Deconstructing Markedness: A Predictability-Based Approach*

ELIZABETH HUME
Ohio State University

0. Introduction
Since first proposed as a linguistic term by Trubetzkoy (1939), the notion of
markedness has come to occupy a position of considerable importance in phonol-
ogy and other areas of linguistics. However, since Trubetzkoy’s time, markedness
has acquired a much broader meaning. The term “unmarked” is generally syn-
onymous with, for example, simpler, more common, easier to produce, acquired
earlier, etc. It is no longer limited to relations between elements on a language-
specific basis, as Trubetzkoy assumed. Rather, markedness has come to refer to
the universals of language (e.g., Jakobson 1963, 1990; Greenberg 1966), deter-
mined by Universal Grammar (Chomsky and Halle 1968, Kean 1975, and many
others following them). Further, it has grown from a simple classificatory term to
a predictive scientific concept (e.g., Kiparsky 1985, Calabrese 1995, Rice 1996,
de Lacy 2002).
There are many serious problems with the notion of markedness, as I outline
in section 1. The result, I suggest, is that markedness is not predictive and hence,
not a scientific concept. The root of the problem is this: due to the vagueness of
the concept, markedness, it is unclear what markedness diagnostics, e.g. neutrali-
zation, simplicity, deletion, are actually diagnosing. What we are lacking is a clear
understanding of the basis of markedness.
I argue that it is predictability; that is, traditional markedness diagnostics are
actually providing evidence for a linguistic element’s predictability within a sys-
tem. An element with greater predictability patterns as less marked than a corre-
sponding, less predictable, one. Predictability is determined by a complex of fac-

*
Special thanks to Mary Beckman, Chris Brew, Nick Clements, Hope Dawson, Kathleen Currie
Hall, Keith Johnson, Hyunsook Kang, Grant McGuire, Jeff Mielke, Anton Rytting, and the OSU
Phonetics/Phonology group (aka Phonies) for their most valuable input. My thanks also to audi-
ence members at the MOT Workshop in Phonology (2004) and BLS, in particular, Juliette
Blevins, Marie-Hélène Côté, Heather Goad, Bruce Hayes, Larry Hyman, John Ohala, Sharon
Peperkamp, and Glyne Piggott. I, of course, am responsible for the views expressed in this paper
as well as for any errors.

182
 Markedness: A Predictability-Based Approach

tors, some language-specific and some universal, as discussed below. This paper
focuses primarily on the contribution of one language-specific factor: language
experience. As I hope to show, this approach is superior to traditional views of
markedness for a number of reasons: first, it correctly predicts observed language
patterns both at the language-specific and universal level; second, it correctly re-
flects the probabilistic nature of markedness; and finally, it is quantifiable, thus
moving us closer to a predictive, scientific theory of sound patterns.

1. Problems with the Traditional Approach to Markedness

Developing a predictive and unified theory of markedness has not been a trivial
undertaking. In fact, a wide range of diagnostics have been proposed (see Rice
1999 for discussion) including: phonetic instability, articulatory simplicity, per-
ceptual salience, neutralization, epenthesis, assimilation, segment deletion, distri-
bution, structural complexity, language acquisition, sound change, creole genesis,
cross-language frequency, and implicational relations. Yet, much of the criteria
assumed to provide evidence for markedness has been brought into question for
the reason that sound patterns contradicting markedness claims are commonly ob-
served at the level of the individual language, e.g., assimilation (Hume and Tser-
danelis 2002, Hume 2003), passive neutralization (Hume 2003), underspecifica-
tion and default values (Steriade 1995; Clements 1988, 1993; Mohanan 1993;
Calabrese 1995), neutralization (Odden, p.c.), epenthesis (Vaux 2002; also, see
below), cross-language phoneme frequency (Battistella 1990, Rice 1999), impli-
cational relations (Rice 1999), creole genesis (Thomason 1993), child language
acquisition (e.g., Menn 1983, Beckman et al. 2003), sound change (e.g., Lass
1975, Ladefoged 1984).
Consider assimilation, for example. When comparing members of a feature
class in assimilation, it is commonly assumed that “the unmarked pole of an op-
position is lost or obscured, with the marked pole remaining…In assimilation, the
marked features within a class are active…the unmarked features, on the other
hand, are passive, or inert…overridden by other features” (Rice 1999:4). Put an-
other way, marked features resist modification while unmarked features are sub-
ject to change. Problematic for this view of markedness is the observation that
virtually any place of articulation can pattern as unmarked in some language
(Hume and Tserdanelis 2002, Hume 2003).
Developing a predictive theory of markedness has also proven difficult due to
apparent inconsistencies and contradictions among the diagnostics themselves.
Consider, for example, perceptual salience, a term commonly used to refer to the
number and quality of the acoustic/auditory cues in the speech signal that a lis-
tener may use to identify a given sound or sound sequence. A sound with weak
phonetic cues, and hence low salience, is more likely to undergo phonological
processes such as assimilation, reduction and deletion, than sounds with better
cues. In cases such as this, the member of the opposition with low salience is con-
sidered unmarked. On the other hand, sounds and sequences with good cues and

