The Internal Organization of Phonological Segments 3110182955, 9783110182958

This book contains a number of studies on modern approaches to phonological segment structure. There are three main sect

234 72 13MB

English Pages 366 [372] Year 2005

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

The Internal Organization of Phonological Segments
 3110182955, 9783110182958

Table of contents :
Contents
Phonological alphabets and the structure of the segment
Part 1: Features and feature geometry
Optimal geometries
Variability in feature affiliations through violable constraints: The case of [lateral]
The geometry of harmony
Piro affricates: Phonological edge effects and phonetic anti-edge effects
On the internal and external organization of sign language segments: some modality-specific properties
Part 2: Nasality
On the ambiguous segmental status of nasals in homorganic NC sequences
Areal and phonotactic distribution of η
Cryptosonorant phonology in Galice Athabaskan
Part 3: Laryngeal features
On the phonological interpretation of aspirated nasals
The representation of the three-way laryngeal contrast in Korean consonants
Diachronic evidence in segmental phonology: the case of obstruent laryngeal specifications
Language index
Author index
Subject index

Citation preview

The Internal Organization of Phonological Segments

W G DE

Studies in Generative Grammar 77

Editors

Harry van der Hulst Jan Köster Henk van Riemsdijk

Mouton de Gruyter Berlin · New York

The Internal Organization of Phonological Segments Edited by

Marc van Oostendorp Jeroen van de Weijer

Mouton de Gruyter Berlin · New York

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter G m b H & Co. KG, Berlin.

The series Studies in Generative Grammar was formerly published by Foris Publications Holland.

© Printed on acid-free paper which falls within the guidelines of the ANSI to ensure permanence and durability.

Bibliographic information published by Die Deutsche

Bibliothek

Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the Internet at .

ISBN 3-11-018295-5 © Copyright 2005 by Walter de Gruyter G m b H & Co. KG, D-10785 Berlin. All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Cover design: Christopher Schneider, Berlin. Printed in Germany.

Contents

Phonological alphabets and the structure of the segment Marc van Oostendorp and Jeroen van de Weijer

1

Part 1: Features and feature geometry Optimal geometries Christian Uffmann Variability in feature affiliations through violable constraints: the case of [lateral] Moira Yip The geometry of harmony Don Salting

27

63

93

Piro affricates: Phonological edge effects and phonetic anti-edge effects Yen-Hwei Lin

121

On the internal and external organization of sign language segments: Some modality-specific properties Els van der Kooij and Harry van der Hulst

153

Part 2: Nasality On the ambiguous segmental status of nasals in homorganic NC sequences Laura J. Downing

183

VI

Areal and phonotactic distribution of η Gregory D.S. Anderson

217

Cryptosonorant phonology in Galice Athabaskan Siri G. Tuttle

235

Part 3: Laryngeal features On the phonological interpretation of aspirated nasals Bert Botma

255

The representation of the three-way laryngeal contrast in Korean consonants Hyunsoon Kim

287

Diachronic evidence in segmental phonology: the case of obstruent laryngeal specifications Patrick Honeybone

317

Language index

353

Author index

358

Subject index

364

Phonological alphabets and the structure of the segment Marc van Oostendorp and Jeroen van de Weijer

1. Introduction Like every living field, phonology is falling apart. Researchers are becoming increasingly specialized, and some phonologists have come to concentrate on particular subfields which they call their own - tonology, vowel harmony, intonation studies, metrical structure, the interface with morphology, or the internal structure of phonological segments. This concentration on subfields seems to have replaced the more traditional specialisation in terms of languages or language families. In addition, of course, the 1990s saw the advent ofOptimality Theory (OT), which all but hid the differences between the various subfields under the blanket of a common notation, by focusing on the issue of how to do phonology rather than on what the primitives of phonology are. This does not mean, however, that these different subfields ceased to exist; as a matter of fact, OT seems to have benefitted certain subfields more than others. Given the centrality of the notion of constraint ranking, OT has proven most successful in domains in which there is some sort of inherent conflict. This may then be one of the reasons why issues concerning the interface between phonology and morphology (such as Prosodic Morphology) or with the interface between phonology and phonetics have been quite successful over the past decade. Considerably less effort has been invested in the study of, for instance, metrical structure, which seems more 'purely' phonological and less influenced by various counterbalancing factors. It is certainly possible to successfully analyze stress phenomena within OT, and various important approaches to these phenomena have appeared, but in many cases there does not seem to be much gain in doing so. Something similar holds for the study of the internal structure of phonological segments, which is one of the classical topics of phonological investigation. The 1980s witnessed a steady interest in topics such as autosegmental phonology and feature geometry, but it seems fair to say that these have not been in the focus of mainstream OT research until very recently. Part of the reason for this may be that the interest in 'small-scale' phonology now is on

2

Marc van Oostendorp and Jeroen van de Weijer

those aspects which can be understood in interaction with phonetics. The study of objects like 'the phonological segment' is an enterprise which is to some extent abstract, since the segment cannot be isolated directly within the phonetic signal or described in purely articulatory or acoustic terms. Most authors assume that the constituent parts of segments are abstract as well for instance, in terms of phonological features. Another reason why the phonological segment has not received a lot of attention in recent work may be related to the fact that OT is mainly a theory of phonological alternations and linguistic variation, not a theory of phonological representation or linguistic universals. OT itself does not impose any restrictions on possible phonological representations - we can combine OT either with abstract Feature Geometry, or with concrete phonetic specifications, or with both. Neither does OT impose any restrictions on the analyst as to the constraints that must be postulated in analysis. It therefore is not a complete theory of phonology: it should be complemented by a theory of phonological primitives, and of a theory of phonological constraints. The articles in this volume - consisting of selected papers presented at the first Old-World Conference on Phonology (OCP1), held in Leiden on January 10-12, 20031 - show that there are still many interesting questions to be asked on segmental structure, that there is quite a lively debate on many of the issues concerned, and that the field is far from monolithic in its methodological approach: some authors use OT as a tool, but others do not; some refer explicitly to the results of phonetics for phonological explanation, while others prefer a purely abstract, cognitive approach. Furthermore, the reader will find contributions from neighbouring disciplines such as language typology and historical linguistics. The articles study topical questions within this particular field from various angles: to what extent do we still need a feature geometry, and to what extent is it universal? What is the relevance of evidence from historical linguistics, typology, etc.? How should we represent the 'complexity' of'complex' segments? In this introduction, we concentrate on the formal theoretical implications of the contributions made in this volume: in this way we hope to shed light on the nature of the 'complementary' theories to Optimality Theory, i.e. the basic requirements of a theory of phonological representations and universals.

1

Only a selection of the papers presented at the Conference have been included in this volume. Furthermore, all of the papers were reviewed and revised before they were included in this book.

Phonological alphabets and the structure of the segment

3

On the assumption that it is indeed possible and desirable to construct something like a 'theory of segmental phonology', we must establish what the constituent parts of such a theory are. In syntax, the 1990s saw a strong interest in the so-called 'Minimalist Program' of Chomsky (1995). This program has not been applied to phonology, but part of the 'program' is that the metagrammar of linguistic derivations and representations can be expressed by elements of (i) 'conceptual necessity' on the one hand, and (ii) 'conditions of the interfaces' ('articulatory-perceptual' and 'conceptual-intentional') on the other (Chomsky, 1995, p. 171). This part of the program is kindred in spirit to Occam's Razor; it is a healthy exercise to check the status of every element of the formal apparatus that constitutes our theory. Which objects are part of the phonological universe and which are not? Every article in this book contributes in some way to an answer to this question. This book consists of three parts. The first part, Features and feature geometry is the most general, and deals with some of the general issues we touched upon above. The second and the third part look more closely into two phenomena which warrant further discussion: Nasality in the second part, and Laiyngeal features in the third. We will introduce each of these parts in turn.

2. Features and feature geometry The first part of this volume is devoted to general issues, and the most general issue of all within the study of segmental structure is perhaps the relation between phonology and phonetics. We propose to look at this issue from the point of view of formal theory-building: the question then is whether the interpretation of the primitives of this theory is purely cognitive, or phonetic. The first two articles in this section deal with the relation between the theories of Feature Geometry and OT; the authors reach conclusions which are diametrically opposed with respect to the question whether such a combination is feasible. The next three articles are concerned with the interpretation of the features themselves, as well as with the issue of temporal ordering within the segment.

2.1. Constraints based on Feature Geometry In his article 'Optimal geometries' (p. 27-62), Christian Uffmann argues that OT and Feature Geometry can indeed be combined. In his view, feature ge-

4

Marc van Oostendorp and Jeroen van de Weijer

ometry can act as a filter on the generator function of OT. This means that constraints on segmental phonology only take into account those representations that are feature-geometrically well-formed. Moreover, constraints may directly inspect the result of autosegmental operations, such as spreading of a Place node from one segment to one or more other segments, and calculate the number of violations for a particular constraint accordingly. As far as we can see, the gist of the argument in favour of this proposal is theoretical restrictiveness. The oldest criticism against OT undoubtedly is that it does not provide us with a theory of what constitutes a possible constraint. It thus needs to be complemented with such a theory, which can be understood as a formal language and thus consists of the following two elements: i. an alphabet, i.e. an exhaustive list of symbols which are allowed within a constraint; 2 ii. a syntax, i.e. a set of rules that specify how these symbols can be combined to form well-formed constraints. By way of illustration, here is the formal language to describe one well-known family of constraints outside the realm of segmental phonology, Alignment, where the syntax consists of one statement (a rewrite rule, where C is the symbol for a constraint): (1)

a.

Constraint lexicon for Alignment constraints {'V', ' 3 ' , 'left', 'right', 'edge of', 'μ', ' σ ' , 'Ft', 'ω', 'stem',...} b. Constraint syntax for Alignment constraints C C -> V + (leftVright) + edge of + (μ V σ V Ft VwV stem V...) + 3 + (leftVright) + 'edge of' + (μ V σ V Ft VwV stem V...)

The grammaticality of (lb) is connected at least partly to the semantics of the elements in the alphabet: the fact that e.g. 'leftVedge of right μ' is not a possible constraint is strongly related to the fact that no sensible interpretation can be assigned to this statement (McCarthy & Prince, 1995). Similarly, Feature Geometry provides us with an alphabet on which constraints can be defined:

2

The term phonological alphabet was coined by Calabrese (1988); we extend its use here to refer to all primitives of (segmental) phonology.

Phonological alphabets and the structure of the segment (2)

5

a.

Constraint lexicon for Feature Geometric constraints {[+voice], [Coronal], [+consonantal], ..., Place, Laryngeal, ..., 'is associated to', 'branches',...} b. Constraint syntax for Feature Geometric constraints C —> V (SupralaryngealVLaryngeal) + 'χ: χ is associated to the root node'

It also provides us with a semantics - the interpretation of feature tree structures - on which we could build a syntax for constraints on the internal structure of segments. In this way, we constrain the set of constraints. It should be noted that there are other ways of arriving at the same effect, and in particular, we could also constrain the theory by deriving our alphabet from phonetics and an appropriate syntax of constraints (which is related to an interpretation - in this case, based on articulatory and perceptual reality). (3)

Constraint alphabet for phonetically grounded constraints {F0, F p ..., 1, 2, 3, 4, 5,..., Hz, ms, ..., 'should be bigger than', 'should be smaller than' ...,}

Even in this case it should be established, however, what the relevant vocabulary items are. This approach does not provide a restrictive theory if everything which is measurable in principle can participate in constraint phrasing. The formalisation and the restrictions are usually left implicit, but essentially every researcher will tacitly assume some restrictions; for instance, an analysis which is based on an alphabet consisting of the features [high], [low], and [open] as well as F values within the same constraint set would probably be unacceptable to everybody. A well-known example of a fairly restricted and formalised theory of phonology based in part on a phonetic alphabet is Archangeli & Pulleyblank (1994). Within an approach such as the one defended by Uffmann, Feature Geometry restricts the notion of what constitutes a possible constraint; at the same time it also gives us a set of possible configurations and results of operations. For instance, while spreading of a Place node could result in a segment acquiring the place feature [Coronal] (and hence an increase or decrease of the number of violations of a constraint involving this feature), spreading of a Laryngeal node by itself could never have this result. It is Uffmann's goal to show that his restricted theory of constraints can analyse various phenomena elegantly. His constraint syntax could be formulated as follows:

6 (4)

Marc van Oostendorp and Jeroen van de Weijer Constraint syntax for Feature Geometric constraints C —> V + {segment, Place, Laryngeal, [coronal], [voice] ...} + χ + 3 {μ, segment, Place, Laryngeal,...} + y : + χ is parsed into y C —» V + {Foot, segment, Place, Laryngeal, ...} + χ : + χ is binary branching

In parallel to this, there should be a theory of Gen which uses partly the same alphabet, i.e. which generates structures that can be referred to using terms such as 'segment' and 'Place' and [+voice].

2.2. Phonetically grounded constraints Moira Yip's article, 'Variability in feature affiliations through violable constraints: The case of [lateral]' (p. 63-92), on the other hand, argues that Feature Geometry is too rigid to understand the phonological behaviour of certain features, in particular of [lateral]. Her point is that [lateral] behaves ambiguously with respect to the Feature Geometry: in some cases, its behaviour seems to indicate that it depends on [Coronal], but in others that it depends on the Sonorant Voicing node. Feature Geometry predicts that these cannot both be true at the same time; therefore, Yip's data can be seen as evidence that Feature Geometry is falsified if taken as a universally valid tree structure of segments. The conclusions which Yip draws from this - we should abandon almost all structure in favour of phonetics 3 - are quite strong. In line with recent work (by Gafos, 1996, among others) she proposes that we accept a completely different interpretation of constraints, viz. one which is based on phonetics. This leaves her with an alphabet of constraints which is in some ways

3

Notice that Yip does not do away with tree structure below the segment altogether. She seems to assume autosegmental spreading, for which at least some rudimentary structure of the following form is necessary: (1)

[coronal]

[lateral]

[sonorant]

etc.

This structure (where χ is the segmental spine or root node) formally is a tree, albeit a very rudimentary one.

Phonological alphabets and the structure of the segment

7

simpler than the one proposed by U f f m a n n (it does not contain vocabulary items such as Place, Laryngeal, and the like), but may be more complex in other ways (e.g. it may need a larger number of constraints). Furthermore, we need an extra device, viz. the universal ranking of constraints, given to us by the phonetics, but still adding to the formal complexity of our machinery. 4 (5)

a.

Constraint syntax for phonetically grounded constraints (Yip-style) C —> V + {[coronal],[continuant],[+sonorant], [-sonorant],...} + x + V + {[coronal], [continuant], [+sonorant], [-sonorant],...} + y : + χ and y should not co-occur in the same segment C —> Segments should be faithful to their underlying specification for {[coronal], [continuant], [+sonorant], [-sonorant], ...}

b.

Ordering statement syntax for phonetically grounded constraint orderings (in the style of Yip) 9t —> Cι v( » Cn7 )*

Yip's problem might be solved in a different way, for instance by putting to work Jaye Padgett's idea of Feature Classes (Padgett, 2000). As she discusses on p. 86-88, this framework inherits many of the properties of Feature Geometry, but it is more flexible. Even though Padgett states a prohibition of overlapping class membership, there is nothing inherent to his approach that implies such a prohibition, hence constraints which would allow [lateral] to be a Place feature or a Sonorant Voicing feature, alternatingly or at the same time, are not formally impossible. We would thus slightly enlarge the set of possible constraints, but the change would not be a radical one. As a matter of fact, nothing in U f f m a n n ' s system precludes such an interpretation in principle either. 5 Yet Yip argues that "even Feature Classes can be dispensed with" (p. 87). If we are willing to accept phonetics as an explanatory force in pho-

4

It is not completely clear whether an abstract theory would be able to abandon 'universal' rankings completely, e.g. to be able to deal with sonority effects, as Yip notes. Yet in such a theory it would be very hard to find a motivation - or an interpretation - for such a device. 5 There is no constraint actively opposing affiliation of a feature to different organizing nodes, since these are assumed not to be generated by Gen (with a few exceptions such as place features that can be linked both to C-place and V-place, as in (Clements 6 Hume, 1995)). There is a constraint against multiple linking which could probably give us the right effect, if it is placed in an appropriate position in the hierarchy.

8

Marc van Oostendorp and Jeroen van de Weijer

nology, this might certainly be true. Scholars such as Uffmann and other contributors to this volume may not opt for this approach, however, so the matter is still open for debate.

2.3. Constraints based on abstract features Also with respect to the alphabet of phonological constraints, Don Salting's article 'The Geometry of Harmony: Evidence against Targeted Constraints', provides arguments in favour of a geometric approach to vowel harmony using abstract features. His arguments against a phonetically oriented approach (on p. 93-120) may sound familiar by now: "the theory becomes less constrained, and thus less explanatory" and "these constraints [...] suffer from [a] lack of of explanatory rigor [...]. If one can posit *[+rd, +lo], one can just as easily posit *[-rd, +lo]". Notice that these arguments hold only under the conception that phonetics should be irrelevant in the evaluation of constraints; authors such as Yip do not need to be convinced by them. Salting goes on to propose an approach to Feature Geometry which is indeed quite abstract; he notes the similarity to Van der Hulst's Radical CV Phonology (Van der Hulst, to appear), which might be the theory of segmental structure that is most remote from phonetics of all theories which are presently available: it is based entirely on the concept of binary opposition. Salting's own approach is slightly less radical, and uses the binary feature [±open] (Clements, 1991). The article therefore focuses on the nature of phonological features rather than on the concept of 'geometry'. The geometry is, as a matter of fact, fairly simple in Salting's proposal, partly because the abstract interpretation of the features involved do not necessitate a complex structure. Salting's proposals are more compatible with those of Uffmann than with those of Yip, but only by virtue of the fact that the interpretation of the phonological elements he proposes are more similar to the type of interpretations that Uffmann uses.

2.4. Using different alphabets and constraint sets Yen-Hwei Lin's article in this volume, 'Piro affricates: Phonological edge effects and phonetic anti-edge effects?' (p. 121-152) sheds interesting light on the debate on the abstractness of phonological representations, based on a thorough discussion of affricates in Piro. This paper is concerned neither with the geometry of segments (in the sense of the presence or absence of organiz-

Phonological alphabets and the structure of the segment

9

ing nodes) nor with the specific interpretation of features, but with the phonological implementation of temporal ordering. Like in other languages, affricates in Piro phonetically start with a stoplike ([+continuant]) part, and end with a fricative-like ([-continuant]) part. The question is whether this sequencing is reflected in their phonological representation: arguments for and against such a mirroring have been put forward in the literature, but Lin shows that in Piro we can find arguments in both directions. She therefore proposes that there is a split between a lexical and a postlexical component. Affricates behave as (strident) stops in the lexical component, but as segments with an ordered sequence of [-continuant, +continuant] postlexically. This corresponds to the postlexical phonology being more 'phonetic' in nature than the phonology, and Lin notes that this type of approach supports the approach of Stratal OT (Kiparsky, to appear), in which the grammar consists of more than one Optimality Theoretic module - at least one for the lexical phonology and one for the postlexical phonology. Notice that in the approach advocated by Lin the two modules would not just be differentiated by different rankings of the same constraints: the constraints would also address different phonological objects. Even though Lin leaves open the possibility that there are languages in which affricates behave as strident stops or as complex segments throughout the grammar, both in the lexicon and in the postlexicon, it seems unlikely that there would be languages in which the ordering of events would be reversed and affricates would start out as ordered [-continuant, +continuant] sequences in the lexicon, and turn into strident stops postlexically. This means that the lexical phonology has a constraint system which is based on more abstract representations, whereas the postlexical constraints are more phonetically based. This means that we can formulate the following restrictions on the constraint systems of the two components: (6)

Within the lexical component, segments are not allowed to have both a [+continuant] and a [-continuant] specification. In the postlexical component, segments are allowed to have such a specification.

(6) implies that a constraint such as O C P - A F F R (do not allow two segments with a specification [-cont, +cont] in a row) does not have a sensible interpretation within the lexical phonology: since these segments are not generated by Genlexjcal, no structure will ever violate this constraint. This opens up the possibility that this constraint is not present in the postlexical component at all, i.e. that the constraint grammars for the two components are different. We believe that this is a line of research that will generate further exploration.

10

Marc van Oostendorp and Jeroen van de Weijer

2.5. The phonological alphabet of sign language Within the debate on the abstractness of the phonological alphabet, sign language plays a crucial role. Under the assumption that phonology is an abstract system, there should be no difference between the formal apparatus of sign language phonology and that of spoken language phonology; the two constraint sets should have exactly the same alphabet and syntax (as a matter of fact, they may be the same in all relevant respects). If, on the other hand, phonology is phonetically grounded, the two sets of constraints could be very diverse. Sign language phonology has been the topic of quite some discussion over the past few years. In the present volume, it is the topic of the article by Els van der Kooij and Harry van der Hulst (p. 153-180): O n the internal and external organization of sign language segments: Some modality specific properties'. 6 As the subtitle suggest, the focus of these authors is more on the differences between the phonologies of different modalities than on the similarities. In particular, Van der Kooij and Van der Hulst argue that the sign in sign language corresponds to a phonological segment rather than to the phonological syllable, as other authors have suggested. This implies that in their view most morphemes of sign language (or at least of the Sign Language of the Netherlands) are monosegmental. These segments may to some extent be different from spoken language segments in the sense that they have some internal ordering: many signs involve a 'movement', e.g. from a relatively low position in front of the body to a relatively high position. This is intuitively very similar of course to the 'movement' in complex segments in spoken language, as noted by Van der Kooij and Van der Hulst, referring to Channon (2002). 7 Interestingly, these authors note that "nothing hinges on making what is essentially a terminological decision. The 'real' proposal is the structure [...], irrespective of how we label or call the nodes." (p. 153-180). Exactly the same point is made by Yip in her article (p. 63-92): "The terms mora, syllable, and foot, are just labels for levels in the hierarchy, and could equally well be stated in numerical terms. So really all we 6

Different from the articles just discussed, this paper is not cast in an Optimality Theory framework. This means that the relevant 'grammar' will not be necessarily of OT constraints, but of something else. Yet this grammar will still consist of an alphabet (listing the atoms) and a syntax (listing the ways to formalize sensible statements about the world in terms of the elements of the alphabet). 7 This similarity disappears altogether if complex segments can be shown not to exist in spoken language, e.g. if affricates are strident stops (cf. Lin, this volume) and prenasalized obstruents are bisegmental (cf. Downing, this volume).

Phonological alphabets and the structure of the segment

11

are saying is that level η is normally directly parsed into level n+1 [...]". This may be construed as a claim about phonological alphabets: (7)

Tree structures (relations between nodes) are part of phonological alphabets, but labels of nodes are not.

The statement in (7) is in fact also compatible with all other articles in this volume: no constraint ever seems to rely exclusively on the node label. This is a property that might then be shared between spoken language and sign language. Since tree structures are popular in syntactic and morphological research as well, we may tentatively generalize that these form the core of every alphabet of grammatical constraints. In the view of Van der Kooij and Van der Hulst there is an essential difference between sign language phonological alphabets and spoken language phonological alphabets: whereas in the latter linearization largely takes place at a level higher than the segment, in the former linearization is supposed to take place at a level lower than the segment. They relate this to a difference between the two modalities: "We submit, tentatively, that the reversed relation between syllable and segment is a result of the fact that in the visual channel perception is 'instantaneous', which then leaves little room for temporal effects. Conversely, we think that perception in the auditory channel proceeds in a predominantly temporal fashion, making horizontal, co-temporal divisions a secondary effect." The theory presented by Van der Kooij and Van der Hulst, then, is a mixed model: there is an abstract core of linguistic constraints (instantiated by the tree structures), but the contents of the linguistic structures and (presumably) the constraints, are determined also by the modality, i.e. by the requirements of the phonetic interface. Another interesting finding of Van der Kooij and Van der Hulst is that meaning may be part of the phonological alphabet: some phonological (or phonetic) constraints may need to refer to the meaning of the sign. A pair such as [open]-[closed] has a 'neutral' order ([open, closed]), and whenever this order is reversed ([closed, open]), this seems to be motivated by the meaning. But this in turn means that sign language phonology constraints need to be able to refer to semantic information, something which spoken language phonology is not usually assumed to be capable of (though cf. onomatopoeia, sound symbolism, etc.). Again, this seems to be an indication that the structure of phonological constraints seems to be related to the phonetics after all.8 8

There is one exception, viz. if we dismiss all of these phenomena as being purely phonetic.

12

Marc van Oostendorp and Jeroen van de Weijer

Thus, even though Van der Kooij and Van der Hulst start their contribution by stating "that [...] we believe that a consideration of certain conceptual issues regarding a potential common organization of languages in different modalities is [...] important", this does not cover all of their findings. In fact, their contribution is among the more phonetically oriented in this volume.

3. Nasality Next to general discussions of segmental structure, this volume also contains articles about two types of segmental modification that are of special interest: nasality and laryngeality. Both of these are interesting because they form a well-defined testing ground for issues concerning the phonological alphabet. One article, the one by Botma (p. 255-286, see also our discussion in section 4.1.) covers the interaction between these two dimensions of segmental structure.

