure shown in figure 15a; the
verb with lexical form (71b) is inserted into the structure shown in figure 15b.
In the structure shown in figure 15a, on John is a constituent, while in the structure
shown in figure 15b, it is not. Moreover, John is an OBLON in figure 15a, but an OBJ in
figure 15b. The former structure accounts for the possibility of It is on John that Mary
relies, in which the constituent on John is clefted; the latter structure accounts for John
is relied on by Mary, in which John as an OBJ has passivized (Bresnan (1982b)). The
theory of control adopted here predicts that functional control of a complement to rely
on should be possible only when the preposition is incorporated into the verb as in figure
15b. Hence, if the functionally controlled complement is present, on John must fail to

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 399

a. S

NP VP

V PP

P NP

Mary relies on John

b. S

NP VP

V NP

Mary relies on John

Figure 15
Verb-Preposition Incorporation

form a constituent. This explains the contrast between (72a) and (72b) and the possibility
of (72c).

(72) a. *It is on John that Mary relies to dress himself.

b. It is John that Mary relies on to dress himself.
c. John is relied on by Mary to dress himself.

It is interesting to compare this explanation with an alternative proposal, due to

Williams (1980). Williams proposes the C-Command Condition, which states that in
obligatory control relations, the controller must c-command 'PRO'. Let us first observe
that Williams's "obligatory control" corresponds to our functional control. That is, the
central properties that Williams takes to be characteristic of obligatory control follow
from our theory of functional control. First, Williams stipulates that in obligatory control
there must be an antecedent. In our theory, this property follows from the completeness
condition (section 5); if the functional controller is omitted from a sentence in which
functional control obtains, the f-structure value of both the controller and the controlled
subject will be missing, producing an incomplete f-structure. Second, Williams stipulates
that in obligatory control, lexical NPs cannot appear in the position of the controlled

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 400 JOAN BRESNAN

subject. As we have seen, this property follows from our theory of categories and
functional control (sections 4, 8.2). Next, in Williams's obligatory control the antecedent
must be thematically or grammatically uniquely determined. In our theory, this property
follows directly from the lexical rule of Functional Control (20). In cases of construc-
tionally induced functional control, as we have seen in examples (28)-(3 1), the choice
of grammatical controller is not uniquely determined; it remains true, however, that any
function which is chosen as a functional controller must be unique (i.e. "split antece-
dents" are impossible). Williams suggests as a further property that the antecedent must
precede the controlled subject position, but indicates that this property may be accidental
to English. This property is not implied by our theory. In fact, the property does not
hold even for English, as examples (32) and (33) show. The final property of Williams's
obligatory control is that the antecedent must c-command the controlled subject position.
This property of course does not follow from our theory, since the restrictions on the
controller are functional, not structural. Nevertheless, because of the way functions are
encoded in English constituent structure, certain effects of the C-Command Condition
already follow from our theory.
Since NP objects of prepositions will fail to c-command nodes outside of the PP,
the C-Command Condition could also account for the above examples. However, it is
easy to distinguish the two explanations empirically: the theory adopted here defines
control in terms of relations between functions; the alternative theory defines it in terms
of relations between constituent structure positions. Now it happens that in English, the
OBLH function is syntactically realized as PP and the OBJ function as NP immediately
dominated by VP, but this is not universally the case. There are languages in which
oblique objects are syntactically realized as case-marked NPs and languages in which
direct objects are syntactically realized as prepositional objects. In such languages, the
predictions of the C-Command Condition diverge from those of the present theory of
control.
In Icelandic, oblique objects can be realized as case-marked NPs which are not
dominated by the node PP (Andrews (1982), Levin and Simpson (1981)). If our theory
is correct, the oblique objects in Icelandic cannot be lexically induced functional con-
trollers; but if the c-command hypothesis is correct, these oblique NPs can be obligatory
controllers, since they c-command the controlled subject. The evidence given in (73)
and (74), cited by Levin and Simpson (1981), confirms our theory. These are examples
of state predication, discussed in section 8, in which a complement is predicated of a
theme of the main verb.

(73) Hann rendi matnum hraum frai mer.

he-nom stole the-meat-dat raw-dat from me-dat
'He stole the meat from me raw.'
(74) *Hann randi mig matnum hraum.
he-nom robbed me-acc the-meat-dat raw-dat
('He robbed me of the meat raw.')

Levin and Simpson argue on independent grounds that in (73) matnum 'meat' is a direct

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 401

object of the verb, while in (74) it is an oblique object of the verb. For example, the
direct object passivizes, while the oblique object does not. Thus, it is the function of the
NP, and not the c-command relation, that correctly predicts its possible control relations.
Spanish provides a case in which an NP does not c-command constituents of the
verb phrase but does have the OBJ function. In Spanish, animate direct objects are
realized as objects of the preposition a. According to the c-command hypothesis, these
animate objects should not be possible obligatory controllers; according to our theory,
they should be. The evidence below (provided by M. Montalbetti, personal communi-
cation) favors the functional theory.

(75) Juan la encontro a ella borracha.

Juan CL-acc met her drunk-fem
'Juan met her drunk.'
(76) *Juan le hablo a ella borracha.
Juan CL-dat spoke to her drunk-fem
('Juan spoke to her drunk.')

In (75), the NP ella 'her' does not c-command borracha 'drunk', but control is never-
theless possible. The presence of the doubled accusative clitic la shows that ella is the
direct object of the verb 'meet' in (75). Example (76) illustrates that the OBLGO fails to
be a possible controller, as both theories would predict.
This evidence from Icelandic and Spanish shows that it is the function and not the
c-structure position that determines the eligibility of an "obligatory" controller in func-
tional control relations. Furthermore, we can see even in English that the C-Command
Condition cannot be a necessary property of obligatory controllers. We have already
seen examples of constructionally induced functional control, such as (77), which show
an obligatory control relation between an NP and a predicate that it fails to c-command.
(In this respect, AP adjuncts differ from participial adjuncts.)

(77) John said he was passed by Mary in the hall yesterday drunk.

Note that a controller of the predicative AP in (77) is required in the sentence, and that
a lexical NP in place of the controlled subject is not possible in the AP construction:

(78) *It was raining outside drunk.

(79) *John was passed in the hall, Mary drunk.

Hence, the C-Command Condition is both too weak and too strong. Where objects and
oblique objects happen to be syntactically encoded as NPs and PPs, respectively, the
C-Command Condition will appear to hold, but the underlying restrictions on "obliga-
tory" controllers are functional, not structural.

9.4. Visser's Generalization

Another consequence of our theory is that the controllers of lexically induced functional
control relations must change under lexical operations on function assignments. Con-

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 402 JOAN BRESNAN

sider, for example, the verb keep as used in John kept laughing; its lexical entry includes
the lexical form shown in (80a).

(80) a. 'KEEP((SUBJ)(XCOMP))'
b. ( t SUBJ) = ( j XCOMP SUBJ)

The lexical rule of Functional Control obligatorily adds the unmarked control equation
shown in (80b) to the lexical entry for keep. Now, like other intransitive verbs in English,
keep undergoes a lexical rule of Causativization. Causativization adds a new AGENT
argument, which is assigned the function SUBJ, and the SUBJ of the intransitive verb
becomes the OBJ of the causativized verb by the rule (OBJ) -> (SUBJ) (cf. Mohanan (1981
1982a)). Because Causativization replaces lexical occurrences of (SUBJ), or ( SUBJ), by
(OBJ), the control equation changes as shown in (81).

(81) a. 'KEEPcaus((SUBJ)(OBJ)(XCOMP))'
b. ( T OBJ) = ( T XCOMP SUBJ)

The result is to shift control to the object of the causativized verb, as in Mary kept John
laughing.
The transfer of control in the examples John kept laughing and Mary kept John
laughing could also be obtained if Causativization simply asserted that the THEME ar-
gument is the controller of predicative complements (cf. Anderson (1977)). However,
this alternative formulation fails to account for the fact that nonthematic subjects can
also become objects under Causativization; compare Tabs started being kept on celeb-
rities and What started tabs being kept on celebrities?
Predictably, control is also transferred under the lexical operation of Passivization:
(OBJ) -> (SUBJ). For example, Passivization of the transitive active lexical entry giv
in (81) produces the intransitive lexical entry shown in (82). (For detailed formulation
and discussion of the rule, see Bresnan (1982b).)

(82) a. 'KEEPcaus((OBLAG)(SUBJ)(XCOMP))'
b. ( T SUBJ) = ( T XCOMP SUBJ)

Accordingly, we find that in John was kept laughing by Mary, control has transferred
to the passive subject.
The theory of control predicts, however, that Passivization of a verb whose SUBJ
is a (lexically induced) functional controller should be impossible, for Passivization shifts
the semantically unrestricted function (SUBJ) to the semantically restricted function
(OBLo) or to +, and these cannot be functional controllers. This explains Visser's gen-
eralization (pointed out in Bresnan (1976b))-the observation that verbs whose com-
plements are predicated of their subjects do not passivize (Visser (1963-1973, part 111.2,
2118)). Consider, for example, the contrast between (83) and (84).

(83) a. His friends regard him as pompous.

b. Aunt Mary made the boys good little housekeepers.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 403

c. Her friends had failed her in some unclear way.

d. The vision struck him blind.
e. Frank persuaded Mary to leave.
(84) a. He strikes his friends as pompous.
b. The boys made Aunt Mary good little housekeepers.
c. Max failed her as a husband.
d. The vision struck him as a beautiful revelation.
e. Mary promised Frank to leave.

The verbs in (83) have objects as functional controllers; those in (84) have subjects as
functional controllers. Only the former type has corresponding well-formed passives:

(85) a. He is regarded by his friends as pompous.

b. The boys were made good little housekeepers by Aunt Mary.
c. She had been failed by her friends in some unclear way.
d. He was struck blind by the vision.
e. Mary was persuaded to leave by Frank.
(86) a. *His friends are struck (by him) as pompous.
b. *Aunt Mary was made good little housekeepers (by the boys).
c. *She was failed (by Max) as a husband.
d. *He was struck (by the vision) as a beautiful revelation.
e. *Frank was promised to leave (by Mary).

Note that the examples in (86) are ill-formed whether or not the agentive by-phrase is
expressed.
Where the complement subject is not functionally controlled, Passivization should
be possible. This accounts for the contrast between (87) and (88).

(87) a. John promised Mary to be on time.

b. *Mary was promised by John to be on time.
(88) a. John promised Mary that he would be on time.
b. Mary was promised by John that he would be on time.

