OPTIMALITY THEORY

ren kager

published by the press syndicate of the university of cambridge

The Pitt Building, Trumpington Street, Cambridge c b 2 1 r p, United Kingdom
cambridge university press
The Edinburgh Building, Cambridge, c b 2 2 r u, UK http://www.cup.cam.ac.uk
40 West 20th Street, New York, NY 10011 4211, USA http://www.cup.org
10 Stamford Road, Oakleigh, Melbourne 3166, Australia
Cambridge University Press 1999
This book is in copyright. Subject to statutory exception and to the provisions
of relevant collective licensing agreements, no reproduction of any part may
take place without the written permission of Cambridge University Press.
First published 1999
Printed in the United Kingdom at the University Press, Cambridge
1
Typeset in Times 92 /13pt

[g c ]

A catalogue record for this book is available from the British Library
Library of Congress cataloguing in publication data
Kager, Ren.
Optimality theory / Ren Kager.
p. cm. (Cambridge textbooks in linguistics)
Includes bibliographical references and indexes.
ISBN 0 521 58019 6 (hardback). ISBN 0 521 58980 0 (paperback)
1. Optimality theory (Linguistics) I. Title. II. Series.
P158.42.K35 1999
410.1 dc21 98 39103 CIP
ISBN 0 521 58019 6 hardback
ISBN 0 521 58980 0 paperback

CONTENTS

Preface

page xi

1 Conflicts in grammars
1.1 Introduction: goals of linguistic theory
1.2 Basic concepts of OT
1.3 Examples of constraint interaction
1.4 The architecture of an OT grammar
1.5 Interactions of markedness and faithfulness
1.6 Lexicon Optimization
1.7 A factorial typology of markedness and faithfulness
1.8 On defining segment inventories
1.9 Conclusion

1
1
3
14
18
27
32
34
43
47

2 The typology of structural changes

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

52
52
59
78

3 Syllable structure and economy

3.1 Introduction
3.2 The basic syllable typology
3.3 Epenthesis and the conflict of well-formedness and
faithfulness
3.4 Generalized Alignment
3.5 The quality of epenthetic segments
3.6 Coda conditions
3.7 Conclusion

91
91
92

83
86

98
117
124
130
139
vii

Contents

viii

4 Metrical structure and parallelism

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

142
142
143
148
161

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

194
194
201
216
223

6 Output-to-output correspondence
6.1 Introduction
6.2 Identity effects in truncation
6.3 Identity effects in stem-based affixation
6.4 The cycle versus base-identity
6.5 Output-to-output correspondence: conclusions

257
257
259
273
277
293

7 Learning OT grammars
7.1 Introduction
7.2 Learning constraint rankings
7.3 Learning the Pintupi grammar of stress
7.4 The learning algorithm: discussion
7.5 Learning alternations and input representations

296
296
297
300
321
324

8 Extensions to syntax
8.1 Introduction
8.2 OT and syntax
8.3 The structure of extended verbal projections
in English
8.4 Typological consequences
8.5 Conclusions

341
341
341

177
189

229
230
248

353
366
369

Contents
9 Residual issues
9.1 Introduction
9.2 Opacity
9.3 Absolute ungrammaticality
9.4 Free variation
9.5 Positional faithfulness
9.6 Underlying Representations versus allomorphy
9.7 Conclusion: future perspectives
References
Index of languages
Index of subjects
Index of constraints

372
372
372
400
404
407
413
420
425
445
447
451

ix

1
Conflicts in grammars

1.1 Introduction: goals of linguistic theory

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

Conflicts in grammars
nor on the notion of possible rule interaction. In the late 1970s and early 1980s,
considerable efforts were put into constraining both rule typology and interactions. The broad idea was to factor out universal properties of rules in the form
of conditions.1 While rules themselves may differ between languages, they must
always respect a fixed set of universal principles. Gradually more and more
properties were factored out of rules and attributed to universal conditions on
rules and representations. Developments came to their logical conclusion in
Principles-and-Parameters Theory (Chomsky 1981b, Hayes 1980), which has as
its central claim that grammars of individual languages are built on a central core
of fixed universal properties ( principles), plus a specification of a limited number of universal binary choices ( parameters). Examples of parameters are the side
of the head (left or right) in syntactic phrases, or the obligatoriness (yes/no)
of an onset in a syllable. At the same time, considerable interest developed in
representations, as a way of constraining rule application, mainly with respect to
locality (examples are trace theory in syntax, and underspecification theory in
phonology). Much attention was also devoted to constraining rule interactions,
resulting in sophisticated theories of the architecture of UG (the T-model) and
its components (e.g. Lexical Phonology, Kiparsky 1982b).
1.1.2 Markedness
What all these efforts to constrain rules and rule interactions share, either implicitly or explicitly, is the assumption that universal principles can only be universal
if they are actually inviolate in every language. This interpretation of universality leads to a sharp increase in the abstractness of both linguistic representations
and rule interactions. When some universal principle is violated in the output of
the grammar, then the characteristic way of explaining this was to set up an
intermediate level of representation at which it is actually satisfied. Each grammatical principle thus holds at a specific level of description, and may be switched
off at other levels.
This absolute interpretation of universality is not the only one possible, however. In structuralist linguistics (Hjelmslev 1935, Trubetzkoy 1939, Jakobson
1941; cf. Anderson 1985), but also in Generative Phonology (Chomsky and Halle
1968, Kean 1975, Kiparsky 1985) and Natural Phonology (Stampe 1972, Hooper
1976), a notion of markedness plays a key role, which embodies universality in
a soft sense. The idea is that all types of linguistic structure have two values,
one of which is marked, the other unmarked. Unmarked values are crosslinguistically preferred and basic in all grammars, while marked values are crosslinguistically avoided and used by grammars only to create contrast. For example,
1