3.1. The structure of prenasalised consonants One longstanding problem in the study of phonological segments is how homorganic nasal-obstruent (NC) sequences should be analysed. Two types of analysis are available in principle: these sequences are monosegmental ('prenasalised obstruents') or they are bisegmental ('clusters'). This topic is important for many reasons, one of them being the question of the internal structure of complex segments. Together with affricates (see the article by Lin on p. 121-152 and our discussion in section 2.4.), prenasalised consonants are the most famous examples of such putative complex segments. But while affricates can still be analysed as sequences of [-continuant,+continuant], this is more problematic for nasals. Whereas both [-continuant] and [+continuant] may be active in the phonology, it is very hard to find evidence that this holds for [-nasal] as well. Based on a wealth of data and detailed analysis, Laura Downing shows in her article 'On the ambiguous segmental status of nasals in homorganic NC sequences' (p. 183-216) that at least in Bantu languages these items should be analysed as bisegmental clusters. Downing differentiates between three different types of arguments for this reasoning: 'phonetic', 'phonemic' and 'phonological'. Phonetic arguments are those based on measurements - bisegmental units have approximately twice the length of monosegmental units; phonemic arguments are based on matters of complementary distribution - there is a difference between monosegmental and bisegmental units i f f we can find

Phonological alphabets and the structure of the segment

13

minimal pairs based on that difference. Downing argues that neither of these classes of arguments can provide us with the definitive answer in this case, and therefore we have to take recourse to phonological arguments, which show that the nasal part of the cluster is in the coda of a separate syllable. It is an interesting observation that phonetic and phonemic arguments in themselves do not provide us with sufficient evidence to decide between those two analyses, since these arguments presumably are the ones which are most easily available to the language acquiring child. The fact that subtle phonological argumentation is necessary in order to reach a decision may be seen as an indication that the issue is entirely dependent on the rest of our theory: we can only decide whether or not the first part of the cluster is 'moraic' if we have established independent criteria to decide whether or not a segment is moraic. In terms of the preceding discussion this means that the difference between monosegmental and bisegmental NC units cannot be expressed in the alphabet of phonetics ({1, 2, 3, ..., ms, ...}) or in the alphabet of phonemics ({'differs in meaning from', ...}). The issue can only be phrased sensibly within a reasonably abstract phonological alphabet (basically again one of tree geometry at the level of the syllable); and here the available evidence points in the direction of a bisegmental analysis for the Bantu languages. Downing leaves open the question whether there are other languages in which prenasalised consonants do behave as monosegmental also in the phonology. She mentions Fijian, referring to Maddieson & Ladefoged (1993), who found that in this language "NC sequences are not only similar in timing to singleton segments, they also pattern phonologically as unit segments". Yet, importantly, Maddieson & Ladefoged also found that "Fijian has no voiced stops, and NC sequences function as the voiced counterpart of voiceless stops". But this means that we may not need to analyse the Fijian 'prenasalised' segments as phonologically prenasalised at all; the 'prenasalisation' may be the phonetic reflex of phonological voicing. In that case, we might conclude the following: 9 (8)

9

The phonological alphabet should not provide for the possibility of monosegmental prenasalised segments.

Notice that this is of course a theory, and could be falsified, viz. by a language in which it can be shown that there are monosegmental nasalised segments.

14

Marc van Oostendorp and Jeroen van de Weijer

A theory which embraces monovalent [nasal] would be an example of a phonological alphabet satisfying (8), since it would not be able to express prenasalisation. Notice that this would have serious implications for a phonetically grounded view of segmental structure. Since phonetically it is indeed possible to express prenasalisation - since it exists phonetically, and the contrast does not seem particularily difficult to produce or hear , the burden of proof is on those proposing such a theory - they need to show that (8) is false (or that there is some other reason why this potential contrast is not exploited).

3.2. The structure of the velar nasal The article by Gregory D.S. Anderson (p. 217-234), 'Areal and phonotactic distribution of /η/' does not present a phonological theory, but it can be used to test existing theories, given the broad range of facts it covers. For instance, Anderson points out that there is a well-known theory about English (and other Germanic languages) relating the fact that /q/ cannot occur at the beginning of a syllable (*qa) to the fact that this segment is historically and/or underlyingly a cluster /ng/, and nasal+obstruent clusters do not occur at the beginning of a syllable (coincidentally, this is an assumption on which Downing bases much of her argument). Anderson shows that there are languages for which there is no basis for assuming a historical cluster, while they still satisfy the same requirement. This in turn means that the ban on */q/ in English is not necessarily related to the history of the language. Apparently, our constraint system needs to be able to ban this segment in this position independently. Another interesting observation is that in some languages this restriction against initial velar nasals concerns only the first position of the word. This gives us various indications about the phonological alphabet. In the first place it shows, once again, that we need to be able to refer to (prosodic) structure. It also shows that we need so-called 'positional markedness' of some form: our linguistic constraints need to be able to refer to the markedness of initial positions in the word separately. As a matter of fact, this shows that we need a parameterized constraint of some form: (9)

C —> *[ o q, where ae {σ, ω,...}

The fact that constraint schemes seem to be used by almost every contributor to this volume - as well as by virtually all other work in phonology - is an

Phonological alphabets and the structure of the segment

15

indication that the constraint set is not random, but has some internal structure (this does not tell us of course whether the source of this structure is functional of cognitive). We may therefore posit the following generalisation: (10)

The rules of constraint syntax may contain variables.

There are at least two types of variables: (i) those which refer to features or equivalents {[coronals], [voice], ...}, (ii) those that refer to organizing nodes {segment, σ, F, ω, ...} and, depending on one's assumptions, also {Place, Laryngeal, ...}. On the other hand, at least the set of feature sets does not seem to be completely randomly organized. This is another conclusion we can draw from Anderson's article: it seems unlikely that it would be possible to gather similar material about restrictions on, for instance, coronal nasals or velar obstruents. There is something special about velar nasals and phonological theory should be able to express what this is. We are not aware of any theory that can give a satisfactory account of all of these data (cf. Van der Torre, 2003, for a theory that may come close).

3.3. Morphosyntactic features in phonology Siri Tuttle's article 'Cryptosonorant phonology in Galice Athabaskan' (p. 235252) describes the behaviour of two segments in one specific language. This article deals with the intricate relations between voicing, sonorancy and nasality, and in this sense it belongs to the next part of this volume (on laryngeal features) almost as much as to the one on nasality. The article also touches on another important topic: the way in which morphology and segmental phonology interact. Tuttle argues that a class of 'phonetically obstruent' segments can best be analysed as being 'phonologically sonorant'. The phonological theory which Tuttle uses is the Sonorant Voicing theory of Rice (1993), among others. Within this theory, there are two possible phonological representations of voicing: a feature [voice] and an organizing Sonorant Voicing (SV) node. The former is typical for voiced obstruents, the latter for sonorants. Sonorant features such as [nasal] dock onto the SV node, and this explains, among other things, why only sonorants are targets for nasal spreading in some languages. Adherents of an abstract view of phonology could argue that this is an argument in favour of their approach, since it is unclear phonetically why [d] should behave as a sonorant. Their opponents could point out that the pho-

16

Marc van Oostendorp and Jeroen van de Weijer

netic sources for Galice are not very clear, and, furthermore, that SV theory does not really explain why some obstruents behave like sonorants, whereas others do not. SV Theory merely provides us with a sufficiently rich phonological alphabet and constraint syntax to describe this state of affairs. But richness can also be seen as a lack of restrictiveness, and therefore does not necessarily count as a virtue (for theories). In any case, it is interesting that there are several examples of phonetic obstruents which behave as phonological sonorants, but as far as we are aware there are no examples of phonetic sonorants which behave phonetically as obstruents; and these would be excluded in SV Theory. Another interesting property of Tuttle's proposal is that it draws a connection between morphological and phonological features. In the first place, she assumes that the class of AuGNMENT-constraints should refer both to phonological features and to morphological features. This means that the constraint grammars for phonology and for morphology share at least one rule (the one in (lb)). Furthermore, she shows that morphological and phonological constraints can refer to the same object. She shows how a feature [nasal] can move to an initial position in the word due to phonological alignment of the nasal feature, while at the same time it has to be linked to its original (vocalic) position due to morphological alignment (which basically says that every phonological feature which is part of the specification of a given morpheme should occur in the slot that is assigned to that morpheme in the template). In order for this proposal to work, morphological and phonological features need to be linked in some way (as is illustrated in Tuttle's example (21), p. 245). We can do this in many ways, but the following may be one of them: (11)

Every element in the phonological alphabet can have an index denoting the morphological affiliation of the element involved.

4. Laryngeal features Laryngeal contrasts can in principle be produced on all phonological segments, but they play the most prominent role on consonants. The way in which laryngeal modifications influence the patterning of segments is the topic of the last three contributions to this book; they all argue in favour of a threeway distinction in the laryngeal dimension and thus seem to converge on one view, in spite of differences in the labelling of these three elements of the alphabet, and their interpretation.

Phonological alphabets and the structure of the segment

17

4.1. Typological restrictions on phonetic possibilities The article by Bert Botma, 'On the phonological interpretation of aspirated nasals' (p. 255-286), investigates how the behaviour of aspirated nasals in a variety of languages can inform us about the phonology of laryngeality. In a survey based on the UPSID Database (Maddieson, 1984; Ladefoged & Maddieson, 1996), Botma shows that languages allow for at most a threefold laryngeal contrast on nasals. A fourfold phonetic contrast - voiced, voiceless (apirated), laryngealised or breathy voice - is possible, and this phonetic space is indeed explored (every possibility is used in some language), but every language uses at most three out of those four possibilities, and furthermore all languages with nasal consonants have voiced nasals. Botma argues that phonological theory has to be structured in such a way that it can account for these restrictions and he proposes a model to do this, combining insights of Element Theory (Harris & Lindsey, 1995) and Dependency Phonology (Anderson & Ewen, 1987). In line with some other work (see, for instance the article by Van der Kooij and Van der Hulst in this volume), Botma proposes a structure for (sub)syllabic constituents which is quite similar to what other people have proposed for segmental structure: (12)

Ο, N, C

Manner

Phonation

I Place (12) looks like a segment, except for the label of its root node, which is one of the subsyllabic nodes O(nset). N(ucleus) or C(oda). The feature geometry in its strictest sense is minimal (it involves Manner dominating Place, a standard assumption in Feature Geometry (see for instance McCarthy, 1988)). This conforms to the observations by Yip and by Van der Kooij and Van der Hulst in this volume, viz. that the labels of phonological trees might be irrelevant. Botma's main concern is with the structure of the constituent elements of the segment. In line with Element Theory and Dependency Phonology, he argues that these are not binary features, but unary elements. He suggests that there are three possible laryngeal ('phonation') elements, H, L and ?, which represent aspiration, voicing and glottalisation, respectively. Phonological glottalisation may be interpreted as either laryngealisation or breathy voice;

18

Marc van Oostendorp and Jeroen van de Weijer

this is the reason why these two interpretations cannot contrast phonologically in a language. The type of approach which is put forward by Botma aims at reducing the number of elements in the phonological alphabet. One way of achieving this is by interpreting these elements differently if they appear in different places in the segmental hierarchy, which therefore attains crucial status. In particular, the elements H, L and ? are also used to describe Manner (while other elements function to describe Place). That is, the research program that he adheres to can be summarised as follows: (13)

Keep the number of primitives in the phonological alphabet as small as possible.

The ultimate conclusion of this program might be that there are only one or two phonological primitives. This will then imply more complicated structures which are built out of those primitives (if we have 60 primitives to describe 60 distinctions, our structures can consist of 1 primitive each; if we have 2 primitives, we need to combine them in various ways to get the required number of distinctions). Botma seems to be able to keep these in reasonable balance: the structures are sufficiently simple and the number of primitives is sufficiently small. It should be noted also that 'as small as possible' is a relative notion. Some scholars who propose a much larger phonological alphabet might include many elements for which they see independent phonetic justification. In that case, they could argue that these elements do not really enlarge the alphabet of their theory, because their ambitions are to build one system for both phonetics and phonology; more abstract views of phonology then need a phonological and a phonetic alphabet which would be bigger, if unified. Furthermore, Botma adduces interesting evidence for this type of approach: since sonorant Manner and voicing Phonation are both represented by the element L, we can establish a direct relation between the two, and furthermore he speculates about similar direct relations if we represent 'nasality' as an L element as well, and both High tone and aspiration as an Η element. Notice that the system does not abstract away from the (acoustic and perceptual) phonetics, as H, L and ? are assumed to have at least some very general phonetic interpretation, and their phonological behaviour can therefore probably be understood at least in part by this very general phonetics.

Phonological alphabets and the structure of the segment

19

4.2. Phonetics and economy of phonological alphabets Hyunsoon Kim's article, 'The three-way laryngeal contrast in Korean' (p. 287-316) provides us with new data on a phenomenon which has been discussed extensively in previous literature, viz. the representation of laryngeal contrasts in the Korean obstruent system, where we find 'lenis', 'aspirated' and 'fortis' segments in a three-way contrast. Kim draws a distinction between two types of theory, which mainly differ in their representation of fortis consonants. While both theories agree that lenis segments are unmarked - do not have a laryngeal feature or only have unmarked values for those features one theory states that fortis consonants are bisegmental, whereas the other theory proposes that fortis segments have a marked feature specification ([+tense]). Kim shows that the latter theory is more successful in describing her phonetic, instrumental data. In some ways, her analysis is reminiscent of what Botma proposes for the laryngeal modifications of nasals, and one could wonder whether the same three elements that he proposes, H, L and ?, could be put to use. However, Kim argues that her data also show that binaiy features are necessary, for instance that we need a binary feature [±tense] in order to describe the facts she discusses. The relevance of Kim's findings for 'purely' phonological research can only be established once we have established a satisfactory notion of'economy' for phonological alphabets. Every theorist wants her theory to be as parsimonious as possible. In terms of our preceding discussion, this means that the alphabet should not contain any superfluous elements, and the constraint syntax should also be as concise as possible. If we now assume that Kim's facts and the more purely phonological processes of Neutralisation etc. have to be described within the same theory, the conclusion must be that unary features are unwanted, and so is a description of fortis consonants in terms of length, since the alternative descriptions are necessary independently. Unfortunately, measures of theoretical economy are never as simple as this. For instance, we cannot rule out in principle the possibility that we need an analysis of Korean fortis stops along the lines of Lin (this volume): one in which their representation in the phonology is slightly different than it is in the phonetics. While this will necessarily make the phonological alphabet more voluminous, it may simplify the set of necessary constraints. We then have to find an overall evaluation measure for alphabet and constraint syntax, but there is no a priori way to establish this. We simply do not have an objec-

20

Marc van Oostendorp and Jeroen van de Weijer

tive, purely theoretical way to determine which grammar would be the simplest. However this may be, Kim's facts show that those proposing a bisegmental and/or a bivalent feature analysis of Korean consonants need to reconsider their arguments. They will have to prove that their preferred theory would still be able to account for these facts in a satisfactory way, or otherwise recognize that they need to change their phonological alphabet.

4.3. Diachronic tests of phonological theories Patrick Honeybone's article, 'Diachronic evidence in segmental phonology: the case of obstruent laryngeal specifications' (p. 317-352) also discusses the relevance of a three-way distinction for phonological theory. The languages which Honeybone discusses, mainly German and English, merely display a binary contrast between two types of voicing. Within a ternary theory, this raises the question which two out of three possible values are selected. Honeybone's article is the only one in this volume which is mainly concerned with markedness: given a pair of phonemes which are written as and

, which should count as the most marked? Logically speaking there are two possibilities, and both seem to be explored in natural languages. In some languages - for instance Romance languages such as Spanish and French - it can be demonstrated that /b/ is the marked segment in the pair: the marked feature is [(+)voice], and in a unary feature account we could say that /p/ does not have any laryngeal feature at all. Based on diachronic data, Honeybone shows that in Germanic languages such as German and English, it is more appropriate to assume that /p/ is the marked segment: the marked feature is [(+)aspirated], and Ibl does not have any laryngeal specifications. Honeybone draws his arguments from diachronic evidence. He shows that both in (certain varieties of) German and in English at some point the 'voiced' and 'voiceless' sets became conflated, and that the new system was interpreted as 'voiced'. Under the assumption that the result of a sound change will be a less marked structure, this means that the 'voiced' (i.e. unaspirated) segments are unmarked. It is important that the two groups of languages are also distinguishable in terms of phonetics: French /p/ is closer to English Ibl than to English /p/. The transcription of IPA symbols within dashes is therefore rather confusing: it is distinct from the phonetic observations, but it also does not conform to the

Phonological alphabets and the structure of the segment

21

phonological reality, /p/ and fb/ are not part of the phonological alphabet, while [aspirated] and [voiced] are.10 The fact that the phonetics and the phonological behaviour converge is interesting, and it could lead us to suppose that the phonetics will be a cue for the language learning child to acquire the phonology. It does not necessarily mean, however, that phonetic information should be part of the phonological alphabet.

Conclusion One of the demands on phonological theory is that it should describe those aspects of sound structure which are universal. In order to be able to do this, we need to know what the formal language of phonology is, and which statements can be made in that language. Together, the articles in this volume contribute to an answer to these questions in the domain of segmental phonology. For instance, it seems clear that we need some form of arboreal structure, even though not everybody agrees on the question whether such structure also extends to the level below the segment. A general trend in the articles in this volume is to relocate the complexity of 'complex' segments to a higher level, so that prenasalised consonants are bisegmental and laryngeal modifications occur at the level of subsyllabic constituents. At the same time, we have seen that several authors declare the actual labeling of trees irrelevant, which blurs the distinction between subsegmental and suprasegmental structure even further. Within our theoretical vocabulary we need both tree structures and features (or elements); the argumentation about the precise nature of the tree structures and the elements will have to be detailed and subtle. We have also seen that some authors crucially use variables in the formulation of constraint: if a constraint can refer to one node, similar constraints should be able to refer to other nodes. Since a debate about the 'abstractness' of phonology seems to be inherent to the discussion of segmental structure, it is interesting that none of the contri-

10

Honeybone draws attention to the fact that there are different possible labels for these features, and that Botma and Kim make different choices among these. These differences may be connected to differences in phonetic interpretation and as such be responsible for different phonological behaviour; but formally it does not make a real difference what the labels of features or elements are (just as it does not make a difference whether we label a certain prosodic node as σ, as Si or as x).

T1

Marc van Oostendorp and Jeroen van de Weijer

butions in this volume abstract away completely from the phonetics, as some work advocates (Hale & Reiss, 2000). On the other hand, no author seems to take a purely phonetic point of view either. We believe that this reflects a common understanding in the field (which does not mean that extreme positions are not worth considering). We have also seen that some authors need to refer to entities which are outside the phonology (or the phonetics) proper, such as morphological structure, and even lexical semantics. The precise form of the phonological alphabet certainly has not been established yet. An interesting development within the phonology of the past decades is that considerable attention is being paid to evidence from neighbouring disciplines, such as experimental psycholinguistics, diachronic linguistics, language acquisition, etc. Theoretical phonology has clearly profited from the confrontation from these data, but it has its own (theoretical) agenda as well. This volume shows that the field is still very lively and full of promising developments.

References Anderson, John & Colin Ewen 1987 Principles of Dependency Phonology. Cambridge University Press. Archangeli, Diane & Douglas Pulleyblank 1994 Grounded Phonology. Cambridge, Mass: The MIT Press. Calabrese, Andrea 1988 Towards a Theory of Phonological Alphabets. Ph.D. thesis, MIT. Channon, R. 2002 Signs are Single Segments: Phonological Representations and Temporal Sequencing in ASL and Other Sign Languages. Ph.D. thesis, University of Maryland. Chomsky, Noam 1995 The Minimalist Program. Cambridge, Mass: The MIT Press. Clements, George N. 1991 'Vowel Height Assimilation in Bantu languages'. Working Papers of Cornell Phonetics Laboratoiy, 5: 37-74. Clements, George N. & Beth Hume 1995 'The Internal Organization of Speech Sounds'. In: Goldsmith, John (ed.), The Handbook of Phonological Theory, pp. 245-306. Cambridge: Blackwell Publishers. Gafos, Diamandis 1996 The Articulator)' Basis of Locality in Phonology. Ph.D. thesis, Johns Hopkins University.

Phonological alphabets and the structure of the segment

23

Hale, Mark & Charles Reiss 2000 'Phonology as cognition'. In: Burton Roberts, N., Philip Carr, & Gerard Docherty (eds.), Phonological Knowledge: Conceptual and Empirical Foundations, pp. 161-184. Oxford: Oxford University Press. Harris, John & G. Lindsey 1995 'The Elements of Phonological Representation'. In: Durand, Jacques & Francis Katamba (eds.), Frontiers of Phonology, pp. 34-79. London and New York: Longman. Hulst, Harry van der to appear 'Molecular Structure of Phonological Segments'. University of Connecticut, ms. Kiparsky, Paul to appear Paradigmatic Effects and Opacity. Stanford: CSLI. Ladefoged, Peter & Ian Maddieson 1996 The Sounds of the Worlds Languages. Oxford: Blackwell. Maddieson, Ian 1984 Patterns of Sounds. Cambridge: Cambridge University Press. Maddieson, Ian & Peter Ladefoged 1993 'Phonetics of Partially Nasal Consonants'. In: Huffman, Marie K. & Rena Krakow (eds.), Nasals, Nasalization and the Velum, pp. 251 -301. San Diego: Academic Press. McCarthy, John 1988 'Feature Geometry and Dependency: A Review'. Phonetica, 42: 84108.

McCarthy, John & Alan Prince 1995 'Generalized Alignment'. Morphology Yearbook. Padgett, Jaye 2000 'Feature Classes in Phonology'. Language, 89: 81-110. Rice, Keren 1993 Ά Reexamination of the Feature [sonorant]: Sonorant Obstruents.' Language, 69, 2: 308-344. Torre, Erik Jan van der 2003 Dutch Sonorants: The Role of Place of Articulation in Phonotactics. Ph.D. thesis, Leiden University.

Part 1 Features and feature geometry

Optimal geometries1 Christian Uffmann

1. Introduction Within Optimality Theory (Prince & Smolensky 1993, McCarthy & Prince 1993), there is no general consensus as to how segment interaction, most notably cases of assimilation and dissimilation, should be modeled. In this paper, I want to contribute to the ongoing debate by proposing a model of segment interaction in Optimality Theory (henceforth OT) which combines the general framework of OT with a representational component, that is feature geometry, in a principled way. More precisely, I am going to argue that feature geometry functions as a phonological primitive and a filter on GEN, the generator function of OT. Under this view, only geometrically well-formed candidates are evaluated, and the candidate which minimally violates a set of ranked constraints, which includes constraints on geometric structure, will be selected as optimal. The paper is organized as follows: In §2, an introduction to theories of segment interaction in OT and in pre-OT days will be provided, followed by a detailed presentation of the proposal and a discussion of the structural constraints that follow from such a proposal (§3). §4 contains a number of case studies which apply these proposals to actual cases of feature interaction. Here, I will discuss cases of assimilation (§4.1), long-distance assimilation (§4.2), spreading and epenthesis (§4.3 and §4.4) and dissimilation (§4.5). A brief outlook concludes the paper.

1

1 would like to thank participants of OCP-1 for comments and especially Eric Bakovic and Markus Hiller for fruitful discussion. Versions or parts of this work were also presented at GGS 2000, NAPhC 2, the Universities of Marburg and Düsseldorf and the University of Southern California. Thanks to audience members for helpful suggestions. Additional thank-yous go to Mary Pearce for her Kera expertise, to Birgit Alber for commenting on various stages of the development of the theory and to the editors and one anonymous reviewer for valuable comments on an earlier version of this paper. All remaining errors are mine.

28

Christian Uffmann

2. Theories of Segment Interaction This section briefly reviews current theories of segment interaction. First, I will discuss how segment interaction was described before OT, emphasizing the importance of autosegmental phonology and feature geometry in particular. Then, I will review how segment interaction has been modeled in OT so far. It will be shown that accounts of how to analyze instances of assimilation and dissimilation in OT are quite diverse, which demonstrates that we are still lacking a general theory of segment interaction within this framework. Moreover, principled statements about the role of autosegments and feature geometry in OT are rare.

