[Dissertation] Reduplication, Feature Displacement, and Existential Faithfulness


249 96 492KB

English Pages [232] Year 2000

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Abstract
Table of Contents
1 Introduction
1.1 Optimality Theory and classic Correspondence Theory
1.2 Existential Faithfulness
1.2.1 Existential faithfulness constraints defined
1.2.2 Segmental preservation and reduplication
1.2.3 Preservation of feature specifications
1.2.3.1 Preservation of feature specifications and reduplication
1.2.3.2 Preservation of feature specifications and feature movement
1.2.3.3 Preservation of feature specifications and dissimilation
1.2.4 Preservation of adjacency and ordering relations
1.3 Fission and surface correspondence
1.4 Conclusion
Appendix I: Overview of existential faithfulness constraints
2 Reduplicative TETU
2.1 Faithfulness relations in reduplication
2.1.1 Broad input-output correspondence15 and Output TETU
2.1.2 Root faithfulness and Reduplicant TETU
2.1.3 Base-reduplicant correspondence
2.1.4 Summary
2.2 Reduplicant TETU: Kwakwala case study
2.2.1 Unreduplicated words and the moraic status of codas
2.2.2 The emergence of WEIGHTbyPOSITION in reduplicated words
2.2.3 Typological predictions: TETU in reduplicative and lexical affixes
2.3 Output TETU: Kwakwala case study
2.3.1 Unreduplicated words and stress clash
2.3.2 The emergence of *CLASH in reduplicative words
2.3.2.1 Type A words
2.3.2.2 Type B and type C words
2.4 Realization of reduplicative morphs and phonological reduplication
2.4.1 Non-realization of /RED/
2.4.2 Forces driving realization of /RED/
2.4.3 Reduplication in the absence of /RED/
2.5 Reduplicant size as a predictor of TETU alternations
2.6 Markedness constraints in Output and Reduplicant TETU
2.6.1 The effect of constraint domain size
2.6.2 Determining the alternation site in Output TETU
2.7 Division of input characteristics between base and reduplicant
2.8 Identifying base and reduplicant
2.9 The emergence of the faithful
2.10 Comparison with other proposals
2.10.1 Comparison with classic Correspondence Theory
2.10.1.1 The Emergence of The Unmarked
2.10.1.2 The Emergence of the Marked
2.10.1.3 Normal application
2.10.2 Comparison with other work assuming broad IO
2.11 Conclusion
3 Feature movement and dissimilation
3.1 Feature movement
3.1.1 Feature movement as fission and coalescence
3.1.1.1 Fission in feature movement
3.1.1.2 Coalescence
3.1.2 Combining fission and coalescence into feature movement
3.2 Dissimilation as a result of fission and coalescence
3.3 Case study: Sanskrit
3.3.1 Ban on laryngeally marked segments
3.3.2 [+Murmur] movement
3.3.2.1 Multiple feature movement?
3.3.3 Bartholomae’s Law
3.3.4 Grassmann’s Law
3.4 Case study: Cuzco Quechua
3.4.1 Feature value preservation
3.4.2 Floating features and the OCP
3.4.3 Further cooccurrence restrictions
3.5 Conclusion
Appendix II - The proximity effect
4 ∃-IDENT[F] and MAX[F] compared
4.1 Correspondence Theory and the status of features
4.2 Similarities between ∃-IDENT[F] and MAX[F]
4.3 Phenomena
4.3.1 Distributing diphthongization
4.3.2 Coalescence (and feature stability)
4.3.3 Feature movement
4.3.4 Dissimilation
4.4 Conclusion
Appendix III Sanskrit ranking
5 Conclusion
References
Recommend Papers

[Dissertation] Reduplication, Feature Displacement, and Existential Faithfulness

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

EXISTENTIAL FAITHFULNESS A Study of Reduplicative TETU, Feature Movement, and Dissimilation

by Caro Struijke

Revision of the dissertation submitted to the Faculty of the Graduate School of the University of Maryland at College Park in partial fulfillment of the requirements for the degree of Doctor of Philosophy August 2000

Advisory Committee: Professor Laura Benua, Chair Professor Luigi Burzio Professor Linda Lombardi Professor John McCarthy Professor Mary Ellen Scullen Professor Paul Smolensky

i

©Copyright by Carolina Maria Struijke 2000

ii

Abstract

The main thesis of this dissertation holds that faithfulness regulates preservation, and faithfulness constraints require input elements be present in the output. They do not demand identity of input and output strings, and are therefore existentially quantified. Preservation is less stringent than identity: only one output correspondent of an input segment needs to reflect the input segment’s feature specifications. Thus, in segmental fission, where an input segment has multiple output correspondents, only one correspondent need resemble the input segment. Other correspondents are free to respond to markedness constraints without violating faithfulness requirements. The dissertation investigates three empirical domains, all of which constitute ways to improve or satisfy markedness requirements while preserving underlying information. They are reduplicative TETU, feature movement, and dissimilation. Reference is also made to 'distributing diphthongization,' in which different features of a single input segment are preserved on two different output segments. The dissertation includes three in-depth case studies. The case study of Kwakwala reduplicated words shows that TETU alternations can affect either the reduplicant or the base. The case studies of Sanskrit and Cuzco Quechua show that feature movement and dissimilation often converge to result in a single alternation, and I claim that these two patterns are formally identical. Chapter 2 argues that reduplication involves segmental fission. Both members of the base-reduplicant pair relate to segments in the input via a general Input-Output faithfulness relation, dubbed 'Broad IO.' The base is in no sense prior to the reduplicant. The strings are of equal status with respect to IO faithfulness, and the existential definition of faithfulness allows TETU alternations affecting either member

iii

of the base-reduplicant pair. The proposed theory makes strong predictions as to which member is affected, depending mostly on the size of the domain evaluated by the emerging markedness constraint. Chapter 3 deals with both dissimilation and feature movement. I argue that both involve segmental fission where features of an input segment are distributed between two output correspondents. The correspondent preserving the 'moving' or 'dissimilating' marked feature coalesces with another segment. This research is in accord with recent work on dissimilation holding that it is driven by markedness requirements. However, existential faithfulness allows an account in which independently needed markedness constraints are at work, rather than constraints specifically banning multiple instances of a given marked feature (see also de Lacy and Struijke 2000). Throughout the dissertation I assume that features are attributes of segments, not independent entities. Chapter 4 compares these different approaches with respect to Correspondence Theory. Since ∃-IDENT[F] constraints of the first approach and MAX[F] constraints of the second approach are existentially defined, unidirectional, and feature value specific, both can account for the presented phenomena.

Key words: existential faithfulness, Correspondence Theory, reduplication, TETU, feature movement, dissimilation, OCP, cooccurence restrictions, (distributing) diphthongization, Kwakwala, Sanskrit, Cuzco Quechua

iv

Acknowledgments

Most thanks in these acknowledgments go to my advisor Laura Benua. She has taught me so much, and I hope her advice and insights find a reflection in this work. Her enthusiasm and support have helped me tremendously. I feel proud to be her student. Many key points in this dissertation were developed during my stay at UMass, where John McCarthy generously shared his wealth of knowledge. I fondly remember many enlightening discussions, fun chats, and reassuring encouragements. I could simply not have wished for a more wonderful committee. Luigi Burzio taught me that there are endless theoretical possibilities. At the same time, Linda Lombardi held my feet firm on the ground and taught me to be precise in the details. Mary Ellen Scullen was an endless source of great advice and editorial tips, and Paul Smolensky showed the power of theoretical subtleties and was a source of constant support. I want to thank my classmates, Juan Carlos Castillo, John Drury, Klea Grohmann, Akemi Matsuya, Julien Musolino, and Acrisio Pires for their friendship and for sharing the ups and downs of the graduate experience. Special thanks go to Haruka Fukazawa, Pat Hyronymous, Viola Miglio, Frida Morelli and Bruce Morén for adopting me in their ‘funology’ world and for being there all the way. Lisa Davidson, Matt Goldrick, Mits Ota, and Colin Wilson deserve thanks for making interdepartmental seminars so much fun to attend. I am grateful to the phonologists and phoneticians at UMass for providing an intellectually stimulating and supportive environment during my dissertation year: in particular Paul de Lacy, John Kingston, Ania Lubowicz, Steve Parker, Joe Pater, Jen Smith, and Lisa Selkirk.

v

I was fortunate to land a job at University of Toronto, where I revised parts of this dissertation. I want to thank everyone at the linguistics department, in particular those I got to talk phonology with: Peter Avery, Elan Dresher, Greg Guy, Wenckje Jongstra, Arsalan Kahnemuyipour, Sara MacKenzie, Jack Panster, Milan Rezac, Keren Rice, Nicole Rosen, and Tom Wilson. Thanks to Daniel Hall for carefully reading the penultimate draft of this work. Finally, I want to thank my friends and family for being existentially faithful.

vi

Table of Contents

1 Introduction

1

1.1 Optimality Theory and classic Correspondence Theory

10

1.2 Existential Faithfulness 1.2.1 Existential faithfulness constraints defined 1.2.2 Segmental preservation and reduplication 1.2.3 Preservation of feature specifications 1.2.3.1 Preservation of feature specifications and reduplication 1.2.3.2 Preservation of feature specifications and F movement 1.2.3.3 Preservation of feature specifications and dissimilation 1.2.4 Preservation of adjacency and ordering relations

18 19 21 23 25 26 28 31

1.3 Fission and surface correspondence

35

1.4 Conclusion

39

Appendix I: Overview of existential faithfulness constraints

40

2 Reduplicative TETU

45

2.1 Faithfulness relations in reduplication 2.1.1 Broad input-output correspondence and Output TETU 2.1.2 Root faithfulness and Reduplicant TETU 2.1.3 Base-reduplicant correspondence 2.1.4 Summary

48 49 52 53 57

2.2 Reduplicant TETU: Kwakwala case study 2.2.1 Unreduplicated words and the moraic status of codas 2.2.2 The Emergence of WEIGHTXPOSITION in reduplicated words 2.2.3 Typological predictions: TETU in red. and lexical affixes

58 59 62 66

2.3 Output TETU: Kwakwala case study 2.3.1 Unreduplicated words and stress clash 2.3.2 The emergence of *Clash in reduplicated words 2.3.2.1 Type A words 2.3.2.2 Type B and C words

68 69 73 73 77

2.4 Realization of reduplicative morphs and phonological reduplication 2.4.1 Non-realization of /RED/ 2.4.2 Forces driving realization of /RED/ 2.4.3 Reduplication in the absence of /RED/

80 80 83 86

2.5 Reduplicant size as a predictor of TETU alternations

88

vii

2.6 Markedness constraints in Output and Reduplicant TETU 2.6.1 The effect of constraint domain size 2.6.2 Determining the alternation site in Output TETU

91 91 93

2.7 Division of input characteristics between base and reduplicant

96

2.8 Identifying base and reduplicant

98

2.9 The emergence of the faithful

100

2.10 Comparison with other proposals 2.10.1 Comparison with classic Correspondence Theory 2.10.1.1 The Emergence of the Unmarked 2.10.1.2 The Emergence of the Marked 2.10.1.3 Normal application 2.10.2 Comparison with other work assuming broad IO

103 104 105 106 108 112

2.11 Conclusion

115

3 Feature movement and dissimilation

117

3.1 Feature movement 3.1.1 Feature movement as fission and coalescence 3.1.1.1 Fission in feature movement 3.1.1.2 Coalescence 3.1.2 Combining fission and coalescence into feature movement

119 122 123 124 128

3.2 Dissimilation as a result of fission and coalescence

131

3.3 Case study: Sanskrit 3.3.1 Ban on laryngeally marked segments 3.3.2 [+murmur] movement 3.3.2.1 Multiple feature movement? 3.3.3 Bartholomae’s Law 3.3.4 Grassmann’s Law

139 140 143 148 152 155

3.4 Case study: Cuzco Quechua 3.4.1 Feature value preservation 3.4.2 Floating features and the OCP 3.4.3 Further cooccurrence restrictions

164 166 171 175

3.5 Conclusion

177

Appendix II: The proximity effect

179

Appendix III: Sanskrit ranking

viii

4 ∃-IDENT[F] and MAX[F] compared

181

4.1 Correspondence Theory and the status of features 4.2 Similarities between ∃- IDENT[F] and MAX[F] 4.3 Phenomena 4.3.1 Distributing diphthongization 4.3.2 Coalescence (and feature stability) 4.3.3 Feature movement 4.3.4 Dissimilation 4.4 Conclusion

182 188 189 189 196 201 205 209

5 Conclusion 211 References

216

ix

Struijke

1 Introduction

Generative phonology assumes a mapping from underlying to surface forms (e.g. Chomsky 1966; Chomsky and Halle 1968; Anderson 1974). This view of the grammar allows us to determine a single form, the underlying or lexical representation, from which it is possible to derive all shapes it might exhibit on the surface. In Optimality Theory (Prince and Smolensky 1993) the nature of the underlying-surface mapping is determined by the interaction of faithfulness constraints, which demand that the two representations match, and well-formedness constraints, which may prefer a change in the surface form. Faithfulness constraints are usually thought to evaluate the degree of identity between the input and the output, so that any deviation between these strings violates such a constraint (Prince and Smolensky 1993; Correspondence Theory - McCarthy and Prince 1995, 1999). Although I assume the basic premises of Correspondence Theory, I argue instead that faithfulness constraints demand preservation of input material in the output, not identity of the two representations. That is, faithfulness constraints merely demand that an input element have some output correspondent. The constraints are thus existentially quantified. In a simple one-to-one input-output mapping, identity and preservation are synonymous. For instance, given an input segment and a single identical output correspondent, both identity and preservation are achieved. However, in 'fission', where an input segment is related to two or more output segments, preservation does not imply identity. If only one output correspondent shares a given feature specification with the input segment, preservation of that input feature specification is achieved, because it is preserved on some correspondent in the output. However, segmental identity is not attained because not all output correspondents are like the input segment.

1

Existential Faithfulness

This idea can be exemplified by diphthongization. In diphthongization, an underlying segment (vowel or consonant) undergoes fission and surfaces as a heterogeneous sequence (cf. Hayes 1990). In one type of diphthongization, the underlying segment divides its features between the two surface segments. I refer to this as 'distributing diphthongization.' Many examples come from the phonology of loan words. In English loan words from French, for instance, [ü] is realized as the diphthong [iu] (Shane 1984 (Middle English); Hawkins 1984 (Modern English)), where the front and round qualities of the input vowel are maintained on different segments. When a French word containing a nasalized vowel is the input to the present-day German and English grammars, a sequence of an oral vowel and nasal arises (Itô and Mester 2000; Hawkins 1984, respectively), preserving the vowel quality and nasality of the input segment on different output segments. Examples of distributing diphthongization outside the domain of lexical borrowing include Spanish diphthongization of stressed vowels (Harris 1969; Halle, Harris and Vergnaud 1991 - quoted in Cole 1995) and Icelandic pre-aspiration (Tráinsson 1978, amongst others; discussed in chapter 4). (1.)

English distributing diphthongization (loans from French) ‘déjà v[ü]’ [deIZ´ viu] German diphthongization (loans from French) ann[ç)]ce ann[çN]ce ‘advertisement’ bal[A)]ce bal[AN]ce ‘balance’ cous[E)] cous[EN] ‘(male) cousin’ English distributing diphthongization (loans from French) [bç)bç)] [bAnbAn] ‘candy’ [deta)t] [deItAnt] ‘relaxation’ Spanish distributing diphthongization c[ç]nt-a@-ba c[ue@]nt-a ‘he counted / counts’ n[e]g- a@-ba n[ie@]g-a ‘he denied / denies’ p[E]ns-a@mos p[ie@]ns-o ‘we / I think’ s[ç]lt-a@mos s[ue@]lt-o ‘we / I release’

In Correspondence Theory, diphthongization is analyzed as multiple correspondence. Thus, the Spanish input segment /ç/ has two output correspondents in

2

Struijke

stressed syllables: [u] and [e]. This is depicted below. Throughout this dissertation, lines and indices indicate correspondence relations. (2.)

Spanish diphthongization unstressed syllables input::

ç1

output;:

ç1

stressed syllables ç1

u1

e1

In Spanish stressed syllables of this type, corresponding segments are not identical. Hence, the universally quantified faithfulness constraints of classic Correspondence Theory are violated. However, the existentially quantified faithfulness constraints proposed in this dissertation are satisfied, because each feature specification associated with the input vowel is preserved on at least one of its output correspondents. The back and round features are preserved on the first vowel, the height feature on the second vowel. Thus, the existential view of faithfulness allows a formalization of the idea that distributing diphthongization ensures faithfulness.1 One of the main claims of this dissertation is that segmental fission allows improved markedness of a phonological structure, while at the same time securing preservation of underlying material. In the Spanish example, diphthongization allows the stressed nucleus to be heavy, probably in satisfaction of the markedness constraint STRESStoWEIGHT. In the French - English borrowing mapping /ü/ - [iu], a marked front rounded vowel is avoided without loss of the underlying front and round features. Thus, distributing diphthongization allows simultaneous satisfaction of a markedness 1

The problem that classic Correspondence Theory faces in distributing diphthongization was previously noted in Keer's (1999) and Morén's (1999) analyses of Icelandic preaspiration. The latter adopts existentially defined faithfulness from Struijke (1998). The former assumes that aspiration is a 'semi-independent segment.'

3

Existential Faithfulness

constraint (“if stressed, then heavy,” “don’t have [ü]”) and faithfulness constraints (“preserve [+round],” etc.). This dissertation argues that a range of phenomena exhibit simultaneous satisfaction of potentially conflicting markedness and faithfulness constraints. Apart from distributing diphthongization, these include The Emergence of The Unmarked in reduplication, feature movement, and dissimilation. I claim that all these phenomena involve segmental fission, in which a single input segment corresponds to two output correspondents. If faithfulness constraints are existentially quantified, only one output correspondent needs to preserve a given feature specification, leaving the other free to respond to markedness requirements. I will now briefly discuss the three patterns in turn.

It is a well-known fact that reduplicated words often include less marked structure than unreduplicated words (McCarthy and Prince 1986, 1994a; Shaw 1987; Steriade 1988). For example, Sanskrit permits complex onset clusters and codas in unreduplicated words, but only CV syllables in reduplicants (Steriade 1988). This pattern is dubbed 'The Emergence of The Unmarked' (McCarthy and Prince 1994a), commonly abbreviated as TETU. In chapter 2, I argue that emergent unmarkedness in reduplicated words can be attributed to existential faithfulness constraints if reduplication is treated as segmental fission. Reduplicative fission is depicted below with an example of Malay full reduplication. In full reduplication, each input segment has a correspondent in the base and a correspondent in the reduplicant.

4

Struijke

(3.)

Malay full reduplication (mata 'eye' - mata-mata 'policeman') Input

Output

RED + m a t a

m a t a

m a t a

Given existential faithfulness, only one member of the base-reduplicant pair must preserve input material to satisfy faithfulness constraints on the input-output relation. The other member can become unmarked without incurring a penalty. For instance, in the Sanskrit mapping /svap/ - [sa:-svap] ('sleep' (intensive)) each input segment is preserved in the output (namely the base), while syllabic markedness is improved in the reduplicant. The Tübatulabal mapping /Sˆ/ˆwˆ/ - [/ˆ˘-Sˆ/ˆwˆ] ('it looks different') constitutes a second example of The Emergence of The Unmarked (Alderete et al. 1999). Here, the marked place and manner features of the initial input segment are preserved in the base correspondent, but lost in the reduplicant correspondent. Thus, the account of emergent unmarkedness in reduplication that is presented in this dissertation formalizes the idea that markedness can be improved without compromising preservation of underlying material (cf. McCarthy and Prince 1994a). It was previously assumed that only reduplicants can alternate to achieve unmarked structure in reduplication. However, I show that bases can sometimes be affected by The Emergence of The Unmarked instead. If reduplication involves fission, and faithfulness is existentially defined, this pattern is expected. Fissioned output segments are of equal status, because either can ensure preservation of input features and satisfy input-output faithfulness constraints. Existential constraints are indifferent as to which member of the base-reduplicant pair is faithful and which changes to decrease markedness: either the reduplicant or the base can undergo a TETU alternation.

5

Existential Faithfulness

Usually, one output correspondent in a reduplicative word is completely like the input segment while the other becomes unmarked. In the Tübatulabal example mentioned above, the reduplicant consonant is unmarked with respect to place and manner features, while the corresponding base consonant remains unchanged. Such asymmetry is also found in some cases of diphthongization. In Mälaren Region Swedish (Hayes 1990), for instance, a full vowel V diphthongizes to V´, where the second output correspondent acquires unmarked vowel features, and the first stays unchanged. In reduplication, it is usually the correspondent in the base that is like the input segment. However, reduplication sometimes involves division of input features between segments in the base and the reduplicant, comparable to fission in English /ü/ -[iu] diphthongization. For instance, in the Japanese mapping /p1i2t3o4/ - [h1i2t3o4-b1i2t3o4] the first correspondent of /p1/ preserves the voicelessness of its input corespondent, while the second correspondent preserves the manner and place features. Whether input features are divided between output correspondents or whether one preserves all features and the other becomes maximally unmarked depends on language-specific constraint rankings. Put differently, the degree of similarity between an input segment and each of its output segments is an emergent property of the constraint ranking: neither correspondent is inherently closer to the input than the other. Throughout this dissertation, I use the term ‘distributing fission’ to refer to any situation in which an input segment divides its feature specifications among two or more output correspondents (including distributing diphthongization and reduplication patterns similar to Japanese). In distributing fission, feature specifications associated with a single input segment surface in different positions in the output word. As a result, it sometimes seems as if a feature has moved from one segment to another. In Esimbi, for example, height

6

Struijke

features seem to move from the root vowel to the prefix vowel (Stalcup 1980a,b, Hyman 1988, Walker 1997). (4.)

Esimbi feature movement /u-se/ osi /u-rE/ çri /i-gbe/ egbi /i-so/ esu /i-b´/ ebˆ

‘laugh’ ‘daub’ ‘bushfowl’ ‘hoe’ ‘cane rat’

I argue in chapter 3 that feature movement is an epiphenomenon. The Esimbi height feature [-hi] is not actually reassociated. Instead, the root vowel undergoes fission and its different feature specifications are preserved on two different output correspondents. One correspondent preserves the underlying color features, while the other maintains the height feature. As a result, the relevant existential faithfulness constraints are satisfied. This is depicted below. (5.)

Esimbi feature movement involves segmental fission (only input features are shown) [-hi] [-back] [- rd]

input:

output:

e1

o1

i1

[-hi]

[-back] [- rd]

7

Existential Faithfulness

So far, feature movement2 resembles distributing diphthongization. However, the two phenomena differ in an important way. In movement, one output correspondent of a fissioned input segment coalesces with another segment in the string. The coalesced segment then preserves features from its two input correspondents, as shown below for the Esimbi example. (6.)

Feature movement is a combination of fission and coalescence [+back][+hi][+rd]

input:

output:

u1

o 1,3

[-back][-hi][-rd]

s2

s2

e3

i3

[+back][-hi][+rd]

[-back][-rd]

Feature movement takes place to improve markedness of a segment. In Esimbi, non-high vowels such as /e/ are marked in non-initial syllables. Feature movement thus improves markedness while at the same time ensuring preservation of the underlying feature specification that is prevented from surfacing in a particular environment. The proposed analysis shows that one does not need to ascribe an autosegmental status to features in order to explain feature movement. Instead, features can be seen as properties of segments. In chapter 4, I compare the features-as-attributes view with the features-as-entities view.3 I refer to the first as 'Featural Attribute Theory,' and the second as 'Featural Independence Theory.'

2

Throughout this dissertation, I use the term ‘feature movement’ as a descriptive term to refer to any mapping in which a feature specification is associated with one segment in the input, but another segment in the output. ‘Feature displacement’ and ‘feature transfer’ are other terms found in the literature to refer to this phenomenon. 3 The terms 'attributes' and 'entities' as applied to features are due to McCarthy (1996).

8

Struijke

Dissimilation, like distributing diphthongization, reduplicative TETU, and feature movement, permits simultaneous preservation of underlying material and improvement of markedness. Again, this is accomplished through segmental fission, and the existential definition of faithfulness constraints plays a crucial role. The proposal is developed in chapter 3 and builds on Struijke (2000b,c) and Struijke & de Lacy (2000). It formalizes the idea that one output segment can ensure preservation of identical feature specifications belonging to two input segments. The other output segment alternates to meet markedness requirements. As in feature movement, dissimilation involves both fission and coalescence. This is depicted below: (7.)

Dissimilation [lab][+nas]

input:

output:

[lab][-cont][-vc]

m1

p2

m 1,2

t3

[lab][+nas]

[cor][-cont][-vc]

In this example, the first output consonant ensures preservation (i.e. existential faithfulness) of the labial feature specifications associated with both its input correspondents. The second output consonant can become unmarked without incurring a faithfulness violation. All phenomena analyzed in this dissertation are accounted for by assuming that featural faithfulness is evaluated over segmental correspondence by IDENT[F] constraints (McCarthy and Prince 1995, 1999), rather than by MAX[F] constraints (Causley 1997a,b; Lamontagne and Rice 1995; Lombardi 1995, 1998; Parker 1997; Walker 1997). This

9

Existential Faithfulness

assumption crucially relies on the idea that featural faithfulness constraints are existentially quantified. In chapter 4, I show that existential IDENT[F] and MAX[F] constraints share important properties, which explains their comparable success in accounting for the aforementioned phenomena. Because this dissertation aims to contribute to the Correspondence Theory of faithfulness, an overview of 'classic' Correspondence Theory (as proposed in McCarthy and Prince 1995, 1999) is given in section 1.1, set in a basic outline of Optimality Theory (Prince and Smolensky 1993). Section 1.2 defines existential faithfulness and the constraints that enforce it. In addition, it gives an outline of the dissertation. This chapter is concluded by an appendix that provides an overview of the existential faithfulness constraints and their formal definitions.

1.1 Optimality Theory and classic Correspondence Theory

The OT grammar consists of three components: GEN, CON, and EVAL (Prince and Smolensky 1993). GEN generates output candidates from a given input. These candidates are evaluated by EVAL to determine which candidate best satisfies the language particular hierarchy of constraints (CON). This most harmonic, or optimal form is the actual output.

10

Struijke

(8.)

OT grammar candidate 1

/input/

G

candidate 2

C

E

candidate 3

O

N

candidate 4

N

EVAL

[optimal candidate]

candidate 5

For a given input, the function GEN produces a set of candidates. Correspondence Theory (McCarthy and Prince 1995) holds that each candidate consists of the input, a possible output, and correspondence relations between the segments in these structures. Correspondence relations permit EVAL to make comparisons between different parts of a candidate - in the most obvious case, between the input and the output. Correspondence is defined as below. I adopt the definition in this dissertation. (9.)

Correspondence (McCarthy and Prince 1995, 1999) Given two related strings S1 and S2, correspondence is a relation ℜ from the elements of S1 to those of S2. Elements α∈S1 and β∈S2 are referred to as correspondents of one another when αℜβ.

GEN is free to supply a potentially infinite set of candidates with a variety of correspondence relations (‘freedom of analysis’). I illustrate this below. Correspondence relations are indicated by indices.

11

Existential Faithfulness

(10.)

Some candidates generated by GEN for the input /lamtI/ 1. /l1a2m3t4I5/ - [l1a2m3t4I5] 2. /l1a2m3t4I5/ - [l1a2t4I5] 3. /l1a2m3t4I5/ - [l1a2m3et4I5] 4. /l1a2m3t4I5/ - [n1a2m3t4I5] 5. /l1a2m3t4I5/ - [l1a2n3,4I5] 6. /l1a2m3t4I5/ - [l1a2m3t4a5i5] 7. /l1a2m3t4I5/ - [l1a2t4m3I5]

Correspondence relations keep track of the degree to which inputs and output candidates match. In candidate 1, correspondence is perfect because the output is exactly like the output. In all other candidates correspondence is imperfect. This imperfection is reflected in the correspondence relations. In candidate 2, /m/ deletes and has no correspondent in the output. Conversely, in candidate 3 epenthesis takes place, so that the output segment [e] has no input correspondent. In candidate 4, all segments have an output correspondent, yet perfect correspondence is not achieved, because the input segment /l/ surfaces as [n]. GEN is also free to establish multiple correspondence, in which a segment in one string is in correspondence with two or more segments in the other string. Candidate 5 shows coalescence (two segments in the input correspond to one segment in the output), while candidate 6 shows fission (one segment in the input corresponds to two segments in the output). In candidate 7, metathesis has taken place. The choice among candidates is determined by their performance against the constraint set CON. CON is a set of universal, violable constraints which are ranked on a language-particular basis. It includes faithfulness constraints, which evaluate the similarity between the input and the output by assessing the correspondence relations between these strings in each candidate. In classic Correspondence Theory (McCarthy and Prince 1995), every possible deviation from perfect identity between the input and the output incurs a faithfulness violation. Individual faithfulness constraints demand completeness of correspondence,

12

Struijke

featural identity of segments, preservation of linear order and contiguity of segments under correspondence, and so forth. Candidate 1 (/l1a2m3t4I5/ - [l1a2m3t4I5]) satisfies all faithfulness constraints. Candidate 2 (/l1a2m3t4I5/ - [l1a2t4I5]) deletes one segment and thereby violates MAX, which demands that an input segment have a correspondent in the output. The definition of this constraint is given below. (All definitions in this section are from McCarthy and Prince 1995.) Some of these familiar faithfulness constraints are to be modified later in this dissertation, while others among those given by McCarthy and Prince (1995) are not adopted at all (see section 1.2 below). I adopt the basic insight behind the constraint MAX. (11.)

MAX: Every segment in S1 has a correspondent in S2

Candidate 3 (/l1a2m3t4I5/ - [l1a2m3et4I5]) epenthesizes a segment and thus violates DEP. This constraint is the counterpart of MAX because it demands that all output segments have a correspondent in the input. I will not adopt this constraint in this thesis (see section 1.2). (12.)

DEP: Every segment of S2 has a correspondent in S1

McCarthy and Prince (1995, 1999) suggest that correspondence relations do not necessarily hold over segments alone. It is conceivable that relations are also established over suprasegmental elements, such as moras (McCarthy 1997a; Morén 1999; Ota 1999; Keer 1999, amongst others), prosodic heads (Alderete 1999), and tones (Bickmore 1996; Yip 1996, Myers 1997, amongst others). It has also been proposed that subsegmental features are independent elements that can enter into correspondence relations (Causley

13

Existential Faithfulness

1997a,b; Lamontagne and Rice 1995; Lombardi 1995, 1998; Parker 1997; Walker 1997). Featural faithfulness is then evaluated by MAX[F] and DEP[F] constraints, on analogy with MAX(seg) and DEP(seg). However, as an implementational assumption, McCarthy and Prince take features to be attributes of segments, and only segments enter in correspondence relations. They assume that faithfulness to subsegmental features is then mediated by segmental correspondence and evaluated by IDENT[F] constraints. In classic Correspondence Theory, IDENT[F] constraints demand that all corresponding segments be featurally identical. The constraints are relativized to a particular feature, each demanding that corresponding segments agree with respect to the value of that feature.

(13.)

IDENT[F]: Correspondent segments have identical values for the feature F If xℜy and x is [γF], then y is [γF]

I adopt the view that featural faithfulness is regulated by IDENT[F] constraints rather than MAX[F] constraints, although I will modify this constraint to reflect existential quantification4 (section 1.2.3). Candidate 4 in (10) above (/l1a2m3t4I5/ - [n1a2m3t4I5]), violates an IDENT constraint relevant to manner features, since the lateral input segment corresponds to a nasal output segment. The coalescing candidate 5 (/l1a2m3t4I5/ - [l1a2n34I5]) also violates IDENT constraints, because correspondents /m/ and [n] do not match in place features, while /t/ and [n] disagree in manner features. In addition, this candidate violates UNIFORMITY, which bans coalescence. I will not adopt the latter constraint here (see the appendix to this chapter).

4

The definition that I propose is in some sense a compromise among MAX[F], DEP[F], and classic IDENT[F]. This is discussed in chapter 4.

14

Struijke

(14.)

UNIFORMITY: ‘No Coalescence’ No element of S2 has multiple correspondents in S1 For x,y ∈ S1 and z ∈ S2, if xℜz and yℜz, then x=y

The fission candidate (/l1a2m3t4I5/ - [l1a2m3t4a5i5]) violates classic IDENT[F] constraints for the mappings /I/ - [a] and /I/ - [i], as well as the INTEGRITY constraint, which prohibits fission. (15.)

INTEGRITY: ‘No Fission’ No element of S1 has multiple correspondents in S2 For x ∈ S1 and w,z ∈ S2, if xℜw and xℜz, then w=z.

Finally, Classic Correspondence Theory assumes faithfulness to adjacency and ordering relations between segments in the input. The definitions of the two relevant constraints are given below: (16.)

CONTIGUITY : I-CONTIG: Domain (ℜ) is a single contiguous string in S1. The portion of S1 standing in correspondence forms a contiguous string. No skipping O-CONTIG: Range (ℜ) is a single contiguous string in S2. The portion of S2 standing in correspondence forms a contiguous string. No intrusion

(17.)

LINEARITY: Let x, y ∈S1 and x', y' ∈ S2. If xℜx' and yℜy', then x < y iff ¬ (y' < x'). S1 is consistent with the precedence structure of S2, and vice versa. "No Metathesis"

15

Existential Faithfulness

O-CONTIG is violated in the epenthesis candidate /l1a2m3t4I5/ - [l1a2m3et4I5], while I-CONTIG is violated in the deletion candidate /l1a2m3t4I5/ - [l1a2t4I5]. Thus, these constraints are violated in cases of string-internal epenthesis and deletion. LINEARITY is violated in the metathesis candidate /l1a2m3t4I5/ - [l1a2t4m3I5] because m3 precedes t4 in the input but not in the output. I will argue in section 1.2.4 that the CONTIGUITY and LINEARITY constraints must be existentially defined. Apart from input-output (IO) representations, correspondence relations can in principle be established between segments in any pair of strings. The most commonly encountered proposals are base-reduplicant (BR) correspondence in reduplicated words (McCarthy and Prince 1995, 1999), and output-output (OO) correspondence in morphological paradigms (Benua 1995, 1997; Burzio 1994). Faithfulness constraints are relativized for the particular correspondence relation they evaluate. For instance, there are MAXIO, MAXBR and MAXOO constraints, which coexist and are separately rankable in the constraint hierarchy. This dissertation mainly focuses on relations between the input and the output, but I will argue that output segments corresponding to a single input segment are in a correspondence relation with each other. The idea that constraints on these relations are separately rankable is crucial in accounting for ‘The Emergence of The Unmarked’ in reduplication, feature movement, and dissimilation. Prominent morphological categories and prosodic positions such as roots, onsets and initial segments are commonly held to be subject to additional faithfulness requirements through Positional Faithfulness constraints (Selkirk 1995; Beckman 1995, 1997; Alderete 1996; Lombardi 1996). These constraints explain why contrasts are often maintained in these positions, while they are neutralized in other environments (a generalization ultimately going back to Trubetzkoy 1939). Root faithfulness plays a crucial role in accounting for emergent unmarkedness patterns in reduplication (chapter 2).

16

Struijke

In addition to faithfulness constraints, the constraint set CON contains markedness constraints that evaluate the phonological well-formedness of output forms. When a markedness constraint dominates a faithfulness constraint, it can force an alternation in a mapping, and hence an imperfection in correspondence. Throughout this dissertation I use the terms ‘marked’ and ‘unmarked’ as defined in Optimality Theory by Prince and Smolensky (1993), and Smolensky (1993): forms are marked with respect to a constraint if they violate it, and unmarked if they do not violate it. CON does not contain constraints that refer to inputs. Constraints refer either only to outputs (markedness constraints) or to corresponding strings such as input-output pairs (faithfulness constraints). Optimality Theory attributes all linguistic generalizations to these constraints, be they evident in overt alternations, distributional facts, or both. Thus, OT does not have the means to restrict inputs in any way. The set of possible inputs is therefore assumed to be universal (‘Richness of the Base’; Prince and Smolensky 1993). Any input that is fed into a language-particular grammar must give rise to an output that is well-formed in that language. This notion plays an important role in the case studies presented in chapter 3. The component EVAL evaluates the candidates provided by GEN against the language particular ranking of the constraints in CON and orders them according to how well they satisfy this constraint hierarchy (‘harmonic ordering’). The actual output or ‘optimal candidate’ is the candidate that best satisfies the constraint hierarchy (the first in the harmonic ordering). All other candidates are ungrammatical.

1.2 Existential Faithfulness

Although I adopt the premises of Correspondence Theory, I depart from the idea that faithfulness constraints demand identity between the input and the output. Instead, I

17

Existential Faithfulness

claim that these constraints require preservation of underlying material in the surface form. Therefore, they are existentially defined, demanding that each characteristic of the input exists in the output. In addition, faithfulness constraints are unidirectional and feature value specific. I argue that these characteristics are shared by all faithfulness constraints evaluating correspondence relations along the input-output dimension, namely general input-output faithfulness and positional faithfulness such as root faithfulness. In section 2.1.3, I discuss the hypothesis that base-reduplicant faithfulness constraints are also existentially quantified. Whether this hypothesis is correct is an empirical question which may not be testable. The remainder of this chapter discusses how specific existential faithfulness constraints evaluate candidates. I first focus on constraints that demand preservation of segments and their featural attributes, namely ∃-MAXIO and ∃-IDENTIO. I then turn to constraints that require preservation of underlying precedence and adjacency relations, namely ∃-LINEARITYIO and ∃-CONTIGUITYIO. In addition to introducing the existential interpretation of faithfulness constraints, this section presents an outline of the dissertation, sketching the main arguments of the individual chapters. A comprehensive list of existential constraint definitions can be found in the appendix at the end of this chapter.

1.2.1 Existential faithfulness constraints defined

Since input-output faithfulness regulates preservation, but not identity of strings, all constraints are existentially quantified, and no faithfulness constraint is universally quantified. This proposal has no effect on the interpretation of the constraint MAX, since it has been existentially defined since its inception. It is satisfied when an input segment has some correspondent in the output. However, the proposal has an important effect on

18

Struijke

the interpretation of the constraints IDENT[F], CONTIGUITY, and LINEARITY in particular. Only one output correspondent needs to be featurally like an input segment to satisfy an ∃-IDENT[F] constraint. Other correspondents are allowed to change. ∃-LINEARITY is satisfied when some pair of output segments preserves the precedence relation of its corresponding input segment pair. Likewise, ∃-CONTIGUITY is satisfied when some pair of output segments preserves the adjacency relation of its corresponding input segment pair. Again, any additional output correspondents may alter these relationships. A key component of the proposed theory of preservation is that faithfulness constraints are unidirectional. That is, they require correspondents of input material in the output, not vice versa (see also Causley 1999). This reflects the idea that the input is logically prior to the output, in the sense that all shapes that a surface form might acquire are derived from the underlying form. The English diphthongization example illustrates why unidirectionality is crucial. In the /u_1/ - [i1u1] mapping, every feature of the input exists in the output, but not vice versa. In particular, the [+back] feature of [u] and the [-round] feature of /i/ are not present in the input. This non-identity is irrelevant to unidirectional ∃-IDENT[F] constraints. A consequence of the idea that faithfulness constraints are unidirectional is the absence of DEP. This constraint cannot be assumed because it demands presence of an output segment in the input. I argue that epenthesis rather violates M-SEG, a markedness constraint that is violated when an output segment has no morphological affiliation (McCarthy 1993)5,6 . 5

Causley (1999) also assumes the absence of DEP. In her proposal, epenthesis is penalized by the constraint *STRUC (a constraint that penalizes structure in general - see also Zoll 1996). Smolenksy (p.c.) points out that, when high-ranking, this constraint not only prefers deletion to epenthesis (the desired result), it also chooses as optimal a candidate in which all input segments are deleted (i.e. the ‘null-candidate’). This is illustrated in the tableau below: /CVC/ NOCODA *STRUC MAX 1 CVC *! *** 2 CV.CV ****! 3  CV **! *

19

Existential Faithfulness

(18.)

M-SEG: Every segment belongs to a morpheme (McCarthy 1993)

The principle of 'Consistency of Exponence' (McCarthy and Prince 1993b) ensures that all and only epenthesized segments lack morphological affiliation. It holds that 'candidate output forms do not permit changes in the exponence of specified[7] morphemes: the underlying and surface segmental affiliation of a given morpheme must be identical' (p. 62). Apart from existentialism and unidirectionality, I assume that featural faithfulness constraints demand preservation of particular feature values (I refer to this property of constraints as 'monovalency'8 ). The constraint ∃-IDENT[-voice] IO, for instance, penalizes voicing, while ∃-IDENT[+voice] IO penalizes devoicing of a segment. I show in section 3.1.1.2 that this characteristic is crucial in accounting for coalescence (following Casali 1996; Gnanadesikan 1995, 1997; McCarthy 1995; Pater 1999, 2000).

1.2.2 Segmental preservation and reduplication

The formal definition of the segmental faithfulness constraint ∃-MAX IO is given below. This constraint is the same as the one proposed in McCarthy and Prince (1995). (19.)

∃-MAX IO: Let seg ∈ input; then there is some seg' ∈ output, such that segℜseg'.

4 Ø

***

Alternatively, one could assume that epenthesis is prevented by ∃-CONTIGUITY (string-internally) and Alignment constraints (at the edges of strings). 7 The caveat 'specified' refers to the fact that segments in the output reduplicant are not considered morphological exponents of the morphemes they copy. 8 Note the feature specifications are binary, and featural faithfulness constraints are feature value specific. 6

20

Struijke

Every segment in the input has some correspondent in the output.

In chapter 2, I illustrate the nature of this constraint by comparing reduplicated and unreduplicated words. As mentioned earlier, I argue that reduplication involves fission such that input segments typically have correspondents in both the base and the reduplicant. In unreduplicated words, however, each input segment usually has one output correspondent. This is illustrated in (20) (lines indicate correspondence relations).

(20.)

Input-output relations unreduplicated word Input:

t a b i

Output:

t a b i

reduplicated word (red. underlined) RED + t a b i

t a b i

t a b i

Both of these hypothetical forms satisfy ∃-MAXIO, because each input segment is maintained in the output. The constraint is indifferent to the number of output correspondents for a given input segment. Likewise it is indifferent to the location of the correspondents. Put differently, it is of no importance to ∃-MAXIO whether the input segment is preserved in the base, the reduplicant, or both. The constraint is therefore satisfied in the forms below. (21.)

