Constraints and Preferences [Reprint 2011 ed.] 9783110881066, 9783110170474

Table of contents :
Constraints and Preferences: Introduction
Parameters and scales in syllable markedness: the right edge of the word in Malayalam
Constraints and processes in phonological perception
Phonotactic constraints are preferences
Striving for optimality: output-oriented models of language change
Preferences as defaults in computational phonology
Devoicing and voicing assimilation in German, English, and Dutch: a case of constraint interaction
Hidden identity, or the double life of segments
Working with licensing constraints
Rules vs. constraints in modeling phonological change: the case of Raddoppiamento Fonosintattico
Are optimality theoretical “constraints” the same as natural linguistic “preferences”?
Abstractness in phonology: the case of virtual geminates
Constraints, preferences, and context-sensitivity in morphology
Old English fricatives: lenition and licensing
Index

Citation preview

Constraints and Preferences

W G DE

Trends in Linguistics Studies and Monographs 134

Editors

Walter Bisang Werner Winter

Mouton de Gruyter Berlin · New York

Constraints and Preferences

Edited by

Katarzyna Dziubalska-Kolaczyk

Mouton de Gruyter Berlin · New York

2001

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter GmbH & Co. KG, Berlin.

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

Library of Congress Cataloging-in-Publication

Data

Constraints and preferences / edited by Katarzyna DziubalskaKolaczyk. p. cm. — (Trends in linguistics. Studies and monographs ; 134) Includes bibliographical references and index. ISBN 3 11 017047 7 (cloth : alk. paper) 1. Grammar, Comparative and general — Phonology. 2. Grammar, Comparative and general - Morphology. 3. Constraints (Linguistics) I. Dziubalska-Kolaczyk, Katarzyna, 1960II. Series. P217.3 .C66 415—dc21

2001 2001044314

Die Deutsche Bibliothek — CI Ρ-Einheitsaufnahme Constraints and preferences / ed. by Katarzyna DziubalskaKotaczyk. — Berlin ; New York : Mouton de Gruyter, 2001 (Trends in linguistics : Studies and monographs ; 134) ISBN 3-11-017047-7

© Copyright 2001 by Walter de Gruyter GmbH & Co. KG, 10785 Berlin. All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced or transmitted 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. Printing & Binding: Hubert & Co, Göttingen. Cover design: Christopher Schneider, Berlin. Printed in Germany.

Contents

Constraints and Preferences: Introduction Katarzyna Dziubalska-Kolaczyk Parameters and scales in syllable markedness: the right edge of the word in Malayalam Eugeniusz Cyran

vii

1

Constraints and processes in phonological perception Patricia Donegan

43

Phonotactic constraints are preferences Katarzyna Dziubalska-Kolaczyk

69

Striving for optimality: output-oriented models of language change Livio Gaeta

101

Preferences as defaults in computational phonology Dafydd Gibbon

143

Devoicing and voicing assimilation in German, English, and Dutch: a case of constraint interaction Janet Grijzenhout

201

Hidden identity, or the double life of segments Edmund Gussmann

229

Working with licensing constraints Jonathan Kaye

251

vi

Contents

Rules vs. constraints in modeling phonological change: the case of Raddoppiamento Fonosintattico Michele Loporcaro 269 Are optimality theoretical "constraints" the same as natural linguistic "preferences"? Nikolaus Ritt

291

Abstractness in phonology: the case of virtual geminates Philippe Segeral and Tobias Scheer

311

Constraints, preferences, and context-sensitivity in morphology Rajendra Singh 339 Old English fricatives: lenition and licensing Jerzy Wojcik

359

Index

391

Constraints and Preferences: Introduction The idea of the present volume stems from the workshop under the same title, Constraints and Preferences, organised by the editor during the Poznah Linguistic Meeting 1999. The volume comprises a selection of the revised workshop papers and additionally invited contributions. The main working question of the workshop was the question concerning compatibility of the notions constraint and preference in modern phonology as well as morphology. The issue is approached in the contributions to this volume from a number of perspectives, shaped by different phonological models, such as Government Phonology, Natural Phonology, Optimality Theory. The articles by E. Cyran, E. Gussmann, J. Kaye and J. Wojcik take up the discussion in the framework of Government Phonology; D. Gibbon elaborates on preferences in computational phonology; J. Grijzenhout and M. Loporcaro consider the role of rules vs. constraints juxtaposing Optimality Theory and derivational approaches; the status of constraints vs. preferences in Natural Phonology and Optimality Theory is analysed in the contributions by P. Donegan, K. Dziubalska-Kolaczyk, L. Gaeta and N. Ritt; a comeback to an old unresolved issue of abstractness in phonology is found in the paper by P. Segeral and T. Scheer; constraints and preferences in morphology are treated by R. Singh. The contributions are ordered alphabetically in the volume and will be summarised in the same order below.1 Eugeniusz Cyran: Parameters and scales in syllable markedness: the right edge of the word in Malayalam Cross-linguistic preferences in the form of asymmetries known as markedness effects occupy a central position in most currently entertained phonological frameworks. The convergence of markedness phenomena observed in language acquisition, typology, historical change, as well as synchronically attested patterns of segment systems and their phonological behaviour ensures that linguists cannot afford to lose sight of the significance of capturing these effects.