2.1. Segment interaction before OT As a response to shortcomings of SPE-type rules especially in the areas of suprasegmental phenomena and non-local interactions, an autosegmental model of segment interaction was developed (Goldsmith 1976, 1990) where features are no longer exclusively associated with single segments but are treated as autosegments instead; they reside on individual tiers and can behave independently of the segment they are linked to. All possible interactions then result from two basic operations: spreading (i.e. insertion of an association line) and delinking (i.e. removal of an association line). Possible and impossible interactions are defined by adjacency conditions: interacting autosegments must be adjacent at least on their respective tier, and line crossing is prohibited. Originally designed to account for phenomena in tonal phonology, the model of autosegmental phonology was gradually extended, resulting in models of feature geometry (Clements 1985, Sagey 1986). Every feature is treated as an autosegment, and features are organized hierarchically into feature classes, accounting for the observation that features can spread as a group and defining which features can spread together, by positing class nodes which themselves function as autosegments. Within feature geometry, two general theoretical strands can be discerned, Articulator Theory (Sagey 1986, for a more recent revision, see Halle, Vaux & Wolfe 2000) and Unified Feature Theory (Clements 1991, Clements & Hume 1995), also known as the Clements/Hume model of feature geometry. In this model, consonantal and vocalic features are unified. Vocalic place features such as [back] and [round] are redefined in terms of more general (consonantal) place features such as [dorsal] and [labial]. A representation of this model can be found in (1). With some minor adjustments to be made in

Optimal geometries

29

due course, I shall assume this model of feature geometry in the remainder of this paper. Note, however, that the model I will propose does not crucially depend on any specific type of geometry, The analyses can be easily recast in terms of Articulator Theory, for example. Similarly, this paper is not making any definitive statements about the exact nature of the geometry. Future research will hopefully shed more light on the question of how a definitive geometry might look like.2 (1)

UFT model of Feature Geometry (after Clements/Hume 1995, Odden 1994) Root {[sonorant], [approximant], [vocalic]}

C-Place

2

See e.g. Yip (this volume) on the lack of consensus regarding the status of [lateral] in feature geometry. This lack of consensus does in no way weaken or invalidate the present argument, however, since it finds parallels in many domains of phonology. There is, for example, no agreement on which set of distinctive features exactly to assume, or which features are binary or unary. Similarly, there are still many open

30

Christian Uffmann

2.2. Segment interaction in O T While feature geometry had become the dominant model for analyses of segment interaction by the early 1990s, the appearance of O T changed this state of affairs considerably. In early OT, little was said about segment interaction, the focus returning to suprasegmental processes such as syllabification and stress assignment instead. 3 Consequently, a plethora of constraints and/or underlying assumptions were invoked in those (relatively few) papers that explicitly dealt with issues such as assimilation. For example, there are papers that use autosegmental representations, although it is not quite clear whether the representations are really seen as part of the phonological system or whether they are rather visualization aids for the reader without subscribing to the theoretical implications of the model, such as tier independence and the possibility of non-local spreading. Their relationship to other parts of the phonology, for example G E N or CON, the set of constraints, is not discussed (see e.g. Lombardi's 1999 analysis of voicing assimilation 4 which uses autosegmental notation at points without clarifying the status of autosegments within the theory). Other researchers use autosegments and feature geometry in a more general way, e.g. by evaluating candidates with their internal (geometric) structure, as in Beckman (1995, 1998) o r i n b o w i c z (1998), or by defining constraints over certain dispreferred geometric configurations, as in Noske's (1997) analysis of velar fricative assimilation in German. Although such papers include feature geometry into the O T framework in a more general way (by referring to it either in the formulation of constraints or in candi-

issues with respect to syllabification (e.g. ambisyllabicity, extrasyllabic segments, codas as onsets etc.). Nevertheless, features and syllables are widely accepted concepts fruitfully used in phonological analyses despite the lack of consensus or even definitive answers. 3 In the "canonical" texts (Prince & Smolensky 1993, McCarthy & Prince 1993), there are no analyses of assimilatory or dissimilatory processes. Prince & Smolensky deal mostly with issues of syllabification, while McCarthy & Prince discuss prosodic morphology. The absence of (sub-)segmental issues from these texts might in part be responsible for the conspicuous lack of consensus in the field with respect to how to view these issues in the light of OT. 4 The mentioning of individual papers is done for illustrative purposes only, not to single out individual researchers for particularly good or bad pieces of work. The papers cited should simply be seen as typical instances of the different approaches towards the topic.

Optimal geometries

31

date evaluation), its exact place within the general framework remains much of an open question. Parallel to this line of research, which retains some notion of autosegments and feature geometry, a different school emerged, however, which is strictly anti-representational. Instead of assuming autosegmental representations, surface-based constraints alone are believed to be sufficient to explain feature interactions. Under this view, autosegmental or feature geometric representations are superseded by OT, and representations can be replaced entirely by output constraints (see Yip this volume for an example discussing laterals). This school has been continuously gaining ground over the last years and is possibly the dominant paradigm today, in its incarnation of strict locality or locality theory (Ni Chiosäin & Padgett 1997, 2000, Gafos 1999; for applications of this theory see, for example, Walker 1998, Bakovic 2000, a m o n g others). In theories of strict locality, segments have no internal structure, there are no tiers, no class nodes, no hierarchies. Consequently, all spreading must be strictly local, as there are no possibilities to relativize the notion of adjacency. Cases which seem like instances of non-local spreading, e.g. vowel harmony, are not; intervening material always participates, although the effect may be only minor (phonetic). The main constraint on this concept of spreading is the "bottleneck effect", which states that vocalic features can be superimposed onto consonants and consequently propagate across them but not vice versa, thus ruling out consonant harmony while at the same time allowing vowel harmony. Finally, spreading is driven by ALIGN or AGREE constraints. While space does not permit to give a detailed critique of strict locality, a few points which might prove problematic for this theory should be addressed. First, it explicitly rules out cases of long-distance assimilation, which are attested (Odden 1994, Rose & Walker 2001; see also §4.2 below). 5 Similarly, tonal processes (one of the main arguments for autosegmental representations) might prove problematic for strict locality, as tone interaction is typically non-local but occurs between tone-bearing units (e.g. syllables, moras, vowels). Second, the bottleneck effect may be too strictly formulated, as spreading of consonantal place features on vowels is attested (see e.g. Hume 1996 for examples). Third, the use of AGREE constraints is somewhat prob-

5

Rose & Walker (2001) attempt to capture long-distance interactions as an effect of segmental correspondence, treating it formally differently from regular (strictly local) spreading. Future research will show how successfully this approach can be implemented.

32

Christian Uffmann

lematic. While AGREE is commonly formulated as a markedness constraint ("adjacent segments agree in F"), it is unclear how agreement is computed (or how agreement domains are set up); the only viable option seems to be by establishing and evaluating a correspondence relation between adjacent segments, since AGREE does not directly evaluate some phonological element (as a regular markedness constraint would) but a relation between elements. The constraint thus has an unclear status within the system, 6 aggravated by the lack of stringent formalizations of AGREE.7 Moreover, the assumption is questionable that assimilation is actively enforced by constraints and not the result of a drive towards featural economy (as in Beckman 1998). Spreading without AGREE will be further discussed in §4.1.

3. Feature Geometry in O T The previous section has shown two things. First, there is no generally accepted standard theory of segment interaction in OT. Second, recent proposals which attempt to delineate such a standard theory as an anti-representational theory of strict locality are not unproblematic either, but raise a number of empirical and formal questions. Therefore, I want to advance a different model in this section, a model which incorporates autosegments and feature geometry into the O T framework. At the core of my proposal lies the assumption that representations are not generally incompatible with O T but that much can be gained from their inclusion in O T instead. One could reverse the question and ask why representations should be abandoned in the first place, since they were a powerful tool in pre-OT days. Two arguments seem to speak against representations in OT. First, one could argue that constraints, which operate on surface strings, alone are sufficient to capture all phonological processes, and that the addition of further levels of representation unnecessarily complicates the machinery. Under this view, the inclusion of autosegments or feature geometry would be a violation of O c c a m ' s Razor, since they do not add any explanatory power to the 6

Krämer (2001) avoids this ambiguity in his theory of syntagmatic identity by formalizing agreement as correspondence. The outcome is identical. Then, however, Rose & Walker's (2001) dichotomy between local spreading and long-distance correspondence collapses. 7 But see McCarthy (to appear) for a more stringent definition couched in Comparative Markedness. His definition, however, also fails to provide a formal account of how agreement could be evaluated.

Optimal geometries

33

system. Second, one could argue against the abstractness introduced into the system by elaborate representations, which are not visible on the surface, and instead call for a maximal surface-orientation of phonology, where we simply find a linear string of segments. Both arguments do not stand up to critical scrutiny, however. The first argument could in fact be reversed. A strong representational component can simplify and streamline the set of constraints needed to describe segment interactions because feature geometry gives an accurate description of attested and unattested types of interaction. It correctly predicts which features or feature classes can interact over which distance (see e.g. Clements 1991, Odden 1994). Including feature geometry into OT can therefore remove a heavy burden from CON, because the geometry filters out impossible configurations in the first place (e.g. consonant harmony or interactions between non-adjacent autosegments). 8 In other words, the inclusion of feature geometry into OT can be a measure against the proliferation of constraints. The argument concerning Occam's Razor becomes void if inclusion of a second component into the phonological system can significantly simplify the first component. In addition, idiosyncratic filters on CON may also be lifted, simplifying the system further. For example, Rose & Walker (2001) limit long-distance correspondence to certain features, since other features simply do not interact over a distance. While they essentially have to stipulate which features can be targeted by correspondence constraints, feature geometry clearly predicts which features may spread over a distance (those for which only some segments are a legitimate anchor). The call for surface-orientation is also problematic, since hierarchical representations are widely accepted in the realm of prosody - the prosodic hierarchy has never been abandoned - or in OT syntax. Consequently, nothing should speak against re-introducing them into subsegmental contexts as well, especially if the same set of constraints is used for the evaluation of prosodic and segmental representations. In this paper, I therefore pursue the idea of such a unified set of structural constraints, aiming at a generalized geometry which comprises the macrocosm of prosody as well as the microcosm of

8

While most researchers do not discuss such candidates because they are clearly suboptimal, there is little evidence that they exclude them from the candidate set. Some actually include them, for example Padgett (1995), who assumes a NOGAP constraint to ensure locality in spreading, although this constraint must obviously be universally undominated. Rose & Walker (2001) explicitly allow interaction across a feature bearer, which is impossible under the model assumed here.

34

Christian Uffmann

subsegmental structure. Geometric structure is thus not just assumed to exist - it can also be targeted by constraints. Finally, the inclusion of autosegments and feature geometry enables us to develop further an idea which was first brought into the discussion by Beckman (1998), namely that assimilation is triggered merely by structural constraints, i.e. constraints on preferred types of feature geometric structure, but never by constraints explicitly enforcing assimilation. This obviates the need for constraint classes such as AGREE (e.g. Lombardi 1999, Bakovic 2000) or S-IDENT (Krämer 2001), further simplifying CON. I therefore propose to reintroduce feature geometry into O T in a most principled and strict way by viewing it as a primitive of Universal Grammar which functions as a filter on GEN. Only candidates which are well-formed with respect to the conditions imposed by feature geometry are generated and added to the candidate set (there are no candidates where association lines cross, no reorderings of the geometric tree, no illegal types of spreading etc.). 9 EVAL consequently evaluates only (geometrically) well-formed candidates. The burden thus lifted from candidate evaluation is twofold. First, the candidate set is streamlined by excluding impossible candidates, as operations performed by G E N adhere to the well-formedness conditions of feature geometry and autosegmental phonology. In addition, the constraint set is also streamlined because a large number of impossible or nonsensical candidates no longer need to be weeded out by the evaluator function. The inclusion of feature geometry into an O T framework simplifies candidate generation and evaluation, which alone improves the explanatory power of OT. A further boost comes from the set of constraints on geometric structure which will be introduced now.

9

While the possibility of filters on GEN has long been accepted, little work has been dedicated to this issue. The proposal to have feature geometry in GEN is not precluded by the theory at all. Prince & Smolensky (1993:5) note that "GEN contains information about the representational primitives and their universally irrevocable relations". Incidentally, Prince & Smolensky invoke feature geometric representations in their manuscript (e.g. §9.2), which suggests that they may themselves have viewed feature geometry as one such "representational primitive". Unfortunately, this line of research was not to be taken up again.

Optimal geometries

35

3 . 1 . PARSE m a t e r i a l

The first constraint to be introduced is a generalized version of the PARSE family of constraints. The constraint PARSE-SYLL, which demands that syllables be parsed into feet, is widely used in the OT literature (e.g. McCarthy & Prince 1993, Kager 1999, Alber 2002). This constraint can be generalized as follows: (2)

PARSE(Cat)

Material must be parsed into the next higher level of representation

This constraint holds at all levels of representation, for syllables to be parsed into feet as well as for features to be parsed into class nodes. For example: (3) (4) (5)

PARSE(O) PARSE(Seg) PARSE(F)

Syllables are parsed into feet Segments are parsed into syllables Features are parsed into segments

While this might at first glance look like a return to Containment, 10 it is not, because the function of PARSE is clearly different from the function of correspondence-based faithfulness constraints, and consequently it cannot replace them. PARSE(G) demands that syllables are footed; violation of PARSE((T), however, does not result in the deletion of a syllable but rather in the creation of an extrametrical syllable. Similarly, a violation of PARSE(Seg) creates an extrasyllabic segment. In prosody, the violability of PARSE thus allows weak layering. Below the segmental level, violation of PARSE can create floating material, which can, for example, opaquely interact with material on the surface."

10

In Containment Theory (Prince & Smolensky 1993, McCarthy & Prince 1993), the input is literally contained within the output. Epenthesis and deletion are thus explained as over- and underpaying of the input, not as the actual insertion or removal of material. Correspondence Theory (McCarthy & Prince 1995) has since superseded Containment. " The possibility of floating material in OT thus opens up a new perspective on the old issue of opacity, which cannot be pursued here further. In counterbleeding opacity, for example, non-application of a process can be explained by the blocking effect of an unparsed, yet present (i.e. floating) feature or segment. While this arguably adds a certain level of abstractness to OT, it also enhances its explanatory power. At any rate, the question whether the positing of floating material adds a higher degree

36 3.2.

Christian Uffmann N o MULTIPLE feature associations

The second constraint to be introduced militates against multiple feature associations. It demands that every feature is associated with one segment only. This constraint is called * M U L T I P L E here. 12 (6)

* MULTIPLE

Nodes are dominated exclusively by one other node

rules out feature spreading, because a feature is no longer uniquely associated with one segment if it spreads. * M U L T I P L E thus penalizes structures of the type in (7). *MULTIPLE

(7)

X

X [F]

Whether spreading can occur in a language, depends on the relative ranking of * M U L T I P L E and some markedness constraint M . This ranking decides whether the marked structure will appear in the output or whether it can be removed through spreading of a less marked feature. (8) (9)

* MULTIPLE » Μ Μ » *MULTIPLE

Spreading is worse than some marked structure A marked structure is worse than spreading

The constraint * M U L T I P L E allows for a straightforward formalization of the observation that some languages allow certain marked elements on the surface, while other languages avoid them by means of spreading a less marked feature. For example, the coronalization of dorsals before front (coronal) vowels 13 results from ranking *DORSAL above *MULTIPLE, which in turn outranks * C O R O N A L : Multiple linkage of [coronal] is less marked than retention of the

of abstractness to the model than, say, Sympathy Theory, two-level constraints or targeted constraints, remains open to (future) debate. 12 Variants of this constraint have been around for a while, and under different names, e.g. NOMULTIPLE (iubowicz 1 9 9 8 ) , UNIQUE (Beckman 1 9 9 8 ) , *MULT-LINK (Rubach 2000). The current proposal attempts to unify these constraints in a more principled way (cf. Bakovic's 2000 opinion that Beckman's use of UNIQUE is ad-hoc). 13 Examples include coronalization in Slovak, velar softening in English or dorsal fricative assimilation in German. For a detailed discussion of coronalization, see Hume (1992).

Optimal geometries

37

feature [dorsal] on the surface, and [dorsal] can be deleted via spreading of a less marked feature. 14 A different example of how low-ranked *MULTIPLE results in spreading rather than retention of a marked feature will be discussed in greater detail in §4.1.

3.3. Do not LINK different nodes A special case of multiple linkage is banned by the *LINK(X,Y) constraint. (10)

*LINK(X,Y)

Nodes attach to the same type of mother node

*LINK(X,Y) captures Hume's (1996) Mother Node Attraction principle, which states that nodes should not be linked to two different class nodes at the same time. It penalizes structures of the type (11)

Χ

Y

where X ^ Y

[F] Most importantly, it rules out feature sharing between C-Place nodes and VPlace nodes; a place feature like [coronal] or [labial] should not be linked to both a C-Place node and a V-Place node. Hence, the more specific constraint *LINK(C,V) militates against the structure (12)

C-Place X^JV-Place [F]

Structures such as in (12) are crosslinguistically marked; place feature sharing between consonants and vowels is relatively rare, although attested (e.g.

14

Admittedly, this is just a very brief analytical sketch. For instance, a complete analysis also needs a high-ranked constraint that ensures that not all dorsals are deleted from the surface, i.e. a constraint against the insertion of features (because [dorsal] would have to be replaced with something), with the effect that [dorsal] can only be deleted via spreading, not via insertion of a less marked feature which has no correspondent in the input.

38

Christian

Uffmann

in coronalization, as opposed to palatalization; see Clements & Hume 1995: 294ff, Hume 1996 for examples). Note that this constraint, while superficially similar to I to, Mester & Padgett's (1995) N O - X Y - L I N K class of constraints, is more stringently formalized, as it makes explicit reference to geometric structure, rather than just stating that association lines should not be established between somehow dissimilar segments. It also explains why place features are less easily shared between consonants and vowels than other features (without reference to feature geometry, such a constraint would be ad hoc, as there would be no structural reason against feature sharing). The importance and effect of the *LINK(X,Y) constraint will be demonstrated in §4.3, in a discussion of vowel epenthesis in loanwords.

3.4. Constraints on node branching Universally, two types of node branching are preferred on all levels of representation: Binary branching structures and non-branching structures. The two relevant constraints evaluating these structures are given in (13)-( 14): (13) (14)

BINARY(Cat) UNARY(Cat)

Mother nodes are binary branching Mother nodes are unary branching (=not branching)

Different categories universally prefer one of the two options, that is, for each category one of the two constraints is active. The function of these constraints is twofold. They first of all generalize prosodic markedness constraints. Compare (15)

FTBIN

vs.

BINARY ( φ )

(16)

ONSET

VS.

BINARY ( σ )

(17)

NOCODA

VS.

UNARY

(18)

*COMPONS

VS.

UNARY ( O n s )

(Rime)

The reformulation of well-established constraints in (15)-( 18) shows one advantage of the proposal: Constraints which are well-supported empirically but nevertheless formalized somewhat impressionistically can be reduced to one class of constraints operating on different levels of the prosodic hierarchy. However, the BINARY/UNARY pair of constraints also operates on the segmental level. A few examples will be provided now in brief sketches. The constraint UNARY(Seg), for instance, bans contour segments by demanding that a segment should comprise one root node only (see also Down-

Optimal geometries

39

ing this volume against prenasalized stops as contour segments). On the tonal level, a constraint like BINARY(H) demands high tone doubling, a process very common in Bantu languages (e.g. in Shona, Carter & Kahari 1972; Chichewa, Goldsmith 1990; Makonde, Odden 1995; Kerewe, Odden 1998). On the subsegmental level, this constraint might also operate on class nodes. With respect to the Laryngeal node, a constraint UNARY(Lar) can be posited which bans complex laryngeal specifications, preferring two-way laryngeal contrasts to more complex systems. In fact, many languages prefer a simple voicing or aspiration contrast to more complex systems where several laryngeal features interact (complex systems are only a subset of all systems which make contrastive use of laryngeal features), a markedness generalization which is neatly accounted for by UNARY(Lar).15 Similarly, UNARY(Aperture) bans complex vowel height systems, penalizing structures (following the Clements/ Hume model of feature geometry which allows multiple specifications for the feature [open]) like (19)

Aperture [op< ,, [open,]

By penalizing such structures, the generalization that 5-vowel systems (which need only one specification of [open]) are less marked than 7- or 9-vowel systems, is accounted for by the constraint UNARY(Aperture).16 It is well possible that similar markedness relations hold also for other nodes; an exploration of this possibility must be left to future research, however.

15

Van de Weijer & Hinskens (2003) note that if more laryngeal features are used contrastively by a language, all tend to be used. In other words: If UNARY(Lar) is violated, it does not matter how many laryngeal features are contrastively used, which lends further support to this constraint, which states that one feature only should be used contrastively but does not entail further constraints if it is violated. "Note that UNARY(Lar) and UNARY(Aperture) assume some version of contrastive underspecification, where noncontrastive feature specifications are not overtly marked on phonological representations. In languages which do not contrast aspirated and unaspirated stops, for example, these stops do not underlyingly bear specifications for the feature [spread glottis].

40

3.5.

Christian

Do not

SKIP

Uffmann

material

The last constraint to be introduced concerns adjacency in spreading. If an autosegmental model of feature interaction is assumed (as opposed to a strictly local model), then spreading need not be strictly local, i.e. root-adjacent. The only condition is tier-adjacency. Nevertheless, local spreading is generally preferred, which is captured by the constraint (20)

* SKIP

No material should intervene between interacting segments

*SKIP enforces adjacency in spreading; however, it needs to be parameterized, in line with Odden's (1994) observation that spreading is constrained by locality requirements. It is either strictly local (root-adjacent), occurs across one syllable boundary only (syllable-adjacent), or it is completely unbounded within the phonological word. Hence, two parameters are needed in order to cover the different levels of adjacency proposed by Odden, tier adjacency, syllable adjacency and segmental adjacency.

(21)

*SKIP(Seg)

(22)

*SKIP(G)

Don't skip segments (enforces root-adjacency) Don't skip syllables (enforces syllable-adjacency)

Universally, root-adjacency is preferred over syllable adjacency (spreading is preferably as local as possible), which follows from the Paninian relation in which both constraints stand. As violation of *SKlP(a) entails violation of *SKiP(Seg), only a subset of grammars which allows to skip a segment will also allow skipping an entire syllable. Ranked with respect to some markedness constraint M, different types of spreading follow: (23)

*SKIP(G), * S K i p ( S e g ) » Μ

(24)

*SKip(a) » Μ » *SKip(Seg)

(25)

Μ » *SKIP(O), * S K i p ( S e g )

spreading must be root-adjacent spreading must be syllable adjacent spreading may be unbounded

If Μ ranks lowest, root-adjacency is required in spreading, because appearance of the feature Μ in the output is less marked than non-local spreading (skipping material). If Μ is sandwiched between *SKIP(O) and *SKip(Seg), syllable-adjacent spreading can occur: Segments may intervene, but not entire syllables. Finally, if Μ outranks both SKIP constraints, then unbounded spreading can occur, for the skipping of segments or syllables is less marked than the output realization of M. In the case studies in §4.2 and §4.5, the effect of *SKIP constraints will be illustrated.

Optimal geometries

41

4. Case Studies In the previous section, a number of constraints were introduced that make direct reference to feature geometric structure. In this section, I will use these constraints in a number of model analyses in order to show how they can motivate different types of segment interaction. In §4.1, a relatively simple type of assimilation, voicing assimilation in clusters, will be discussed in order to demonstrate the general mechanism of the system proposed here. In §4.2, a case of long-distance assimilation, voicing assimilation in Kera, will be analyzed, and it will be shown that the difference between local and nonlocal assimilation can be straightforwardly accounted for by the minimal reranking of constraints. §4.3 discusses different strategies of vowel epenthesis in loanwords: Epenthetic vowels can be default vowels or result from vocalic or consonantal spreading. This typology is also captured elegantly by the model, which can moreover explain variability in epenthesis strategies, which is the topic of §4.4. The case studies are concluded by a look at dissimilation and the sketch of an analysis of Dahl's Law in Kikuria (§4.5).

4.1. Voicing Assimilation: Assimilation without

AGREE

The first example discusses voicing assimilation. Its aim is to show how assimilation can be understood without reference to a constraint like AGREE which explicitly enforces assimilation, but that assimilation can be recast in terms of featural markedness alone, assuming an autosegmental model of feature association. This analysis is still very much in the spirit of Beckman (1998) in obviating the need for a pro-assimilation constraint, but at the same time also forms a point of departure from Beckman's proposal, as will be shown by the elaboration of the principles established here in subsequent analyses. In a number of languages, heterosyllabic obstruent clusters agree in voicing (e.g. Catalan, Beckman 1998; Yiddish, Bakovic 1999, Lombardi 1999; Dutch, Grijzenhout & Krämer 2000; for an overview, see Lombardi 1999 or Wetzels & Mascaro 2001). In Yiddish, obstruent clusters agree in voicing; in coda-onset clusters, obstruents assimilate to the onset (Lombardi 1999, Bakovic 1999, Wetzels & Mascaro 2001), as seen in the examples in (26). (26)

vog briv bak bux zis

'weight' 'letter' 'cheek' 'book' 'sweet'

vok/ol briftregar bagbejn buxgajeft zizvarg

'scales' 'postman' 'cheekbone' 'bookstore' 'sweets'

42

Christian Uffmann

The generalization that can be drawn from the data is that the [voice] specification of the onset spreads to the preceding coda. 17 This is graphically represented in (27), under the additional assumption that laryngeal features are dependents of syllable constituents (onset, nucleus, coda), not of the segmental root node (Kehrein 2002). Hence, onset-coda clusters, irrespective of their segmental make-up, contain maximally two laryngeal nodes. 18 (27)

Coda

Onset

I Lar

! Lar

[ a voi]

[ß voi]

How can this process be motivated? Note that the spreading process in (27) removes one [voice] autosegment from the output. Structurally, the configuration is thus more economical with respect to the number of instances of the feature [voice], although it also gains complexity by linking one autosegment to several laryngeal nodes. This is the situation described in §3.2 above where the conflict between *MULTIPLE, the constraint prohibiting spreading, and some other markedness constraint Μ was described. In this example, Μ is the constraint *[±voice] which incurs one violation mark for every [voice] autosegment that occurs in the output. The relative ranking of the two constraints is provided in (28): (28)

* [ ± v o i c e ] » *MULTIPLE

This ranking states that spreading - which results in multiple associations - is better than several [voice] autosegments within the word. The reverse ranking will result in a non-assimilation grammar, treating spreading as more marked than the occurrence of a larger number of [voice] specifications.

17

This analysis assumes that [voice] is a binary feature. Wetzels & Mascaro (2001) provide evidence that [voice] cannot be a privative feature since both values may spread. l8 Note that this is a good example how the use of a stringent geometry can streamline the candidate set and the constraint set. As the geometry requires tautosyllabic clusters to agree in voicing, non-agreeing candidates are not generated, and consequently there is no need for constraints that enforce agreement in such clusters.

Optimal geometries

43

In addition, some statement must be made about directionality of assimilation: The ranking must ensure that codas assimilate to onsets, not vice versa. Without making a commitment to any specific theory like Positional Faithfulness (Beckman 1998), I shall simply assume two blanket faithfulness constraints for onset and coda positions, which are sufficient for this purpose. (29) (30)

FAITH(Onset) FAITH(Coda)

Onset features are faithful to the input Coda features are faithful to the input

There is a presumably universal ranking of the two constraints, given in (31). It states that onsets are universally more faithful to their underlying specifications than codas (see also Beckman 1998). (31)

FAITH(Onset) » FAITH(Coda)

Codas assimilate to onsets, not vice versa

Finally adding the constraint P A R S E ( F ) to our analysis (see ( 5 ) again), we arrive at the tableau in (32), which shows how the correct candidate is selected, using the example [bagbejn] 'cheekbone' (from underlying /bak+bejn/). (32)

Voicing assimilation in Yiddish 1

input:/bak+bejn/

FAITH(OHS)

a. [bagbejn] b.[bagpejn]

*!