For example, Subjacency was proposed as a universal condition on syntactic movement rules
and the Obligatory Contour Principle as a universal condition on phonological rules.

1.2 Basic concepts of OT

all languages have unrounded front vowels such as [i] and [e], but only a subset
of languages contrast these vowels with rounded front vowels such as [y] and [].
Hence, the unmarked value of the distinctive feature [round] is [round] in
front vowels. At a suprasegmental level, markedness affects prosodic categories.
For example, the unmarked value for syllable closure is open since all languages have open syllables (CV, V), while only a subset of languages allow closed
syllables (CVC, VC).2 The notion of markedness is not only relevant to sound
systems. Markedness principles have been proposed for morphological and
syntactic systems as well (Chomsky 1981a).
The markedness approach of linguistic universality is built on two assumptions. First, markedness is inherently a relative concept: that is, a marked linguistic element is not ill-formed per se, but only in comparison to other linguistic
elements. Second, what is marked and unmarked for some structural distinction is not an arbitrary formal choice, but rooted in the articulatory and perceptual
systems. By this combination of two factors, markedness allows an interpretation
of universality that is fundamentally different from Principles-and-Parameters
Theory, in which markedness has no substantive status in the grammar, but functions as an external system of annotations on parameter values, evaluating a
grammars complexity.3
1.2 Basic concepts of OT
Optimality Theory (Prince and Smolensky 1993, McCarthy and Prince
1993a,b) turns markedness statements into the actual substance of grammars.
Markedness is built into grammars in the form of universal output constraints
which directly state marked or unmarked patterns, for example: front vowels are
unrounded or syllables are open. The universal interpretation of markedness
constraints is reconciled with the observation that languages, to a certain extent
at least, tolerate marked types of structures. Universal markedness constraints can
be literally untrue for a grammars output, or to phrase it in optimality-theoretic
terms: constraints are violable. Violation of a constraint is not a direct cause of
ungrammaticality, nor is absolute satisfaction of all constraints essential to the
grammars outputs. Instead what determines the best output of a grammar is the
least costly violation of the constraints. Constraints are intrinsically in conflict,
hence every logically possible output of any grammar will necessarily violate at
least some constraint. Grammars must be able to regulate conflicts between universal constraints, in order to select the most harmonic or optimal output form.
2

Markedness may also involve scales. For example, the higher a consonants sonority value, the
more likely its occurrence in the syllable coda.
For the view of markedness as a criterion external to the grammar, evaluating its complexity,
see Chomsky and Halle (1968) and Kean (1975, 1981).