183
 Elizabeth Hume

thus higher salience are also considered unmarked. The CV syllable is a proto-
typical example, widely considered to be the universally unmarked syllable type.
Perceptual salience is thus used in apparently contradictory ways to refer to the
same markedness value: an unmarked element can have low salience or it can
have high salience.
Another example illustrating seemingly contradictory predictions of the diag-
nostics involves epenthesis and deletion. The epenthetic vowel or consonant is
generally taken to correspond to the unmarked segment of a language. Deletion of
one member of a class of sounds also denotes unmarked status. What this means
is that an unmarked sound is not only more apt to delete, it is also more apt to be
inserted. To put another way, the unmarked segment is both the most preferred
and the least preferred segment in the language.
Due to problems such as these, an objective definition of what would validate
or falsify markedness is lacking. Is markedness falsified if the evidence from dif-
ferent criteria conflict? If not, how do we determine the markedness status of an
element? For example, a sound that is perceptually salient may be relatively sim-
ple from a production perspective. Does this make it marked in a system or un-
marked? Given the lack of an objective statement of what markedness is, the con-
cept tends to be used as is convenient; patterns that support a particular view of
markedness are used as evidence for that view while those that do not, tend to be
labeled “irregular,” “exceptional,” or “irrelevant.”
Following Menn (1983), it is clear that markedness does not explain anything;
it is something that needs to be explained. As I argue below, the concept predict-
ability is able to do just this.

2. Predictability
Markedness is best considered a probabilistic notion with predictability positively
correlated with unmarkedness. As I show below, observed markedness patterns
follow from this approach in a straightforward manner. In short, an element that is
predictable within a system is less crucial to successful communication than one
that is less predictable. Consequently, a predictable element is a more likely can-
didate for reduction, deletion, change, etc. Recall that these are all traditional di-
agnostics for unmarkedness. Interestingly, a more predictable element is also
more likely to be the epenthetic segment, another diagnostic of unmarkedness,
since language users are biased towards predictable elements (see below). Thus,
by taking into account the predictability of a segment in a language system, we
are able to predict both its instability and a speaker/hearer’s bias towards it. Fur-
ther, unlike traditional approaches to markedness, cross-linguistic variability is
not only accounted for, it is predicted, as I briefly discuss below.
There are several factors that come into play in determining the predictability
of a given linguistic element. As I lay out in greater detail in Hume (2004b,c),
these include perceptual salience, articulatory simplicity, functional load, social
factors (e.g., prestige value), and the speaker/hearer’s experience with the usage

184
 Markedness: A Predictability-Based Approach

of linguistic elements, e.g. sounds, words. All are crucial in determining how pre-
dictable an element will be in a language system. It is the influence of language
experience that I focus on in this paper. Cross-linguistic variability and hence,
language-specific markedness patterns result from the contribution of the system-
dependent factors: functional load, social factors, and experience.

3. Experience
Experience is important because it is correlated with expectation. As described by
Feather (1982), expectation, like other cognitive concepts, recognizes “the indi-
vidual’s capacity to process information so that it becomes organized and set
within a context of meaning. These organized residues of experience then become
important filters for future information processing, serving as benchmarks, crite-
ria, or reference frames against which new information can be tested. They also
guide the form that behavior takes.” Formally, expectation is the probability (P) of
x given y.

(1) EXPECTATION: P (x|y)

The more experience that one has with some element, the greater is the expecta-
tion that that element will occur. Experience is thus also correlated with predict-
ability.
The impact of experience on language learning, language change, language
processing, and language production is well-established (see, among others, Zipf
1932; Bybee, e.g. 1985, 2001; Jurafsky et al. 2001; Frisch 1996; Frisch et al.
2000; Luce 1986; Lindblom 1990; Pierrehumbert 1994; Pitt and McQueen 1998;
Raymond, Dautricourt, and Hume to appear; Saffran, Aslin, and Newport 1996;
Saffran, Newport, and Aslin 1996; Vitevitch and Luce 1999; Dell et al. 2000;
Makashay 2001). In language acquisition, for example, experience with the ambi-
ent language shapes an infant’s expectations regarding language, revealed as a
preference for aspects of his/her native language (Aslin et al. 1981; Best et al.
1988; Jusczyk 1997; Polka and Werker 1997; Werker et al. 1981; Werker and
Tees 1984). Psycholinguistic research in speech and word processing also shows
that the ability to process speech is facilitated by a listener’s familiarity with vari-
ous dimensions of the native language’s phonological system including: the lan-
guage’s sounds (Pitt and Samuel 1990), phonotactics (Hallé et al. 1998, Massaro
and Cohen 1983, Pitt 1998, Pitt and McQueen 1998), patterns of contrast (Lahiri
and Marslen-Wilson 1991, Otake et al. 1996, Dupoux et al. 1997, Harnsberger
2001, Hume and Johnson 2003), and syllable structure (Cutler and Norris 1988,
Pallier et al. 1993, Pitt et al. 1998, Treiman and Danis 1988). For example, listen-
ers are biased to parse consonant clusters that are phonotactically impermissible
into permissible sequences (Hallé et al. 1998, Massaro and Cohen 1983, Pitt
1998). Pitt (1998) found that an epenthetic schwa was more likely to be per-
ceived between the consonants of phonotactically illegal consonant clusters

185
 Elizabeth Hume

(e.g., [tl3] o [t‹l3]) than legal clusters (e.g., [tr3] o [t‹r3]). It is also uncontro-
versial that how words are processed is influenced by factors such as their fre-
quency of occurrence and the predictability of sequences of sounds in the word
(e.g., Savin 1963, Luce 1986, Luce and Pisoni 1998, Pitt and McQueen 1998,
Vitevitch and Luce 1999, Frisch et al. 2000). For example, the higher the fre-
quency of the word, the greater the likelihood is that the listener will identify the
word correctly (Luce 1986). As I discuss below, experience also plays a role in
synchronic phonological patterns.
Language experience and predictability are useful concepts when it comes to
understanding markedness since, as it turns out, most of the traditional diagnostics
fit into one of two categories: (a) they provide evidence for the result of predict-
ability, or (b) they contribute to one’s experience with a particular element and, as
a result, influence what is predictable within the system of an individual speaker/
hearer or the language more generally. I illustrate these two points just below.