PARSE(F)

FAITH(CODA)

Ι

=1=

=1=

=1=

I

*

1 1 1 1 1 1

* *

**

** |

c.fbakbejn] d.fbakpejn]

* [±voice]

*!

=1=

=1=

*MULTIPLE

=1=

*

The four candidates in tableau (32) display all four possibilities with respect to obstruent voicing in the cluster under discussion. Candidate a is optimal; the coda assimilates to the onset with respect to voicing. Although *MULTIPLE and P A R S E ( F ) are violated, the candidate wins because the higher-ranked constraint *[±voice] incurs only one violation (violations are only counted for the cluster under discussion, to enhance readability). Candidate b is the worst of the four; it is unfaithful in both coda and onset without reducing the number of [voice] autosegments. Candidate c is faithful to the input, but it is suboptimal because it has two [voice] autosegments in the output, as opposed to candidates a and d, which have only one. Finally, candidate d is like candidate a with respect to [voice] specifications. However, assimilation is pro-

44

Christian Uffmann

gressive: T h e onset a s s i m i l a t e s to the c o d a . Violation of h i g h - r a n k e d FAITH(Onset) rules this candidate out. Notice that the selection of the optimal output candidate depends on two constraints only whose violation may be fatal, FAITH(Onset) and *[±voice]; low-ranked *MULTIPLE, PARSE(F) and FAITH(Coda) remain inactive. Ranking PARSE(F) or *MULTIPLE above * [±voice] would generate a grammar in which codas do not assimilate to onsets. 19 Under the assumption that FAITH(Onset) universally outranks FAITH(Coda), only two grammars can thus be generated: A faithful one, and an onset-controlled assimilation grammar. Note also that no AGREE constraint is necessary, which sets this analysis apart from the analyses of voicing assimilation in Beckman (1998), Bakovic (1999) and Lombardi (1999); assimilation does not follow from the satisfaction of a constraint that explicitly enforces it but is the result of minimal violation of a markedness constraint which prohibits occurrences of [voice] autosegments. The assimilation candidate is provided by GEN anyway and emerges as optimal if a markedness constraint outranks the prohibition against multiple associations. Assimilation thus becomes epiphenomenal like other processes in OT 20 which are also not explicitly called for by specific constraints but result from the interaction of more general constraints. Epenthesis and deletion, for example, result from the interaction of syllable structure constraints with the faithfulness constraints MAX and DEP but not from an ADD or REMOVE c o n s t r a i n t .

The principle outlined here is in no way restricted to the problem at hand but can be extended to many types of assimilation. If a markedness constraint outranks *MULTIPLE, the marked feature or element can be done away with via spreading. This holds for both syntagmatic markedness constraints (like the above where a marked feature is involved) and paradigmatic markedness constraints which ban certain sequences of sounds - regressive nasalization for example could be conceived of as satisfaction of a constraint against sequences of an oral vowel and a nasal consonant. Similarly, vowel harmony could be captured by the same mechanism (e.g. ATR-harmony: *[±ATR] outranks *SKip(Seg) and *MULTIPLE). Space does not permit to give details of other, similar, analyses here. The applicability of this principle to other cases

19

So far, there seems to be a fair amount of redundancy with respect to the constraints used. In this example, P A R S E ( F ) and *MULTIPLE have a similar effect. That there is a need to distinguish between the two will be shown in §4.3 where epenthetic vowels are discussed. 20 1 would like to thank one anonymous reviewer for nudging me into this direction.

Optimal geometries

45

of assimilation should be sufficiently clear by now, however. Instead, I would like to demonstrate how non-local assimilation can be formalized with only minor adjustments, viz. the inclusion of *SKIP constraints.

4.2. Long-distance voicing assimilation: Kera Voicing assimilation is not necessarily strictly local but can also occur over a distance. The standard example of such long-distance voicing assimilation (e.g. in Odden 1994, Walker 2000) is Kera, a Chadic language, as described in Ebert (1979). According to Odden and Walker, all explosives within a Kera word agree in voicing (fricatives are exempt because they generally do not contrast in voicing). This is true for simplex words (33a) as well as affixed words (33b), at least those words with a £-affix (Ebert 1979): (33)

(a) (b)

dege 'to stomp' tepe 'to gather' k+sir+ki —» kisirki k+taata+w —> kstaataw k+jir+ki —> gijirgi k+dayga+w —> gadaygaw

dabsrga kupurki 'black (m.)' 'cooking pots' 'colorful (m.)' 'jugs'

'chicken' 'billy goat'

On the basis of the available data it is not clear whether voicing assimilation can actually skip a syllable (it does not in the above examples) or whether the process only applies syllable-adjacently (see also Odden 1994 for a similar observation). While clear examples of long-distance agreement (skipping a syllable) are rare,21 so are counterexamples, as the majority of Kera stems are mono- or disyllabic.22 For the analysis, I will nevertheless assume that the

21

Mary Pearce (p.c.) finds three examples only in her fieldnotes and consequently argues against a view of the process as being unbounded. It would be interesting to see how many counterexamples, i.e. long-distance disharmonic forms, one can find to come to a more definitive solution to the problem. 22 The synchronic status and the scope of voicing assimilation do not seem to be clear, either. Several points were disregarded in previous analyses, which render a straightforward analysis somewhat problematic. First, most affixes do not participate (e.g. locative -da), although they participate in vowel harmony. Second, there are a few disharmonic stems in Ebert (1979), e.g. gogloki ~gogoloki 'rooster' or kagay 'hoes' (although Mary Pearce (p.c.) finds that many of Ebert's disharmonic forms are harmonic in her fieldnotes, perhaps due to ongoing change, perhaps due to different

46

Christian Uffmann

process is unbounded, using one of the few examples found in Ebert, which support this view, viz. tssraqka 'old (f.)', from tooroi]+ga. Consider the tableau in ( 3 4 ) . The relative ranking of * M U L T I P L E and *[±voice] remains the same as in Yiddish: The number of [voice] specifications is minimized at the expense of multiple associations, resulting in voicing assimilation. However, *[±voice] is ranked even above *SKip(Seg) and * S K I P ( G ) in Kera (as opposed to Yiddish where voicing assimilation is strictly local, i.e. root-adjacent). Hence, assimilation can also occur long-distance (as long as the interacting features are tier-adjacent; sonorants are not specified for [voice]). Other necessary constraints have been omitted from the tableau in order to make the interaction of * S K I P with * M U L T I P L E and *[±voice] as transparent as possible. 2 3 (34)

Voicing assimilation in Kera

input:/t33raij+ga/ a. [taaraijga] «*·

b.ftaaraijka]

* [±voice]

*SKip(a)

SKip(Seg)

* MULTIPLE

*

*

*

** |

=1=

The tableau shows that local assimilation vs. assimilation over a distance can be straightforwardly described by including *SKIP constraints into the overall ranking. If a markedness constraint outranks both *SKIP and *MULTIPLE, then non-local interaction is possible. Other types of non-local assimilation could thus be formalized similarly to the analysis proposed here to account for longdistance voicing assimilation in Kera. The interaction of *[±voice], *SKIP and * M U L T I P L E yields the factorial typology in ( 3 5 ) . There are three meaningful rankings, each of which corresponds to an attested grammar.

dialects investigated). Third, the (diachronic) devoicing of fricatives obscured the process somewhat. More research (and probably fieldwork) seems to be necessary to come to any final conclusions with respect to the exact status of voicing assimilation in Kera. It exists but the exact environment in which it occurs is still somewhat vague. " F o r example, constraints which determine the directionality of spreading have been omitted. In Kera, voicing assimilation is stem-controlled. Directionality could thus be captured by the constraints FAITH(Stem) and FAITH(Affix) where FAITH(Stem) » FAITH(Affix). Similarly, PARSE(F) has been omitted here because it does not contribute to the selection of the optimal candidate.

Optimal geometries (35)

*MULTIPLE

*SKIP

» *[±voice] 24

» *[±voice] »

*[±voice] »

*MULTIPLE

*SKIP » *MULTIPLE

47

no voicing assimilation (e.g. English) 25 local assimilation (e.g. Yiddish, Dutch) non-local assimilation (e.g. Kera)

By replacing *[±voice] with other markedness constraints, factorial typologies for other kinds of assimilations, local or non-local, can be construed easily. An overview of such assimilations is given, for example, in Odden (1994) and Rose & Walker (2001). The observation that only a subset of features interacts over a distance needs not be formulated independently, since it follows from the nature of the geometry: Only features that are not specified contrastively for all segments may spread. Without feature geometry, some additional mechanism must be assumed to explain the failure of certain features (e.g. C-Place features) to spread. Instead of applying the typology in (35) to other assimilating features, which would only reiterate the same rankings for different markedness constraints, I would like to discuss a different phenomenon next, namely alternations between local and non-local spreading as the interaction of * S K I P and * L I N K ( C , V ) , analyzing different types of vowel epenthesis in loanwords.

4.3. Vowel Epenthesis in Loanwords In many languages, vowels are epenthesized in loanword adaptation if the borrowing language has tighter phonotactic constraints than the donor language. In (36) some examples of epenthetic vowels in different languages are given. Vowels are inserted either to satisfy a ban against codas or to break up illegal consonant clusters.

24

For this ranking, the position of * S K I P is irrelevant as spreading is blocked. *SKIP only has an effect if *MULTIPLE is sufficiently low-ranked, as violation of *MULTIPLE is a prerequisite for *SKIP to be violated. 25 This does not include voicing assimilation within onsets or codas, which of course occurs in English. However, the position was taken above that onsets or codas license only one laryngeal node anyway, following Kehrein (2002). This view disallows conflicting laryngeal specifications in a tautosyllabic cluster on principled grounds, outside constraint rankings.

48 (36)

Christian Uffmann Yoruba Japanese SeTswana Samoan

kilääsi sutoraiku keresemose sikauti

'class' 'strike' 'Christmas' 'scout'

(Akinlabi 1993) (Park 1987) (Batibo 1995) (Cain 1986)

Generally, three general strategies are conceivable to determine the quality of this epenthetic vowel: insertion of a default vowel (the same vowel is invariably inserted), vowel copy or harmony (the epenthetic vowel agrees in some way with a neighboring underlying vowel) or consonantal spreading (the epenthetic vowel shares its place of articulation with a neighboring consonant, most prominently as labial attraction). The different epenthesis strategies can be represented autosegmentally as in (37), where (a) shows default vowel insertion, (b) shows vocalic spreading and (c) shows local consonantal spreading. (37)

Epenthesis strategies a. F V C (V)

c. V C (V)

V C (V)

Each of these strategies is attested. In Japanese, for example, we find default vowel insertion, /u/ is almost invariably epenthesized 26 (Shinohara 1997, Katayama 1998), as shown by the examples in (38). (38)

fesutibaru disuku

'festival' jiguzagu 'zigzag' 'disc' furutaimu 'full-time' (Itö & Mester 1995, Hancin-Bhatt & Bhatt 1997)

In Sranan, a Surinamese Creole language, a different strategy is pursued. Here, the epenthetic vowel harmonizes with the preceding vowel (back/round harmony): ill is inserted after /i,e/; /u/ is inserted after/u,o/ (Uffmann 2002,2004). (39)

26

wiki buku

'week' 'book'

srefi forku

'self' 'fork' (Uffmann 2002)

Unless the preceding consonant is /t,d/; then, /o/ is epenthesized because /u/ triggers affrication of /t,d/.

Optimal

geometries

49

The third strategy is exemplified by the Bantu language Shona where the epenthetic vowel assimilates to the preceding (tautosyllabic) consonant: i\! is inserted after coronals, /u/ is inserted after labials (dorsals behave differently and will be discussed in §4.4 below). (40)

saradhi timu

'salad' 'team'

zoni kirabhu

'zone' 'club' (Uffmann 2002)

The three different strategies can be viewed as the result of constraint interaction. Different rankings lead to the preference of one of the three strategies. First of all, two constraints are in basic opposition which determine whether we find default vowel insertion or some kind of spreading. (41)

DEP(F)

(42)

*MULTIPLE

Features have a correspondent in the input (prohibits feature insertion) N O multiple feature associations (prohibits spreading)

These constraints are in conflict which is resolved by ranking the two constraints with respect to each other: High-ranked * M U L T I P L E prohibits spreading; high-ranked D E P ( F ) prohibits the insertion of features. (43)

*MULTIPLE » D E P ( F )

(44)

D E P ( F ) » * MULTIPLE

default V insertion (Japanese) some kind of spreading (Sranan, Shona)

In the latter case, it must also be established which type of spreading is preferred. Here, the two constraints * S K I P and * L I N K ( C , V ) come into play. On the one hand, *SKIP(Seg) prohibits skipping of a segment, i.e. it enforces local, root-adjacent spreading, which will be spreading from a consonant. On the other hand, * L I N K ( C , V ) prohibits feature sharing between consonants and vowels because the shared feature is linked to two different types of mother nodes (the C-Place node and V-Place node, respectively). Again, these two constraints are in conflict, which can be resolved by two possible rankings: (45)

*SKIP » *LINK

(46)

*LINK»*SKIP

spreading is local, from the consonant (enforces adjacency) spreading occurs from the vowel, across the consonant

50

Christian Uffmann

(47) combines the autosegmental representations in (38) with their violations of the four constraints just introduced. (47)

Epenthesis strategies and constraint violations a. F b. F V

C

(V)

V

Default vowel DEP(F) *MULTIPLE *SKIP *LINK ( C , V )

C

c.

(V)

V C

V V V

(V)

V-spreading

C- spreading *

F

DEP(F)

V

DEP(F)

V

•MULTIPLE

*

•MULTIPLE

*

*SKIP

V

*SKIP

*

*LINK ( C , V )

*

•LINK ( C , V )

V

This translates directly into rankings and tableaux. The example tableaux in (48)-(50) all use the same candidates - in the same order - and the same hypothetical input word team - assuming that the borrowing language does not allow codas and therefore has to add a final epenthetic vowel. The three candidates are (in descending order) a candidate with vocalic spreading (vowel harmony), i.e. timi, a candidate with consonantal spreading, here as labial attraction, i.e. timu, and a candidate with default vowel insertion where the default vowel, being language-specific, is simply indicated by V, i.e. tim V. Note that the three tableaux are not intended to match real data or existing languages perfectly. They are abstractions in order to demonstrate the three basic strategies of epenthesis. The more complicated picture that emerges once we take real data into account will be discussed in §4.4. The first tableau in (48) illustrates a language with default vowel insertion, such as Japanese. * MULTIPLE, the anti-spreading constraint, outranks DEP(F), the anti-insertion constraint. Hence, insertion of a feature is less costly than spreading, and this is why the first two candidates, both of which contain some kind of multiple feature association, are suboptimal. The relative ranking of * S K I P and * L I N K ( C , V ) is irrelevant, since these constraints are only active if spreading occurs. (48)

High-ranked

*MULTIPLE

forces default feature insertion (Japanese): ι

input:/tim/

•a?

•MULTIPLE

DEP(F)

*SKip(Seg)

Ι

*

1

a. [timi]

*!

t

b. [timu]

*!

1

c. [timV]

*

I

Ι

1

1

1

I

1

*LINK(C,V)

*

Optimal geometries

51

The second tableau (49) exemplifies a pattern like the Shona one where the default epenthetic strategy is local assimilation to the place of articulation of the consonant. Here, D E P ( F ) outranks * M U L T I P L E : Feature insertion is less optimal than spreading. Which type of spreading occurs is established by the relative ranking of * S K I P and * L I N K ( C , V ) . As * S K I P ranks above * L I N K ( C , V ) , local linkage will be preferred although two dissimilar segments are linked. (49)

Local (consonantal) spreading: D E P ( F ) » * M U L T I P L E and * S K I P I ι

*SKip(Seg)

a. [timi]

ι

*!

b. [timu]

I

input:/tim/

«·

» *LINK

c. [timV]

DEP(F)

*!

(Shona) •MULTIPLE

*LINK(C,V)

ι =1=

* =1=

ι ι

The third tableau (50) describes languages like Sranan in which the quality of the epenthetic vowel is determined by vowel harmony. Again, as in the previous tableau, D E P ( F ) ranks high, thus rendering feature insertion a suboptimal process. In contrast to ( 4 9 ) , however, the ranking of * S K I P and * L I N K ( C , V ) is reversed. Now the constraint against linkage between a C-Place node and a V-Place node ranks above the constraint demanding strictly local spreading. Consequently, the vowel-harmonic candidate will be preferred over the locally linked one. (50)

Vowel harmony:

D E P ( F ) » *MULTIPLE

and

*LINK » *SKIP

(Sranan)

I

input:/tim/ «*·

DEP(F)

ι

a. [timi]

I

b. [timu]

ι

c. [timV]

*!

*LINK(C,V)

*j

*SKip(Seg)

* MULTIPLE

*

*

1

*

ι ι

In sum, this case study has shown how a general typology of epenthetic vowels can be formalized elegantly and economically within the framework proposed in this paper. The typology of epenthetic vowels falls out from the interplay of only four constraints on feature insertion and feature spreading. Conversely, all possible rankings of the four constraints yield one of the three attested types of vowel epenthesis. In the following section, I will have a closer look at epenthesis patterns in two selected languages, Shona and Sranan. In both languages, we find variation regarding the epenthetic vowel - differ-

52

Christian Uffmann

ent strategies are found in different environments. I will show that this variation is also accounted for by the model suggested here.

4.4. Variability in loanword epenthesis Often languages employ more than one epenthesis strategy, in different environments. In Sranan, for example, vowel harmony is the general strategy but if the vowel is /a/, then we find consonantal spreading, because /a/ is opaque in the harmony process. This variation is shown in (51). In (a), we find two examples of the general process (vowel harmony) again. In (b), however, we can see that if the vowel is /a/, insertion of /u/ after a labial consonant and insertion of /i/ after a coronal consonant follow. (51)

Variation in Sranan (a) luku 'to look' (b) tapu 'top'

sribi hati

'to sleep' 'hot'

In Shona, on the other hand, the general process is consonantal spreading, but not if the consonant is dorsal, in which case vowel harmony is attested. In (52a), the general process is shown again where /i/ is inserted after coronals and /u/ is inserted after labials. In (b), however, we see that after dorsals the pattern is different: /i/ is inserted after /i/, and /u/ is inserted after /u/. 27 (52)

Variation in Shona (a) chuni 'girlfriend' > * L A T S O N

Languages with no laterals (18.6%, Maddieson 1984) Common language type, with sonorant laterals Languages with both obstruent and sonorant laterals

Examples of obstruent laterals include not only the obvious affricates and clicks, but also languages in which [1] patterns as a voiced obstruent, such as Min, which has [1] instead of [d]. For example, /p,t,k/ voice to [b,l,g] footinternally (Hsu 1996) and /b,l,g/ nasalize to [m,n,q] before nasal vowels. In some Bantu languages, like Ikalanga, historical *d has become III, but under velarization /l/ becomes the stop [gw], suggesting that it may still be an obstruent. The prediction of the fixed ranking given here is that no language can have only obstruent laterals and no sonorant laterals. While this is certainly the usual case, there are some possible counter-examples, including Min if its [1] is an obstruent. However, Min has no other oral sonorant consonants - no Irl - so high ranked S O N = N A S could be invoked. An alternative might be to say that this [1] is not phonologically [lateral] at all, but is just an oral stop. Another possible counter-example is Tlingit, which has fricative and affricate laterals, but no voiced approximant. This needs further investigation. Finally, the existence of voiceless lateral sonorants as in Tahltan [1] (Shaw 1 9 9 1 ) simply implies that S O N = V O I C E can be low-ranked.

70

Moira Yip

3.1.2. Preference for Coronals I now turn to the preference for Coronal place. The typology is given below: (8)

*LATLAB »

* L A T D O R S » * L A T C O R » FAITH

*LATLAB »

* L A T D O R S > > FAITH »

*LATLAB »

FAITH »

*LATDORS > > * L A T C O R

*LATLAB »

*LATDORS > > * L A T C O R

FAITH »

*LATCOR

Either no laterals, or placeless ones, Common type, with Coronal laterals3 N e w Guinea type, with velar and coronal laterals Unattested

An example of placeless laterals comes from Cambodian (Nacaskul 1978). The co-occurrence restrictions on identical Place features do not treat /l,r/ as Coronal, even though Place restrictions cross-cut obstruents and sonorants, stops and fricatives, nasals and glides, as the following table shows. Instead, they behave like [h ?] in co-occurring freely with all other sounds. (9)

C1\C2 ρ j. C

k

bpmw Ί· Ί· ^

tdn *** *79

rl

csfiy

kqh?

*** * *

Languages with Dorsal laterals will include those like Mid-Waghi (see (1) in §2.1), and also perhaps languages with palatal laterals, which have been argued to be both Coronal and Dorsal by Sagey 1986 and others. The last grammar in (8), which predicts the existence of the unattested labial laterals, is an unexplained gap. One possibility is that the perceptual effects of lateral re-

3

Walsh (1997) argues that all laterals have both Coronal and Dorsal Place. This is certainly true phonetically in some languages, and perhaps phonologically too (in English, for example, /l/ vocalizes to the Dorsal [u] in many dialects, and children often turn coronals into velars before [1]), but in other languages there is no evidence of a phonologically active Dorsal component. Palatal laterals may also be Coronal and Dorsal, and contrast with plain Coronal laterals. This analysis of palatals is problematic for the view taken here, as a reviewer points out, since it seems to require a positive constraint LAT=COR, but I have no room to explore this further here.

Variability in feature affiliations through violable constraints

71

lease in labials would be too subtle to make such a contrast functionally effective. I shall have nothing further to say about labial laterals. The main prediction of this typology is that no language should have only Dorsal laterals. Either it must have complex corono-dorsal laterals, or Dorsal and Coronal ones in contrast. Blevins argues that many of the cases of apparent velar laterals, such as Yagaria and Kunite, are in fact phonologically complex, being both Coronal and Dorsal. The fact that they have h but no III is then not a problem. However, Mid-Waghi has a Dorsal Id that (contra Blevins) does not seem to be in any way Coronal, 4 but the language also has contrasting IV and / 1/. For Blevins, committed to [lateral] under the Coronal node, this is a problem, since all laterals must be Coronal, but for the approach outlined here plain Dorsal laterals are fine, so long as they contrast with plain Coronal ones. Putting together the results of this section, a language with only coronal sonorant laterals will have the grammar in (10): many of the languages discussed in this paper are of this type. (10)

*LATOBS, *LATDORS »

IDENT-LAT »

*LATSON, *LATCOR

The relevant faithfulness constraint here is IDENTLAT, which requires that segments that are lateral in the input must be lateral in the output. 5 The Richness of the Base hypothesis means that inputs such as Id will inevitably arise, but they cannot survive under this grammar. IDENT-LAT says nothing about the creation of new laterals, but these will violate two of the markedness constraints in the above grammar, so that a new instance of IV will violate *LATSON and *LATCOR. AS a result, if a constraint such as SHARE-F that could create new laterals is ranked below the grammar in (10), no new laterals can be created, although underlying /l/ can survive. In the next section we look more closely at assimilation.

2.2. Targets of spreading Let us assume that assimilation involves a violation of the IDENT family of faithfulness constraints, such as IDENT-PLACE, and IDENT-SON, under pressure

The fact that Id assimilates to a following coronal as in /ai_-to/ > [alto] 'eastwards' does not demonstrate that it has an underlying Coronal node of its own. 5 Here I assume that non-laterals, with the possible exception of Irl, are unspecified for laterality underlyingly.

4

72

Moira Yip

from higher ranked constraints such as SHARE-F (or AGREE) and SYLLABLE CONTACT. Any assimilation process that creates the ordinary sonorant Coronal lateral [1] from an underlying non-coronal or non-sonorant will thus violate at least one of IDENT-PLACE and IDENT-SON (and of course also IDENTLAT). The ranking of these constraints with respect to the constraints causing assimilation, here abbreviated as ASSIM, will determine which segments may undergo the process. If IDENTSON » ASSIM, targets must be sonorant. If IDENTPLACE » ASSIM, targets must be Coronal. If the output is always Coronal and sonorant, * L A T O B S and *LATDORS are always high ranked, and *LAT COR and *LATSON are always low-ranked. The following typology results: (11)

a.

Target must be sonorant:

a'.

Target need not be sonorant, but output will be:

*LATOBS, IDENT-SON » *LATOBS »

ASSIM »

ASSIM > > * L A T S O N I D E N T - S O N , *LATSON

b.

Target must be Coronal:

b'.

Target need not be Coronal, but output will be:

*LATDORS, I D E N T - P L A C E » *LATDORS »

ASSIM »

ASSIM »

*LATCOR

I D E N T - P L A C E , *LATCOR

By combining one of the sonorancy rankings with one of the Place rankings, we get the following mini-grammars (with low-ranked *LATCOR and *LATSON omitted for space reasons). (12)

a & b.: Target must be sonorant and Coronal: Flemish, Toba Batak *LATOBS, *LATDORS, I D E N T - P L A C E , I D E N T - S O N » A S S I M

a & b': Target must be sonorant, but need not be Coronal: Selayarese *LatObs, *LatDors, Ident-Son » A s s i m » Ident-Place a'&b: Target must be Coronal, but need not be sonorant: Sanskrit, Yanggu *LATOBS, *LATDORS, I D E N T - P L A C E » A S S I M »

IDENT-SON

a ' & b ' : Target need not be Coronal or sonorant, but output will be both:? *LATOBS, *LATDORS, ASSIM »

IDENT-SON, IDENT-PLACE

Finally, rankings with *LATSON, *LATCOR ranked above ASSIM (and thus *LATOBS, *LATDORS even higher) would not allow laterality to surface at all on the target, so we would observe either failure of assimilation before laterals (Javanese), or possibly assimilation of other lateral properties, such as voicing (Polish), or coronality (Chukchi), but not laterality.