Conflicts in grammars
This conflict-regulating mechanism consists of a ranking of universal constraints. Languages basically differ in their ranking of constraints. Each violation
of a constraint is avoided; yet the violation of higher-ranked constraints is avoided
more forcefully than the violation of lower-ranked constraints. Accordingly, the
notion of grammatical well-formedness becomes a relative one, which is equivalent to the degree of satisfaction of the constraint hierarchy, or harmony.
OTs viewpoint of UG is fundamentally different from that of classical rulebased generative theory, where UG is defined as a set of inviolate principles and
rule schemata (or parameters). OT defines UG as a set of universal constraints
(markedness relations and other types of constraints, as we will see below), and
a basic alphabet of linguistic representational categories. In its interactions, it is
limited to a single device: constraint ranking. OT still shares with its rule-based
generative ancestors the central position taken by UG, as described above. OT is
a theory of the human language capacity.
The remainder of this chapter is organized as follows. Section 1.2 will introduce
basic notions of OT: conflict, constraints, and domination, which will be exemplified in section 1.3. In section 1.4, we will discuss the architecture of an OT
grammar. Section 1.5 will deal with interactions of markedness and faithfulness,
relating these to the lexicon in section 1.6. A factorial typology of constraint
interactions will be developed in section 1.7 and applied to segment inventories
in section 1.8. Finally, section 1.9 presents conclusions.
1.2.1 Language as a system of conflicting universal forces
At the heart of Optimality Theory lies the idea that language, and in fact every
grammar, is a system of conflicting forces. These forces are embodied by constraints, each of which makes a requirement about some aspect of grammatical
output forms. Constraints are typically conflicting, in the sense that to satisfy one
constraint implies the violation of another. Given the fact that no form can satisfy
all constraints simultaneously, there must be some mechanism selecting forms that
incur lesser constraint violations from others that incur more serious ones.
This selectional mechanism involves hierarchical ranking of constraints, such
that higher-ranked constraints have priority over lower-ranked ones. While constraints are universal, the rankings are not: differences in ranking are the source
of cross-linguistic variation.
But before discussing actual constraints and their rankings, let us first find out
in a general way about the two major forces embodied by constraints. Two forces
are engaged in a fundamental conflict in every grammar. The first is markedness,
which we use here as a general denominator for the grammatical factors that exert
pressure toward unmarked types of structure. This force is counterbalanced by
4

1.2 Basic concepts of OT

faithfulness, understood here as the combined grammatical factors preserving
lexical contrasts. Let us focus on both general forces to find out why they are
inherently conflicting.
In sound systems, certain types of structure segments, segment combinations,
or prosodic structures are universally favoured over others. For example, front
unrounded vowels are unmarked as compared to front rounded vowels, open
syllables as compared to closed syllables, short vowels as compared to long
vowels, and voiceless obstruents compared to voiced obstruents. As was observed
above, marked structures are avoided by all languages, while they are completely
banned by some languages. Therefore the notion of markedness is inherently
asymmetrical.
Most phonologists agree that phonological markedness is ultimately
grounded in factors outside of the grammatical system proper. In particular,
the systems of articulation and perception naturally impose limitations on which
sounds (or sound sequences) should be favoured. Yet explaining markedness relations by phonetic factors does not amount to denying the basis of phonology as
a grammatical system, for two reasons. The first reason is that phonetic factors
are gradient, and add up to numerical patterns, while phonological factors are
categorical, producing patterns whose boundaries are clearly cut by categorical
distinctions. The symmetry of phonological systems cannot be captured by the
interaction of raw phonetic factors. The second reason is that the relative
strength of the individual markedness factors varies from language to language,
which entails that there must be a language-specific system defining the balance
of factors. This is the grammar, a system of ranked constraints, of which phonology is an integral part.
The major force counterbalancing markedness is faithfulness to lexical contrasts.
A grammar that is maximally faithful to a lexical contrast is one in which output
forms are completely congruent with their lexical inputs with respect to some
featural opposition. Or to put it differently, the total amount of lexically contrastive variation of some feature is realized in all of the grammars output forms. For
example, a lexical contrast of voicing in obstruents is preserved in output forms
regardless of their phonological context (at the end of a word, between vowels,
etc.). Thus one may think of faithfulness as the general requirement for linguistic
forms to be realized as close as possible to their lexical basic forms. From a
functional angle, the importance of faithfulness is clear: to express contrasts of
meaning, any language needs a minimal amount of formal contrast. Formal contrasts should be preserved in realizations of lexical items, and not be eroded (or
at least, not too much) by factors reducing markedness. In the realm of sound
5