4. Diagnostics that Feed Experience

Consider first traditional markedness diagnostics that contribute to one’s experi-
ence with a linguistic element. As shown in Figure 1 these include distribution,
neutralization, articulatory simplicity, perceptual salience, and structural simplic-
ity. Each of these diagnostics provides evidence of a speaker/hearer’s experience
with some element of his/her language, and thus the degree of predictability of the
element in question.

distribution

neutralization

articulatory
simplicity EXPERIENCE PREDICTABILITY

perceptual
salience

structural
simplicity

Figure 1. Traditional markedness diagnostics that feed experience

with a particular linguistic element

186
 Markedness: A Predictability-Based Approach

Distribution has long been considered a diagnostic of markedness: the element

with the wider distribution is deemed unmarked. Linking distribution to an ele-
ment’s predictability is straightforward: given two elements x and y, if x has a
wider distribution than y, chances are that the language user will have more ex-
perience with x than with y. Thus, the probability of x occurring is greater than the
probability of y occurring. This is nicely illustrated in Mielke’s (2003) cross-
linguistic study of the perception of /h/ by native listeners of Turkish, Arabic,
English, and French, calculated as a function of the sensitivity measure d'. A
higher d' value indicates that the listener is more sensitive to the presence of /h/;
that is, /h/ is more easily identified. All subjects listened to identical stimuli pro-
duced by a male Turkish speaker. The results, shown in Figure 2, reveal that
Turkish and Arabic listeners have a high degree of sensitivity to /h/, and are able
to detect it in most contexts. English and French listeners, on the other hand, had
much lower sensitivity to /h/. Interestingly, these findings correctly reflect the lis-
tener’s experience with the sound. In Turkish and Arabic, /h/ is widespread, oc-
curring in most contexts. In English, on the other hand, /h/ has a restricted distri-
bution, occurring only in prevocalic position. /h/ does not occur in the French sys-
tem at all. In sum, listeners with the greatest experience with /h/ are best able to
detect the sound’s presence. The observation that there is no significant difference
between Turkish and Arabic, and between English and French, may reflect ceiling
and floor effects, respectively.

3.500
3.000
2.500
sensitivity (d')

T urkis h
2.000
A ra bic
1.500
E nglis h
1.000
F re nc h
0.500
0.000
-0.500
ds

es

ry
s
ps
ps

ic
s

es
s

es

l
sa
te
te

ui

id

da
al
sto
sto

tiv
iv

ca
a

na

liq

gl

oc

un
ic

at

ca
fri
ed
ss

ffr

ric

rv

bo
fri
af
le

ic

te
sa

sf

d
e

vo

in
ed

ed
ic

or
es

es
vo

ic

ic

w
el

el

vo

vo
ic

ic

follow ing c onte xt (X )

vo

vo

Figure 2. Sensitivity (dƍ) to [h] before another sound or word

boundary (VhX) (Mielke 2001, 2003)

187
 Elizabeth Hume

Since first proposed by Trubetzkoy (1939), neutralization has also been

widely acknowledged as a markedness diagnostic. Neutralization involves the loss
of contrast among elements in some context. For example, in languages with final
devoicing such as German, Polish, and Russian, the contrast between voiced and
voiceless consonants is arguably neutralized in word-final or coda position; only
one member of the opposition survives and this member is considered unmarked.
Like distribution, neutralization influences predictability. If only one member, x,
of an opposition, x-y, occurs in a particular context, the language user only has
experience with x in that context. It is therefore this member that is most probable
and thus, predicted.
Phonetic criteria such as articulatory simplicity and perceptual salience are
also equated with unmarkedness. As noted further above, high perceptual salience
has been proposed as an explanation for the pervasive presence of, e.g., CV sylla-
bles in languages, or clusters with [s] as opposed to other fricatives. The reason is
because sound sequences with richer cues tend to be more stable in a language
and consequently, typically occur in more words than those with weaker cues
(Makashay 2001). Thus, all else being equal, speakers/hearers will generally have
more experience with sounds/sequences with higher salience. As a result, high
salience can also provide evidence for the predictability of an element.
A similar conclusion can be drawn with respect to articulatory simplicity:
consonants produced with less complex articulations tend to occur more fre-
quently in a language than those with more complex articulations. As a result,
speakers/hearers will have more experience with these sounds and their predict-
ability will be higher. Of course, articulatorily complex segments do occur in lan-
guages and can be frequent. For example, the [&V] sequence has a high token fre-
quency in English, occurring word-initially in many common function words, e.g.
though, this, the, that. In this context, therefore, the speaker/hearer has consider-
able experience with the sequence and it is accordingly highly predictable.
This is precisely what we expect given that language experience, and hence
predictability, is dependent upon two factors: (a) the elements that are in a lan-
guage system’s inventory, and (b) the extent to which these elements are used.
Thus, even articulatorily complex sounds or sequences with low salience can be
predictable within a system provided that they are used a great deal. The reason
why articulatorily simple sounds and perceptually salient sequences are typically
unmarked is then simply because, due to their inherent phonetic nature, they tend
to occur more frequently in systems and, as a result, are used more.
The same point can be made with respect to structural complexity. Language
users are generally more familiar with simpler structures since they tend to be
used more. Polish syllable structure, for example, contains word-initial onsets of
up to four consonants. An examination of their frequency reveals that two-
consonant clusters occur in 88% of words, three-consonant clusters occur in 10%
of words, and four-consonant clusters are limited to only 2% of words (Bethin