Variability in feature affiliations through violable constraints

73

3.3, Possible outputs of assimilation What about the possible outcomes? If *LATDORS, * L A T O B S » ASSIM, the outputs must be coronal sonorants, and this is the most common case. If A S S I M » *LATDORS, assimilation could create velar laterals. Rather surprisingly, this seems to be unknown, but palatal laterals, which may be thought of as both Coronal and Dorsal, can certainly be created, as in English welch [Xtf]. Lastly, if ASSIM » *LATSON, assimilation could create lateral obstruents. I am not aware of such cases, but some reports of failure of assimilation in /t1/ inputs could perhaps actually be reinterpreted as [tl 1] outputs, which would be hard to distinguish from simple [tl] clusters. McCarthy (2002) argues that markedness must distinguish between underlying marked segments ('old' markedness) and derived marked segments ('new' markedness). He calls preservation of underlyingly marked segments coupled with failure to tolerate creation of new marked ones the 'grandfather' effect. When a segment is highly marked, like [tl ], any grammar that creates new instances of this must have 'new' markedness for *LATSON ranked very low. If the language also has no underlying /tl /, the grammar has * O L D t l » IDENT, ASSIM » * N E W tl. I will not explore this further here, but see Yip (in press) for more details. In the rest of this paper I show how this approach deals with a representative subset of the processes in Table I.

4. How this account deals with a sample of the languages in Table I 4.1. Does lateral spread: the second row of Table I 4.1.1. Selayarese: When Coronal spreads in Place assimilation, so does [lateral] Selayarese has a rule of nasal Place assimilation pictured below. If the trigger is [lateral], the velar nasal becomes not just Coronal but also lateral (Mithun andBasri 1985): (13)

/annaq/ 'six'

anna[mp]oke anna[nj]arang anna[nt]au anna[nr]upa anna[l l]oka

'six 'six 'six 'six 'six

spears' horses' persons' kinds' bananas'

74

Moira Yip

In a feature-geometric analysis, Place spreading spreads all the lower nodes, including Coronal, and this drags its subsidiary [ant, distr, lateral] as well. In the approach outlined here, it is an instance of a language where the target must be sonorant, but need not be Coronal, as in (11 a&b'): (14)

* L A T O B S , * L A T D O R S , IDENT-SON »

SHARE-F »

IDENT-PLACE,

* L A T S O N , *LATCOR

In order to locate each language within the language typology outlined in §2, here and in the rest of the paper I give a full ranking as in (14), but in the tableaux I show only the most relevant constraints. Not every ranking is motivated by a full tableau, for reasons of space. For example, the high ranking of *LATDORS here is motivated by the complete lack of [L] in the inventory of Selayarese, and the high ranking of IDENT-SON by the fact that only sonorants can be targets. Lastly, some rankings are posited as universally fixed, such as *LATDORS »

*LATCOR.

The tableau therefore shows only the crucial constraints: SHARE-F is violated once for each unshared lateral or Place feature, along the lines of the gradient use of SPREAD-F in Padgett (2000). (15) INPUT:/IJL/ Ι®"

A.

*LATDORS

SHARE-F

11

IDENT-PLACE

*LATSON

*

**

*

B.IJL C. NL D.

L!

*! *!

*

*

*

**

In candidate (d), an illicit Dorsal lateral has been created. In (c) the feature [lateral] is not shared, and in (b) neither Coronal nor [lateral] is shared. In (a) both are shared, so although this violates IDENT-PLACE it is the optimal candidate. Note that for a /tl/ input, IDENT-SON » SHARE-F will block lateralization.

4.1.2. Chukchi: Place spreads, but laterals spread only their coronality We now turn to a case where Place assimilation does not create laterals. The data here come from Chukchi (Clements & Hume 1995:270). Similar facts hold in English intrusive stop formation, and also in Catalan and Yoruba.

Variability in feature affiliations through violable constraints

(16)

tor]-9l?-3n tam-pera-k tam-vairgin tam-waysrx-an tan-tsai

'good'(/tEq-/) 'to look good' 'good being' 'good life' 'good tea'

ten-leut tan-ran tcn-yolqot-ak

75

'good head' 'good house' 'to sleep well'

As in Selayarese, the target is a sonorant. The failure to spread lateral is clearly not due to high-ranked IDENT-PLACE since the input velar nasal does become a Coronal nasal. Instead, it seems that new laterals cannot be created by assimilation because of high-ranked *LATSON (or perhaps *LATCOR). (17)

*LATOBS, *LATDORS »

*LATSON, (*LATCOR) »

SHARE-F »

IDENT-

PLACE

input:/ij-l/ •a?

•LatDors

•LatSon

Share-F

Ident-Place

=1=

a.nl

=1=

=1=

b.ijl

=1=

** I

c. 11 d. J

=1=

* * |

*!

**

=1=

As before, candidate (d) creates an illicit Dorsal lateral. Now, however, the Coronal [1] in candidate (c) is ruled out by *LATSON. (a) wins out over (b) because it at least shares Place features. One point of clarification might be useful. A reviewer suggests that *LATSON could be replaced by IDENT-LAT here, and in many other cases in this paper. This is not in fact true, for reasons explained at the end of section 3 . 1 . 1 . IDENT-LAT looks only at underlying instances of [lateral], whereas the *LAT constraint family looks at both old and new cases. To block lateralization of non-laterals, *LAT constraints are essential.

4.1.3, Sanskrit: Sonorant voicing spreads, taking [lateral] with it The third case is again one that creates new laterals, like Selayarese, but this time as part of a process that spreads voicing, nasality, and laterality, creating new sonorants. In Sanskrit (Whitney 1889) stops optionally assimilate in nasality to the following segment, (18a), [Whitney: 161], and coronals lateralize, (18b). If the target was a nasal /n/, as in (18c), it lateralizes but stays nasal [Whitney: 65], Before other voiced segments, as in (18d), obstruents voice.

76

Moira Yip

(18)

a.

b. c. d.

tat namas vak m e tri s tup nunam tat labhate trin lokan /..kl.../

—> tan namas / tad namas —» vaq me / vag me —» tristum nunam / tristub nunam tal labhate —;> tril lokan —> [...gl...] No examples, but Whitney: 54 is explicit on this point

In a feature-geometric analysis, this process has been analyzed as spreading sonorant voicing, dragging nasality and laterality with it. 6 Empirically, this is problematic, for two reasons: (i) /nl/ should become *[11], with no nasalization left on the target (ii) the spreading of obstruent voicing must be separately handled. Let us assume that this process is triggered by the Syllable Contact Law, as argued by Davis and Shin 1996 for Korean. The unacceptable sequences are those with a rising sonority incline across the syllable boundary: nl, dl, dn, tl, tn, td. These are adjusted to create the best possible syllable contact, but using the available features only: neither nasality nor laterality can be inserted, although they can be spread. There is no need to see the spreading as taking a bundle of features, just whatever is needed to improve syllable contact (cf. Padgett 1995). Since the target of lateralization is Coronal but not necessarily sonorant, we have the following grammar, an instance of (1 l a ' & b ) . T h e g r a m m a r w e n e e d is t h i s : * L A T O B S , *LATDORS, IDENT-PLACE, D E P F »

SYLLCONTACT » I D E N T - S O N , * L A T S O N , *LATCOR. SYLLCONTACT v i o l a -

tions are counted by the number of steps up on the sonority scale between C1 and C2, where the scale is t < d < η < 1. The first tableau shows an input with a coronal stop.

6

Korean has usually been cited as the classic case of this type, but Kang (2002) shows that contra most reports obstruents do not usually lateralize, so that han+path +lo > hanpanno 'Hanpat Street', *[hanpallo].

Variability in feature affiliations through violable constraints (19)

Lateralization of coronals

input :/tl/ ι®"

77

•LatDors

a. 11

I 1 I

SyllContact

Dep-F

ΦΦ

b. nl 1

c. dl

I

d. tl

•LatSon

1

=1=

=1=

** I

=1=

*** I

=1=

Although not shown in the above tableau, the winner also violates IDENT-SON and *LATCOR (twice), so these must be low-ranked. The language has no lateral obstruents, so *LATOBS must be undominated. The next tableau shows an input with a non-coronal. (20)

Voicing of non-coronals

input :/kl/ •a?

a.

•LatDors

gl

b. Μ c. ljl d. 11 e. J

*!

' Ident-Place I 1 1

' I 1 1

1

1

I 1 1 ι

1 1 1 ι

*|

Dep-F

SyllContact

•LatSon =1=

**

=1= =1=

*!

=1= **

**

Note that candidate (a) above does not violate DEP-F, since the underlying [voice] feature of IV has simply spread left onto the velar. In the next tableau, I show what happens to an /nl/ input, in which the nasal lateralizes, but remains nasal: (21)

Preservation of nasality is a MAX-F effect

input :/nl/ ι®"

MAX-F

a. tl b. 11 c. nl

*!

I 1 I1 ι 1

DEP-F

SyllContact

* LatSon **

**

*!

*

Lastly, for a /tn/ input [nn] will win out over [tn] or [dn] by virtue of better satisfying SYLLCONTACT.

78

Moira Yip

4.1.4. Polish: Laterals cause voicing assimilation but not

lateralization

The fourth case, like the second, involves a failure to create new laterals, but this time when voicing spreads from a sonorant to a preceding consonant. The following data show a post-lexical process in Krakow and Poznan Polish (Glowacka, p.c., Madelska et al 1998): (22)

sybax trop brat syn

—> sybafy] lesriicufki —> tro[b] lisa —> bra[d] Doroty / Natalji / Iwony / Luizy -> sy[n] Luizy

'brother of X ' 'son of Luiza'

Similar facts hold in all dialects between verbal prefixes and roots: (23)

s-kojitf+te

'to end'

z-bitc

'to break'

z-litf+tc

'to count'

The first thing these facts demonstrate is that voicing in obstruents and sonorants must be handled by the same feature, as discussed in footnote 2. The second thing they show is that if voicing assimilation caused by sonorants involves spreading the SV node, we cannot explain why nasality and laterality do not spread. I assume that voicing assimilation results if IDENT-PLACE » SHARE-F » IDENT-VOICE. IDENT-PLACE here can either be construed as a reference to a feature class, following Padgett 2 0 0 0 , or as a shorthand for IDENT-LAB, IDENTCOR, IDENT-DORS etc. See § 4 for more discussion. Here I consider only coronal inputs, starting with /tl/. Candidate (b) in (24) creates an illicit lateral obstruent. Candidate (c) violates IDENT-SON. Candidate (a) with voicing assimilation does better than (d) on SHARE-F. (24)

*LATOBS, I D E N T - S O N , * L A T S O N , IDENT-PLACE »

SHARE-F »

IDENT-

VOICE

input:/tl/

1

•a? a. dl b.dl-1

IdentSon

•LatObs

•LatSon

Share-F

=1=

ΦΦ [lat,son] =1= [son]

=1=

*!

c. 11 d. tl

1

1

=1=

* *

*!

1

*

Ident-voice =1=

ilülül: [lat,son,voice]

Variability in feature affiliations

through violable constraints

79

In the account proposed here, we can again attribute the failure of lateral spread to high-ranked *LATSON. It is the /nl/ inputs below that demonstrate the need for this constraint. (25) input :/tl/ •a?

•LATOBS

IDENTSON

•LATSON

SHARE-F

a. nl

=1=

** [lat.nas]

b. 11

** I

IDENT-VOICE

3.2, Three superficially different targets of spreading I now turn to how the proposal handles different targets of spreading. The examples are from the third row of Table I.

3.2.1, Flemish: Coronal sonorants as target of [lateral]

spreading

In Teralfene Flemish (Blevins 1994) [lateral] spreads rightwards to Coronal sonorants. Other Coronals appear to be transparent, as in (a), but non-coronals are said to block and not undergo the process, as in (b). (26)

a. b.

smelt-η vals-n elp-n zwolme

smeltl valsl elpen zwoleme

'to melt' 'filings' 'to help' ?

Blevins' feature geometric analysis says that [lateral] can only spread to adjacent Coronal nodes. Any non-coronal interrupts the adjacency, [-son] segments also accept [lateral], and it is either subsequently delinked, or possibly stays and is present phonetically but not perceived. I begin with a re-interpretation of the facts. Suppose that the assimilation, like most consonantal assimilations, is a fact about clusters only. Then notice that the examples in (b) that purport to show blocking by non-coronals all surface with a vowel between the lateral and the nasal target, so that they are

80

Moira Yip

no longer in a cluster together. I therefore suppose that only vowels block the process. 7 Given this preamble, the targets of assimilation must b e both sonorant and Coronal, so the basic g r a m m a r is as follows (an instance of (12 a&b)): (27)

* L A T O B S , * L A T L A B , IDENT-PLACE, I D E N T - S O N »

SHARE-F »

NOGAP,

*LATSON, *LATCOR

Showing only the crucial constraints, the following tableaux demonstrate the rankings, l m stands for a labial lateral, a non-existent sound for which there is no IPA symbol! SHARE-F violations are calculated pairwise. For example, in [lpl], [1] and [p] differ by two features (Place and lateral), as do [p] and fl], and [1] and [1] do not differ at all, so w e assign four stars to [lpl]. The first tableau shows the failure of non-coronals to assimilate.

(28) input:/lm/ a. 1 lm

•LATLAB

1

IDENT-PLACE

1 1 ι

=I=|

b. 11

SHARE-F =1=

*!

•a? c. lm

* *

Since obstruents are transparent, it appears that targets must also be sonorant, and that gapping is allowed: (29)

input :/ltn/ a. lfll b. Ill c. Itn ι®" d. ltl

7

•LATOBS

'

=I=|

'

IDENT-SON

SHARE-F

NOGAP

**

*

I 1 1 1

*!

I 1

Lastly, many dialects of this area have a small excrescent or epenthetic vowel between a lateral and a following non-coronal, so that [elpen] is likely to be more like [elopen] (John Harris, p.c.). In that case laterals are never in a cluster with anything except a following coronal. However, it could be argued that this vowel is only a

Variability' in feature affiliations through violable constraints

81

In the case of non-coronals sandwiched between two coronals, obligatory epenthesis takes place after the non-coronal. My data is insufficient to be sure what causes the epenthesis. Perhaps Place reversals in clusters from Coronal to Labial and back to Coronal are banned, triggering epenthesis, an idea formalized here as *CORLABCOR. In any case, what is crucial in the tableau below is that SHARE-F applies to the [lp] clusters only, so there is no motivation for creating a new lateral from the now solitary [n]. (30) input :/lpn/

•CORLABCOR

a. lpl

DEP

*!

SHARE-F

•LATSON

ilülülül:

* *

b. lpel

=1=

* *

c. lpen

=1=

* *

=1=

The final grammar is shown below: (31)

*LATOBS, *LATLAB, IDENT-PLACE, IDENT-SON, * C O R L A B C O R » DEP »

SHARE-F »

N O G A P , *LATSON, *LATCOR

4.2.2. Yanggu Chinese: All Coronals as targets of [lateral]

association

The data discussed in this section add an extra complication, in that the [lateral] feature is floating, and must acquire its Place from another segment in order to surface. Even so, it can be handled with the tools developed here, with high-ranked *LAT LAB, *LATDORS again being crucial. Yanggu Chinese has a very unusual affix (Yip 1992). It may be rhotic or lateral, or both, surfacing variously as [r, 1, R, and J], where the last two are retroflexed. All that concerns us here is that the lateral variants only surface in words with surface coronals, and they are attracted to the rightmost coronal. This mobility suggests that [lateral] is a floating feature:

phonetic side-effect of release, so in what follows I will assume it is absent. Note that these excrescent vowels differ from the [e] in the second syllable of [elpen], which is obligatory, and must be phonologically present.

82

Moira Yip a.

N o Coronals:

b.

Initial Coronal:

c.

Initial and final coronals:

xou ρε tu na ts h uon

xour per tlur nlar ts h uslο

'monkey' 'card, board' 'rabbit' 'to press' 'village'

The facts are quite complex: see Yip 1992 and Chen 1992 for full details. Here I stick to the core of the phenomenon. Since the laterals appear to form onset clusters, I start by assuming that clusters must be homorganic, or that SHARE-F holds in clusters. Coda clusters are disallowed by *COMPLEXCODA, and will not be considered here. I also assume that placeless consonants are n o t a l l o w e d , SPECIFYPLACE. Lastly, *LATLAB, *LATDORS are u n d o m i n a t e d ,

meaning that laterals must be Coronal. The combination of these means that the [lateral] feature must seek out a Coronal node, and if none is available, it cannot surface. Where there are two Coronal nodes, ALIGN-R selects the rightmost. The 'Coronal target' effect is here achieved by *LATLAB, *LATDORS » MAX-LAT, and is exactly parallel to cases like Flemish where spreading to Coronal targets is achieved by *LATLAB, *LATDORS » SHARE-F.8 The full g r a m m a r is SHARE-F, *LATLAB, *LATDORS, SPECIFYPLACE »

MAX-LAT-

E R A L » *LATCOR., ALIGN-R. In the first t w o t a b l e a u x final [1] in the o u t p u t

candidates is a placeless lateral The first tableau shows an input stem with no Coronal segments. SPECIFYPLACE means that [lat] needs a Place node, and *LATDORS forces it to be Coronal. Candidate (b) below, in which [lat] has no Place node, is thus ruled out. I assume that high-ranked DEP-F (not shown) rules out insertion of new Coronal nodes. The affixal /r/, on the other hand, comes with its own Place features, so candidate (a) satisfies SPECIFY-PLACE.9

8

9

An alternative to MAX-LAT would be MAX-AFFIX.

There is one outstanding puzzle: why can't the Place node of /r/ host [lat], creating a retroflex lateral, since this is how [lat] surfaces on a stem-final Coronal, as in (35)? It is possible that /r/ is [-lat] underlyingly, but I leave this question for future research.

Variability in feature affiliations through violable constraints

(33)

Failure to surface in the absence of Coronals

input:/xou-r[lat]/

SHARE-F

b. xoul c. XLOur d. xlour

•LATDORS

'

*'

1 1 1 Ι 1 1

1 I 1 Ι 1 1

a. xour

*!

I

SPECIFY PLACE

MAX-LAT

=1= *!

I

If the input has a coronal node, [lat] seeks it out to satisfy shown below. (34)

83

SPECIFYPLACE,

as

Attraction to Coronal nodes

input:/na-r[lat]/

SHARE-F

1 1 1 *LATDORS 1

a. nlar

1

1

b. nar

1 Ι

1 I

I

I

c. nal

SPECIFY PLACE

1 MAX-LAT

*LATCOR

I ALIGN-R

Φ

*

I I

*!

*!

1

Where there are two Coronal nodes, the rightmost wins, and here the final [1] at last has a Place node of its own, and can surface at the right edge. (35)

Preference for right-most coronal

input:/ts h u3n-r[lat]/ «·

SHARE-F

1 1 1 *LATDORS 1

SPECIFY PLACE

1 MAX-LAT

*LATCOR

a. ts^ial

1

1

*

b. tsluar

1 Ι

1 I

*

1

1

c. ts^iar

*!

4.2.3. Tahltan: Harmony that targets Coronals ignores [lateral]

I ALIGN-R

1

*F

1

affricates

The next case is one where lateral segments fail to undergo a harmony that otherwise targets coronals. Tahltan (Shaw 1991) has five series of coronals, shown here by their voiced affricate members: d, dl, dö, dz, (%. At each place of articulation, there are voiced/voiceless and glottalized as in - [dz, ts, ts']. At the last four places, there also voiced and voiceless fricatives, as in [z, s].

84

Moira Yip

Within a word, the last three series harmonize with the rightmost participant. In the following data, the first-dual subject prefix /6i(d)/ stays [Θ] before noncoronals or the /d/ and /Θ/ series, but becomes [s] before the /dz/ series, and /J7 before the /&,/ series: (36)

9i:t6asdi ni-si-t'a:ts u-ficfe,e

'we ate it' 'we got up' 'we are called'

The lateral /dl/ series are not triggers or targets, and are transparent: na6iba:tl 'we hung it'. My proposal here is extremely simple. The failure of the lateral series to participate in harmony results from high-ranked *LAT » SHARE-F. Since all the other series are non-lateral, they are free to harmonize. *LAT here is shorthand for *LATOBS, *LATSON, *LATDORS and *LATCOR. Note that underlying laterals will still survive, if IDENT-LAT » *LAT. (37)

Harmony with /s/

input:/0i-s/

*LAT

a. 0i-s «·

(38)

*!

b. si-s

No harmony with/l/

input:/0i-l/ •a?

SHARE-F

*LAT

=1=

a. θϊ-1 b. li-1

SHARE-F

*!

This suggestion has one drawback (Clements, p.c.). It predicts the possibility of a system in which one of the other features resists change as the result of high-ranked * ANTERIOR or *DISTRIBUTED. In the former case, [Θ] and [s] could harmonize, since both are [anterior] and only the feature [distributed] would change, but [s] could not. In the latter case, [s] would be the one left out. I know of no cases that fit either description, although to make the point one needs a language with the three-way contrast between /Θ, s, s/, and consonant harmony, a rare combination. See Gafos (1996) for a useful survey of coronal harmony systems, and a rather different view as to the spreading feature.

Variability in feature affiliations through violable constraints 4.2.4. Toba Batak: [Sonorant

Voice] as the target of lateral

85

spreading

One of the classic cases that has been used to argue the case for [lateral] being under an SV node is Toba Batak (Hayes 1986). Coronal sonorants assimilate to a following liquid: (39)

Spreading nr —> rr nl —> 11 rl —> 11

No change In, rn, lr mr, ml, ql, qr rr, 11, nn.

The target conditions are thus identical to the Teralfene Flemish case, and the same basic grammar (an instance of (12a&b)) suffices: 10 (40)

*LATOBS, *LATDORS, IDENT-PLACE, IDENT-SON »

SHARE-F,

*LATSON, *LATCOR

input :/nl/

•LATDORS

IDENT-PLACE

IDENT-SON

SHARE-F

IDENT-PLACE

IDENT-SON

SHARE-F

ι®" a. 11 b. nl (41)

Only Coronals lateralize

input:/ijl/ •a?

•LATDORS

a. ljl

=I==I==I=

b. 11 c. Li (42)

•LATOBS

IDENT-PLACE

IDENT-SON

a. tl b. tl-1 c. 11

10

=1=

*!

Only sonorants lateralize

input :/tl/ •a?

*!

SHARE-F * *

=1=

*! *!

The assimilation here may be caused by the Syllable Contact Law, not SHARE-F, except for the fact that /lr/ is unchanged.

86

Moira Yip

4.3. Laterals as targets The cases above have concerned the behaviour of lateral as the trigger of some process, but the complete picture requires us to understand their behaviour as targets, and in particular whether laterality survives or is lost. For reasons of space the full set of cases cannot be covered here, so I will just sketch the problem and proposal. In a fixed feature geometry, if a superordinate node is eliminated by spreading of a neighbouring node, [lateral] should be lost. So if [lateral] is under Coronal, it should be lost under Place assimilation. However, Row four of Table I shows cases where laterals as the target of rules that spread Place or voicing may or may not keep their laterality, as in English where /l/ assimilates to a following dental, but without ceasing to be lateral: we[16] 'wealth'. Similarly, if a superordinate node is lost by some process like debuccalization in coda position, [lateral] should be lost. However, Row five shows that when place or voicing contrasts are neutralized for such reasons, laterality may or may not survive. Caribbean Spanish allows only velars in coda position, so that /tren/ —> treq 'train', but laterals survive: /tonel/= [tonel] 'barrel'. The core of an OT account is that if IDENTLAT is ranked above whatever causes neutralization, laterality will survive. In the case of spreading, we have IDENT-LAT » S H A R E - F , and in the case of neutralization due to place markedness, we have IDENT-LAT » *CORONAL. The full details are worked out in Yip (in prep.)