VÜi

Katarzyna

Dziubalska-Kolaczyk

Thus, the fundamental goal is similar to most approaches, even if the means may differ. This paper attempts to capture the implications concerning the relative markedness of syllable structures within the model of Government Phonology by referring to the interaction between the licensing potential of different types of nuclei and the increasing licensing demand made by syllabic structures of growing complexity. The licensing of syllabic complexities will be shown to correspond to licensing of segmental complexity, which is discussed in, for example, Harris (1997). It is proposed that a single mechanism, that is, the interaction between the complexity of structure and licensing potential of nuclei may be able to explain the crosslinguistic preferences that characterise the distribution of segments and syllable types. Patricia Donegan: Constraints and processes in phonological ception

per-

In Natural Phonology (Stampe 1969, 1973; Donegan and Stampe 1979), phonology is conceived as a system for the production and perception of speech. Generative Phonology has been neutral with respect to production and perception, but the implied orientation of modern generative phonology (e.g., Kenstowicz 1994) is production; generative derivations are usually understood to change underlying representations to surface representations, and perception is rarely considered. Optimality Theory (OT) was developed (Prince and Smolensky 1993) in response to the generative model, and consequently most of its literature has focused on lexical phonology (morphonology), but OT shares with Natural Phonology (NP) an interest in universal constraints and the phonetic motivation of such constraints. Both OT and NP base phonological production and perception on conflicting but coexisting universal principles (constraints in OT, processes in NP), which are attributed to human phonetic abilities. OT's universal Well-Formedness constraints (or constraint families) may be thought to specify, in general, the undesirable configurations eliminated by NP's natural phonological processes, and OT's Faithfulness con-

Introduction

ix

straints might be taken to correspond to the limitations that speakers must place on natural processes. OT is nonderivational and based on evaluation of alternative candidates; NP has derivations, but they involve no extrinsic ordering (Donegan and Stampe 1979). In OT, the terms "input" and "output" continue the generative orientation toward modeling production, but OT, like NP, offers an opportunity to consider phonology as a system for the perception (as well as production) of speech. Smolensky (1996) considered perception within OT. He attempted, in particular, to show how a single constraint ranking can allow for asymmetries in children's acquisition, where perception of a distinction often seems to precede the production of the distinction. Pater (1998) offered a revision of Smolensky's model, and Lassettre and Donegan (1998) attempted a different revision, incorporating into OT a Stampean view of phonological representation (Stampe 1987). This paper considers these proposals regarding perception in OT and attempts to determine if a constraint-based, non-derivational analysis of perception can account for the same kinds of data as the NP derivational model. Katarzyna Dziubalska-Kolaczyk: Phonotactic constraints are preferences The main objective of this paper is to demonstrate that phonotactic constraints on consonant clusters should be understood as preferences in the epistemological sense. The paper starts with a discussion of the notion of a preference against the notion of a constraint in phonology. The framework of analysis is Beats & Binding Phonology, embedded in the theory of Natural Linguistics, which is an explicit preference theory (cf. Dressier 1999a). A model of phonotactics presented in the paper is derived from the Beats & Binding Phonology (cf. Dziubalska-Kolaczyk 1995, 1996). The phonotactic well-formedness conditions specified by this model according to the proposed Optimal Sonority Distance Principle (OSDP) define the cluster spaces representing universally preferred distributions of clusters. The graphic representation of the cluster spaces helps to see that (a) the clusters defined as preferred in