188
 Markedness: A Predictability-Based Approach

1992). Two-consonant clusters are thus the most predictable clusters in the lan-
guage.
To summarize, the traditional markedness diagnostics noted above provide
evidence for a speaker/hearer’s experience with some element of his/her lan-
guage. As we have seen, some elements are more apt to occur in a language sys-
tem than others due to their inherent phonetic nature. However, presence within a
system is not sufficient to determine the predictability (unmarkedness) of an ele-
ment; we must also take into account the extent to which the elements are used.

5. Predictability Effects
In this section I consider another group of well-known markedness diagnostics
and show how they follow from the proposed model. They differ from the set
above in that they provide evidence for the effects of predictability. These diag-
nostics can be further divided into two sets, as shown in Figure 3. One illustrates
the instability of the predictable and includes diagnostics such as reduction, dele-
tion, and assimilation. The other set exemplifies the bias towards the predictable
as evidenced in processes like epenthesis, metathesis, and dissimilation.

INSTABILITY OF THE PREDICTABLE BIAS TOWARDS THE PREDICTABLE

e.g., undergoer of reduction, e.g., output of metathesis,
deletion, assimilation epenthesis, dissimilation

PREDICTABILITY

Figure 3. Observed effects of predictability

5.1. Instability
Instability of the predictable incorporates the observation that elements that are
predictable within a system are more likely to undergo change, such as reduction,
deletion, and assimilation. The reason for this is because predictability reduces an
element’s surprise value in the system and hence, the amount of information that
it contains (Zipf 1932, Shannon 1949, Applebaum 1996). As a result, the greater
the predictability of an element, the less information content it has and, I argue,
the more expendable it is. The most expected, i.e. unmarked, category is thus the
one with the least information content. The view of information content as a quan-
tifiable alternative to markedness is developed more fully in Hume (2004b,c). I

189
 Elizabeth Hume

offer some illustrations below but refer the reader to the works just noted for more
detailed discussion.
It is widely assumed that the phonologically unmarked segment in a system is
the segment which is least stable phonetically. That is, it is most likely to undergo
processes such as reduction, deletion, assimilation, etc. According to this view,
the alveolar coronal stop can be considered the unmarked consonant type in Eng-
lish. Support for this proposal comes, for example, from data from the Buckeye
Speech Corpus of over 100,000 words of conversational speech (Pitt et al. 2005),
which shows that 17% of all word variants involve t/d changes, the most of any
consonant type. Similarly, in Sri Lankan Portuguese Creole, labial m can be con-
sidered unmarked since it is the least stable nasal in the language (nasals include
labial, coronal, dorsal) (Hume and Tserdanelis 2002). The word-final labial nasal
optionally deletes; dorsal and coronal nasals do not. The labial nasal also under-
goes place assimilation to a following consonant; the coronal nasal does not.
The observed instability of t/d in English and m in Sri Lankan Portuguese
Creole finds a straightforward explanation when we take into account their pre-
dictability within each language. In Sri Lankan Portuguese Creole, users have
more experience with m than with other nasals. The labial nasal is twice as fre-
quent as its nearest competitor n (Hume and Tserdanelis 2002). Interestingly, the
velar nasal occurs in only a few words in word-final position and like the highly
frequent labial, it undergoes assimilation. The fact that one of the words in which
it occurs is the highly frequent indefinite article is perhaps not surprising. I sus-
pect that the high token frequency of the velar nasal led to its instability and thus
resulted in its patterning with the labial nasal rather than with the coronal. With
respect to English, it is clear that users have more experience with t/d than with
any other consonant in the language. In the Buckeye corpus, 40% of all tran-
scribed words have t or d, the highest by far of any consonant. Further, some very
frequent words contain t/d, including the three most frequent words in the corpus:
and, to, that. In addition, it is well-established that higher rates of t/d reduction in
English correlate with factors such as higher lexical frequency (Bybee 2001, Ju-
rafsky et al. 2001, Patterson and Connine 2001) and predictability from the fol-
lowing word (Raymond, Dautricourt, and Hume to appear).