The following observations explain the contrast in more detail. The lexical form for
promise in (88) is 'PROMISE((SUBJ)(OBJ)(COMP))'; the COMP (closed complemept) is not a
functionally controllable clause. In contrast, the lexical form for promise in (87) has an
open XCOMP, which must be grammatically controlled. However, this is a marked control
relation, for even though promise has an object, it evinces subject control.

(89) 'PROMISE((SUBJ)(OBJ)(XCOMP))'
( T SUBJ) = ( t XCOMP SUBJ)

In (89), Passivization would substitute (OBJAG) or + for ( t SUBJ), yielding an impossible

functional control relation, since only the semantically unrestricted functions can be
functional controllers.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 404 JOAN BRESNAN

Certain apparent exceptions to the conditions on grammatical control relations are

actually instances of anaphoric control. Examples are given in (90).

(90) a. Mary was never promised to be allowed to leave.

b. It was never promised to Mary to be allowed to leave.
c. To be allowed to leave was never promised to Mary.

(90a) appears to contradict Visser's generalization, in that a subject control verb (prom-
ise) passivizes, but (90b,c) show that to be allowed to leave behaves like a closed COMP,
which cannot be functionally controlled. Unlike XCOMPS, closed COMPS may undergo It
Extraposition ((90b)) and may appear as the subjects of finite clauses ((90c)). Thus, (90)
is analogous to (91).

(91) a. Mary was never promised that she would be allowed to leave.
b. It was never promised to Mary that she would be allowed to leave.
c. That she would be allowed to leave was never promised to Mary.

In contrast, the infinitival complement in (92) does not behave like a COMP.
(92) a. John promised Mary to be on time.
b. *It was promised to Mary to be on time.
c. *To be on time was promised to Mary.

The reason that to be on time in (92) is interpretable as an XCOMP but not as a COMP
may be that there is a subtle difference between the meaning of promise . . . to and
promise . . . that. To promise ... to ... is to commit oneself to someone to act in some
way. In X promised Y to Z, X is an agent, Y is a goal or beneficiary, and Z is an action.
If the action Z is not in X's power to perform, the result is odd: cf. #John promised
Mary to be tall, #John promised Mary to be allowed to feed himself. But to promise
. . .that . . . is to commit oneself to a (possibly abstract) transfer of some benefit to
someone. In X promised Y that Z, X is an agent and perhaps source as well, Y is a goal,
and Z is a theme. This sense of promise is comparable to its sense in John promised
Mary an apple, which means 'John promised Mary that she could have an apple'. With
promise . . . that, if the theme Z is not transferable to the goal, the result is odd. This
would account for the difference between (90a), Mary was never promised to be allowed
to leave, and (87b), *Mary was promised by John to be on time; to be allowed to leave
can be interpreted as an abstractly transferable benefit ("permission to leave"), but to
be on time cannot. In short, the semantic content of the complement biases the choice
of predicates promise-to or promise-that. Thematic constraints on anaphoric control
imply that a theme will be anaphorically controlled by its goal, experiencer, or possessor
(Bresnan (1979)); hence, this analysis of promise-that subsumes the interpretation of
(90) under the same rule that interprets (93). 12

12 Although the passives Mary was never promised permission to leave and Mary was never promised to
be allowed to leave are both judged grammatical by many speakers, the actives Fred promised Mary permission
to leave and Fred promised Mary to be allowed to leave differ, the latter being less acceptable for many. It
is unclear exactly why this should be so. Perhaps the infinitival marker to creates a strong lexical bias toward
the promise-to reading, which is incompatible with the content of the infinitival phrase (cf. Ford, Bresnan,
and Kaplan (1982)).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 405

(93) a. No one ever promised Mary permission to leave.

b. Mary was never promised permission to leave.
c. Permission to leave was never promised to Mary.

The unexpected deviance of some examples of the type in (87) had occasionally
been noticed in the literature of transformational grammar (Chomsky (1965, 229), Jenkin
(1972, 200ff)). However, its significance has only recently been appreciated, since it was
an assumption in transformational grammar so long and widely maintained as to seem
a truism, that however passives may differ from actives in scope relations, discourse
function, or stylistic effect, the grammatical relations relevant to the semantic inter-
pretation of passive sentences are identical to those of corresponding active sentences. 13
Indeed, this is the fundamental reason for postulating a Passive transformation. Visser's
generalization shows that this truism is false. In the lexical analysis of passivization
required by our theory, grammatical relations do vary under passivization while the
basic predicate argument structure remains invariant (Bresnan (1982b), Roberts (1982),
Mohanan (1982a), Halvorsen (1981)). This offers a straightforward explanation of Vis-
ser's generalization, as we have just seen.
Since the significance of Visser's generalization was first pointed out (Bresnan
(1976b)), there have been several attempts to provide an account for it that is consistent
with transformational theories of passivization (Wasow (1977)), Anderson (1977), Chom-
sky (1977b; 1980b), Williams (1980)), but none of them are satisfactory. The first attempt
in transformational grammar to explain the deviance of examples like (87) appears to
be that of Jenkins (1972, 200ff), who proposed a constraint stating that the object of by
cannot be coreferential with an implicit or expressed subject of a complement. If this
were true, it would itself require explanation, but the following examples show that this
by-phrase constraint does not express the correct generalization:

(94) a. John had been promised by Mary that she would meet him at the station.
b. John expects a promise by Mary to remain faithful to him.
c. An attempt by the gang of four to advance themselves now would be
foolhardy.

In all of these examples, the object of by is or can be understood as coreferential with

the subject of the complement. In (94a) the complement subject she is expressed, while
in (94b,c) the subjects of the infinitival complements are implicit. Moreover, the by-phrase
constraint fails to account for the persisting deviance of examples (86a-e) and (87b)
when the by-phrases are omitted. In contrast, our theory correctly distinguishes examples
like (94a-c) from the deviant example (87b). In (94a), the that-complement is a (closed)
COMP, which cannot be functionally controlled. In (94b,c), the nominals promise and
attempt cannot have undergone Passivization, since derived nominals lack both SUBJ

13 Recall that we use the term grammatical relations in a theory-neutral way to refer to the mapping
between semantic predicate argument structure and syntactic constituent structure. Thus, grammatical re-
lations are to be distinguished from grammatical functions.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 406 JOAN BRESNAN

and OBJ (Rappaport (1980)). '4 The anaphoric relations between the objects of by and the
implicit subjects of the infinitives are similar to those in (95).

(95) a. John awaits a promise from Mary to remain faithful to him.

b. An attempt on the part of the gang of four to advance themselves now
would be foolhardy.

Anaphoric control in nominals is analyzed in Kisala (in preparation).

Anderson (1977) also attempts to explain Visser's generalization. He formulates a
thematic control rule which states that a complement is attributed to the theme of its
clause (Anderson (1977, 375-376)) and implies that the passivized subject in (87b) is not
thematic. However, this account provides no insight into the contrast between Mary
was promised to be allowed to leave and *Mary was promised to be on time, for in
neither example can Mary be considered the theme. Moreover, Anderson's hypothesis,
like Jenkins's, fails to explain why the infinitival clause can be predicated of a nontheme
in (94b,c). Finally, as mentioned above in the discussion of Tabs started being kept on
celebrities, complements can be predicated of nonthematic subjects and objects. As
Borkin (1974) has observed, in many idiolects of English, examples like (96a) are found
alongside the construction (96b).

(96) a. There struck me as being too few women in positions of power.

b. Too few women struck me as being in positions of power.

Our theory explains why these examples fail to passivize:

(97) a. *1 was struck (by there) as being too few women in positions of power.
b. *1 was struck (by too few women) as being in positions of power.

However, Anderson's hypothesis offers no explanation for the contrast between (96a)
and (97a) and for the parallel behavior of (96a,b) and (97a,b), for in neither (96a) nor
(97a) is the complement predicated of a theme.
Still another recent attempt to explain Visser's generalization in a way consistent
with transformational theory is found in Chomsky (1977b; 1980b). Chomsky proposes
that passive sentences have the structure shown in figure 16 (see Chomsky (1977b, 13)).
As indicated in the figure, an NP Movement transformation has shifted the object Bill
of the verbs promise, persuade into position as subject of be; NP Movement leaves a
trace NP2 which is coindexed with the moved NP2. The NP, in the S-complement t
promise and persuade is a 'PRO' which must be coindexed with a subject or object NP

'4 According to Rappaport's analysis, passivization appears to apply in derived nominals because (1)
nominalized verbs inherit the predicate argument structures of their base verbs, but not the grammatical
function assignments of those verbs; (2) the grammatical function assignments of nominals include only
Poss(essor) and the semantically restricted functions. Thus, in the destruction of the city by the Romans the
of-phrase is an OBLTH and the by-phrase is an OBLAG. Rappaport's analysis explains why both types of phrases
appear with nominalizations of intransitive verbs: the arrival of John, a loud sneeze by Bill. John is the THEME
of the nominal arrival and the AGENT of the nominal sneeze. It also explains why the nominalizations corre-
sponding to Mary destroyed John and Mary surprised John differ in form: Mary's destruction of John, *Mary's
surprise of John, John's surprise at Mary. John is the TH of destroy, but the EXP of surprise.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 407

NP2 VP1

V AP

VP2

V NP2 S

NPx VP

Bill was EN persuade e COMP e to ieave

T promise
In this structure, promise has no subject, so co
subject is impossible; the object of persuade controls its complement's
subject.

Figure 16
A configurational account of Visser's generalization (Chomsky 1977b)

of these verbs. Chomsky (1977b, 13-14) writes:

The rule of control for persuade applies within VP2, assigning to x of NP, the value 2, as
before, persuade being the main verb of VP2 with NP2 as its object. But the rule of control
for promise is inapplicable since promise in [figure 16] is not the main verb of VP1 with the
surface subject Bill; compare in contrast [figure 17], where promise [is] the main verb of the
VP sister to the matrix subject.

Since NPx cannot be assigned subject control for promise in figure 16, the passive
sentence Bill was promised to leave is ungrammatical. Chomsky concludes, "More
generally, it follows that a verb with subject control (promise, strike, etc.) cannot appear
in the passive" (p. 14). This account is technically inadequate in certain ways that are

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
408 JOAN BRESNAN

NP1 VP1

V NP2S

NPZ VP

John persuaded Bill COMP e to leave

promised

Control of complement's subject is possible for both promise and per-

suade in this structure.