5. Conclusions I have argued that traditional universal feature geometry is too rigid to handle variation like that seen with laterals. It is a desirable property of OT that it allows for cross-linguistic variation in affinities between features, while also expressing universal preferences as fixed rankings of constraints governing feature-combinations. These fixed rankings are grounded in phonetic dictates. The preference for Coronal laterals is the phonologization of the articulatory fact that lateral release is most readily produced with the blade of the tongue not the dorsum or the lips. The preference for voiced sonorant laterals is the phonologization of the fact that in a laterally released sound the airflow is never obstructed enough to hinder spontaneous voicing. The need for admitting this sort of variation makes any attempt to incorporate a fixed feature geometry into OT a retrograde step. It is also unnecessary: the advantages of feature geometrical theories can be achieved by con-

Variability in feature affiliations through violable constraints

87

straints on feature co-occurrence, along the lines of Padgett (1995, 2000). The arguments for representational approaches to feature combinatorics are rendered moot. The arguments here have been based entirely on the feature [lateral], but what of other features? Variable behaviour might be seen whenever the features are most readily produced on a certain type of segment, but can nonetheless be produced on other sounds too. For example, [strident] sounds, in which the turbulence produced at the point of constriction is sufficiently strong, and/or where the ensuing airstream then hits a sharp obstacle like the teeth, is easy to produce with the tip or blade of the tongue, but hard to produce elsewhere. We derive from this a constraint hierarchy * [Labial, strident] » * [Coronal, strident]. Languages which contrast [f] and [φ], like Ewe, arguably violate the former as well as the latter. Turbulent airflow also requires a period of incomplete closure, or continuancy, so we also derive *[-cont, strident] » *[+cont, strident]. Languages that violate the former have strident affricates, which have often been argued to be strident stops (see Lin, this volume, and references therein). In principle, then, the interactions of these constraints might also produce comparable variation to that we have seen with laterals. For other features, no such variation is to be expected, [anterior] and [distributed] refine the type of contact the tip or blade of the tongue makes with the roof of the mouth. As such they can only be present in Coronals, and a sound that is [Dorsal, +ant] is phonetically uninterpretable. The approach taken here is in many ways close to that of Padgett (2000), but there are two differences. First, Padgett uses feature classes, sets of features such as Place, which is defined as the union of the sets Pharyngeal and Oral, where Oral is in turn the union of Labial, Coronal, Dorsal, and V-Place, and V-Place in turn is the union of Height and Color. Lateral is not mentioned in Padgett's paper. I have tried to demonstrate here that even feature classes can be dispensed with. Second, Padgett says explicitly (2000: 99) that feature classes do not overlap, which in our context means that [lateral] cannot be a member both of Place and SV, contra Yip (1990). It is difficult to see how the variable behaviour of [lateral] can be captured if this latter stipulation is maintained, so it would probably be necessary in his approach to loosen the grammar such that the SINGLE MOTHER NODE CONDITION, which bans feature class overlap, is a violable constraint. If one accepts the case made here against a fixed feature geometry, one might ask whether all representational restraints might be relaxed at the level of GEN, so that anything is a possible candidate. The answer is unclear. Consider the prosodic hierarchy. Although moras are usually grouped into syllables before joining any higher level constituent, there are cases where this

88

Moira Yip

has been argued not to be the case, such as Bella Coola (Bagemihl 1991). The constraint that forces moras to be parsed into syllables may thus be extremely low-ranked in this language, and should arguably not be part of GEN. On the other hand reversals of the hierarchy are presumably not found: it is hard to envisage a language in which syllables dominate feet, instead of vice-versa, suggesting that this may be located in GEN. I suspect, however, that this data gap is more apparent than real. The terms mora, syllable, and foot, are just labels for levels in the hierarchy, and could equally well be stated in numerical terms. So really all we are saying is that level η is normally directly parsed into level n+1, and that this constraint seems to be usually high-ranked, but by no means always surface true. It is in fact frequently violated in languages which do not allow degenerate feet, for example, and where the stray syllables (level n) are then directly dominated by the prosodic word (level n+2). A more complex case is presented by languages in which moraic feet could potentially break up syllables, so that a bi-moraic foot would be composed of a monomoraic syllable, plus the first mora of a heavy syllable. The undesirability of this configuration has been dubbed Syllable Integrity, and if this is a universal, then it must somehow be incorporated into GEN. Resolution of these issues is beyond the scope of this paper.

Acknowledgements Thanks to an anonymous reviewer, whose comments greatly improved this paper, and also to members of the audience at OCP 1 in Leiden, the Phonology Reading Group at UCL, and the Department of Linguistics at Edinburgh University, especially Nick Clements, Abigail Cohn, Stuart Davis, Laura Downing, Dorota Glowacka, John Harris, Rene Kager, Robert Ladd, Mario Saltarelli, James Scobbie, Jeroen van de Weijer, and Marc van Oostendorp. All errors are of course my own responsibility.

References Bagemihl, B. 1991 Syllable structure in Bella Coola. Linguistic Inquiry 22: 589-646. Bao, Z.M. 1992 A note on [Lateral], Ms. Ohio State University. Blevins, J. 1994 A place for lateral in the feature geometry'. Journal of Linguistics, 30: 301-4.

Variability in feature affiliations

through violable constraints

89

Brown, Cindy 1995 The feature geometry of lateral approximants and lateral fricatives. In H. van der Hulst and J. van de Weijer, eds. Leiden in Last: HIL Phonology Papers I. Holland Academic Graphics. The Hague, pp 41-88. Chen, M. 1992 The Chameleon [-r] in Yanggu: Morphological Infixation or Phonological Epenthesis? Journal of East Asian Languages 1.2: pp 197-213. Cho, Y.-M.Y. 1988 Korean assimilation. In H. Borer, ed. WCCFL 7: 41-52. Cho, Y.-M.Y. and S. Inkelas 1994 Major class alternations. WCCFL 12: 3-18. Clements, G.N. and E.V. Hume 1995 The internal organization of speech sounds. In J. Goldsmith, (ed.) The handbook of phonological theoiy. Blackwells. Oxford, pp 245-306. Cohn, A. 1992 The consequences of dissimilation in Sundanese. Phonology 9.2: 199220. Davis, S. and S-H Shin 1999 The syllable contact constraint in Korean: An Optimality-theoretic analysis. Journal of East Asian Linguistics. 8.4: 285-312. Dudas, K. 1976 The phonology and morphology of Modern Javanese. Ph.D. dissertation, U. of Illinois. Gafos, A.I. 1996 The articulatoiy basis of locality in phonology. Johns Hopkins University Ph D dissertation. Hayes, B.P. 1986 Assimilation as spreading in Toba Batak. Linguistic Inquiiy 17: 467500. Hegarty, M. 1989 An investigation of laterals and continuancy. Ms., MIT. Hsu, C.S. 1996 A phonetically-based optimality-theoretic account of consonant reduction in Taiwanese. UCLA WPP 92. Hualde, J.I. 1991 Basque phonology. Routledge, New York and London. Kang, H. 2002 On the Optimality-Theoretic analysis of Korean nasal-liquid alternations. Journal of East Asian Linguistics. 11: 43-66. Ladefoged, P., A. Cochran, and S. Disner 1977 Laterals and trills. Journal of the IPA 7: 46-54. Ladefoged, P. and I. Maddieson 1996 Sounds of the World's Languages. Blackwell, Oxford.

90

Moira Yip

Lin, Y.H. this volume Piro affricates: Phonological edge effects and phonetic anti-edge effects? Maddieson, I. 1984 Patterns of sounds. Cambridge University Press. Madenska, L. and Witaszek-Samborska, M. 1998 Zapis fonetyczny. Poznan: Wydawnictwo Naukowe UAM. Mascarö, J. 1976 Catalan phonology and the phonological cycle. MIT PhD Dissertation. McCarthy, J.J. 2002 Comparative Markedness. In A. Carpenter, A. Coetzee, and P. De Lacy, eds. University of Massachusetts Occasional Papers in Linguistics 26: Papers in Optimality Theory II, Amherst, Ma, GLS. [ROA-489], McCarthy, J. J. and A.Taub 1992 Carole Paradis and Jean-Francois Prunet, (eds): the special status of coronals: internal and external evidence. Review in Phonology 9.2: 363-370. Mithun, M. and H. Basri 1985 The phonology of Selayarese. Oceanic Linguistics 25.1 /2: 210-254. Nacaskul, K. 1978 The syllable and morphological structure of Cambodian words. In Jenner, P. ed. Mon-Khmer Studies VII: 183-200. U. of Hawaii Press. Padgett, J. 2000 Feature classes in phonology. Language. 78.1: 81-110. Padgett, J. 1995 Feature classes. In J. Beckman, S. Urbanczyk and J. Walsh, eds. Papers in Optimality Theory. UMOP 18. Pater, J. 1999 Austronesian nasal substitution and other NC effects. In Kager, R., H, van der Hulst, and W. Zonneveld, eds. The Prosody-Morphology Interface. Cambridge University Press, pp 310-343. Piggott, G.L. 1994 Feature dependency in Optimality theory: Optimizing the phonology of sonorants. Ms., McGill University. Pineros, C. 2002 Instability of voiced stops in Palenquero. Ms. Pulleyblank, D.G. 1997 Optimality Theory and features. In D. Archangeli and T. Langendoen, eds., Optimality Theory: an overview. Blackwell, Oxford. Rice, K. D. and P. Avery. 1991 On the relationship between laterality and coronality. In C. Paradis and J.F. Prunet, eds. The Special Status of Coronals. Academic Press. Pp 1124.

Variability in feature affiliations through violable constraints Sagey, E.C. 1986 Shaw, P.A. 1991

Spencer, A. 1984 Steriade, D. 1987

Trigo, R. 1988

The representation of features and relations in non-linear MIT Ph D Dissertation.

91

phonology.

Consonant harmony systems: The special status of coronal harmony. In C. Paradis and J.F. Prunet, eds. The Special Status of Coronals. Academic Press, pp 125-158. Eliminating the feature [lateral]. Journal of Linguistics. 20. 23-43. Redundant Values. In A. Bosch, A. Need and E. Schiller (eds.) Papers from the 23rd Annual Regional Meeting of the Chicago Linguistic Society - Part Two: Parasession on Autosegmental and Metrical Phonology, 339-62. Chicago Linguistic Society, Chicago, Illinois. On the phonological thesis, MIT.

derivation and behaviour of nasal glides. PhD

Tuttle, S. this volume Cryptosonorant phonology in Galice Athabaskan. Uffmann, Ch. this volume Optimal geometries. Walsh, L.D. 1997 The phonology of liquids. UMass PhD Dissertation. GLSA, Amherst. Whitney, William Dwight 1889 Sanskrit Grammar: including both the classical language, and the older dialects, of Veda and Brahmana. 2d (rev. and extended) ed. Leipzig: Breitkopf & Härtel; Boston: Ginn & company. Yip, M. 1990 Two cases of double-dependency in feature geometry. Ms., Brandeis University. Yip, M. 1992 The Prosodic Morphology of Four Chinese Dialects. Journal of East Asian Languages 1.1. pp 1-35 Yip, M. in prep Lateral survival: an OT account. Invited contribution to monograph, Advances in Optimality Theoiy, Paul Boersma and Juan Antonio Cutillas, eds. special issue of International Journal of English Studies. Yip, M. 2003 Some real and not-so real consequences of comparative markedness. Theoretical Linguistics 29, pp 53-64.

The geometry of harmony1 Don Salting

1. Introduction This paper examines neutral vowels in vowel harmony and illustrates five strategies languages will employ with these vowels. The analysis will be couched in a framework that invokes a marriage of an abstract geometric hierarchy and general well-formedness constraints. With this approach, neutral vowels are predicted, and the compensating strategies can be analyzed without recourse to targeted feature constraints. Like the two previous works in this volume, this article addresses the role of representation and feature geometry within/alongside a constraint based framework. Following Uffmann (this volume), I hold that the historical need for conditions (constraints) on representational models does not militate against the entire idea of representations as a component of the phonology. Also in line with Uffmann, I propose that an explanatorily accurate geometry can constrain a constraint-based approach, possibly contributing to fewer conditions on the analysis. I will argue that a failure of geometries was the assumption (a lä SPE) that the smallest components - phonetic gestures - are the primary and universal informants of phonological processes. As with Yip (this volume), a test for the efficacy of a geometry is its ability to account for (and predict) cross-linguistic variability within a constrained framework. Whereas Yip (this volume) argues that the variability of laterals provides evidence against feature geometry, I will suggest, for vowel harmony, that a constrained representational hierarchy can not only account for 1

Portions of this paper have appeared in previous publications. The goal of this paper is to attempt to tie the notions of the paradigm I propose to other frameworks in phonology, specifically ranked constraints. I am grateful to the staff and attendees of WECOL 2002 and of Colloque: theories linguistique et langues subsaharienne at the University of Paris. I am especially grateful to Long Peng, Stuart Davis, and the anonymous reviewer of this article for comments and suggestions. Thank you to Jeroen van de Weijer and Marc van Oostendorp for their support and patience. I am grateful to Harry van der Hulst for materials on RCVP. Any and all errors are mine.

94

Don Salting

but also predict patterns. Fundamental to the model I propose is a rethinking of the nature of distinctive features and their role in informing the phonology. I will argue against the idea of distinctive vowel features as put forth in SPE and will argue that vocalic features are abstract, and their phonetic implementation the by-product of an interface between the segmental inventory and an abstract, universal hierarchy. After explicating the theory, I offer evidence of the predictive and explanatory capability of the hierarchy in the analysis of asymmetries in a number of languages. I will close with some comments about the implications of this approach toward what I argue is the other motivation for harmonic asymmetry, and will suggest a direction for subsequent research.

1.1. Features and Feature Geometry In Feature Geometry theory (hereafter FG) (Clements 1985, Sagey 1986, Steriade 1987, et al.) the goal was to develop groupings and subgroupings of features that would account for linguistic and cross-linguistic phenomena natural classes of features as indicated by phonological phenomena. For the most part, feature geometries relied on the set of features put forth in SPE. The SPE features are almost entirely defined by articulation. In so doing, the geometers were implying that all languages accessed the same articulatory gesture(s) as defmers of phonological phenomena. An explanatory failing of Feature Geometry with SPE features was its inability to capture real world natural classes and their typological variations (see Yip this volume) as witnessed by the variety of geometric configurations to account for vowel harmony (Odden 1991, Cahill 1996, Chumbow 1982, Clements 1991, et al.). With the exception of Clements (1991), all prior geometries employed SPE features. The exit of FG and the advent of Optimality Theory (Prince & Smolensky (1993) hereafter OT) have brought great changes to the mechanisms employed to account for phonological (and syntactic) patterns in languages. For the most part, however, research in vowel harmony still relies on the same SPE features that proved inadequate in feature geometry. A key component to many OT analyses is the targeted markedness constraint. In vowel harmony analyses, targeted constraints are used to account for disharmonic patterns. Some markedness constraints reference basic vowelheight levels (*MID, *HI, *LO in Beckman 1997). Others - Feature Cooccurrence Constraints (hereafter FCC's) - mark certain combinations of features. Padgett (2002), citing Kirchner (1996) argues for *[rd, -hi]. This appears to

The geometry of harmony

95

be the same markedness constraint as *RO/LO in Beckman (1997) from Kaun (1995). Others include * [hi/RTR] (Archangeli & Pulleyblank 1994) and *[lo, ATR] (Bakovic 2002). Kiparsky & Pajusalu (2003) include constraints that target specific segments, such as *[ö,ü]. While the analyses cited here represent the highest standards of scholarly research in our field, I suggest that these constraints are suboptimal for a number of reasons. One argument against such classes of constraints is that, by creating a set of constraints against commonly occurring sounds, the theory becomes less constrained, and thus, less explanatory (Casali 1998). Secondly, these constraints, by design, suffer the same lack of explanatory rigor as was leveled against rule notation: if one can posit a rule A —> Β / C, one can just as easily posit Β —»A/ C. If one can posit *[+rd, +lo], one can just as easily posit *[-rd, +lo]. Though violability and frequency of occurrence can offer a systematic account for typological variation, left unchecked it can create the same empirical conundrum as that criticized in rule notation. Thirdly, the use of combinations of phonetic features to describe phonological phenomena carries a post hoc ergo propter hoc implication that the articulations inform the processes, regardless of the framework. That said, in the analysis I will offer, the only reference to a specific feature will be in the definition of the active harmony feature. I will attempt to account for all other factors by reference to general conditions on well-formedness. The phonetic particulars obtain from the interface of the segmental inventory and a phonological representational hierarchy.

1.2. Abstract Features The Aperture Node model (hereafter AN) of Clements (1991) offers a significant departure from SPE features, positing gradient demarcation of an abstract phonological space with successive degrees of [±open]. Fundamental to this model is the idea that the phonology subdivides the height continuum with no reference to specific articulatory or acoustic parameters. This model accounts for chain-shift (scalar) harmonies in a straightforward manner. However, most harmony systems are not scalar. To account for these, Clements (1991) suggests the notion of a subregister- a bifurcation of one of the points along the continuum. It is implied that the subregisters can vary depending on the language, and most importantly, that stratification is the foundation of phonological division of vowel space. Salting (1998) argues for a different, universal paradigm to account for all non-scalar harmony systems. It is based on notions put forth in Clements (1991):

96

Don Salting

a branching, non-exhaustive hierarchy subdividing the phonological space (in the case, the height continuum). It differs from Clements (1991) in that it is fully binary. The hierarchy is presented in (1) below along with the inventories of languages whose harmony systems the hierarchy can account for. This hierarchy suggests that languages exhibiting non-scalar harmony divide the height continuum, and thus the segment inventory along a relative mid-line ([±openA], and then subdivide each of the 'halves' created by the primary division into secondary subdivisions [±openB]. The secondary subdivisions are 'nested' within the primary divisions of the vowel space, thus, it is called the Nested Subregister model (hereafter NS). In all cases, phonetics is acknowledged in that the inventories are always arrayed highest-to-lowest, left-to-right. I consider the NS model a typological co-variant with the AN model of Clements (1991). The implication is that UG has access to two fixed harmony hierarchies, the Aperture Node model for chain-shift systems, and the Nested Subregister model for non-scalar systems. (1)

Nested Subregister Model Aperture

openB 5 5 5 6 7 7 7 7 7 8 8 9 9 10 12

Chewa, Ganda Shona, Thumbuka Pulaar Obolo Nande N. Salentino, Yoruba Lucanian, Kuyu Braz. Port. Ogori2 Igbo Wolof Maasai, Turkana Okpe, Konni Akan, Igede Lhasa Tibetan



i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u i,u,ü

+

Ι,ϋ

e,o Ι,ϋ Ι,ϋ Ι,ϋ Ι,ϋ Ι,ϋ,ϊ



+

e,o e,o [e,o] e,o [e,o,(a)] e,o e,o e,o ε, ο ε,ο e,o,a e,o e,o e,o,a e,o,ö

a a ε,ο,a a,ο ε,ο,a ε,ο,a ε,ο,a ε,ο,a a a,o ε,ο,a ε,ο,a ε,ο,a ε,ο,a ε,ο,a

The geometry of harmony

97

The interface of four possible heights with different inventory sizes carries interesting theoretical implications. In languages with smaller inventories, one might expect that there is a concomitant lower number of features operative, and thus little or no need for compensating strategies. That is, the active features in such a language would be unlikely to create an unattested feature combination. Specifically, in the 5-vowel Bantu languages listed above, [ATR] ([openB]) does not operate, eliminating the potential for a feature combination that would define the empty slot. This is indeed what we find in the 5vowel inventories in (1). Conversely, in larger inventories ( 9 - 1 2 segments), all possible feature combinations have corresponding segments, eliminating any need for compensating strategies. That is, any feature combination resulting from harmony will have a corresponding segment in the inventory, and this is what we find in the languages with 9-12 segments. It is with the 7-8 vowel inventories that harmony can reference enough features to create a feature combination unattested in the phonemic inventory - a feature combination predicted and defined by an empty slot in the hierarchy in (1). And it is precisely in these languages that we encounter a variety of compensating strategies, which I address in detail in §2 below. Very critically, the Nested Subregister model contains no feature or feature combination analogous to SPE [low]. This has strong theoretical implications for analyzing height harmony (note the markedness constraints that reference [low] in §1.2 above). I address this in §5 below. For the most part, [openA] is analogous to SPE [high] and [openB] is analogous to [ATR], but not always. An example of the divisions not always being analogous to SPE features, and thus not definable by universal phonetic correlates, can be seen in Igbo (Anyanwu 1998). The segmental inventory is phonetically arrayed in (2a) and phonologically arrayed (harmonic pairing) in (2b):

2

Larry Hyman suggests that Ogori vowels might be arrayed as the inventory for Yoruba, which is typologically more frequent. The debate hinges largely on the direction of harmony in Ogori. That is, is [e] the [+ATR] form of /a/, or is [a] the [-ATR] realization of /e/. The former would predict a Yoruba type array, while the latter predicts the array given above. For an explicit analysis using the configuration given above, see Salting (1998a,b). Data are from Chumbow (1982).

98 (2)

Don Salting a.

Igbo Phonetic Segments i I

b.

i I ε a

u υ 0 0

ε

Igbo Harmonic Pairs u υ 0 0

a However, when viewed in light of the NS model (see Igbo in (1) above), the harmonic pairings are a natural outcome of the superimposition of the hierarchy onto the language-specific segmental inventory. This parallels Kim (this volume) in that a phonological constant can subsume varying phonetic implementations cross-linguistically. Further motivation for an abstract phonological paradigm can be found in the discussions on the phonetic and phonological behavior of Bantu vowels. Hyman (1999: §3) citing earlier work by Juliana Kuperus points out that, in Londo (Guthrie A. 11), the orthographic segments e,o "...sound like the [+ATR] vowels usually written e,o, but they function like the [-ATR] vowels written ι,υ." Peng (2000: fnl), citing previous research, mentions the same issues regarding Kuyu. Notice also the array for the Ogori inventory above. I take this to indicate that the phonetics of a segment are not a priori determiners of phonological patterning. 3 From this, it only follows that phonology must access and employ its own determiners. (3)

Feature Geometry Representation of Nested Subregister Model [e] Aperture

I emphasize here that the NS hierarchy in (1) is not feature geometry in its traditional sense; association lines are not exhaustive. A feature geometric

3

A la Clements (1985: 230)

The geometiy of harmony

99

representation of the segment [e] in Lucanian using the NS model would be as in (3) above. The association lines in (1) demarcate categorial divisions of the segment inventory, not groupings of associated features.

1.3. Parallels: Radical CV Phonology A striking parallel to the NS model can be found in Radical CV Phonology (hereafter RCVP) (van der Hulst, forthcoming). Both models suggest fixed, paradigmatic divisions of abstract phonological domains. The features used are also abstract, and their phonetic interpretation can vary across (and in the case of RCVP, within) languages. As with the NS model, the hierarchy in RCVP is a fixed constant, culminating in four possible feature combinations for each domain. RCVP advocates two unary features: C and V. These features coexist in a binary, branching hierarchy which can account for segment Place, Manner and Laryngeal properties, depending on the domain in and of the hierarchy (syllable onset, rhyme, etc.). In terms of vowel height, roughly put, the feature C references closedness (high), and V openness (low). The RCVP hierarchy for vowel height is given in (4) below. The dotted lines demark components not relevant to the current discussion. (4)

RCVP rendition of vowel height (van der Hulst, forthcoming, p. 18) V (Rhyme - Manner)

V (Vowel) (Head)

lo

hi

lo

C (ApproximÄnt) (Dependent)

hi

glide

liquid

glide

liquid

While both RCVP and the NS model present binary branching on every level, the asymmetrical array in the RCVP diagram above marks Head (left set) from Dependent (right set).4 A given vowel will be specified vis ä vis both nodes under the Vowel node. The Head specification of the RCVP hierarchy

4

1 take liberties with RCVP and indicate Dependent features in lower case.

100 Don Salting is exactly analogous to the primary division in the NS model ([openA]) and the Dependent node specifies the subregister - NS [openB]. The most open segment - /a/ - would be specified as V for both nodes, the least open - /i,u/ - would be C for both nodes. Van der Hulst (forthcoming) further restricts the model with the idea of ENHANCEMENT, which states that for a given Head specification, an identical Dependent specification is redundant and can be left unspecified. That is, if the Head is V, a Dependent ν is redundant. A fully specified /a/ would be portrayed Vv, non-redundantly it would be V. As mentioned above, both the NS model and RCVP argue that the phonetic realization of a given feature can vary across languages. Nonetheless, the two models appear to have some different ideas about the height of /a/.5 Following the idee fixe of the NS model that there are only four possible heights, and that there is no single feature analogous to SPE [low], I suggest that the segments /ε,a,a/ are all the same height (the lowest) in most languages. 6 To demonstrate the similarity of the NS model and RCVP, I re-present the NS hierarchy from (1) above employing the features of RCVP. The array in (5) below differs representationally from RCVP in that the Dependent node occurs twice as subsets of the two Head features. It appears that this difference is only cosmetic, however, and has no effect on empirical applications. (5)

Nested Subregister with RCVP features Aperture

openA

5

C

openB

c

ν

c

ν

Maasai

i,u

ι,υ

e,o

e,o,a

Because van der Hulst (forthcoming) considers /a/ to be alone in its height, distinguishing [ATR] properties of vowels requires the adjunction of an additional head to account for all possible Height manifestations. I suggest that this may be suboptimal in that it employs four extra feature specifications to account for a two-way distinction. In the array in (5), [ATR] is addressed in the Dependent specification. 6 1 will make the case in section 5 that /a/ asymmetries are not a product of its lowness but of its marked Place.

The geometry of harmony

101

1.4. Summary of Introduction Arguments against the empirical validity of phonetic features and against targeted constraints have been made. A constrained, abstract hierarchy - the Nested Subregister model - has been proposed and corroboration from another hypothesis - Radical CV Phonology - has been provided. Following the lead of RCVP, but more specifically referencing vowel height, the binary features [±openA] and [+openB] of the NS model could be given unary identities as HIGH/LOW and high/low respectively. These unary labels are more salient and thus easier to remember. Further, there is no empirical basis for choosing binary over unary for the hierarchy in question. The only basis for staying with the binary features is to distance them from any assumptions of phonetic universality associated with SPE [high] and [low]. For that reason alone, I employ [openA] and [openB] for the time being. The paradigm for vowel height in (1) conceptually parallels work by Avery & Idsardi (2001) on laryngeal features. They propose three hierarchical levels of representation: Articulators > Dimensions > Gestures. In their approach, phonological representations access the Dimensions level, which contains "very general phonetic information" (p.41). The model in (1) differs from their approach in that, while their dimensions organize "antagonistic pairs" of gestures, the NS model delineates a more abstract dimension. Most importantly, the phonetics of phonological subdivisions can vary from language to language.