5.2. Bias
A speaker/hearer is also biased towards the predictable. Since predictability is a
function of experience, this means that patterns that are more familiar to the lan-
guage user will have a greater chance of being produced and perceived. A listener
is especially biased towards the more frequent pattern in a system when informa-
tion specifying a sound or sound sequence is indeterminate (Pitt and McQueen
1998).
In metathesis, for example, Hume (2004a) shows that the knowledge of the
sound patterns of one’s language influences how the speech signal is processed
and thus, the order in which a sequence of sounds is parsed. To be specific, the

190
 Markedness: A Predictability-Based Approach

order inferred from the signal, and thus the output of metathesis, is the order that
occurs most frequently in the language. In this view, the reason that improved
perceptual salience is characteristic of many cases of metathesis becomes an arti-
fact of the nature of sequences subject to the process and of those that influence
speech processing. Sequences with poorer cues are more likely to undergo me-
tathesis, while those with good cues tend to be more frequent in a system and thus
have a greater impact on how the speech signal is parsed (Hume 2004a:227).
With respect to epenthesis, recall from the discussion of Pitt (1998) above that
a vowel is often perceived between the consonants of phonotactically illegal con-
sonant clusters in English (e.g., [tl3] is perceived as [t‹l3]). It should not be sur-
prising that the vowel in question is schwa, the most frequent vowel in the lan-
guage.
This approach to epenthesis departs from the common view that the epenthetic
segment is the consonant or vowel with the weakest phonetic cues, i.e., the least
salient segment. Support for the low-salience approach comes from the observa-
tion that in many languages schwa or [i] is the epenthetic vowel, and the coronal
stop [t] or glottal stop is the epenthetic consonant. In the approach taken here,
however, I argue that while the phonetic nature of a sound is an important con-
tributing factor (along with functional load, social prestige, and experience), it is
more generally a segment’s predictability in a given context that is crucial in de-
termining whether it will be perceived as the epenthetic vowel.
Epenthesis in French illustrates the mismatch between epenthesis and pho-
netic salience and thus provides support for the predictability approach. While the
epenthetic vowel in French is commonly referred to as schwa it is in fact the mid
front rounded vowel, [1] (Adda-Decker et al. 1999). This is of particular interest
since it is commonly assumed that roundness is marked in front vowels (e.g.,
Chomsky and Halle 1968). Yet, it is the front rounded vowel [1] that is the epen-
thetic vowel rather than the arguably simpler and less salient ['], also a sound in
the French inventory. Not surprisingly, [1] is also commonly deleted, e.g.,
[R1VK?~ [RVK?petit ‘small’.
The explanation for [1]’s behavior is not that it is the least phonetically salient
vowel in the language, but that it is predictable, a consequence (in part) of its high
frequency. For example, it occurs in many highly frequent function words, e.g., je
‘I’, le ‘the (masc.)’, me/te/se ‘1st/2nd/3rd pers. pronoun’. Further evidence comes
from corpus-based studies (Adda-Decker et al. 1999). The BREF corpus of read
speech contains 66,500 sentences from 120 speakers. In the corresponding word
lexicon, 37% of the words contain optional schwas, as in petit ‘small’. The count
of all schwas would actually be higher since this does not include those schwas
that are not in the appropriate phonological context for deletion, e.g., vendredi
‘Friday’. Similarly, in a 38,000 word subset of the MASK corpus of spontaneous
speech (409 speakers), of the 2,000 entries in the word list 35% have schwa.
Thus, both token and type frequency point to [1] as a highly frequent vowel in the
language.

191
 Elizabeth Hume

The bias towards the predictable can also be observed in patterns of language
acquisition. For example, Quiché-learning children master /V5/ at an earlier age
than English-learning children. The explanation for this difference can be related
to the observation that the sound occurs in many words that children are exposed
to in Quiché whereas it is relatively infrequent, compared to the other plosives, in
English (Pye et al. 1987). That is, Quiché-learning infants have more experience
with the sound than do English-learning children.
Substitution errors provide further evidence. Results from a study of Japanese-
learning children 2-5 years old show that they made more than twice as many
“backing” errors for /t/ (i.e., /t/ pronounced as /k/) as they made “fronting” errors
for /k/ (/k/ pronounced as /t/) (Yoneyama et al. 2003). This runs counter to the
claim that back consonants like /k/ are universally marked and likely to be re-
placed by front consonants like /t/. However, the patterns are straightforwardly
predicted by the observation that /k/ occurs more frequently than /t/ in Japanese
(adult lexicon and words that a Japanese child is most apt to hear) (Yoneyama et
al. 2003; Beckman et al. 2003). As Beckman et al. point out, the higher frequency
of /k/ in Japanese also correctly predicts the earlier acquisition of /k/ reported for
Japanese and the larger number of errors for /t/. In English, on the other hand, /t/
occurs more frequently than /k/.
Patterns of language development in creole genesis also support the predict-
ability model. Thomason (1993) shows that the properties of pidgins/creoles de-
pend on the properties of the source languages. Those that are common to both
varieties, i.e., familiar to the users, are more likely to be preserved, regardless of
how phonetically complex they might be. Chinook Jargon (a pidgin), for example,
contains a stable phonemic inventory which includes glottalized, labialized, and
uvular (vs. velar) stops.

6. Problems with Markedness Revisited and Explained

At the beginning of this paper I noted a number of problems with the traditional
approach to markedness. In this section, I return to these problems and show how
the proposed model accounts for the patterns in a straightforward manner.
One problem concerned the observation that given standard assumptions,
markedness only predicts the patterns that are supposed to be universal; it does
not provide predictions regarding language-specific markedness patterns. Yet,
many of the criteria assumed to provide evidence for markedness have been
brought into question precisely because patterns contradicting markedness are
commonly observed at the level of the individual language. In the present ac-
count, both language-specific and language-universal patterns are expected.
Briefly put, language-universal patterns result from the phonetic factors due to the
shared physiological make-up of humans. To the extent that languages differ in
terms of the elements that make up their systems and how they are used, expecta-
tions of language users and predictability of the elements will differ. We thus cor-
rectly predict cross-linguistic variability in observed language patterns.