Input-output relations (reduplicant underlined) (a) deletion in the reduplicant (b) deletion in the base Input

Output

RED + t a b i

t a

RED + t a b i

t a b i

t a b i

21

t a

Existential Faithfulness

(c) deletion in both reduplicant and base Input

Output

RED + t a b i

t a b

t

a

i

In (21c) the base and reduplicant each delete a segment. Yet, no ∃-MAXIO violation is incurred, because they delete different segments: every input segment has an output correspondent. Chapter 2 examines the idea that there is one chance to satisfy ∃-MAXIO in unreduplicated words, but two chances to satisfy it in reduplicated words. With undominated ∃-MAXIO, we predict that deletion is prohibited in unreduplicated words, but allowed in one part of a reduplicated word. Put differently, in unreduplicated words marked structure cannot be repaired by deletion, but in reduplicated words deletion is permitted in one member of the base-reduplicant pair without incurring a faithfulness penalty. This is 'The Emergence of The Unmarked' (TETU; McCarthy and Prince 1994a).9 In unreduplicated words of Tohono O’odham (Uto-Aztecan) (Fitzgerald 1998), for instance, high-ranking ∃-MAXIO ensures that adjacent vowels are both parsed, even though they create hiatus, which is a marked structure. This is seen in the first column of (22) below. In the reduplicated words of the second column, hiatus is resolved by preserving one vowel in the base and one vowel in the reduplicant, as allowed by ∃-MAXIO.

9

As I show in section 2.10.1, the proposed approach to reduplicative TETU differs from that presented in McCarthy and Prince 1994a in crucial ways.

22

Struijke

(22.)

The Emergence of The Unmarked in Tohono O’odham (formerly Papago) unreduplicated word reduplicated word gloss hi.o.sig hi.-ho.sig ‘flower’ ˜ˆ.ok ˜ˆ.-˜ok ‘talking’ do.a do.-da ‘to be healthy’ to.a.ya to.-ta.ya ‘towel’

The fact that a TETU alternation can affect the base as well as the reduplicant shows that reduplication involves fission. Or, to put it in reverse terms, the fact that either can ensure satisfaction of undominated ∃-MAXIO shows that both the base and the reduplicant are in direct correspondence with the input. Inputs must be related to entire output words, including both the base and the reduplicant in reduplicated words. I argue that IO correspondence is therefore broadly defined to relate inputs to output words as a whole (‘Broad Input-Output Correspondence’) (cf. McCarthy and Prince 1995, 1999). The base and the reduplicant enjoy equal status for the purposes of IO Faithfulness requirements (i.e., IO constraints do not privilege the base or the reduplicant). Which member deletes a segment in TETU alternations is determined by the ranking of other constraints, as I will show in chapter 2.

1.2.3 Preservation of feature specifications

The definition of the featural faithfulness constraint ∃-IDENT[F] IO is given below: (23.)

∃-IDENT[F] IO: Let seg ∈ input be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an input segment preserves the feature specification [αF] of that input segment.

This constraint requires that a feature specification associated with an input segment be also associated with some corresponding output segment. In the absence of a

23

Existential Faithfulness

corresponding output segment, ∃-IDENTIO constraints cannot demand preservation of an input feature specification (McCarthy and Prince 1995). That is, when an input segment is not in the domain of ℜ, (and thus has no output correspondent), ∃-IDENTIO constraints are vacuously satisfied. Therefore, all mappings given below satisfy ∃-IDENT[+voice]IO. (24.)

∃-IDENT[+voice] IO is satisfied in each of the following: (a)

Input:

g

Output:

g

(b) [+voice]

g

[+voice]

d

(c) [+voice]

g

[+voice]

[+voice]

When a segment has only one output correspondent, ∃-IDENT[F]IO has the same effect as classic (universally quantified) IDENT[F]IO constraints in that the output segment must be identical to the input segment. However, when a segment has multiple output correspondents, only one needs to be match the relevant feature value of the input segment to satisfy an ∃-IDENT[F] constraint. (25.)

∃-IDENT[+voice] IO is satisfied in each of the following: (a)

(b)

g

Input:

– Output:

g

[+vc]

(c)

[+vc]

g

— g

– [+vc]

g

24

[+vc]

[+vc]

g

— k

– [-vc]

d

[+vc]

[+vc]

— [-vc]

k

Struijke

(d)

g

Input:

– Output:

t

[-vc]

[+vc]

— g

[+vc]

Mapping – in (25b and c) and mapping — in (25d) ensure satisfaction of ∃-IDENT[+voice]: the [+voice] specification associated with the input segment perseveres on some output correspondent. Because the faithfulness constraint is existentially quantified and unidirectional, mappings — in (25b and c) and mapping – in (25d) do not cause a violation of this constraint. I now turn to the role of ∃-IDENT[F] in reduplication, feature movement, and dissimilation.

1.2.3.1 Preservation of feature specifications and reduplication

This section briefly illustrates the role of ∃-IDENT[F] constraints in reduplication by examining The Emergence of The Unmarked in Lushootseed diminutive reduplication. In Lushootseed (Salish), unreduplicated words are faithful. Each vowel in the input preserves its marked underlying features, regardless of whether it is stressed or unstressed.10 Thus, ∃-IDENT[F]IO constraints on vowel quality features dominate the markedness constraint penalizing full unstressed vowels.

10

In fact, unstressed vowels in unreduplicated words optionally reduce. Since this has no bearing on the argument presented, it is ignored here (but see Struijke 1998 for an account of variation that includes existential faithfulness constraints).

25

Existential Faithfulness

(26.)

Lushootseed faithful unreduplicated words //idigWat/ /i@digWa$t ‘say something’ //agWal-´b/ /a@gWal-´b ‘yawn’

In words reduplicated for diminutive effect (Bates 1986; Broselow 1983; Urbanczyk 1996), the unstressed member of the base-reduplicant pair undergoes vowel reduction, while the stressed member preserves the underlying vowel quality features. (27.)

Vowel reduction in reduplicated words of Lushootseed /RED-/agWal-´b/ /a@-/´gWa$l-´b ‘yawn (dim)’ /RED-tadz-´d/ t´-ta@dz-´d ‘little dance’

The undominated ∃-IDENT[F]IO constraints on vowel quality features are satisfied because one correspondent of each vowel preserves these features. Which member of the base-reduplicant pair preserves them depends on whether the reduplicant syllable is stressed or the base-initial syllable. Again, the unmarked can emerge in reduplication without violating existential faithfulness.

1.2.3.2 Preservation of feature specifications and feature movement

Diphthongization examples such as the French - English loan mapping /ü/ - [iu] show that feature specifications associated with an input segment can be distributed among output correspondents. In fission, ∃-IDENT[F] constraints cannot force feature specifications to stay together on one output correspondent. Put differently, existentially defined featural faithfulness constraints allow features of a single input segment to surface in different locations in the output. I argue in chapter 3 that apparent feature movement should therefore be analyzed as segmental fission.

26

Struijke

Chapter 3 includes two case studies which exemplify ostensible feature movement, namely murmur movement in Sanskrit (section 3.3), and movement of aspiration and glottalization in Cuzco Quechua. Here I illustrate the main argument presented there with the Sanskrit case. In Sanskrit, a murmured segment cannot surface before an obstruent. Instead of losing the feature specification [+murmur] in the output form, the language sometimes opts to associate it with a different segment. (28.)

Sanskrit murmur throw-back /bud˙/ bod˙-ati b˙ot-syati

‘to wake (present)’ ‘to wake (future)’

I argue that /d˙/ in /bud˙/ undergoes fission in the future form and is in correspondence with the output segments [b˙] and [t]. The first correspondent preserves the murmur feature of the input segment, while the second preserves the place feature [cor]. Thus ∃-IDENT[+MURM]IO and IDENT[COR]IO are satisfied. The segment preserving the murmur feature coalesces with the root initial segment of the input /bud˙/, which contributes its own place feature. The existential character of ∃-IDENT[F] constraints allows for an account of apparent feature movement in which features are taken to be properties of segments, rather than autosegments that can enter into separate correspondence relations, and move independently of segments. In chapter 4, I argue that the proposed account is preferred to the alternative, because segment-based fission and coalescence are independently needed in phonology.

27

Existential Faithfulness

1.2.3.3 Preservation of feature specifications and dissimilation

The existential definition of ∃-IDENT[F] constraints also predicts segmental dissimilation. This is analyzed in chapter 3 and, like feature movement, it is illustrated in the case studies of Sanskrit and Cuzco Quechua. Languages often allow marked segments, but do not permit multiple instances of those segments in a given domain. For instance, many morphemes of Palauan and Ponapean (Austronesian) contain a single labial consonant, but never do we find morphemes that contain two or more labials (Suzuki 1998 and references quoted therein). When two such segments are contained in the input form, one must change or delete in the output form. I argue that these alternations are driven by segmental markedness constraints, in this case *LABIAL (cf. Alderete 1997; Alderete and Pater in prep.; Itô and Mester 1996; Suzuki 1998; Yip 1998b). This markedness constraint is able to emerge in the presence of two labial segments, because one of these segments can ensure faithfulness (i.e. preservation) of both. As mentioned, the repair strategy in dissimilation is sometimes deletion. I argue that deletion is apparent. Instead, coalescence takes place. An abstract example is given below:

(29.)

Dissimilation through coalescence [lab][-cont][+vc]

input:

output:

b1

b 1,2

[lab][-cont][+vc]

i

b2

i

a

a

[lab][-cont][+rc]

28

Struijke

Here, the coalesced output segment [b1,2] ensures preservation of both input segments and their feature specifications (so that ∃-MAX and all ∃-IDENT[F] constraints are satisfied). At the same time, markedness is improved, because the occurrence of marked labial segments is minimized. Thus, the combination of coalescence and *[F] constraints can ensure dissimilation. In fact, this does not only account for dissimilation of identical segments, it also accounts for OCP effects in morphological haplology (de Lacy 1998) and the behavior of geminates (Keer 1999). However, in many languages total coalescence is not the preferred repair, and two segments must surface in the output, one of which is unmarked. The existential definition of IDENT[F] plays a pivotal role in accounting for these patterns. Total coalescence as seen above might be disallowed because it would compromise syllable structure or incur a fatal violation of ∃-CONTIGUITY. In such cases, languages have to revert to 'partial coalescence.' Rather than merging entirely with another segment, a given input segment undergoes fission, and only one of its output correspondents coalesces with the other segment. This is depicted below: (30.)

Dissimilation through featural alternation (only input features are shown) [lab][-cont][+vc]

input:

output:

b1

b 1,2

[lab][-cont][+vc]

i

b2

i

d3

[lab][-cont][+rc]

a

a [-cont][+vc]

Because IDENT[F] constraints are existentially defined, featural faithfulness is achieved. The marked feature specification [lab] associated with the two relevant input

29

Existential Faithfulness

segments is preserved on some output correspondent, and the fact that one output segment is unmarked is irrelevant to evaluations of IDENT[lab]. Dissimilation of non-identical segments relies even more on the existential quantification of IDENT[F]. Consider, for instance, the mapping /p1m2/ - [p12 n2] below. Despite the fact that the two non-identical input segments coalesce, all their feature specifications can be preserved in the output, because there are two ouput segments that can carry them. Thus, in addition to the constraints mentioned above, featural faithfulness constraints may prevent total coalescence. (31.)

Dissimilation through featural alternation [lab][-cont][-vc]

input:

output:

[lab][+nas][+vc]

p1

m2

p 1,2

n3 [lab][-cont][-rc]

[+nas][+vc]

This concludes the introduction of ∃-IDENT[F] and its role in the different phenomena discussed in this dissertation. I now turn to the definitions of the remaining existential faithfulness constraints.

1.2.4 Preservation of adjacency and ordering relations

Preservation of adjacency and ordering relations of the input is regulated via ∃-LINEARITYIO and ∃-CONTIGUITYIO. Like other input-output faithfulness constraints, these constraints are existentially quantified. They demand that some correspondent of an

30

Struijke

input segment respects its linear order and adjacency relations with respect to other input segments. (32.)

∃-LINEARITYIO: Let seg-x, seg-y ∈ input be in the domain of ℜ, such that seg-x precedes seg-y; then there is some seg-x' ∈ output such that seg-xℜseg-x' which precedes some seg-y' ∈ output such that seg-yℜseg-y' When one segment precedes another segment in the input, some output correspondent of the first segment precedes some output correspondent of the second segmen

Note that this constraint penalizes string-internal epenthesis, but not deletion (cf. I-Contiguity which penalizes string-internal deletion, and O-contiguity, which penalizes string-internal epenthesis). (33.)

∃-CONTIGUITY IO: Let seg-x, seg-y ∈ input be in the domain of ℜ, such that seg-x is adjacent to seg-y; then there is some seg-x' ∈ output such that seg-xℜseg-x' which is adjacent to some seg-y' ∈ output such that seg-yℜseg-y'. When two segments are adjacent in the input, some output correspondent of one is adjacent to some output correspondent of the other

Given these definitions of ∃-CONTIGUITYIO and ∃-LINEARITYIO, the constraints are satisfied in most cases of reduplication, feature movement, and dissimilation. To facilitate exposition, I demonstrate the workings of these constraints in reduplication with structures in which each segment is represented by a number. The form below shows that partial reduplication does not normally disrupt preservation of underlying segmental adjacency and precedence relations.

31

Existential Faithfulness

(34.)

No ∃-LINEARITYIO and ∃-CONTIGUITYIO violations in partial reduplication input:

output:

RED +

1

2

1

1

2

2

3

3

4

4

Even though segment 2 of the reduplicant precedes segment 1 of the base in this form, ∃-LINEARITYIO is satisfied because some correspondent of 1 is followed by some correspondent of 2. Similarly, the fact that 2 in the reduplicant and 3 in the base are not adjacent is irrelevant. ∃-CONTIGUITYIO is satisfied because some correspondent of 2 is adjacent to some correspondent of 3. In fact, these constraints allow segmental scrambling, deletion and epenthesis in one member of the base-reduplicant pair. For example, ∃-CONTIGUITYIO is satisfied in the following Sanskrit forms (data analyzed by Steriade 1988; Gnanadesikan 1995, amongst others), where a segment is ‘skipped’ in the reduplicant, and in Semai (Diffloth 1976ab; Nelson 2000) where the reduplicant contains only material from the stem’s edges. Note, though, that CONTIGUITY on the base-reduplicant dimension is violated in both cases, because segments that are adjacent in the base are not adjacent in the reduplicant. (35.)

Satisfaction of ∃-CONTIGUITYIO in Sanskrit /svap/ sa:-svap ‘sleep’ (intensive) ˙ /grab / ga:-grap ‘seize’ (intensive) ˙ /vyad / va:-vyap ‘pierce’ (intensive)

(36.)

Satisfaction of ∃-CONTIGUITYIO in Semai /paya¯/ p¯ -paya¯ ‘appearance of being disheveled’ /cayEm/ cm-cayEm ‘contracted fingers of human animal, not moving’ /cruha:w/ cw-cruha:w ‘sound of waterfall, monsoon rain’

32

Struijke

∃-LINEARITYIO is satisfied in the following Rotuman forms (data analyzed by Mester 1986, McCarthy 1995, amongst others), where segments metathesize in the reduplicant. Again, the constraint on the base-reduplicant dimension is violated because segments in the base and reduplicant display a different linear ordering. (37.)

Satisfaction of ∃-LINEARITYIO in Rotuman /toka/ toak-toka ‘to cease’ /pure/ puer-pure ‘to decide’ /jika/ jiak-jika ‘narrow’

I would like to stress that reordering of segments and deletion or epenthesis in reduplicated forms as seen in these examples must be forced by markedness requirements. In the absence of such requirements it is prevented by base-reduplicant faithfulness constraints. We therefore do not expect unmotivated alternations of this kind, even though they are allowed by the existentially defined IO faithfulness constraints. Both feature movement and dissimilation involve segmental splitting and coalescence. In these patterns, ∃-LINEARITYIO and ∃-CONTIGUITYIO are satisfied also.11 This is true regardless of the fact that, in isolation, fission and coalescence cause violations of these constraints when they occur non-locally. Again, I will illustrate this point with abstract examples. (38.)

(a) input:

non-local fission /1 2

(b)

3 4 /

non-local coalescence /1

output: 1 2 4 3 4

1,4 2

11

2

3 4 /

3

I focus here on dissimilation which is resolved by featural change, rather than segmental deletion. In the latter case, ∃-LINEARITYIO and ∃-CONTIGUITYIO are violated, as the reader may verify by considering diagram (29).

33

Existential Faithfulness

In the fission example, underlying adjacency of segments is not preserved because 2 and 3 are not adjacent in the output (i.e., ∃-CONTIGUITYIO is violated). Note that the linear order of segments is preserved because some correspondent of /4/ follows [3] (∃-LINEARITYIO is satisfied). In the coalescence example, ∃-CONTIGUITYIO is violated twice because 3 or 4 are not adjacent in the output. In addition, ∃-LINEARITYIO is violated because no correspondent of /2/ and /3/ precedes a correspondent of /4/. In feature movement and dissimilation, ∃-LINEARITYIO and ∃-CONTIGUITYIO are satisfied. Some correspondent of each input segment is adjacent to some correspondent of each segment to which it is adjacent in the input. Also, all precedence relations are maintained. Consider the abstract form below. Because some correspondent of /5/ follows the correspondent of /4/, ∃-LINEARITYIO is satisfied. Likewise, because some correspondent of 5 follows the correspondents of /3/ and /4/, ∃-CONTIGUITYIO is satisfied and the fact that another correspondent of /5/ precedes them is irrelevant. (39.) Non-local feature movement input:

output:

/ 1

1 2,5 3

2

4

3

4

5 /

5

1.3 Fission and surface correspondence

Since the difference between universally quantified faithfulness constraints of classic Correspondence Theory and existentially quantified constraints is apparent primarily in segmental fission, all phenomena discussed in this dissertation involve such multiple correspondence. I have assumed above that fissioned segments in reduplicated

34

Struijke

words are subject to base-reduplicant faithfulness. In this section, I argue that segments resulting from non-reduplicative fission are also in a faithfulness relation with each other, dubbed 'surface faithfulness.' Whether surface faithfulness and base-reduplicant faithfulness can be united in a 'generalized surface relation' is an empirical question. I discuss this issue further at the end of this section and throughout the dissertation. Surface faithfulness was introduced in Struijke and de Lacy (2000), and finds its origins in Kang (2000), Kawu (1999), Kitto and de Lacy (1999), Pin)eros (2000), Raimy and Idsardi (1997), Rose and Walker (2000), and Walker (to appear). The proposal has functional grounding in the idea that common 'parenthood' of output segments must be recoverable from the output: the more alike surface correspondents are, the easier it is to determine that they are derived from the same input segment. Surface faithfulness constraints demand that fissioned segments be identical. They are violated in (distributing) diphthongization, dissimilation, and feature movement, because they all involve non-identical fissioned segments. The constraints thus play an important role in preventing such patterns from arising when they are not required for the satisfaction of markedness constraints. Although the main focus of this dissertation is on input-output faithfulness and its existential quantification, this role of surface faithfulness is discussed in the main chapters of this dissertation. The constraint IDENT[F]∑∑ is defined below.

(40.)

IDENT[F]∑∑ Let seg ∈ output be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an another output segment preserves the feature specification [αF] of that segment.

35

Existential Faithfulness

Evaluation of this constraint is identical to that familiar from base-reduplicant faithfulness. For instance, IDENT[+vc]∑∑ is satisfied in (41a) and (b) because the output correspondents related to the same input segment are identical with respect to voicing. It is violated in (41c).

(41.)

(a) satisfaction of IDENT[vc]∑∑ input:

output:

(b) satisfaction of IDENT[vc]∑∑

g

g

g

g

g

b

(c) violation of IDENT[vc]∑∑ input:

output:

g

g

k

(existential input-output faithfulness constraints satisfied)

Two issues that fall outside the cope of this dissertation, but must be addressed in future research are the following. First, in order to determine whether base-reduplicant faithfulness and surface faithfulness can be united, we need to compare the roles they play in reduplicative and non-reduplicative fission. Second, we need to determine whether surface faithfulness constraints are defined existentially or universally. I briefly lay out the issues here, starting with the first.

36

Struijke

The featural faithfulness constraints IDENT[F]BR and IDENT[F]ΣΣ behave exactly alike. They both demand featural identity of two segments that correspond to the same input segment. In the absence of fissioned segments they are vacuously satisfied. However, the constraint MAX only seems to perform a function in the BR relation. It is usually assumed that MAXBR automatically comes into play whenever the morpheme /RED/ is present in the input. The constraint demands that every segment in the base has a correspondent in the reduplicant. In effect, it can thus force overt realization of the reduplicative morpheme (i.e., it can force copying/fission, see section 2.4). In non-reduplicative contexts, MAXΣΣ has no role to play. Unification of BR and surface faithfulness can cover these functions of IDENT and MAX. Generalized IDENTΣΣ is relevant in fissioned segments of any kind, and MAXΣΣ is relevant in the presence of /RED/, just like MAXBR. Unification into generalized surface faithfulness predicts that fissioned segments behave alike in reduplicative and non-reduplicative contexts, all other things being equal. Thus, the issue of unification is in large part an empirical one. This dissertation remains agnostic to the issue, and I follow most contemporary literature on reduplication by utilizing BR faithfulness constraints. Whether surface faithfulness and BR faithfulness constraints (or generalized constraints) are quantified existentially is also in part an empirical question. Recall that we need to study cases of multiple correspondence in order to distinguish existential quantification from universal quantification. Thus, in order to resolve the issue at hand, we need to study the behavior of surface faithfulness constraints in one-to-many surface correspondence. That is, we need to examine forms in which an input segment corresponds to three or more output segments, so that each of these output segments is in a surface relation with multiple output segments. Since the main focus of this dissertation is on existential quantification in input-output faithfulness and its role in

37

Existential Faithfulness

phenomena involving two-way fission, I only discuss this issue briefly, and leave it for future research.12 On conceptual grounds it does seems likely that correspondence relations between input and output material fundamentally differ from surface correspondence. On the one hand I have argued that the function of input-output relations is to regulate preservation (an asymmetrical concept). On the other hand, surface faithfulness, and particularly base-reduplicant correspondence seem to regulate identity (i.e. symmetry) of the segments (in the spirit of Wilbur 1973 and classic Correspondence Theory). Indeed, input-output relations are inherently asymmetrical in nature. The output can be forced to be like the input, but the input is fixed and cannot be forced to resemble the output. The base-reduplicant relation is symmetrical, in that either member of the base-reduplicant pair can change or resist change to be like the other. Thus, it seems reasonable to assume that IO and BR and/or surface faithfulness constraints are differently quantified: IO constraints are existentially quantified, and BR constraints might be universally quantified. Even if future research shows that IO and BR/surface relations do indeed differ in quantification, I believe that they should both be seen as instances of faithfulness relations, adopting the insights of Correspondence Theory that different pairs of strings are related through correspondence and evaluated by faithfulness constraints.

1.4 Conclusion

This concludes the introductory chapter of this dissertation. Each of the following main chapters deals with a particular aspect of existential faithfulness and its 12

However, following Urbanczyk (1995), I argue in section 2.1.3 that double reduplication does not involve multiple BR correspondence, and hence does not constitute the state of affairs we are looking for.

38

Struijke

consequences for phonological theory. They are written such that they can be read independently of each other. Before I give the in-depth arguments and language-specific analyses, I present the proposed faithfulness constraints and their formal definitions in an appendix to this chapter.

39

Existential Faithfulness

Appendix I: Overview of existential faithfulness constraints

Under the proposal presented in this thesis, IO faithfulness constraints are satisfied when, roughly, each property of the input is preserved somewhere in the output. These properties are segments (∃-MAX IO), featural specifications of segments (∃-IDENT[F] IO), linear order of segments (∃-LINEARITY IO), and contiguity of segments (∃-CONTIGUITY IO). In addition I assume ∃-MORPHDIS, which demands that the separateness of segments belonging to different morphemes be preserved. INTEGRITY demands preservation of a segment’s integrity. The definitions given below convert the universally defined faithfulness constraints of Classic Correspondence Theory (McCarthy and Prince 1995, 1999) to existentially defined constraints. Here they are specified for the general input-output relation, but they hold true of any relation between the input and output (e.g. Positional Faithfulness such as root faithfulness), and hypothetically of relations along other dimensions, such as surface correspondence relation. Since I assume that segments (but not features) enter into correspondence relations, variables refer to segments. In the definitions below I therefore use variables such as ‘seg-x.’ ℜ stands for a correspondence relation.

∃ -MAX IO demands that each input segment have an output correspondent,13 as discussed in section 1.2.2.

13

As noted, all faithfulness constraints are unidirectional, going from the input to the output. Therefore, there is no constraint DEP. Instead, I adopt the markedness constraint M-SEG which penalizes epenthesis (M-SEG: Every segment belongs to a morpheme; McCarthy 1993). String-internal epenthesis normally also violates ∃-CONTIGUITY (defined below).

40

Struijke ∃-MAX IO: Let seg ∈ input; then there is some seg' ∈ output, such that segℜseg'. Every segment in the input has some correspondent in the output.

∃ -IDENT[F] IO demands that some output correspondent of a given input segment preserve a feature specification associated with that input segment (see section 1.2.3 for discussion). ∃-IDENT[F] IO: Let seg ∈ input be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an input segment preserves the feature specification [αF] of that input segment.

∃ -CONTIGUITYIO prohibits epenthesis of segments string-medially, while allowing it at edges. It also prohibits reordering of non-adjacent segments. The constraint is not normally violated by reduplication, feature movement or dissimilation (see section 1.2.4). This is due to the fact that it is existentially quantified: some correspondent of segment x needs to be adjacent to some correspondent of segment y (where x and y are adjacent input segments). ∃-CONTIGUITY IO: Let seg-x, seg-y ∈ input be in the domain of ℜ, such that seg-x is adjacent to seg-y; then there is some seg-x' ∈ output such that seg-xℜseg-x' which is adjacent to some seg-y' ∈ output such that seg-yℜseg-y'. When two segments are adjacent in the input, some output correspondent of one is adjacent to some output correspondent of the other

41

Existential Faithfulness

∃ -LINEARITYIO prohibits both metathesis and coalescence. Like ∃-CONTIGUITYIO it is existentially quantified and is therefore not normally violated by reduplication, feature movement or dissimilation (see section 1.2.4). ∃-LINEARITYIO: Let seg-x, seg-y ∈ input be in the domain of ℜ, such that seg-x precedes seg-y; then there is some seg-x' ∈ output such that seg-xℜseg-x' which precedes some seg-y' ∈ output such that seg-yℜseg-y' When one segment precedes another segment in the input, some output correspondent of the first segment precedes some output correspondent of the second segment.

I do not assume a constraint such as UNIFORMITY which would specifically prohibit coalescence, because LINEARITY already prohibits it (Pater 1999). Also, coalescence of non-identical segments incurs featural faithfulness violations (Keer 1999).

∃-MORPHDIS penalizes coalescence of segments belonging to different morphemes. (The constraint ∃-LINEARITYIO does not penalize such coalescence under the assumption that morphemes are unordered in the input. That is, in this type of coalescence, there is no linear order to be faithful to.) ∃-MORPHDIS (Morphemic Disjointness) Let seg-x ∈ input Morph1, and seg-y ∈ input Morph2, and seg-x', seg-y' ∈ output, and seg-xℜseg-x', seg-yℜseg-y'. Then there is some seg-x'≠ seg-y'. (McCarthy and Prince 1995 (adapted).) The disjointness of input morphemes is preserved in the output. (segments belonging to different morphemes do not coalesce)

42

Struijke

Highranking of this constraint can prevent feature movement and dissimilation across morpheme-boundaries, because it would involve coalescence of segments that are part of different morphemes. I hypothesize, by analogy with the constraints discussed above, that this constraint is existentially quantified. In order to see the effect of this existential quality, we need to study a specific state of affairs. I illustrate this below with a hypothetical form. Two morphemes must each contain a segment that undergoes fission. In the example they are /i/ and /a/. Existential MORPHDIS is then satisfied if one correspondent of /i/ and one correspondent of /a/ are disjoint. The other correspondents are then allowed to coalesce without violating existential MORPHDIS.

(42.)

input:

s

p

output:

s i1 p

i1

+

a2 k

e1,2

t

k a2 t

INTEGRITY penalizes fission (as in classic Correspondence Theory). The constraint is violated in diphthongization, feature movement, dissimilation, and phonological reduplication, and any other phenomena involving fission. I argue that GEN always establishes multiple correspondence in the presence of a reduplicative morpheme, and therefore INTEGRITYIO is not relevant in morphological reduplication. Since this constraint specifically penalizes fission, quantification is not an issue (recall that quantification is only an issue if fission is actually established).

43

Existential Faithfulness

INTEGRITYIO: For seg-x ∈ input and seg-y, seg-z ∈ output, if seg-xℜseg-y and seg-xℜseg-z, then seg-y = seg-z No fission. The integrity of an input segment is preserved in the output.

I suggest in the appendix to chapter 3 that this constraint is evaluated gradiently to account for locality effects. It is satisfied when no fission takes place; violated once when fissioned output segments are adjacent; and multiple times when the output correspondents of an input segment are separated by other segments (with an additional violation for each additional intervening segment).

44

Struijke

2 Reduplicative TETU

The core idea developed in this dissertation is that phonological faithfulness requires preservation of underlying elements in surface forms, rather than identity of input and output. To formally implement this idea, I propose that input-output (IO) faithfulness constraints are existentially quantified, such that they require each underlying element to have some faithful correspondent in the surface form. This is a departure from classic Correspondence Theory (McCarthy and Prince 1995), in which faithfulness constraints are universally quantified and require all output elements to be faithful to their input correspondent. The effects of existentially quantified faithfulness are visible when an input segment has more than one correspondent in the surface form. I argue in this chapter that such multiple correspondence, or ‘fission’, occurs in reduplication. In a reduplicated word, underlying segments typically undergo fission and have two correspondents in the output: one in the base and one in the reduplicant. I argue that segments in both members of the base-reduplicant pair are subject to IO faithfulness constraints. Because these constraints are existentially quantified, however, only one of these strings needs to be like the input, even when IO faithfulness constraints are undominated. I show that the ranking of other constraints in a grammar determines whether the other member stays unchanged, or alternates to (better) satisfy some lower-ranked markedness constraint. In the latter case, the result is ‘The Emergence of The Unmarked,’ henceforward abbreviated as TETU (McCarthy and Prince 1994a - for precursors see McCarthy and Prince 1986; Shaw 1987; Steriade 1988). In reduplicative TETU, a markedness constraint is inactive in unreduplicated words due to dominant IO faithfulness, but emerges in reduplication such that it forces an alternation in one member of the base-reduplicant pair. Thus, even though unreduplicated words are marked, reduplicated forms are less marked.

45

Existential Faithfulness

The main focus of this chapter is on The Emergence of The Unmarked, and the factors determining which output string preserves input information, and which undergoes a change. An important result of this chapter is that either the base or the reduplicant can alternate in TETU. This is a departure from McCarthy and Prince 1994a, who argue that it is always the reduplicant that changes. This result was reported earlier in Struijke 1998. The more general idea pursued in this chapter pertains to the nature of reduplication. Underlying this specific study of reduplicative TETU is the idea that the base and reduplicant are generated in parallel. This is of course a consequence of the basic architecture of standard Optimality Theory, but I take the idea to its extreme by claiming that the base is in no sense prior to the reduplicant (see also Inkelas and Zoll 2000). As mentioned earlier, I claim that both strings are related to the input and are equally subject to IO faithfulness constraints. Thus, the reduplicant is not 'copied' from the base, but is directly mapped from the input. As a result, the reduplicant can contain input information that is not preserved in the base. Even though I assume that reduplication is established through fission along the input-output dimension the base-reduplicant correspondence relation (McCarthy and Prince 1995) or a more general surface faithfulness relation is anything but superfluous. It is crucial in the theory presented in this chapter because (typically) it accounts for any identity between base and reduplicant. Like IO faithfulness, BR/surface faithfulness appears to assign equal status to these strings. Although the base and reduplicant are equivalent with respect to both this type of correspondence and IO correspondence, they tend to differ morphologically. Reduplicants are often affixes while bases are usually roots. Therefore only bases tend to be subject to Positional Faithfulness requirements specific to root morphemes (root faithfulness; Beckman 1997). The faithfulness dimensions relevant in reduplication are laid out in section 2.1.

46

Struijke

Section 2.2 introduces ‘Reduplicant TETU.’ In this type of TETU, the base is faithful and preserves marked input material, while the reduplicant changes to become unmarked. It is familiar from the literature following McCarthy and Prince (1994a), but is accounted for in a different manner in order to allow for an explanation of 'Output TETU,' which is introduced in section 2.3. In Output TETU the emerging markedness constraint14 determines in which part of the output word a TETU alternation takes place. I show that either the base or the reduplicant can be affected. Both types of reduplicative TETU are illustrated with a case study of Kwakwala reduplication. At the ends of both section 2.2 and section 2.3, I draw parallels between TETU in reduplicated words and unreduplicated words, and argue that reduplicative affixes are more likely to be involved in TETU alternations than lexically specified affixes. In the remaining sections of this chapter, I point to some important predictions of the proposed theory of faithfulness with respect to reduplication, many of which are illustrated by Kwakwala reduplication. In section 2.4, I show that an existential faithfulness constraint and a markedness constraint can conspire to either prevent phonological realization of a reduplicative morpheme or, conversely, force reduplication in the absence of such a morpheme (i.e. ‘phonological reduplication’). Section 2.6 focuses on the different roles played by the emerging markedness constraints in Reduplicant and Output TETU. I explain why the size of the domain evaluated by the markedness constraint may determine which type of TETU occurs (i.e. which member of the base-reduplicant pair is able to change). Section 2.7 shows that input segments and input feature specifications can be distributed between the base and the reduplicant such that these strings are equally unlike the input. Section 2.8 shows that reduplicative infixation may be less frequent than previously assumed. In section 2.10, the proposed

14

An ’emerging markedness constraint’ is a constraint that is (typically) inactive in unreduplicated words of a given language, but active in reduplicated words.

47

Existential Faithfulness

view of reduplication is compared with other theories, in particular classic Correspondence Theory (McCarthy and Prince 1995). Section 2.11 concludes this chapter.

2.1 Faithfulness relations in reduplication

The diagram in (1) displays the correspondence relations to which reduplicated words are subject. They are input-output faithfulness and base-reduplicant faithfulness. I have included the positional faithfulness relation root faithfulness, because it usually distinguishes bases and reduplicants, and because it plays a crucial role in Reduplicant TETU. Following a brief description of the terms ‘base’ and ‘reduplicant’, each of these three correspondence relations is discussed in turn. (43.)

Correspondence relations relevant in reduplication input:

/RED + Root/ Input-Output Faithfulness Root Faithfulness

output:

Reduplicant

Base

Base-Reduplicant Faithfulness

Because root faithfulness typically selects the base (or part of the base), but not the reduplicant in a reduplicated word, it is useful to differentiate these strings. Following most research on reduplication, I assume that the reduplicant is the phonological realization of a segmentally empty reduplicative morpheme: /RED/ (see e.g. McCarthy and Prince 1986; but see e.g. Inkelas and Zoll 2000 and Yip 1998b for a different view). Given 'consistency of exponence' (McCarthy and Prince 1993b) the output reduplicant is

48

Struijke

the exponent of /RED/. Thus, if the morphological identity of /RED/ is an affix, the morphological status of the reduplicant is an affix also, even when segments in the reduplicant are in correspondence with the input root. The base is the string that complements the reduplicant. It includes root material and any segmentally specified affixes.

2.1.1 Broad input-output correspondence15 and Output TETU

Input-output correspondence relates elements of the input to elements in the output (McCarthy and Prince 1995). I argue that it is a general relation which is insensitive to the morphological make-up of words. In unreduplicated words, each input element usually has one output correspondent, as illustrated below (lines indicate correspondence). (44.)

IO correspondence in unreduplicated words Input:

Output:

/t a b i-s/

t a b i s

The idea that entire inputs are related to entire output words also applies to reduplicated words. Thus, I claim that IO correspondence relates inputs to reduplicated words as a whole, including both the base and the reduplicant (cf. Classic Correspondence theory which assumes an input-base relation only ('Basic Model of Correspondence') or separate input-base and input-reduplicant relations ('Full Model of Correspondence'). Under the present proposal, input segments split and have a 15

In Struijke 2000a I referred to this relation as ‘Word Faithfulness’ (but see Borowsky 2000 for a different use of that term).

49

Existential Faithfulness

correspondent in the base and a correspondent in the reduplicant, as depicted below. This hypothetical form is an instance of full reduplication and therefore each input segment undergoes fission. (45.)

Reduplication involves fission Input:

Output:

/t a b i /

t a b i

t a b i

Segments in the two copies of a reduplicative word are of equal status with respect to the input-output relation and are therefore equally subject to IO faithfulness constraints. This proposal is not new: similar ideas are developed in Struijke (1997) (where I dubbed the general IO relation 'broad IO'), and were independently presented in Raimy and Idsardi (1997), Spaelti (1997), Fitzgerald (1998, 1999), and Yip (1998a, 2000) (cf. Cole 1997). All these proposals assume fission in reduplication (see also Inkelas and Zoll 2000). Section 2.10 discusses how these proposals differ from the ideas presented in this dissertation. The main innovation of this dissertation is that faithfulness constraints relating input elements to output elements are existentially defined. (46.) Existential Broad Input-Output Faithfulness • each segment of the input must be preserved somewhere in the output (∃-MAXIO) • some output correspondent of a given input segment must preserve the feature specifications associated with that input segment (∃-IDENT[F]IO)

In unreduplicated words, there is typically one chance to satisfy IO faithfulness constraints. However, in reduplicated words, input segments potentially have two output correspondents, so in these words there are two chances to satisfy IO faithfulness

50

Struijke

constraints. Thus, when IO constraints are undominated, unreduplicated words must be exactly like the input, but in reduplicated words only one member of the base-reduplicant pair needs to be like the input. As mentioned earlier, constraint interaction determines whether the other member remains unchanged or undergoes an alternation to better satisfy markedness requirements. When base-reduplicant (BR) faithfulness constraints are sufficiently high-ranking, segments in the base and reduplicant must be alike, and no change takes place. However, when constraints on the BR relation are low-ranking, one of the members in a base-reduplicant pair can change or delete material in response to some markedness constraint. This is ‘The Emergence of The Unmarked’ (‘TETU’ McCarthy and Prince 1994a): undominated input-output faithfulness constraints ensure that a lower-ranked markedness constraint is inactive in unreduplicated words, but the markedness constraint emerges and forces alternations in reduplicated words. Because the base and reduplicant are of equal status with respect to IO correspondence, broad input-output constraints are indifferent as to whether it is the base or the reduplicant that preserves input material. As a result, TETU can affect either string. In section 2.3 this is demonstrated by a certain type of reduplication in Kwakwala (Wakashan; Boas 1947; Zec 1988; Rodier 1989). The broad interpretation of the input-output correspondence relation and the existential nature of constraints thereon account for the fact that a reduplicant can ensure satisfaction of IO faithfulness constraints, while a base can be affected by TETU. If one assumed fission (broad IO correspondence) but universal faithfulness, TETU could not occur because output correspondents in both the base and the reduplicant should be like the input element. If one assumes that only bases are subject to input-output faithfulness (‘narrow’ IO correspondence) (McCarthy and Prince 1995, 1999), undominated constraints on this relation force these strings to be faithful, and only reduplicants can be affected by TETU.

51

Existential Faithfulness

2.1.2 Root faithfulness and Reduplicant TETU

Root faithfulness is a type of Positional Faithfulness (Beckman 1997). Roots are believed to participate in a special faithfulness relation in order to account for the fact that these morphemes are more likely to preserve underlying distinctions than affixes. Constraints on this relation require elements of the morphological input root to be faithfully realized in the output root. They are existentially quantified because they evaluate corresponding segments in the input and output, and hence demand preservation of input elements in the surface form (see chapter 1). (47.)

Existential Root Faithfulness: • Every segment in the input root has some correspondent in the output root (∃-MAXRT) • Some root output correspondent of each root input segment preserves the feature specifications associated with that segment (∃-IDENT[F]RT)

Bases of reduplicated words typically consist of a root (or contain root material). Therefore bases (or root segments in bases) are subject to root faithfulness in addition to general input-output faithfulness. In partial reduplication, reduplicants are usually affixes, and hence they are only subject to IO constraints.16,17 Since bases are subject to a superset of faithfulness constraints compared to reduplicants, they are more likely to be faithful to the input than reduplicants. As a result, most cases of TETU involve an alternation in the reduplicant. When root faithfulness constraints ensure that bases are 16

Even when they are in correspondence with input roots, affixal reduplicants are not subject to root faithfulness, because root faithfulness only holds over segments that are affiliated with both the input root and the output root. Given an affixal reduplicative morpheme in the input, the morphological status of the output reduplicant is an affix, because the reduplicant is the phonological realization of that morpheme. Correspondence relations do not affect the intrinsic morphological status of a morpheme (consistency of exponence; McCarthy and Prince 1993b). 17 Urbanczyk (1996) shows that reduplicants are sometimes roots. Such reduplicants are subject to both IO and root faithfulness constraints. In this dissertation I focus on forms containing affixal reduplicants.

52

Struijke

faithful and consequently only reduplicants change to satisfy an emergent markedness constraint, we have ‘Reduplicant TETU.’ Like Output TETU, Reduplicant TETU will be illustrated with Kwakwala reduplication (section 2.2).

2.1.3

Base-reduplicant correspondence

I suggested in chapter 1 that base-reduplicant faithfulness should be subsumed under general surface faithfulness. However, awaiting further research, I use base-reduplicant faithfulness throughout this chapter. I assume that GEN requires segmental splitting in overt morphological reduplication. That is, the presence of a reduplicative morpheme in the input causes GEN to establish fission of input segments, and hence overt morphological reduplication always involves multiple correspondence.18 Even though reduplication is not crucially established through copying of the base, a relation between the output base and the output reduplicant is important in the proposed framework. (48.) Reduplication involves fission of input segments and BR correspondence Input:

Output:

/t a b i /

t a b i

t a b i

18

Thus, a mapping such as /RED+kat/ - [k-at], where the initial consonant is the phonolgical exponence of RED is prohibited. Note that existential IO faithfulness constraints do not penalize this form. In addition to the formal ban on this mapping via universal requirements of GEN, functional considerations explain why such structures are not found. In the output [k-at], the reduplicant is not recoverable, and the form would cause learnability problems.

53

Existential Faithfulness

BR correspondence relates the output reduplicant to the output base, and BR constraints demand faithfulness between these strings. In all cases of The Emergence of The Unmarked, at least some BR faithfulness constraint must be low-ranking, because, by definition, one of the base-reduplicant strings undergoes a change. Conversely, only faithfulness constraints along the BR dimension can force identity of base and reduplicant (i.e. absence of TETU).19 In fact, any degree of similarity between the base and the reduplicant is typically due to BR faithfulness requirements. In the absence of BR faithfulness constraints (or general surface faithfulness constraints), we would expect rampant TETU. Because IO faithfulness is existentially defined, only one string in the reduplicated word needs to be like the input. Thus, unrestrained by BR requirements, markedness constraints would be able to affect the reduplicated word heavily, even to such extremes that the reduplicant would not be realized at all.20 However, BR constraints may be sufficiently low-ranking in a language that phonological realization of a reduplicative morpheme is indeed prevented, and we will see that this is sometimes the case in Kwakwala (section 2.4). In addition to preventing TETU alternations, BR constraints are required to account for over- and underapplication phenomena in which base and reduplicant are unexpectedly identical (McCarthy and Prince 1995 after Wilbur 1973). In these phenomena, a configuration disfavored by a markedness constraint is found in one member of the base-reduplicant pair, yet both members undergo an alternation (in overapplication), or neither does (in underapplication). BR faithfulness requirements formalize Wilbur's (1973) idea that these unexpected patterns are due to a pressure for the base and the reduplicant to be identical. 19

Alternative constraints could also demand identity of base and reduplicant, such as REPEAT (Yip 1995, 1998b; Rose 2000b), RHYME and ALLITERATE (Yip 1998a, 2000). Raimy & Idsardi (1997) and Kawu (1999) posit a surface relation in reduplicated words between output correspondents of a fissioned input segment. 20 Provided that MORPHREAL (‘realize a morpheme in an overt and detectable manner’) is low ranking.