Figure 17
Examples of control in active structures

corrected in Chomsky (1980b). However, the basis for Chomsky's explanation of Visser's
generalization remains that "there is no subject under passive, hence no way for control
to be assigned [to passivized verbs of obligatory subject control]" (Chomsky (1980b)),
and this account is carried over in Chomsky's subsequent government and binding theory
(Chomsky (1981, 75-76)).
Thus, for Chomsky, the apparent subject of a passive verb-e.g. Bill in figure 16-
is actually only the subject of be; it is related to the deep object of the passivized main
verb only by the trace left by NP Movement. Hence, when a passive verb is embedded
in the complement to any verb that assigns control to its complement subject, what is
assigned control is the PRO subject NP, of be, as figure 18 illustrates. The binding relati
between the subject of be and the object of elect is accomplished by the transformation
of NP Movement. The rule of control for try will thus assign both occurrences of x in
figure 18 the value 1.
The key assumption in Chomsky's explanation for Visser's generalization is that
passivized VPs by themselves have no subject. From this assumption it follows straight-
forwardly that such VPs could not be direct complements to verbs that assign comple-
ment subject control; this situation is schematized in figure 19. Here, neither NP1 nor

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 409

NP2 can control the subject of VP2, because VP2, by hypothesis, has no subject. Thus,
the ungrammaticality of (98a,b) is consistent with Chomsky's assumption.

(98) a. *Bill tried elected.

b. *Mary forced Bill examined by a doctor.

The ill-formed example *Bill tried elected contrasts with the well-formed example Bill
was elected because try, unlike be, assigns a thematic role to its subject, precluding NP
Movement into the subject position. (Formally, this would be ruled out by the Theta
Criterion (Chomsky (1981)).)
However, examples like (98a,b) could be ill-formed simply because try and force
require infinitival complements in their intended uses here, while be clearly does not.
To control for this possibility, we should find verbs that resemble try and force in having
thematic subjects and objects, but differ from these verbs in permitting noninfinitival
phrasal complements to be predicated of their arguments. Sample verbs that would fit

NP1 VP,

NP/ VP

V AP

V NPx

Bill tried COMP e to be EN elect e

Figure 18
Passivization by NP Movement in the complement

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
410 JOAN BRESNAN

NP1 VP1

(lexical) V (NP2) AP

(lexical) EN VP2

V NP3

Figure 19
An impossible configuration

into the main verb position of structures like figure 19 with ordinary AP complements
would be stay, keep, feel, and get in examples like the following.

(99) a. The room stayed empty.

b. She kept the room empty.
c. John felt glum.
d. That got John glum again.

These verbs have dyadic or triadic predicate argument structures, assigning thematic
roles to their subjects and to their objects (in addition to their predicative complements):

(100) a. *The tables have stayed turned on us.

(= Our roles have stayed reversed.)
b. The tables have been turned on us.
(= Our roles have been reversed.)
(101) a. *The reporters will keep the cat out of the bag.
(= The reporters will keep the secret exposed.)
b. The reporters will consider the cat out of the bag.
(= The reporters will consider the secret exposed.)
(102) a. *Our goose feels cooked.
(= We feel in trouble with no way out.)
b. Our goose is cooked.
(= We are in trouble with no way out.)
(103) a. *Trying to help his friends got excessive advantage taken of John.
(= Trying to help his friends got John taken advantage of excessively.)
b. I consider excessive advantage to have been taken of John by his friends.
(= I consider John taken advantage of excessively by his friends.)

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 411

(Example (103b) is ill-formed without to have been because consider takes AP comple-
ments, and a thematic condition on adjectival passives rules out semantically empty
subjects (see Bresnan (1982b) for a detailed account). In contrast, the addition of to be
in (103a) does not improve the example.)
Of course, the selectional restrictions of the AP-complement verbs are not identical
to those of the infinitival-complement verbs; the latter often require animate arguments
where the former require only themes. Contrast The situation stayedlfelt unbearable
with #The situation tried to be unbearable, and She keptlgot the room warm with #She
forced the room to be warm. Nevertheless, in their imposition of selectional restrictions
on their subjects and objects, these verbs behave like try and force rather than seem
and expect:

(104) a. Our goose seems to be cooked. (idiomatic)

b. Our goose tries to be cooked. (unidiomatic)
(105) a. I expected excessive advantage to be taken of John by his friends.
b. *1 forced excessive advantage to be taken of John by his friends.

If Chomsky's explanation of Visser's generalization is correct, then passivized VPs

should not appear as direct complements to the verbs stay, keep, feel, and get as used
above. NP Movement from the complement object position is impossible; the pas-
sivized VP itself lacks a subject, and the matrix verb assigns a thematic role to its own
subject and object, precluding NP Movement into these positions. Consider then the
crucial examples (106a-d).

(106) a. The wall will stay painted black.

b. She kept the wall painted black.
c. John felt betrayed and made to look like a fool by Susan.
d. That got John handed a can of worms.

Indeed, we have already seen passives embedded in some of the examples given above
(e.g. Trying to help his friends got John taken advantage of). The passivized VPs or
APs in each of these well-formed examples show that Chomsky's explanation for Visser's
generalization cannot be correct.'5 (The lexical-functional theory of passivization cor-
rectly predicts the existence of such examples (see Bresnan (1982b)).)
One might think that Chomsky's proposal could be rescued by assuming that a rule
of To Be Deletion has applied in (107a-d), deleting the verb be from the immediate

'5 It is easy to verify that the VP complements in these examples contain passivized verbs: only Vs and
not As in English can take direct NP complements (Wasow (1977), Bresnan (1978)) and made in made to look
like a fool has the verbal sense 'caused' and not the adjectival sense 'created'. In any event, however, adjectival
passives pose exactly the same problem for the NP Movement theory as do verbal passives, for Bresnan
(1982b) has shown that adjectival passives are formed by conversion from verbal passives and that there is
no separate lexical process of adjectival passivization distinct from verbal passivization. This means that the
NP Movement approach can express the relation between adjectival and verbal passives only by assuming
that both are derived by the same NP Movement rule, as Fiengo (1977) proposes (in violation of the lexicalist
hypothesis; see Bresnan (1982b) for a discussion of problems inherent in Fiengo's analysis). Thus, whether
they contain adjectival or verbal passives, examples like John felt betrayed by Mary are inconsistent with
Chomsky's explanation of Visser's generalization.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 412 JOAN BRESNAN

complement of the main verb. However, there is evidence against an underlying be in

these examples:

(107) a. *The wall will stay to be painted black.

b. *She kept the wall being painted black.
c. *John felt to be betrayed and to be made to look like a fool by Susan.
d. *That got John to be handed a can of worms.

Moreover, when the phrasal complement of one of these verbs does contain be, as in
John felt eager to be on the list, it may not delete: *John felt eager on the list. Under
the To Be Deletion hypothesis, the rule of To Be Deletion would have to apply obligatorily
and locally to verbs like stay, feel, keep, and get and yet be inapplicable to verbs like
try and force. These properties of lexical specificity and strict locality are of course
precisely the characteristics of lexical subcategorization. In fact, Baker (1979) has argued
from considerations of language acquisition that replacing the To Be Deletion analysis
with the direct subcategorization hypothesis yields a more explanatory theory of gram-
mar. His arguments indicate that rules such as To Be Deletion should be eliminated in
principle from grammars.
Since Chomsky (1970; 1980b) recognizes the theoretical deficiencies of the To Be
Deletion analysis, the only other recourse would seem to be to assume that the verbs
stay, keep, feel, and get in (106) have sentential complements containing be in logical
form, as Chomsky (1980b) has proposed for verbs like strike. Accordingly, (106a) would
have the logical form sketched in figure 20. Note that the circled portion of this figure
was not present in the syntactic structure shown in figure 19; this is structure that has
been created in logical form by lexically governed structure-building rules. A rule of
control for the verb stay could then assign control to the PRO subject of the logical form
S which was not present in the transformational derivation of the example. Observe,
however, that since the PRO subject NP2 of be was not present in the transformational
derivation of the example, it remains unbound to the PRO object of paint. Whereas the
structure-preserving transformation of NP Movement established this crucial connection
in figure 18, NP Movement cannot have applied in figure 20, because, by hypothesis,
no subject NP position existed for the object NP to be moved into in the transformational
derivation of the sentence. Hence, there is no explanation for the fact that in The wall
will stay painted black, the phrase painted black has the passive interpretation. The
same argument applies, mutatis mutandis, to each of the other examples of (106). We
might think of adding a new rule of NP Coindexing to logical form to rescue this example
by binding the syntactic object of painted in figure 20 to its logical form subject. However,
this would of course make the transformational rule of NP Movement redundant; more-
over, it would raise the question of why examples like *John tried examined by a doctor,
*Mary forced John examined by a doctor, and even *Bill was promised to leave cannot
be rescued in the same way, by NP Coindexing in logical form.
Thus, the To Be Insertion analysis is just as deficient as the To Be Deletion analysis.
In fact, in Chomsky's current government and binding theory, both analyses must be

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 413

NP1 AUX VP

z:' \s 0Mp os ' ' V NP3 N P

the wall will stay to be N paint e black

Figure 20
Structure-building in "logical form"

rejected as incompatible with the Projection Principle. However, this leaves the gov-
ernment and binding theory with no explanation at all for Visser's generalization.
Another recent attempt to explain Visser's generalization in a transformational
theory of passivization is Wasow's (1977, 352-353). Wasow suggests that the general-
ization could be explained if passives were transformationally derived from the 'logically
subjectless" deep structures proposed in Bresnan (1972, 139-143) and in Emonds (1970).
Except for the assumption that NP Movement leaves a trace, this analysis of passivi-
zation is essentially the one adopted in Chomsky (1977a; 1980b; 1981), and it suffers
from the same deficiencies.
Williams (1980) offers an account of Visser's generalization based on his C-Com-
mand Condition. He observes that in examples like (108a), Tom fails to c-command
smart.

(108) a. *Mary was struck by Tom as smart.

b. *Mary was struck as smart.