192
 Markedness: A Predictability-Based Approach

A further problem with the traditional approach to markedness has to do with

apparent inconsistencies and contradictions among the diagnostics. Recall that an
unmarked sound can have either low or high salience. Also, an unmarked sound is
not only more apt to delete, it is also more apt to be inserted.
The solution to the problem lies in understanding that the diagnostics are actu-
ally providing evidence for the predictability of an element. Recall that some tra-
ditional diagnostics feed experience and thus increase an element’s predictability,
while others provide evidence for the effects of predictability. Thus, an element
with good cues (e.g., CV syllable) is more likely to be predictable than one with
poor cues simply because the former tends to occur more frequently in language
systems. All else being equal (e.g., quality of phonetic cues), a more predictable
element is also more apt to delete than a less predictable one (instability). Simi-
larly, a predictable sound is more likely to be epenthesized than a less predictable
one (bias). Within the current approach, there is no contradiction; all observations
fall out of the single metric, predictability.

7. Conclusion
In this paper I have argued that predictability is at the basis of markedness. Un-
marked elements are those that have a high degree of predictability within a sys-
tem (or a given context). While I have focused on only one of the factors crucial
to determining the predictability of a linguistic element, the promise of this ap-
proach to explaining markedness patterns should be evident.
First, a single metric, predictability, provides an explanation for both the in-
stability of an element and the speaker/hearer’s bias towards that same element.
In both cases, the element in question has high predictability. Markedness, on the
other hand, is a descriptive label; it doesn’t explain anything.
Second, the predictability of an element is dependent upon both universal
properties and the language system in which the element occurs. Thus, both uni-
versal and language-specific patterns are predicted. Given standard assumptions,
markedness only predicts the patterns that are supposed to be universal; it has
nothing to say about language-specific markedness patterns.
Third, predictability is quantifiable, thus moving us closer to a scientifically
rigorous theory of the observations (see Hume 2004b,c). Markedness is not.

References

Adda-Decker, Martine, Philippe Boula de Mareuil, and Lori Lamel. 1999.

Pronunciation variants in French: Schwa and liaison. Proceedings of ICPhS,
San Francisco. 2239-2242.
Applebaum, David. 1996. Probability and Information: An Integrated Approach.
Cambridge: Cambridge University Press.

193
 Elizabeth Hume

Aslin, Richard, David Pisoni, Beth Hennessy, and Alan Perey. 1981. Discrimina-
tion of voice onset time by human infants: New findings and implications for
the effects of early experience. Child Development 52:1135-1145.
Battistella, Edwin. 1990. Markedness: The Evaluative Superstructure of Lan-
guage. Albany: State University of New York Press.
Beckman, Mary, Kiyoko Yoneyama, and Jan Edwards. 2003. Language-specific
and language-universal aspects of lingual obstruent productions in Japanese-
acquiring children. Journal of the Phonetic Society of Japan 7(2):18-28.
Best, Catherine T., Gerald McRoberts, and Nomathemba Sithole. 1988. Examina-
tion of perceptual reorganization for nonnative speech contrasts: Zulu click
discrimination by English-speaking adults and infants. Journal of Experimen-
tal Psychology: Human Perception and Performance 14:345-360.
Bethin, Christina. 1992. Polish Syllables: The Role of Prosody in Phonology and
Morphology. Columbus, OH: Slavica Publishers.
Bybee, Joan. 1985. Morphology: A Study of the Relation between Meaning and
Form. Philadelphia: Benjamins
Bybee, Joan. 2001. Phonology and Language Use. Cambridge: Cambridge Uni-
versity Press.
Calabrese, Andrea. 1995. A constraint-based theory of phonological markedness
and simplification processes. Linguistic Inquiry 26:373-463.
Chomsky, Noam, and Morris Halle. 1968. The Sound Pattern of English. New
York: Harper & Row.
Clements, G. N. 1988. Toward a substantive theory of feature specification. Pro-
ceedings of NELS.
Clements, G. N. 1993. Underspecification or nonspecification. Proceedings of
ESCOL. 58-80.
Cutler, Anne, and Dennis Norris. 1988. The role of strong syllables in segmenta-
tion for lexical access. Journal of Experimental Psychology: Human Percep-
tion and Performance 14:113-121.
de Lacy, Paul. 2002. The formal expression of markedness. Ph.D. diss., Univer-
sity of Massachusetts, Amherst.
Dell, Gary S., Kristopher D. Reed, David R. Adams, and Antje S. Meyer. 2000.
Speech errors, phonotactic constraints, and implicit learning: A study of the
role of experience in language production. Journal of Experimental Psychol-
ogy: Learning, Memory and Cognition 26(6):1355-1367.
Dupoux, Emmanuel, Christophe Pallier, Nu"ria Sebastian, and Jacques Mehler.
1997. A destressing ‘deafness’ in French? Journal of Memory and Language
36:406-421.
Feather, Norman T. 1982. Expectations and Actions: Expectancy-Value Models in
Psychology. Hillsdale, NJ: LEA.
Frisch, Stefan. 1996. Similarity and frequency in phonology. Ph.D. diss., North-
western University.