54

Struijke

Whether BR constraints (or general surface faithfulness constraints) are existentially quantified, by analogy with IO constraints, or whether they are universally quantified, as proposed by McCarthy and Prince (1995), remains an open question (as I pointed out in chapter 1). In order to distinguish these alternatives in reduplication, we would need to study cases of multiple correspondence along the BR dimension. This is because universally and existentially quantified faithfulness can be distinguished in instances of fission only: in cases of one-to-one correspondence, identity and preservation are equivalent (see chapter 1). Rose 2000a argues that BR multiple correspondence is found in double reduplication, where a reduplicated word contains two reduplicative morphemes. (49.)

Fission along the BR dimension in double reduplication (hypothetical form) [t1a2b3i4 - t1a2b3i4 - t1a2b3i4] Output:

t a b i

t a b i

t a b i

However, there is reason to believe that double reduplication involves two separate instances of base-reduplicant correspondence, one for each reduplicant. Urbanczyk (1995) demonstrates that different reduplicative morphemes in a language show different degrees of faithfulness in their phonological exponents, regardless of whether they are the only reduplicative morpheme in the word, or one of several. She presents data from Lushootseed (Salish), in which the reduplicant that encodes distributive meaning is a closed syllable, while the diminutive reduplicant is an open syllable. These different morphemes must be subject to different types of base-reduplicant correspondence. BR faithfulness constraints relevant to the distributive must dominate NOCODA to allow codas

55

Existential Faithfulness

in these reduplicants. In turn, NOCODA must dominate BR faithfulness constraints relevant to the diminutive to ensure that these reduplicants are an open syllable. When a single word contains both types of reduplicants, both faithfulness relations come into play. One relates the base to the distributive reduplicant, the other relates the base to the diminutive reduplicant. Thus, no fission takes place where a single type of BR correspondence relates the base to multiple reduplicants. Rather, reduplicants relate to the base via separate types of BR correspondence. In fact, Urbanczyk presents evidence that the innermost reduplicant copies root material, while the outermost reduplicant copies the innermost reduplicant (such that the base of this reduplicant itself contains a reduplicant). This is depicted below for our hypothetical form: (50.) Different instances of BR correspondence in double reduplication [taabbcid - t1aa 2bb 3ci4d - t1a2b3i4] Output:

t a b i

t a b i

red. 1 red. 2

t a b i

base of red. 1

base of red. 2

Because double reduplications do not seem to provide examples of fission along the base-reduplicant correspondence dimension, they cannot serve as a testing ground for the question of whether BR constraints are existentially quantified.

2.1.4 Summary

I have argued that three faithfulness relations play an important role in reduplication: IO faithfulness, root faithfulness and BR faithfulness. The first two relate underlying and

56

Struijke

surface forms and are existentially quantified. Both types of underlying-surface relations are essential in accounting for the two different types of The Emergence of The Unmarked to be studied in this chapter, namely Output TETU and Reduplicant TETU. In the following two sections, I discuss each type of TETU in turn and illustrate them with a particular kind of Kwakwala reduplication. In this instance of reduplication a prefixal reduplicant co-occurs with one of two lexical suffixes: [m’u:t] ~ [mu:t]21 ‘useless refuse’ or [-(g)i:sa:we:/] ‘left over’. Examples are at first restricted to forms containing the former suffix, although reduplicated forms containing either suffix behave the same with respect to Output and Reduplicant TETU. Unless indicated otherwise, data containing the suffix [-mu:t] are taken from pages 339-340 of Boas’ (1947) grammar.

2.2 Reduplicant TETU: Kwakwala case study

Reduplicant TETU in Kwakwala involves the distribution of obstruent codas in reduplicated words. The language shows an asymmetry between obstruents and sonorants: only sonorants contribute to syllable weight in coda position (Zec 1988; Yip 1992). Obstruents are always non-moraic (see evidence below) and when in coda, they violate the constraint WEIGHTbyPOSITION (WxP). (51.)

WEIGHTbyPOSITION (WxP): Coda consonants must be moraic (after Hayes 1989)

Because obstruents generally surface in coda without being moraic, this constraint is inactive in non-reduplicative forms.22 However, it is active in reduplicative forms, because

21

I ignore the variation in glottalization in this suffix, as it has no bearing on the analysis. In fact the constraint is only partially inactive in unreduplicated words, because it forces sonorants to be moraic (as discussed below). 22

57

Existential Faithfulness

it prevents non-moraic obstruents from surfacing in reduplicants. Thus the constraint emerges in reduplication. This is TETU. The emergent pattern is shown below: (52.)

Obstruents are not allowed to surface in reduplicants k'a˘xW k ’ a˘-k’axW-m’u:t * k ’ a˘xW-k’axW-m’u:t ts’a:s ts’a:-ts’´s-m’u:t *ts’a:s-ts’´s-m’u:t te:¬ te:-ta¬-m’u:t *te:¬-ta¬-m’u:t

‘shavings’ ‘old eel-grass’ ‘remains of bait’

I show in this section that obstruent deletion is blocked in bases by root faithfulness. TETU alternations are restricted to reduplicants and therefore this is an instance of Reduplicant TETU.

2.2.1 Unreduplicated words and the moraic status of codas

Evidence that Kwakwala sonorant codas are moraic and obstruent codas are non-moraic comes from the assignment of main stress, and the syllable canon (Zec 1988; Yip 1992).23 The clash avoidance pattern analyzed in section 2.3 below provides more evidence for this claim. In Kwakwala, eight syllable types surface, as given below. (O stands for obstruent, S for sonorant; C and V stand for onset consonant and vowel respectively. The relevant syllable is underlined).

23

Zec (1988) argues that glottalized sonorants should be included in the non-moraic class of Kwakwala coda segments. However, I will argue in section 2.3.2.2 that these segments do not surface in coda position. Therefore the question of their moraicity is moot.

58

Struijke

(53.)

Kwakwala syllable canon syllable type example CV b´.xo˘t CVV dze˘.daqW CVO Gas.xa˘ CVOO24 ha˘.l’a˘.maxs.ta˘ CVVO ya˘x.k’a˘ CVVOO ts’´.da˘xs.t´.w´.l´.la˘ CVS d´l.xa˘ CVSO t’ ´ls.ta˘s

gloss ‘torch’ ‘milky sea eggs’ ‘to carry on fingers’ ‘to eat quickly’ ‘to hop on one foot’ ‘woman representative’ ‘damp’ ‘to eat crabapples’

No word-medial syllables contain a long vowel followed by a sonorant, or a short vowel followed by two or more sonorants: *CVVS; *CVSS. This suggests that syllables are at most bimoraic and sonorants contribute to weight. Obstruents do not contribute to weight and can therefore co-occur with sonorants or other obstruents in a coda and follow a long vowel without exceeding the bimoraic maximum25 . Further evidence for the moraic status of consonants comes from the assignment of main stress. Through inspection of the open-syllabled words in (12) we can conclude that main stress falls on the leftmost heavy syllable. (54.)

Main stress is initial and weight sensitive ha@˘.dza˘.pa˘.ma˘ ‘yarrow’ x´.sa@˘.¬a˘ ‘those who have disappeared’

The forms in (13) below show that initial short-vowelled syllables closed by a sonorant attract main stress (13a), but when such a syllable is closed by an obstruent it does not attract main stress (13b). This again indicates that obstruents do not contribute to weight.

24

Onset clusters are prohibited, but coda clusters are allowed (Wilson 1978; see also Boas (1947) who does not report this generalization, but claims that word-initial clusters do not occur, but final clusters do). 25 Sonorants do occasionally follow a long vowel or coda sonorant word-finally. Apparently, these segments are extrametrical.

59

Existential Faithfulness

Initial syllables closed by obstruents attract main stress only when the vowel is long (form (13c)).

(55.)

Stress assignment in closed syllables a. xW´@l.dzo˘s ‘Hexagrammus superciliaris’ b. p´X.d´@m * p´@X.d´m ‘time’ c. me@˘x.ts’a˘s ‘dreamer’

The moraicity of coda sonorants is forced by the constraint WEIGHTbyPOSITION, and the non-moraicity of obstruents is due to a constraint against the association of moras to obstruents. The latter is part of the constraint family *µ/SEG (Morén, 1997).26 (56.)

*µ/OBSTR: obstruents are non-moraic (shorthand for *µ/t >> *µ/v, etc.) *µ/SON: sonorants are non-moraic (shorthand for *µ/m >> *µ/r, etc.) *µ/VOWEL: vowels are non-moraic (shorthand for *µ/i >> *µ/a, etc.)

These constraints are ranked universally to implement Zec’s (1988) finding that the more sonorous a coda segment, the more likely it is to be moraic:

(57.)

Universal ranking of constraints against segment/mora associations (Morén, 1997; cf. Peak and Margin Hierarchy, Prince and Smolensky 1993) *µ/OBSTRUENT

>>

*µ/SONORANT >>

*µ/VOWEL

Kwakwala sonorants are moraic because WEIGHTbyPOSITION dominates *µ/SONORANT. As candidate 3 in tableau (16) shows, both of the markedness 26

Cf. positive constraints, which license mora-segment associations (Gnanadesikan 1995).

60

Struijke

constraints could be satisfied if coda sonorants delete. However, such a repair is prohibited by high-ranking ∃-MAXIO. (58.) Sonorants are obligatorily moraic in a non-reduplicative context /CVµS/ ∃-MAXIO WxP *µ/SON µ µ * 1 CV S µ 2 CV S *! 3 CVµ *!

WEIGHTbyPOSITION cannot force coda obstruents to be moraic, because it is outranked by *µ/OBSTR. Nor can WxP force deletion and avoid violations of either of these constraints, because ∃-MAXIO ranks above WxP.27 (59.)

Obstruents are obligatorily non-moraic /CVµO/ ∃-MAXIO *µ/OBSTR µ µ 1 CV O *! µ 2 CV O 3 CV *!

WxP *

Due to the Kwakwala ranking of WxP within the universal hierarchy of constraints against segment/mora associations (*µ/OBSTR >> WxP >> *µ/SON), sonorant codas must be moraic and obstruent codas must be non-moraic in Kwakwala. Coda consonants are prevented from deleting by high-ranking ∃-MAXIO. With this background, I turn now to the distribution of obstruent codas in reduplicated words.

27

Given Richness of the Base (Prince and Smolensky 1993; see chapter 1), we need to assume the possibility of inputs in which obstruents are underlyingly moraic. Such forms will not surface faithfully due to the ranking *µ/obstr >> ∃-IDENTWeight. For a definition of the latter constraint see section 2.3.1.

61

Existential Faithfulness

2.2.2 The emergence of WEIGHTbyPOSITION in reduplicated words

Even though WEIGHTbyPOSITION cannot force coda deletion in unreduplicated words, in reduplicated words it is able to trigger deletion of obstruent codas in one member of the base-reduplicant pair.28 This is because undominated existential MAXIO only demands one output correspondent for each input segment. The configuration in (18) shows an input and its optimal reduplicated output form. Because the input obstruent is present in the base, existential MAXIO is satisfied. The reduplicant is able to respond to the lower-ranked markedness constraint WEIGHTbyPOSITION by deleting its coda obstruent. Thus, in reduplication it is possible to minimize WEIGHTbyPOSITION violations without violating ∃-MAXIO. (60.)

Fission in reduplication and TETU input:

output:

RED ts’ a: s + m’ u˘ t

ts’ a:

ts’ ´ s m’ u˘ t

The argument is summarized in tableau (19). Candidate 1 parses the root obstruent of the input into both the base and the reduplicant, satisfying ∃-MAXIO. WEIGHTbyPOSITION is violated three times, because the reduplicant, root, and suffix syllables all end in obstruents, which are necessarily non-moraic. Candidate 2 deletes the obstruent in the reduplicant and hence only incurs two violations of WxP. This candidate is optimal because deletion does not result in a violation of ∃-MAXIO (both input obstruents are preserved in the base). 28

This discussion addresses the behavior of obstruent codas only. The behavior of sonorant codas in reduplication will be discussed in section 2.3.

62

Struijke

(61.)

Obstruents cannot be parsed in reduplicant codas (L indicates obstruent codas) (ranking established above) ∃-MAXIO *µ/OBSTR /RED + ts’a µµs + muµµt/ 1

WxP

ts’a µµs - ts’´µs - m’uµµt L

L

***!

L

2 ts’aµµ - ts’´µs - m’uµµt L

**

L

So, high-ranking ∃-MAXIO forces each input obstruent to be parsed in the output, and WxP ensures that it appears in coda only once. It is important to note that neither of these constraints indicates a preference for the location in which the obstruent is parsed. That is, IO faithfulness constraints and WxP do not differentiate a candidate in which the obstruent is present in the reduplicant from a candidate in which it is present in the base. This is illustrated in the tableau below. (62.)

∃-MAXIO and WxP do not indicate a preference for a TETU alternation in base or reduplicant ∃-MAXIO *µ/OBSTR WxP /RED + ts’aµµs + muµµt/ ** 1 ts’ a@µµ - ts’´µs - m’uµµt µµ

L

µ

µµ

L

2 ts’ a@ s - ts’´ -m’u t L

**

L

There are two reasons why the two candidates in this tableau fare equally well. First, WxP evaluates base and reduplicant syllables individually, and demands that both have moraic codas. Second, base and reduplicant are of equal status with respect to input-output faithfulness, so ∃-MAXIO is satisfied no matter where the obstruent appears. However, the two candidates are crucially differentiated by root faithfulness. Root faithfulness constraints are violated when a segment is deleted in the base, but satisfied when a segment is deleted in the reduplicant. As shown in the tableau below, a candidate

63

Existential Faithfulness

which shows a TETU alternation in the reduplicant incurs a subset of violations compared to a candidate which shows a TETU alternation in the root (i.e. base-initial) syllable. Put differently, candidate 1 harmonically bounds candidate 2: there is no ranking of these constraints under which candidate 2 will be optimal. (63.) ∃-MAX RT: Let seg ∈ input root; then there is some seg' ∈ output root, such that segℜseg' Every segment in the input root has some correspondent in the output root. (64.)

Emergence of the Unmarked affecting the reduplicant29 *µ/OBSTR ∃-MAXI WxP /RED + ts’aµµs + µµ O mu t/ ** 1 ts’ aµµ - ts’aµs - m’uµµt ** 2 ts’ aµµs - ts’aµ - m’uµµt

∃-MAXRT

*!

Since the base and reduplicant are not identical in The Emergence of The Unmarked, BR faithfulness constraints must be low-ranking.30

29

This tableau shows that a violation of a root faithfulness constraint does not necessarily entail a violation of its more general input-output faithfulness counterpart (cand. 2). This is due to the fact that fission takes place in reduplication, and the hypothesis that these constraints are existentially quantified. If one assumes that faithfulness constraints are universally quantified, a violation of the specific constraint does entail a violation of the general constraint. For a discussion on the interaction of special and general constraints, see Prince 1997. 30 If BR constraints are universally quantified to demand identity between base and reduplicant (rather than preservation of base material in the reduplicant) as described in section 1.3, they are symmetrical. We may then assume that they are bi-directional, and include both MAXBR and DEPBR. Note that these constraints also differentiate the two candidates in tableau (22). However, the fact that deletion occurs in the reduplicant rather than in the base does not seem to be caused by the ranking DEPBR >> MAXBR. Although empirically adequate, such a ranking does not reflect the generalization that roots tend to be more faithful than reduplicants and other affixes. In fact, it seems plausible that base and reduplicant are equal with respect to BR faithfulness, and are only differentiated by root faithfulness. That is, perhaps deletion in the base and deletion in the reduplicant are not differentiated by BR faithfulness, and DEPBR and MAXBR may be one constraint (cf. IDENT[F]BR constraints, which equally punish an alternation in the base and an alternation in the reduplicant). Further research is needed to investigate this issue (see Spaelti (1997) for arguments against the constraint DEPBR). In this dissertation I will assume the standard constraints MAXBR and DEPBR.

64

Struijke

In this section, I explained Reduplicant TETU, relying on the hypotheses that reduplication involves fission and that input-output constraints are existentially defined. These assumptions are not crucial in accounting for this type of TETU. McCarthy and Prince (1994a, 1995) explain TETU alternations affecting the reduplicant by assuming that only the base is related to the input, and the reduplicant is unfaithfully related to the base. A comparison of the two approaches is given in section 2.10.1. Section 2.3 focuses on the second type of TETU, in which either the base or the reduplicant can be affected. An account of such Output TETU crucially relies on the notions of broad IO correspondence and existential faithfulness. I show in section 2.10.1 that classic Correspondence Theory cannot capture this phenomenon. Before turning to Output TETU, I address the typological predictions that follow from the existential faithfulness account of Reduplicant TETU.

2.2.3 Typological predictions: TETU in reduplicative and lexical affixes

Under the present proposal, reduplicant TETU affects reduplicants because they are affixes. This raises the question of whether affixes whose segmental content is lexically specified are subject to TETU in languages where affixal reduplicants show emergent unmarkedness. In this section, I argue that we should find languages in which TETU affects reduplicative affixes only, lexical affixes only, or reduplicative affixes and lexical affixes alike. In Kwakwala, the emergent markedness constraint WEIGHTbyPOSITION affects reduplicative affixes, but not lexically specified affixes. That is, obstruent codas are deleted in reduplicative affixes, but not in lexical affixes, as illustrated by the form ts’a:-ts’´s-m’u:t 'old eel-grass.' Lexical affixes are subject to the broadly defined input-output constraints, which must be high-ranking in Kwakwala (as assumed).

65

Existential Faithfulness

Because segments in suffixes do not reduplicate in this language, they only have one chance to be faithful in the output and satisfy high-ranking ∃-MAXIO. Therefore obstruents cannot delete in affixal codas, even though they are non-moraic. The following ranking schema holds in a language exhibiting TETU in reduplicated words only:31 (65.)

TETU affecting reduplicated strings only IO Faith >> Markedness C >> BR Faith

(Root Faith ranking irrelevant)

In some languages the unmarked emerges in both reduplicative and lexical affixes. Such a state of affairs arises when root faithfulness constraints dominate a markedness constraint (so that roots must be faithful), while the markedness constraint dominates general IO faithfulness constraints (so that both reduplicative and lexical affixes are unmarked). (66.)

TETU affecting reduplicative and non-reduplicative affixes Root Faith >> Markedness C >> BR Faith, IO Faith

We also expect languages in which only lexical affixes show TETU. The ranking Root Faith >> Markedness C >> IO Faith ensures that such affixes are unmarked, but roots are faithful. Reduplicants are prevented from showing TETU when BR constraints are high-ranking.

31

In the following ranking schemes the usual caveats hold: general names are used for the constraints at hand (e.g. I-O Faithfulness may indicate an IDENT constraint on a particular feature disfavored by the relevant markedness constraint). The sub-hierarchies shown are assumed to be responsible for the canonical case under discussion (that is, there are no higher-ranked constraints which nullify the crucial interaction).

66

Struijke

(67.)

TETU affecting non-reduplicative affixes, but not reduplicative affixes BR Faith, Root Faith >> Markedness C >> IO Faith

2.3

Output TETU: Kwakwala case study

This section illustrates that either the base or the reduplicant can undergo a TETU alternation in Output TETU. We have seen that violations of an emergent markedness constraint are minimized in Reduplicant TETU. In Output TETU, markedness violations can often be entirely eliminated. I will exemplify this with the same Kwakwala forms seen in the previous section. These reduplicated words are unmarked in the sense that they repair stress clash. Unreduplicated words are marked because they allow it. Kwakwala reduplicated words can be categorized into four types, according to how they repair stress clash. Types A and B involve bimoraic input roots. In order to avoid stress clash, one lightens the base-initial syllable, while the other lightens the reduplicant syllable. Types C and D involve monomoraic roots. Type C avoids clash without needing an alternation in either the base or the reduplicant. Type D suppresses the reduplicant in order to avoid clash. This last type will be discussed in the section that focuses on realization of reduplicants (section 2.4).

67

Existential Faithfulness

(68.)

Overview of reduplicative types S = sonorant (moraic); O = obstruent; S’ = glottalized sonorant; Ov = voiced obstruent (non-moraic) bimoraic roots type

root shape: bimoraic syllable

reduplicative form

section

A

• •

CVS, CVSO, CVV, CVVO w´n

• • •

trisyllabic base lightens w´@n.w´.mu$˘t

2.3.2.1



CVSS’, CVSOv, CVVS’ (final C laryngeally marked) qW’a˘l’

• • •

quadrisyllabic reduplicant lightens qw’ ´.qW’a@˘.l’´.mu$˘t

B •

2.3.2.2

monomoraic roots type

root shape: monomoraic syllable reduplicative form • • •

quadrisyllabic faithful base and redupl. ts’´.ts’´$.m’´.mu@˘t

2.3.2.2



CVS’ (final C laryngeally marked) ts’´m’

• •

CVO q’ax

• • •

bisyllabic no reduplicant q’ax.m’u@ ˘ t

2.4

• C

D

section

After analyzing unreduplicated words in section 2.3.1, I will focus on types A and B in section 2.3.2, and show that the markedness constraint against stress clash determines whether the base or the reduplicant lightens a syllable (i.e. undergoes the TETU alternation).

2.3.1 Unreduplicated words and stress clash

Clash arises when heads of feet are adjacent in a string, a structure that is dispreferred cross-linguistically, and violates the following markedness constraint:

68

Struijke

(69.)

*CLASH: Adjacent heads of feet are prohibited (cf. Hung’s (1994) RHYTHM constraint - based on Prince 1983)

In Kwakwala, unreduplicated words often violate this constraint, because they can consist entirely of heavy syllables, which are footed and stressed individually, as shown below.32

(70.) Clash in non-reduplicative words (g´@lt)-(k’o$:)(di$:¬) ‘longer one side’ (ts’o@:)(l’-´$m)(y’a$:) ‘black cheek’ (he@:)(¬-o$˘)(m’a$:)(la$:) ‘to be in time’ (te@:)(n-o$:s)(ta$:)(la$:) ‘to pole up river’ Kwakwala is an iambic language (Zec 1988), and hence allows the following right-headed feet (L stands for a light syllable, H for a heavy syllable; syllables heading feet are in bold):

(71.)

Iambic foot types (McCarthy and Prince, 1986 et seq.; Hayes 1987) (LH) (LL) (H)

Stress clash is avoided when there is an alternating pattern of stressed and unstressed syllables. Given strings of heavy syllables, clash could therefore be resolved by lightening of syllables (see ungrammatical forms below), so that a sequence of (H)(H) feet can become a single, quantitatively unbalanced (LH) foot. Lightening can be achieved through vowel shortening or deletion of moraic coda consonants (in Kwakwala these are sonorants, as argued in section 2.2.1 above).

69

Existential Faithfulness

Clash in non-reduplicative words33

(72.)

trisyllabic words surface form (ts’o@:).(l’´$m).(y’a$:) (g´@lt).(k’o$:).(di$:¬) (H)(H)(H)

gloss ‘black cheek’ ‘longer one side’

no σ lightening *(ts’o@:)(l’´.y’a$:) *(g´@lt).(k’´di$:¬) *(H)(LH)

quadrisyllabic words (he@:).(¬o$˘).(m’a$:).(la$:) (te@:).(no$:s).(ta$:).(la$:) (H)(H)(H)(H)

‘to be in time’ ‘to pole up river’

*(h´.¬o@˘).(m’a.la$:) *(t´.no@:s).(ta.la$:) *(LH)(LH)

In the trisyllabic words, lightening of the medial syllable would result in a less marked form, because that syllable would be footed with the following heavy syllable in a (LH) foot. The head of this foot is not adjacent to the head of the initial (H) foot, and no clash would occur. In the quadrisyllabic forms, the first and third syllables would have to lighten to create two (LH) feet. Because such lightening does not occur, faithfulness constraints banning vowel shortening and segmental deletion must rank above *CLASH and prevent such repairs in Kwakwala unreduplicated words. The relevant faithfulness constraints are ∃-MAX IO and ∃-IDENT[weight] IO. The tableaux below illustrate the analysis. The first deals with a trisyllabic form. Candidate 1 deletes a sonorant in the medial syllable to avoid a *CLASH violation, but by doing so it fatally violates ∃-MAXIO, because a segment in the input is not present in the output.

32

Some evidence for clash and lapse patterns in languages comes from secondary stress. Boas does not indicate secondary stress in his data, but Bach (1975) claims that secondary stresses are found (p. 17). I therefore assume iterative footing. 33 Throughout this chapter I consider only candidates which satisfy FOOTBINARITY and PARSESYLLABLE. All optimal reduplicative forms satisfy these constraints.

70

Struijke

(73.) No clash resolution through sonorant deletion in unreduplicated words ∃-MAXIO *CLASH /ts’oµµl’-´µmy’aµµ/ (ts’oµµ).(l’´µ.y’aµµ) (H) (LH) 2 (ts’oµµ).(l’´µm).(y’aµµ) (H) (H) (H) 1

*! **

Tableau (33) deals with a quadrisyllabic form. Candidate 1 avoids clash by lightening the first and third syllables through vowel shortening. It fatally violates high-ranking ∃-IDENT[weight]IO because the moraic quality of two input vowels is not preserved on corresponding output vowels. Faithful candidate 2 is therefore optimal. (74.) ∃-IDENT[weight] IO:34 Let seg ∈ input be in the domain of ℜ, and seg is [αµ] then there is some seg' ∈ output, such that segℜseg' and seg' is [αµ] Some segment corresponding to an input segment preserves the moraic quantity of the input segment. (75.) No clash resolution through vowel shortening in unreduplicated words teµµn-oµµstaµµlaµµ ∃-IDENT[weight]IO *CLASH µ µµ µ µµ 1 (t´ .no s).(ta .la ) **! (LH) (LH) µµ µµ µµ µµ 2 (te ).(no s).(ta ).(la ) *** (H) (H) (H) (H)

Thus, clashes are tolerated in unreduplicated words, because ∃-IDENT[weight]IO and ∃-MAXIO dominate *CLASH.

34

I assume that moras are attributes of segments. Recall, though, that I take no position on the question whether suprasegmental features such as moras or tones are attributes of segments or autosegments. For an analysis using MAX[µ] see Struijke 1998. The ∃-IDENT[weight] IO constraint used here is adapted from McCarthy (1995).

71

Existential Faithfulness

2.3.2 The emergence of *CLASH in reduplicative words 2.3.2.1 Type A words

When a root consisting of a single heavy syllable (CVV(O) or CVS(O)) is reduplicated and concatenated with the bimoraic suffix [m’u:t], one might expect the reduplicated to exhibit clash: if left unchanged, the reduplicant, root and suffix syllables would all be heavy and stressed. However, this is not the case, as the examples below show. The root syllable of the base lightens through deletion of the moraic coda sonorant or shortening of the long vowel/diphthong to avoid clash. The reduplicant is a heavy syllable, and retains the long vowel or coda sonorant of the input. (76.)

Reduplication with bimoraic roots ending in a laryngeally unmarked segment35 root reduplicated word gloss w´@n-w´-mu˘t ‘refuse of drilling’ a w´n k´n k´@n-k´-mu˘t ‘what is left after scooping up’ b y´nt y´@n-yat-m’u˘t ‘gnawings of a large animal’ q´ns q´@n-qas-m’u˘t ‘chips’ c ka˘xW k ’ a@˘-k’axW -m’u˘t ‘shavings’ qa˘s qa@˘-qas-m’u˘t ‘tracks’ d d´y de@˘-d´-mu˘t ‘refuse of wiping’ x´w xo@˘-x´-mu˘t ‘refuse of splitting wood’ (H)(LH)

35

Recall that reduplicant codas never contain obstruents (Reduplicant TETU; see section 2.2). Vowel quality alternations are not analyzed here.

72

Struijke

Because the relevant IO faithfulness constraints dominate *CLASH (section 2.3.1), stress clash can be avoided only if these faithfulness constraints are satisfied. That is, under all circumstances each input segment and its moraic quality must be preserved in the output. In the cases above, the reduplicant contains the moraic quality of a vowel or the sonorant lost in the base. This indicates that the reduplicant is in direct correspondence with the input and can ensure input-output faithfulness. The pattern therefore constitutes strong support for the idea that IO correspondence is broadly defined to relate inputs to both the base and the reduplicant, and the idea that IO faithfulness constraints are existentially quantified so as to be satisfied when only one of these strings preserves input material. With the established constraint ranking (IO faith >> *CLASH), the Kwakwala pattern makes sense under the proposals of this dissertation. Given that reduplication involves fission, and IO faithfulness constraints are existentially quantified, one member of the base-reduplicant pair can change to resolve clash without incurring violations of high-ranking faithfulness constraints. This is depicted below in diagram (35). A sonorant coda is deleted in the root syllable, which becomes light. This syllable joins the suffix to form a (LH) foot. Because the head of this foot is not adjacent to the head of the initial foot, clash is avoided. Crucially, faithfulness to the input is not compromised because the input sonorant is present in the reduplicant. (77.)

Clash avoidance in reduplication through deletion of a sonorant in the base input:

output:

RED

(w

´µ @ n µ )

w

(

´µ nµ

w

(H)

´µ m

m

uµµ t ) (LH)

73

uµµ t

Existential Faithfulness

In diagram (36) below, the root syllable shortens its vowel and becomes light. Again clash is avoided, and again faithfulness to the input is achieved, because the correspondent of the input vowel that surfaces in the reduplicant preserves its moraic content, and shortening of the base vowel does not incur a faithfulness violation.

(78.)

Clash avoidance in reduplication by shortening of a vowel in the reduplicant input:

output: ( q

RED

aµµ@

)

q aµµ s

( q (H)



m

s

uµµ t

m uµµ t ) (LH)

Clash resolution in reduplicated words is clearly an instance of The Emergence of The Unmarked: a markedness constraint is inactive in unreduplicated words, but its effect emerges in reduplicated words. In these trisyllabic forms, it is always the base-initial syllable that undergoes the TETU alternation. If the reduplicant were to lighten, clash would not be resolved. This is shown in the two tableaux below. Tableau (37) deals with deletion of a coda sonorant, and (38) deals with vowel shortening. Faithful candidates 1 contain three (H) syllables and violate *CLASH twice. Candidates 2 lighten their reduplicant syllable, repairing one of the violations seen in the faithful candidate. However, the neighboring stressed base syllables still create clash. Clash is only avoided in candidates 3, in which the initial syllable of the base lightens.

74

Struijke

(79.)

Lightening of the base to avoid clash: deletion of sonorants ∃-MAXIO *CLASH /RED- +w´µnµ + -muµµt/ (w´µnµ)-(w´µnµ)-(muµµt) (H) (H) (H) 2 (w´µ-w´µnµ)-(muµµt) (LH) (H) µ µ µ 3 (w´ n )-(w´ -muµµt) (H) (LH) 1

(80.)

**! *!

Lightening of the base to avoid clash: shortening of a vowel ∃-IDENT[weight]IO *CLASH /RED- + qaµµs + -muµµt/ (qaµµs)-(qaµµs)-(muµµt) (H) (H) (H) µ µµ 2 (qa -qa s)-(muµµt) (LH) (H) 3 (qaµµs)-(qaµs-muµµt) (H) (LH) 1

**! *!

Lightening of the initial syllable in the base not only avoids clash, it also results in less marked foot forms because quantitatively unbalanced (LH) feet are less marked than (H) feet (Hayes 1995). Although markedness constraints on foot form could explain why syllables lighten, they cannot explain which syllable lightens. Only *CLASH favors the form (H)(LH) over (LH)(H). Because * CLASH forces deletion of a segment or shortening of a vowel in the root, the root faithfulness constraints ∃-MAXRT and ∃-IDENT[weight]RT must be ranked below it. We thus have evidence for the following constraint ranking. (81.)

Kwakwala constraint ranking ∃-MAXIO , ∃-IDENT[weight]IO >> *CLASH >> ∃-MAXRT , ∃-IDENT[weight]RT

Even though root faithfulness is dominated by *CLASH, unreduplicated roots are not affected by the markedness constraint because they must satisfy higher ranked ∃-MAXIO and ∃-IDENT[weight]IO.

75

Existential Faithfulness

Again, base-reduplicant faithfulness constraints must be low-ranking to facilitate The Emergence of The Unmarked (see also McCarthy and Prince 1994a) because it results in non-identity of the base and reduplicant. DEPBR is violated when a segment is present in the reduplicant but not in the base. IDENT[weight]BR is violated when corresponding base-reduplicant segments are not identical with respect to moraic quality (i.e. when one shortens its vowel). The constraint ranking allowing Output TETU is summarized in the schema below.

(82.)

Constraint ranking allowing Output TETU ∃ IO faith >> markedness >> ∃ Root faith, BR faith

2.3.2.2 Type B and type C words

It is not always the base that responds to *CLASH in Kwakwala. In quadrisyllabic reduplicated words derived from bimoraic roots (type B), it is the reduplicant that lightens. Reduplicated words are quadrisyllabic when a root ends in a laryngeally marked segment (a voiced obstruent or glottalized consonant). Boas points out that these segments “cannot be followed by a [consonant] without having an [[´]] or [a] following, that is to say, the voicing [or] glottalization are continued as a vocalic vibration of the vocal cords after the consonantic closure” (1947: p. 209). Put differently, laryngeally marked segments trigger epenthesis of a vowel, resulting in an additional syllable. Examples are given below.

76

Struijke

(83.)

Forms with bimoraic roots ending in laryngeally marked consonants36 root reduplicative form gloss m´.m´@n.dz´.mu$˘t ‘leavings after cutting kindling wood’ m´ndz qW’a˘l’ qW’ ´.qW’a@˘.l’´.mu$˘t ‘embers’ sa˘qW’ s´.sa@˘.qW’´.mu$:t ‘peelings’

Crucial here is the fact that the epenthesized vowel heads an additional syllable.37 This syllable is light and can be footed with the heavy suffix syllable [mu:t] into a (LH) foot. With no change in the first two syllables, both the reduplicant and base-initial syllable would be heavy and stressed, creating clash. However, clash is resolved because the reduplicant becomes light. No IO faithfulness violations are incurred, because the input material lost in the reduplicant is present in the base, as can be seen in the data above. The tableau below shows that only lightening of the reduplicant can prevent violation of *CLASH while at the same time ensuring preservation of input material. If the base-initial syllable lightened (candidate 3), the reduplicated word would be footed as the clash violating structure (H)(LL)(H). Clash could be avoided if both the reduplicant and base-initial syllable lightened (candidate 4). However, this would lead to loss of input material.

36

The form glossed ‘peelings’ is taken from Boas, 1948. Thanks to Stephen Anderson for making this dictionary available to me. 37 I will not provide a formal analysis of Kwakwala vowel epenthesis. However, the analysis of epenthesis in Sanskrit (chapter 3) can be extended to the present case.

77

Existential Faithfulness

(84.)

*CLASH prefers an alternation in reduplicants of quadrisyllabic forms ∃-IDENT[weight]IO *CLASH /RED + qw’aµµl’ + muµµt/ (q w ’aµµ)-(qW’aµµ)(l’´µ-muµµt) (H) (H) (LH) w µ µµ µ 2 (q ’´ -qW’a )(l’´ -muµµt) (LH) (LH) w µµ µ 3 (q ’a )-(qW’´ l’´µ)-(muµµt) (H) (LL) (H) w µ µ µ 4 (q ’´ -qW’´ )(l’´ -muµµt) (LL) (LH) 1

*!

*! *!

In this analysis of reduplicative types A and B, I have presented evidence for the claim that input-output faithfulness relates inputs to both bases and reduplicants, and that constraints on the relation are indifferent as to whether the base, the reduplicant, or both resemble the input. Because these constraints are existentially quantified, other constraints must determine the shape of reduplicated forms. In the case at hand, the markedness constraint *CLASH emerges, and depending on the syllabic structure of the word, it requires either the base or the reduplicant to shorten a vowel or delete a sonorant coda. Output TETU differs crucially from Reduplicant TETU in that the emerging markedness constraint determines the location of the TETU alternation, and can be fully satisfied in reduplicated words. Before concluding this section, I briefly turn to reduplicative forms C of chart (26) above, an example of which is given below. (85.)

Type C: faithful base and reduplicant root reduplicative form ts’´m’ (ts’´.ts’´)(m’´.m’u@˘t) (LL) (LH)

gloss ‘left after melting’

Forms like these do not undergo a TETU alternation, because they can faithfully realize the reduplicant and base-initial syllable without creating clash. They contain monomoraic roots ending in a laryngeally marked consonant, and therefore induce vowel epenthesis

78

Struijke

(as seen in type B reduplicated words). As a result, the reduplicated form is quadrisyllabic, and can be footed as (LL)(LH). I now turn to reduplicative type D (which does not show overt reduplication), and the general issue of phonological realization of reduplicative morphemes.

2.4 Realization of reduplicative morphs and phonological reduplication

This section looks into the factors determining whether reduplication takes place. I show that the existential definition of faithfulness constraints makes important predictions with respect to this issue. First, existential IO faithfulness requirements can conspire with a markedness constraint to force realization of a reduplicative morpheme (section 2.4.2) or prevent it (section 2.4.1). Second, the theory presented in this dissertation predicts that reduplication can take place in the absence of a reduplicative morpheme (i.e. phonological reduplication) in order to achieve simultaneous satisfaction of a markedness constraint and an IO faithfulness constraint (section 2.4.3). The empirical focus of this section remains the analysis of Kwakwala Output TETU. It analyzes reduplicative type D, in which the reduplicant is not phonologically realized, and addresses the question of what forces reduplication in types A, B, and C.

2.4.1 Non-realization of /RED/

Kwakwala reduplicative words of type D illustrate how an emergent markedness constraint can prevent realization of a reduplicative morpheme. Given a Kwakwala input containing the reduplicative morpheme, a monomoraic root, and the suffix [m’u:t], the grammar produces a form in which no reduplicant is present in the output (type D of table 26).

79

Existential Faithfulness

(86.)

Reduplication with monomoraic roots (Type D) input output RED-/ax-m’u˘t (/ax- m’u@˘t) RED-q’´k-m’u˘t (q’ax-m’u@˘t) RED-/aXW-m’u˘t (/aXW- m’u@˘t) RED-ts’´x-m’u˘t (ts’ax-m’u@˘t) RED-y’´xW-m’u˘t (y’axW-m’u@˘t) (LH)

gloss ‘waste left after some work’ ‘piece bitten out’ ‘waste scum’ ‘hair singed off’ ‘high water mark’

This pattern follows directly from the TETU constraint ranking established above (IO Faith >> *CLASH >> Root Faith, BR Faith). The demands of undominated existential input-output faithfulness constraints are met in the above forms because there exists a faithful output segment for every input segment. *CLASH is also satisfied, because these forms consist of a single (LH) foot. Of the relevant constraints, only low-ranking MAXBR is violated, because no segment in the base has an output correspondent. (87.)

MAXBR: Every segment of the base has a correspondent in the reduplicant

If the reduplicative morpheme were realized, the base-reduplicant faithfulness constraint would be better satisfied. However, a violation of the higher-ranked constraint *CLASH would result, as the tableau below shows.

(88.)

Reduplicant is not realized given monomoraic roots that do not trigger epenthesis (type D) IO FAITH *CLASH MAXBR RED-/ax-m’u@˘t (/a./ax)(m’u@˘t) (LL) (H) 2 (/ax.m’u@˘t) (LH) 1

*!

**** ******

80

Struijke

Clash could be avoided if the reduplicant were realized with a long vowel or sonorant coda. However, the base does not contain such material, and vowel lengthening and epenthesis are ruled out by IDENT[weight]BR and DEPBR. (89.)

No vowel lengthening or epenthesis in reduplicants to satisfy *CLASH *CLASH DEPBR IDENT[weight]BR MAXBR RED-/aµx-m’u@µµt 1 (/aµµx)(/aµx.m’u@µµt) (H) (LH) 2 (/aµlµx)(/aµx.m’u@µµt) (H) (LH) 3 (/ax.m’u@˘t) (LH)

*! *!

*** *** ***** *

In conclusion, if reduplication in Kwakwala would create a clash violation, it does not take place. An emergent markedness constraint can thus prevent over realization of a reduplicative morpheme.38 An important question to ask is why reduplicant morphemes are realized at all in languages exhibiting TETU deletions. After all, MAXBR is crucially dominated in such languages. This is one of the issues addressed in the next section.

2.4.2 Forces driving realization of /RED/

38

Even though the reduplicant is not phonologically realized, the morphological status of these words is recoverable from the fact that they contain the suffix [m’u:t], which always co-occurs with the reduplicative morpheme. Such a state of affairs is not uncommon cross-linguistically. For instance, German also shows non-realization of a morpheme driven by footing requirements, namely in the formation of the past participle (Kiparsky 1997): ge-glaubt versus *glaub-t ‘believed’ *ge-telefonier-t versus telefonier-t ‘phoned’. Even when the German prefix [ge] is not realized, the morphological status of the word is recoverable, because the morpheme /-t/ always co-occurs with the prefix. Alternatively, one might view Kwakwala /RED - m’ut/ and German /ge - t/ as circumfixes, and the morpheme is always realized.

81

Existential Faithfulness

There are at least three distinct forces that can drive reduplication. The first is a constraint that explicitly demands the phonological realization of underlying morphemes, including reduplicative morphemes. (90.)

MORPHREAL: Realize a morpheme in an overt and detectable manner (Samek-Lodovici 1992)

The second cause of reduplication is the constraint MAXBR, which is generally assumed to be active whenever a reduplicative morpheme is present. It requires that segments in the base have correspondents in the reduplicant, and thus prefers realization of a reduplicative morpheme over non-realization. In languages that permit deletion in the reduplicant due to The Emergence of The Unmarked, this constraint is low-ranking. Despite its low rank, it is still able to force overt reduplication when constraints dominating it are indifferent as to whether the morpheme is realized or not. Thirdly, and most importantly, I show that reduplication can be forced jointly by existential input-output faithfulness constraints and markedness constraints.

(91.)

Constraints forcing reduplication 1.

MORPHREAL

2.

MAXBR

3.

∃-IO Faith and Markedness

I will now exemplify the last two causes of reduplication by examining reduplicative types A, B, and C of Kwakwala. In type B words (e.g. [qW’ ´qW’a˘l’´mu˘t]), realization of the reduplicant is

forced by MAXBR. Tableau (50) compares two candidates, both of which faithfully parse

82

Struijke

the root and suffix, but only candidate 1 realizes the reduplicant. Since the output root and suffix of both candidates are identical to the input root and affix, IO faithfulness constraints are satisfied, regardless of whether the reduplicant is realized. *CLASH is also satisfied in both, because the base-initial syllable is heavy and forms the head of the initial foot (either LH or H), which is not adjacent to the head of the following (LH) foot. The crucial difference between candidate 1 and 2 is phonological realization of the reduplicant. Since candidate 2 does not realize the morpheme, it incurs more violations of MAXBR than candidate 1. Thus, even though the constraint is low-ranking and cannot force reduplication in words of type D, it is active in candidates that are otherwise equally harmonic with respect to input-output faithfulness and *CLASH. (92.)