Conflicts in grammars
systems (or phonologies), lexical contrasts are carried by oppositions between
sounds, as well as by their combinations. Phonological elements are not the only
carriers of lexical contrast. (Although phonology is what we will focus on in this
book.) Lexical contrasts are also expressible by word structure (morphology) or
phrase structure (syntax).
Closely related to faithfulness (or preservation of lexical contrasts) is the pressure towards the shape invariability of lexically related items in various grammatical contexts. This was known in pre-generative linguistics as paradigm
uniformity. Shape invariance of lexical items is understandable as another priority of linguistic communication: there should be a one-to-one relation between
lexical items, the atoms of meaning, and the shapes which encode them.
1.2.2 Conflicts between markedness and faithfulness
Markedness and faithfulness are inherently conflicting. Whenever some lexical
contrast is being preserved, there will be some cost associated in terms of markedness since in every opposition one member is marked. For example, consider
the fact that English limits the possible contrasts in its vowels with respect to the
dimensions of backness and rounding: no rounded front vowels stand in contrast
to unrounded front vowels. This correlation of rounding and backness in vowels
is not idiosyncratic to English, but it reoccurs in a great majority of the worlds
languages. In fact it is grounded in properties of the articulatory and perceptual
systems. Yet this restriction is certainly not universal in the sense that all of
the worlds languages respect it. Many languages do allow a contrast of rounding
in front vowels, thus increasing the potential amount of lexical contrast at the
expense of an increase in markedness.
Generally we find that the larger the array of means of encoding lexical contrasts, the larger the complexity of the sound system, either in terms of segmental
complexity, or in terms of the combinatory possibilities between segments (phonotactics). A language can be maximally faithful to meaningful sound contrasts
only at the expense of an enormous increase in phonological markedness. Conversely, a language can decrease phonological markedness only at the expense of
giving up valuable means to express lexical contrast.
First consider what a hypothetical language would look like at one extreme of the
spectrum: a language giving maximal priority to the expression of lexical contrasts, while imposing no markedness restrictions. We endow this language with
the combined segment inventories of the worlds languages, roughly 50 consonants and 30 vowels (Ladefoged and Maddieson 1996). We drop combinatory
markedness restrictions, allowing all logically possible segment combinations to
form a lexical item. Permutation of these 80 segments into lexical items of two
6

1.2 Basic concepts of OT

segments already produces some 6,400 items, including [ph], [x], and [.],
all highly marked. But why stop at two segments per item? By sheer lack of
phonotactic limitations, nothing rules out lexical items of 37 or 4,657 segments,
or even longer. Now consider the fact that the number of possible lexical items
increases exponentially with the number of segments (80n) so that at segmental
length 6 we already approximate an awesome 300 billion potential lexical items.
Clearly no human language requires this number of lexical contrasts, hence there
is room to impose markedness restrictions on segments and their combinations in
lexical items. Since such restrictions make sense from an articulatory and perceptual point of view, we expect to find them.
Let us now turn the tables to find out what a language at the other extreme
would look like, a language giving maximal priority to markedness, and minimal
priority to the expression of lexical contrasts. Let us assume that this language
limits its lexical items to the general shape of CV* (sequences of consonant
vowel), with C {p,t,k} and V {,a}.4 The complete set of potential monosyllables contains 6 items {p, pa; t, ta; k, ka}, the set of disyllables contains 36 (or
62) items ({pp, pap, kp . . . }), trisyllables 216 (or 63), etc. But stop! We are
overlooking the fact that the unmarked length of lexical item is two syllables (this
is the minimum size in many languages and by far the most frequent size in most
languages). Since we are assuming that this language is maximally concerned
about markedness, we should limit word size to two syllables. The bitter consequence is a mini-lexicon containing at most 36 items. Now consider the fact that
the lexicon of an average natural language contains some 100,000 items.5 It is
clear that giving maximal priority to markedness implies an acute shortage of
lexical contrasts, which no language can afford.
This comparison of two extremes shows that languages may, in principle at
least, go astray in either of two ways: by giving blind priority to expression of
lexical contrast, resulting in massive costs in terms of markedness or, at the other
end of the spectrum, by giving unlimited priority to markedness reduction, resulting in a fatal lack of contrast.
4

These limitations are actually grounded in speech production and perception: every consonant
is maximally different from a vowel (hence, all consonants are voiceless stops). Every vowel
is maximally different from other vowels (a 2-vowel set, a). Every consonant is maximally
different from other consonants (place of articulation restricted to labial, alveolar, and velar).
Every vowel is preceded by a consonant (no word-initial vowels, no hiatus). Every consonant
precedes a vowel for optimal release (hence no consonant clusters nor word-final Cs).
Suppose that our hypothetical language would not respect word size restrictions, having at its
disposition all possible CV*-shaped items. Here, with a maximal density of lexical contrast,
all potential items up to seven syllables long would not suffice to build the required size of
lexicon. This would only reach to a moderate total of (46,656 + 7,776 + 1296 + 216 + 36 +
6) = 55,986 lexical items. The average item in this language would be over six syllables long.
Without doubt, speaking would become a rather time-consuming activity.