Williams also observes that the ill-formedness of (108a) persists when the by-phrase is

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 414 JOAN BRESNAN

omitted ((108b)), and he therefore adds the stipulation that 'PRO' must have an antecedent
in such examples. However, we have already seen in section 9.3 that the C-Command
Condition provides neither a necessary nor a sufficient characterization of obligatory
control and predication relations; furthermore, the need for an antecedent in (108b) is
precisely what must be explained.
We see, then, that despite a concerted effort, no adequate explanation for Visser's
generalization has yet been found within transformational theories of passivization. The
fact that this generalization is among the consequences of our theory counts as very
strong support. Before leaving this topic, let us consider a very different approach to
the problem.
Bach (1979; 1980) has proposed a novel explanation for Visser's generalization
within the framework of Montague grammar as developed by Bach, Dowty, Thomason,
and others (Partee (1976)). In this framework, syntactic structures are represented by
a categorial grammar, which projects combinatorial information about sentence structure
onto basic lexical categories. For example, the transitive verb hit is assigned to the
lexical category IVP/NP (something that combines with an NP to form an intransitive
verb phrase), while the intransitive verb die is assigned to the lexical category IVP
(intransitive verb phrase, that is, something with which an NP combines to form a
sentence). Similarly, the verb persuade is assigned to the lexical category (IVP/NP)/to
VP (something which takes an infinitival VP to form a phrase which takes an NP to
form an intransitive VP), while the verb promise is assigned to the lexical category (IVP/
to VP)/NP (something that takes an NP to form a phrase that takes an infinitival VP to
form an intransitive VP). The motivation for this categorial difference between persuade
and promise is in part semantic: unlike persuade and like try, promise is a verb whose
complement is predicated of its subject. Bach comments on this classification (1979,
518): "Just as persuade to go is a complex representative of the category of transitive
verb, so promise Mary is a complex representative of the same category as try." In
Bach's terms, persuade to go and hit are transitive verb phrases and promise Mary and
try are intransitive verb phrases. Bach's explanation for Visser's generalization is then
that Passivization applies only to transitive verb phrases and that its effect is to turn
them into intransitive verb phrases.
Like the lexical-functional framework, a categorial framework encodes a large
amount of surface syntactic information into lexical entries. Because this syntactic
information is encoded in the syntactic categories of individual lexical items in a cate-
gorial framework, however, many cross-categorial lexical generalizations cannot be
expressed directly in the lexicon. For example, there is no uniform lexical characteri-
zation of the notion transitive verb, for hit and persuade have very different lexical
categories (owing to their different combinatorial properties). Hence, a rule such as
Passivization that applies to transitive verbs cannot be formulated as a lexical rule.
However, the cross-categorial nature of Passivization can be captured by formulating
it as a phrasal rule that operates on transitive verb phrases, for at the phrasal level of
representation, both persuade to go and hit are transitive verb phrases.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 415

As evidence in favor of this phrasal analysis of Passivization over a lexical approach,

Bach (1979; 1980) cites the existence of infinitivals that may modify transitive verb
phrases and passive verb phrases, but not intransitive verb phrases:

(109) a. Mary bought it to drive to work in.

b. It was bought to drive to work in.
c. *The Dean came in to talk to.

However, the assumption that these infinitivals can modify only transitive verb phrases
leaves examples like (1 lOa-c) unexplained:

(110) a. My car is always available (for you) to drive to work in.

b. I am always on hand for students to talk to.
c. It will serve nicely to fix drinks on.

It is clear that these infinitivals are not extraposed infinitival relative clauses, for pos-
sessed NPs and definite personal pronouns resist relative clauses in English: *My car
(for you) to drive to work in is always available, *Ifor students to talk to am always on
hand, *It to fix drinks on will serve nicely. Nor can the verb phrases containing the
adjective available, the idiomatic prepositional phrase on hand, and the intransitive verb
serve be naturally analyzed as transitive. Moreover, many transitive verb phrases fail
to permit these infinitival modifiers (e.g. *Mary ate it to amuse John with). Thus, it
appears that these infinitival modifiers are neither necessary nor sufficient indicators of
transitivity. More generally, Halvorsen (1981) has shown that a variety of arguments for
a phrasal passive in the recent literature (Bach (1980), Keenan (1980)) dissolve when
the lexical analysis of passivization is coupled with a theory of the semantic interpretation
of functional structures.
Is there then any evidence at all to choose between Bach's explanation of Visser's
generalization and the explanation proposed here? The fundamental difference between
the two is that on our account, Passivization fails because the functional control relation
between the subject and the complement is destroyed, while on Bach's account, Pas-
sivization fails because intransitive verb phrases cannot be passivized. Thus, according
to our analysis, the transitivity or intransitivity of the verb undergoing Passivization is
irrelevant; the crucial factor is that a (subject or object) AGENT controller of the com-
plement is not syntactically available. Therefore, our analysis would predict that an
intransitive verb whose subject functionally controlled a predicative complement could
not be passivized, even if intransitive verbs otherwise undergo Passivization. Indeed,
in the lexical theory of passivization (Bresnan (1982b)), intransitive Passivization can
be formulated as a subcase of transitive Passivization, resulting from omission of the
optional operation that replaces OBJ by SUBJ:

(III ) (SUBJ) (OBLAG)4+

((OBJ) (SUBJ))

In Bach's account, by contrast, transitivity is the crucial factor; the syntactic

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
416 JOAN BRESNAN

of the AGENT controller is not important. In Bach's theory of passivization, the AGENT
is existentially quantified over the passivized verb phrase, as in (112); hence, the AGENT
is semantically available as a controller, even if it is not syntactically expressed.

(112) John is regarded as happy.

j*( Xx 3y [[regard' ('happy')] (PP(x))](y))
= 3y [[[regard' ('happy')] (PP(^))](y)]

Thus, Bach's account of Visser's generalization would not generalize to the case of
intransitive passives; in particular, if intransitive verbs could undergo Passivization, an
intransitive verb whose subject controlled a predicative complement could be passivized.
(More detailed reasoning in support of this conclusion is given below.)
Intransitive Passivization does exist in a number of languages, Icelandic and Nor-
wegian among them. (The following evidence from Icelandic was provided by Joan
Maling.) In Icelandic, intransitive agentive verbs undergo Passivization, as in (1 13)-(114).

(113) Allir keyra hattulega i baenum.

all-nom-masc-pl drive-pl dangerously in town-dat
'Everyone drives dangerously in town.'
(114) Pa6 er keyrt hittulega i baenum.
it be-sg-3 driven-nom-neut-sg dangerously in town-dat
'There is dangerous driving in town.'
[Lit: 'It is driven dangerously in town.']

The morphology of the intransitive passive is the same as that of the transitive passive,
except that the passive participle appears in the neuter nominative singular form (the
unmarked gender, case, and number). Expression of the agent in intransitive Passivi-
zation is marginal. There is evidence that pad is a syntactic placeholder rather than a
subject (Maling and Zaenen (1978)).
In Icelandic, a predicative state complement controlled by the subject can appear
with intransitive verbs. Unlike an adverb, this complement agrees in gender, case, and
number with the subject:

( 15) Allir keyra fullir i benum.

all-nom-masc-pl drive drunk-nom-masc-pl in town
'Everyone drives drunk in town.'

However, no form of this subject-controlled complement can appear in the passivized

form of ( 15), as (116) shows:

(116) *Pao er keyrt fullt/fullum/fullir

it be-sg-3 driven-nom-neut-sg drunk-neut-sg/drunk-dat-pl/drunk-nom-pl
i banum.
in town
('There is driving drunk in town.')
[Lit:" 'It is driven drunk in town.']

The ungrammaticality of (1 16) is predicted by our theory, but not by Bach's. (See below.)

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 417

Similar evidence exists in Norwegian. (The following examples from Norwegian

were provided by Per-Kristian Halvorsen.) In Norwegian, as in Icelandic, agentive
intransitive verbs may undergo Passivization as transitive verbs do, except that the a
is not syntactically expressed.

(117) De kjorte gjennom byen.

they drove through the-town
'They drove through the town.'
(118) Det ble kj0rt gjennom byen.
it was driven through the-town
'There was driving through the town.'
[Lit: 'It was driven through the town.']

In Norwegian, a predicative state complement modifying the subject may appear in the
active but not the passive form of the sentence:

(119) De kj0rte fulle gjennom byen.

they drove drunk-pl through the-town
'They drove drunk through the town.
(120) *Det ble kj0rt fulle/full/fullt gjennom byen.
it was driven drunk-pl/drunk-sg/drunk-neut-sg through the-town
('There was driving drunk through the town.')
[Lit: 'It was driven drunk through the town.']

Thus, the evidence from intransitive passivization in Icelandic and Norwegian confirms
our theory that Visser's generalization follows from a syntactic restriction on subject
control rather than a condition on transitive verb phrase passivization.
The following reasoning (suggested by P.-K. Halvorsen) supports this conclusion
in detail. In general, Bach handles control of infinitival VPs by forming a lambda abstract
of the open sentence consisting of the VP and a variable for the subject. Thus, John
persuaded Mary to go is built up from persuade to go and Mary, John being added last.
Simplifying inessential details, the translation of persuade to go is an abstract:
Xx[persuade'('go'(x))(x)]. When this is combined with the translation of Mary, the x is
bound and Mary will be interpreted to be the subject of the complement go. Passivization
existentially quantifies the subject and abstracts over the object, producing Xz[3y Xx
[persuade'(go'(x)) (x)] (z)(y)], which reduces to Xz[3y persuade'('go'(z))(z)(y)]. What
would happen in the case of passive intransitives such as Det ble kj0rt 'There was
driving' [Lit: 'It was driven (intrans)'] in Norwegian? The passive intransitive could be
translated as 3y[drive'(y)] (again simplifying inessential details). This is of the wrong
type for an IVP, so let us assume that the translation is Xx3y[drive'(y)], using vacuous
lambda abstraction to perform a type adjustment. When the placeholder subject det 'it'
is combined with the IVP, the lambda disappears and the translation 3y[drive'(y)] re-
mains for the sentence Det ble kj0rt. What if there were a complement? An example
such as *Det ble kjirtfulle [Lit: 'It was driven drunk'] would be built up from kj0rtfulle
and det. The active phrase kj0rte fulle would be translated Xx[drive'('drunk'(x))(x)], in

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 418 JOAN BRESNAN

a fashion similar to try to go. (Note that, like try, drive can be used intransitively,
transitively, or with a complement. For evidence that state predicates like drunk are
(controlled) complements rather than adjuncts, see section 8 and Neidle (1982).) The
translation rule for passive would then give Az[3y^Ax[drive'(^drunk'(x))(x)](y)], which
reduces to Xz3y[drive'(Cdrunk'(y))(y)], which again combines with the placeholder subjec
det to yield the well-formed interpretation 3y[drive'(^drunk'(y))(y)] 'There was someone
driving drunk'.
In conclusion, Visser's generalization follows in its full generality from our theory.
Despite proposals by Jenkins (1972), Anderson (1977), Chomsky (1977b; 1980b), Wasow
(1977), Williams (1980), and Bach (1979; 1980), other theories leave it unexplained.

9.5. Bach's Generalization

In his study of passivization, Bach (1979) has also observed that where the object of a
verb is an obligatory controller, intransitivization is impossible. This is illustrated in
(121) and (122). In (121), the subject is the controller, and the object may be omitted;
in (122), the object is the controller, and may not be omitted.