2. Data In this section I examine harmony in five different languages, each with a seven or eight vowel inventory. As mentioned above, when inventories of this size are delineated with the Nested Subregister hierarchy, the resulting empty slot predicts a potential compensating strategy. We will examine five different compensating strategies in some detail. After this I will offer a constraint-based analysis that does not require targeted constraints on specific features or feature combinations, but rather, can account for the featural interaction with general constraints on well-formedness. Kiparsky & Pajusalu (2003) infer that, in harmony, segments are neutral fail to undergo harmony - for one of two reasons: (i) the harmonic counterpart is prohibited - context-free neutralization, or (ii) there is a distributional restriction -positional neutralization. In generative terms, context-free neutralization refers to target vowels only, whereas positional neutralization re-

102 Don Salting fers to restrictions on both trigger and target. I will expand on this delineation and suggest that context-free neutralization in Height harmony can be accounted for via the paradigm above. Most work on neutral segments focuses on instances of segment opacity or segment transparency (Bakovic 2002, Krämer 1998, Padgett 2001, Cole & Kisseberth 1997, Kiparsky & Pajusalu 2003, Beckman 1997, et al.). Bakovic (2002) paralleling earlier work by Calabrese mentions a third possibility, a strategy he terms re-pairing in which surface forms alter in more than one feature specification (the harmonic feature plus another change). I submit that there are at least five types of pattern that can emerge from the process of harmony defining the empty slot. I call them Compensating Strategies and they are: opacity, transparency, new segment, cleanup and epenthesis. We will examine five languages that exhibit these strategies, and in every case, the hierarchy in (1) correctly predicts the languages and segments within them that will manifest in compensating strategies. I repeat the hierarchy in (6) with the pertinent languages. (6)

Nested Subregister Model Aperture

openB 7 7 7 8



Kuyu N. Salentino, Lucanian Nande Wolof

2.1. Opacity - KiKuyu

i,u i,u i,u i,u

+



e,o e,o I,υ e,o,a

+ e,o,a e,o,a e,o,a ε,o,a

/i,e,e,a,o,o,u/

Kuyu has a seven-vowel inventory. Common to Bantu languages, derivational suffixes in the verb phrase will harmonize with the root vowel. In Kuyu, the operative feature is [openB] (SPE [ATR]). In (7) below, the Applicative and Tolerative suffixes contain [e] (7a-e) or [ε] (7f,g) dependent upon the [openB] status of the root.

The geometry of harmony (7)

103

Kuyu front vowel suffixes ROOT + APP + TOL + vrb suff (Peng 2000) a. b. c. d. e.

ι u e 0 a

tiy-er-ek-a tum-er-ek-a yer-er-ek-a ho0-er-ek-a ßa6-er-ek-a

'abandon or be left over' 'join, intrude' 'have something fetched for' 'be used' 'become rich'

f.

ε 3

tem-er-ek-a ßoy-er-ek-a

'cut into specific shapes' 'calm down, slow down'

Ε-

Opacity in Kuyu obtains when a high vowel /i,u/ occurs within a suffix (8a) or in the root (8b). That high vowel not only fails to undergo harmony, but also blocks harmony to subsequent segments. The Kuyu pattern is segmentally the same as that in Yoruba (Archangeli & Pulleyblank 1989). In NS terms, the operative harmony feature is [+openB] (-ATR). The high vowels have no [+openB] counterpart in the inventory (see (6)), and thus, harmony is blocked. (8)

Kuyu Opaque High Vowels a. b.

i i u u

tern - ίθ - er - a ßoor - ϊγ - er - a eqgut-er-a hooruh - er - a

*tem-i0-er-a

'become cutable' 'feel dizzy at' 'move away for' 'become blistered for'

In item (7e) above, the root vowel /a/ does not trigger lax vowels in the suffixes, though featurally it should. I take this as evidence that the input form for the suffixes is /-er-ek-/. Because /a/ is a trigger vowel in this, its neutrality is motivated differently than the segments in question. Its behavior here is not neutrality, but asymmetry, which I define and address in §5.

2.2. Transparency - Wolof

/i,e,3,e,a,o,o,u/

As with Kuyu, the inventory of Wolof contains an empty node in the NS hierarchy and also like Kuyu, the vowels /i,u/ do not harmonize. In this case, they will trigger harmony if they occur in the root, but will be transparent to harmony if in a derivational suffix. As with Kuyu, the operative feature is [openB] (ATR). In (9) below, the root controls harmony in the 2pp Imperative

104 Don Salting suffix -lEEn. In items (a-d), the root vowel is [-openB], and all subsequent vowels are tense. In (e-h), the root vowel is [+openB] and this quality spreads to the suffix unaffected by the intervening high vowel. (9)

Wolof(Ka 1993: 31) a.

musluleen

/musl-u-lEEn/

b.

gimmileen

/gim-mi-lEEn/

c.

letuleen

/let-tu-lEEn/

d.

toxileen

/tox-i-lEEn/

e.

seetuleen

/seet-u-lEEn/

f.

tekkileen

/tek-ki-lEEn/

g.

wooyileen

/wooy-i-lEEn/

h.

waxsileen

/wax-si-lEEn/

protect-Refl.-2pp Imp. 'protect yourselves!' c/o.se-Revers. -2pp Imp. 'open your eyes!' /?ra;i/-Refl.-2pp Imp. 'braid your hair!' smoke-away-2pp Imp. 'go and smoke! look ai-R.efl.-2pp Imp. 'look at yourselves!' ?/e-Reversive-2pp Imp. 'untie!' ca//-away-2pp Imp. 'go and call!' say- to w ard s - 2 ρ ρ Imp. 'come and say!'

At this point I address the difference between opacity and transparency. In both cases above, the target vowel fails to surface with the harmonic featural specification. What differs is whether or not harmony 'continues beyond' the non-participating segment. This raises issues of domain, direction, iteration, locality and mechanism (spreading, alignment, agreement) (Bakovic & Wilson 2000, Ni Chiosäin & Padgett 2001). Thus, the difference between transparency and opacity is arguably not featural, but rather processual. OT literature has produced a variety of constraints to define harmony including: A G R E E [ F ] (Bakovic 2 0 0 2 , Krämer 1 9 9 8 , Kiparsky & Pajusalu 2 0 0 3 ) ; SPREAD[F] (Padgett 2 0 0 1 ) ; IDENT (POSITIONAL FAITHFULNESS) (Beckman ( 1 9 9 7 ) ; EXPRESSION (Cole and Kisseberth 1 9 9 7 ) , etc. The purpose of this paper is to offer a theory of features that will account for harmony and asymmetry without invoking targeted constraints on features and feature combinations, and from that perspective, opacity and transparency are variations of the same process, apparently differentiated by domain. However, in §3.2 a case is made that the difference is the product of antagonism between harmony and faithfulness.

The geometiy of harmony 2.3. New Segment - Nande

105

/i,i,e,a,D,u,u/

Nande exhibits two different harmony patterns. Like Kuyu above, Nande is a Bantu language, and derivational suffixes in the verb phrase harmonize with the vowel of the root. As can be seen in (11 e,f) below, the Applicative suffix will harmonize to the [openA] (hi) status of the root vowel. In addition, all vowels except /a/ harmonize for [ openB] (+ATR) if a tense vowel occurs anywhere in the verb phrase. 7 Of interest are items (11 a-b) below. The surface vowel [e] in the Infinitive [eri-] does not occur in the phonemic inventory. In other words, as a compensating strategy, the unattested feature combination created by harmony (describing the empty slot in the hierarchy) manifests in a non-phonemic segment. To militate against this rather rare occurrence in harmony, I suggest a highly ranked general constraint INVENTORY: (10)

INVENTORY: Every F combination of the output will have a corresponding F combination in the phonemic inventory, (no new segments)

(11)

Kinande (Mutaka 1995, Hyman 1989) ROOT V INFINITIVE a. i eri-ßin-a 'to dance' b. u eri-kümb-a 'to grasp'

7

Applicative Suffix ir / er APPLICATIVE eri-ßin-ir-a 'to dance for' eri-kümb-ir-a 'to grasp for'

c. d.

I υ

επ-lim-a επ-tum-a

'to work' 'to send'

επ-lim-ir-a επ-tüm-ir-a

e. f.

ε 0

εή-hek-a επ-bg-a

'to carry' επ-1ιε^εΓ-3 'to bewitch' er'i-bg-er-a

'to carry for' 'to bewitch for'

g-

a

επ-hat-a

'to peel'

'to peel for'

επ-hat-ir-a

'to work for' 'to send for'

The Causative suffix also contains a tense vowel, and harmony triggered by it will include the root vowel: /en-rim-a/ 'to cultivate' [eri-rim-is-i-a] 'to cause to cultivate'. This has no effect on our analysis so is not included for the sake of space.

106 Don Salting 2.4. Cleanup - Lucanian

/i,e,e,a,o,o,u/

The following two languages - Lucanian and Northern Salentino - are dialects of Italian. In them, plural nouns and masculine adjectives exhibit metaphony, which Walker (2001) analyzes as the raising of a stressed stem vowel in the context of an unstressed high-vowel suffix. In (12) below, the raised forms of /ε,ο/ are [i,u] respectively. For the /ε,ο/ segments to surface as [i,u], an additional feature change must take place. They must both 'change' in the specification for [openB] as well. Without this extra change, harmony would describe the empty slot. (12)

Lucanian Metaphony (Calabrese 1987) SING.

a. ρέδε b. δέηδε

PLURAL

ρίδί δίηδϊ

'foot' 'tooth'

2.5. Epenthesis - Northern Salentino

FEM SING.

MASC SING,

c. d. e.

bbilu nuvu lintu

bb£lla nova tenta

/i,e,e,a,o,o,u/

Northern Salentino is another dialect of Italian with the same inventory as Lucanian and with the same conundrum posed by metaphony. That is, the harmonic feature, when applied to the lax vowels /ε,ο/ (13 below), creates a feature combination that describes the empty slot in the hierarchy. Whereas in Lucanian this is circumvented by an additional feature change in the output, in Northern Salentino, an epenthetic segment is introduced so that harmony may occur without violating the phonemic inventory or invoking any secondary feature changes. (13)

Northern Salentino (Calabrese 1987) PLURAL 8 NOUNS

a. /pet-i/ b. /dent-i/ c. /kor-i/

8

[pieti] [dienti] [kueri]

MASC. ADJECTIVES

'feet' 'teeth' 'hearts'

d. /lent-u/ e. /b6n-u/ f. /mort-u/

[lientu] 'slow' [buenu] 'good' [muertuj 'dead'

The fact that both the singular and plural forms end with /—i/ but only the plural form exhibits harmony could be evidence of a morphemic floating feature (see Salting 1998a and Archangeli & Pulleyblank 1994 for other cases).

The geometry of harmony

107

3. Analysis In the five languages above, harmony creates a feature combination unattested in the phonemic inventory but nonetheless delimited by the hierarchy in (1). This unattested combination defines an empty slot in the hierarchy. That empty slot predicts that in these languages, harmony will invoke a compensating strategy and we have examined five different strategies in some detail: opacity, transparency, segment creation (new F combo), cleanup, and epenthesis. In this next section, I offer an analysis of this cross-linguistic variation as a marriage of the representational hierarchy in (1) and general constraints. Rather than explicitly describe the harmony process as a constraint in the tableaus, I will describe harmony in each language in detail here. In the tableaus, a general constraint - HARMONY - will substitute. (14)

Harmonies a. b. c. d. e.

Kuyu: Wolof: Nande: Lucanian:

Root controlled, suffixes agree for [openB] (ATR). Root controlled, suffixes agree for [openB]. Feature driven, [-openB] (+ATR). Morphemic [-openA] (+hi) unstressed final vowels trigger raising in stressed, word-internal vowels. Northern Salentino: Morphemic [-openA] (+hi) unstressed final vowels / floating feature trigger raising in stressed, word-internal vowels.

3.1. Constraints As mentioned in § 1.2, the goal is to eliminate reference to distinctive features other than in the definition of harmony itself. All other constraints will reference general properties. That said, I reintroduce the constraint from (10) above: (15)

Every F combination of the output will have a corresponding F combination in the phonemic inventory, (no non-phonemic segments) INVENTORY:

This militates against a surface segment that does not occur in the phonemic inventory. I also include D E P - I O which references the potential of epenthesis.

108 Don Salting (16)

DEP-IO:

Output segments must have input correspondents (no

epenthesis) I include a general faithfulness constraint on featural correspondence: (17)

IDENT-IO[F]:

Correspondent segments have identical values for

[F]

In this paper, I D E N T - I O [ F ] does not include epenthetic segments (no input correspondence) but does include both feature changes described in HARMONY as well as any others. Because some languages (here, Lucanian) exhibit feature changes other than that specified in Harmony, I include a constraint against non-harmonic feature changes. I call this constraint IDENT-IOEH. (18)

(Extra-Harmonic): with the exception of feature conditions specified in Harmony, correspondent segments have identical values for [F] (no cleanup)

IDENT-IO-EH

3.2. OT Analyses I assume an underlying form and that harmony invokes the surface form. While I acknowledge that underlying segments may be underspecified or unspecified (Wolof?) for a given feature, these issues do not appear to be directly relevant to the analyses here. That said, I offer tableaus of the languages presented in this paper below: (19)

Kuyu 9 - Opacity

/tem - ίθ - er - a/

INVENTORY

DEP-IO

a. tern - ίθ - er - a

IDENT-IO

IDENT-EH

*

!*

ι®" b. tem - ίθ - er - a c. tern - ίθ - er - a d. tem - ίεθ - er - a e. tem - εθ - er - a 9

HARMONY

* *

* *

!* !*

1***

Φ

In the Bantu examples - Kuyu and Nande - I do not consider word-final /a/ as it complicates but does not affect analysis.

The geometry of harmony (20)

109

Wolof - Tranparency

/seet - u - lEEn/ seet - Ϋ - Ιεεη

INVENTORY

DEP-IO

IDENT-EH

HARMONY

IDENT-IO

*

* *

!*

ι®" seet - u - Ιεεη β ε ε ΐ - u - leen

* *

βεεΐ - uo - leen

* *

!*

βεεΐ - ο - Ιεεη

!*

In the two languages above, the antagonistic relationship between HARMONY and I D E N T - I O [ F ] seems to differentiate between opacity and transparency. I leave a definite statement on this to future work. (21)

Nande - New Segment

/εή - kiimb - ir - a/

DEP-IO

HARMONY

IDENT-EH

επ - kiimb - ir - a TS" eri - kiimb - ir - a

(22)

!*

INVENTORY

DEP-IO

ρέδί

HARMONY

ρίέδί

!*

piöi

IDENT-IO

*

* *

*

* *

* *

!*

Northern Salentino - Epenthesis

/bon-u/

INVENTORY

bonu

HARMONY

IDENT-EH

DEP-IO

IDENT-IO Φ

!*

Φ

TS" bu6nu

bunu

IDENT-EH

!*

ι» piöi

bunu

Φ

!*

Lucanian - Cleanup

/ρέδ-ί/

(23)

IDENT-IO

*

irl - kiimb - ir - a ίεπ - kiimb - ir - a

INVENTORY

1***

!* !*

* *

110 Don Salting The tableaus above can be schematized as follows: (24)

Constraint Schema a. Kuyu

opacity

INVENTORY, D E P - I O , I D E N T - I O [ F ] » I D E N T - E H , HARMONY

b. Wolof

transparency INVENTORY, D E P - I O , I D E N T - E H » I D E N T - I O [ F ] , HARMONY

c. Nande

new segment D E P - I O , HARMONY, I D E N T - E H » INVENTORY, I D E N T - I O [ F ]

d. Lucanian

cleanup

INVENTORY, D E P - I O , HARMONY » I D E N T - E H , I D E N T - I O [ F ]

e. N. Salentino

epenthesis

INVENTORY, HARMONY, I D E N T - E H » D E P - I O , I D E N T - I O [ F ]

As can be seen, the varying patterns in the harmony systems of the above languages, when viewed in the context of a paradigmatic motivation for distinctive features at the phonological level (not the phonetic), can be effectively accounted for without recourse to targeted constraints against specific features or feature combinations. What supports the NS hierarchy is (i) the fact that the empty slots predict compensating strategies, and (ii), that the hierarchy and a small set of general constraints on well-formedness can account for the cross-linguistic variation.

4. Other Implications of the NS Model: Place harmony Recall from §1 above that Clements (1991) argues for a scalar delineation of vowel height as a universal with the modification of additional subregisters for non-scalar harmony systems. With the NS model I offer a variation to the idea, proposing as a universal the binary division and subdivision of the vowel height continuum along language-specific lines (segment inventory). I suggest here that corroboration for the universality of the NS model can be seen in Place harmony systems. I offer the inventories of some languages that exhibit Place harmony in (25) below. Note that there do not appear to be languages with chain-shift Place harmony10 - evidence against the universality of a scalar division of the space (Clements 1991) - whereas the one paradigm of the NS model can account for both cross-linguistic variety and multiple dimensions. 10

1 am grateful to Long Peng of State University of New York, Oswego for this observation.

The geometry of harmony (25)

111

Nested Subregister Hierarchy for Place Harmony Systems Place

backB Somali Turkish Hungarian Finnish Lhasa Tibetan Class. Mongolian Eastern Cheremis

-

+

+

i,e,ae i,e i,e i,e,ä ί,ε,ι,ε i,e i,e

ι,ε,α

ϋ,Ο

u,o

ü,ö ü,ö ü,ö ü,ö ü,ö ü,ö

t,Q

U,0

α α ί,α α a?

u,o U,0 U,0,ü,D U,0 U,0

With the exception of Somali, the feature [backA] is analogous to SPE [back] while [backB] is analogous to [round]. For most languages, [round] defines the subregister. That is, [round] distinguishes the subsets within each primary division. For Somali, it's more complex; [round] is a co-occurring property of one of the primary divisions ([+backA]) of the horizontal axis. While [backA] still serves as the major divider of the front-back axis, the feature [round] is also subsumed within it." As with the NS model for Height, in the above representation, phonetics is acknowledged by the phonology in that the inventory is arrayed front-to-back, left-to-right. However, abstraction is also evident in the cross-linguistic variation between the [round] and [back] implications of the nodes (Somali vs. the rest). Importantly, one hierarchy accounts for all variation with no recourse to self-modification.

4.1.RCVP Revisited I revisit RCVP (§1.3 above) here because an important parallel occurs in the Place systems. Recall that RCVP employs two features - C and V - to ac-

" See Salting (1998a) for a detailed explication of Somali harmony.

112 Don Salting count for all phonological dimensions. For vowel Place, C corresponds to Front, and V corresponds to both Back and Round. Specifically, the Head V corresponds to [back] and the dependent ν corresponds to [round]. This is illustrated in (26) below from van der Hulst (forthcoming). (26)

RCVP rendition of vowel Place, (van der Hulst, forthcoming: p.22) V (Rhyme - Place)

V

front

(Vowel)

back

front

round

When the featural notions in (26) are incorporated into the organizational paradigm of the NS model, the constituents are arrayed the same: (27)

Place

Head

C Front

V Back

Dependent

c front

ν round

c front

ν round

Somali Turkish Hungarian Finnish Lhasa Tibetan Class. Mongolian Eastern Cheremis

i,e,ae i,e i,e i,e,ä ί,ε,ι,ε i,e i,e

ι,ε,α ü,ö ü,ö ü,ö ü,ö ü,ö ü,ö

tt,o

u,o u,o u,o u,o

t,Q

α α t,a α a?

U,0,ü,0

u,o u,o

Notice here also that the secondary cv tier does not accurately capture the [round] distinctions for Somali. However, I do not see this as problematic for the RCVP model. Given the inherent abstractness of the features in RCVP, the language-specific allowance would be parametric, and does not conflict

The geometry of harmony

113

with previous notions about the nature of the features. That is, V and ν still reference backness, the difference is in the register that subsumes roundness. Again, the parallels between the NS model and RCVP suggest glimmerings of universal representation. The case is made that an approach that treats phonological dimensions as abstract but highly constrained can account for linguistic and cross-linguistic patterns, and can do so in a straightforward manner. The important point to be drawn here is that two different approaches with differing pedigrees interface perfectly in their descriptions of cross-linguistic patterns.12

5. Further Research: Asymmetry and the X/Y Factor The languages in §2 and §3 involve neutral target vowels where harmony creates an unattested F-combo for that language's segmental inventory. Here I distinguish between neutrality and asymmetry. I define asymmetry as the absence of harmony even though the trigger and target vowels are appropriately specified in input for harmony (see also Cole & Kisseberth 1997 and Kiparsky & Pajusalu 2003 et al.). Asymmetry is demarked by a feature in the opposite axis of the one that defines harmony. Height harmony will be blocked by a Place specification in the target-trigger relation, and Place harmony will be blocked by a Height specification. In other words, harmony on the Y axis is informed by conditions on the X axis and vice versa. In the majority of cases, it is the specification of the trigger that delimits asymmetry.13 I suggest here that this process also informs ubiquitous /a/ opacity as a trigger in situations where, by its featural makeup, it should trigger harmony (see (22a) below). I argue that it is not the [low] specification of /a/ but rather its marked Place configuration. A parallel to (28a) can be seen in Bantu Reversives (28b) where the target and trigger must agree in Place for Height harmony to occur.

12

1 see the NS model as a natural extension of the Aperture Node model of Clements (1991), whereas van der Hulst (forthcoming) acknowledges the influence of John Anderson and Dependency Phonology as a primary informant of RCVP. 13 One exception can be seen in Turkish harmony where the height of the target demarks [round] harmony.

114 Don Salting (28)

Height Harmony Asymmetries a.

Low vowel triggering opacity - Nande Applicative -ir / -er /eri-hat-a/

[επ-hat-ir-a]

*[eri-hat-er-a] 'to peel for'

b. Bantu Reversive - Thumbuka -ul / -ol (personal field notes) 'to blanch' 'to untie'

i. /i/ ku-ind-ul-a ii. /u/ ku-sut-ul-a iii. lol ku-soqg-ol-a iv. Id ku-cek-ul-a v. /a/ ku-land-ul-a

* [ku-cek-ol-a]

'to sharpen a point' 'to grow old' 'to explain'

A mirror image to this pattern obtains in Place harmony systems. In Turkish, high-vowel suffixes will harmonize for both [back] and [round] but non-high suffixes only for [back]. In Yawelmani (Cole & Kisseberth 1 9 9 7 ) , if the suffix is underlying high, only a high root vowel will occasion Place harmony, if the suffix vowel is non-high, only a non-high root vowel will occasion Place harmony.14 In other words, root and suffix must agree in Height for Place harmony to obtain. In (29) below I explicate these conditions for a sampling of languages. In the majority of the cases in (29), the necessary conditions are that target and trigger must agree on one axis for harmony to occur on the other. Cole & Kisseberth ( 1 9 9 7 ) suggest a family of constraints on UNIFORMITY to analyze this in Yawelmani. It is beyond the scope of this paper, but I would like to suggest that, though these cases are motivated by featural specifications in the participants, these cases can be analyzed as well without recourse to targeted constraints. It will have to wait on further research.

14

Interestingly, the mid vowels /e,o/ pair phonologically with /a/ in Yawelmani. I take this as further evidence against the idea of SPE [low].

The geometry of harmony (29)

115

The X/Y Factor

HEIGHT HARMONY Thumbuka: trigger & target must be [around] for [-hi] harmony (reversive suffix) trigger & target must be [around] for [-hi] harmony Nande: (reversive suffix) for [-ATR] harmony trigger & target must be [around] Kuyu: (reversive suffix) PLACE HARMONY Okpe: trigger & target must be [ahigh] for Place harmony (past tense suffix) Okpe: trigger & target must be [ahigh] for [-round] harmony (infinitive suffix) Yawelmani: target & trigger must be [ahigh] for Place harmony (deriv. suffixes) Ogori: the trigger must be [+hi] for Place harmony to occur (reduplication) Turkish: the target must be [+hi] for [+rd] harmony to occur (inflectional suffixes) Most importantly, the robustness of this axis-interface provides a compelling arena for analyzing /a/ asymmetries (7e, 28a above) cross-linguistically. The elimination of notions of [low] as a definer of /a/ forces reanalysis of its ubiquitous asymmetry. When viewed in light of the data above, the highly marked Place configuration of /a/ makes it a natural candidate for X/Y asymmetry.

6. Conclusion 1 have demonstrated that harmony in a number of languages can be analyzed as the interface of a highly constrained but abstractly motivated feature hierarchy in concert with general well-formedness constraints. The analysis did not require recourse to targeted constraints on feature cooccurrence. In addition, the hierarchy used to define vowel features required no theory-internal modification to account for cross-linguistic variability. The Nested Subregister model suggests, and the data support an abstract but highly constrained phonological division of the vowel space for languages that exhibit harmony. The phonetic specifics of vowels are determined in part by the size of the segmental inventory and its dispersion in the vowel space to facilitate great-

116 Don Salting est salience. The physical factors that define harmony patterns result from the interface of the segmental inventory and the abstract, harmonic division and subdivision of the vowel space. Parallels with the notions put forth here occur in multiple treatments and multiple domains. As has been discussed several times, the parallels between the NS model and Radical CV Phonology (van der Hulst, forthcoming) are striking. Both models invoke a single binary, branching hierarchy of abstract elements to account for multiple patterns and definitions. Further, both models recognize an ultimate array of four possible feature combinations to define a dimension. The key idea to both is that the nature of subsegmental factors, though ultimately physical and measurable, is essentially abstract. Nonetheless, these abstract parameters are couched within a very constrained framework and ultimately reflective of universal phonetic parameters. When features are considered as abstract components with room for cross-linguistic variation factored in, entire classes of analytical machinery are rendered unnecessary. Here, I am specifically referring to the notion of targeted constraints against features and feature combinations. Parallels outside phonology also lend support to the ideas presented here. Hopper & Thompson (1984) argue that the morphological and syntactic properties of a word are equally if not more determined by its position in discourse than by its lexical category. The parallel with this paper is that phonetic descriptors are in part the product of a segment's relative position in an inventory, and not "a priori considerations of vocal tract anatomy and the like" Clements (1985: 230). Another possible parallel can be seen in Levinson & Meira (2003) which explores universale in the semantics of adpositions across languages. The article suggests the possibility of "prototype categories with elastic, ill-defined boundaries" as a tool for capturing cross-linguistic variation, (p.511) Paraphrasing the conclusion, literal assumptions (read 'phonetic features') fall down in the presence of typological variation. However, this typological variation is an important tool in discovering universal tendencies. Though in a somewhat different sense, Levinson & Meira (2003) suggest "...fractionation of categories formed over primary foci in a fixed ... conceptual space may prove a very helpful heuristic for understanding the patterns in the diversity" (p. 513). The descriptive and explanatory efficacy of the NS model indicates that the hierarchical configuration of the "fractionation" may be the universal.