RED must be realized with bimoraic roots which trigger epenthesis (type B) ∃-IDENT[weight]IO *CLASH MAXBR /RED- + qWa˘l’ + -mu:t/ 1 (qW’ ´.qW’a˘).(l’´.mu˘t) (LH) (LH) 2 (qW’a˘).(l’´.mu˘t) (H) (LH)

***** ***** **!

Consider now the trisyllabic reduplicated forms (type A; e.g. [w´nw´mu˘t]). Here existential IO faithfulness and the emerging markedness constraint conspire to force reduplication. The tableau below shows that realizing /RED/ (candidate 1) is more harmonic than failing to realize it (candidates 2 and 3). Suboptimal candidate 2 faithfully parses the root to satisfy input-output constraints, but it thereby incurs a fatal violation of *CLASH. Suboptimal candidate 3 parses the root unfaithfully as a light syllable to avoid this violation. However, high-ranking ∃-MAXIO is violated, because the sonorant of the input is not parsed in the output. Optimal candidate 1 realizes the reduplicative morpheme as a heavy syllable and the base as a light syllable. As a result, both existential

83

Existential Faithfulness

faithfulness and the emerging markedness constraint are satisfied. Thus ∃-MAXIO and *CLASH conspire to force realization of the reduplicant in candidate 1. (93.)

RED must be realized with bimoraic roots which do not trigger epenthesis ∃-MAXIO MAXBR *CLASH /RED- +w´n + -mu:t/ 1 (w´n).(w´.mu˘t) (H) (LH) 2 (w´n).(mu˘t) (H) (H) 3 (w´.mu˘t) (LH)

*** **! *!

****** *****

*CLASH and IO faithfulness requirements likewise force reduplication in the quadrisyllabic reduplicated forms derived from monomoraic roots (type C, such as [ts’´ts’´m’´mu˘t]). Suboptimal candidate 2 of tableau (52) faithfully parses the root and does not phonologically realize the reduplicant. The foot structure of the output form is (LL)(H) and incurs a fatal *CLASH violation. In order to avoid this violation, candidate 3 lengthens its root vowel. However, we have seen above that ∃-IDENT[weight]IO is undominated in Kwakwala. Thus, candidate 3 is ruled out by this constraint. Realization of the reduplicant (candidate 1) avoids both ∃-IDENT[weight]IO and *CLASH violations, and is therefore optimal. (94.)

RED must be realized with monomoraic roots which trigger epenthesis *CLASH MAXBR /RED- + ts’´m’ + -mu:t/ ∃-IDENT [weight]IO 1 (ts’´ -ts’´)(m’´-mu˘t) (LL) (LH) ***** 2 (ts’´m’´)-(mu˘t) (LL) (H) *! ****** * 3 (ts’e˘)(m’´-mu˘t) (H) (LH) *! ****** *

84

Struijke

The next section shows that existential IO constraints can conspire with markedness constraints to force reduplication in the absence of a reduplicative morpheme.

2.4.3 Reduplication in the absence of /RED/

The table below repeats the three driving forces of reduplication exemplified in the preceding sections.

(95.)

Constraints forcing reduplication 1.

MORPHREAL

reduplicative morpheme necessary

2.

MAXBR

reduplicative morpheme necessary

3.

IO Faith and Markedness

no reduplicative morpheme necessary

These first two constraints can only come into play when the input contains a reduplicative morpheme. That is, these constraints cannot force reduplication when no such morpheme is present. Whether a reduplicative morpheme is present or absent in the input is irrelevant for the third driving force of reduplication, because the constraints involved do not refer to reduplication. The theory presented here therefore makes an important prediction: reduplication can take place in the absence of a reduplicative morpheme, in order to improve performance on markedness constraints without causing loss of input information. In particular, a phonological copy can preserve input material that is lost in the base. We have just seen that reduplication in Kwakwala sometimes takes place to ensure preservation of input segments and their moraic specifications, while at the same time avoiding clash. That is, reduplication in word types A and C is driven by the third force in the table above. It is clear that a reduplicative morpheme is present in the inputs

85

Existential Faithfulness

of these forms, by analogy with reduplicative words of type B (where MAXBR forces reduplication). In all relevant reduplicative forms of Kwakwala, reduplication takes place in a certain morphological environment, namely one that indicates the meaning ‘refuse of’ or ‘left over.’ Outside this environment reduplication is not used as a repair strategy for satisfying *CLASH. In fact, words violate this constraint in the absence of a reduplicative morpheme, as shown in section 2.3.1. Thus, in Kwakwala, morphologically driven reduplication is allowed, but phonologically driven reduplication is prohibited. Clearly, the grammar must differentiate morphological reduplication from phonological reduplication. As mentioned in section 2.1.3, I assume that GEN always produces multiple input-output correspondence in the presence of a reduplicative morpheme. Because fission in morphological reduplication is universally required by GEN,39 it does not violate the constraint INTEGRITYIO. Phonological reduplication, however, does incur violations of this constraint. (96.)

INTEGRITYIO: For seg-x ∈ input and seg-y, seg-z ∈ output, if seg-xℜseg-y and seg-xℜseg-z, then seg-y = seg-z No fission. The integrity of an input segment is preserved in the output.

This constraint must be high-ranking in Kwakwala to prevent phonological reduplication. However, we expect it to be violated in some languages, so that reduplication occurs for phonological reasons. Further research is needed to identify languages in which multiple segments undergo fission to create the effect of reduplication. However, in chapter 3 and 4 I analyze languages in which single segments

39

More precisely, there is no morphological 'copying' without segmental fission. Of course, fission is not established when /RED/ is present in the input, but not realized in the output.

86

Struijke

undergo fission to ensure faithfulness and simultaneous improvement of markedness (in feature ‘movement,' dissimilation, on analogy with distributing diphthongization).

2.5 Reduplicant size as a predictor of TETU alternations

The main goal of this section is to point out that reduplicant size is an important factor in determining which segments of a reduplicated word can be affected by an emerging markedness constraint. Before turning to this issue, I will briefly discuss the idea that reduplicant size (as well as the overall shape of a reduplicated word) is an emergent property of constraint interaction. Reduplicant size was previously argued to be determined by a fixed template (McCarthy 1979, Marantz 1982, McCarthy and Prince 1986). Such templates cannot capture the idea that the Kwakwala reduplicant is sometimes a light syllable, sometimes a heavy syllable, and sometimes not realized at all. OT accounts of reduplicant size rely on the interaction of markedness and faithfulness constraints and can account for variable reduplicant size. Two different proposals have been presented in the literature. The first is Generalized Template Theory (McCarthy and Prince, 1994b; Urbanczyk 1996). It holds that reduplicative morphemes are underlyingly specified for morphological category (i.e. affix or root) and are subject to morphology-prosody interface constraints specifying the unmarked prosodic shape of each morpheme category. For instance, the Kwakwala reduplicant is an affix, and under this view is subject to the constraint AFX≤σ, which states that an output affix is maximally a syllable. This constraint is satisfied in Kwakwala because it outranks MAXBR, which prefers full reduplication. The second proposal holds that reduplicant size is an emergent effect of other, more general constraints (Spaelti 1997; Walker 2000). Spaelti argues that reduplicant

87

Existential Faithfulness

size is limited by emerging alignment constraints (McCarthy and Prince 1993b), most notably ALIGN-σ-R/L. This constraint demands all syllables to be aligned with a designated edge of the word, and thus prefers forms that comprise not more than one syllable. Monosyllabic reduplicants are therefore preferred to multisyllabic reduplicants. Better still are reduplicants consisting of material that can be incorporated into syllables of the base; however, this option is often ruled out by constraints on syllable structure. Walker (2000) adopts Spaelti's proposal but utilizes emergent *STRUC-σ as the size restrictor, rather than alignment constraints. Either of these two constraint-interaction proposals can account for the fact that Kwakwala reduplicants are maximally a syllable. To complement these ideas, I have argued that the shape of both the reduplicant and the base in part depends on the overall shape of the reduplicated word. Let us now turn to reduplicant size as a predictor of TETU alternations. TETU alternations arise in reduplicated words because input segments have two chances to surface in the output. The fact that the Kwakwala reduplicant is only one syllable long therefore has important consequences. The only input segments that potentially undergo fission (i.e. reduplicate) are those that can be syllabified into the reduplicant syllable and the base-initial syllable. Subsequent input segments only have one chance to surface in the output. These segments are predicted never to delete or change to meet the demands of a markedness constraint ranked below existential IO faithfulness constraints. That is, we do not expect them to participate in reduplicative TETU alternations. This prediction cannot be tested for the emergence of *CLASH in words containing the suffix [m’u:t], because alternations in this syllable will never improve the rhythmic pattern of a word. However, the prediction can be tested and is borne out in reduplicative forms containing the longer suffix [(g)i:sa:we:/]. This suffix consists of three heavy syllables. Lightening of the medial syllable would avoid clash. However, this

88

Struijke

does not take place. Instead, the final and penultimate syllables of the suffix are footed and stressed individually and hence cause a * CLASH violation. (97.)

Clash in reduplicative forms with the morpheme /(g)i:sa:we:/ / input root reduplicative form gloss ‘left-over drilling’ w´n (w´n).(w´.gi˘).(sa˘).(we˘/) d´y (de˘).(d´.gi˘).(sa˘).(we˘/) ‘left-over wiping’ gi˘x (gi˘).(ga.xe˘).(sa˘).(we˘/) ‘left-over filings’ (H) (LH) (H) (H)

Thus, because the suffix is not reduplicated, suffixal segments only have one chance to surface in the output. Therefore, high-ranking existential IO faithfulness constraints force them to surface faithfully, and a * CLASH violation is unavoidable. The constraint schema allowing Output TETU is repeated below:

(98.)

Ranking schema allowing Output TETU IO Faith >> Markedness C >> Root Faith, BR Faith

Given this ranking, lexical suffixes will be affected by TETU only when they are reduplicated.

2.6 Markedness constraints in Output and Reduplicant TETU

This section discusses the role of markedness constraints in Output TETU and Reduplicant TETU. I first show that the domain size evaluated by an emerging markedness constraint may determine which kind of TETU takes place (section 2.6.1). When a markedness constraint evaluates a small domain, it is more likely to be involved in Reduplicant TETU. Conversely, when a markedness constraint evaluates a large domain, it is more likely to be involved in Output TETU. Section 2.6.2 discusses the fact

89

Existential Faithfulness

that the emerging markedness constraint determines the alternation site in Output TETU. I provide examples of Output TETU from a variety of languages confirming that the alternation site is sometimes the base, sometimes the reduplicant, and at other times both of these strings are simultaneously affected. These examples indicate that it is often a prosodic markedness constraint - and thus one that evaluates a large domain - that emerges in Output TETU.

2.6.1 The effect of constraint domain size

We have seen the roles of *CLASH and WEIGHTbyPOSITION in The Emergence of The Unmarked in Kwakwala. Because these markedness constraints differ crucially in domain size, they are involved in two different kinds of TETU. The domain of WxP is relatively small because it evaluates segments within a syllable, while *CLASH assesses the interaction between heads of feet, and thus evaluates a domain larger than a syllable. This difference is important in Kwakwala, because the size of the reduplicant is one syllable. That is, the size of the reduplicant coincides with the size of WxP’s domain. For this reason WxP evaluates the reduplicant independently from syllables in the base, and both reduplicant and base syllables may have to undergo alternations in order to satisfy it. When WxP is dominated by input-output faithfulness constraints, its demands for alternations are not fully granted: only one member of the base-reduplicant pair can alternate. It is irrelevant to WEIGHTbyPOSITION where the alternation or the faithful parsing takes place, regardless of its rank. Root faithfulness ensures that the base surfaces faithfully, and the reduplicant alternates. *CLASH evaluates a larger domain. In Kwakwala, this domain includes material from both the base and the reduplicant. In other words, a string which includes base and reduplicant material can violate *CLASH as a whole, and an alternation in either can avoid

90

Struijke

such a violation. This means that this constraint indicates a preference for the site of alternations: depending on the syllabic structure of a word, it either requires the base to change, or the reduplicant, but not both. Because it demands only one string to change, it can be entirely satisfied. In summary, in Reduplicant TETU the domain of the emerging markedness constraint is always small (i.e. equal to or smaller than the reduplicant), and as a result it demands alternations in both members of the base-reduplicant pair. In Output TETU, the emerging markedness constraint demands an alternation in either the reduplicant, or the copied part of the base. Often this is because the domain it evaluates is larger than the reduplicant.

2.6.2 Determining the alternation site in Output TETU

A markedness constraint that emerges in Output TETU determines the location of the TETU alternation and does not always show variable preferences for the location, as in Kwakwala (and Lushootseed, discussed briefly in section 1.2.3.1). Sometimes markedness constraints show a fixed preference. Languages in which it is always the base that undergoes Output TETU alternations include Bella Coola (Salish), Klamath (Petunia) and Tohono O’odham (Uto-Aztecan) (Raimy and Idsardi 1997; Cole 1997; Fitzgerald 1999, respectively). In Bella Coola and Tohono O’odham, syncope is not found in unreduplicated words, but is attested in bases of reduplicated words. The base undergoes the TETU alternation rather than the reduplicant, because it is the base that is penalized by the emerging markedness constraint.

91

Existential Faithfulness

(99.)

TETU affecting the base in Bella Coola unreduplicated word reduplicated word kap’ay ka-kp’ayi k’inax /i-k’naax silin sil-slin

(100.) TETU affecting the base in Tohono O’odham unreduplicated word reduplicated word ko@:s ko@:-ks wo@son wo@-pson ma@wid ma@-mwid

gloss ‘humpback salmon dim.’ ‘crab dim.’ ‘kidney dim’

gloss ‘sleeping (plural)’ ‘sweeping (plural)’ ‘mountain lion’

Cole (1997) points out that, in Klamath, syncope and vowel reduction are also only attested in bases of reduplicated words (distributive reduplication).

(101.) TETU affecting the base in Klamath unreduplicated word reduplicated word paga pa-pga nt’op’a nt’o-nt’a tSonwa tSo-tS´nwa sditga sdi-sd´tga

gloss ‘barks’ ‘spoils’ ‘vomits’ ‘cheat’

Fitzgerald argues that syncope in the base-initial syllable of Tohono O’odham words takes place to make the word-initial (i.e. stressed) syllable heavy. Raimy & Idsardi and Cole ascribe the alternations in Bella Coola and Klamath to constraints in the *STRUC family, such as *σ and *SEGMENT. However, their similarity to the Kwakwala and Tohono O’odham patterns, as well as the Lushootseed pattern, suggests that the patterns in these languages are due to emerging prosodic constraints. Further research into these cases is needed to draw conclusions. Other Northwest American languages that seem to display similar patterns are Chilliwack Halkomelen (Salish - Galloway 1977), Saanich (Salish - Montler 1989), Shuswap (Salish - Kuipers 1974; Bell 1983; Broselow and McCarthy 1984), and Quileute (Chimakuan - Andrade 1993; Broselow and McCarthy 1984).

92

Struijke

These examples show typical characteristics of Output TETU: the emerging markedness constraint is usually a prosodic constraint which evaluates a relatively large domain, and demands an alternation in only one member of the base-reduplicant pair. However, the Tohono O’odham example discussed in chapter 1 (Fitzgerald 1998) shows that these properties are not defining characteristics of Output TETU. In this instance of Output TETU, ONSET emerges and causes both the base and the reduplicant to alternate. (102.) TETU affecting both the base and the reduplicant in Tohono O’odham unreduplicated word reduplicated word gloss hiosig hi-hosig ‘flower’ ˜ˆok ˜ˆ-˜ok ‘talking’ doa do-da ‘to be healthy’ toaya to-taya ‘towel’

The data above indicate that ONSET (or *HIATUS) is freely violated in unreduplicated words, because a sequence of two input vowels surfaces faithfully in two consecutive syllables. ∃-MAXIO prevents deletion of one of the vowels in these words. However, in reduplicated words, both constraints can be satisfied, because one vowel surfaces in the reduplicant and the other surfaces in the base. Such division of input material between the base and reduplicant is the topic of the next section. Summarizing this section, I have identified the following characteristics of Output and Reduplicant TETU:

93

Existential Faithfulness

(103.) Reduplicant TETU and Output TETU compared Reduplicant TETU Output TETU Domain evaluated by emerging Domain evaluated by emerging markedness constraint ≤ size of markedness constraint includes reduplicant. material of both the base and the reduplicant Emerging markedness constraint Emerging markedness constraint demands the same alternation in the demands an alternation in one member base and reduplicant of the base-reduplicant pair, or different alternations in each Root faithfulness determines the The emerging markedness constraint location of a TETU alternation determines the location of a TETU alternation Performance of emerging markedness Emerging markedness constraint is constraint is improved often entirely satisfied

2.7 Division of input characteristics between base and reduplicant

In this section I show that the existential definition of IO constraints allows input segments and feature specifications associated with segments to be distributed among the base and reduplicant. That is, both members of the base-reduplicant pair can be unlike the input, and still satisfy existential IO constraints. We have just seen division of two input segments between the base and reduplicant in Tohono O'odham. The same pattern can be observed in Kwakwala. Kwakwala reduplicated words often show deletion as a repair strategy in both Reduplicant TETU and Output TETU. In reduplicative words of type A whose roots contain a sonorant and obstruent coda, the base and the reduplicant are equally unfaithful to the input.

94

Struijke

(104.) Reduplicated forms in which base and reduplicant are equally unfaithful40 y´nt y´@n-yat-m’u˘t ‘gnawings of a large animal’ q´ns q´@n-qas-m’u˘t ‘chips’

The obstruent is deleted in the reduplicant through Reduplicant TETU of WEIGHTbyPOSITION, and the sonorant is deleted in the base through Output TETU of *CLASH (see sections 2.2 and 2.3 above). ∃-MAXIO is satisfied because the segment that is lost in one copy is present in the other. In addition to segments themselves, feature specifications of segments can be divided between the base and reduplicant. This can be seen in Japanese, where the voicing feature value of an input /p/ is preserved on the first copy in a reduplicated word, while the place and manner feature values are preserved on the second copy.41,42 (105.) Feature value distribution in Japanese reduplication unreduplicatedreduplicated gloss /pito/ hito hito-bito people /pasi/ haSi haSi-baSi edges /poso+i/ hosoi hoso-boso very slender

This instance of feature value distribution is illustrated in the diagram below.

Note that ∃- CONTIGUITY is violated in these forms. For instance, in /y´nt/ - [y´@n-yat-m’u˘t], /n/ and /t/ are adjacent in the input, but no correspondents of these segments are adjacent in the output. 41 As I will point out in section 2.9, these Japanese forms are not an instance of TETU, but are rather instances of normal application (or ‘the emergence of the faithful’). 42 Thanks to Markus Hiller for drawing my attention to this pattern, and to Haruka Fukazawa and Mits Ota for providing examples. 40

95

Existential Faithfulness

(106.) Feature value distribution in Japanese reduplication (only input features are shown) [lab] [-cont] [-voice]

input:

output:

[h]

i

RED [p]

i

o

[b]

t

[-voice]

t

i

o

t

o

[lab] [-cont]

Because each feature value associated with the input /p/ is associated with one of the segment’s output correspondents, existential IDENT[F]IO constraints are satisfied. This is, of course, the same state of affairs as distributing diphthongization. I give an analysis of the Japanese pattern in section 2.9. Bases and reduplicants can be equally unfaithful to inputs because they are generated in parallel and are both related to the input through a single correspondence relation (broad IO correspondence). In addition, the existential nature of IO faithfulness allows one member of the base-reduplicant pair to be unlike the input, as long as the other resembles it. If we adopt the more standard assumption that the base is prior to the reduplicant, however, or that only the base is related to the input, we wrongly predict that the base is more like the input than the reduplicant. In the next section I show that the different theories (base priority versus base-reduplicant equality43 ) therefore make different assumptions as to which string is the base and which is the reduplicant in a given reduplicated word.

43

Base-reduplicant equality is also assumed in the ‘Double Stem Selection’ theory of reduplication, proposed by Inkelas and Zoll. In this theory different ‘co-phonologies’ determine the shape of each base-reduplicant string.

96

Struijke

2.8 Identifying base and reduplicant

The unmarked, most common placement of a reduplicant is at the right or left edge of the word. However, reduplicants can also occur word-internally. Infixation comes in two varieties (Broselow and McCarthy 1984; McCarthy and Prince 1990). First, like lexical affixes, reduplicants may be placed adjacent to some phonological domain, such as the syllable carrying main stress. Second, reduplicants may be placed near (but not at) the edge of a word to improve the syllabic structure of that word (Anderson 1972), as seen in for instance Chamorro (Austronesian). (107.) Chamorro intensifying reduplication unreduplicatedreduplicated gloss metgot metgogot ‘strong’ na)laN na)lalaN ‘hungry’ cf. bunita bunitata 'pretty'

In claiming that the reduplicant of these Chamorro words is the medial string, rather than the final string, researchers (implicitly) relied on the idea that bases are more like the input than reduplicants. However, we have seen that both bases and reduplicants can be unlike the input. I have argued that the base and reduplicant are related to the input by a single correspondence relation, and either string can be more faithful to the input than the other. As a result, either string of a reduplicated word can in principle be identified as the reduplicant (modulo Root Faithfulness). In Chamorro, we could identify either the final or the penultimate syllable as the reduplicant: (108.) Chamorro intensifying reduplication unreduplicatedred. as infixation red. as suffixation metgot metgogot metgogot n‚alaN n‚alalaN n‚alalaN

97

gloss ‘strong’ ‘hungry’

Existential Faithfulness

Whichever theory of faithfulness one assumes, it is clear that the constraint NOCODA emerges in Chamorro reduplicated words and rules out the form [metgotgot] (this is in effect the analysis given in Prince and Smolensky 1993). The syllabic markedness constraint emerges and ensures that the input-final coda consonant appears only once in the reduplicated form. The fact that it appears in the final member of the base-reduplicant pair is probably due to the constraint FINAL-C.

(109.) FINAL-C: Every prosodic word is consonant final (McCarthy and Prince 1994b) (110.) Emergence of NoCoda in Chamorro reduplication /metgot + RED/ ∃-MaxIO NOCODA FINAL-C 1 metgotgot ***! 2 metgotgo * * *! * * 3 metgogot

Under the proposed theory, two output candidates compete: metgogot which contains an infixed reduplicant and metgogot which contains a suffixed reduplicant. The former better satisfies root faithfulness requirements, in particular ∃-MaxRT, while the latter satisfies the constraint demanding reduplicants to be edge-tropic (i.e. ALIGN(RED, R, STEM, R)). (111.) ALIGN(RED, R, STEM, R): The right edge of a reduplicant coincides with the right edge of the stem. (the reduplicant is a suffix) (McCarthy and Prince 1993b, after Prince and Smolensky 1993)

Thus, in the proposed theory, the ranking of root faithfulness constraints and alignment constraints determine whether a reduplicant is edge-tropic or infixed in languages such as Chamorro. In the alternative theories discussed above, reduplicants are less faithful than bases, and therefore must be infixed in such languages.

98

Struijke

Summarizing, under the assumption that IO faithfulness is broadly defined to relate inputs to bases and reduplicants alike, resemblance to the input does not determine which string is the base and which string is the reduplicant. This is happy result, because some reduplicative forms contain bases and reduplicants that are equally unlike the input (see previous section). In such forms, resemblance to the input cannot be the basis for identification of each string.

2.9 The emergence of the faithful

The issue of infixation versus edge-tropic affixation is not the only one that is cast in a new light by the existential faithfulness framework. Because existential faithfulness is satisfied when one member of the base-reduplicant pair preserves input material, the faithful can sometimes emerge in reduplication. In some languages, unreduplicated words neutralize underlying distinctions which emerge in reduplicated words. This is because reduplication sometimes creates an environment in which the underlying marked structure is allowed to surface - an environment that is absent in unreduplicated words. An enlightening example comes from Yidiny (Australian) discussed in Evans (1995). (112.) Yidiny emergence of the faithful input unreduplicatedreduplicated gloss m payka-® payka® payka®mpayka® 'feeling very sore'

Here, the nasal property of a pre-nasalized stop is lost at the beginning of the unreduplicated word, but emerges in the beginning of the reduplicant in the reduplicated word.

99

Existential Faithfulness

Japanese exhibits a more intricate example of a language in which unreduplicated words are unfaithful, but reduplicated words are faithful. Words of the Yamato lexical stratum never contain a singleton [p]. The segment is only found in geminated form (nippon ‘Japan’) or in ‘partially geminated’ form (kampai ‘cheers’). In unreduplicated words, single underlying /p/ usually surfaces unfaithfully as [h]. (113.) Japanese unreduplicated words /pito/ hito person /pasi/ haSi joint /poso+i/ hosoi slender

I assume that this fact is due to the following markedness constraint (Itô and Mester 1995): (114.) *p: no singleton [p]

In the /p/ - [h] alternation, the place and manner feature specifications of the input segment are lost. Only the voicing specification is preserved. Thus the markedness constraint *p dominates ∃-IDENT[lab] and ∃-IDENT[-cont]. Higher-ranked ∃-IDENT[-voice] ensures that /p/ does not change to [b]. (115.) Unfaithful realization of /p/ in Japanese unreduplicated words44 /pito/ ∃-IDENT *p ∃-IDENT ∃-IDENT [-voice] [lab] [-cont] 1 pito *! * * 2 hito 3 bito *!

44

There is a candidate that satisfies all constraints in this tableau, namely /h1it2o/ - [b1it12o]. In this candididate, input /h/ undergoes fission and its correspondent preserving [-voice] coalesces with /t/. The resulting output [t] then preserves the [-voice] specifications of both its contributing input segments. This candidate is suboptimal in Japanese because it violates high-ranking INTEGRITY.

100

Struijke

As Fukazawa (1999) points out, [h] is the voiceless segment of Japanese most similar to /p/. [f] is prohibited from surfacing in the language altogether, and [∏] only appears before the high back vowel [µ] (e.g. /pµka+i/ - [∏µkai] ‘deep’). In reduplication, the second member of the base-reduplicant pair is subject to rendaku: in the Yamato stratum obstruents are voiced when they are the initial segment of the second member in a compound. Thus when /pito/ is reduplicated, /p/ surfaces as [b] in the second member of the base-reduplicant pair and as [h] in the first member. All input features associated with /p/ are now preserved in the output. The voicing feature is preserved on the first correspondent, while the place and manner feature specifications are preserved on the second correspondent. (116.) Emergence of the faithful in Japanese reduplication reduplicated gloss /pito/ hito-bito people /pasi/ haSi-baSi edges /poso+i/ hoso-boso very slender

Given that IDENT[F] constraints are existentially quantified, and are satisfied when input feature specifications are preserved on some output correspondent of the input segment, all relevant IDENT[F] constraints are satisfied. At the same time, *p is satisfied also, because [p] is not contained in the output. This is shown in the tableau below.45 (117.) The emergence of the faithful in Japanese reduplication /RED+pito/ RENDAKU ∃-IDENT *p ∃-IDENT [-voice] [+lab] 1 pito-pito *! ** 2 hito-hito *! * 3 hito-bito

45

∃-IDENT [-cont] *

For the sake of convenience, I use the descriptive constraint RENDAKU to force voicing in the second compound. For analyses of rendaku see Itô and Mester (1995) and references quoted therein.

101

Existential Faithfulness

The Japanese forms examined exemplify the emergence of the faithful. Some IO faithfulness constraints are inactive (i.e. violated) in unreduplicated words, but are active (and in this case entirely satisfied) in reduplication. Low-ranked faithfulness constraints can emerge, because they can be satisfied without violating dominant *p. The conflict between ∃-IDENT constraints and *p has disappeared completely in reduplicative words.

2.10 Comparison with other proposals

In this section, I first compare the proposals presented in this chapter with those of classic Correspondence Theory (McCarthy and Prince 1995, 1999). I then compare them with other research in which a broad IO relation has been proposed.

2.10.1 Comparison with classic Correspondence Theory

The faithfulness relations I have argued for are repeated below. They were discussed in section 2.1. (118.) Correspondence relations assumed in this dissertation input:

RED + Root IO Faithfulness Root Faithfulness

output:

Reduplicant

Base

BR Faithfulness

102

Struijke

McCarthy and Prince (1995) assume different correspondence relations. Their ‘Full Model’ of correspondence is depicted below:

(119.) Full Model of Correspondence (McCarthy and Prince 1995) input:

RED + Stem IR Faithfulness IO Faithfulness (IB Faithfulness)

output:

Reduplicant

Base BR Faithfulness

Input-output faithfulness of the Full Model relates the lexically specified input to the unreduplicated word, or the base in the reduplicated word. Therefore this relation is referred to as input-output correspondence in unreduplicated words, but input-base (IB) correspondence in reduplicated words. Despite these two different names, it is conceived as a single correspondence relation. Note that the concept ‘base’ includes both roots and affixes and therefore differs from the root faithfulness relation of the model proposed in this thesis. The IO/IB relation of the Full Model does not establish correspondence between inputs and reduplicants. That is, no fission takes place in reduplication. The reduplicant is related to the input by a separate relation, dubbed input-reduplicant (IR) correspondence. The reduplicant is also related to the base by base-reduplicant correspondence (BR). The differences between these two models have consequences for the analysis of TETU (section 2.10.1.1), The Emergence of the Marked (section 2.10.1.2), and normal application (section 2.10.1.3).

103

Existential Faithfulness

2.10.1.1 The Emergence of The Unmarked

The Full Model of Correspondence and the model of correspondence assumed in this dissertation differ in how they account for The Emergence of The Unmarked in reduplication. The Full Model cannot capture patterns like those exhibited in Kwakwala, in which bases alternate to facilitate TETU. This is due to the assumption that unreduplicated words and bases in reduplicated words are related by the same correspondence relation. Because these strings are subject to the same faithfulness constraints they must behave in the same manner. When input-output constraints are undominated and cause unreduplicated words to be faithful and marked, bases in reduplicated words cannot become unfaithful and unmarked. Even though the Full Model does not capture Output TETU in which bases alternate, it does account for Reduplicant TETU. Reduplicants are not subject to IO/IB faithfulness constraints and can therefore behave differently from unreduplicated words and bases. Reduplicants are subject to input-reduplicant and base-reduplicant constraints instead, and when these rank below a markedness constraint, markedness-driven alternations are allowed in reduplicants. Thus, in the Full Model, it is possible for a markedness constraint to be active in reduplicants only, particularly given the following constraint ranking:

(120.) TETU ranking schema for Full Model (only reduplicant can change) IO Faith >> Markedness >> BR Faith, IR Faith

Because IO correspondence of classic Correspondence Theory regulates faithfulness to all lexically specified morphemes in the input (roots and affixes alike), The Emergence of The Unmarked is restricted to reduplicants in this ranking, and does not extend to lexical affixes. In order to capture the generalization that roots tend to be

104

Struijke

more faithful than affixes cross-linguistically, root faithfulness must be assumed in addition to the IO/IB relations.

2.10.1.2 The Emergence of the Marked

In hypothetical ‘Emergence of the Marked’ (term coined by McCarthy and Prince 1995), a reduplicant would contain more marked structure than unreduplicated words. Such a state of affairs is unattested cross-linguistically. For instance, if a language does not allow coda consonants in unreduplicated words, it will not exhibit them in reduplicants either. The full model of correspondence and the model of correspondence proposed here both account for the absence of The Emergence of the Marked, but in different ways. In the Full Model, IO/IB faithfulness must universally dominate input-reduplicant faithfulness to rule out the Emergence of the Marked, as McCarthy and Prince (1995) demonstrate. If IR faithfulness constraints could dominate IO/IB constraints, such a pattern would be expected, as shown in the tableaux below. If a markedness constraint ranked between input-reduplicant and input-output faithfulness constraints, unreduplicated words and bases in reduplicated words would be subject to the markedness constraint, but reduplicants would not be.

(121.) If IR >> Markedness >> IO, the marked could emerge in reduplicants (a) unreduplicated word is unmarked /lab/ MAX IR NOCODA MAX IO/IB 1 lab *! * 2 la (b) reduplicant is marked / RED + lab/ MAX IR 1 la-la *! 2 lab-la 3 lab-lab

NOCODA * **!

105

MAX IO/IB * *

Existential Faithfulness

Given the assumption made in classic Correspondence Theory that input-output faithfulness constraints universally dominate input-reduplicant constraints, such Emergence of the Marked is ruled out. Any markedness constraint that ranks above IO faithfulness constraints and causes phonological alternations in unreduplicated words must also rank above input-reduplicant faithfulness and cause alternations in reduplicants also. The broad input-output relation that I propose in this chapter rules out The Emergence of the Marked without stipulating a universally fixed constraint ranking.46 Broad IO correspondence combines the IR and IO/IB relations of the Full Model, in the sense that it relates inputs to unreduplicated words and to both the base and the reduplicant in reduplicated words. Thus, a single constraint family regulates input faithfulness for entire output strings, regardless of their morphological make-up. When a markedness constraint forces unfaithfulness in unreduplicated words containing a certain marked structure, it must therefore also force unfaithfulness in reduplicated words containing the same marked structure, as shown below:

(122.) With broad IO, the marked cannot emerge in reduplicants (a) unreduplicated word is unmarked /lab/ NOCODA ∃-MAX IO ∃-MAX RT 1 lab *! * * 2 la (b) reduplicant must be unmarked also / RED + lab/ NOCODA ∃-MAX IO * 1 la-la 2 lab-la *! 3 lab-lab **! 46

∃-MAX RT * *

Universally fixed constraint rankings are quite common in the realm of markedness, where they provide a formalization of universal markedness implications (e.g. the ranking *lab >> *cor expresses thegeneralization that languages containing labial segments also contain coronal segments, but not vice versa). Universally fixed rankings of faithfulness constraints, however, are not commonly assumed. Entailments of faithfulness are usually taken to be the result of positional faithfulness.

106

Struijke

Thus, under the proposed model of correspondence, reduplicants (potentially) containing a marked structure cannot be more faithful to the input than unreduplicated words (potentially) containing the same marked structure, simply because these strings are subject to the same faithfulness constraints: broad IO constraints.47

2.10.1.3 Normal application

The third mode of application that the two models of correspondence account for differently is normal application. In normal application, reduplicated and unreduplicated words both repair marked structures. For example, in Pangasinan (Austronesian; Rose 2000b), /d/ cannot surface intervocalically. Instead, its allophone [R] appears. In the reduplicative examples (81b and c) below, the base and reduplicant contain different allophones. (123.) Pangasinan flapping (normal application) a. dabok ‘dust’ ma-Rabok b. ma-Rakep ‘nice’ ma-Rag-dakep c. dalikan ‘clay stove’ da-Ralikan

‘dusty’ ‘quite nice’ ‘clay stoves’

Regardless of whether one takes /d/ to be the input segment or /R/, these examples show that reduplicants must be directly related to the input. That is, if the input contains /d/ it is faithfully preserved in the reduplicant of (81c), and if the input contains /R/ it is faithfully preserved in the reduplicant of (81b). In these cases the reduplicant cannot be

47

The emergence of the faithful is not an instance of the emergence of the marked. It is a special type of normal application, where the reduplicated word as a whole or one member of the base-reduplicant pair happens to lack a marked structure found in unreduplicated words.

107

Existential Faithfulness

established by means of the BR relation alone, because it contains material absent from the base. Because reduplicants sometimes contain input material lost in the base due to normal application, McCarthy and Prince (1995) see themselves forced to incorporate the input-reduplicant relation in their model of correspondence. However, it is clear that they consider this relation an unsatisfactory concept (because without the stipulated fixed ranking it predicts the richer, partially unattested, typology discussed in the previous section). In classic Correspondence Theory, normal application is achieved when a markedness constraint dominates both input-output and input-reduplicant constraints. The tableaux below show an alternation in an unreduplicated word of Pangasinan, the base of a reduplicated word, and the reduplicant of a reduplicative word respectively (assuming input segment /d/). (124.) Normal application (a) unreduplicated word is unmarked *VdV / ma-dakep / 1 madakep 2 maRakep

IO/IB Faith

IR Faith

*! *

(b) reduplicated word is unmarked (alternation in the base) / RED - dalikan / *VdV IO/IB Faith IR Faith *! 1 da-dalikan * 2 da-Ralikan (c) reduplicated word is unmarked (alternation in the reduplicant) / ma-RED-dalikan / *VdV IO/IB Faith IR Faith *! 1 ma-dag-dakep * 2 ma-Rag-dakep

McCarthy (1997b) points out that cases such as these do not necessarily call for an input-reduplicant relation. Instead, the reduplicant could be in correspondence with the base only (as in the ‘Basic Model’ of correspondence, McCarthy and Prince 1995). We

108

Struijke

then deal with an opaque output whenever a reduplicant contains material present in the input but not the base. McCarthy (1998: 'sympathy' theory of opacity) argues that opaque outputs require reference (more explicitly, faithfulness) to a form that is not the actual input or output. In the optimization /RED + dalikan/ - [da-Ralikan] this form contains a [d] in the base (da-dalikan). Thus, the Sympathy theory of opacity can account for cases of normal application in which the base, but not the reduplicant is affected. However, this pattern does not constitute an instance of opacity if one assumes that the output strings in a reduplicative word are generated in parallel and are both related to the input by a single correspondence relation: broad input-output correspondence. When a markedness constraint dominates constraints on this relation, any marked structure penalized by the constraint alternates, regardless of whether it is in the base, the reduplicant, or both. (125.) Normal application affecting the base in reduplicated words (a) unreduplicated word is unmarked *VdV ∃-IO Faith ∃-Root Faith / ma-dakep / 1 madakep 2 maRakep

*! *

*

(b) reduplicated word is unmarked (alternation in the base) / RED - dalikan / *VdV ∃-IO Faith ∃-Root Faith *! 1 da-dalikan * 2 da-Ralikan (c) reduplicated word is unmarked (alternation in the reduplicant) / ma-RED-dalikan / *VdV ∃-IO Faith ∃-Root Faith *! 1 ma-dag-dakep 2 ma-Rag-dakep

This concludes the comparison of Classic Correspondence Theory and the theory proposed in this dissertation with respect to reduplicative patterns. I have shown

109

Existential Faithfulness

that the latter covers a larger empirical domain than the former, even though they contain the same number of correspondence relations. It does so without resorting to a universally fixed ranking of faithfulness constraints. Instead it relies on the idea that faithfulness relations are sometimes in a subset relation with each other, namely when general and positional faithfulness constraints are involved (broad IO and Root Faithfulness).

2.10.2 Comparison with other work assuming broad IO

Several researchers have independently proposed ideas similar to broad input-output correspondence. In addition to Struijke (1997), they are Cole (1997), Raimy and Idsardi (1997), Spaelti (1997) and Yip (1998a, 2000). I will briefly discuss these proposals, as well as that of Fitzgerald (1998, 1999) who adopts the basic idea.48 Raimy and Idsardi propose a broad IO relation motivated by principles of minimalism (Chomsky 1995). They argue that phonology should exclusively involve phonological elements and hence faithfulness relations should be defined in terms of phonological entities only: they should not make reference to both morphological and phonological constructs. Thus, they claim that input-output correspondence is conceptually superior to input-base and input-reduplicant correspondence of the Full Model, simply because it relates a phonological construct (the input) to another

48

Comparison is restricted to research adopting general assumptions about reduplication made here (i.e. the presence of a morpheme RED and some form of duplication). Therefore this section excludes the proposal presented in Inkelas and Zoll (2000), even though they postulate a relation similar to broad IO. The proposal assumes three separate ‘co-phonologies’ (constraint rankings) relevant in reduplicated words: one for each member of the base-reduplicant pair, and one for the reduplicated word as a whole. All contain input-output faithfulness constraints. Inkelas and Zoll argue that this is empirically comparable to a grammar consisting of a single constraint ranking with three types of IO faithfulness constraints: input-base, input-reduplicant, and broad IO constraints.

110

Struijke

phonological construct (the output), rather than a ‘quasi-morphological’ construct (the base or reduplicant). Using examples from Bella Coola reduplication given in section 2.6 above, Raimy and Idsardi show that broad input-output correspondence can account for languages in which certain processes can affect bases, but not unreduplicated forms. This is, of course, the same argument as the one put forward in this dissertation (i.e. this is a case of Output TETU affecting the base). Spaelti proposes the ‘Reduplicate! Model of Correspondence,’ consisting of base-reduplicant and broad input-output faithfulness (which he calls ‘Lexical form-Surface form Faithfulness’). He argues that the broad IO relation is superior to the IO/IB and IR relations of the Full Model because no universal ranking of faithfulness constraints is needed. Spaelti argues that all TETU predictions of the Full Model (i.e. Reduplicant TETU) can be maintained by eliminating the input-reduplicant relation. However, he does not address the question of why it is always the reduplicant that changes in this type of TETU. Contrary to Raimy and Idsardi (1997) and Spaelti (1997), the present chapter of this dissertation has shown that root faithfulness plays an important role in TETU, since it accounts for the fact that reduplicants often contain less marked material than roots (Reduplicant TETU). Fitzgerald (1998, 1999) adopts the broad input-output relation to account for TETU affecting the base in Tohono O’odham (see section 2.6 above). In addition she assumes the input-reduplicant and input-base faithfulness constraints of the Full Model, which she argues are freely rankable. Although these constraints can determine which member of the base-reduplicant pair preserves input material, and which undergoes the TETU alternation, their free ranking prevents an account of the fact that all other things being equal, bases are more faithful to the input than reduplicants. As I have shown, the

111

Existential Faithfulness

IR relation is entirely superfluous if one assumes broad IO correspondence. Affixes, including reduplicants, are subject to general IO constraints and no special faithfulness constraints need to refer to them. Although Yip’s proposal (1998a, 2000 - see also Rose 2000b) differs from the one presented in this chapter in important ways, it is similar with respect to the analysis of The Emergence of The Unmarked. Yip argues that TETU involves high-ranking input-output constraints which relate both the base and the reduplicant to the input. Positional faithfulness constraints determine which member of the base-reduplicant pair reflects the input, and which is free to change in The Emergence of The Unmarked. She presents data from Chinese languages which show that the relevant positional faithfulness constraints are not necessarily root faithfulness constraints. For instance, faithfulness constraints on word-initial segments may ensure that a prefixal reduplicant is more faithful than the base.49 All these researchers rely on the standard assumption that MAX is existentially quantified. They do not discuss featural TETU alternations in bases. This type of alternation forms pivotal evidence for existentially defined featural faithfulness constraints in the present dissertation. Raimy & Idsardi (1997) and Fitzgerald (1998, 1999) do not analyze featural alternations, and the question of how IDENT[F] constraints are quantified is not addressed. Yip (1998a, 2000) and Spaelti (1997) discuss featural alternations but do not explicitly address the question either. However, from their analyses it can be inferred that Yip does not assume these constraints to be existentially quantified (see tableaux (25) and (26) in Yip 2000), whereas Spaelti assumes they are existentially defined (see tableau (53) in chapter 2 of Spaelti 1997). Cole (1997) also aims to explain TETU which affects bases. However, rather than assuming broad input-output constraints, she proposes disjointive constraints 49

See also the analysis of Chamorro in section 2.8.