Conflicts in grammars
In sum, we have seen that every grammar must reconcile the inherently competing forces of faithfulness to lexical contrasts (the inertness which draws output
forms back to their basic lexical shapes) and markedness (minimization of
marked forms). However, as we are about to find out, Optimality Theory recognizes no unitary or monolithic forces of faithfulness or markedness: the picture
is more fragmented. In the grammars of individual languages, the overall conflict
between both forces assumes the form of finer-grained interactions of individual
constraints. At this level, where individual constraints compete, languages are
quite diverse in their resolutions of conflicts between markedness and faithfulness. A language may give priority to faithfulness over markedness with respect
to some opposition, but reverse its priorities for another opposition.
Let us now turn to the implementation of these basic ideas in Optimality
Theory.
1.2.3 The OT grammar as an inputoutput device
The basic assumption of OT is that each linguistic output form is optimal, in the
sense that it incurs the least serious violations of a set of conflicting constraints.
For a given input, the grammar generates and then evaluates an infinite set of
output candidates, from which it selects the optimal candidate, which is the actual
output. Evaluation takes place by a set of hierarchically ranked constraints (C1
C2 . . . Cn), each of which may eliminate some candidate outputs, until a point
is reached at which only one output candidate survives. This elimination process
is represented schematically:6
(1)

Mapping of input to output in OT grammar

C1

Input

Candidate a
Candidate b
Candidate c
Candidate d
Candidate . . .

C2 . . . Cn

Output

The optimal output candidate is the one that is most harmonic with respect to
the set of ranked constraints. Harmony is a kind of relative well-formedness,
taking into account the severity of the violations of individual constraints, as
determined by their hierarchical ranking. That is, violation of a higher-ranked
6

Elimination of less-harmonic candidates is portrayed in (1) as a serial filtering process, but we

will learn to view it as a parallel process, with higher-ranked constraints taking priority over
lower-ranked constraints.

1.2 Basic concepts of OT

constraint incurs a greater cost to harmony than violation of a lower-ranked constraint. Some violations must occur in every output candidate, as constraints
impose conflicting requirements. Accordingly, a lower-ranked constraint can be
violated to avoid the violation of a higher-ranked one, but violation is always kept
to a minimum, given the requirement of maximal harmony.
With the basic assumptions of OT in our minds, let us now turn to a finergrained discussion of the core notions constraints, conflict, domination, and
optimality.
1.2.4 Constraints: universality and violability
Our preliminary definition of constraint is: a structural requirement that may
be either satisfied or violated by an output form. A form satisfies a constraint if
it fully meets the structural requirement, while any form not meeting this requirement is said to violate it. For the moment we will assume no degrees of violation, so that output forms are simply categorized by a crude binary criterion as
either satisfying or violating a constraint. Forms may satisfy constraints vacuously, which is the case if a constraint makes a requirement about some structural
element that is not present in a particular candidate.
OT recognizes two types of constraints, faithfulness constraints and markedness
constraints. Each individual constraint evaluates one specific aspect of output
markedness or faithfulness. Let us now look into the general properties of both
types of constraints, and into their functions in the grammar.
Markedness constraints require that output forms meet some criterion of structural well-formedness. As the examples below illustrate, such requirements may
take the form of prohibitions of marked phonological structures, including segment types (2a), prosodic structures (2b), or occurrences of segment types in
specific positions (2c).
(2)

Examples of markedness constraints

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

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

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

Conflicts in grammars
As we have seen in section 1.1, markedness is an inherently asymmetrical
notion. Hence, the universal constraint inventory lacks the antagonist constraints
of (1ae), which make opposite requirements syllables must have codas, sonorants must be voiceless, etc.7
Faithfulness constraints require that outputs preserve the properties of their basic
(lexical) forms, requiring some kind of similarity between the output and its
input.
(3)

Examples of faithfulness constraints

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

Faithfulness constraints are, strictly speaking, not pure output constraints, since
they take into account elements at two levels: input and output. In contrast,
markedness constraints never take into account elements in the input.8 The
important thing is, however, that both kinds of constraints refer to the output
(exclusively so in markedness, and in relation to the input in faithfulness). OT has
no constraints that exclusively refer to the input. (This is a crucial difference from
classical generative phonology, as we will see in chapter 2.)
From a functional viewpoint, faithfulness constraints protect the lexical items
of a language against the eroding powers of markedness constraints, and thereby
serve two major communicative functions. First, they preserve lexical contrasts,
making it possible for languages to have sets of formally distinct lexical items to
express different meanings. Phrasing it slightly differently, with an emphasis on
contrast, we may say that faithfulness is what keeps the shapes of different lexical
items apart. Second, by limiting the distance between input and output, faithfulness constraints restrict the shape variability of lexical items. Faithfulness thus
keeps the contextual realizations of a single morpheme (called its alternants)
from drifting too far apart. This enhances the one-to-one relations of meaning and
form. In sum, the overall function of faithfulness is to enforce the phonological
shape of lexical forms in the output, as a sort of inertness limiting the distance
between outputs and their basic shapes.
Two more assumptions are to be made about constraints in OT: they are universal and violable requirements on some aspect of linguistic output forms. Let
us now focus on each of these properties of constraints. The first property is
7
8

We will see later that some markedness constraints do have antagonists.