194
 Markedness: A Predictability-Based Approach

Frisch, Stefan, Nathan Large, and David Pisoni. 2000. Perception of wordlike-
ness: Effects of segment probability and length on the processing of non-
words. Journal of Memory and Language 42:481-496.
Greenberg, Joseph. 1966. Language Universals. The Hague: Mouton.
Halle", Pierre A., Juan Segui, Uli Frauenfelder, and Christine Meunier. 1998.
Processing of illegal consonant clusters: A case of perceptual assimilation?
Journal of Experimental Psychology: Human Perception and Performance
24(2):592-608.
Harnsberger, James. 2001. The perception of Malayalam nasal consonants by
Marathi, Punjabi, Tamil, Oriya, Bengali, and American English listeners: A
multidimensional scaling analysis. Journal of Phonetics 29:303-327.
Hume, Elizabeth. 2003. Language specific markedness: The case of place of ar-
ticulation. Studies in Phonetics, Phonology and Morphology. 295-310.
Hume, Elizabeth. 2004a. The indeterminacy/attestation model of metathesis. Lan-
guage 80(2):203-237.
Hume, Elizabeth. 2004b. Language specific and universal markedness: A unified
information-theoretic approach. Ms. in preparation.
Hume, Elizabeth. 2004c. Predicting sound change: An information-theoretic ap-
proach. Presentation given at LabPhon9, University of Illinois, Urbana-
Champaign. (Handout of .ppt slides available at www.ling.osu.edu/~ehume)
Hume, Elizabeth, and Keith Johnson. 2003. The impact of impartial phonological
contrast on speech perception. Proceedings of the 15th International Congress
of Phonetic Sciences. 2385-2388.
Hume, Elizabeth, and Georgios Tserdanelis. 2002. Labial unmarkedness in Sri
Lankan Portuguese Creole. Phonology 19(1):441-458.
Jakobson, Roman. 1949. The sound laws of child language. In R. Jakobson, Stud-
ies on Child Language and Aphasia. The Hague: Mouton.
Jakobson, Roman. 1963. Implications of language universals for linguistics. In
Roman Jakobson: Selected Writings II, 580-592. The Hague: Mouton.
Jakobson, Roman (with Krystyna Pomorska). 1990. The concept of mark. In L.
Waugh and M. Monville-Burston, eds., On Language: Roman Jakobson, 134-
140. Cambridge, MA: Harvard University Press.
Jurafsky, D., A. Bell, M. Gregory, and W. Raymond. 2001. Probabilistic relations
between words: Evidence from reduction in lexical production. In J. Bybee
and P. Hopper, eds., Frequency and the Emergence of Linguistic Structure,
229-254. Philadelphia: John Benjamins.
Jusczyk, Peter. 1997. The Discovery of Spoken Language. Cambridge MA: MIT
Press.
Kean, M.-L. 1975. The theory of markedness in generative grammar. Ph.D. diss.,
MIT.
Kiparsky, Paul. 1985. Some consequences of Lexical Phonology. Phonology
Yearbook 2:85-138.

195
 Elizabeth Hume

Ladefoged, Peter. 1984. ‘Out of chaos comes order’; Physical, biological, and
structural patterns in phonetics. In M. P. R. Van den Broecke and A. Cohen,
eds., Proceedings of the Tenth International Congress of Phonetic Sciences,
83-95. Dordrecht: Foris Publications.
Lahiri, Aditi, and William Marslen-Wilson. 1991. The mental representation of
lexical form: A phonological approach to the recognition lexicon. Cognition
38:245-294.
Lass, Roger. 1975. How intrinsic is content? Markedness, sound change, and
‘family universals’. In D. Goyvaerts and G. Pullum, eds., Essays on The
Sound Pattern of English. Ghent: E. Story-Scientia.
Lindblom, Björn. 1990. Explaining phonetic variation: A sketch of the H and H
theory. In W. Hardcastle and A. Marchal, eds., Speech Production and Speech
Modeling, 403-439. Dordrecht and Boston: Kluwer.
Luce, Paul. 1986. Neighborhoods of words in the mental lexicon. Research on
speech perception technical report no. 6. Bloomington, IN: Indiana Univer-
sity.
Luce, Paul, and David Pisoni. 1998. Recognizing spoken words: The neighbor-
hood activation model. Ear and Hearing 19:1-36.
Makashay, Matthew. 2001. Lexical effects in the perception of obstruent ordering.
In E. Hume and K. Johnson, eds., Studies on the Interplay of Speech Percep-
tion and Phonology. Ohio State University Working Papers in Linguistics 55:
88-116.
Massaro, Dominic, and Michael Cohen. 1983. Phonological constraints in speech
perception. Perception and Psychophysics 34:338-348.
Menn, Lise. 1983. Language acquisition, aphasia and phonotactic universals. In F.
Eckman, E. Moravcsik, and J. Wirth, eds., Markedness, 241-256. New York
and London: Plenum Press.
Mielke, Jeff. 2001. Explaining directional asymmetry in Turkish [h] deletion: A
crosslinguistic study of perceptibility. Ohio State Working Papers in Linguis-
tics 55:117-171.
Mielke, Jeff. 2003. The interplay of speech perception and phonology: Experi-
mental evidence from Turkish. Phonetica 60(3):208-229.
Mohanan, K.P. 1993. Fields of attraction in phonology. In Goldsmith, J. (ed.),
The Last Phonological Rule. Chicago: University of Chicago Press. 61-117.
Otake, Takashi, Kiyoko Yoneyama, Anne Cutler, and Arie van der Lugt. 1996.
The representation of Japanese moraic nasals. Journal of the Acoustical Soci-
ety of America 100(6):3831-3842.
Pallier, Christophe, Nu"ria Sebastian-Galles, Teodora Felguera, Anne Christophe,
and Jacques Mehler. 1993. Attentional allocation within the syllabic structure
of spoken words. Journal of Memory and Language 32:373-389.
Patterson, David, and Cynthia M. Connine. 2001. Variant frequency in flap pro-
duction: A corpus analysis of variant frequency in flap production. Phonetica
58:254-275.