112

Struijke

involving input-base and input-reduplicant constraints of the Full Model. A disjointive constraint is satisfied when either of the ‘member constraints’ is satisfied (Smolensky 1993; cf. Hewitt and Crowhurst 1996). Like Yip, Cole argues that positional faithfulness constraints determine whether the base or the reduplicant surfaces faithfully.

2.11 Conclusion

This chapter introduced a model of reduplication in which the base and reduplicant are considered equal with respect to the phonology. This is established formally such that both strings are related to the input by input-output correspondence. In reduplication, each input segment has the potential to undergo fission and have a correspondent in both the base and the reduplicant. The distinct behavior often exhibited by the strings results from the difference in their morphological status. It is formally derived from the idea that roots in bases, but not affixal reduplicants, are subject to faithfulness requirements of the phonology-morphology interface, namely input-root faithfulness (usually referred to as ‘Root Faith’; Beckman 1997). This chapter identified a novel generalization about reduplication: it is not only reduplicants that tend to be less marked than unreduplicated words; sometimes reduplicated words as a whole are less marked. Such a pattern is predicted to occur under the proposed model and the claim that input-output faithfulness constraints are existentially defined. In reduplication, one member of the base-reduplicant pair can change in response to a markedness constraint without violating faithfulness requirements. I have identified languages in which either the base or the reduplicant can alternate, even though such alternations are prohibited in unreduplicated words (The Emergence of The Unmarked).

113

Existential Faithfulness

In fact, both strings can be equally unlike the input, but along different dimensions. Still, it is usually possible to identify which is the base and which is the reduplicant (cf. Inkelas and Zoll 2000, who claim that they cannot be distinguished under these circumstances). When the markedness constraint driving the TETU alternation demands alternations in both strings, the reduplicant changes. When it demands only one to change, it single-handedly determines which.50 In previous theories, TETU alternations in the base constituted instances of opacity. This is because bases were taken to be prior to reduplicants; either literally (in serial, rule-based theories) or conceptually (in classic Correspondence Theory). In these theories, the reduplicant is derived from or related to the base, not directly to the input. Thus, under these assumption, opacity arises when the reduplicant contains material not present in the base. The proposed framework eliminates this particular type of opacity, which was traditionally seen as problematic for output-oriented theories such as OT. I have compared the proposed model to classic Correspondence Theory (section 2.10.1). The two frameworks differ empirically only in their account of The Emergence of The Unmarked. In the earlier theory only the reduplicant could alternate; in the theory proposed here eitherthe reduplicant or the base can. The proposed model is able to account for this slightly larger number of reduplicative patterns without complicating its architecture. The correspondence relations and faithfulness constraints are equal in number to those assumed in classic Correspondence Theory. Instead of input-reduplicant correspondence, I assume broad IO faithfulness. The fact that bases tend to be more faithful than reduplicants results from the subset relation between this general IO relation and the specific Root Faithfulness, rather than the universal ranking root faithfulness/input-base faithfulness >> affix faithfulness.

50

Provided Root Faith is low-ranking.

114

Struijke

3 Feature movement and dissimilation

The main thesis of this dissertation holds that faithfulness regulates preservation of underlying material, and faithfulness constraints are existentially quantified. Empirically, existential faithfulness differs from universal faithfulness (McCarthy and Prince 1995) in segmental fission. ∃-IDENT[F] constraints do not require all output segments to be identical to a given corresponding input segment. Instead, only one must resemble it. The other can change to become less marked, without compromising faithfulness. This chapter looks at cases where the marked output correspondent of a fissioned input segment coalesces with another segment. I will show that this combination of fission and coalescence results in phonological patterns usually referred to as ‘feature movement’ and ‘dissimilation.’51 Like distributing diphthongization and TETU in reduplication, these phenomena show simultaneous preservation of underlying elements and decrease of markedness. The chapter first focuses on feature movement (section 3.1). Descriptively, it occurs when a feature value underlyingly associated with one segment surfaces on another. Such a state of affairs has been taken as evidence for the idea that features are entities independent of segments, and can literally reassociate - from one segment to another. However, I show that the existential definition of IDENT[F] constraints predicts an alternative, segment-based, explanation of this phenomenon. In the proposed account, an underlying segment undergoes fission and divides its feature specifications between two surface correspondents, in the same vein as Middle English sometimes divides input features in diphthongization when borrowing words from French (chapter 1), and Japanese divides features in reduplication (chapter 1; section 2.7). By means of such 51

Other terms used to refer to these phenomena (or special cases of them) are ‘floating features’, ‘feature displacement’; and ‘cooccurence restrictions’, ‘OCP’, and ‘Morpheme Structure Constraint’. These terms were coined in derivational theories, and I will use them in a descriptive fashion.

115

Existential Faithfulness

fission, markedness of a form is improved, while each underlying feature specification of the fissioned segment is preserved in the surface form. I argue that feature movement differs from distributing diphthongization only in that the more marked surface segment resulting from distributing fission coalesces with another segment in the word. Section 3.1 includes subsections dealing with fission and coalescence. The main idea of this proposal was previously presented in Struijke 2000b,c. In chapter 4, I compare the segment-based account of feature movement with the feature-based account. The second part of this chapter (section 3.2) deals with cooccurrence restrictions. Many languages do not allow identical or similar segments in a domain. For instance, syllables in Seri (Hokan) never contain more than one glottalized segment, and words in Georgian (Caucasian) do not normally contain more than one rhotic segment (for an overview of these and other dissimilation cases, see Suzuki 1998). Following recent research, I argue that dissimilation is driven by a need to reduce segmental markedness (Alderete 1997; Alderete and Pater, in prep.; Itô and Mester 1996; Suzuki 1998). However, I depart from this research in claiming that it also ensures faithfulness. That is, dissimilation is a way to improve on the markedness of a form, while simultaneously achieving faithfulness of underlying material. The argument is based on the idea that one output segment can preserve identical feature specifications of two coalesced input segments, and only one correspondent of a fissioned segment must preserve its marked features. Thus, like feature movement, dissimilation involves both coalescence and distributing fission. This can be briefly illustrated by labial dissimilation in the Akkadian (Semitic) mapping /ma-ereb/ - [ne-ereb]52 'entrance' (data analyzed earlier in Hume 1992; McCarthy 1981; Odden 1994; Suzuki 1998). In the proposed analysis, the input labial /m/ undergoes fission. Only its second output correspondents is labial. This correspondent coalesces with the labial /b/, which ensures faithfulness to both [lab] 52

Underscore points to the relevant segments.

116

Struijke

feature specifications. The first correspondent of /m/ can then acquire the unmarked place specification to better satisfy the markedness constraint *LAB. The main idea of this proposal was previously presented in Struijke 2000b,c and Struijke and de Lacy 2000. This chapter includes two case studies of languages which show feature movement and dissimilation, namely Sanskrit (section 3.3) and Cuzco Quechua (section 3.4). The relatedness of the two phenomena is discussed in these sections. In particular, I will show that movement and dissimilation are sometimes indistinguishable in that they can both be a means to avoid one particular marked structure (i.e. both are driven by the same markedness constraint).

3.1 Feature movement

As mentioned above, the term ‘feature movement’ refers to a situation in which a feature specification is associated with a segment with which it is not associated underlyingly. For example, in Esimbi (Bantoid) the feature specification [-hi] can only be associated with words in initial syllables. When a [-hi] input vowel surfaces in a non-initial syllable, its height feature specification is preserved on the vowel in the initial syllable (Stallcup 1980a,b; Hyman 1988; Walker 1997). The underlying [-hi] vowel defaults to [+hi] in the surface form.53 (126.) Esimbi feature transfer ([-hi] vowels are underlined) /u-se/ osi ‘laugh’ /u-rE/ çri ‘daub’ /i-gbe/ egbi ‘bushfowl’ /i-so/ esu ‘hoe’ /i-b´/ ebˆ ‘cane rat’

53

I show in section 3.1.2 that this is not an instance of feature metathesis.

117

Existential Faithfulness

Obviously, some markedness constraint demands that vowels in non-initial syllables be [+hi] (perhaps because they are unstressed and must be of low sonority, or perhaps there is a more general constraint against non-high vowels which is not active in initial or stressed syllables).54 The fact that the feature specification is preserved is clearly due to some faithfulness requirement. Yet, feature movement is problematic for the classic Correspondence Theory of faithfulness (McCarthy and Prince 1995). This is because it assigns two properties to featural faithfulness constraints. First, it assumes that featural faithfulness is regulated through segmental correspondence. Second, constraints on this relation are universally quantified. It is the combination of these properties that makes feature movement problematic for the theory. Because classic IDENT[F] constraints demand featural faithfulness under segmental correspondence, they cannot enforce feature value preservation outside of the segment. Any output segment that preserves a feature value associated with a given input segment must be in correspondence with that segment. For instance, in the Esimbi example, /e/ of /u-se/ must be in correspondence with both vowels of the output [osi] because the first preserves the [-hi] feature value, and the second preserves the color feature specifications. However, classic IDENT[F] is universally quantified, demanding all corresponding segments be featurally identical. Thus, even though the mapping /e/ - [o] satisfies the classic IDENT constraint on height features in the input-output pair /u-se1/ [o1si1], the mapping /e/ - [i] violates this constraint. Hence, the idea that a feature specification moves to ensure faithfulness of that feature cannot be formalized using classic IDENT[F]. Instead, classic Correspondence Theory creates a paradox: feature value preservation must be due to a faithfulness constraint, yet faithfulness constraints are violated because not all corresponding segments are featurally identical.

54

Thanks to Larry Hyman for information on this point. In the tableau below I use the descriptive constraint *[-hi]/non-initial σ.

118

Struijke

Some have chosen to employ MAX[F] constraints to deal with this problem (e.g. Parker 1997; Walker 1997). These researchers assume that features are entities independent of segments and enter into correspondence relations of their own (Lamontagne and Rice 1995; Lombardi 1995, 1998; Causley 1997a,b). MAX[F] constraints enforce preservation of features independently of segmental correspondence. While the MAX[F] approach is empirically adequate, I argue that the original insight in Correspondence Theory can be maintained: features are properties of segments, and featural faithfulness is mediated by segmental correspondence. This insight is modified, however, by employing existentially defined IDENT[F] constraints. The definition of existential IDENT[F] is repeated below.55 (127.) ∃-IDENT[F] IO: Let seg ∈ input be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an input segment preserves the feature specification [αF] of that input segment.

Because an existential IDENT[F] constraint is satisfied when a feature value associated with an input segment is preserved on some corresponding output segment, apparent feature movement actually involves segmental fission. An input segment corresponds to multiple output segments, and some of its feature specifications are preserved on one of these correspondents, while other specifications are preserved on the other. This resolves the paradox described above: an ∃-IDENT[F] constraint can ensure preservation of an input feature specification through segmental correspondence, without imposing the stringent requirement that all corresponding segments must be featurally

55

This constraint shares important characteristics with the MAX[F] constraint. I show in chapter 4 that both are unidirectional, existentially quantified, and demand faithfulness to a particular feature value. The crucial difference between the two is whether features are seen as independent entities or properties of segments.

119

Existential Faithfulness

identical to the input segment. Under the present analysis, feature movement is thus an epiphenomenon of segmental faithfulness, rather than a primitive notion. I will now discuss fission and coalescence, which combine to produce feature movement (and dissimilation) in the existential faithfulness framework. I show that classic IDENT[F] needs to be modified to account for coalescence (Pater 1999).

3.1.1 Feature movement as fission and coalescence

Feature movement within the existential faithfulness framework is a combination of two patterns independently found in phonology: segmental fission and coalescence. To see this, consider again the Esimbi example /u-se/ - [osi]. I suggested above that the input segment /e/ undergoes fission and has two output correspondents, realized as [o] and [i]. All feature specifications of the input segment /e/ are preserved on one of these output segments. The feature values [-back] and [-round] are preserved on the second vowel [i], and the feature value [-hi] is preserved on the first vowel [o]. The vowel [o] also preserves the color specifications of the input segment /u/. Thus, the first correspondent of /e/ coalesces with /u/. This is depicted below. It bears repeating that lines are indications of correspondence relations, equivalent to the numerical indices, not autosegmental association lines. (128.) Feature value transfer as fission and coalescence (only input features are shown) [+rd] [+back] [+hi]

[-rd] [-back] [-hi]

input:

u1

s2

e3

output:

o1,3

s2

i3

[+rd] [+back] [-hi]

[-rd] [-back]

120

Struijke

Because feature transfer involves both fission and coalescence, I discuss these phenomena in turn, building on the existent body of research in these areas.

3.1.1.1 Fission in feature movement

The general topic of fission has been addressed in chapter 1. Phonological fission in feature movement (and distributing diphthongization) occurs under a specific set of circumstances. First, some markedness constraint prohibits feature specifications associated with an input segment from being preserved together on a single output correspondent. A second, and equally important prerequisite is that feature values must be maintained due to highranking of certain ∃-IDENT[F] constraints. Third, fission itself must be allowed: INTEGRITY has to be low-ranking. Finally, surface faithfulness constraints evaluating output segments related to the same input segment must be of low rank. This is because, at a minimum, one feature specification is found on one but not the other segment (i.e. the moved feature value). The constraint schema expressing these circumstances is given below: (129.) Ranking schema for fission in feature movement (and distr. diphthongization) Markedness constraint, ∃-IDENT[F] >> INTEGRITY, ∃-IDENT[F]∑∑

3.1.1.2 Coalescence

Coalescence is the second component of feature value transfer. It involves fusion of two underlying segments into a single surface segment. The resulting segment

121

Existential Faithfulness

typically combines feature specifications of both participants.56 For example, coalescence of /n/ and /b/ might result in /m/, preserving the nasal feature value of the first segment and the labial feature value of the second segment. Within Correspondence Theory, coalescence involves multiple correspondence: two input segments are in correspondence with one output segment. MAXIO is satisfied because each input segment has an output correspondent. However, when the contributing segments are not identical, not all input feature specifications can be preserved in the output. That is, a conflict arises between IDENT[F] constraints. Which feature specifications are preserved in the output is determined by the ranking of these constraints. I follow most research on coalescence by assuming that IDENT[F] constraints are unidirectional and must be specified for feature values, as set out below (Gnanadesikan 1995, 1997; McCarthy 1995; and Pater 1999, 2000; see also Casali 1996 in the Parse/Fill model of faithfulness (Prince and Smolensky 1993)). This departure from the original definition of IDENT[F] constraints is necessary to distinguish the following types of structures: (130.) Classic IDENT[F] constraints cannot distinguish (a) and (b) (a) input:

n –

output:

(b) d

n

d

—

–

—

n

d

Classic IDENT[F] constraints are defined symmetrically and can therefore not distinguish representations such as those in (5a) and (5b) (Pater 1999; McCarthy and Prince 1995). 56

Coalescence sometimes preserves features of one of the participating segments only, such that on the surface, coalescence is indistinguishable from deletion of one of the segments (Gnanadesikan 1995; Causley 1997b; McCarthy 1995; Pater 1999).

122

Struijke

(131.) Classic IDENT[F]: Correspondent segments have identical values for the feature F If xℜy and x is [γF], then y is [γF] (McCarthy and Prince 1995)

Classic IDENT constraints are symmetrical in two ways. First, they are bidirectional: the output must be like the input, and vice versa. Second, they are value-blind: a change from a minus value to a plus value incurs a violation, as does the reverse change. Thus, mappings — in (5a) and – in (5b) equally violate classic IDENT[nasal]: this constraint cannot distinguish nasalization in the former from denasalization in the latter. The first part of the solution to this problem is the proposal that IDENT[F] constraints refer to specific feature values (Pater 1999). That is, the universal constraint set CON contains both IDENT[+nasal]IO and IDENT[-nasal]IO. The former penalizes denasalization, while the latter penalizes nasalization. However, this solution is only viable if constraints are unidirectional. If they were bidirectional, feature value specificity would have no effect, and denasalization and nasalization would be evaluated equally after all. For instance, the /d/ - [n] mapping (—) in (5a) violates bidirectional IDENT[+nasal] because the [+nasal] output segment has no [+nasal] input correspondent. The mapping /n/ - [d] (–) in (5b) violates the same constraint, because the [+nasal] input segment has no [+nasal] output segment. Both mappings also violate Ident[-nasal] for the opposite reasons. Only the combination of feature value specificity and unidirectionality can distinguish nasalization from denasalization (Pater 1999). One of the core proposals in this dissertation is that faithfulness mandates preservation of input material in the output, and that faithfulness constraints are therefore unidirectional, going from the input to the output (chapter 1). Therefore I assume that ∃-IDENT[F] constraints demand feature specifications associated with an input segment

123

Existential Faithfulness

be preserved on an output correspondent of that segment. There are no additional IDENT[F] constraints that require feature specifications associated with an output segment be associated with an input correspondent (e.g. IDENT[+nasal]OI (O→I), Pater 1999). Having IDENT[F] constraints for each direction would introduce considerable redundancy in the grammar (see Bakovic 1999 for a discussion of this point). For instance, the mapping /t/ - [n] would incur violations of both IDENT[-nasal]IO and IDENT[+nasal]OI , while /n/ - [t] would incur violations of both IDENT[+nasal]IO and IDENT[-nasal]OI. Instead, I assume only the following IDENT[F] constraints for nasality: (132.) ∃-IDENT[+nasal] IO : Let seg ∈ input be in the domain of ℜ, and seg is [+nasal]; then there is some seg' ∈ output, such that segℜseg' and seg' is [+nasal] Some segment corresponding to an input segment preserves the feature specification [+nasal] of the input segment. (133.) ∃-IDENT[-nasal] IO : Let seg ∈ input be in the domain of ℜ, and seg is [-nasal] then there is some seg' ∈ output, such that segℜseg' and seg' is [-nasal] Some segment corresponding to an input segment preserves the feature specification [-nasal] of the input segment.

Existential quantification of these constraints is of no consequence in coalescence, because each input segment has only one output correspondent. What is crucial for coalescence is unidirectionality ( I → O) and feature value specificity ([αF]). An important question is how privativity might be incorporated into this discussion. The theory of privativity assumes that a feature or feature specification has no plus and minus values. For example, a nasal segment is associated with the feature [nasal], but an oral segment is not specified for or associated with a nasal feature. Privativity can be incorporated in an IDENT model of featural faithfulness and can

124

Struijke

successfully differentiate denasalization from nasalization, voicing from devoicing, etc. (Lombardi 1996). However, it is only able to do so if there are IDENT[F] constraints in both the input-output and output-input directions. That is, the mapping /n/ - [t] violates privative ∃-IDENT[nasal]IO (I→O) and the mapping /t/ - [n] violates privative ∃-IDENT[nasal]OI (O→I) (Pater 1999, 2000). This view is inconsistent with my claim that faithfulness constraints require preservation of input material in the output, and I will therefore not address privativity in ∃-IDENT[F] constraints again. Since there are faithfulness constraints relevant to all feature values in the input, coalescence of non-identical segments always incurs ∃-IDENT[F]IO violations. For this reason I do not assume the additional constraint UNIFORMITY which penalizes coalescence in classic Correspondence Theory (see chapter 1; Keer 1999). However, coalescence of segments belonging to different morphemes is often disallowed in languages, even when they otherwise permit such multiple correspondence. In these languages the constraint ∃-MORPHDIS is highranking. (134.) ∃-MORPHDIS (Morphemic Disjointness) Let seg-x ∈ input Morph1, and seg-y ∈ input Morph2, and seg-x', seg-y' ∈ output, and segxℜseg-x', seg-yℜseg-y'. Then there is some seg-x’≠ seg-y’. (McCarthy and Prince 1995 (adapted)). The disjointness of input morphemes is preserved in the output. (segments belonging to different morphemes do not coalesce)

Feature movement involves coalescence, and hence can only take place when the ∃-IDENT[F]IO constraints violated in coalescence are ranked sufficiently low. Features can move from a segment in one morpheme to a segment in a different morpheme only when ∃-MORPHDIS is also low-ranking.

125

Existential Faithfulness

3.1.2 Combining fission and coalescence into feature movement

We return now to the Esimbi example to conclude the proposed analysis of feature value transfer. Here I will also present evidence that the pattern is indeed the result of fission and coalescence, and should not be analyzed as literal feature movement, or an instance of feature metathesis. Diagram (3) is repeated below for convenience. (135.) Feature value transfer as fission and coalescence (only input features are shown) [+back][+hi]

[-back][-hi]

input:

u1

s

output:

o1, 2

s

[+back][-hi]

e2

i2 [-back]

Tableau (11) illustrates the ranking of constraints that make the form depicted here optimal. Esimbi has a highranking markedness constraint banning the feature [-hi] from surfacing on the segment with which it is underlyingly associated. The height value of /e/ is not lost (as in candidate 2), but preserved in the initial syllable, due to the high rank of ∃-IDENT[-hi] IO. This existential constraint forces fission of /e/ (i.e. it forces a violation of INTEGRITY). Since one output correspondent of /e/ carries the [-hi] feature specification, the ∃-IDENT[-hi] IO constraint is satisfied, so the other correspondent can be unmarked [+hi]. The first correspondent of root segment /e/ coalesces with affix segment /u/, which loses its [+hi] specification. Apparently, preservation of feature value [-hi] associated with /e/ is of more importance than preservation of the value [+hi] associated with /u/ (i.e. ∃-IDENT[-hi] dominates ∃-IDENT[+hi]). Even though one correspondent of /e/ coalesces with /u/, there is no conflict of ∃-IDENT constraints for

126

Struijke

color features since the color features of both input vowels can be preserved. Note that this instance of coalescence involves segments from different morphemes. Hence ∃-MORPHDIS is violated and must be low-ranking also. Because fission was established so that the feature specifications of /e/ could be divided between its output correspondents, these two segments are not identical, and the surface faithfulness constraint ∃-IDENT[-hi]∑∑ is low-ranking (as are surface constraints on color features). (136.) Feature movement as segmental fission and coalescence /u1-se2/ *[-hi]/non- ∃-IDENT ∃-IDENT INTEGRITY IDENT initial σ [-hi]∑∑ [-hi] [+hi] 1 u1se2 *! 2 u1si2 *! * * * 3 o12si2

∃-MORPH DIS

*

I now briefly discuss some data from Esimbi which provide evidence for the claim that features in this language do not literally reassociate, but rather appear to reassociate through a combination of fission and coalescence. In addition, these data show that feature value transfer is not in fact an instance of feature metathesis, which one might be tempted to conclude when considering alternations such as /u-se/ - [osi]. In this example, an underlying sequence of a high and a mid-high vowel is seemingly reversed to yield a sequence of a mid-high and a high vowel57 . However, forms containing a different prefix (indicating plural class 6) show that a single feature specification moves, rather than two feature values being switched. The infinitive prefix vowel in the Esimbi examples discussed so far is underlyingly high. However, the class 6 prefix vowel is underlyingly low (Hyman 1988). Concatenation of this low vowel with a

57

The Esimbi vowel system is as follows: high i ˆ u mid-high e ´ mid-low E ç low a

o

127

Existential Faithfulness

root containing a mid-high vowel does not result in a surface form with a mid-high - low sequence, as one would expect if feature metathesis were involved. Thus, /a-t´/ ‘ear’ does not surface as *[´-ta]. Instead, the corresponding surface form is [ç-tˆ], where the initial vowel is a ‘compromise’ in height between the underlying root vowel and the prefix vowel ([ç] is a mid-low vowel). The second vowel defaults to unmarked high. This form shows that features do not reassociate. Instead, segments coalesce. It is fairly common for a coalesced vowel to be a compromise between its contributing segments (see Casali 1996; Gnanadesikan 1997 for recent accounts). Before concluding this section, I would like to point out an important property of the proposed movement analysis. In Esimbi, [-hi] vowels can surface in word-initial syllables, and are banned in all following syllables. There is, then, only one vowel per word that can be specified as [-hi]. Thus, a cooccurence restriction on [-hi] vowels is an emergent characteristic of Esimbi. An important connection between feature movement and cooccurrence restrictions can be drawn. Both are driven by conspiring markedness and faithfulness constraints and both involve fission and coalescence. I return to this issue later. In summary, when we assume an existential definition of IDENT[F]IO faithfulness constraints, we predict that feature movement is due to a combination of segmental fission and coalescence. Feature movement is crucially made possible by the assumption that, in fission, only one output correspondent needs to preserve a given input feature specification to satisfy an existentially defined IDENT[F] constraint. The following sections discuss the mechanics of fission and coalescence (3.2) and show how they combine to produce dissimilatory feature movement in Sanskrit (3.3) and Cuzco Quechua (3.4).

128

Struijke

3.2 Dissimilation as a result of fission and coalescence

Two segments dissimilate when they are specified for one or more identical marked features. One of the segments usually stays unchanged, while the other seems to delete or become unmarked. I will focus here on repairs involving feature changes, because they show the importance of the existential quantification of faithfulness constraints (but see also section 1.2.3.3 in which both repairs are discussed). Let us return to the Akkadian example mentioned in the introduction of this chapter (Hume 1992; McCarthy 1981; Odden 1994; Suzuki 1998; data from Von Soden 1969). Labial segments cannot normally co-occur in words of this language. As mentioned earlier, I argue that this cooccurrence restriction follows from a basic segmental markedness requirement (*LAB), rather than a constraint that explicitly bans multiple labial segments such as an OCP-type constraint (Yip 1998b, following Goldsmith 1976, Leben 1973), or self-conjoined *LAB&*LAB (Alderete 1997; Itô and Mester 1996; MacEachern 1999; Suzuki 1998). *LAB is better satisfied when one of two labials changes to an unmarked coronal. I claim that this alternation does not incur a faithfulness violation, because the segment that remains labial in the output ensures preservation of the labial input specifications associated with both labial input segments. As in feature movement, dissimilation involves segmental fission and coalescence. This is depicted below.

129

Existential Faithfulness

(137.) Dissimilation as fission and coalescence (lines indicate relevant correspondence relations) [+nas][lab]

input:

output:

n1

[-cont][lab]

m1

a

e

r

e

e

e

r

e

b 12

[+nas][cor]

b2

[-cont][lab]

In the mapping above, all existential faithfulness constraints are satisfied. Input segment /m/ preserves its manner specification on its first output correspondent, and its place specification on the second correspondent. Input segment /b/ preserves both its place and manner specifications on its single output correspondent (coalesced with the second correspondent of /m/). The output now contains only one labial; hence violations of *LAB are minimized. The fact that the segmental inventory of Akkadian contains labials indicates that ∃-IDENT[lab] dominates *LAB. Thus, the markedness constraint is usually inactive in the language. However, it is able to emerge in the presence of two (or more) labial segments in the input. In essence, then, dissimilation is like The Emergence of The Unmarked in reduplication: segments undergo fission, and, instead of demanding two identical output segments, existentially defined faithfulness constraints permit one correspondent to conform to a lower-ranked markedness constraint. I discuss this parallelism further at the end of this section. The basic interaction of constraints in Akkadian is illustrated in the tableau below with the input just discussed. Candidate 1 is identical to the input. It contains two labials and hence incurs two violations of *LAB. In the optimal candidate (number 2), fission

130

Struijke

and coalescence applies as described above. This candidate is less marked than the first, because it contains one labial only: it violates *LAB once. At the same time, IO faithfulness is achieved because both labial specifications of the input are preserved on the remaining labial segment. The form violates INTEGRITY, because it involves segmental fission. Candidate 3 is phonetically identical to candidate 2: both contain one labial and one coronal. However, in candidate 3 fission and coalescence are not established, and so the final input labial does not have a labial correspondent. Hence it causes a fatal violation of undominated ∃-IDENT[lab]. (138.) Dissimilation as segmental fission and coalescence /m1a-ereb2/ ∃-IDENT *LAB ∃-IDENT [lab] [+nas] 1 m1a-ereb2 **! * 2 n1a-ereb12 3 n1a-ereb2 *! *

∃-IDENT [-cont]

INTEGRITY

*

Note that it is only the feature specification [lab] that dissimilates. Consider for instance the mapping /ma-lmenu/ - [ne-lmenu] 'loss, damage.' Marked manner and voice are not dissimilated to obtain the output *[te-lmenu]. Just as base-reduplicant faithfulness constraints can block The Emergence of The Unmarked in reduplication, constraints on surface correspondence can block ‘dissimilation to the unmarked’ (as suggested in Struijke and de Lacy 2000).

(139.) IDENT[F]∑∑ Let seg ∈ output be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an another output segment preserves the feature specification [αF] of that segment.

131

Existential Faithfulness

Blocking of dissimilation is illustrated in the tableau below for the input /ma-lmenu/. For reasons of space this tableau is divided into three sub-tableaux. The first illustrates dissimilation of labial place. The second shows blocking of manner dissimilation, and the third blocking of voice dissimilation. The candidate set discussed includes the faithful candidate and candidates showing different degrees of dissimilation. Candidate 2 dissimilates place only; candidate 3 dissimilates both place and manner; and candidate 4 dissimilates place, manner and voicing specifications. Following Struijke and de Lacy (2000), I will indicate surface relatedness by the subscript ∑. Labial dissimilation is allowed because *LAB dominates IDENT[lab]∑∑: it is more important to reduce labial markedness than it is to maintain labial identity between output segments related to the same input segment. Nasal dissimilation is not permitted, because IDENT[+nasal]∑∑ dominates *NASAL: it is more important to preserve nasal identity in a surface relation than it is to reduce nasal markedness. Similarly, voice dissimilation is prohibited because IDENT[+voice]∑∑ dominates ∃-IDENT[+voice].

(140.) No dissimilation to the entirely unmarked ∃-IDENT[lab] *LAB /ma-lmenu/ 1 ma-lmenu **! * 2 n∑a-lm∑enu 3 d∑a-lm∑enu * 4 t∑a-lm∑enu * /ma-lmenu/ 1 ma-lmenu 2 n∑a-lm∑enu 3 d∑a-lm∑enu 4 t∑a-lm∑enu

58

∃-IDENT[lab] *LAB IDENT[lab] ∑∑

IDENT [+nasal]∑∑

!

*! *!

IDENT[lab]∑∑ * * * ∃-IDENT [+nasal]

*NASAL58 ** ** * *

For reasons of clarity, I show only thos violations incurred by the segments under discussion.

132

Struijke

/ma-lmenu/ 1 ma-lmenu 2 n∑a-lm∑enu 3 d∑a-lm∑enu 4 t∑a-lm∑enu

∃-IDENT[lab] *LAB IDENT[lab] ∑∑

IDENT [+voice]∑∑

!

*!

∃-IDENT [+voice]

*VOICE59 ** ** ** *

When all relevant surface faithfulness constraints are low-ranking we expect the least marked segment to emerge in dissimilation. This is the main point made in Struijke and de Lacy (2000). Note, however, that a change to the completely unmarked is unlikely. It would require a language in which manner, place, and laryngeal dissimilation occur independently from one another throughout the language. To my knowledge, no such language exists.60 In languages not exhibiting dissimilation, fission might be prohibited by highranking of INTEGRITY, or by a ranking in which all surface faithfulness constraints dominate all segmental markedness constraints, and no markedness constraint is able to force fission and coalescence. Consider for example the lack of labial dissimilation in English, exemplified in the tableau below. Dissimilation in candidate 1 is prevented by the ranking IDENT[lab]∑∑ >> *LAB. No dissimilation has taken place in candidate 2 and 3. These forms are segmentally identical, but suboptimal candidate 3 involves gratuitous fission and coalescence. In this candidate fission is not distributing features: the marked place feature is preserved on both output segments. Hence, fission does not serve to reduce markedness. Coalescence of the first correspondent is not needed to guarantee faithfulness, since the second correspondent of fissioned /p/ preserves the labial feature specification. Because fission and coalescence are gratuitous, this candidate is 59

Again, I show only those violations incurred by the segments under discussion In both Konni (Gur; Cahill 1999) and Yimas (Austronesian; Foley 1991), rothic dissimilation results in /rr/ - [rt] alternations. This, however, does not appear to constitute counter examples to this observation (contrary to Struijke and de Lacy 2000). Like [t], [r] seems to be specified as [-cont] in these languages, and the segments are also alternants outside the domain of dissimilation. 60

133

Existential Faithfulness

harmonically bounded by optimal candidate 2, which does not involve such instances of multiple correspondence. Candidate 3 can never be optimal because it incurs a violation of INTEGRITY. Thus, fission and coalescence are never generated unless they are coerced by markedness constraints.

(141.) No unmotivated fission and coalescence ∃-IDENT[lab] IDENT[lab]∑∑ /p1ap2a/ 1 p12at2a *! 2 p1ap2a 3 p12ap2a

*LAB * ** **

INTEGRITY * *!

To conclude this introductory section on dissimilation, I compare it with reduplicative feature TETU. Conceptually these phenomena are similar in that they both prevent a sequence of potentially identical segments from surfacing. I have formalized this similarity through similar constraint interactions. The rankings that force reduplicative feature TETU and dissimilation are summarized in the table below.

(142.) Dissimilation to the unmarked compared to TETU in reduplication dissimilation

∃-IDENT[F] >> *[F] >> IDENT[F]∑∑

blocking of dissimilation

IDENT[F]∑∑ , ∃-IDENT[F] >> *[F]

reduplicative TETU

∃-IDENT[F] >> *[F] >> IDENT[F]BR

blocking of reduplicative TETU

IDENT[F]BR , ∃-IDENT[F] >> *[F]

I suggested in chapter 1 that base-reduplicant faithfulness may be subsumed under a general surface faithfulness relation (what I will call 'generalized surface faithfulness'). Under that proposal the table above can be reduced to the one below.

134

Struijke

(143.) Dissimilation and reduplicative TETU under generalized surface dissimilation

faithfulness

∃-IDENT[F] >> *[F] >> IDENT[F]∑∑

reduplicative TETU blocking

IDENT[F]∑∑ , ∃-IDENT[F] >> *[F]

Given generalized surface faithfulness, dissimilation and reduplicative TETU require the same ranking of IO faithfulness, surface faithfulness, and featural markedness constraints. Yet the presence of one of these patterns in a language does not necessarily imply the presence of the other. This is because they are evaluated differently by some other constraints. Fission in dissimilation always incurs a violation of INTEGRITY. When segments belonging to different morphemes coalesce to facilitate dissimilation, ∃-MORPHDIS is violated in addition. However, fission in reduplication does not incur an INTEGRITY violation (see section 2.4.3), and coalescence does not take place. Thus, assuming that reduplicative feature TETU and feature dissimilation incur the same surface faithfulness violations, reduplicative TETU incurs a subset of the violations incurred by dissimilation. Consequently, dissimilation implies the presence of feature TETU in reduplication, but not vice versa.61 Two languages mentioned in this chapter support one side of this prediction. Sanskrit aspirated and murmured segments dissimilate in roots, and are never copied faithfully into reduplicants (i.e. bases and reduplicants cannot both contain marked segments specified for [+SG]). Similarly, liquids in Yimas (Foley 1991; Odden 1994) dissimilate and are never copied faithfully into reduplicants (i.e. bases and reduplicants cannot both contain marked liquids). Support for the other side of the prediction comes from languages that do not exhibit dissimilation, but show reduplicative

61

TETU affecting reduplicants that are morphologically roots incurs violations of root faithfulness constraints. The violations incurred by TETU affecting root reduplicants are therefore not a subset of those incurred by dissimilation, and dissimilation does not imply TETU in root reduplicants.

135

Existential Faithfulness

TETU which leads to segments that are entirely unmarked. For instance, Alderete et al. (1999) analyze Tübatulabal (Uto-Aztecan; Voegelin 1958) reduplication in which any copied onset consonant defaults to glottal stop, with unmarked place and manner feature specifications. Outside of reduplication, marked place and manner can cooccur, as shown in the bases of [/i:-bi:bi:win] 'to play jew’s harp' (containing three labials) and [/i:-mˆ:hli:n] 'to hurt her' (containing two nasals). These observations suggest that reduplicative feature TETU is a specific type of cooccurrence restriction. Further research must determine whether the suggested implicational relation between reduplicative TETU and dissimilation is an accurate one. If we find languages in which marked feature dissimilation is found outside the domain of reduplication, but not in reduplicative TETU, we have evidence that reduplicated words and unreduplicated words are subject to different faithfulness cosntraints holding over output segments (BR faithfulness and surface faithfulness). Regardless of the outcome, it is clear that dissimilation and reduplicative TETU are similar in important ways. Both are driven by a need to improve markedness, which can be satisfied only because input-output faithfulness is achieved through multiple correspondence. This concludes the general introduction to dissimilation. I have argued that dissimilation is a way to improve markedness without loss of underlying information. The existential definition of faithfulness constraints ensures that dissimilation does not incur faithfulness violations. A full account of generalizations associated with dissimilation, such as the similarity effect, 62 falls outside the scope of this dissertation,

62

In some languages, cooccurence restrictions are enforced on segments that agree in multiple feature specifications, but not on segments that agree in a single feature specification, even though only one feature dissimilates (Frish 1996; Frish, Broe and Pierrehumbert 1997; MacEachern 1999; Padgett 1992; Pierrehumbert 1993; Yip 1989). Surface faithfulness constraints can account for this effect. For instance, in Russian, coronals only dissimilate when they agree in continuency and sonorancy (Padgett 1992). Surface faithfulness constraints on these features must be highranking.

136

Struijke

and I leave them for future research.63 I now turn to case studies of Sanskrit (section 3.3) and Cuzco Quechua (section 3.4).

3.3 Case study: Sanskrit

This section gives an analysis of two patterns in Classical Sanskrit. First, in a number of roots the location of a murmured segment is entirely predictable. In some previous accounts this was accounted for by feature reassociation. Second, [h], aspirated segments, and murmured segments are not allowed to cooccur in roots. In previous analyses this was accounted for by means of the OCP or dissimilation alternations. Thoughout this section I refer to these patterns descriptively as ‘movement’ and ‘dissimilation.’ I show that feature movement is most likely to take place when the segments involved are alike. In addition, this section shows that surface faithfulness and IO faithfulness constraints can prevent simultaneous movement of multiple feature specifications. It also gives a brief comparison of the proposed Sanskrit analysis with previous accounts. The predictable location of murmured segments and some of the cooccurence facts can be explained by the ban on laryngeally marked segments in certain environments. I turn to this issue first. Section 3.3.2 focuses on murmur movement, known as ‘throwback’ in the Sanskrit literature. Murmur throwback is not required in certain environments (Bartholomae’s Law) this issue is addressed in section 3.3.3. Section 3.3.4 discusses the laryngeal cooccurrence restrictions (Grassmann’s Law).

3.3.1 Ban on laryngeally marked segments

63

But see appendix II on the proximity effect.

137

Existential Faithfulness

Sanskrit shows several laryngeal contrasts in stops and affricates. These sounds can be voiceless, voiced, aspirated, or murmured (breathy voice). I will use the term 'stops' to refer to both stops and affricates. (144.) Sanskrit consonant inventory voiceless voiced voiceless aspirated voiced murmured

p b pH b˙

t d tH d˙

ˇ Í ˇH Í˙

c& j& c&H j&˙

m v

s n l r

ß ˜

C ¯ j

k g kH g˙ h N

Laryngeally marked segments cannot surface word-finally or before obstruents. In the examples in (20) below, they neutralize to unmarked plain voiceless stops. The forms in (21) show that the marked segments are allowed before sonorants. Data are from Whitney (1885, 1889), unless otherwise indicated. (145.) No marked laryngeal feature values word-finally or before obstruents -> neutralization /suhrd/ suh«rt ‘goodhearted’ /agnimath/ agnimat ‘being near the fire’ virut ‘obstruction’ /virud˙/ /svid+ syati/ svetsyati ‘sweat’ (future) sa:tsyati ‘succeed’ (future) /sa:d˙+ syati/ (146.) Marked laryngeal feature values are allowed before sonorant segments /svid/ svidya:t ‘sweat’ (aorist) /svid/ svedate ‘sweat’ (present) dag˙nuyat ‘reached to’ /dag˙/ /ba:d˙/ ba:d˙ate ‘oppress’ (present) /paˇH/ paˇHyate ‘read’ (passive) /paˇH/ paˇHati ‘read’ (present)

138

Struijke

Consonants in all medial clusters are heterosyllabic, as evidenced by versification and reduplication (McCarthy and Prince 1986; Steriade 1997 and references quoted therein). Thus, the laryngeally marked stops in the forms below surface in coda. (147.) Syllabification of stop-sonorant clusters (σ boundaries indicated by periods) /svid/ svid.ya:t ‘sweat’ (aorist) ˙ ˙ /dag / dag .nu.yat ‘reached to’ /paˇH/ paˇH.ya.te ‘read’ (passive)

These data indicate that laryngeal neutralization does not affect segments in coda position, but segments which do not precede sonorants. I attribute this behavior to a highranking markedness constraint LARREL, loosely based on constraints proposed in Steriade 1997. (148.) LARREL A laryngeally marked obstruent (cluster) must release into a sonorant

In avoiding a violation of this constraint, laryngeally marked stops usually become plain voiceless. However, other repairs are also found. The feature value [+murmur] is preserved on another segment whenever possible, and aspirated segments induce vowel epenthesis word-internally. This is summarized in the table below. (149.) Alternations of laryngeally marked stops in pre-obstruent / word-final environments murmured stops

throwback - else neutralization

aspirated stops

epenthesis / neutralization

voiced stops

neutralization

139

Existential Faithfulness

The different repairs affecting aspirated and murmured segments suggest that murmur and aspiration are different features, and it reflects the fact that they differ phonetically (Ladefoged 1971; Ladefoged and Madieson 1996). I therefore assume the distinct features murmur and aspiration. Aspirated and murmured segments behave like a class with respect to cooccurrence restrictions: roots cannot contain more than one segment with a [+murmur] or [+aspiration] feature value. I refer to this class as [+SpreadGlottis]. I adopt Padgett’s (1995) Feature Class Theory. Features are specified for class membership, and constraints can refer to laryngeal features as a class, to spread-glottis features as a class, or to each laryngeal feature individually. In their basic effect, feature classes replace nodes in Feature Geometry.64 The constraint LARREL must outrank faithfulness constraints on marked laryngeal features in order to drive neutralization. These faithfulness constraints are ∃-IDENT[+lar], ∃-IDENT[+SG], ∃-IDENT[+murmur], ∃-IDENT[+asp], ∃-IDENT[+voice]. The tableaux below contain the faithfulness constraints on the individual features.

(150.) Neutralization of voiced stops LARREL /suhrd/ 1 suhrd 2 suhrt

∃-IDENT [+voice]IO

∃-IDENT [+murmur]IO

∃-IDENT [+asp]IO

∃-IDENT [+murmur]IO

*! *

(151.) Neutralization of aspirated stops /agnimath/ LARREL ∃-IDENT [+voice]IO h 1 agnimat *! 2 agnimat

64

∃-IDENT [+asp]IO

*

However, there are some differences between the two models (see Padgett 1995).