(121) a. Louise promised Tom to be on time.

b. Louise promised to be on time.
(122) a. Louise taught Tom to smoke.
b. *Louise taught to smoke.

In fact, however, not all cases of obligatory control by an object rule out intransitivization
(a fact not explained by Bach's (1979) analysis):

(123) a. Louise signaled Tom to follow her.

b. Louise signaled to follow her.

What is the difference between (122) and (123)? Note that an NP subject can be expressed
in the case of (123) but not (122):

(124) *Louise taught Tom for him to smoke.

(125) Louise signaled Tom for him to follow her.

On the basis of this fact, we may tentatively hypothesize that the object in (122a) is a
functional controller of its complement, while the object in (123) is an anaphoric con-
troller of its complement. Thus, we can amend Bach's generalization to state that where
the object of a verb is a functional controller, intransitivization is impossible. In this
form, the generalization appears quite generally to be true:

(126) a. John called someone.

b. John called.
c. John called someone a fool.
d. *John called a fool.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 419

(127) a. John painted something.

b. John painted.
c. John painted something black.
d. *John painted black.

This generalization is the counterpart for objects of Visser's generalization for

subjects: since only subjects and objects can be functional controllers, it follows that
any lexical operations that eliminate subjects or objects will be inapplicable where they
would destroy functional control relations. Consider the control relation in example
(122). Assuming that teach has a basic lexical form 'TEACH((SUBJ)(OBJ)(XCOMP))', the
lexical rule of Functional Control (section 8.2) obligatorily adds the control relation
( T OBJ) = ( T XCOMP SUBJ) to the lexical entry. The lexical rule of Intransitiviza
replaces (OBJ) by + (Bresnan (1980)); it therefore could not apply to teach without
destroying the functional control relation, since only the semantically unrestricted func-
tions can be functional controllers, and + is not a function (see footnote 1).

9.6. Raising

Raising refers to the type of construction illustrated in (128). The italicized NPs are
sometimes referred to as "raised" NPs.

(128) a. John considered himself for a long time to be inferior to his brother.
b. Mary seems to like witty women.

Two properties characterize these examples of raising constructions: first, the verb has
a predicative complement whose understood subject is referentially dependent upon (in
fact, identical to) a pre- or postverbal NP; and second, this pre- or postverbal NP is not
a "logical", or thematic, argument of the verb. From the first property we see that
raising is a type of control relation (section 8). From the second property, we see that
it must be a functional control relation. To see this, suppose that raising involved
anaphoric control. The sentences would then be semantically incoherent, containing a
subject or object (the "raised" NP) that cannot be bound to any argument position
(Halvorsen (1981)). Thus, raising must be a functional control relation. From our theory
of functional control it follows at once that only the SUBJ of a predicative complement
can be "raised" and that "raised" NPs can appear only as SUBJ, OBJ, or OBJ2.
Let us now consider the problem of how raising constructions are expressed in
c-structure. The government and binding theory of Chomsky (1981) incorporates the Pro-
jection Principle and Theta Criterion as fundamental representational assumptions; to-
gether, they imply that all representations of syntactic structure are isomorphic to the
representation of thematic structure and reflect the subcategorization properties of lex-
ical items. Given these assumptions, raising constructions must have syntactic structures
of the form indicated in (129) and (130). The motivation for this analysis is that in their
thematic structures, raising verbs have a propositional argument, and S is the syntactic
form of such an argument.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
420 JOAN BRESNAN

(129) [s John considered [s hi

(130) [s Mary seems [S[NP e] to like witty women]]

However, this analysis is impossible in our theory, for by the principles of lexical and
syntactic encoding (sections 2 and 4), S-complement clauses can only be assigned closed
functions such as coMP(lement), which cannot be functionally controlled. (Moreover,
of course, we are constrained from using null structure to represent local dependencies
(section 6).) Furthermore, if we assumed that there is no functional control at all in
object-raising constructions like (129), we would incorrectly predict that raised objects
could not passivize. John is believed to be smart would be ill-formed for the same reason
that *John is believed is smart is ill-formed: namely, because John is not the object of
believe in such structures (Bresnan (1982b)). Hence, the c-structures shown in (129) and
(130) are inconsistent with our theory. This analysis also poses a theoretical problem for
Chomsky's government and binding theory, because the complement subject NPs must
be governed and assigned case, the S-complements lack finite INFL to assign case, and
government from a matrix verb normally cannot cross a maximal projection such as S.
Chomsky therefore postulates a rule which deletes the S, leaving only S, across which
the matrix verbs are supposed to be able to govern the NP subject of a complement
(Chomsky (1981, 66)). However, S Deletion has just the properties of strict locality and
lexical specificity that characterize lexical subcategorization. It is evident that the use
of such lexically governed "pruning transformations" (the term is due to Ross (1969)),
like the use of transformational exception features elsewhere (Lakoff (1970)), weakens
the theory of structure-dependent rules (Bresnan (1975)) and circumvents the strong
constraint imposed by the Projection Principle. Since such syntactic restructuring pro-
cesses are prohibited from our much more restrictive theory of constituency relations
(section 6) in any event, it might seem preferable simply to hypothesize that the raising
verbs directly subcategorize for S rather than S:

(131) [s John considered [s himself for a long time to be inferior to his brother]]
(132) [s Mary seems [S[NP e] to like witty women]]

But this option, too, is closed to us, for it follows from our theory of lexically and
syntactically encoded functions that in configurational structures, only maximal projec-
tions can be subcategorized for, and S is submaximal (section 3).
Thus, our theory requires that the infinitival complements to raising verbs like
consider and seem be maximal projections other than S. Since a verb heads these
complements in the above examples, the natural choice of category is VP, the maximal
projection of V. However, the category VP cannot contain a structural subject (section
4). Hence, the only analysis open to us is the one shown in (133), in which the NP
himself is generated as the object of considered outside the predicative VP complement.

(133) [s John considered [NP himselfl for a long time [vp to be inferior to his
brother]]
(134) [SINP Mary] seems [vP to like witty women]]

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 421

Note that our theory requires only that the complement to a raising verb be a maximal
projection other than S. Nothing in our theory requires that it be a VP (as opposed to
another phrasal category which is a maximal projection). Hence, in our theory, raising
should be possible with phrasal complements other than VP; in fact, both consider and
seem also allow phrasal complements of categories other than VP:

(135) John considers himself [AP inferior to his brother]

(136) Mary seems [AP fond of witty women]

Moreover, for many speakers, consider and seem in these uses may have nonthematic
objects and subjects (respectively); compare Consider your goose cooked, which can
be idiomatic, and Close tabs seem likely to be kept on celebrities.
Now according to our analysis, the raising verbs consider and seem take phrasal
complements; they differ in that consider is transitive, having a nonthematic object,
while seem is intransitive, having a nonthematic subject. The phrasal complements in
(133)-(136) have the function xcoMp, and the lexical entries for consider and seem will
therefore include the lexical forms given in (137) and (138).

(137) ( T PRED) 'CONSIDER((SUBJ)(XCOMP))(OBJ)'

(138) ( T PRED) 'SEEM((XCOMP))(SUBJ)'

The lexical rule of Functional Control (section 8.2) will expand these entries with the
control equations shown in (139) and (140).

(139) (T PRED) 'CONSIDER((SUBJ)(XCOMP))(OBJ)'

( T OBJ) = (T XCOMP SUBJ)
(140) (T PRED) 'SEEM((XCOMP))(SUBJ)'
(T SUBJ) ( T XCOMP SUBJ)
Note that it is because the OBJ of consider and the SUBJ of seem are nonthematic, that
they can designate expletives such as there and idiom chunks; see Bresnan (1982b) for
a detailed analysis. In sum, the c-structures and f-structures of our raising constructions
must be as shown in figures 21 and 22.
However, the analysis of raising required by our theory is incompatible with the
basic representational assumptions of Chomsky's (1981) theory (the Projection Principle
and the Theta Criterion). The assumptions of each theory require a phrase structure
representation of raising constructions which is inconsistent with the assumptions of the
other theory. Evidence bearing on the choice of phrasal representations of these con-
structions would therefore provide grounds for deciding which theory is more explan-
atory, in the sense of limiting the hypothesis space of grammar construction to the
descriptively optimal grammars. In fact, there is strong evidence favoring the analysis
of raising constructions given in figures 21 and 22 over Chomsky's sentential analyses
in (129)-(132).
First, because the raised NP in figure 21 is an object of the matrix clause, it
f-commands the subject in the matrix clause nucleus. If we had adopted the sentential

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
422 JOAN BRESNAN

c-structure

NP VP

V NP VP

V NPPAP

A PP

P NP

John considered himself to be inferior to his brother

f-structure

SUBJ EPRED JOHN]

TENSE PAST

PRED 'CONSIDER((SUBJ)(XCOMP))(OBJ)

OBJ rPRED PRO

-(reflexive)
XCOMP SUBJ

FIN

PRED BE((XCOMP))(SUBJ) )

XCOMP SUBJ 7

PRED INFERIOR-TO ((SUBJ)(OBLGO))

L OBLGO ["his brother"]

Figure 21
A raising-to-object construction

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 423

analysis, himself would be contained solely within the coMP(lement), in a position where
it could not f-command the subject of the matrix clause nucleus. We therefore predict
that anaphoric control of a subject clause by the object in a raising construction should
be possible (cf. section 8.3). Examples (141a-c) bear out this prediction.

(141) a. Contradicting himself will prove Mr. Jones to be a liar.

b. *Contradicting himself will prove that Mr. Jones is a liar.
c. Contradicting himself will discredit Mr. Jones.

We see that the object of the raising construction in (141a) behaves like the object of
the simple transitive verb in (141c), and not like the subject of the S-complement in
(141b). 16
Further confirmation for the correctness of our analysis comes from the postposing
of a "heavy" postverbal NP to VP-final position (noted in Bresnan (1972)). This rule
is bounded, in the sense that it can only postpose the postverbal NP over its sister
phrases (as observed in Bresnan (1976a)):

(142) a. I discussed [all those women riding motorcycles nowadays] with my

sister.
b. I discussed with my sister [all those women riding motorcycles nowadays].
c. *1 discussed [riding motorcycles nowadays] with my sister all those
women.