196
 Markedness: A Predictability-Based Approach

Pierrehumbert, Janet. 1994. Knowledge of variation. Chicago Linguistic Society,

vol. 2, Parasession on Variation in Linguistic Theory, 232-256.
Pitt, Mark. 1998. Phonological processes and the perception of phonotactically
illegal consonant clusters. Perception and Psychophysics 60:941-951.
Pitt, Mark, and Arthur Samuel. 1990. Attentional allocation during speech percep-
tion: How fine is the focus? Journal of Memory and Language 29:611-632.
Pitt, Mark, and James McQueen. 1998. Is compensation for coarticulation medi-
ated by the lexicon? Journal of Memory and Language 39:347-370.
Pitt, Mark, Katherine Smith, and James Klein. 1998. Syllabic effects in auditory
word recognition: Evidence from the structural induction paradigm. Journal of
Experimental Psychology: Human Perception and Performance 24:1596-
1611.
Pitt, Mark, Keith Johnson, Elizabeth Hume, Scott Keisling, and William Ray-
mond. 2005. The Buckeye Corpus of Conversational Speech: Labeling con-
ventions and a test of transcriber reliability. Speech Communication 45:89-95.
Polka, Linda, and Janet Werker. 1997. Adult and infant perception of two English
phones. Journal of the Acoustical Society of America 102:3742-3753.
Pye, C., D. Ingram, and H. List. 1987. A comparison of initial consonant acquisi-
tion in English and Quiche". In K. E. Nelson and A. Van Kleeck, eds., Chil-
dren’s Language 6, 175-190. Hillsdale, NJ: Lawrence Erlbaum.
Raymond, William, Robin Dautricourt, and Elizabeth Hume. To appear. Medial
t/d deletion in spontaneous speech: The effects of extra-linguistic, lexical and
phonological factors. Language Variation and Change.
Rice, Keren. 1996. Default variability: The coronal-velar relationship. Natural
Language & Linguistic Theory 14(3):493-543.
Rice, Keren. 1999. Featural markedness in phonology: Variation. Glot Interna-
tional 4(7):3-6, 4(8):3-7.
Saffran, Jenny, Richard Aslin, and Elissa Newport. 1996. Statistical learning by 8-
month-old infants. Science 274:1926-1928.
Saffran, Jenny, Elissa Newport, and Richard Aslin. 1996. Word segmentation:
The role of distributional cues. Journal of Memory and Language 35:606-621.
Savin, Harris B. 1963. Word-frequency effect and errors in the perception of
speech. Journal of the Acoustical Society of America 35:200-206.
Shannon, Claude. 1949. The Mathematical Model of Communication. Urbana:
University of Illinois Press.
Steriade, Donca. 1995. Underspecification and markedness. In J. Goldsmith, ed.,
Handbook of Phonological Theory, 114-174. Malden, MA: Blackwell.
Thomason, Sarah G. 1993. On identifying the sources of creole structures. A dis-
cussion of Singler’s and Lefebvre’s papers. In S. Mufwene, ed., Africanisms
in Afro-American Language Varieties, 280-295. Athens and London: The
University of Georgia Press.
Treiman, Rebecca, and Danis, Catalina. 1988. Syllabification of intervocalic con-
sonants. Journal of Memory and Language 27:87-104.

197
 Elizabeth Hume

Trubetzkoy, Nicolai. 1939. Grundzüge der Phonologie. Göttingen: van der Hoeck
& Ruprecht. Translated 1969 by Christine Baltaxe as Principles of Phonology.
Berkeley and Los Angeles: University of California Press.
Vitevitch, Michael S., and Paul A. Luce. 1999. Probabilistic phonotactics and
neighborhood activation in spoken word recognition. Journal of Memory and
Language 40:374-408.
Werker, Janet, John Gilbert, Keith Humphrey, and Richard Tees. 1981. Develop-
mental aspects of cross-language speech perception. Child Development 52:
349-355.
Werker, Janet, and Richard Tees. 1984. Cross-language speech perception: Evi-
dence for perceptual reorganization during the first year of life. Infant Behav-
ior and Development 7:49-63.
Vaux, Bert. 2002. Consonant epenthesis, unnatural phonology, and the problem of
rules vs. constraints (and gestures). Handout of presentation given at Depart-
ment of Linguistics. September 2002.
Yoneyama, Kiyoko, Mary E. Beckman, and Jan Edwards. 2003. Phoneme fre-
quencies and acquisition of lingual stops in Japanese. Ms. under review.
Zipf, George Kingsley. 1932. Selected Studies of the Principle of Relative Fre-
quency in Language. Cambridge, MA: Harvard University Press.

Dept. of Linguistics
Ohio State University
222 Oxley Hall
Columbus, OH 43202

[email protected]

198