140

Struijke

(152.) Neutralization of murmured stops LARREL ∃-IDENT /virud˙/ [+voice]IO ˙ *! 1 virud * 2 virut

∃-IDENT [+asp]IO

∃-IDENT [+murmur]IO *

I now turn to preservation of murmur specifications in ‘throwback.’

3.3.2 [+Murmur] movement

In Sanskrit, a [+murmur] feature specification is not always lost when the stop with which it is underlyingly associated surfaces before an obstruent or word-finally. In particular, when an underlying murmured consonant is root-final, its murmur specification can be preserved on a root-initial voiced stop. Thus, in surface forms related to /bud˙/ ‘to wake’ and /bad˙/ ‘to oppress’ given below, [+murmur] is associated with the root-final segment when the following suffix starts with a sonorant, but with the root-initial segment when the suffix starts with an obstruent. (153.) Words containing the root /bud˙/ ‘to wake’ (root printed in bold) bod˙ati present ˙ bud yate passive ˙ bod ayati causative bub˙utsati b˙otsyati

desiderative future

Words containing the root /bad˙/ ‘to oppress’ (root printed in bold) bad˙ate present ˙ bad yate passive ˙ bad ayati causative bib˙atsate

desiderative

141

Existential Faithfulness

I propose that ‘movement’ takes place in the desiderative and future forms (among others) because it avoids a violation of LARREL without incurring a violation of ∃-IDENT[+murmur]. The basic schema for feature value transfer was laid out in section 3.1. The specific analysis of Sanskrit is illustrated in the following diagram: (154.) Sanskrit feature value transfer (only relevant feature specifications present in the input are shown) [-murmur]

input:

output:

b



[+murmur]

u

u

t



s

s

u

u

[+murmur]

Candidate 1 of tableau (30) embodies the faithful parse of the /bud˙+su/ input. The feature specification [+murmur] surfaces before an obstruent and fatally violates LARREL. Candidate 2 repairs this violation by not preserving the [+murmur] value. By doing so, it fatally violates ∃-IDENT[+murmur]. Candidate 3 is the optimal form. It avoids both of these violations, because the [+murmur] value is preserved on a segment that precedes a sonorant. Feature value transfer can take place because the fission and coalescence it involves are allowed. First, INTEGRITY is low-ranking. Second, the fissioned output segments are not identical with respect to murmur values, and cause a violation of the surface faithfulness constraint relevant to this feature, but this constraint is low-ranking. (In this particular form, the surface constraint on place is also violated.) Third, coalescence resulting in a murmured segment is permitted in the optimal candidate because IO faithfulness to [-murmur], which is relevant for the root-initial segment, is

142

Struijke

low-ranking. To ensure that murmur throwback of candidate 3 is more harmonic than vowel epenthesis of candidate 4, all these constraints must be ranked lower than M-SEG. (155.)

Sanskrit feature value transfer ˙

/b1ud 2 + su/

LARREL

1 bud˙su 2 butsu 3 b˙12ut2su 4 bud˙Vsu

*!

∃-IDENT M-SEG INTEG [+murm]IO

IDENT[+murm]∑∑ IDENT[cor]∑∑

∃-IDENT [-murm]IO

*! *

*

*

*!

[+Murmur] cannot be thrown back onto consonants other than voiced stops. When the root-initial segment is not a voiced stop, the [+murmur] specification is lost. (156.) [+murmur] cannot be thrown back onto segments other than voiced stops /yud˙/ yotsyati *y˙otsyati ‘fight’ (future) ˙ /lab˙/ lapsyate *l apsyate ‘take’ (future) ˙ /sid˙/ setsyati *s etsyati ‘repel’ (future)

On the one hand, this could be due to a licensing constraint, undominated in Sanskrit, that allows only voiced stops and affricates to be murmured. (157.) LICENSE[+murmur]: If a segment is specifified as [+murmur], then it must be specified as [-sonorant, -continuant, +voice]

On the other hand, the same effect is achieved when the surface faithfulness constraints IDENT[+voice]∑∑ and IDENT[-cont]∑∑ dominate ∃-IDENT[+murm]. With this ranking, it is more important that fissioned segments agree in voicing and manner specifications, than it is for murmur to be preserved. Thus, surface faithfulness constraints can account for the generalization that feature movement is most likely when the segments involved are featurally alike.

143

Existential Faithfulness

The tableaux below summarize the two possible explanations of the fact that murmur cannot be thrown back onto a fricative, or a voiceless segment. The faithful form (candidate 1) is ungrammatical because [+murmur] surfaces before an obstruent. Candidate 2 shows throwback of [+murmur] onto a voiceless fricative, so that both LARREL and ∃-IDENT[+murm] are satisfied. However, throwback onto such a segment results in a fatal violation of the licensing constraint in tableau (33), and fatal violations of IDENT[+voice]∑∑ and IDENT[-cont]∑∑ in tableau (34). Candidate 3 is optimal even though it does not preserve the murmur specification: it satisfies the highranking constraints at the expense of ∃-IDENT[+murmur]. (158.) No throwback onto a voiceless consonant due to licensing constraint LARREL ∃-IDENT /s1id˙2+ sya + ti / LICENSE [+murmur] [+murmur] IO

˙

1 sid syati 2 s˙12et2syati 3 setsyati

*! *! *

(159.) No throwback onto a voiceless consonant due to surface faith requirements IDENT LARREL ∃-IDENT /s1id˙2+ sya + ti / IDENT [+voice]∑∑ [-cont]∑∑ [+murmur] IO

˙

1 sid syati 2 s˙12et2syati 3 setsyati

*! *!

* *

The feature [+murmur] can only be thrown back onto segments in the root. For instance, given the input /ba-lad˙su/, [+murmur] does not surface on the voiced stop in the prefix: *[b˙a-latsu]. Instead, the feature specification is not preserved: the surface form is [balatsu]. Apparently, coalescence of segments belonging to different

144

Struijke

morphemes is prevented by ∃-MORPHDIS. The definition of this constraint is repeated below.65 (160.) ∃-MORPHDIS (Morphemic Disjointness) Let seg-x ∈ input Morph1, and seg-y ∈ input Morph2, and seg-x', seg-y' ∈ output, and segxℜseg-x', seg-yℜseg-y'. Then there is some seg-x'≠ seg-y'. (McCarthy and Prince 1995 (adapted)). The disjointness of input morphemes is preserved in the output. (segments belonging to different morphemes do not coalesce)

In conclusion, Sanskrit murmur throwback is analyzed straightforwardly as a combination of segmental fission and coalescence. This next section addresses the question of why feature movement is restricted to one member of the laryngeal class.

3.3.2.1 Multiple feature movement?

Given the fact that murmur can move in Sanskrit, one might expect that [+asp] and [+voice] can be transferred as well. After all, voiced, aspirated and murmured segments belong to the same class of segments prohibited from pre-obstruent and word-final positions, and movement is allowed. However, [+asp] and [+voice] are not thrown back. In this section, I show that movement of these features can be prevented by input-output faithfulness constraints or IDENT [F]∑∑ constraints. When an aspirated segment is followed by an obstruent, vowel epenthesis takes place, rather than aspiration throwback. When an aspirated stop occurs word-finally, it neutralizes to a plain voiceless stop.

65

Highranking of this constraint also accounts for the fact that dissimilation does not occur across morpheme boundaries.

145

Existential Faithfulness

(161.) No throwback of aspiration or neutralization before an obstruent - epenthesis instead input grammatical *throwback *neutralization gloss /paˇH-syati/ pa:ˇHißyati *pHaˇsyati *paˇsyati ‘read’ (future) /putH-syati/ potHißyati *pHotsyati *potsyati ‘crush’ (future) The [+voice] feature specification is never thrown back either. Voiced stops always neutralize to plain voiceless when preceding obstruents or occurring word-finally. (162.) No throwback of voicing or epenthesis - devoicing instead input grammatical no throwback no epenthesis /pad/ patsyati *batsyati *padisyati /tud/ totsyati *dotsyati *todisyati /svid/ svetsyati *zvetsyati *svedisyati

gloss (future) ‘go’ ‘push, thrust’ ‘sweat’

The prohibition against [+asp] throwback cannot be due to a general prohibition on fission because INTEGRITY is low-ranking. Rather, coalescence of segments differing in specification for the feature [aspiration] must be prohibited (by highranking ∃-IDENT[-aspirate]IO) or the particular instance of distributing fission must be prohibited (because it would result in surface correspondents differing in specifications for the feature [aspiration]). Since epenthesis (not throwback) is the optimal repair, at least one of these faithfulness constraints must dominate the anti-epenthesis constraint M-SEG. This is shown in the tableau below: (163.) [+asp] is preserved through epenthesis66 ∃-IDENT[+asp]I ∃-IDENT[-asp] IO / M-SEG /putH-syati/ IDENT [+asp]∑∑ O 1 putsyati 2 putHisyati 3 pH12ut2syati

INTEGRITY

*! * *! /*!

66

*

Recall that epenthesis does not take place when an underlyingly aspirated segment surfaces word-finally (/agnimath/ - [agnimat], *[agnimathi]). This must be due to an undominated Alignment constraint (McCarthy and Prince 1993b) requiring segments at the morpheme edge to be aligned to the word edge.

146

Struijke

In contrast to [+murmur] and [+asp], the specification [+voice] is never preserved before obstruents or word-finally. This indicates that ∃-IDENT[+voice] ranks below constraints banning voice throwback and vowel epenthesis. That is, it must be dominated by ∃-IDENT[-voice] RT67 (which bans coalescence of a voiced segment with a voiceless segment resulting in a voiced segment) or IDENT [+voice]∑∑ (which bans fission resulting in output segments that do not agree in voicing). In addition, it must be dominated by M-SEG (which prevents epenthesis). (164.)

[+voice] is not preserved /tud-syati/ ∃-IDENT IDENT [-voice]RT [+voice]∑∑

1 totsyati 2 todißyati 3 d12ot2syati

M-SEG

∃-IDENT

INTEGRITY

[+voice] * *!

*!

*!

*

We can now explain why several feature values cannot be transferred together. For instance, [+murmur] and [+voice] cannot be transferred in tandem to a voiceless stop.68 (165.) Murmur and voice cannot both be thrown back onto a voiceless stop krotsyati *g˙rotsyati ‘be angry’ (future) /krud˙/ ˙ ˙ /triߡub / triߡup *triß∂Í up ‘a metre of 4 x 11 syllables’

The feature values [+murmur] and [+voice] cannot move together because [+voice] cannot be transferred. Recall that murmur cannot be thrownback onto a voiceless

67

This constraint must be specific to roots. I show in the section on Bartholomae’s Law that an affixal [-voice] specification can be lost. 68 The impossibility of throwback in the first example is not due to a prohibition on throwback in a complex onset, as is evidenced by /drug˙/ -> [d˙roksyati] - ‘be hostile’ (future). In this form, the marked feature value on the first consonant would be licensed by the following sonorant.

147

Existential Faithfulness

segment on its own because of a licensing constraint or a surface faithfulness constraint on voicing (tableaux 33 and 34). Thus, feature values do not automatically get a ‘free ride’ once multiple correspondences are established. It is entirely dependent on the constraint ranking of a language whether a second or third feature can piggyback on a first feature. Feature movement is expected to take place most easily when the segments involved are alike. The more two segments resemble each other, the fewer violations of IO faithfulness constraints are incurred by colalescence, and the fewer violations of surface faithfulness constraints are incurred by fission. However, the theory does not rule out transfer of multiple feature specifications. Given the right constraint ranking, multiple feature movement is expected to take place. Obviously, transfer of each feature must be forced by a markedness constraint banning it from remaining in its original context. Thus, features belonging to a class banned in a given environment by means of a single markedness constraint are more likely to move together than features that are banned in the same environment by means of different markedness constraints. This discussion brings us back to the idea expressed in chapters 1 and 2 that the two output correspondents resulting from distributing fission are of equal status, even though one of these segments often preserves only one of the input’s feature specifications, and the other preserves all other specifications. Neither of the output correspondents is secondary to the other, and the fact that one shares more feature specifications with the input than the other has no particular status in the theory. As I have just argued, there is pressure to move as few features as possible, because coalescence incurs a violation of ∃-IDENT[F]IO constraints69 and distributing fission

69

Unless coalescing segments share the relevant feature specification. See the analysis of Grassmann's law below.

148

Struijke

incurs violations of an IDENT[F]ΣΣ constraint. Thus the ranking of such constraints determines which segment is more faithful in fission. The discussion of Sanskrit laryngeal features clearly illustrates the necessary conditions for feature transfer. It shows that movement of feature values is permitted by a very specific constraint ranking. A grammar must contain a highranking markedness constraint banning a certain feature specification in a particular environment, as well as a highranking IO faithfulness constraint demanding preservation of that feature. These constraints must dominate those banning fission in general (INTEGRITY), a particular instance of distributing distributing fission (IDENT[F]ΣΣ), and a particular instance of coalescence (∃-IDENT[F]IO). If any of these specific ranking properties is not found in a grammar, feature movement is not observed. This explains, perhaps, why this phenomenon is relatively rare across the world’s languages. The following sections complete the analysis of murmur distribution in Sanskrit. Section 3.3.3 discusses the absence of murmur throwback when a root such as /bud˙/ is followed by a stop-initial suffix (Bartholomae’s Law). Section 3.3.4 gives an account of the fact that no root contains more than one murmured segment (Grassmann’s Law).

3.3.3 Bartholomae’s Law

In Sanskrit, throwback of [+murmur] does not take place when a root-final murmured segment is followed by a stop-initial suffix. In the clusters created through such morpheme concatenation, only the second stop is phonetically murmured. The phonetic transcriptions of such clusters are given below.

149

Existential Faithfulness

(166.) Bartholomae’s Law70 - phonetic transcription input phonetic form gloss budd˙a ‘awakened, enlightened’ (nom. infl.) /bud˙+ta/ /bud˙+ti/ budd˙i ‘insight, understanding’ (nom. infl.) ˙ ˙ /dug +te/ dugd e ‘milk’ (3rd sg mid) ˙ ˙ /dug +tas/ dugd as ‘milk’ (3rd dual act) ˙ ˙ /dug +tHas/ dugd as ‘milk’ (2nd dual act) ˙ ˙ /dug +tHa/ dugd a ‘milk’ (2nd pl act) ˙ ˙ ˙ /dug +d i/ dugd i ‘milk’ (2nd sg pres imp) ˙ ˙ ˙ /bud +d i/ budd i ‘wake’ (2nd sg pres imp) ˙ ˙ ˙ /dag +d i/ dagd i ‘reach to’ (2nd sg pres imp)

Even though only the second stop is phonetically murmured, there is evidence that both segments in these clusters are assimilated entirely with respect to laryngeal feature specifications, and hence both are murmured phonologically (Borowsky and Mester 1983; Lombardi 1991). This assumption is supported by Whitney's statement (1889, p. 155) that murmured geminates are sometimes spelled C˙C˙. (167.) Bartholomae’s Law (phonological transcription) input phonological form gloss ˙ ˙ ˙ /bud +ta/ bud d a ‘awakened, enlightened’ (nom. infl.) ˙ ˙ ˙ /bud +ti/ bud d i ‘insight, understanding’ (nom. infl.) ˙ ˙ ˙ /dug +te/ dug d e ‘milk’ (3-sing-mid) ˙ ˙ ˙ /dug +tas/ dug d as ‘milk’ (3-dual-act) ˙ ˙ ˙ /dug +tHas/ dug d as ‘milk’ (2-dual-act) ˙ ˙ ˙ /dug +tHa/ dug d a ‘milk’ (2-pl-act) ˙ ˙ ˙ ˙ /dug +d i/ dug d i ‘milk’ (2nd sg pres imp) ˙ ˙ ˙ ˙ /bud +d i/ bud d i ‘wake’ (2nd sg pres imp) ˙ ˙ ˙ ˙ /dag +d i/ dag d i ‘reach to’ (2nd sg pres imp)

70

Concatenation of a root-final murmured stop and an affix-initial non-murmured voiced stop is not attested (Anderson 1974; Schindler 1976).

150

Struijke

I assume that markedness constraints can refer to clusters containing segments that agree in feature specifications (cf. Smolensky 1995; Cole and Kisseberth 1994 on feature domains). LARREL is such a constraint. Its definition is repeated below. (168.) LARREL A laryngeally marked obstruent (cluster) must release into a sonorant

Each cluster in (42) satisfies the constraint, and there are no phonological reasons for the root-final segments to become laryngeally unmarked (cf. Itô 1986; Itô and Mester 1993). In fact, there is pressure for them to be marked. Laryngeal assimilation is forced by the constraint AGREE[lar]. (169.) AGREE[lar]: Adjacent segments must have the same value for laryngeal features (Lombardi 1996, 1999; Beckman 1997; Bakovic 1999)

Following most work on assimilation, I assume that positional faithfulness constraints determine the direction of assimilation. In Sanskrit, assimilation is regressive because segments in the onset must be faithful71 . Tableau (45) considers different output candidates for the input /bud˙+ta/, all of which satisfy highranking LARREL and ∃-IDENT[+murmur]. In candidate 1, throwback has taken place. The candidate violates constraints banning murmur movement, namely INTEGRITY, IDENT[+murmur]∑∑, and ∃-IDENT[-murmur]IO. Movement takes place in candidate 2 also. Rather than being thrown back, the feature is thrown forward to the initial segment of the suffix. In addition to the violation marks associated with throwback, this candidate violates the agreement constraint, because the segments in the consonant

71

The positional faithfulness constraint must be defined such that if an input segment has a correspondent in an onset position, that correspondent must be faithful.

151

Existential Faithfulness

cluster do not agree in their specifications for the feature [murmur]. Candidate 3 is optimal because it satisfies the agreement constraint, and incurs no violations associated with fission in feature movement. However, assimilation causes a violation of ∃-IDENT[-murmur]IO because an unmurmured segment becomes murmured. We have seen repeatedly that this faithfulness constraint ranks below constraints banning fission. ∃-IDENT[-voice]IO is violated also and must be low-ranking.72, 73 (170.) Bartholomae's Law AGREE /b1ud˙2+t3a/ [lar]

IDENT [+voice]∑∑

˙

1 b 12ut2t3a 2 b1ud 2d ˙23a 3 b1ud˙2d˙ 3a

INTEGRITY IDENT [+murm]∑∑ *! *

*!

∃-IDENT

∃-IDENT

[-vce]IO

[-murm]IO

* *

I now turn to the cooccurrence restrictions on aspirated and murmured segments.

72

Thus, an affix-initial voiceless stop can obtain voicing from a root-final segment. Recall, however, that a root-initial voiceless stop cannot do so (i.e. voicing cannot be thrown back onto a voiceless stop). This difference is due to the different status of roots and affixes with respect to faithfulness. Both roots and affixes are subject to IO faithfulness, but roots are subject to extra faithfulness requirements. ∃-IDENT[-voice]RT apparently ranks higher than the general ∃-IDENT[-voice]IO. 73 While the majority of suffixes with initial murmured segments cause regressive assimilation of murmur, there are three such suffixes which condition throwback in roots. They are /-d˙ve/, /-d˙vam/, and /-b˙is/. Example mappings are /bud˙+b˙is/ - [b˙udb˙is]; /dug˙+d˙vam/ - [d˙ugd˙vam]; /dug˙+d˙ve/ [d˙ugd˙ve]. Some researchers have assumed that these elements form separate words, so that murmur assimilation is blocked (Sag 1976; Borowsky and Mester 1983). Others have claimed they are simply exceptions (Sag 1976; Schindler 1976). One could also assume that these suffixes form a separate affix class, which I dub ‘-d˙..’. This class of affixes is then subject to the following alignment constraint (Generalized Alignment, McCarthy and Prince 1993b): ALIGN[-d˙..] LEFT, [+MURMUR]DOMAIN LEFT (Align the left edge of the morphemes /-d˙ve/, /-d˙vam/, and /-b˙is/ to the left edge of some segment (cluster) specified as [+murmur]). This constraint must dominate AGREE[lar] in order to prevent assimilation of [+murmur] across the root-suffix boundary. ALIGN-d˙.. AGREE INTEGRITY ∃-IDENT ˙ ˙ /b1ud 2+b 3is/ [+murm] L [lar] [-murm] 1 bud˙b˙3 is 2 b˙12ud2 b˙3 is

*! *

152

*

*

Struijke

3.3.4 Grassmann’s Law

This section gives an account of laryngeal cooccurrence restrictions in Sanskrit (Grassmann’s law - Grassmann 1863; see MacEachern 1999 for a recent account). I show that in some forms, dissimilation is indistinguishable from feature movement. Sanskrit roots are not allowed to contain two segments specified as [+SpreadGlottis] (i.e. murmured and aspirated segments as well as [h]). The specifics are summarized in the table below. (171.) Cooccurrence restrictions in Sanskrit 1. 2. 3.

Only one murmured segment per root Only one aspirated segment per root Only one [h] per root

4.

Murmured and aspirated segments do not cooccur in roots

5. 6.

[h] and aspirated segments do not cooccur in roots [h] and murmured segments do not cooccur in roots

Restrictions do not hold across root-suffix boundaries: - [bib˙r-tHa] ‘bear’ (2nd. pl. active present) - [da-d˙a:tHe] ‘put’ (2nd dual middle present)74 - Bartholomae's law

I assume that aspirated segments, murmured segments, and [h] are specified as [+SpreadGlottis] and argue that the cooccurrence restrictions are caused by the general markedness constraint *SG. The focus of this section is first on forms containing murmured segments. The murmured examples extend straightforwardly to the aspirated

74

Data from Anderson (1970).

153

Existential Faithfulness

examples. At the end of this section the remaining concurrence restrictions are discussed briefly. Although researchers agree that laryngeal cooccurrence restrictions exist in Sanskrit roots, there is disagreement as to which alternations should be seen as instances of dissimilation. I have postulated that the [+murmur] specification in forms such as ‘to wake’ originates on the root-final segment: /bud˙/ (as did Whitney 1889; Allen 1951; Sag 1974, 1976; Borowsky and Mester 1983; Lombardi 1991; following Panini). However, a number of early generative accounts assumed that these roots contain two murmured consonants underlyingly, one of which loses its murmur feature on the surface (Zwicky 1965; Kiparsky 1965; Anderson 1970, 1974; Phelps 1975; Hoard 1975, amongst others). This is a dissimilation analysis of the murmur distribution facts analyzed earlier in section 3.3.2. It assumes the underlying form /b˙ud˙/ for the root ‘to wake,’ and /b˙ad˙/ for the root ‘to oppress’ (rather than /bud˙/ and /bad˙/ as assumed in section 3.3.2). Under this analysis, murmur is sometimes lost on the root-initial segment, and sometimes on the root-final segment. (172.) Dissimilation analysis words containing the root ‘to wake’ - /b˙ud˙/ root-final murmur bod˙-ati root-initial murmur bu-b˙ut-sati words containing the root ‘to oppress’ - /b˙ad˙/ root-final murmur bad˙-ate root-initial murmur bib˙at-sate

present desiderative

present desiderative

This analysis is based on the generalization that no Sanskrit root surfaces with two murmured segments, a fact that must be explained by any study of the laryngeal distribution patterns. Under Richness of the Base (Prince and Smolensky 1993; see chapter 1), an OT analysis must assume inputs which contain two murmured consonants.

154

Struijke

Such forms never surface, and must be filtered out by the grammar of Sanskrit. Put differently, the grammar must generate a well-formed output even for an input with two murmured segments. I argue (following a suggestion by John McCarthy) that /C˙VC˙/ inputs of the rich base always surface as [C˙VC]. Conversely, inputs of the shape /CVC˙/ either surface as [CVC˙] or [C˙VC] (depending on whether a sonorant follows, as argued in section 3.3.2). Examples of either type of output paradigm are given in (48). (173.) Consistent [C˙VC] forms derivable from / C˙VC˙/ b˙idyate ‘to split’ (passive) b˙edayati ‘to split’ (causative) b˙etsyati ‘to split’ (future) bib˙itsati ‘to split’ (desiderative)

[CVC˙] ~ [C˙VC] alternates derivable from / CVC˙/ bud˙yate 'to wake' (passive) ˙ bod ayati 'to wake' (causative) b˙otsyati 'to wake' (future) ˙ bub utsati 'to wake' (desiderative)

The proposal that /C˙VC˙/ inputs always surface as [C˙VC] and never as [CVC˙] is natural if one considers that, cross-linguistically, final segments tend to neutralize and root-initial segments preserve contrast. In terms of Correspondence Theory, root-initial segments are more faithful than other segments in the morpheme because they are subject to Positional Faithfulness constraints (Beckman 1997). Richness of the Base forces us to consider not only /C˙VC˙/ and /CVC˙/ input roots, but also /C˙VC/. Since no constraint bans such structures on the surface, these inputs always surface faithfully. Thus, /C˙VC˙/ and /C˙VC/ inputs neutralize to [C˙VC] outputs. An overview of possible inputs and their corresponding outputs is given below. It includes both murmured and aspirated inputs.

155

Existential Faithfulness

(174.) Sanskrit inputs containing [+SG] segments and corresponding outputs inputs a b

c

/C˙VC/ /CHVC/ /C˙VC˙/ /CHVCH/ /CHVC˙/ /C˙VCH/ /CVC˙/ /CVCH/

output [C˙VC] [CHVC] [C˙VC] [CHVC] [CHVC] [C˙VC] [CVC˙] ~ [C˙VC] [CVCH] ~ [CVCHV]

repairs --dissimilation dissimilation dissimilation dissimilation -- / throwback -- / vowel epenthesis

Let us now turn to the formal analyses of the input-output pairs in (a) and (b) of this table (pairs in (c) were analyzed in section 3.3.2). The tableau below shows that /C˙VC/ inputs are allowed to surface faithfully because the initial murmured segment precedes a sonorant, and because preservation of the feature specification [+SG] is of more importance than the need for segments to be unmarked [-SG] (∃-IDENT[+SG] >> *SG). (175.) /C˙VC/ inputs retain their initial [+SG] segment faithfully LARREL ∃-IDENT *SG ˙ [+SG]IO /b id-syati/ 1 bet-syati *! ˙ * 2 b et-syati

In inputs containing two murmured segments, dissimilation takes place. Tableau (51) deals with an input root that is followed by a vowel-initial suffix. Candidate 1 is identical to the input and incurs two violations of *SG. Loss of both murmur specifications rids candidate 2 of these violations. However, failure to preserve them results in fatal violations of ∃-higher-ranked IDENT[+SG]IO. The optimal candidate contains one murmured segment. This segment is the product of coalescence, and it

156

Struijke

preserves the [+SG] specifications of both its contributing input segments. Therefore faithfulness is achieved, and violations of *SG are minimized to one. The candidate also incurs a violation of INTEGRITY and surface faithfulness constraints such as IDENT[+SG]∑∑ because dissimilation involves fission in which the resulting segments are not identical. (176.) /C˙VC˙/ inputs dissimilate ∃-IDENT *SG /b˙1id˙2- eda/ [+SG] **! 1 b˙ed˙eda 2 bededa **! ˙ * 3 b 12ed2eda

INTEGRITY

*

IDENT [+SG]∑∑

*

Dissimilation in this example must result in the pair [b˙12d2] rather than [b1d˙12], even though these pairs incur the same violations of the constraints shown in this tableau. The sub-optimal dissimilation repair [b1d˙12] violates the positional faithfulness constraint on root-initial segments ∃-IDENT[+SG]RT-in. (177.) ∃-IDENT[+SG] RT: Let root-initial seg ∈ input be in the domain of ℜ, and seg is [+SG]; then there is some root-initial seg' ∈ output, such that segℜseg' and seg' is[+SG]. Some output root-initial segment corresponding to the root-initial input segment preserves the feature specification [+SG] of that input segment.

As is clear from the diagram below, in the ungrammatical output the [+SG] specification associated with the root-initial segment in the input is not preserved on the root-initial segment.

157

Existential Faithfulness

(178.) Dissimilation in a sub-optimal candidate given the input /b˙1id˙2- eda/ *

input:

output:

b1

[+SG]

[+SG]

b˙1

i

d˙2

e

d

i

d˙12

e

d

a

a

[+SG]

In the optimal candidate depicted below, the faithfulness requirement specific to root-initial segments is met. Thus, positional faithfulness constraints will often determine which output segment is marked in dissimilation, and which is unmarked.

(179.) Dissimilation in the optimal candidate given the input /b˙1id˙2- eda/75 [+SG]

input:

output:

b˙12

[+SG]

b˙1

i

d˙2

e

d

i

d2

e

d

a

a

[+SG]

When a /C˙vC˙/ input is word-final or precedes an obstruent-initial suffix, the markedness constraint *SG is not the only constraint accounting for the cooccurrence restriction. Also, root-initial faithfulness is not the only factor distinguishing [C˙VC] and 75

Voicing is not dissimilated because IDENT[+voice]∑∑ >> *VOICE.

158

Struijke

[CVC˙] candidates for a /C˙vC˙/ input. When roots surface in this particular environment, LARREL and ∃-IDENT[+murmur] conspire to force the root-final murmur specification to be associated with the root-initial segment.76 Thus, in this particular circumstance, feature movement and dissimilation are indistinguishable. I show in section 3.4 that in Cuzco Quechua a single markedness constraint drives both feature movement and dissimilation. I now sketch an analysis of cooccurrence restrictions 3 - 6 in table (49). The ban on roots containing both murmured and aspirated segments follows straightforwardly when ∃-IDENT[+SG] and *SG are at the top of the Sanskrit constraint hierarchy. The diagram below shows the potential input-output mapping /ph1ad˙2/ - [ph12ad2]. Note that the [+SG] specifications of both input consonants are preserved on the first output consonant, and the markedness constraint *SG is violated minimally. The licensing constraint on murmur ensures that the [+murmur] specification cannot be preserved on the voiceless consonant (LICENSE[+murmur] >> ∃-IDENT[+murmur]), as argued in tableau (33)). Since there is also a licensing constraint for the feature [+aspiration] (“if a segment is specified for [+asp], then it must be specified for [-son], [-cont], [-voice]”), we expect a potential input like /d˙1a ph2/ to map to [d˙12a p2].

76

Here feature value transfer is vacuous, because the root-initial segment is itself murmured.

159

Existential Faithfulness

(180.) Dissimilation in roots containing a murmured stop and an aspirated stop (only fatures present in the input are shown) [+asp] [+SG]

input:

output:

ph 12

[+murm] [+SG]

ph1

a

a

t2

[+SG] [+asp]

d˙2

[-murm]

Since [h] is specified for [+SG], it cannot cooccur with murmured and aspirated segments or another [h] in roots. Again, this follows from the ranking established above. When an input root with an initial /h/ and an aspirated or murmured segment is presented to the Sanskrit grammar, the latter segment undergoes distributing fission and preserves its [+SG] feature on the correspondent that coalesces with [h] (e.g. the hypothetical form /h1a:d˙2/ - [h12a:t2]). If a murmured or aspirated segment is root-initial, an /h/ occurring later in the root undergoes fission and its output correspondent preserving [+SG] coalesces with the initial segment. The other correspondent preserves the manner feature and presumably surfaces as the least marked fricative not associated with [+SG], namely [s] (thus, [d˙12vas2] ‘scatter’ might be derived from /d˙1vah2/). This concludes the case study of Sanskrit murmur throwback and Grassmann's Law. The ranking of the relevant constraints is summarized in appendix III at the end of this chapter. I have presented an account of the predictable location of murmured segments in roots such as ‘to wake’ (in some previous studies attributed to literal feature movement), and the cooccurrence restriction on segments sepecified as [+SG] (in previous studies taken to result from the OCP or to involve dissimilative alternations). I

160

Struijke

have shown that these patterns are related in the proposed system, in the sense that they both involve fission and coalescence. This similarity accounts for the fact that both predictable location and cooccurrence restrictions hold over segments in roots. Features move within morpheme boundaries, but not across. Also, words may contain multiple [+SG] segments, but morphemes may not. These facts are the result of undominated ∃-MORPHDIS, which prevents coalescence of segments belonging to different morphemes.

3.4 Case study: Cuzco Quechua

In Cuzco Quechua, roots may contain at most one laryngeally marked segment (MacEachern 1999; Parker and Weber 1996, following Rowe 1950; Orr and Longacre 1968; Carenko 1975; Mannheim 1991). There is also evidence that the laryngeal specification of a segment is not always associated with that segment underlyingly (McCarthy 1996; Parker 1997). Thus, like Sanskrit, Cuzco Quechua shows both dissimilation and feature movement. (Recall that these terms are used descriptively.) We have seen that dissimilation and feature movement are sometimes indistinguishable in Sanskrit. In Cuzco Quechua this is always the case. This supports the idea that the phenomena are intimately related: both are driven by a need to preserve underlying material and simultaneously improve markedness. They are alike formally in that they both involve segmental fission and coalescence. When a laryngeally marked segment occurs in a Cuzco Quechua root, its position is entirely predictable. Working within Autosegmental Theory, Parker and Weber (1996) assume that predictable information is absent from underlying forms (underspecification), including predictable positional information. They therefore argue that a Cuzco Quechua laryngeal feature is not underlyingly associated with the stop it

161

Existential Faithfulness

surfaces on. In fact, they claim that laryngeal features are not associated to any segment, but rather ‘float’ in the underlying form. Their surface positions, being fully predictable, are determined by a rule of autosegmental association. Although previous theories could impose language-particular restrictions on what could be specified in a lexical form, in Optimality Theory such restrictions cannot be made. Again, OT holds the strong claim that all linguistic generalizations can be captured by conditions on outputs. It is a theory of output constraints that does not have the means to restrict inputs in any way. Hence, the set of possible inputs is universal (Richness of the Base), and any input that is fed into a language-particular grammar must give rise to an output that is well-formed in that language (Prince and Smolensky 1993). The OT grammar of Cuzco Quechua must therefore deal with all kinds of inputs, including those containing more than one laryngealized segment and those in which a segment other than an initial stop is laryngealized. I will present evidence that the feature specifications for glottalization and aspiration are always preserved. It is the grammar’s job to ensure that they are realized on the first stop. Section 3.4.1 lays out the basis of the analysis, and makes the important point that glottalization and aspiration feature specifications are preserved in the output root at all times. Section 3.4.2 focuses on the predictable location of glottalized and aspirated segments. The groundwork of the movement analysis within Optimality Theory was laid in McCarthy (1996) and Parker (1997). The analysis presented here follows these proposals in its essentials, but crucially differs from them in assuming that featural correspondence is mediated by segments, rather than that features enter into correspondence relations independently.77 In section 3.4.3 the focus shifts to dissimilation, and the link between movement and dissimilation is drawn.

77

A comparison of feature-based and segment-based analyses of movement can be found in chapter 4.

162

Struijke

3.4.1 Feature value preservation

In Cuzco Quechua, glottalization and aspiration are contrastive in stops and affricates, but not in fricatives and sonorants. I will refer to the relevant obstruents as ‘stops.’ (181.) Glottalization and aspiration contrast in stops only tanta ‘collection, combination’ t’anta ‘bread’ tHanta ‘old, used up, worn out’ manta *m’anta *mHanta (182.) Cuzco Quechua consonant inventory78 voiceless stops aspirated stops glottalized stops

p p’ s m

t pH t’

c& tH c&’

s& n l r

n) ¥

k c&H k’

q kH q’

qH /

h

w

y

The consonant inventory does not include voiced obstruents or murmured segments. Voicing and murmur must be prohibited by undominated markedness constraints *VOICE/OBSTR and *MURMUR. Following McCarthy (1996) and Parker (1997), I posit licensing constraints to account for the fact that only stops, affricates and [h] can be specified as [+SG], and stops, affricates and glottal stop can be specified as [+CG].

78

Glottal stop is epenthetic, and found only word-initially. The language allows voiced stops b, d, and g in Spanish loans.

163

Existential Faithfulness

(183.) LICENSE[+CG]: If a segment is specified as [+CG], then it is specified as [- sonorant, -continuant] (184.) LICENSE[+SG]: If a segment is specified as [+SG], then it is specified as [- sonorant, -continuant], or [h]

In order to prevent any underlying glottalized or aspirated sonorants from surfacing faithfully, these licensing constraints must dominate the ∃-IDENT constraints on glottalization and aspiration. This is shown in tableau (61). The input considered in this tableau contains no stops, but has two sonorants, one of which is glottalized. The faithful candidate is sub-optimal, because LICENSE[+SG] bans glottalized sonorants. The feature specification [+CG] is lost due to the lower rank of ∃-IDENT[+CG]. (185.) ∃-IDENT[+CG] Let seg ∈ input be in the domain of ℜ, and seg is [+CG] then there is some seg' ∈ output, such that segℜseg' and seg' is [+CG] Some segment corresponding to an input segment preserves the feature specification [+CG] of the input segment.

(186.) Glottalized sonorants are disallowed /m’aru/ LICENSE[+CG] 1 m’aru *! 2 maru

∃-IDENT[+CG] *

For reasons of space, I restrict the analysis to the glottalization facts. However, for every proposed constraint specific to [+ConstrictedGlottis] there is a constraint specific to [+SpreadGlottis] for which the same ranking arguments hold. When a stop specified as [+ConstrictedGlottis] or [+SpreadGlottis] occurs in a root, its position is entirely predictable. As the data below show, it is always the first stop in the root.

164

Struijke

(187.) Only left-most stops can be glottalized/aspirated a. p’atay ‘to bite’ pHatay ‘to explode, blow up’ c&’ac& u ‘treacherous, tricky’ c&Hac&u ‘ragged, tattered’ k’anka ‘rooster’ kHanka ‘slimy, clammy’ b. n) u k’u ‘withered’ hap’iy ‘to light (a fire), grab, hold supHu ‘body hair’ wirp’a ‘lip’ ranqHa ‘forgetful’ warak’a ‘sling made of wool’ hamawt’a ‘learned, wise’ wa¥wak’u ‘armpit’ c. rank’ukuy ‘to get twisted up’ akHakaw ‘How hot it is’ muspHapakuy ‘to be delirious’ lap’otiyay ‘to walk in the mud’ wic&’ic&ic&iy ‘onomatopoeic sound made by pigs’ ayk’uc&iy ‘to tire, wear out (transitive)’ wisk’ac&a ‘vizcacha rodent’ melq’ote ‘oesophagus’

Thus, a glottalized stop can surface anywhere in the root, as long as it is not preceded by another stop.79 Since the location of the laryngeally marked segment is variable, preservation of the marked specification cannot be due to enhanced faithfulness in a prosodically prominent position (Positional Faithfulness, Beckman 1997). Instead, McCarthy (1996) and Parker (1997) assume that glottal features are left-tropic because of alignment constraints demanding that these features lie at the left edge of the prosodic word.80 I assume ALIGN[+CG] and ALIGN[+SG], which are violated when a glottalized or aspirated segment does not lie at the left edge of the word. 79

Modulo restrictions on stops preceding obstruents (i.e. stops in coda). In the following words the first stop is in coda, hence the second rather than the first stop is glottalized: rakt’a, moqc&’ikuy, lapt’ay. I assume glottalized codas are ruled out by the ranking LARREL >> ALIGN[+CG]. Rowe (1950) and MacEachern (1999) argue that this issue is moot. They claim that coda consonants transribed as stops are in fact fricatives. 80 Affixes never contain glottalized or aspirated stops (this is derived by the ranking ∃-IDENT[+CG]RT >> *CG >> ∃-IDENT[+CG]IO: Beckman 1997). Yet the alignment constraint can refer to the left edge of the prosodic word, because the language does not have prefixes.

165

Existential Faithfulness

(188.) ALIGN[+CG]: A segment specified as [+CG] lies at the left edge of the prosodic word. Align([+CG],L,PrWd,L) (Generalized Alignment: McCarthy and Prince 1993b) (189.) ALIGN[+SG]: A segment specified as [+SG] lies at the left edge of the prosodic word. Align([+SG],L,PrWd,L) (Generalized Alignment: McCarthy and Prince 1993b)

The words in (62b and c) contain aspirated and glottalized segments that do not lie at the left edge of the word and hence violate these constraints. The constraints would be satisfied if the initial segment were glottalized or aspirated, but in each case this segment is not a stop and cannot license the glottalization or aspiration feature. The alignment constraints must thus be dominated by the licensing constraints. The ranking argument is illustrated in the tableau below. The optimal candidate violates ALIGN[+CG] twice (the constraint is evaluated gradiently such that each segment intervening between the glottalized stop and the left edge of the word causes a violation mark). (190.) Left-alignment is blocked if it results in glottalized sonorants /n)uk’u/ License[+CG] ALIGN[+CG] 1 n)’uku *! 2 n)uk’u **

It is important to note that a glottal segment does not neutralize to a plain stop when it is not the first segment in the word. That is, the marked feature specification is not lost in order to avoid an alignment violation. It is crucial to preserve the glottal feature value, even if it results in misalignment. This shows that the constraint regulating featural faithfulness dominates ALIGN[+CG].

166

Struijke

(191.) Glottal feature specification is preserved at the expense of alignment /n)uk’u/ ∃-IDENT[+CG] ALIGN[+CG] 1 n)uku *! 2 n)uk’u **

The constraint ranking established so far is given below: (192.) Intermediate ranking fir Cuzco Quechua *VOICE/OBSTR

, *MURMUR

∃-IDENT[+CG]

LICENSE[+CG] >> LICENSE[+SG]

∃-IDENT[+SG]

ALIGN[+CG] >> ALIGN[+SG]

With this ranking, an underlying feature specification [+CG] must be preserved on a stop. It cannot be preserved on a sonorant (LICENSE[+CG] >> ∃-IDENT[+CG]), even if that would result in left-alignment of the feature specification (LICENSE[+CG] >> ALIGN[+CG]). I show in the next section that this ranking promotes feature movement to the leftmost stop in the word. As always, feature movement ensures simultaneous satisfaction of (or improved performance on) conflicting faithfulness and markedness constraints. Here, it is driven by the need to preserve a glottal feature specification and the need for that specification to be as close to the beginning of the word as possible (without violating licensing constraints). Following McCarthy (1996) and Parker (1997), I show that feature movement must be assumed, regardless of the fact that there are no overt alternations. I then go on to claim that the 'floating features'81 and OCP patterns described by Parker and Weber (1996) are in fact formally indistinct in Cuzco Quechua.

81

The term 'floating feature' is used here to refer to features that are not associated with a segment in the underlying representation. It does not refer to featural morphemes, for which the term has been used also (see e.g. Goldsmith 1990).

167

Existential Faithfulness

3.4.2 Floating features and the OCP

In order to illustrate feature movement / dissimilation in Cuzco Quechua, I use the example [melq’ote] ‘esophagus.’ Several inputs in the rich base can map to this output, and I discuss a few of them below. The input /melq’ote/, where the first stop is glottalized, surfaces faithfully. Glottalization moves in the case of any input in which a segment other than the first stop is glottalized, e.g. /m'elqote/ and /melqot'e/. Crucially, the feature specification is not lost. When two segments are underlyingly glottalized, e.g. /m'elqot’e/, both glottal feature specifications are preserved on the first stop. I now discuss these inputs in turn. Nothing new needs be said for the input /melq’ote/, where the first stop is glottalized. The established ranking ensures that the first stop is faithfully glottalized in the output, as shown in the tableau below. This tableau reiterates the generalization that a glottal feature specification must be preserved on a stop.