X

Katarzyna

Dziubalska-Kolaczyk

a given position in a word are not preferred in other positions; (b) the preferred clusters are preferred to a different degree, so that a direction from the least to the most preferred among the preferred can be established. Those two features of the cluster spaces clearly point to the fact that the conditions defining them have the nature of preferences, and not absolute constraints. Universal cluster spaces constitute a matrix against which language-specific phonotactics can be described. More importantly, they also allow for the evaluation of a given language-specific phonotactics. In particular, more and less stable clusters can be distinguished, and thus predictions can be drawn as to their "behaviour" in language use, acquisition and change. Livio Gaeta: Striving for optimality: output-oriented models of language change In the paper, rule-based and constraint-based approaches to phonological change are investigated. In particular, the characteristics of a constraint-based approach such as OT are highlighted, for which language change is immediately accounted for in terms of a re-ranking of constraints (cf. Cho 1998). Since constraints are considered general statements concerning markedness, language change is explained in terms of a redistribution of the markedness values within the grammar, completely independent of problems related to the format of the grammar, as in classical generative linguistics. This makes OT much closer to functionalist approaches to phonology, that view language change as motivated by a local decrease of markedness independent of the general improvement, i.e. simplification, of the grammar (e.g., Vennemann 1993; Wurzel 1994). However, the essence and the role of the constraints must be carefully inspected in order to establish their plausibility and their explanatory power. On the one hand, a markedness statement represented by a constraint must be proved to be significant for a given language at a given time. That is, we must be able to find a bundle of phenomena vouching for the relevance of a process for that language, as shown by the rphenomena in English. Moreover, OT seems unable to deal with

Introduction

XI

typically derivational issues such as counterbleeding order, as in the case of i-epenthesis in Algares Catalan. This points to an excessive rigidity of the model. On the other, an alleged re-ranking of constraints must be scrutinized in all its functional consequences in order to provide a robust explanation rooted in general properties of language, as shown by the history of the glides in the Germanic languages. Dafydd Gibbon: Preferences as defaults in computational phonology A preference is intuitively an element of a set which is valued more highly than other elements; the preferred element can be interpreted as an object or a property, a state, an event or a course of action. Similarly, a default is a typical state or typical choice which is valued more highly than other choices, other things being equal (other meanings of "default" are of no concern here). Neither a preference nor a default is a fixed value: a preference may not be favoured under all conditions, and a default may be overridden by other considerations in specific situations. The concept of once-for-all fixed rules is gradually being replaced in linguistics by a notion of preference, starting perhaps with Trubetzkoy's notion of privative opposition, though various notions of markedness to Vennemann's Preference Laws and Optimality Theory, as well as the generalisation and irregularity hierarchies of computational lexicon theory. Quantitative expressions of markedness range from Labov's variable rules to probabilistic automata as used in corpus tagging and in automatic speech recognition. The mechanisms used to express preference valuations range from marked/unmarked values in feature structures through default inheritance hierarchies with orthogonal and prioritised inheritance, default automata, to fully fledged default logics and non-monotonic reasoning. Perhaps the simplest gloss for default is typical, a word which deserves more extensive empirical and phenomenological analysis than it can be given here. In everyday life, it is a vehicle for stereotypes and many forms of prejudice, the Occam's razor wielded in the daily struggle to avoid as far as possible the chore of thinking rationally. In

XÜ

Katarzyna

Dziubalska-Kolaczyk

this pragmatic context, the allies of typical are adverbs of subjective frequency, such as usually, mostly, probably, which are used not as quantifiable statements of frequency and probability, but as adverbs of degree which qualify stereotypic convictions. The main goal of the present study is to not to argue for the use of default mechanisms for explicating preferences in phonology and prosody, since this is quite widely accepted. Rather, the author tries to collect a wide range of different aspects of phonology and prosody, and show, often in novel ways, the relevance of defaults to them, with the longer term goal of developing a unified theory of defaults and preferences in phonology and prosody. Janet Grijzenhout: Devoicing and voicing assimilation in German, English and Dutch: a case of constraint interaction This paper shows the advantages of Optimality Theory (OT) over traditional derivational approaches to voicing phenomena in Germanic languages. It is argued that the presence or absence of final devoicing and progressive and regressive voicing assimilation in German, Dutch, and English are best attributed to the different rankings of universal constraints. With respect to modern Dutch and English, a further interesting aspect is that stem-final obstruents behave differently in derivation, compounding and cliticization. An underlyingly voiced obstruent in Dutch is (i) voiced before the schwa-initial suffix -er (cf. he[b]er 'greedy person'), (ii) assimilated to a following obstruent in compounding (cf. hand + palm —> han[tp]alm), and (iii) voiceless before the schwa-initial clitic -er (heb + 'r —> he[p]er 'have her'). An underlying voiced obstruent in English is voiceless before a voiceless monosegmental suffix (five + th —> fifth), and voiced in compounds {bird + cage —• bir[dk~\ase). Final-devoicing and voicing assimilation have always been regarded as evidence for level-ordering in Dutch and English (e.g., Booij 1995; 1996). This paper argues that these phenomena are best analysed in a constraint-based framework which does not employ level-ordering (e.g., Kiparsky 1998), but rather domain specific constraints such as IDENTITYONSET (VOICE) and