See chapter 9 for OT models which weaken this assumption.

10

1.2 Basic concepts of OT

(4)

Universality: constraints are universal.

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

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

11

Conflicts in grammars
language-specific (since it lacks both phonetic grounding and cross-linguistic
motivation), this will be indicated.
However, the universality of constraints should directly be relativized somewhat. We will find that in special cases, language-specific elements may occur in
constraints of otherwise universal formats. This option is typical for a class of
constraints defining the interface of morphology and phonology, so-called alignment constraints, matching up the edges of specific morphemes and prosodic
categories. (See chapters 3 and 5.) Such interface constraints define schemata in
which individual languages may substitute their specific morphemes.
We now move on to the second major property of OT constraints: their softness,
or violability. Violability of constraints must be understood in a specific way: the
general requirement is that it must be minimal:
(5)

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

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

another higher-ranked constraint. And even if a constraint is violated, violation
must be kept to a minimum. Everything else being equal, forms with lesser
violations are more harmonic than forms with greater violations. (Exactly how
degree of violation is determined will be taken up in section 1.4.3.) Violability of
constraints is an essential property of OT, representing a radical break away from
derivational models, as well as from constraint-based theories, such as Declarative Phonology (Bird 1990, Scobbie 1991), which assume that constraints are hard
or inviolate. (For a broad comparison with derivational theory, see chapter 2.)
This discussion of violable constraints gives rise to an important new question,
to which we now turn: what is the optimal candidate?
1.2.5 Optimality: domination and conflict
As mentioned before, optimality is the status of being most harmonic with respect
to a set of conflicting constraints. It is now time to take a closer look at the
concept of optimal in OT. The general idea is that the grammar evaluates an
infinite set of candidate output forms, all analyses of a given input. From this
candidate set it selects the optimal output, the one which best matches the set
of conflicting constraints. But what precisely does it mean for an output to be
optimal? Does it involve some sort of compromise between constraints of different strengths? Or is it perhaps the case that weaker constraints are rendered
inactive when they come into conflict with stronger constraints?
In fact optimality involves neither compromise nor suppression of constraints,
but instead it is built on (strict) domination of constraints in a hierarchy.

12

1.2 Basic concepts of OT

(6)

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

violations of a set of constraints, taking into account their hierarchical ranking.

So we assume that each output form of the grammar is by definition the best
possible in terms of the hierarchy of constraints, rather than the form which
matches all constraints at the same time. Perfect output forms are principally
non-existent, as every output form will violate at least some constraints. Therefore the selection of the optimal output form involves setting priorities.
This is where a hierarchy comes into play. Conflicts are resolved by domination:
(7)

Domination: the higher-ranked of a pair of conflicting constraints

takes precedence over the lower-ranked one.

This tentative definition will be refined below in section 1.4, on the basis of more
complex cases.
The ranking of constraints can be demonstrated by a tableau: this lists two
(or any number of) output candidates vertically in random order, and constraints
horizontally, in a descending ranking from left to right. The cells contain violation
marks * incurred by each candidate for the constraint heading the column.
Schematically:
(8)

A tableau for simple domination

C1
a. + candidate a
b.

candidate b

C2
*

*!

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

13

Conflicts in grammars
1.3 Examples of constraint interaction
1.3.1 Neutralization of voicing contrast in Dutch
Among the universal and violable constraints is the following:
(9)

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

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

a. /bd/
bt
bed
b. /bd-n/ bdn beds

Dutch has no voicing contrast in final obstruents, neutralizing it towards voicelessness.

Next consider the evaluation of two candidate outputs for the input /bd/, [bt],
and [bd], with respect to *Voiced-Coda:
(11)

Evaluation of two candidates with respect to *Voiced-Coda

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

If this constraint were the only one relevant for these forms, then things would
be simple. Violators could be dismissed without second thoughts. But in actual
grammars things are not that simple since constraints may make conflicting
requirements about output forms.
A second constraint of the universal inventory is a typical faithfulness constraint, requiring that the input value of the feature [voice] be preserved in the
output.
(12)

10

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

Actually *Voiced-Coda can be interpreted as the conjunction of two markedness statements,

an idea to which we will return in chapter 9.

14

1.3 Examples of constraint interaction

This faithfulness constraint mentions a notion correspondent, which is tentatively (and very informally) defined as follows:
(13)

Correspondent: the output segment that is the realization of an

input segment.