(193.) [+CG] is associated with the first stop in the word / melq’ote / LICENSE[+CG] ∃-IDENT[+CG] ALIGN[+CG] 1 m’elqote *! 2 melqote *! 3 melq’ote ***

Consider now inputs in which a segment other than the first stop is glottalized, such as /melqot’e/. We know that this input cannot surface faithfully: there are no Quechua words in which the second stop is glottalized (unless the first stop is in coda position). Two possible repairs come to mind: non-preservation of the feature specification, and feature movement. Candidates with these repairs are shown in tableau (69), which contains the constraint ranking established earlier. Candidate 1 involves deletion of the glottal feature. It vacuously satisfies the alignment constraint, but fatally

168

Struijke

violates higher-ranked ∃-IDENT[+CG]. Candidate 2 involves movement of the [+CG] specification to the first stop, attained through segmental fission and coalescence. Since this stop is not word-initial, the candidate incurs violations of the alignment constraint. However, it satisfies higher-ranked ∃-IDENT[CG] by preserving the glottal feature specification. Faithful candidate 3 preserves the feature value on the second stop, and incurs more violations of the alignment constraint. Thus, candidate 2 is the winning candidate. It embodies the optimal compromise between feature value preservation and alignment of [+CG] to the beginning of the word. (194.) Feature value transfer to the first stop in the word /melq1ot’2e/ LICENSE[+CG] ∃-IDENT[+CG] 1 melqote *! 2 melq’12ot2e 3 melqot’e

ALIGN[+CG] *** *****!

Because the optimal form involves feature movement through fission and coalescence, it incurs violations of constraints banning fission in general, fission that results in non-identical output correspondents, and constraints banning coalescence of segments with different feature specifications. These are INTEGRITY, ∃-IDENT[-CG], and ∃-IDENT[+CG]∑∑.82 Obviously, these constraints must be low-ranking. Had feature movement been prohibited by highranking of any of these constraints, the faithful candidate would have been optimal. That is, the glottal feature would have to surface in its original position, on the second stop in the word. This is shown in tableau (70). Since such a form is not attested in Cuzco Quechua, feature movement must be permitted.

82

At a minimum, IDENT[+CG]∑∑ is violated in glottalization movement. When the segments involved differ along other dimensions, additional surface faithfulness constraints are violated (in this example one pertaining to place specifications).

169

Existential Faithfulness

(195.) Ban on feature value transfer does not account for Quechua generalizations /melq1ot’2e/ INTEGRITY / LICENSE ∃-IDENT ALIGN ∃-IDENT[-CG] / [+CG] [+CG] [+CG] ∃-IDENT[+CG]∑ ∑

1 melqote 2  melq’12ot2e 3  melqot’e

*! *! /*!/*!

*** *****

Thus, even though Cuzco Quechua does not show any overt alternations involving feature movement, we can conclude that a glottal feature specification can surface in a location that is different from its location in the input (McCarthy 1996; Parker 1997). The cooccurrence ban on glottalized segments is now an emergent property of the established constraint ranking. All glottal specifications must be associated with the first stop in the surface form. Below, I depict an input-output mapping in which two feature specifications move. Any segment specified as [+CG] in the input other than the first stop undergoes fission, and one of its output correspondents coalesces with the first stop in the word.

(196.) Cooccurrence restriction and feature movement are formally identical input:

m'1

e

l

q2

o

t’ 3

output: m1

e

l

q'123

o

t3

e

e

Given highranking ∃-IDENT[+CG], both glottal feature specifications in the above form must be realized in the output, because it contains a segment that licenses them. Tableau

170

Struijke

(72) considers three candidates for this particular input. Candidate 1 deletes both features, and is therefore suboptimal. In candidate 2, the features surface in their underlying positions. This results in a form containing a glottalized sonorant, and the licensing constraint is fatally violated. In addition, the alignment constraint is violated five times because the second glottal feature is five segments removed from the beginning of the word. Candidate 3 is most harmonic because it preserves both glottal features on the first stop. It satisfies the licensing constraint and ∃-IDENT [+CG], while violating the alignment constraint only three times. (197.) Dissimilatory feature movement / m'1elq3ot’2e / 1 2

melqote m'elqot'e

LICENSE [+CG]

∃-IDENT ALIGN [+CG] [+CG] **!

INTEG

∃-IDENT IDENT [-CG] [+CG]∑∑

*! ***** ***

3 m1elq’123ot3e

**

*

* *

This concludes the section that unifies feature movement and dissimilation in Cuzco Quechua. I have shown that feature movement can have a dissimilatory effect. While previous analyses of the Cuzco Quechua pattern have presented a dissimilation/OCP account or a reassociation/floating feature account without identifying a common cause underlying the cooccurrence patterns and predictable location of laryngeally marked segments, I have established that these generalizations are in fact two sides of the same coin. They are emergent generalizations driven by the need to preserve feature specifications while simultaneously improving markedness with respect to alignment. The following section gives an analysis of the remaining cooccurrence restrictions.

171

Existential Faithfulness

3.4.3 Further cooccurrence restrictions

A complete overview of Cuzco Quechua cooccurrence restrictions is given below (based on Hornberger and Hornberger 1977; MacEachern 1999; Parker and Weber 1996). So far I have analyzed the cooccurrence restriction on glottalization, and I have mentioned that the same analysis accounts for the identical restriction on aspiration (generalizations 1 and 2 in the table). (198.) Cooccurrence restrictions in Cuzco Quechua 1. 2.

Only one glottalized segment per morpheme Only one aspirated segment per morpheme

3. 4.

[h] and aspirated stops do not co-occur [/] and glottalized stops do not co-occur

5. 6.

Only one [h] per morpheme Only one [/] per morpheme

7.

Aspirated stops and glottalized stops do not co-occur

but: - Aspirated stops and [/] do co-occur - Glottalized stops and [h] do co-occur - [h] and [/] co-occur

The analysis just presented to account for (1) and (2) also explains generalizations (3) and (4). I assume that [/] is specified for [+CG] and [h] is specified for [+SG]. I will illustrate the explanation with inputs containing /h/ and an aspirated stop: /hatHun/ or /pHahi/. These inputs cannot surface faithfully: dissimilation must take place. Given /hatHun/, the aspirated stop undergoes distributing fission, and its correspondent specified for [+SG] coalesces with [h], resulting in the surface form

172

Struijke

[hatun] 'big, large, long'. Both [+SG] specifications of the input are preserved on the initial segment. With the reverse order of aspirated stop and /h/, as in /pHahi/, we expect /h/ to undergo fission and coalesce with the aspirated stop. Presumably, the intervocalic consonant appears as an unmarked fricative, preserving the manner feature of input /h/: [pHasi] 'cooked in steam.' Generalizations (5) and (6) state that forms do not contain two instances of [h] or two instances of [/]. Again, this is expected under the feature movement analysis: all instances of [+CG] and [+SG] must be as far to the left of the root as possible. We expect an input such as /h ah ut'iy/ to surface as [h asut'iy] ('whip, punish with a whip') with an unmarked fricative, and an input such as //a/aw/ to surface as [/ataw] ('happiness, good fortune, good luck'), with an unmarked stop. The relevant correspondence relations in the first form are shown below.

(199.) Dissimilation of [h] input:

h

a

h

u

t'

i

y

output:

h

a

s

u

t'

i

y

This form shows that [h] can co-occur with a glottalized stop. Cuzco Quechua also allows words containing [/] and an aspirated stop ([/apHay] 'carry a child'), and words containing an [h] and a [/] ([/ahoya] 'wild duck'). Again, these patterns are expected under the proposed analysis. Recall that highranking ∃-IDENT[+CG] and ∃-IDENT[+SG] force preservation of glottalization and aspiration whenever a form

173

Existential Faithfulness

contains a segment that licenses them. Under the assumption that segments cannot be both glottalized and aspirated, these features must be preserved on separate segments.83

3.5 Conclusion

I have presented the segmental analyses of feature movement and dissimilation predicted by existentially defined IDENT[F] constraints, arguing that both are driven by the need to satisfy general markedness constraints and the need to preserve underlying material. Because the two phonological patterns are driven by these same requirements, this proposal gives a straightforward account of dissimilatory feature movement seen in Sanskrit and Cuzco Quechua. The proposal imposes important restrictions on feature movement and dissimilation. First, only marked features move or dissimilate. There is a rising consensus that this is in fact an accurate description of these patterns (e.g. Alderete 1997; Alderete and Pater (in prep.); Itô and Mester 1996, MacEachern 1999; Suzuki 1998, who argue that dissimilation is driven by self-conjoined markedness constraints). Second, movement and dissimilation can only take place when the relevant elements can be involved in coalescence. For instance, featural attributes can move or dissimilate because they can be indirectly involved in coalescence through segments. We do not expect dissimilation or movement of, for instance, syllables or onsets, because they do not coalesce.84 We have seen the important role of surface faithfulness constraints. They can prevent feature movement and dissimilation from taking place when the segments 83

In light of these data, it is puzzling that no word of Cuzco Quechua contains both a glottalized and an aspirated stop (restriction (7) of table 73). 84 As McCarthy (1996) points out, the self-conjunction account of dissimilation wrongly predicts that any marked structure could be involved in dissimilation, such as onsetless syllables or syllables with codas via the conjunctions NOCODA&NOCODA and ONSET&ONSET.

174

Struijke

involved in these patterns are not alike with respect to some feature specification(s). This gives a straightforward account of the generalization that feature movement and dissimilation are most likely to occur when the segments involved are alike (the similarity effect). In the appendix to this chapter I briefly address the issue of locality in feature movement and dissimilation and other phonological patterns involving segmental fission, and sketch a possible account for the proximity effect.

175

Existential Faithfulness

Appendix II - The proximity effect

Feature movement and dissimilation are usually restricted to a certain domain. Often they can only take place within a morpheme. I have argued that this is due to the ban on coalescence in which the participating segments belong to different morphemes (∃-MORPHDIS). In general, we can conclude that these phenomena are more likely to occur locally than non-locally. I suggest that this proximity effect is due to gradient evaluation of INTEGRITY. The constraint is satisfied when no fission takes place; violated once when fissioned output segments are adjacent; and multiple times when the output correspondents of an input segment are separated by other segments (with an additional violation for each additional intervening segment).85 The constraint then accounts for the fact that segments are adjacent in diphthongization, feature specifications are displaced to the nearest possible segment, and the closer two (near-)identical segments are, the more likely they undergo dissimilation. The locality effect is also evident in copy-epenthesis, which is commonly argued to involve fission (Lamontagne and Rice 1995, McCarthy 1996; cf. Kitto and de Lacy 1999).86 Languages showing copy-epenthesis duplicate a segment in a word to create a 85

Alternatively, we may assume a universal hierarchy of constraints banning specific numbers of segments intervening the output correspondents (where constraints banning higher numbers of intervening segments universally dominate those banning lower numbers of intervening segments). This alternative would allow us to implement the idea that there is a window of opportunity in which movement or dissimilation can take place (e.g. 3 or more intervening segments could form a barrier to movement or dissimilation). 86 Kitto and de Lacy argue that copy-epenthesis is due to a correspondence relation between output segments in the same string. This allows an account of over and underapplication in copy-epenthesis, similar to these modes of application in reduplication. These authors argue that copy-epenthesis should therefore not be seen as fission. However, as I have argued, fission and surface correspondence are not mutually exclusive. In fact, treating both reduplication and copy-epenthesis as fission in which the output correspondents are related by a surface correspondence relation allows a formal characterization of the similarities between the two phenomena. Even though one employs copying for morphological reasons, and the other for phonological reasons, both are limited by the same or similar constraints. (Recall, though, that only phonologically driven fission induces violations of INTEGRITY. In reduplication, Anchoring constraints on the base-reduplicant relation ensure the base and the reduplicant are adjacent.)

176

Struijke

new output segment, instead of epenthesizing a segment. Apparently, the constraint banning epenthesis (M-SEG) dominates the constraint banning fission (INTEGRITY) in these languages. Hawaiian shows copy-epenthesis in words borrowed from English in order to avoid consonant clusters. (200.) Hawaiian copy-epenthesis (Kitto 1997) (‘epenthesized’ segment underlined) pe1le1kania ‘Britannia’ pa1la1ni ‘brandy’ The correspondence relations involved in the fission analysis of Hawaiian copy-epenthesis are illustrated below:

(201.)

b

p

e

r

I

t

a

n

i

a

l

e

k

a

n

i

a

This Hawaiian form shows that the copied vowel is the one nearest to the epenthesis site (Kitto and de Lacy, 1999). In Hawaiian, vowels of any quality can fulfill the role of the epenthesized segment, but it is the vowel closest to the ‘epenthesis’ site that undergoes fission, not the least marked vowel in the word. The candidates *[pa2leka2nia] and *[pi3lekani3a] are suboptimal because they incur more violations of INTEGRITY than the optimal form [pe1le1kania]. The fact that [a] and [i] are less marked than [e] is irrelevant here. Apparently better performance with respect to INTEGIRTY is of more importance than satisfying markedness constraints that disfavour [e].

177

Existential Faithfulness

4 ∃ -IDENT[F] and MAX[F] compared

With the advent of faithfulness and Correspondence Theory (Prince and Smolensky 1993; McCarthy and Prince 1995), the status of features in phonological theory must be recosidered. Whether features are properties of segments or independent entities has important consequences for the way we view faithfulness. Throughout this dissertation I have assumed that features are attributes of segments87 that cannot be preserved independently of their segmental sponsors. Featural faithfulness is evaluated through segmental correspondence, by IDENT[F] constraints. Some researchers have argued that features are independent entities that can be preserved outside of their sponsoring segments. Under this view, features enter into correspondence relations separately from segments. Featural faithfulness is then evaluated by the constraints MAX[Feature] and DEP[Feature] (Lamontagne and Rice 1995; Lombardi 1995, 1998; Causley 1997a,b; Parker 1997; Walker 1997). I refer to the first theory as ‘Featural Attribute Theory,’ and the second as ‘Featural Independence Theory.’ I compare them in this chapter, and conclude that both can account for phenomena that have been thought to rely crucially on the idea that features are autosegments, namely distributing diphthongization, feature movement, coalescence, feature stability, and dissimilation. Their comparable success is due to two factors. First, the featural faithfulness constraints in these theories share important characteristics: existentialism, unidirectionality, and feature value specificity. Second, both theories assume segmental correspondence and therefore allow the same segmentally based analyses of the aforementioned phenomena through fission or fusion of segments, or a combination thereof. Because the Featural Independence Theory assumes feature correspondence in addition to segment correspondence, it allows a featurally based explanation for each of 87

The terms 'attributes' and 'entities' as referring to features are due to McCarthy 1996.

178

Struijke

the phenomena explored, in addition to the segmentally based analyses. For instance, feature movement can be analyzed as feature reassociation, and coalescence as feature stability (i.e. deletion of a segment and reassociation of one or more of its features on another segment - also known as ‘phantom limb’). I discuss the differences between the two theories in section 4.1, the similarities of the constraints ∃-IDENT[F] and MAX[F] in section 4.2, and I compare the different analyses of the relevant phenomena in section 4.3. As mentioned, there are two different ways to analyze these phenomena in the Featural Independence Theory, but only one in the Featural Attribute Theory. The latter is thus more restrictive. Whether this restrictiveness is to be preferred or dispreferred is in large part an empirical question. Some implications are discussed in the concluding section of this chapter. I would like to stress that the discussion is restricted to subsegmental features; no claims are made about suprasegmental features such as tones, moras, and accents.

4.1 Correspondence Theory and the status of features

Early work in generative phonology, following structuralism, assumed that forms are sequences of segments, like beads on a string (e.g. Jakobson 1939; Jakobson, Fant and Halle 1952; Chomsky and Halle 1968). Segments were considered to be the smallest phonological units. Larger units were assumed to exist (morphemes and perhaps prosodic units like syllables), but smaller units were not. Features were seen as properties of segments, and were associated with them in an unordered fashion. Every phonological rule could be formulated such that it referred to segments and their featural attributes. Because features as independent elements were unnecessary, they were not assumed: it was for reasons of parsimony that segments were considered the smallest units in phonology.

179

Existential Faithfulness

One of the main claims of Autosegmental Theory (Goldsmith 1976, 1990; following Leben 1971 and Williams 1976) is that features are independent entities.88 Several phonological patterns were believed to provide evidence for this claim. These include feature stability and feature movement. Whether the assumptions of Autosegmental Theory about the independence of features should be carried over into Correspondence Theory is subject to debate (Causley 1997a,b; Lamontagne and Rice 1995; Lombardi 1995; 1998; McCarthy 1995, 1996; McCarthy and Prince 1995, 1999; Pater 1999; Walker 1997; Zoll 1996, among others). At the outset of Correspondence Theory, it was tentatively assumed that features are properties of segments, and that only segments enter into correspondence relations, as depicted below (McCarthy and Prince 1995). (202.) Correspondence relations in the Featural Attribute Theory (lines indicate segmental correspondence) ([+cor][+cont][-vc]) ([+hi][-bk][+rd]) ([+lab][-cont][+voice])

s

input:

output: s

ü

b

ü

b

([+cor][+cont][-vc]) ([+hi][-bk][+rd]) ([+lab][-cont][+voice])

In this theory, faithfulness to both segments and feature specifications is evaluated on the basis of segmental correspondence. The definition of the featural faithfulness constraint assumed in this dissertation is repeated below:

88

See Hocket 1947, Bloch 1948, and Firth 1948 for precursors.

180

Struijke

(203.) Featural faithfulness constraint in the Featural Attribute Theory ∃-IDENT[F] IO: Let seg ∈ input be in the domain of ℜ, and seg is [αF]; then there is some seg' ∈ output, such that segℜseg' and seg' is [αF]. Some output segment corresponding to an input segment preserves the feature specification [αF] of that input segment.

Those who argue that features are independent entities assume that correspondence relations are established both between features and between segments. In addition, features need to be associated to segments. An illustration is given below.

(204.) Correspondence relations in the Featural Independence Theory (assuming privativity) (solid lines indicate correspondence relations, broken lines indicate the association of features to segments). input:

[cor]

output: [cor]

[cont]

[hi]

[rd]

[lab] [voice]

s

ü

b

s

ü

b

[cont]

[hi]

[rd]

[lab] [voice]

In the Featural Independence Theory, segmental faithfulness is evaluated through segmental correspondence, and featural faithfulness is evaluated through featural correspondence. The featural faithfulness constraints are typically defined on analogy with MAX(seg) and DEP(seg) to demand presence of an input feature in the output (MAX[F]IO) or presence of an output feature in the input (DEP[F]IO).

181

Existential Faithfulness

(205.) Featural faithfulness in the Featural Independence Theory (to be revised) MAX[F]IO: Every feature of the input has a correspondent in the output. DEP[F]IO: Every feature of the output has a correspondent in the input.

Research in the Featural Independence Theory has implicitly assumed that these constraints not only demand the presence of a feature in a string, but also require identity of features in correspondence (as Bakovic (1999) and Keer (1999) point out).They are violated when corresponding features are different. Thus, in the mapping /l/ - [n] below, MAX[lat] is violated.

(206.) Correspondence between non-identical features (solid lines indicate correspondence relations, broken lines indicate the association of features to segments) [lat] input:

output:

l

n [nas]

No research adopting MAX[F] and DEP[F] has assumed that the above structure satisfies these constraints. It is thus clear that the featural faithfulness constraints are taken to

182

Struijke

demand both the presence of a feature in a string, and identity of corresponding features. I will therefore assume the following definitions89 :

(207.) MAX[F] IO: Let x ∈ input, and x is [F], then there is some x' ∈ output such that xℜx', and x' is [F] Every feature of the input has an identical correspondent in the output. (208.) DEP[F] IO: Let x ∈ output, and x is [F], then there is some x' ∈ input such that xℜx', and x' is [F] Every feature of the output has an identical correspondent in the input. In addition to segmental and featural faithfulness constraints, the Featural Independence Theory presupposes faithfulness constraints on the association of features to segments, which are violated when underlying associations are altered. They are NO-DELINK([F], seg) and NO-SPREAD([F], seg) (McCarthy 1997a). The first prohibits delinking of a feature from its input segment. The second prohibits linking of a feature to a segment with which it is not associated underlyingly. They are sometimes conflated into the constraint ‘NOFLOP/STAY’ (McCarthy 1996; Parker 1997; Walker 1997; Alderete 1999).

(209.) NO-DELINK([F], seg) Let seg ∈ input, and seg' ∈ output, and segℜseg'. Let [F] ∈ input, and [F'] ∈ output, and [F]ℜ[F']. If [F] is associated with seg, then [F'] is associated with seg'. Feature-segment associations in the input must be reflected by the corresponding elements in the output. (McCarthy 1997a) 89

Under the Featural Independence Theory, MAX(seg) is essentially a constraint mandating the preservation of timing slots. Like MAX[F], it requires both preservation and idendity (it is not satisfied, for example, by a candidate in which an input segment is in correspondence with an output feature).

183

Existential Faithfulness

(210.) NO-SPREAD([F], seg) Let seg ∈ input, and seg' ∈ output, and segℜseg'. Let [F] ∈ input, and [F'] ∈ output, and [F]ℜ[F']. If [F'] is associated with seg', then [F] is associated with seg. Feature-segment associations in the output must be reflected by the corresponding elements in the input. (McCarthy 1997a) (211.) STAY[F] Let the feature Fi be associated with the segment Si in the input and let Fo be associated with the segment So in the output. If FiℜFo, then SiℜSo. (McCarthy 1996)

It is clear from this discussion that the Featural Independence Theory assumes a number of correspondence relations and constraints thereon that are not assumed in the Featural Attribute Theory utilized throughout this dissertation. An important question is whether both segmental and featural correspondence relations are necessary in the former theory. It could be argued that segments are epiphenomenal, and do not exist as independent entities. That is, segments could be taken as the sum of their features. Correspondence could then be restricted to features, which are bundled into segments by association conventions (cf. ‘path conditions’ in Archangeli and Pulleyblank 1994). This would eliminate the segmental correspondence relations in diagram (3) above. However, previous work in Optimality Theory, including research in the Featural Independence Theory, has taken segments to be primitives. To my knowledge, all work within Correspondence Theory presupposes that segments can enter into correspondence relations and can be referred to by markedness constraints such as NOCODA, ONSET, and *COMPLEX. I show in section 4.3 that the extra correspondence relation (i.e. featural correspondence) allows for an extra analysis of distributing fission, coalescence, feature movement, and dissimilation.

184

Struijke

4.2 Similarities between ∃ -IDENT[F] and MAX[F]

The main thesis of this dissertation is that faithfulness constraints evaluating the correspondence relation between input and output elements are existentially quantified. Working withing the Featural Attribute Theory, this claim has particular impact on the function of the featural faithfulness constraint IDENT[F]IO. Not only is the proposed ∃-IDENT[F] constraint existentially defined, I have argued that it is unidirectional, requiring presence of a feature specification associated with an input segment to be preserved on an output segment, not vice versa (cf. Keer 1999; Pater 1999). Also, I have argued that these constraints demand preservation of feature values. That is, they refer to positive and negative values separately, so that, for instance, denasalization and nasalization incur violations of separate constraints (namely ∃-IDENT[+nasal]IO and ∃-IDENT[-nasal]IO respectively). The constraints MAX[F]IO and DEP[F]IO of the Featural Independence Theory are also existentially defined: they require that a feature present in one string has some (identical) correspondent in the other string. They are also unidirectional; MAX[F]IO going from the input to the output, and DEP[F]IO going from the output to the input. Because there are constraints in both directions, the Featural Independence Theory can implement privativity of features, so that a feature is either present in or absent from a string. For instance, an oral segment is not associated with a nasal or a [-nasal] feature, but a nasal segment is. This privativity in the Featural Independence Theory amounts to feature value specificity seen in the Featural Attribute Theory. Again, feature value specificity and unidirectionality allow us to distinguish such alternations as nasalization from denasalization.

185

Existential Faithfulness

Because the featural faithfulness constraints ∃-IDENT[F]IO, MAX[F]IO, and DEP[F]IO share these characteristics, and because the two theories of features assume (at least) segmental correspondence, they both allow segmental analyses of phonological alternations. In particular, their existential and unidirectional nature explains their success in accounding for phenomena that can be argued to involve segmental fission, while the unidirectional and feature value specific aspects of the constraints explain their success in accounting for phenomena that can be argued to involve coalescence (see section 3.1.1.2 and Pater 1999 for extensive discussion, and section 4.3.2 for a more consice discussion). All three characteristics are crucial in feature movement and dissimilation, because they combine fission and coalescence under the segmental analysis.

4.3 Phenomena

Although this section compares the behavior of ∃-IDENT[F], MAX[F], and DEP[F] in different phonological patterns, reference will also be made to classic IDENT. This is because doing so emphasizes the similarities between the constraints under discussion. Repetition of some arguments given in chapter 3 will therefore be unavoidable. Following most research in the MAX[F] framework, I assume that features are privative in the Featural Independence Theory. However, I assume full specification in the Featural Attribute Theory, as I have done throughout this dissertation.

4.3.1 Distributing diphthongization

I argued in chapter 1 that classic IDENT constraints (McCarthy and Prince 1995) fail to account for featural faithfulness in distributing diphthongization. This is because the constraints are universally quantified and hence demand faithfulness of all output

186

Struijke

segments that correspond to a given input segment. Instead, the existentially defined faithfulness constraints ∃-IDENT[F] and MAX[F] are satisfied when one output segment is featurally like the input segment, thereby allowing a straightforward account of distributing diphthongization. Icelandic pre-aspiration can briefly illustrate this point. I will show that the Featural Attribute Theory must analyze this pattern as segmental fission, while the Featural Independence Theory can analyze it either as fission or a combination of segmental epenthesis and feature reassociation. Aspirated stops in Icelandic cannot surface in certain environments. I assume that this is due to some markedness constraint or constraint constellation which I will not identify here (but see Keer 1999 and Morén 1999 for recent accounts set in Optimality Theory). One of the repairs exhibited by Icelandic is pre-aspiration. Some examples are given below: (212.) Icelandic pre-aspiration /haph/ hahp Yhpi /Yphi/ vahtna /vathna/

‘luck’ ‘upstairs’ ‘water’

Following Thráinsson (1978), Clements (1985), and Hayes (1990), Morén (1999) assumes that pre-aspiration involves segmental fission, as depicted below. (213.) Fission in Icelandic preaspiration (only relevant features shown) [+asp][-cont][lab]

input:

output:

pH

h [+cont][+asp]

p [-cont][lab]

187

Existential Faithfulness

As both Keer (1999) and Morén (1999) point out, classic IDENT constraints do not allow a fission account of Icelandic diphthongization, because they are universally quantified. They create the paradox discussed in chapter 1: persistance of input material in the output must be due to faithfulness, yet faithfulness is not satisfied by persistance of input features on only one out of two output correspondents. (214.) Classic IDENT[F]: Correspondent segments have identical values for the feature F. If xℜy and x is [γF], then y is [γF]. (McCarthy and Prince 1995)

Given this classic definition of IDENT, loss of either aspiration or place should be preferred to segmental fission in Icelandic, even when IDENT[asp] and IDENT[place] are high-ranking. This is shown in the tableau below. In the fission candidate (number 3), one of the output correspondents is not identical to the input segment with respect to place, while the other is not identical to the input segment with respect to aspiration. Thus, fission incurs a violation of both classic IDENT[asp] and IDENT[lab]. However, loss of only one of these features in the non-fission candidates incurs only one IDENT violation. Hence candidates 1 and 2 incur a subset of violations compared to the actual surface form: [Yhpi] is harmonically bounded by [Ypi] and [Yhi]. It will never be optimal given the universally quantified classic IDENT[F], especially since the actual surface form incurs a violation of INTEGRITY and surface faithfulness constraints (and additional markedness constraints such as NOCODA). (215.) Classic IDENT[F] disfavors distributing diphthongization IDENT[asp] IDENT[lab] INTEGRITY /Yph1i/ SURFACE FAITH * 1  Yp1i * 2  Yh1i 3

 Yh1p1i fission

*

*

188

*/*

Struijke

∃-IDENT[F] and MAX[F] are existentially defined and unidirectional, and can therefore capture the idea that distributing diphthongization ensures preservation of the input features. Consider the tableaux below. ∃-IDENT[+asp] and MAX[asp] are satisfied in the Icelandic fission candidate, as are ∃-IDENT[lab] and MAX[lab]. This is because each feature (specification) associated with the aspirated input segment is preserved (on some correspondent of the segment). Note, however, that segmental fission incurs a violation of INTEGRITY and surface faithfulness constraints on aspiration, place, and manner features, which must be low-ranking in Icelandic.90 In the Featural Independence Theory, the constraint DEP[cont] is violated in addition (under the assumption that [h] is a placeless fricative). The constraint I-CONTIGUITY, defined below is also violated. Firstly, epenthesis of [cont] disrupts the underlying adjacency of the vowel features associated with [Y] and the [lab] feature associated with [p]. Secondly, reassociation of the feature [asp] disrupts the underlying adjacency of this feature and the features associated with the final vowel. (216.) I-CONTIGUITY[F]:91 Let feature-x, feature-y ∈ input be in the domain of ℜ, such that feature-x is adjacent to feature-y; then there is some feature-x' ∈ output such that feature-xℜfeature-x' which is adjacent to some feature-y' ∈ output such that feature-yℜfeature-y'. When two features are adjacent in the input, some output correspondent of one is adjacent to some output correspondent of the other. (cf. McCarthy and Prince 1995 - for definition see (1.1); Causley 1997)

90

I have argued for the surface relation within the Featural Attribute Theory assumed in this dissertation. Presumably, the Featural Independence theory also assumes such a relation to account for the similarity effect in feature movement and dissimilation. I will not address the question as to how the surface relation is established in this theory (but see Walker (to appear) who argues for constraints establising a surface relation). Since it is not a central issue to this chapter, I will use the shorthand 'SURFACE FAITH' in the tableaux, and assume that the same violations are incurred in the two theories. 91 For easy comparision to the definition assumed throughout this dissertation, this definition is existentially quantified.

189

Existential Faithfulness

(217.) Distiributing diphthongization with ∃-IDENT[F] in the Featural Attribute Theory INTEGRITY /Yph1i/ ∃-IDENT ∃-IDENT surf. faith [+asp] [lab] *! 1 Yp1i *! 2 Yh1i 3 Yh1p1i */* fission (218.) Distributing diphthongization with MAX[F] in the Featural Independence Theory92 INTEGRITY /Yph1i/ MAX[asp] MAX[lab] I-CONTIG[F] DEP[cont] surf. faith *! 1 Yp1i *! 2 Yh1i 3 Yh1p1i

*

fission

*

*/*

In the Featural Independence Theory, features can be preserved independently of their sponsoring segments. Therefore this theory allows an account of Icelandic diphthongization that involves segmental epenthesis, with reassociation of the feature [asp] onto the epenthesized segment. This is illustrated below.

92

I use indices to indicate segmental correspondence, not featural correspondence.

190

Struijke

(219.) Distributing diphthongization as epenthesis and feature reassociation (Featural Independence Theory) [asp] [lab]

input:

output:

pH

h [cont] [asp]

p [lab]

Tableau (19) compares the analyses possible under the Featural Independence Theory. Candidate 2 corresponds to the epenthesis-and-reassociation form depicted above. Like the segmental fission candidate (number 1), it satisfies all relevant MAX[F] constraints (all features in the input are present in the output), but violates surface faithfulness, DEP[cont], and I-CONTIGUITY[F]. Instead of the INTEGRITY violation of the fission candidate, candidate 2 violates STAY[asp] because [asp] is underlyingly associated with one segment, but associated with another on the surface. In addition, segment epenthesis causes violations of DEP(seg)93 and I-CONTIGUITY(seg). Thus, in the Featural Independence Theory, the ranking of INTEGRITY with respect to STAY[asp], DEP(seg), and

93

Since the Featural Independence Theory assumes faithfulness constraints in both the I→O and O→I directions, the constraint DEP(seg) can be assumed (McCarthy and Prince 1995), rather than M-seg (McCarthy 1993.)

191

Existential Faithfulness

I-CONTIGUITY(seg) determines whether this instance of distributing diphthongization involves segment fission or segment epenthesis and feature reassociation.94

(220.) Distributing diphthongization as segment fission or epenthesis in the Featural Independence Theory ICONTIG INTEG /Yph1i/ MAX[asp] SURF. DEP / MAX[lab] FAITH [cont] [F] 1 Yh1p1i fission

2 Yhp1i epenthesis

*

*

*

*

*

*

STAY[asp] DEP(seg) ICONTIG(seg)

* */*/*

In the Featural Attribute Theory, feature specifications cannot be preserved outside of the segment with which they are associated in the input. Even though the epenthesis candidate (number 2) in tableau (20) is specified as [+asp], undominated ∃-IDENT[+asp] is fatally violated because the epenthesized segment does not correspond to the aspirated input segment. (221.) Distributing dipthongization cannot be due to epenthesis in the Featural Attribute Theory ∃-IDENT[+asp] ∃-IDENT[+lab] /Yph1i/ 1 Yh1p1i fission

2

Yhp1i epenthesis

*!

Thus, in accounting for diphthongization, the Featural Attribute Theory permits both segmental fission and epenthesis with feature reassociation. The Featural Attribute 94

In this example of dipthongization, the resulting output segments are not associated to identical features. In many cases of diphtongization, though, such a state of affairs is found. For instance, in the French-English borrowing diphthongization /ü/ - [iu], both output vowels are high. In the Featural Independence Theory, four possible analyses are (in theory) allowed: 1. segment fission and feature fission; 2. segment fission and epenthesis of one of the [high] features; 2. segment epenthesis and feature fission; 4. segment epenthesis and feature epenthesis. Which of these analyses is found in a language depends on the ranking of INTEGRITY, INTEGRITY[F], DEP(seg), and DEP[F].

192

Struijke

Theory allows fission only. Which theory should ultimately be adopted will in part depend on empirical findings (more on this in section 4.4).

4.3.2 Coalescence (and feature stability)

Like the first type of multiple correspondence (fission), the second type (coalescence) forces us to redefine classic IDENT[F]. Constraints on featural faithfulness must be unidirectional and evaluate a single value of a feature specification to account for featural faithfulness in coalescence (Pater 1999). Both ∃-IDENT[F] and MAX[F] have these characteristics. Causley (1997a) and Pater (1999) point out that classic IDENT[F] constraints cannot determine which input features are preserved in coalescence, and which are lost. A full discussion of this fact was given in section 3.1.1.2. I briefly illustrate it again with the diagrams below. For expository purposes, I use input segments that differ in manner features only. (222.) Coalescence input:

output:

a.

n1

l2

–

—

b.

n1

l2

–

— l1 2

n12

In both of the above input-output mappings, classic IDENT[nas] and IDENT[lat] are violated, because both contain corresponding segments that are not identical with respect to the feature [nasal] or the feature [lateral], namely the pairs /l/ - [n] in (21a) and /n/ - [l] in (21b). Thus, classic IDENT constraints cannot distinguish these two representations. In the tableau below I indicate for each violation which mapping incurs it.

193

Existential Faithfulness

(223.) b

Classic IDENT constraints cannot distinguish candidate a from candidate a b

/ n1 l2 / n12 l12

IDENT[lat] * (—) * (–)

IDENT[nas] * (—) * (–)

Because they cannot distinguish (21a) from (21b), classic IDENT constraints are unable to determine which features are preserved in coalescence, and which are lost. Two different solutions have been proposed: MAX[F] constraints assuming featural independence (Lamontagne and Rice 1995; Causley 1997ab) and unidirectional IDENT[F] constraints specified for feature value (IDENT[+F] and IDENT[-F]; Gnanadesikan 1995, 1997; Pater 1999, McCarthy 1995; discussed extensively in section 3.1.1.2). The latter approach is compatible with the existential constraints proposed here. Both solutions rely on the idea that correspondence constraints are unidirectional and regulate faithfulness to a particular feature value. The constraints MAX[F] and ∃-IDENT[F] are satisfied if a given input feature value is preserved on the coalesced output segment. They do not demand that an output feature (specification) be present in the input. Note that DEP[nas] and DEP[lat] of the Featural Independence Theory are irrelevant here, since no nasal or lateral features are added in the output. The tableaux below show that the mapping /n1l2 / - [n12] is optimal when faithfulness to input [+nasal] (via ∃-IDENT[+nas] or MAX[nas]) is of more importance than faithfulness to input [+lateral] (via ∃-IDENT[+lat] or MAX[lat]). With the reverse ranking the mapping /n1l2 / - [l12] is optimal. (224.) Coalescence in the Featural Attribute Theory / n1 l2 / ∃-IDENT[+nas] ∃-IDENT[lat] * 1 n12 2 l12 *!

194

Struijke

(225.) Coalescence in the Featural Independence Theory / n1 l2 / MAX[nas] MAX[lat] * 1 n12 2 l12 *!

The Featural Attribute Theory provides a single, segment-based analysis of coalescence, shown below for /fd/ - [v]:

(226.) Coalescence in the Featural Attribute Theory ([lab][+cont][-vc]) ([cor][-cont][+vc])

input:

output:

f

d

v ([lab][+cont][+vc])

Constraint evaluations associated with this account are summarized in table (26). Faithfulness constraints on the feature specifications preserved in the output are satisfied, while those contraints on the feature specifications that are lost in the output are violated. (227.) Coalescence in the Featural Attribute Theory ∃-IDENT[cor] ∃-IDENT[lab] ∃-IDENT[+cont ∃-IDENT[-cont] / f1 d 2 / ] ∃-IDENT[-vc] ∃-IDENT[+vc] v12 */*/* coalescence

In the Featural Independence Theory, this pattern can be analyzed as segmental coalescence or as feature stability (as McCarthy (1996) points out). Both are shown below.

195

Existential Faithfulness

(228.) Coalescence in the Featural Independence Theory [lab][cont]

input:

output:

f

[vc][cor]

d

v

[lab] [cont] [vc]

In the feature stability analysis for /fd/ - [v], the second input segment deletes, but its voicing feature is preserved and surfaces on the first segment.

(229.) Feature stability in the Featural Independence Theory [lab][cont]

input:

output:

f

[vc][cor]

d

v

[lab][cont][vc]

196

Struijke

Thus, in the Featural Independence Theory, feature stability competes with segmental coalescence. This competition is illustrated in tableau (30). Both candidates satisfy MAX[voice], MAX[lab], and MAX[cont], but violate MAX[cor]. In a mapping like /ifdu/ - [ivu], they violate O-CONTIGUITY, defined below, because in the output the feature [cont] is adjacent to the features associated with [u], but in the input these features are not adjacent. (230.) O-CONTIGUITY[F]:95 Let feature-x, feature-y ∈ output be in the domain of ℜ, such that feature-x is adjacent to feature-y; then there is some feature-x' ∈ input such that feature-xℜfeature-x' which is adjacent to some feature-y' ∈ input such that feature-yℜfeature-y'. When two features are adjacent in the output, some input correspondent of one is adjacent to some input correspondent of the other. (cf. McCarthy and Prince 1995; Causley 1997)

The feature stability candidate violates two additional constraints, namely MAX(seg) and the feature-segment association constraint STAY[voice]. In the absence of a constraint explicitly banning coalescence (such as UNIFORMITY, as argued in this dissertation and Keer 1999, Pater 1999), the coalescence analysis incurs a subset of the constraint violations incurred by feature stability. As a result, coalescence should always be preferred to feature stability. This is shown in the tableau below. (231.) Coalescence rather than feature stability in the Featural Independence Theory / f1 d2 / MAX[voice] MAX[cor] MAX(seg) STAY[VC] MAX[lab] MAX[cont] 1 v1 * * * [F] stability 2 v12 * coalescence

Again, I use the existential definition of the constraint for easy comparison with the definition of ∃contiguity defined in appendix I.

95

197

Existential Faithfulness

In conclusion, both feature theories can accomodate a segmental analysis of phenomena known as coalescence or feature stability.96 In principle, the Featural Independence Theory allows a featural analysis in addition. However, under present assumptions the featural analysis is harmonically inaccessible.

4.3.3 Feature movement

Here I compare the account of feature movement in the Featural Attribute Theory proposed in chapter 3 with the accounts allowed in the Featural Independence Theory (McCarthy 1996, Parker 1997, Walker 1997). Again, both theories are successful in capturing the phenomenon because their respective featural faithfulness constraints share certain characteristics. Existential quantification, unidirectionality, and feature value specificity are all important in accounting for this phenomenon. In chapter 3, I focused on laryngeally marked feature movement in Sanskrit and Cuzco Quechua, and marked vowel height movement in Esimbi. Here I illustrate the arguments with marked place feature movement attested in Tyneside English (McCarthy 1996), and Kirundi (Goldsmith 1990). In both languages, the oral place feature of a voiceless stop moves to an adjacent nasal. The stop neutralizes to a glottal. (232.) Tyneside English place feature movement hå/m« ‘happen’ b√/n« ‘button’ tSI/N« ‘chicken’

96

See Stahlke (1976) for a coalescence analysis of French vowel nasalization traditionally seen as feature stability (pl[e]nitude ‘fullness’ - pl[E)] 'full (masc.)’).

198

Struijke

(233.) Kirundi place feature movement ku-temera ku-n-hemera gu-korera ku-N -horera

'to cut for someone / me' 'to work for (me)'

In the Featural Attribute Theory, the only way feature specifications can ‘move’ from one segment to another is through segmental correspondence by means of fission and coalescence, as I have argued in chapter 3. The input segment /p/ in Tyneside /håpn/ must have two output correspondents, one of which carries the place feature and coalesces with the /n/. The result of this coalescence is [m]. This state of affairs is depicted below. The argument is laid out in detail in section 3.1. (234.) Tyneside English feature ‘movement’ as segmental splitting and coalescence [lab] [cor] [-cont] [+nas] [-vc] [+vc]

input:

h

å

p

n

output:

h

å

/

m

[-cont] [-vc] [+CG]

[+lab] [+nas] [+vc]

With the existential and unidirectional definition of IDENT[F]IO, fission of the input stop is allowed because all of its features are preserved on some output correspondent (see section on distributing diphthongization above). The fact that ∃-IDENT[F] constraints are unidirectional and evaluate particular feature values ensures that coalescence is a possible operation in the Featural Attribute Theory (see section on coalescence above). However, we have seen that fission and coalescence are not gratis.