The geometry

of harmony

117

References Andrzejewski, B.W. 1955 The problem of vowel representation in the Isaaq dialect of Somali. Bulletin of the School of Oriental and African Studies University of London 17.3: 567-580 Anyanwu, Rose-Juliet 1998 Aspects of Igbo Grammar. Hamburg: Lit Verlag. Archangeli, Diana, & Douglas Pulleyblank 1994 Grounded phonology. Cambridge, MA: MIT Press. Archangeli, Diana, & Douglas Pulleyblank 1989 Yoruba vowel harmony. Linguistic Inquiry, 20.2: 173-217. Avery, Peter & William J. Idsardi 2001 Laryngeal dimensions, completion and enhancement. In Hall (ed.) Distinctive Feature Theory. 41-70. Berlin: Mouton de Gruyter. Bakovic, Eric 2002 Vowel harmony and stem identity. Ms. UCSD. Rutgers Optimality Archive. Bakovi6, Eric & Colin Wilson 2000 Transparency, Strict Locality, and Targeted Constraints. WCCFL 19:43-56. Beckman, Jill 1997 Positional faithfulness, positional neutralization and Shona vowel harmony. Phonology, 14: 1-46. Cahill, Mike 1996 ATR harmony in Kooni. OSU Working Papers in Linguistics, 48:13-30. Calabrese, Andrea 1987 The interaction of phonological rules and filters. NELS, 17:79-98. Casali, Roderic F. 1998 Predicting ATR activity. In M. Catherine Gruber Derrick Higgins, Kenneth S. Olson and Tamra Wysocki (eds.), CLS 34: 55-74. Chumbow, B.S. 1982 Ogori vowel harmony: an autosegmental perspective. Linguistic Analysis, 10.1: 61-93. Clements, George N. 2001 Representational economy in constraint-based phonology. In Hall (ed.) Distinctive Feature Theory. 71-146. Berlin: Mouton de Gruyter. Clements, George N. 1991 Vowel height assimilation in Bantu languages. Working Papers of Cornell Phonetics Laboratory, 5: 37-74. Clements, George N. 1985 The geometry of phonological features. Phonology Yearbook, 2:225-252. Clements, George N. & Elizabeth Hume 1995 The Internal Organization of Speech Sounds. In John Goldsmith (ed.)

118

Don

Salting

The Handbook of Phonological Theory. Blackwell, Cambridge, Mass. 245-306. Cole, Jennifer & Charles Kisseberth 1994 An optimal domains theory of harmony. Studies in the Linguistic Sciences (24)2. Cole, Jennifer & Charles Kisseberth 1997 Restricting multi-level constraint evaluation: opaque rule interaction in Yawlemani vowel harmony. In K. Suzuki and D. Elzinga (eds.) Proceedings of the Arizona Phonology Conference. 18-38. Faraclas, Nicholas 1984 A grammar of Obolo. Bloomington, Indiana: Indiana University Linguistics Club. Ghini, Mirco 2001 Place of articulation first. In Hall (ed.) Distinctive Feature Theoiy. 147176. Berlin: Mouton de Gruyter. Hall, T. Alan. (Ed.) 2001 Distinctive Feature Theoiy. Berlin: Mouton de Gruyter. Harris, John & Geoff Lindsey 2000 Vowel patterns in mind and sound. In N. Burton-Roberts, R Carr & G. Docherty (eds.) Phonological knowledge: its nature and status. 185205. Oxford: Oxford University Press. Hayes, Bruce 1996 Phonetically driven phonology: the role of optimality theory and inductive grounding. Ms. for proceedings of The 1996 Milwaukee Conference on Formalism and Functionalism in Linguistics. Hopper, Paul J. and Sandra A. Thompson 1984 The discourse basis for lexical categories in universal grammar. Language 60.4: 703-752. Hulst, Harry van der Forthcoming Molecular Structure of Phonological Segments. Hyman, Larry 2002 'Abstract' vowel harmony in Kälöng: in support of underlying representations. Paper to appear in Proceedings of Colloque: theories linguistique et langues subsaharienne. Universite de Paris 8. Hyman, Larry 1999 The historical interpretation of vowel harmony in Bantu. In Hombert & Hyman (eds.) 1999, Recent Advances in Bantu Historical Linguistics. Stanford: CSLI. Hyman, Larry 1989 Advanced tongue root in KiNande. Ms., UC Berkeley. Ka, Omar 1993 WolofPhonology and Morphology. Lanham, Maryland: University Press of America.

The geometry of harmony 119 Kager, Rene 1999 Optimality Theory. Cambridge University Press. Cambridge, UK. Kaun, Abigail 1995 The typology of rounding harmony: an Optimality Theoretic approach. PhD dissertation, UCLA. Kaze, Jeffery 1989 Metaphony in Italian and Spanish Dialects Revisited. Ph.D. dissertation. Champaign-Urbana: University of Illinois. Kiparsky, Paul & Karl Pajusalu 2003 Towards a typology of disharmony. Ms., Stanford University. Krämer, Martin 1998 A correspondence approach to vowel harmony and disharmony. Ms. Heinrich-Heine-Universität Düsseldorf and University of Ulster. Laniran, Yetunde 1985 Vowel merger and Emälhe vowel harmony. Journal of the Linguistic Association of Nigeria 3: 3-11. Levinson, Stephen, Sergio Meira, and The Language Cognition Group 2003 'Natural concepts' in the spatial topological domain - adpositional meanings in crosslinguistic perspective: an exercise in semantic typology. Language 78.3: 485-516. McCarthy, John and Alan S. Prince 1995 Faithfulness and identity in prosodic morphology. Ms. excerpted from Beckman, et al. (eds.) 1995, University of Massachusetts Occasional Papers in Linguistics, 18: Papers in Optimality theory, Amherst: Graduate Linguistics Student Organization. Mutaka, Ngessimo M. 1995 Vowel harmony in KiNande. The Journal of West African Languages. 25.2:41-55. Nornang, W. 1978 Tibetan vowel process: a case for the feature advanced tongue root. In D. Malsch, J. Heard and C. Sleat (eds.) Proceedings of the 8th Annual Meeting of the Western Conference on Linguistics. 106-112. Odden, David 1991 Vowel geometry. Phonology, 8: 261-289. Padgett, Jaye 2001 The unabridged feature classes in phonology. Ms. to appear in Language. University of California, Santa Cruz. Peng, Long 2000 KiKuyu vowel harmony. South African Journal of African Lingusitics, 20.4: 370-384. Polgärdi, Krisztina 1998 Vowel harmony. An account in terms of government and optimality. Ph.D. dissertation, Leiden University.

120 Don Salting Pulleyblank, Douglas 1986 Underspecification and low vowel harmony in Okpe. Studies in African Linguistics 17( 1): 119-153. Pulleyblank, Douglas 1994 Vowel harmony and optimality theory. Ms. to appear in Proceedings of Workshop on Phonology. University of Combra, Portugal. Pulleyblank, Douglas 2001 Defining features and constraints in terms of complex systems: Is UG Too Complex? Paper given at Workshop on Early Phonological Acquisition,, Carry-le-Rouet, Marseilles, France. Saeed, John I. 1993 Somali Reference Grammar. Dunwoody Press, Kensington, MD. Sagey, Elizabeth 1986 The Representation of Features and Relations in Non-Linear Phonology. PhD dissertation: MIT. Sagey, Elizabeth 1990 The Representation of Features in Non-Linear Phonology. New York: Garland Publishing. Salting, Don 1998a. The nested subregister model of vowel height. Ph.D. dissertation, Indiana University. Salting, Don 1998b. Vowel height: reconsidering distinctive features. Proceedings of the 24th Annual Conference of the Berkeley Linguistic Society . Salting, Don 2002 Traits commandes par un inventaire pour la hauteur vocalique. Proceedings of Colloque: theories linguistique et langues subsaharienne. Universite de Paris 8, St. Denis. Salting, Don To appear Toward a typology of vowel height features. WECOL 2002. Schlindwein Schmidt, Deborah 1996 Vowel raising in Basaa: a synchronic chain shift. Phonology, 13: 239267. Steriade, Donca 1987 Locality conditions and feature geometry. NELS 17: 595-617. Stewart, John 1983 Akan vowel harmony: The word structure conditions and the floating vowels. Studies in African Linguistics, 14: 111-139. Walker, Rachel 2001 Two kinds of vowel harmony in Italian dialects. Ms. UC Irvine. Wetzeis, W. Leo 1995 Mid-vowel alternations in the Brazilian Portuguese verb. Phonology, 12: 281-304.

Piro affricates: Phonological edge effects and phonetic anti-edge effects1 Yen-Hwei Lin

1. Introduction In current phonological theory, how to formally characterize affricates is a debated issue (Clements and Hume 1995:256). There are essentially three different views regarding the phonological representation of affricates. The first treats affricates, at both phonological and phonetic levels, as contour segments that contain ordered stop and fricative units or ordered [-cont] and [+cont] specifications (Clements and Keyser 1983, Sagey 1986).2 On the second view, the [-cont] and [+cont] components of affricates are phonologically unordered but become ordered phonetically (Hualde 1988, Lombardi 1990, 1995, Schäfer 1995, van de Weijer 1996). The third view is referred to as the Stop Hypothesis: phonologically affricates are strident stops or stops contrasting with plain stops in place features but a fricative release or the [+cont] feature is added at the phonetic level (Jakobson, Fant and Halle 1952, Steriade 1989, 1993, Shaw 1991, Rice 1994, LaCharite 1993, Rubach 1994, Kim 1997, Clements 1999, Kehrein 2002).3 In all three proposals, an affricate phonetically has a fricative release component on the right edge although they differ in the phonological representation of affricates. These proposals make different predictions regarding EDGE EFFECTS, which refer to cases where affricates behave as stops with regard to rules sensitive to their left edge and as fricatives with regard to rules sensitive to their right edge. If [-cont] and [+cont] are ordered in sequence, edge effects are expected: affricates should pattern with fricatives with respect to the right-side context

1

1 thank the participants of OCP-I for their comments and questions. Special thanks go to the reviewers, the editors, Stuart Davis and Nick Clements for their valuable comments and suggestions. 2 The following abbreviations for features are used: [cont] for [continuant], [strid] for [strident], and Cor, Lab, and Dor for the Coronal, Labial, and Dorsal nodes respectively. 3 Note that Yip (this volume) also assumes that a lateral affricate is a lateral stop.

122

Yen-Hwei Lin

and pattern with stops with respect to the left-side context. If [-cont] and [+cont] are unordered, affricates should pattern both with stops and fricatives on either side of the environment. If affricates are strident stops, they should pattern with stops and strident fricatives on either side of the environment. Since affricates have a fricative release phonetically, one predicts that affricates may show edge effects in the phonetic implementation component. In recent years, the Stop Hypothesis for affricate representation seems to be gaining grounds. In this connection, the affricates in Piro (Matteson 1965), an Arawakan language of Peru, are of interest in that they seem to exhibit edge effects at the phonological level but anti-edge effects in phonetic implementation: phonologically, the cooccurrence restrictions on adjacent obstruent consonants clearly exhibit edge effects but a subclass of affricates and fricatives constitutes systematic exceptions; phonetically, affricates and stops pattern together and show an anti-edge effect with respect to phonetic implementation on their right edge. The cooccurrence distribution of the affricates contradicts the prediction made by the strong version of the Stop Hypothesis that no edge effects are possible at the phonological level; on the other hand, as we will see below, the systematic exceptions to edge effects pose problems for the phonological presence of [+cont]. The phonetic behavior of these affricates also begs for explanations as no anti-edge effects are predicted to exist. Piro has nine obstruents as given in (la). 4 N o consonant is allowed word finally, and permissible consonant clusters are found both word initially and

4

A reviewer correctly points out that, as in many languages, the affricates in Piro share the same place of articulation as the fricatives. This insight can be captured directly by a model that includes the fricative phase of the affricate in the phonological representation and associates the place feature with the fricative component, as in van de Weijer (1996). Then how can this insight be expressed if affricates are phonological stops? When an affricate is a strident stop, the surface fricative release tends to be identical in features to a strident fricative already allowed in the language due to markedness consideration or feature economy (Clements 2003) or structure preservation (Kiparsky 1982, 1985), hence the same place features between the affricates and the fricatives. By the same token, when an affricate is a nonstrident stop, the affricate is predicted to share the same place features with one of the nonstrident fricatives in the language. The reviewer also comments that in the case of Piro an underlying inventory like /p, t, alveolar strident stop, palatoalvelar strident stop, palatal stop, k, s, J, ς/ seems highly asymmetrical. It is, however, not uncommon for a language to have more stops than fricatives and more Coronal stops than other stops. In terms of feature economy (Clements 2003), the strident and palatal stops (i.e. the affricates) do make use of the features already available in the obstruent system.

Piro affricates

123

word medially. The cooccurrence restrictions on consonant clusters apply both underlyingly on m o r p h e m e structures and derivationally to repair nonpermissible clusters. For example, no [s?] cluster is allowed in underlying representation. A s the example in ( l b ) shows, w h e n such a cluster arises during the derivation, the first consonant deletes with compensatory lengthening of the preceding vowel. In addition, Piro has no sonority restrictions on onset consonant clusters. For example, both [ks] and [sk] are possible onset clusters as the examples in (1 c) illustrate. 5 In each C C cluster, either a transitional vowel is inserted or the first consonant is m a d e syllabic b y phonetic implementation rules (Levin 1987, Lin 1993, Lin 1999), as the examples in ( I d ) show. (1)

Piro: An A r a w a k a n language of Peru (Matteson 1965) 6 a. Obstruents: [p t k s J ς ts t j ίς] 7 b. kose + + ta —> kos-9eta —> ko:$eta to pull - always - verb suffix 'to always pull' c. ksu '...'s h o u s e ' skota 'low abdomen' d. k ' s u '...'s h o u s e ' skota 'low abdomen'

T h e first problem that Piro poses for affricate representation is that obstruent cooccurrence restrictions seem to show edge effects and therefore favor a representation in which affricates contain ordered [ - c o n t ] [+cont]. However, this analysis incorrectly rules out a set of systematic permissible clusters. T h e examples in (2ab) show that affricates are allowed to precede but not to follow the coronal stop [t] and that affricates are in general allowed to follow but not to precede a fricative. The edge effects shown b y (2ab) can b e accounted for by (i) representing affricates with ordered [ - c o n t ] [+cont] components and (ii) positing two constraints as in (2c) to rule out a d j a c e n t C o r o n a l [ - c o n t ] features and adjacent [+cont] features. If this analysis is correct, the

5

In Piro, all words and all morphemes (except single-consonant morphemes) end in a vowel. Matteson (1965: 23) argues that all consonant clusters belong to the onset. Lin (1997b) agrees with Matteson that all Piro syllables are open but proposes that Piro has only CV syllables and any preconsonantal consonants are extrasyllabic but moraic. The syllabic affiliation of the consonant clusters is not crucial to the analysis and arguments in this paper. 'Matteson (1954) contains earlier descriptions of Piro phonology and morphology, which were later revised and incorporated into Matteson (1965). I therefore take the data only from Matteson (1965). Thanks to Jeroen van de Weijer for reminding me of the existence of Matteson (1954). 'Matteson (1965) uses [x] and [tx] to represent [9] and [ίς] respectively.

124

Yen-HweiLin

Stop Hypothesis is called into question. However, representing affricates with ordered [-cont] [+cont] components fails to account for the permissible clusters in (2d), which have adjacent [+cont] features and therefore would have been ruled out by the constraint *[+cont]-[+cont]. (2)

Phonological edge effects a. Cooccurrence restrictions on [t] and affricates permissible: affricate-t: ts-t tj-t Ις-t nonpermissible: *t-affricate: *t-ts *t-tj *t-tg b. Cooccurrence restrictions on fricatives and affricates permissible: fricative-affricate: s-ts J-tJ ς-t? nonpermissible: affricate-fricative: *ts-s *tJ"-J" c. Phonologically ordered [-cont] [+cont] for affricates *[-cont, Cor]-[-cont, Cor] *[+cont]-[+cont] d. permissible: tg-s tg-J ΐς-ς tj-g

The second problem presented by Piro is that phonetic implementation of transitional vowels and syllabic consonants seems to show anti-edge effects for affricates. As shown in (3a), the phonetic implementation refers to the right-edge context of the first C in a CC cluster: a transitional vowel is inserted after the first stop in a [stop-C] cluster, and a fricative becomes syllabic when followed by a stop. If the affricate has [-cont] on the left edge and fricative release on the right edge, one expects it to behave like a fricative in an [affricate-C] sequence and like a stop in a [C-affricate] sequence; that is, edge-effects are expected. However, Piro affricates pattern exactly like stops in both [affricate-C] and [C-affricate] sequences, as the examples in (3b) show. Crucially, an affricate followed by a stop behaves like a [stop-stop] rather than a [fricative-stop] sequence: a transitional vowel but not a syllabic consonant occurs. This looks like a case of anti-edge effects since the fricative release of the affricate seems invisible to the phonetic implementation rule. (3)

Phonetic anti-edge effects a. transitional vowel in a [stop-stop] sequence: [t-1po] transitional vowel in a [stop-fricative] sequence: [k3su] syllabic fricative in a [fricative-stop] sequence: [skota] b. [p'tjowi] 'an edible root' cf. stop-stop [yoftgita] 'he changes course' cf. fricative-stop [tpkotu] 'cebus monkey' cf. stop-stop [petjitj^eta] 'he always lies in wait' cf. stop-fricative

Piro affricates

125

In this paper I analyze the behavior of Piro affricates within Optimality Theory (OT), and argue that the Stop Hypothesis is essentially correct. I pursue an analysis in which the affricates are stops in the lexical phonology but become contour segments with a fricative release in the postlexical phonology, and the cooccurrence restrictions are considered output constraints rather than morpheme structure constraints. The phonological edge effects are then accounted for by output constraints that are active postlexically. By treating [ts, s, tj, J] as stridents and [t?, 9] as nonstridents, the systematic exceptions to the edge effects in (2d) are attributed to the suggestion that affricates pattern with fricatives to the extent that both are stridents (Rubach 1994). What is crucial in this analysis is that we have to adopt some version of derivational OT in which lexical and postlexical levels may have different constraint rankings (Kiparsky 1997, 2000, to appear; 8 cf. Rubach 1997, 2000). 9 The phonetic anti-edge effects are explained by grouping affricates and stops in terms of sonority and by showing that the fricative release is irrelevant to the phonetic implementation process in question; that is, the phonetic 'syllabicity' implementation process is sensitive to the relative sonority of adjacent consonants rather than the adjacent stricture features. The remainder of the paper is organized as follows. Section 2 presents the cooccurrence restrictions on Piro obstruents and proposes an analysis compatible with the Stop Hypothesis. I also argue that even if the fricative release of an affricate must be present lexically, [-cont] must be designated as the primary stricture in order to account for the Piro data. In section 3 , 1 present an analysis in which the fricative release of the affricate is rendered irrelevant and suggest a phonetic interpretation for the phonetic 'syllabicity' process. The concluding section discusses the theoretical implications of this study and a potential alternative analysis that could further strengthen the support for the Stop Hypothesis.

8

See McCarthy (2002:185) for more references on implementation of OT along the lines of the theory of Lexical Phonology. 9 As we will see below, the analysis in this paper is more like Kiparsky's derivational model of OT where different rankings are possible at different strata; in Rubach's model, the derivational stages of different rankings do not necessarily correspond to the strata of Lexical Phonology. Thanks to Stuart Davis for pointing this out to me.

126

Yen-Hwei Lin

2. Phonological edge effects In this section, I first introduce the data of obstruent cooccurrence restrictions (section 2.1) and then present the p r o p o s e d analysis (section 2.2). Section 2.3 argues that any alternative analyses that allow [+cont] to b e specified in the lexicon m u s t treat [ cont] as the primary stricture of affricates in order to account for the Piro data.

2.1. Obstruent cooccurrence restrictions in Piro T h e table in (4) shows the possible and impossible word-initial and w o r d medial obstruent clusters in Piro, with the shaded areas indicating the n o n e x istent combinations. 1 0 (4)

Permissible and nonpermissible obstruent clusters in P i r o " C1/C2

Ρ

Ρ

10

k

t

ts

tf

+

+

+

+

+

+

+

s

J

9

+

+

+

+

+

+

+

+

+

+

k

+

t

+

+

ts

+

+

+

tf

+

+

+

t?

+

+

+

s

+

+

+

+

+

+

J

+

+

+

+

+

+

9

+

+

+

+ +

+

+

+

The table in (4) is based on the chart given in Matteson (1965:30) listing all permissible and nonpermissible clusters in Piro. This paper discusses only obstruent clusters. I have verified the accuracy of the chart by checking examples throughout the grammar, lexicon and texts sections of the source, and made a few changes. In the chart, Matteson admits *[ts-9] but excludes [tj-ς] and [ίς-ς]; on the contrary, I have

Piro affricates

127

A s m e n t i o n e d earlier, Piro c o o c c u r r e n c e restrictions on c o n s o n a n t clusters hold not only of u n d e r l y i n g representation but also t h r o u g h o u t the derivation. W h e n a n o n p e r m i s s i b l e cluster is created b y suffixation that triggers stem final v o w e l deletion, the m o s t c o m m o n solution is to delete the first c o n s o nant; if consonant deletion occurs w o r d medially, the preceding vowel is lengthened. S o m e e x a m p l e s are given in (5). 12 O n l y t w o derived n o n p e r m i s s i b l e obstruent clusters d o not u n d e r g o c o n s o n a n t deletion. T h e s e q u e n c e *[?-tJ] a l w a y s surfaces as [s-tj] ( M a t t e s o n 1965:34) and a n y *[t-s] s e q u e n c e surfaces as the alveolar a f f r i c a t e [ts] t h r o u g h obligatory affrication ( M a t t e s o n 1965:34). 1 3

found that [tf-9] and [ΐς-ς] are attested sequences but *[ts-9] is not. In the description following the chart, Matteson in fact states that [ts] does not precede [ς], and my search throughout the source confirms that there are indeed no examples of *[ts-9]. There are, however, instances of [tf-9] and many examples of [Ις-ς], e.g. [petfitj^eta] 'he always lies in wait' (p. 130), [tjanitggctna] 'invitation' (p. 180) and [kanmit^eru] 'enemy' (p.427). " A reviewer points out that one can claim that no consonant clusters are permissible on the surface phonetic representation in the language since there is vowel epenthesis/ consonant syllabicity, as illustrated in (3). Piro vowel epenthesis/consonant syllabicity, however, is phonetic and transitional in nature, so the transitional vowel does not constitute a segment and is considered the release of the first consonant in a CC sequence (Lin 1999; see section 3). In addition, when the first consonant is phonetically syllabic, the two consonants are still adjacent segments. Even if one maintains that no two consonants in Piro are segment-adjacent, we still have to explain why certain [C 3 C] sequences are attested but some others are not. With surface epenthesis and syllabicity, the reviewer asks if it is possible to have classic one-level OT evaluation. Whereas the phonetic constraints proposed in section 3 and the postlexical phonological constraints in section 2 may be present at the same level (see footnote 21), the distinction between the lexical and postlexical levels for constraint evaluation is still needed unless one believes that affricates must have ordered [-cont] [+cont] components lexically (see sections 2.2 and 2.3 and footnote 18). 12 For more examples, see Matteson (1965: 33-34) and Lin (1997a,b). The capitalized suffix in (5b) indicates a suffix that never triggers vowel deletion. Which suffixes trigger vowel deletion is morphologically conditioned and arbitrarily marked (Matteson 1965, Lin 1997ab). 13 In Piro, there is no contrast between a [t-s] sequence and the affricate [ts]; on the other hand, as opposed to affricates [tj] and [ίς], there are [t-J] and [t-9] clusters (Matteson 1965:26).

128 (5)

Yen-Hwei Lin a.

— » hitsrukate + tji ...'s chief - absolute b. kose+ ςε + t a —> to pull - always - verb suffix c. ςίίςί + tji —» ...'s foot - absolute d. 0 - nika + ka —> he - to eat - passive

hitsrukattji

— >

kosseta ?it?tji

— »

nikka

— >

hitsruka:tji 'chief ko:9eta 'to always pull' 9i:t/i 'foot' ni:ka 'he is eaten'

The generalizations of Piro obstruent occurrence restrictions are summarized in (6). First, identical stop sequences are prohibited (6a). Second, no sequences of two fricatives or two affricates are allowed (6b). Third, there is an asymmetry regarding the interaction between affricates and [t]: [t] occurs to the right but not to the left of the affricates (6c). Fourth, in general, a fricative may precede an affricate, with only two exceptions: *[