This informal definition is precise enough for our present purposes. (We will
return to the important notion of correspondence, particularly in chapters 2 and
5.)
In a correspondence diagram of the Dutch word [bt] bed the input and output segments that are correspondents of one another are connected by vertical lines.
(14)

Correspondence diagram of [bt]

/b d/ Input
[b t] Output

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

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

Evaluation of two candidate outputs for the input /bd/

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

Observe the conflict: the evaluation of both output forms is different for each
constraint.
This conflict requires resolution, which is the task of the constraint hierarchy.
The form [bt] emerges as the optimal output of the grammar, given the following
fragment of the phonology of the language:
(16)

Coda devoicing in Dutch

*Voiced-Coda Ident-IO(voice)

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

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

Conflicts in grammars
well-formedness with respect to the ranking in (16). Or stated differently, [bt] is
more harmonic than [bd] with respect to the ranking in (16).
(17)

Harmonic ranking of two output candidates for the input /bd/ in

Dutch
[bt] [bed]

Since we are only considering two candidates here, the harmonic ranking directly
gives us the optimal output: [bt].
The correctness of this constraint ranking can be represented in a tableau-format:
(18)

Tableau for the input /bd/, assuming the Dutch ranking

Candidates:

*Voiced-Coda

a. + [bt]
b.

Ident-IO(voice)
*

[bd]

*!

The optimal candidate in the top row, [bt], incurs a violation of Ident-IO(voice)
while it satisfies *Voiced-Coda. Suboptimal [bd] has exactly the reverse pattern
of violations: it has a violation mark for *Voiced-Coda, but none for IdentIO(voice).
Being presented with these two output candidates, the grammar (whose only
goal is selecting an optimal output) must settle for a candidate that has a violation
of a lower-ranked constraint, simply because no perfect output candidate is available, satisfying both constraints. This point can be made more general: constraints
are intrinsically conflicting, hence perfect output candidates will never occur in
any tableau:
(19)

Fallacy of perfection: no output form is possible that satisfies all

constraints.

An output is optimal since there is no such thing as a perfect output: all that
grammars may accomplish is to select the most harmonic output, the one which
incurs the minimal violation of constraints, taking into account their ranking.
Nothing better is available.
Observe that the result of the constraint interaction in Dutch is a neutralization
of the voicing contrast in a specific context: the syllable coda. That neutralization
indeed takes place can be easily shown by the following set of examples:
(20)

16

a.i
a.ii
b.i
b.ii

/bd/
/bd-n/
/bt/
/bt-n/

bt
b.dn
bt
b.tn

bed
beds
(I) dab
(we) dab

1.3 Examples of constraint interaction

Neutralization of the sound shapes of two lexical items is the ultimate consequence of the domination of markedness over faithfulness. The lexical contrast
between /bd/ and /bt/, residing in the value of voicing of their final stem consonants, might (in principle at least) have been preserved in all morphological
contexts in which they occur. But this is not the case, and a complete neutralization occurs, into [bt].
1.3.2 Preservation of voicing contrast in English
In English, as opposed to Dutch, an analogous input /bd/ is mapped to an output
[bd], preserving the voicing in the final consonant. Accordingly, English contrasts words such as bed and bet. This is due to the following fragment of the
phonology of this language:
(21)

Preservation of voicing contrast in English

Ident-IO(voice) *Voiced-Coda

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

ranking of the one we established for Dutch. Accordingly, the harmonic ranking
of the output candidates under discussion is reversed, as compared to the one of
Dutch:
(22)

Harmonic ranking of two output candidates for the input /bd/ in

English
[bd] [bt]

That is, assuming an input /bd/, [bd] is more harmonic than [bt] with respect
to the ranking in (21).
Again, we illustrate this ranking with the help of a tableau, evaluating the same
candidates as we used in tableau (18) for Dutch. Observe that Ident-IO(voice)
and *Voiced-Coda have changed places:
(23)

Tableau for the input /bd/, assuming the English ranking

Candidates:
a.

[bt]

b. + [bd]

Ident-IO(voice)