(The boundedness property follows at once if the rule operates on the VP phrase structure
rule, reordering the categories in a function-preserving manner.) Note in particular that
the postverbal NP all those women cannot be shifted outside the clause that immediately
contains it ((142c)). In our analysis of object raising constructions, the raised NP is the
object of the raising verb, and hence may undergo Heavy NP Shift over material in the
matrix VP without violating the boundedness of NP Shift. However, if the NP were
dominated by an S- or S-complement of the raising verb, shifting it over material in th
matrix verb phrase would violate the boundedness of NP Shift. Both count and consider
are object raising verbs in the author's dialect: cf. I counted there as being more than
20 women on the payroll, I consider there to be too many fools in this committee. The
following examples therefore confirm our analysis.

(143) a. I'll count [as being dead] from now on any corpse more than 3 days old.
b. I will consider [to be fools] in the weeks ahead all those who drop this
course.

These examples are naturally paraphrased as "From now on, I'll count any corpse
... as being dead" and "In the weeks ahead, I will consider all those ... to be fools,"

16 These facts also show that it is insufficient to state the constraint on anaphoric control as a condit
that the controller must be an argument of the matrix of the controlled complement (cf. Chomsky (19
Clearly, Mr. Jones is not a semantic argument of prove in Contradicting himself will prove Mr. Jones to
a liar.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
424 JOAN BRESNAN

c-structure

NP VP

V VP

'/VP

V NP

Mary seems to like witty women

f-structure

SUBJ ]PRED MARY']

TENSE PAST

PRED SEEM((XCOMP))(SUBJ)'

XCOMP SUBJ

FIN -

PRED LIKE ((SUBJ)(OBJ))'

LOBJ ["witty women"]

Figure 22
A raising-to-subject construction

with the adverbial material modifying the matrix clauses as showr. They demonstrate
that in c-structure, the raised NP must not form a constituent with the verbal phrase
that is predicated of it. (Postal (1974) makes a slightly different argument from NP Shift;
his (1977) response to objections by Bresnan (1976c) indicates that his argument for a
"raised" NP position in the VP is also valid.)
A third source of confirmation for our analysis is the rule of Right Node Raising,
which permits a constituent of any category to be omitted from two conjoined clauses
and positioned to their right, as shown in (144).

(144) a. Mary admired Peter the Great, and Catherine disliked Peter the Great.
b. Mary admired, and Catherine disliked, Peter the Great.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 425

In the author's dialect, nonconstituent sequences of categories cannot undergo Right

Node Raising.'7 Accordingly, examples (145a,b) contrast with (144a,b).

(145) a. Mary gave Peter the grape, and Catherine did not give Peter the grape.
b. *Mary gave, and Catherine did not give, Peter the grape.

Right Node Raising is a very useful test for constituency because it applies to all cat-
egories, including S (Bresnan (1974)):

(146) It is possible that, but it isn't obvious that, your theory is right.

In our analysis of object raising constructions, the raised NP and the predicative phrase
that it controls do not form a constituent in c-structure; hence, we correctly predict the
following contrasts in the restrictive dialect (cf. footnote 17):

(147) a. Mary believes Peter to be fat, but Catherine doesn't believe Peter to be
fat.
b. *Mary believes, but Catherine doesn't believe, Peter to be fat.

A similar argument has been made by Postal (1974).

To account for these facts in the government and binding framework, it has been
proposed that a verb can assign case only to an adjacent NP; Right Node Raising is
held to destroy the connectivity of case assignment by violating the Adjacency Condit
(Stowell (1981), Chomsky (1981)). Since, in fact, an object NP can be right-node-raised
away from its governing verb (as in (144b)), the claim must be that case is transmitted
to such an NP from its trace under Right Node Raising, and that there is no trace left
of an NP (such as Peter in (147b)) which is properly contained within a right node rai
constituent. Thus, we have two competing hypotheses to account for the ungrammat
cality of examples like (147b). Our theory claims that Peter to be fat is a nonconstitu
NP VP sequence; it therefore fails to undergo Right Node Raising (in the restrictive
dialect) for exactly the same reason that other nonconstituent sequences (e.g. (144b))
fail to do so. The alternative theory claims that (147b) and (144b) are ill-formed for
unrelated reasons. Example (144b) is said to be ill-formed (in the restrictive dialect)
because it is a nonconstituent sequence, but Peter to be fat in (147b) is a constituent S,
according to this theory. However, Peter is not case-marked by believes because Right
Node Raising destroys the connectivity of case assignment.
Can we find evidence to distinguish between the constituency hypothesis and the
connectivity hypothesis? Such evidence could take the form of a construction which
required exceptional case marking (in terms of the government and binding theory) and
yet behaved as a constituent under Right Node Raising. This evidence exists. Consider
the "acc-ing" construction, illustrated in (148).

17 Unlike the author, some speakers do weaken the condition on Right Node Raising to allow non-
constituent sequences (cf. Abbott (1976), Peterson (1981)). Thus, the Right Node Raising judgments of these
speakers cannot be used as a test for constituency.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 426 JOAN BRESNAN

(148) a. All his friends worry about Fred losing again.

b. Doctors disapprove of children smoking.

"Acc-ing" constructions are similar in their interpretation to "poss-ing" (gerund)

clauses, but they lack the genitive marker 's to assign case to their subject NPs (Fred
in (148a), children in (148b)). In the government and binding theory, these NPs must be
assigned case. Given that the acc-ing clauses lack a finite INFL, there are only two
possible mechanisms for doing this: either the position of the subject NP is inherently
case-marked, or it receives case exceptionally from a governing verb or preposition
outside the clause. Now if the NP position were inherently case-marked, the acc-ing
construction would occur in positions where exceptional government by a verb or prep-
osition does not hold, such as the subject position of a finite clause. However, if the NP
position were exceptionally case-marked, the acc-ing construction would not occur in
subject position. Examples like (149) indicate that the subject of the acc-ing construction
must be exceptionally governed.

(149) a. *Fred losing again is worried about by all his friends.

(cf. Fred's losing again is worried about by all his friends.)
b. *Children smoking is disapproved of by doctors.
(cf. Children's smoking is disapproved of by doctors.)

It is generally true that an acc-ing construction cannot function as a subject, even in

nonfinite clauses:

(150) *I'd like for Fred losing again not to bother you.
(cf. I'd like for Fred's losing again not to bother you.)

In this respect, then, acc-ing constructions resemble object raising constructions. They
differ crucially, however, in that the acc-ing construction gives every evidence of form-
ing a single constituent:

(151) a. It's Fred losing that I can't stand the thought of.
b. What doctors really disapprove of is children smoking.

Indeed, this assumption (together with some version of the A-over-A Condition (Bresnan
(1976a)) is needed by the government and binding theory to explain the failure of NP
Movement in examples like *Children are disapproved of smoking by doctors. We see
now, though, that both clefting (151a) and pseudoclefting (15 lb) preserve the connectivity
of exceptional case marking. And in fact, so does Right Node Raising:

(152) a. Mary's not looking forward to, and I can't stand the thought of, Fred
losing again.
b. Doctors disapprove of, and parents worry about, children smoking.

Since these examples are perfectly well-formed, we must reject Stowell's and Chomsky's
hypothesis, and conclude that Right Node Raising, like other dislocation rules, does
preserve the connectivity of exceptional case marking. This conclusion gains plausibility

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 427

because Right Node Raising preserves connectivity in general; for example, it preserves
the lexically governed control relations in (153a,b).

(153) a. Mary regards Bill, but Sue doesn't regard Bill, as too fond of himself.
b. Mary strikes Bill, but she doesn't strike Peter, as too fond of herself.

Consequently, failure of connectivity cannot be the explanation for the ungrammaticality

of (147b) *Mary believes, but Catherine doesn't believe, Peter to be fat. The non-
constituency hypothesis, which is the simpler account, provides a superior explanation.
In summary, our theory requires an analysis of raising constructions in which raising
verbs take nonthematic subjects or objects and predicative phrasal complements. There
is strong evidence that the analysis of raising constructions required by our theory is
optimal for the grammatical description of English. Recent work by Mohanan (1982a)
shows that the essential features of this analysis are supported by raising constructions
in Malayalam, a non-Indoeuropean language which differs strikingly from English in its
constituent structure properties. Since the optimal grammatical descriptions of raising
constructions are required by the representational assumptions of our theory, but in-
consistent with those of the government and binding theory, we conclude that our theory
is the more explanatory.

9.7. Pro-Drop

Pro-drop is a widespread linguistic phenomenon in which, under certain conditions, a

structural NP may be unexpressed, giving rise to a pronominal interpretation. Given our
constraints on the use of null c-structure categories (section 6) and our postulate that
all semantic forms are introduced lexically (section 8.3), it follows that in the unmarked
case, pro-drop should arise for the functions SUBJ, OBJ, and OBJ2. The reasoning is exactly
analogous to that given for the representation of the controlled ([+ u]) 'PRO' in the
beginning of section 8.3. Moreover, it follows from our theory of syntactic encoding that
when these functions are nonconfigurationally encoded, pro-drop should always be pos-
sible. The reason is that nonconfigurational encoding pairs functional schemata with
arbitrary X expanded by rules of the form C -> X* (section 4). Hence, the categorial
expression of the SUBJ, OBJ, or OBJ2 is always optional. From functional completeness
(section 5) it follows that when the categorial expressions of SUBJ, OBJ, or OBJ2 are
omitted, the optional functional anaphors carried by the governing lexical forms will be
required. Thus, in the unmarked case, pro-drop of SUBJ, OBJ, and OBJ2 will always be
possible in nonconfigurational languages.
Investigating this consequence will require careful descriptive studies of non-
configurational encoding, distinguishing between pro-drop and the rather different phe-
nomena of morphologically incorporated oblique pronouns (as in Navajo (Roberts
(1981))) and oblique pronominal clitics (as in Warlpiri auxiliaries (Simpson (in prepa-
ration))). However, Hale (to appear) has already observed that extensive pro-drop is a
hallmark of nonconfigurational languages, and research on nonconfigurational encoding

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 428 JOAN BRESNAN

in Malayalam (Mohanan (1981b; 1982a)) and Warlpiri (Simpson and Bresnan (in prep-
aration), Simpson (in preparation)) bears out this consequence of our theory.