199

Existential Faithfulness

This particular instance of fission incurs a violation of INTEGRITY and some surface faithfulness constraints, while the coalescence incurs a violation of ∃-IDENT[cor]. These constraints must be low-ranking to permit feature movement in the Featural Attribute Theory. (235.) Tyneside English feature movement in the Featural Attribute Theory SURFACE INTEGRITY /håp1n2/ ∃-IDENT ∃-IDENT FAITH [lab]IO [cor]IO * * * hå/1m 12 seg. fission and coalecence

In the Featural Independence Theory, both segmental and featural analyses are allowed. They are shown below. (236.) Feature movement in the Featural Independence Theory (a) segmental fission (b) [F] movement [lab]

input:

output:

[cor][nas]

[lab]

[cor][nas]

Q p

n

h

Q p

n

h Q /

m

h

Q

m

h

[CG]

[lab][nas]

/ [CG]

[lab][nas]

The first is analogous to the Featural Attribute Theory analysis and incurs comparable constraint violations (with additional violations of DEP[CG] and I-CONTIGUITY[F]

200

Struijke

(because the feature [lab] is not adjacent to the vowel features, as in the input97 )). In the second analysis, segmental fission and coalescence do not take place. Instead a feature delinks from its original segment and relinks to another segment, so violating STAY[lab]. The relative ranking of STAY with respect to INTEGRITY determines whether the segmental or featural pattern is preferred. (237.) Tyneside English feature movement in the Featural Independence Theory MAX MAX DEP SURF. I-CONTIG INTEG STAY /håp1n2/ [lab] [cor] [CG] FAITH [F] [lab] * * * * * hå/1m 12 segm fission and coalescence

hå/m

*

*

*

*

*

reassociation

Thus, in the Featural Independence Theory, fission and coalescence do not have to take place to effect feature movement; it can also be accomplished through feature reassociation. The Featural Attribute Theory is more restrictive in that it only allows the former strategy.

4.3.4 Dissimilation

In chapter 3, I argued that dissimilation takes place to ensure faithfulness and reduce markedness. Both the Featural Attribute Theory and the Featural Independence Theory allow analyses that accord with this claim. As in feature movement, the Featural Attribute Theory account involves segmental fission and coalescence, and the Featural Independence Theory allows an additional feature reassociation account. I illustrate these

97

When features move across segments (see e.g. Esimbi, Sanskrit, Cuzco Quechua), LINEARITY[F] is violated in addition. I showed in section 1.2.4 that feature movement does not incur ∃-LINEARITY and ∃-CONTIGUITY violations in the Featural Attribute Theory.

201

Existential Faithfulness

analyses with rhotic dissimilation in Georgian (Fallon 1993; Odden 1994; Walsh Dickey 1997; Struijke and de Lacy 2000), exemplified below with forms containing the suffix /-uri/. (238.) Rhotic dissimilation in Georgian forms containing the suffix /-uri/ dan-uri 'Danish' somX-uri 'Armenian' asur-uli 'Assyrian' ungr-uli 'Hungarian' tSer k'ez-uli 'Circassian'

The core idea underlying the proposed dissimilation analysis is that the single output [r] in these words ensures faithfulness of the [+rhotic] feature specifications of all instances of input /r/. The existential, unidirectional, and value specific character of featural faithfulness constraints ensures that this is possible. The segment-based analysis of the Featural Attribute Theory is depicted below. One rhotic segment undergoes fission, and one of its correspondents coalesces with the other rhotic segment. Since the feature specifications of both input segments are preserved on some output correspondent, existentially defined ∃-IDENT[F]IO constraints such as ∃-IDENT[+rhotic]IO and ∃-IDENT[-lat]IO are satisfied. (239.) Dissimilation in the Featural Attribute Theory (only relevant correspondence relations and features are shown) [+rhotic]

[+rhotic]

input:

u

s

u

r

u

r

output: u

s

u

r

u

l

i

[+rhotic]

202

[+lat]

i

Struijke

The segmental fission involved in this dissimilation analysis incurs violations of INTEGRITY and surface faithfulness. The unidirectional (I¡O) and value specific character of featural faithfulness constraints ensures that the mapping /r/ - [l] does not incur a violation of ∃-IDENT[+lat]. Note also that coalescence can take place without being penalized by ∃-IDENT[F] constraints, because the contributing segments are identical. The constraint evaluations are summarized in the table below. (240.) Constraint evaluation of dissimilation in the Featural Attribute Theory /asur1-ur2i/ ∃-IDENT [+rhotic] ∃-IDENT [-lateral] SURFACE FAITH INTEGRITY asur12ul2i * * seg fission and coalescence

The segment-based analysis of the Featural Independence Theory is shown below. In addition to coalescence of the two rhotic segments it involves vacuous coalescence of the two [rhotic] features. Since every segment and feature of the input is preserved in the output, existentially defined MAX(seg) and MAX[F] constraints are satisfied. As in the Featural Attribute Theory analysis, INTEGRITY and surface faithfulness are violated. In this theory, however, the mapping /asururi/ - [asuruli] incurs additional violations of the following constraints: DEP[lat] (because the feature [lat] of the output is not present in the input); LINEARITY[F] (because in the input the second [rhotic] feature follows the features associated with the second [u], but precedes them in the output); and I-CONTIGUITY[F] (because the second [rhotic] feature is adjacent to features associated with [i] in the input, but not in the output).

203

Existential Faithfulness

(241.) Segment-based analysis of dissimilation in the Featural Independence Theory (only relevant correspondence relations and features are shown)

[rhotic]

[rhotic]

input:

u

s

u

r

u

r

i

output:

u

s

u

r

u

l

i

[rhotic]

[lat]

The INTEGRITY violation of the segment-based analysis can be avoided in the Featural Independence Theory if the [rhotic] feature of the second segment reassociates to the first [r], resulting in [rhotic] coalescence without segment coalescence or fission, as shown below. (242.) Feature analysis of dissimilation in the Featural Independence Theory (only relevant correspondence relations and features are shown)

[rhotic]

[rhotic]

input:

u

s

u

r

u

r

output: u

s

u

r

u

l

i

[rhotic]

204

[lat]

i

Struijke

Although this structure avoids an INTEGRITY violation, it causes a violation of STAY[rhotic]. With respect to all other faithfulness constraints, however, these two structures perform equally. Thus, whether segmental fission and coalescence take place in the Featural Independence Theory to produce dissimilation, or whether a feature reassociates depends on the relative ranking of STAY[rhotic] and INTEGRITY. (243.) Dissimilation in the Featural Independence Theory MAX DEP SURF LIN[F] /asur1-ur2i/ [rhot] [lat] FAITH ICONTIG[F] 1 as12urul2i * * */* seg. fission and

INTEG

STAY [rhot]

*

coalescence

2

asuruli

feature coalescence

*

*

*/*

*

In conclusion, as in the other phenomena discussed in this chapter, the Featural Independence Theory allows two analyses of dissimilation, while the Featural Attribute Theory allows only one. 4.4 Conclusion In this chapter, I have investigated the status of features in Correspondence Theory, assuming that faithfulness is existentially defined. The comparison of the Featural Independence Theory and the Featural Attribute Theory has shown that the former assumes a plurality of correspondence relations and faithfulness constraints, allowing an additional analysis of each of the phonological phenomena discussed. As I mentioned in the introduction to this chapter, it is unclear at this point which theory is to be preferred. In order to distinguish the two empirically, we need to study languages exhibiting several of these phenomena. For instance, under the Featural Attribute Theory, all patterns involving phonologically driven segmental fission should be banned when INTEGRITY is undominated (i.e. dissimilation, feature movement and

205

Existential Faithfulness

diphthongization, as well as e.g. copy-epenthesis98 ). When the constraint is undominated in the Featural Independence Theory, the segmental versions of these patterns are not allowed either, but the featural versions could still be attested. Thus, more research is needed. particularly in the additional areas that were believed to support the autosegmental character of features (e.g. Goldsmith 1990). One such area is assimilation and harmony (feature agreement). Recent research (e.g. Bakovic 1999; Walker to appear; Walker and Rose 2000) has accounted for agreement in the Featural Attribute Theory, by means of IDENT[F] constraints. In the concluding chapter of this dissertation I will suggest a possible analysis of feature agreement that links it formally to the phenomena discussed earlier. I will show that this link allows an account of, for instance, the fact that movement and agreement of a given feature cooccur in many languages. This research needs to be extended also to suprasegmental features such as tones. Many arguments and case studies presented in support of autosegmental theory come from the study of tones, and future study must determine whether tones can be considered attributes of segments on analogy with subsegmental features, or whether they are true autosegments.

98

Recall that morphological reduplication does not incur a violation of INTEGRITY.

206

Appendix III Sanskrit ranking ∃-MORPH Dis

LICENCE [+murm]

ID E N T [-cont]∑∑

IDENT [+vc]∑∑

∃-IDENT [+murm]

LARREL

∃-IDENT [+asp]

∃-IDENT [-asp]

∃-IDENT [+SG]

IDENT [+asp]∑∑

IDENT [+murm]∑∑

∃-IDENT [+vc]

∃-IDENT[-vc] 1

*SG

∃-IDENT [-VC]RT

M-Seg

IDENT∑∑ place1

IDENT [+SG]∑∑

∃-IDENT [+SG]RT-IN

INTEGRITY

∃-IDENT[-murm]

Since murmur can be thrown back onto consonants of any place, all surface faithfulness constraints relevant for place (summarized here as IDENT∑∑-place-) must be low-ranking.

Struijke

5 Conclusion

In this concluding chapter, I present the main results of this dissertation, and some of its predictions. In addition, I identify a number of remaining issues, and point to areas of further research. The core proposal of this dissertation holds that faithfulness regulates recoverability, and constraints thereon are existentially quantified. The main consequence of this proposal is that segmental fission can ensure preservation of underlying material, while simultaneously satisfying markedness requirements. I have studied three phenomena which utilize this property of segmental fission. Reduplication always involves fission. As a result the unmarked can emerge in reduplicated words. In the other phenomena discussed, IO faithfulness and markedness constraints conspire to force fission. Apart from feature movement and dissimilation, these include distributing diphthongization and copy epenthesis. We also expect phonological reduplication to take place: IO faithfulness and markedness constraints may force copying of multiple segments, giving the false impression that morphological reduplication has taken place. I have argued for a general, ‘broad’ input-output correspondence relation. No relation between the input and the output is specific to affixes or reduplication. In reduplicated words, both members of the base-reduplicant pair are related directly to the input. Therefore, the base is not logically or temporally prior to the reduplicant. Because the base and reduplicant are truly generated in parallel through segmental fission, a reduplicated word in which the reduplicant is more like the input than the base does not constitute an instance of opacity. The proposed account of feature movement does not involve opacity either. In some previous accounts, movement was believed to involve assimilation of a feature,

211

Existential Faithfulness

followed by deletion of the feature on the segment that was underlyingly associated with it. Again, the present proposal does not involve opacity because GEN can establish multiple correspondence freely, and fissioned segments can differ from the input and from each other. Thus, this dissertation contributes to the growing body of research showing that certain patterns previously thought to display opacity (and therefore appear problematic for a parallel model of the grammar) can be analyzed straightforwardly in Optimality Theory. The explanation presented for dissimilation is also well-grounded in current research. I have built on work arguing that dissimilation is driven by general markedness requirements, and have argued that these constraints need not refer to multiple instances of a given marked element. I have introduced a general surface relation, established between segments related to a single input correspondent. Every (morpho-) phonological pattern involving segmental fission is subject to surface faithfulness constraints. In reduplication they are responsible for any degree of identity between the reduplicant and base. In feature movement and dissimilation, they account for the similarity effect, among others. At this point, the nature of the surface relation is not entirely clear. In particular, more study is needed to determine whether constraints on this relation are existentially or universally defined. As I explained in chapter 2, we need to look into patterns involving multiple correspondence along the surface relation in order to determine the correct quantification. If surface faithfulness constraints are existentially quantified, they must be unidirectional and feature value specific. If they are universally quantified (on analogy with classic base-reduplicant constraints), they are bidirectional and neutral with respect to feature value.

212

Struijke

Further research is also needed to determine whether the status of features (independent entities or properties of segments) has any important consequences for Correspondence Theory and its ability to explain a wide range of phonological patterns. With respect to the phenomena discussed in this dissertation, the nature of their status seems of no consequence to empirical coverage. Finally, I would like to point to one further area of investigation, namely assimilation and harmony, or 'agreement.' Although it does not bear on the main thesis of this dissertation, I believe there are interesting parallels to be drawn between this phenomenon and both feature movement and dissimilation. In feature agreement, multiple segments in the surface word are specified for a particular feature specification, even though the underlying representation often contains only one segment bearing that specification. It seems plausible that all segments involved in feature agreement are related to that particular input segment. For instance, in the place agreement seen in the Luganda examples below (taken from Katamba (1989), both labial output segments are related to the single labial input segment. Both preserve the feature specification [labial].

(1.)

/n + bala/

m-bala

I count

/n + pa/

m-pa

I give

/n + mala

m-mala

I finish

Again, the relation between a single input segment and multiple output segments is formalized as segmental fission. As in feature movement and dissimilation, fission and coalescence go hand in hand. This is depicted below. (A similar account of agreement was given in Bakovic (1997)).

213

Existential Faithfulness

(2.)

Luganda agreement as segmental fission and coalescence [cor]

[lab]

input:

n

b

a

l

a

output:

m

b

a

l

a

[lab]

[lab]

The crucial difference between feature agreement on the one hand, and feature movement and dissimilation on the other is the number of output correspondents that is featurally alike an input segment. In agreement, all surface segments resemble the underlying segment, while only one of these segments does in the latter two phenomena. This approach to feature agreement has several advantages. First, the formal similarity between feature agreement and feature movement may bring us to an understanding of languages in which a feature sometimes moves and sometimes assimilates. Also, it allows us to see harmony and dissimilation as two sides of the same coin (see also Krämer (1999)). Lastly, and perhaps most importantly, segmental fission ensures that fissioned output segments are in a surface relation with each other. Constraints on this relation demand that these segments are faithful to each other, not just with respect to the harmonizing feature, but with respect to other features as well. Therefore, agreement is most harmonic with respect to surface faithfulness constraints when the segments involved share multiple feature specifications, and is thus more prone to occur when segments are alike. As Walker (to appear) and Rose and Walker (2000) have shown, a

214

Struijke

surface relation is crucial in accounting for the similarity effect in feature agreement1 (as it is in feature movement and dissimilation).

1

See also Krämer (1999), who argues that agreement is driven by surface faithfulness (dubbed 'Syntagmatic Identity').

215

Existential Faithfulness

References Alderete, John. 1996. Prosodic Faithfulness in Cupeño. To appear in University of Massachusetts Occasional Papers in Linguistics, Papers on Indigenous Languages, ed. by E. Benedicto. Amherst, GLSA. [ROA 131]. Alderete, John. 1997. Dissimilation as local conjunction. Proceedings of the North East Linguistics Society 27, ed. by K. Kusumoto. 17-32. Amherst: GLSA. [ROA 175]. Alderete, John. 1999. Morphologically Governed Accent in Optimality Theory. Ph.D. dissertation University of Massachusetts, Amherst. Alderete, John, Jill Beckman, Laura Benua, Amalia Gnanadesikan, John McCarthy and Suzanne Urbanczyk. 1999. Reduplication with fixed segmentism. Linguistic Inquiry, 30. 327-364. Alderete, John and Joe Pater. (in prep.) Dissimilation as Cumulative Markedness. Ms., Swarthmore College and University of Massachusetts. Allen, W.S. 1951. Some prosodic aspects of retroflexion and aspiration in Sanskrit. Bulletin of the School of Oriental and African Studies, 13. 939-946. Anderson, Stephen R. 1970. On Grassmann’s Law in Sanskrit. Linguistic Inquiry, 1. 387-396. Anderson, Stephen R. 1972. On nasalization in Sundanese. Linguistic Inquiry, 3. 253268. Anderson, Stephen R. 1974. The Organization of Phonology. New York: Academic Press. Andrade, Manuel. 1933. Quileute. New York: Columbia University Press. Archangeli, Diana and Douglas Pulleyblank. 1994. Grounded Phonology. Cambridge, Mass: MIT Press. Bach, Emmon. 1975. Long vowels and stress in Kwakiutl. Texas Linguistic Forum, vol. 2. 9-19. Bakovic, Eric. 1997. Paper presented at the Rutgers-Maryland OT Workshop (RUMD). New Brunswick: Rutgers University. Bakovic, Eric. 1999. Harmony, Dominance and Control. Ph.D. dissertation Rutgers University, New Brunswick. Bates, Dawn. 1986. An analysis of Lushootseed diminutive reduplication. Proceedings of the Twelfth Annual Meeting of the Berkeley Linguistics Society, ed. by V. Nikiforidou, M. VanClay, and M. Niepokuj. 1-12. Berkeley Linguistics Society, University of California, Berkeley. Bates, Dawn, Thom Hess and Vi Hilbert. 1994. Lushootseed Dictionary. Seattle: University of Washington Press. Beckman, Jill. 1995. Shona height harmony: Markedness and positional identity. University of Massachusetts Occasional Papers in Linguistics 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. Amherst: GLSA. Beckman, Jill. 1997. Positional Faithfulness. Ph.D. dissertation University of Massachusetts, Amherst. (Published by Garland, New York, 2000).

216

Struijke

Bell, Sarah. 1983. Internal C reduplication in Shuswap. Linguistic Inquiry, 14. 332338.Benton, Richard. 1971. Pangasinan Reference Grammar. Honolulu: University of Hawaii Press. Benua, Laura. 1995. Identity effects in morphological truncation. University of Massachusetts Occasional Papers in Linguistics 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. 77-136. Amherst: GLSA. [ROA 74]. Benua, Laura. 1997. Transderivational identity: phonological relations between words. Ph.D. dissertation University of Massachusetts, Amherst (Published by Garland, New York, 2000). Bickmore, Lee. 1996. Bantu tone spreading and displacement as alignment and minimal misalignment. Ms., University of Albany. [ROA 161]. Bloch, Bernhard. 1948. A set of postulates for phonemic analysis. Language, 24. 3-46. Boas, Franz. 1947. Kwakiutl Grammar with a Glossary of the Suffixes, ed. by Helene Boas Yampolsky. Transactions of the American Philosophical Society. New Series, Vol. 37, part 3. 201-377. Boas, Franz. 1948. Kwakiutl Dictionary, ed. by Helene Boas Yampolsky. Ms., American Philosophical Society. Borowsky, Toni. 2000. Word-Faithfulness and the direction of assimilation. The Linguistic Review, 17. Borowsky, Toni and Armin Mester. 1983. Aspiration to roots: Remarks on the Sanskrit diaspirates. Proceedings of CLS 19. Chicago: Chicago Linguistic Society. Broselow, Ellen. 1983. Salish double reduplications: Subjacency in morphology. Natural Language and Linguistic Theory, 1. 317-346. Broselow, Ellen and John McCarthy. 1984. A theory of internal reduplication. The Linguistic Review, 3. 25-88. Burzio, Luigi. 1994. Principles of English Stress. Cambridge: Cambridge University Press. Cahill, Michael. 1999. Aspects of the phonology and morphology of Konni. Ph.D. dissertation, Ohio State University. Carenko, E.I. 1975. On Laryngealization in Quechua. Linguistics, An International Review, 146. 5-14. Casali, Roderic F. 1996. Resolving Hiatus. Ph.D. dissertation University of California Los Angeles. Causley, Trisha. 1997a. Identity and featural correspondence: The Athapaskan case. Proceedings of the North East Linguistics Society 27, ed. by K. Kusumoto. Amherst: GLSA. Causley, Trisha. 1997b. Featural correspondence and correspondence constraints. Paper presented at HILP3, Free University, Amsterdam. Causley, Trisha. 1999. Complexity and Markedness in Optimality Theory. Ph.D. dissertation University of Toronto. Chomsky, Noam. 1966. Topics in the Theory of Generative Grammar. The Hague: Mouton & Co. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Chomsky, Noam and Morris Halle. 1968. The Sound Pattern of English. New York: Harper and Row.

217

Existential Faithfulness

Clements, G.N. 1985. The geometry of phonological features. Phonology Yearbook, 2. 225-52. Cole, Jennifer. 1995. The cycle in phonology. The Handbook of Phonological Theory, ed. by John A. Goldsmith. Oxford: Blackwell Publishers. Cole, Jennifer. 1997. Deletion and Recoverability in Klamath. Ms., University of Illinois. Cole, Jennifer, and Charles Kisseberth. 1994. An optimal domains theory of harmony. Studies in the Linguistic Sciences, 24. Diffloth, Gérard. 1976a. Minor-syllable vocalism in Senoic languages. Austroasiatic Studies, Part I, ed. by P.N. Jenner, L.C. Thompson, and S. Starosta. 229-247. University Press of Hawaii. Diffloth, Gérard. 1976b. Expressives in Semai. Austroasiatic Studies, Part I, ed. by P.N. Jenner, L.C. Thompson, and S. Starosta. 249-264. University Press of Hawaii. Evans, Nick. 1995. Current issues in the phonolgy of Australian languages. In: The Handbook of Phonological Theory, ed. by J.A. Goldsmith. 723-761. Fallon, Paul. 1993. Liquid dissimilation in Georgian. Ms., Ohio State University. Firth, J.R. 1948. Sounds and prosodies. Transactions of the Philological Society. 127152. Fitzgerald, Colleen. 1998. Vowel hiatus and faithfulness in Tohono O’odham reduplication. Ms., State University of New York at Buffalo. Fitzgerald, Colleen. 1999. Unfaithful bases and syncope in Tohono O’odham reduplication. Paper presented at WCFL. University of Arizona, Tucson. Foley, W. 1991. The Yimas Language of New Guinea. Stanford: Stanford University Press. Frisch, Stefan. 1996. Similarity and frequency in phonology. Ph.D. dissertation Northwestern University. Frisch, Stefan, Michael Broe, and Janet Pierrehumbert. 1997. Similarity and phonotactics in Arabic. Ms., Northwestern University. [ROA 223]. Fukazawa, Haruka. 1999. Theoretical Implications of OCP Effects on Features in Optimality Theory. Doctoral dissertation University of Maryland, College Park. Galloway, Brent Douglas. 1977. A Grammar of Chilliwack Halkomelem, vol. 1. Gnanadesikan, Amalia. 1995. Markedness and faithfulness constraints in child phonology. Ms., University of Massachusetts, Amherst. Gnanadesikan, Amalia. 1997. Phonology with Ternary Scales. Ph.D. dissertation University of Massachusetts, Amherst. Goldsmith, John. 1976. Autosegmental Phonology. Ph.D. dissertation MIT, Cambridge, Mass. Goldsmith, John. 1990. Autosegmental and Metrical Phonology. Oxford: Blackwell Publishers. Grassmann, Herman. 1863. Ueber die Aspiranten und ihr gleichzeitiges Vorhandensein in An- und Auslaute der Wurzeln. Zeitschrift fuer vergleichende Sprachforschung auf dem Gebiete des Deutschen, Griechischen und Lateinischen, 12. 81-138. Hale, Kenneth. 1959. A Papago Grammar. Ph.D. dissertation. Indiana University. Halle, Morris, James W. Harris and Jean-Roger Vergnaud. 1991. A reexamination of the Stress Erasure Convention and Spanish stress. Linguistic Inquiry, 22. 141159.

218

Struijke

Harris, James W. 1969. Spanish Phonology. Cambridge, Mass.: MIT Press. Hayes, Bruce. 1987. A revised parametric metrical theory. Proceedings of the North East Linguistics Society 17, ed. by J. McDonough and B. Plunkett. Amherst: GLSA. Hayes, Bruce. 1989. Compensatory lengthening in moraic phonology. Linguistic Inquiry, 20. 253-306. Hayes, Bruce. 1990. Diphthongization and coindexing. Phonology, 7. 31-71. Hayes, Bruce. 1995. Metrical Stress Theory: Principles and Case Studies. Chicago: University of Chicago Press. Hawkins, Peter. 1984. Introducing Phonology. London: Hutchinson University Library. Hewitt, Mark S. and Megan J. Crowhurst. 1996. Conjunctive constraints and templates in Optimality Theory. Proceedings of the North East Linguistics Society 26, ed. by K. Kusumoto. 101-116. Amherst: GLSA. Hill, Jane and Ofelia Zepeda. 1992. Derived words in Tohono O’odham. International Journal of American Linguistics, 58. 355-404. Hoard, J. 1975. On incorporating Grassmann’s Law into Sanskrit phonology. Proceedings of the first Annual meeting of the Berkeley Linguistics Society, ed. by Cogen et al. Berkeley: University of California. 207-220. Hockett, Charles. 1947. Componential analysis of Sierra Popoluca. International Journal of American Linguistics, 13. 259-67. Hornberger, Esteban and Nancy H. Hornberger. 1977. Diccionario tri-lingüe, Vol. 2. Quechua-English, quechua-español. Hume, Elizabeth V. 1992. Front Vowels, Coronal Consonants, and their Interaction in Nonlinear Phonology. Ph.D. dissertation Cornell University, Ithaca. Hung, Henrietta. 1994. The Rhythmic and Prosodic Organization of Edge Constituents. Ph.D. Dissertation Brandeis University, Waltham. Hyman, Larry. 1988. Underspecification and vowel height transfer in Esimbi. Phonology 5. 255-273 Inkelas, Sharon and Cheryl Zoll. 2000. Reduplication as Double Stem Selection. Ms, University of California Berkeley, and Massachusetts Institute of Technology, Cambridge Mass. [ROA 412] Itô, Junko. 1986. Syllable Theory in Prosodic Phonology. Ph.D. dissertation University of Massachusetts, Amherst. (Published by Garland, New York, 1988). Itô, Junko and Armin Mester. 1993. Licensed segments and safe paths. Canadian Journal of Linguistics, 38. 197-213. Itô, Junko and Armin Mester. 1995. Japanese Phonology. The Handbook of Phonological Theory, ed. by J. Goldsmith. 817-838. Oxford: Blackwell Publishers. Itô, Junko and Armin Mester. 1996. Rendaku 1: constraint conjunction and the OCP. Ms., University of California Santa Cruz. [ROA 144] Itô, Junko and Armin Mester. 2000. Covert generalizations in Optimality Theory. Talk presented at NELS 31. Georgetown University, Washington DC. Jakobson, Roman. 1939. Observations sur le classment phonologique des consonnes. Proceedings of the Third International Congress of Phonetic Sciences. 34-41.

219

Existential Faithfulness

Jakobson, Roman, Gunnar Fant and Morris Halle. 1952. Preliminaries to Speech Analysis. Cambridge, Mass.: MIT Press. Kang, O. 2000. A correspondence approach to glide-formation in Korean. ROA384. Katamba, Francis. 1989. An Introduction to Phonology. London: Longman. Kawu, Ahmadu. 1999. Paper presented at RUMD II. University of Maryland, College Park and Johns Hopkins University, Baltimore. Keer, Edward. 1999. Geminates, the OCP and the Nature of Con. Ph.D. dissertation Rutgers University, New Brunswick. Kiparsky, Paul. 1965. Phonological change. Ph.D. dissertation MIT, Cambridge, Mass. Kiparsky, Paul. 1997. Class handouts Linguistics Institute. Cornell University, Ithaca. Kitto, Catherine. 1997. Epenthesis and Polynesian loan-words. MA thesis University of Auckland. Kitto, Catherine and Paul de Lacy. 1999. Correspondence and epenthetic quality. Proceedings of AFLA VI, ed. by C. Kitto and C. Smallwood. Krämer, Martin. 1999. A correspondence approach to vowel harmony and disharmony. Ms., University of Düsseldorf. Kuipers, Aert H. 1974. The Shuswap Language. The Hague: Mouton. Ladefoged, Peter. 1971. Preliminaries to Linguistic Phonetics. University of Chicago Press. Ladefoged, Peter and Ian Maddieson. 1996. The Sounds of the World’s Languages. Oxford: Blackwell Publishers. Lamontagne, Greg and Keren Rice. 1995. A correspondence account of coalescence. University of Massachusetts Occasional Papers in Linguistics 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. 211-223. Amherst: GLSA. Leben, Will. 1971. The morphophonemics of tone in Hausa. Papers in African Linguistics, ed. by C.-W. Kim and H. Stahlke. 201-218. Albert: Linguistic Research, Inc. Leben, Will. 1973. Suprasegmental phonology. Ph.D. dissertation MIT, Cambridge Mass. Lombardi, Linda. 1991. Laryngeal Features and Laryngeal Neutralization. Ph.D. dissertation University of Massachusetts. Amherst. (Published by Garland, New York). Lombardi, Linda. 1995. Why place and voice are different: Constraint interactions and featural faithfulness in Optimality Theory. Ms., University of Maryland, College Park. Lombardi, Linda. 1996. Positional Faithfulness and voicing assimilation in Optimality Theory. Ms., University of Maryland, College Park. Lombardi, Linda. 1998. Evidence for Max-Feature constraints from Japanese. University of Maryland Working Papers in Linguistics, 7, Papers in Phonology, ed. by H. Fukazawa, F. Morelli, C. Struijke, and Y. Su. 41-62. Lombardi, Linda. 1999. Positional faithfulness and voicing assimilation in Optimality Theory. Natural Language and Linguistic Theory, 17. 267-302. Mannheim, Bruce. 1991. The Language of Inka Since the European Invasion. Austin: University of Texas Press.

220

Struijke

MacEachern, Margaret R. 1999. Laryngeal Coocurrence Restrictions. New York: Garland (Ph.D. dissertation University of California Los Angeles). Marantz, Alec. 1982. Re Reduplication. Linguistic Inquiry, 13. 483-545. McCarthy, John. 1979. Formal Problems in Semitic Phonology and Morphology. Doctoral dissertation MIT, Cambridge, Mass. (Published by Garland, New York, 1985). McCarthy, John. 1981. The theory of nonconcatenative morphology. Linguistic Inquiry 12. 373-418. McCarthy, John. 1993. The parallel advantage: containment, consistency, and alignment. Talk presented at Rutgers Optimality Workshop I. Rutgers University, New Brunswick. McCarthy, John. 1995. Extensions of faithfulness: Rotuman revisited. Ms., University of Massachusetts, Amherst. (to appear in NNLT 2000). McCarthy, John. 1996. Class handouts from Linguistics 603. University of Massachusetts, Amherst. McCarthy, John. 1997a. Faithfulness and prosodic circumscription. [ROA]. (to appear in Optimality Theory: Phonology, Syntax and Acquisition, ed. by J. Dekkers, F. van der Leeuw and J. van de Weijer. Oxford: Oxford University Press). McCarthy, John. 1997b. Prosodic Morphology in Optimality Theory. Course handouts Linguistics Institute. Cornell University, Ithaca. McCarthy, John. 1998. Sympathy and phonological opacity. Ms, University of Massachusetts, Amherst. McCarthy, John. 1998b. Constraints on word edges. Paper presented at Johns Hopkins University, Baltimore. McCarthy, John and Alan Prince. 1986. Prosodic Morphology. Ms., University of Massachusetts, Amherst, and Brandeis University, Waltham. McCarthy, John and Alan Prince. 1990. Foot and word in prosodic morphology: the Arabic broken plurals. Natural Language and Linguistic Theory, 8. 209-282. McCarthy, John and Alan Prince. 1993b. Generalized alignment. Yearbook of Morphology, ed. by G. Booij and J. van Marle. 79-153. McCarthy, John, and Alan Prince. 1994a. The Emergence of the Unmarked: Optimality in Prosodic Morphology. Proceedings of the North-East Linguistic Society 24, ed. by Merce Gonzalez. 333-379. Amherst: GLSA. McCarthy, John, and Alan Prince. 1994b. An Overview of Prosodic Morphology: Parts I and II. Talks presented at the OTS/HIL Workshop on Prosodic Morphology. Utrecht University. McCarthy, John and Alan Prince. 1995. Faithfulness and reduplicative identity. University of Massachusetts Occasional Papers in Linguistics 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. 249-384. Amherst: GLSA. McCarthy, John and Alan Prince. 1999. Faithfulness and identity in Prosodic Morphology. The Prosody Morphology Interface, ed. by R. Kager, H. van der Hulst, and W. Zonneveld. Cambridge: Cambridge University Press. Mester, Armin. 1986. Studies in tier structure. Ph.D. dissertation University of Massachusetts, Amherst. Montler, Timothy. 1989. Infixation, reduplication and metathesis in the Saanish actual aspect. Southwest Journal of Linguistics.

221

Existential Faithfulness

Morén, Bruce. 1997. Markedness and Faithfulness Constraints on the Association of Moras: Vowel Length and Consonant Weight in Three English Dialects. Paper presented at the Hopkins Optimality Theory Workshop/Maryland Mayfest, Baltimore. Morén, Bruce. 1999. Distinctiveness, Coercion and Sonority: a Unified Theory of Weight. Ph.D. dissertation University of Maryland, College Park. Myers, Scott. 1997. OCP effects in Optimality Theory. Natural Language and Linguistic Theory, 15. 847-892. Nelson, Nicole. 1998. Right Anchor, aweigh. Ms, Rutgers University, New Brunswick. Nelson, Nicole. 2000. Revisiting EDGE-ANCHOR: further restricting asymmetric anchoring. Paper presented at HUMDRUM, Rutgers University. Odden, David. 1994. Adjacency parameters in phonology. Language 70. 289-330. Orr, Caroline and Robert E. Longacre. 1968. Proto-Quechumaran. Language, 44. 528555. Ota, Mitsuhiko. 1999. Phonological Theory and the Acquisition of Prosodic Structure: Evidence from Child Japanse. Ph.D. dissertation Georgetown University, Washington, DC. Padgett, Jaye. 1992. OCP subsidiary features. In Proceedings of the North East Linguistic Society, 22. Amherst: GLSA. 335-346. Padgett, Jaye. 1995. Feature Classes. University of Massachusetts Occasional Papers in Linguistics, 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. Amherst: GLSA. Parker, Steve. 1997. An OT account of laryngealization in Cuzco Quechua. Work Papers of the Summer Institute of Linguistics, 41, University of North Dakota Session. 1-11. Parker, Steve and David Weber. 1996. Glottalized and aspirated stops in Cuzco Quechua. International Journal of American Linguistics 62. 70-85. Pater, Joe. 1999. Austronesian nasal substitution and other NC effects. The Prosody Morphology Interface, ed. by R. Kager, H. van der Hulst, and W. Zonneveld. Cambridge: Cambridge University Press. [ROA 160]. Pater, Joe. 2000. Austronesian nasal substitution revisited: What’s wrong with *NC (and what is not). Ms, University of Alberta and University of Massachusetts, Amherst. Phelps, Elaine. 1975. Sanskrit diaspirates. Linguistic Inquiry, 6. 447-464. Pierrehumbert, Janet. 1993. Dissimilarity in the Arabic verbal roots. In Proceedings of the North East Linguistic Society, 23, ed. by A.J. Schafer. Amherst: GLSA. 368-381. Pin)eros, C.E. 2000. Surface identity and ternary scales in Spanish voicing assimilation. ROA370. Prince, Alan. 1983. Relating to the grid. Linguistic Inquiry, 14. 19-100. Prince, Alan. 1997. Endogenous constraints on Optimality Theory. Talk presented at Hopkins Optimality Theory Workshop/Maryland Mayfest, Baltimore. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Ms., Rutgers University, New Brunswick, and University of Colorado, Boulder.

222

Struijke

Raimy, Eric and William Idsardi. 1997. A minimalist approach to reduplication in Optimality Theory. Proceedings of the North East Linguistics Society 27, ed. by K. Kusumoto. Amherst, GLSA. Rodier, Dominique. 1989. Prosodic affixation in Kwakiutl. Proceedings of the Eastern States Conference on Linguistics 6. 218-228. Rose, Sharon. 1993. Palatalization, underspecification, and plane conflation in Chaha. Proceedings of WCCFL, 12. 101-116. Rose, Sharon. 2000a. Multiple correspondence in reduplication. Proceedings of BLS. 315-326 Rose, Sharon and Rachel Walker. 2000. Long-distance consonant agreement. Talk presented at NELS. Georgetown University, Washington D.C. Rowe, John Howland. 1950. Sound Patterns in three Inca dialects. International Journal of American Linguistics, 16. 137-148. Sag, Ivan. 1974. The Grassmann’s Law ordering pseudo-paradox. Linguistic Inquiry, 5. 591-607 Sag, Ivan. 1976. Pseudo-solutions to the pseudo-paradox: Sanskrit diaspirates revisited. Linguistics Inquiry, 7. 609-622. Samek-Lodovici, Vieri. 1992. Universal constraints and morphological gemination: a crosslinguistic study. Ms, Brandeis University, Waltham. Schane, Sanford. 1968. French Phonology and Morphology. Cambridge, Mass.: The MIT Press. Schane, Sanford. 1984. The fundamentals of particle phonology. Phonology Yearbook 1. 129-155. Schindler, Jochem. 1976. Diachronic and synchronic remarks on Bartholomae’s and Grassmann’s Laws. Linguistic Inquiry, 7. 622-637. Selkirk, Elisabeth. 1995. Surface restrictions in the distribution of lexical contrasts: the role of Root Faithfulness. Handout Ling751, University of Massachusetts, Amherst. Shaw, Patricia. 1987. Non-conservation of melodic structure in reduplication. CLS 23, Parasession on Autosegmental and Metrical Phonology, ed. by A. Bosch, B. Need, and E. Schiller. 291-306. Smolensky, Paul. 1993. Harmony, markedness, and phonological activity. Paper presented at Rutgers Optimality Workshop I. Rutgers University, New Brunswick. [ROA 87]. Smolensky, Paul. 1995. On the internal structure of the constraint component Con of UG. Talk presented at UCLA. Spaelti, Philip. 1997. Dimensions of Variation in Multi-Pattern Reduplication. Ph.D. Dissertation University of California, Santa Cruz. Stahlke, Herbert F.W. 1976. Segment sequences and segmental fusion. Studies in African Linguistics, 7. 41-63. Stallcup, Kenneth L. 1980a. A brief account of nominal prefixes and vowel harmony in Esimbi. L’expansion bantoue, vol. 2. 435-411. Paris: Société d’Etudes Linguistiques et Anthropologiques de France. Stallcup, Kenneth L. 1980b. Noun classes in Esimbi. Noun classes in the Grassfields Bantu Borderland (Southern California Occasional Papers in Linguistics, 8). 139-153. University of Southern California, Los Angeles.

223

Existential Faithfulness

Steriade, Donca. 1988. Reduplication and syllable transfer in Sanskrit and elsewhere. Phonology 5: 73-155. Steriade, Donca. 1997. Phonetics in phonology: The case of laryngeal neutralization. Ms., University of California, Los Angeles. Struijke, Caro. 1997. Input-Output Correspondence and Kwakwala reduplication. Papers from the 1997 Mid-America Linguistics Conference, ed. by X. Li, L. Lopez and T. Stroik. 108-119. Struijke, Caro. 1998. Reduplicant and Output TETU in Kwakwala. University of Maryland Working Papers in Linguistics, 7. Papers in Phonology, ed. by H. Fukazawa, F. Morelli, C. Struijke, Y. Su. 150-178. [ROA 261]. Struijke, Caro. 2000a. Why constraint conflict can disappear in reduplication. Proceedings of the North East Linguistics Society 30, ed. by M. Hirotani. Amherst: GLSA. Struijke, Caro. 2000b. Feature ‘Movement’ and Segment-based Feature Correspondence. Talk presented at Massachusetts Institute of Technology, Cambridge Mass. Struijke, Caro. 2000c. Feature Movement as splitting. Talk presented at HUMDRUM, Rutgers University, New Brunswick. Struijke, Caro and Paul de Lacy. 2000. Overkill in Dissimilation. Talk presented at NELS. Georgetown University, Washington DC. Suzuki, Keiichiro. 1998. A typological investigation of dissimilation. Ph.D. dissertation University of Arizona, Tucson. Thráinsson, Hoskuldur. 1978. On the Phonology of Icelandic Preaspiration. Nordic Journal of Linguistics, 1. 3-54. Trubetzkoy, N.S. 1939. Grundzüge der Phonologie. Travaux du Cercle Linguistic de Prague, 7 (Principles of Phonology, translated by C.A.M. Baltaxe. Berkeley and Los Angeles: University of California Press. 1969). Urbanczyk, Suzanne. 1995. Double reduplications in parallel. University of Massachusetts Occasional Papers in Linguistics 18: Papers in Optimality Theory, ed. by J. Beckman, L. Walsh Dickey and S. Urbanczyk. 499-532. Amherst: GLSA. Urbanczyk, Suzanne. 1996. Patterns of Reduplication in Lushootseed. Doctoral dissertation University of Massachusetts, Amherst. Urbanczyk, Suzanne. 1999. Why stressed shwa is marked. Paper presented at the annual LSA meeting. Los Angeles. Von Soden, Wolfram. 1969. Grundriss der Akkadischen Grammatik. Analecta orentalia 33.2. Rome: Pontificum Institutum Bilblicum. Walker, Rachel. 1997. Faith and markedness in Esimbi feature transfer. Phonology at Santa Cruz, ed. by R. Walker, M. Katayama and D. Karvonen. 103-115. [ROA 336]. Walker, Rachel. 2000. Nasal reduplication in Mbe affixation. Phonology, 17. 65-115. Walker, Rachel. (to appear). Long-distance consonantal identity effects. Proceedings of WCCFL 19. Walsh Dickey, Laura. 1997. The phonology of liquids. Ph.D. dissertation, University of Massachusetts, Amherst.

224

Struijke

Whitney, William Dwight. 1885. The Roots, Verb-Forms and Primary Derivatives of the Sanskrit Language. Republished by American Oriental Society, 1945. Whitney, William Dwight. 1889. Sanskrit Grammar. Republished by Harvard University Press, 1960. Wilbur, Ronnie. 1973. The Phonology of reduplication. Ph.D. dissertation, University of Rochester. Williams, Edwin. 1976. Underlying tone in Margi and Igbo. Linguistic Inquiry, 7. 463-484. Wilson, Peter J. 1978. Syllable structure in Kwakwala and Heilstuk. Proceedings of the XIII International Conference on Salishan Languages. 275-299. Yip, Moira. 1989. Feature geometry and cooccurrence restrictions. Phonology, 6. 349375. Yip, Moira. 1992. Sonorant vs. obstruent codas: A prosodic distinction. Ms., University of California, Irvine. Yip, Moira. 1995. Repetition and its avoidance: the case of Javanese. Ms, University of California, Irvine. [ROA 83]. Yip, Moira. 1996. Feet, tonal reduction and speech rate at the word and phrase level in Chinese. Paper presented at the Utrecht Conference on Phrasal Phonology. [ROA 159]. Yip, Moira. 1998a. The role of markedness in onset change. Proceedings of ‘On the Formal Way to Chinese Languages’, ed. by C.S. Liu and S.W. Tang. Stanford: C.S.L.I. Yip, Moira. 1998b. Identity avoidance in phonology and morphology. Morphology and its relation to syntax and phonology, ed. by S. Lapointe, D. Brentari and P. Farrell. Stanford: C.S.L.I. 216-246. Yip, Moira. 2000. Segmental unmarkedness versus input preservation in reduplication. Segmental Phonology in Optimality Theory, ed. by L. Lombardi. Cambridge: Cambridge University Press. [ROA 378] Zec, Draga. 1988. Sonority Constraints on Prosodic Structure. Ph.D. Dissertation Stanford University, Palo Alto. (Published by Garland, New York) Zoll, Cheryl. 1996. Parsing below the Segment in a Constraint Based Framework. Ph.D. dissertation University of California, Berkeley. Zoll, Cheryl. 1998. Positional asymmetries and licensing. Paper presented at the Annual Meeting of the LSA, New York. Zwicky, Arnold. 1965. Topics in Sanskrit phonology. Ph.D. dissertation MIT, Cambridge Mass.

225