*Voiced-Coda

*!
*

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

Conflicts in grammars
[bt]. (This contrast is actually reinforced by a subsidiary vowel length difference
between both words: [b]d] versus [bt].)
1.3.3 The relation between universal and language-specific
What we have just witnessed in the examples from Dutch and English is the
universal pan-grammatical conflict of markedness and faithfulness taking place
on a micro-scale. In both languages, the same conflict arises with respect to
preservation of a contrastive property (the feature [voice]), and its neutralization
in a specific context (syllable coda). However, the outcome of this conflict is
different for both languages. Dutch resolves it in the favour of markedness
whereas English favours faithfulness. This shows that universal constraints are
ranked in language-specific ways. OT clearly marks off the universal from the
language-specific. Both constraints and the general principles of their interaction
are universal, while constraint hierarchies are language-specific.
Speaking of forces of faithfulness and markedness is somewhat misleading,
since this suggests that conflicts between these forces are resolved on a superordinate level in the grammar of a single language. This is clearly not the case.
For example, the fact that Dutch ranks markedness above faithfulness with
respect to voice in coda obstruents does not imply that it selects the same ranking
(M F) with respect to voice in other contexts, nor that it selects this ranking
with respect to other features in the syllable coda. In Dutch, voice is contrastive
in obstruents in onsets (even though voiced obstruents are universally marked).
Also, place features are contrastive in obstruents and nasals in codas (in spite of
the markedness of labials and velars). This shows that there are no monolithic
forces of faithfulness and markedness, but that instead finer-grained interactions
occur between the context- and feature-specific versions of these classes of constraints. Still, for expository purposes, the classification of constraints into faithfulness and markedness constraints remains useful, as are shorthand notations
such as M F.
These remarks bring us back to our starting point in this section: the conception of universal grammar in OT. But what exactly do we mean by grammar in
the first place? The OT grammar, and its architecture, will be the topic of the next
section.
1.4 The architecture of an OT grammar
The OT grammar is an inputoutput mechanism that pairs an output form to an
input form (such that each input has precisely one output). To accomplish this
function, the grammar contains a division of labour between a component
which maps the input onto an infinite set of candidate output forms, and another
component that is burdened with evaluating the candidate output forms by a set
18

1.4 The architecture of an OT grammar

of ranked constraints, and selecting the optimal output among these. These two
components are known under the names of Generator (or Gen) and Evaluator
(or Eval ). This grammatical organization is schematically represented in a function notation as follows:
(24)

The grammar as an inputoutput mechanism

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

That is, Gen is a function that, when applied to some input, produces a set of
candidates, all of which are logically possible analyses of this input. Similarly,
Eval is a function that, when applied to a set of output candidates, produces an
output, the optimal analysis of the input. In addition to Gen and Eval, the grammar contains a lexicon storing all lexical forms that are input to Gen. Recapitulating, we find the following model of the grammar:
(25)

Components of the OT grammar

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

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

1.4.1 The LEXICON, and Richness of the Base
The Lexicon contains all contrastive properties of morphemes (roots, stems,
and affixes) of a language, including phonological, morphological, syntactic, and
semantic properties. The Lexicon provides the input specifications which are to
be submitted to the Generator. In this connection, perhaps the most striking property of the Lexicon, as conceived of in OT, is that no specific property can be
stated at the level of underlying representations:
(26)

Richness of the Base: no constraints hold at the level of underlying

forms.

In OT grammatical generalizations are expressed as interactions of constraints

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

Conflicts in grammars
representation, is attributed to interactions at the output level in OT. Whether or
not a feature is contrastive in some language depends on interactions of outputoriented markedness and faithfulness constraints, either preserving or overruling
input specifications (see section 1.5).
OT thus abandons Morpheme Structure Constraints (MSCs), which in classical
generative phonology (Chomsky and Halle 1968) account for prohibitions against
specific types of structure at the level of the morpheme, in specific languages.
MSCs were used, for example, to express prohibitions against front rounded vowels,
or sequences of three or more consonants, or two labial consonants occurring
within a morpheme. In the early 1970s MSCs were argued to be theoretically
problematic in the sense that they duplicate information which is, independently,
expressed by phonological rewrite rules, or that they globally guide the application of rules, a property called structure-preservingness.11 By locating the burden
of explanation of the lack of specific kinds of structure at the level of the output,
OT, in principle at least, circumvents this Duplication Problem.
1.4.2 The GENERATOR , and Freedom of Analysis
The essential property of the Generator is that it is free to generate any conceivable output candidate for some input. This property is called Freedom of Analysis.
(27)

Freedom of Analysis: Any amount of structure may be posited.

The only true restriction imposed on all output candidates generated by Gen is
that these are made up of licit elements from the universal vocabularies of linguistic representation, such as segmental structure (features and their grouping
below the level of the segment), prosodic structure (mora, syllable, foot, prosodic
word, etc.), morphology (root, stem, word, affix, etc.), and syntax (X-bar structure, heads/complements/specifiers, etc.). Within these limits, anything goes.
Since Gen generates all logically possible candidate analyses of a given input,
the OT grammar needs no rewrite rules to map inputs onto outputs. All structural
changes are applied in one step, in parallel. The evaluation of these candidate
analyses is the function of the Evaluator, the component of ranked constraints,
discussed in section 1.4.3. There we will also discuss the issue of whether or not
Eval is able to deal with an infinite candidate space.
1.4.3 The EVALUATOR: economy, strict domination, and parallelism
The Evaluator (henceforth Eval ) is undoubtedly the central component of
the grammar since it is burdened with the responsibility of accounting for all
11

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

20