10. Conclusion

All theories of grammar can be regarded as theories of grammatical relations in the

broad sense, that is, theories of the relations between the surface constituents of sen-
tences and their semantic predicate argument structures. One of the most important
issues in current linguistic theory concerns the nature of grammatical relations: are they
completely derivative of representations of constituent structure and semantic predicate
argument structure, or do they have an independent representation definable in terms
of syntactically primitive grammatical functions (e.g. SUBJ, oBJ)? While this issue will
undoubtedly continue to be debated, one conception of the problem must be rejected
as fundamentally mistaken. That is the view that theories which take only structuralist
concepts as syntactic primitives are inherently simpler and more explanatory than the-
ories which take functions as syntactic primitives, because the latter theories involve
an extra level of representational concepts. This view is mistaken because in constructing
a theory of grammar, we are free to take any set of concepts as primitive, defining in
terms of these a larger set of derivative concepts; and the linguistic concepts of function,
category, government, argument, case, and constituent structure are so closely inter-
related as to be interdefinable. In particular, the structuralist concept of category can
be reduced to more primitive functional concepts. Thus, in the theory presented in this
study, the grammatical functions (SUBJ, OBJ, etc.) are taken as syntactic primitives and
the phrase structure categories (NP, VP, etc.) are derivative notions. Contrary to this
common but erroneous view, there is simply no a priori basis for preferring a structuralist
theory of grammatical relations to a theory which employs grammatical functions as
syntactic primitives. On the contrary, there appears to be good reason to prefer the
function-based theory, since only in this type of theory has it been shown explicitly that
there is a universal representation of grammatical relations for both configurational and
nonconfigurational language types, and that the radical differences in the structural
expressions of subcategorization, government, and control relations in these languages
are predictable from a single parameter of variation: the surface realization of gram-
matical functions by c-structure configuration or by morphological case.
It is the axioms, or postulates, of a theory which give meaning to its primitive
concepts. For example, our theory postulates a narrowly constrained set of relations
between functions and lexical predicate argument structures, on the one hand, and
between functions and phrasal and morphological structures, on the other hand. From
these we have derived several significant generalizations. For example, it follows from
these assumptions of our theory that in configurational structures (a) the head of a phrase
governs the maximal projections that are immediately dominated by any of the projec-
tions of the head and (b) lexical items will appear to subcategorize for maximal projec-
tions of categories. In nonconfigurational structures, it follows (a) that lexical items will

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 429

appear to subcategorize for morphological case and not for phrase structure order, and
(b) that (in the unmarked case) pro-drop of SUBJ, OBJ, and OBJ2 will always be possible.
Similarly, our theory postulates that the set of grammatical functions is partitioned
into classes which differ in their lexical encoding properties; in particular, the seman-
tically restricted functions are lexically encoded in much more limited ways than the
semantically unrestricted functions. All lexically represented grammatical relations are
thereby constrained, including lexically induced control relations and lexically carried
functional anaphors. Among the various consequences we have enumerated, it follows
in our theory (a) that only SUBJ, OBJ, and OBJ2 are possible as lexically induced functional
controllers; (b) that morphosyntactically unexpressed 'PRO'S can bear only these func-
tions; and (c) that lexical transformations of control relations-such as Passivization
(Visser's generalization) and Intransitivization (Bach's generalization)-are severely
restricted.
Finally, our theory postulates that the SUBJ function is universally the controlled
element in functional control relations, and that functional control must be distinguished
from anaphoric control. Evidence has been given that in anaphoric control relations,
objects as well as subjects can be controlled, as our theory predicts. While other theories
of control have been based on rather different assumptions, our theory appears to offer
a much more constrained and explanatory account of government, anaphoric control,
and raising. For example, we have obtained an explanation of the distribution of gov-
erning morphemes, a uniform cross-configurational characterization of government re-
lations, and the predictability of the actual structural configurations in which government
relations can appear. This is achieved within a narrowly constrained theory of constit-
uency relations, which precludes the use of "Affix Hopping" and other structure-
deforming rules. Similarly, we have shown that anaphoric control relations are subject to
a condition definable on f-structure, namely, f-command, but not on c-structure. And
last, we have seen that our theory predicts the essential properties of raising construc-
tions: that only subjects undergo raising, that the subjects are raised only into subject
or object positions in the matrix clause, and that the propositional complement of the
raising predicate is expressed in c-structure discontinuously, in the form of a nonthematic
subject or object constituent and a separate phrasal complement constituent. This last
result is of particular interest, since it is inconsistent with the assumption that phrase
structure relations directly reflect semantic predicate argument structure (as in the Pro-
jection Principle and Theta Criterion of Chomsky (1981)).

References

Abbott, B. (1976) "Right Node Raising as a Test for Constituenthood," Linguistic Inquiry 7,
639-642.
Amritavalli, R. (1980) "Expressing Cross-Categorial Selectional Correspondences: An Alter
to the X Syntax Approach," Linguistic Analysis 6, 305-343.
Anderson, S. (1977) "Comments on the Paper by Wasow," in P. Culicover, T. Wasow, and
A. Akmajian, eds. (1977).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 430 JOAN BRESNAN

Andrews, A. (1981) "The Rudiments of Coordinate Structure in LFG," ms., Department of

Linguistics, Australian National University, Canberra, Australia.
Andrews, A. (1982) "The Representation of Case in Modern Icelandic," in J. Bresnan, ed. (1982a).
Andrews, A. (to appear) "Long Distance Agreement and Functional Interpretation," in
P. Jacobson and G. Pullum, eds. (to appear).
Bach, E. (1979) "Control in Montague Grammar," Linguistic Inquiry 10, 515-531.
Bach, E. (1980) "In Defense of Passive," Linguistics and Philosophy 3, 297-341.
Baker, C. L. (1979) "Syntactic Theory and the Projection Problem," Linguistic Inquiry 10,
533-581.
Bloomfield, L. (1933) Language, Holt, Rinehart and Winston, New York.
Borkin, A. (1974) Raising to Object Position: A Study in the Syntax and Semantics of Clause
Merging, Doctoral dissertation, University of Michigan, Ann Arbor, Michigan; in Univer-
sity of Michigan Papers in Linguistics 2.1.
Bresnan, J. (1972) Theory of Complementation in English Syntax, Doctoral dissertation, MIT,
Cambridge, Massachusetts; published in 1979 by Garland Press, New York.
Bresnan, J. (1974) "The Position of Certain Clause-Particles in Phrase Structure," Linguistic
Inquiry 5, 614-619.
Bresnan, J. (1975) "Transformations and Categories in Syntax," presented at the fifth International
Congress of Logic, Methodology and Philosophy of Science, London, Ontario, Canada,
1975; in R. Butts and J. Hintikka, eds. (1977) Basic Problems in Methodology and Lin-
guistics, Reidel, Dordrecht.
Bresnan, J. (1976a) "On the Form and Functioning of Transformations," Linguistic Inquiry 7,
3-40.
Bresnan, J. (1976b) "Toward a Realistic Model of Transformational Grammar," presented at the
Convocation on Communications, March 9-10, 1976, MIT, Cambridge, Massachusetts.
Bresnan, J. (1976c) "Nonarguments for Raising," Linguistic Inquiry 7, 485-501.
Bresnan, J. (1978) "A Realistic Transformational Grammar," in M. Halle, J. Bresnan, and
G. Miller, eds., Linguistic Theory and Psychological Reality, MIT Press, Cambridge, Mas-
sachusetts.
Bresnan, J. (1979) "A Theory of Grammatical Representation," duplicated lecture notes,
Department of Linguistics and Philosophy, MIT, Cambridge, Massachusetts.
Bresnan, J. (1980) "Polyadicity," in T. Hoekstra, H. van der Hulst, and M. Moortgat, eds. (1980);
in revised form in J. Bresnan, ed. (1982a).
Bresnan, J. (1981) "An Approach to Universal Grammar and the Mental Representation of Lan-
guage," Cognition 10, 39-52.
Bresnan, J., ed. (1982a) The Mental Representation of Grammatical Relations, MIT Press, Cam-
bridge, Massachusetts.
Bresnan, J. (1982b) "The Passive in Lexical Theory," in J. Bresnan, ed. (1982a).
Bresnan, J. and J. Grimshaw (1978) "The Syntax of Free Relatives in English," Linguistic Inquiry
9, 331-391.
Bresnan, J., R. Kaplan, P. S. Peters, and A. Zaenen (in preparation) "Cross-Serial Dependencies
in Dutch," forthcoming in Linguistic Inquiry 13.4.
Chomsky, N. (1965) Aspects of the Theory of Syntax, MIT Press, Cambridge, Massachusetts.
Chomsky, N. (1969) "Quine's Empirical Assumptions," in D. Davidson and J. Hintikka, eds.,
Words and Objections: Essays on the Work of W. V. Quine, Humanities Press, New York.
Chomsky, N. (1970) "Remarks on Nominalization," in R. Jacobs and P. Rosenbaum, eds.,
Readings in English Transformational Grammar, Ginn, Waltham, Massachusetts.
Chomsky, N. (1977a) "On Wh-Movement," in P. Culicover, T. Wasow, and A. Akmajian, eds.
(1977).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 431

Chomsky, N. (1977b) Essays on Form and Interpretation, North-Holland, New York.

Chomsky, N. (1980a) "On the Representation of Form and Function," presented at the C.N.R.S.
Conference at Royaumont, France, June 1980.
Chomsky, N. (1980b) "On Binding," Linguistic Inquiry 11, 1-46.
Chomsky, N. (1981) Lectures on Government and Binding, Foris, Dordrecht.
Chomsky, N. and M. Halle (1968) The Sound Pattern of English, Harper and Row, New York.
Comrie, B. (1977) "In Defense of Spontaneous Demotion: The Impersonal Passive," in P. Cole
and J. Sadock, eds., Syntax and Semantics, Vol. 8: Grammatical Relations, Academic
Press, New York.
Craig, C. (1977) The Structure of Jacaltec, The University of Texas Press, Austin, Texas.
Culicover, P., T. Wasow, and A. Akmajian, eds. (1977) Formal Syntax, Academic Press, New
York.
Dowty, D. (to appear) "Grammatical Relations and Montague Grammar," in P. Jacobson and
G. Pullum, eds. (to appear).
Emonds, J. (1970) Root and Structure-Preserving Transformations, Doctoral dissertation, MIT,
Cambridge, Massachusetts.
Emonds, J. (1976) A Transformational Approach to English Syntax, Academic Press, New York.
Evers, A. (1975) The Transformational Cycle in Dutch and German, Doctoral dissertation, Rijks-
universiteit, Utrecht, Holland.
Falk, Y. (1980a) Word Order and Rules of Syntax, undergraduate thesis, Brandeis University,
Waltham, Massachusetts.
Falk, Y. (1980b) "The English Auxiliary System: A Lexical-Functional Analysis," ms., Depart-
ment of Linguistics and Philosophy, MIT, Cambridge, Massachusetts.
Fassi Fehri, A. (1981) Complementation et Anaphore en Arabe Moderne: Une Approche Lexicale
Fonctionnelle, These de Doctorat d'Etat, Universite de Paris III.
Fiengo, R. (1977) "On Trace Theory," Linguistic Inquiry 8, 35-61.
Ford, M. (1982) "Sentence Planning Units: Implications for the Speaker's Representation of
Meaningful Relations Underlying Sentences," in J. Bresnan, ed. (1982a).
Ford, M., J. Bresnan, and R. Kaplan (1982) "A Competence-Based Theory of Syntactic Closure,"
in J. Bresnan, ed. (1982a).
Gazdar, G. (in press) "Phrase Structure Grammar," in P. Jacobson and G. Pullum, eds. (to
appear).
Grimshaw, J. (1979) "Complement Selection and the Lexicon," Linguistic Inquiry 10, 279-326.
Grimshaw, J. (1981) "Subcategorization of Grammatical Relations," presented at the Harvard
Mini-Conference on the Representation of Grammatical Relations, Harvard University,
December 12, 1981.
Grimshaw, J. (1982) "On the Lexical Representation of Romance Reflexive Clitics," in J. Bresnan,
ed. (1982a).
Hale, K. (1978) "Obviation in Modern Irish," ms., Department of Linguistics and Philosophy,
MIT, Cambridge, Massachusetts.
Hale, K. (1979) "On the Position of Walbiri in a Typology of the Base," published in 1981 by the
Indiana University Linguistics Club, Bloomington, Indiana.
Hale, K. (to appear) "Preliminary Remarks on Configurationality," Proceedings of the Twelfth
Annual Meeting of NELS, MIT, Cambridge, Massachusetts.
Halliday, M. (1967) "Notes on Transitivity and Theme," Journal of Linguistics 1, 37-81.
Halvorsen, P.-K. (1981) "An Interpretation Procedure for Functional Structures," ms., The Cen-
ter for Cognitive Science, MIT, Cambridge, Massachusetts.
Hoekstra, T., H. van der Hulst, and M. Moortgat, eds. (1980) Lexical Grammar, Foris, Dordrecht.
Hornstein, N. (1977) "S and X' Convention," Linguistic Analysis 3, 137-176.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 432 JOAN BRESNAN

Hornstein, N. and A. Weinberg (1981) "Case Theory and Preposition Stranding," Linguistic
Inquiry 12, 55-91.
Jackendoff, R. (1972) Semantic Interpretation in Generative Grammar, MIT Press, Cambridge,
Massachusetts.
Jackendoff, R. (1976) "Toward an Explanatory Semantic Representation," Linguistic Inquiry 7,
89- 150.
Jackendoff, R. (1977) X Syntax: A Study of Phrase Structure, Linguistic Inquiry Monograph 2,
MIT Press, Cambridge, Massachusetts.
Jacobson, P. and G. Pullum, eds. (to appear) On the Nature of Syntactic Representation, Reidel,
Dordrecht.
Jenkins, L. (1972) Modality in English Syntax, Doctoral dissertation, MIT, Cambridge, Massa-
chusetts.
Kaplan, R. and J. Bresnan (1982) "Lexical-Functional Grammar: A Formal System for Gram-
matical Representation," in J. Bresnan, ed. (1982a).
Keenan, E. (1980) "Passive Is Phrasal Not (Sentential or Lexical)," in T. Hoekstra, H. van der
Hulst, and M. Moortgat, eds. (1980).
Kisala, J. (in preparation) "Thematically-Conditioned Control Phenomena in Nominals," ms.,
Department of Linguistics and Philosophy, MIT, Cambridge, Massachusetts.
Lakoff, G. (1970) Irregularity in Syntax, Holt, Rinehart and Winston, New York.
Lapointe, S. (1980) A Theory of Grammatical Agreement, Doctoral dissertation, The University
of Massachusetts, Amherst.
Lehiste, I. (1970) Suprasegmentals, MIT Press, Cambridge, Massachusetts.
Levin, L. (1982) "Sluicing: A Lexical Interpretation Procedure," in J. Bresnan, ed. (1982a).
Levin, L. and J. Simpson (1981) "Quirky Case and Lexical Representations of Icelandic Verbs,"
presented at the Chicago Linguistic Society, April 30, 1981.
Lieber, R. (1980) The Organization of the Lexicon, Doctoral dissertation, MIT, Cambridge, Mas-
sachusetts.
McCloskey, J. (1979) Transformational Syntax and Model Theoretic Semantics: A Case Study in
Modern Irish, Reidel, Dordrecht.
McCloskey, J. (1980) "Is There Raising in Modern Irish?" Eriu 31, 59-99.
McCloskey, J. (1981) "Notes on Case and Control in Modern Irish," ms., Department of Irish,
University College Dublin, Ireland.
Maling, J. (to appear) "Transitive Adjectives: A Case of Categorial Reanalysis," in F. Heny, ed.,
Papers from the Fourth Groningen Round Table, July 1980.
Maling, J. and A. Zaenen (1978) "The Nonuniversality of a Surface Filter," Linguistic Inquiry
9, 475-497.
Manzini, R. (1980) "On Control," ms., Department of Linguistics and Philosophy, MIT, Cam-
bridge, Massachusetts.
Marantz, A. (1981) A Theory of Grammatical Relations, Doctoral dissertation, MIT, Cambridge,
Massachusetts.
Miller, G. A. and P. Johnson-Laird (1976) Language and Perception, Harvard University Press,
Cambridge, Massachusetts.
Mohanan, K. P. (198 la) "Move NP or Lexical Rules: Evidence from Malayalam Causativization,"
ms., Department of Linguistics and Philosophy, MIT, Cambridge, Massachusetts.
Mohanan, K. P. (1981b) "Grammatical and Anaphoric Control," ms., Department of Linguistics
and Philosophy, MIT, Cambridge, Massachusetts.
Mohanan, K. P. (1982a) "Grammatical Relations and Clause Structure in Malayalam," in
J. Bresnan, ed. (1982a).

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 CONTROL AND COMPLEMENTATION 433

Mohanan, K. P. (1982b) "Infinitival Subjects, Government, and Abstract Case," Linguistic In-
quiry 13.2.
Mohanan, K. P. (1982c) Lexical Phonology, Doctoral dissertation, MIT, Cambridge, Massachu-
setts.
Montague, R. (1973) "On the Proper Treatment of Quantification in Ordinary English," in
J. Hintikka, J. Moravcsik, and P. Suppes, eds., Approaches to Natural Language, Reidel,
Dordrecht.
Montalbetti, M. (1981) "Consistency and Clitics," ms., Department of Linguistics and Philosophy,
MIT, Cambridge, Massachusetts.
Nash, D. (1980) Topics in Walbiri Grammar, Doctoral dissertation, MIT, Cambridge, Massachu-
setts.
Neidle, C. (1982) "Case Agreement in Russian," in J. Bresnan, ed. (1982a).
Neidle, C. (in preparation) Topics in Russian Syntax, Doctoral dissertation, MIT, Cambridge,
Massachusetts.
Partee, B., ed. (1976) Montague Grammar, Academic Press, New York.
Perlmutter, D., ed. (to appear) Studies in Relational Grammar.
Peterson, P. (1981a) "Conjunction in LFG," ms., The Center for Cognitive Science, MIT, Cam-
bridge, Massachusetts.
Peterson, P. (1981b) The Progressive Construction in English: An Analysis and Its Implications,
Doctoral dissertation, University of Newcastle, NSW, Australia.
Pinker, S. (1982) "A Theory of the Acquisition of Lexical-Interpretive Grammars," in J. Bresnan,
ed. (1982a).
Pinker, S. and D. Lebeaux (1981) Language Learnability and Language Development, forthcoming
from Harvard University Press, Cambridge, Massachusetts.
Postal, P. (1974) On Raising: One Rule of English Grammar and Its Theoretical Implications, MIT
Press, Cambridge, Massachusetts.
Postal, P. (1977) "About a 'Nonargument' for Raising," Linguistic Inquiry 8, 141-154.
Rappaport, M. (1980) "On the Derivation of Derived Nominals," ms., Department of Linguistics
and Philosophy, MIT, Cambridge, Massachusetts.
Reinhart, T. (1976) The Syntactic Domain of Anaphora, Doctoral dissertation, MIT, Cambridge,
Massachusetts.
Roberts, J. (1981) "Toward a Unified Analysis of the Passive in Navajo," ms., Department of
Linguistics and Philosophy, MIT, Cambridge, Massachusetts.
Robertson, J. (1976) The Structure of Pronoun Incorporation in the Mayan Verbal Complex,
Doctoral dissertation, Harvard University, Cambridge, Massachusetts; published in 1980,
Garland Press, New York.
Ross, J. (1969) "A Proposed Rule of Tree-Pruning," in D. Reibel and S. Schane, eds., Modern
Studies in English, Prentice-Hall, Englewood Cliffs, New Jersey.
Rothstein, S. (1981) "On Preposition Stranding," ms., Department of Linguistics, MIT, Cam-
bridge, Massachusetts.
Selkirk, E. (1981) "English Word Structure," ms., Department of Linguistics, The University of
Massachusetts, Amherst.
Simpson, J. (in preparation) Control and Predication in Warlpiri, Doctoral dissertation, MIT,
Cambridge, Massachusetts.
Simpson, J. and J. Bresnan (in preparation) "Control and Obviation in Warlpiri." [Paper presente
at the First Annual West Coast Conference on Formal Linguistics, Stanford University,
January 1982.]
Sridhar, S. (1979) "New Evidence for Spontaneous Demotion," South Asian Languages Analysis 1.

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms
 434 JOAN BRESNAN

Stowell, T. (1981) Origins of Phrase Structure, Doctoral dissertation, MIT, Cambridge, Massa-
chusetts.
Stump, G. (1981) The Syntax and Semantics of Free Adjuncts and Absolutes, Doctoral dissertation,
Ohio State University, Columbus, Ohio.
Thompson, S. (1973) "On Subjectless Gerunds in English," Foundations of Language 9, 374-383.
Visser, F. Th. (1963-1973) An Historical Syntax of the English Language, E. J. Brill, Leiden.
Wasow, T. (1977) "Transformations and the Lexicon," in P. Culicover, T. Wasow, and
A. Akmajian, eds. (1977).
Williams, E. (1980) "Predication," Linguistic Inquiry 11, 203-238.
Zaenen, A. (1981) "Characterizing Syntactic Binding Domains," Occasional Paper #17, The
Center for Cognitive Science, MIT, Cambridge, Massachusetts.

20D-219
MIT
Cambridge, Massachusetts 02139

This content downloaded from

99.4.123.47 on Wed, 24 Feb 2021 17:48:55 UTC
All use subject to https://about.jstor.org/terms