Introduction Xlll IDENTITYSTEM(VOICE). The use of domain specific constraints enables us to account for data which are problematic (and often ignored) in recent OT-literature and allows for an analysis of voicing alternations in different languages by means of a small set of universal constraints. Edmund Gussmann: Hidden identity, or the double life of segments The paper takes up the issue of the nature of representations within a monostratal model such as Government Phonology. Since the model explicitly bans any serialism or derivational strategies, it forces a situation where all phonological regularities must be contained within a single representation. The existence of empty expressions both nuclear and non-nuclear - is a well-known feature of this approach; the recognition of skeletal points dominating no melodic elements or melodies severed from skeletal points requires full phonological justification. In this paper it is argued that phonological regularities enforce representational differences which are not manifested phonetically. This view, clearly reminiscent of absolute neutralisation in classical generative phonology is argued for on the basis of Russian labial fricatives, Polish dorsal obstruents and Welsh vowels. These cases, along with others reported in the literature, suggest that the traditional notion of the phonetic representation frequently disguises or distorts phonological regularities. These must be established and defended on the basis of the phonology of each language with only marginal attention paid to the actual physical realisation. Jonathan Kaye: Working with licensing constraints In this paper, the author explains the role of licensing constraints within a GP (Government Phonology) framework. Licensing constraints were originally conceived as language-specific constraints on the combinatorial possibilities of the elements A, I and U. It is assumed that any combination of these elements is possible (specifically in nuclear systems) unless explicitly excluded by a licensing

XIV

Katarzyna

Dziubalska-Kolaczyk

constraint. For example the constraint, I and U may not appear in the same nuclear expression, excludes nuclear expressions such as ({I},U) = [Ü], ({A,I},U) = [ö], and so forth. Licensing constraints are what replaced "charm theory" of earlier versions of GP. Following recent work of Monik Charette and Asli Göksel, the author explores other ways in which licensing constraints interact with other parts of the phonology. Charette and Göksel show that Turkish vowel harmony behaviour (I may spread into any harmonically governed position; U can only spread into an empty nuclear position; A does not spread at all) follow directly from the licensing constraints used to define the set of lexical nuclear expressions in Turkish. Following that line of research, the author shows that the three harmonic groups of nuclei in Finnish, harmonising, nonharmonising and neutral, fall out from the set of licensing constraints necessary to define Finnish nuclear expressions. The role of licensing constraints can be extended to tonal systems. The author shows that Putonghua (standard Mandarin Chinese) has not 4 tones but rather 4 tonal patterns. These patterns include one level "tone", 2 rising "tones" and one falling "tone". He also shows that this precise configuration may be derived by a simple tonal licensing constraint to the effect that Putonghua tonal patterns have an intrinsic head. Michele Loporcaro: Rules vs. constraints in modeling phonological change: the case of Raddoppiamento Fonosintattico For those who have been trained as linguists during the second half of the twentieth century, the notion of rule has come to be considered as a matter of course, as something intrinsical to our understanding of linguistic structure. Things have been changing recently, though. While the term "rule" in itself continues to be used, there is now an increasing shift towards non-dynamic (i.e. static, or "declarative") models of linguistic description. In these models, whose most successful representative is nowadays Optimality Theory (OT), the grammar of any specific language is described as the product of a

Introduction

XV

ranking of universal violable constraints selecting among candidate outputs. Thus, the rule component is dispensed with. In this paper, the author tackles the question from the vantage point of phonological change. Many instances of change, which were previously described as changes in the rule component have been recently reanalyzed, within OT, as the product of constraint reranking. One of the arguments which are often mentioned when comparing the relative merits of static vs. dynamic descriptive models, is that the former directly allow for the expression of what is universal in language, whereas rule-based descriptions, it is claimed, are confined to the language-specific dimension. From the point of view of the historical linguist, there is not anything inconvenient about being language-specific. It is of course part of the task of historical linguistics to try to detect constant factors, found to be at work cross-linguistically in different instances of change. But there is also a highly language-specific aspect about doing historical linguistics, which is rooted in the object itself: history is by definition "specific", not general or universal. In the paper, the author presents some considerations as to how we should best describe phonological change, whether with rules or with constraints. In particular, he considers an example drawn from the diachronic phonology of Romance languages, viz. the historical development of Italian Raddoppiamento Fonosintattico (RF). Nikolaus Ritt: Are optimality theoretical "constraints" the same as natural linguistic "preferences "? This paper discusses the meanings which the terms "constraint" and "preference" have in Optimality Theory (henceforth OT) on the one hand and Natural Linguistics (henceforth NL) on the other. It attempts to show that both terms, which at first sight might appear to play comparable roles in the two frameworks, actually refer to different concepts within each of them. The reason why these differences in usage are seldom made explicit and tend to be pasted over and overlooked is that they reflect fundamental tenets of OT on the one hand and NL on the other, which are taken for granted by repre-

xvi

Katarzyna

Dziubalska-Kolaczyk

sentatives of each of the two camps but often fail to be come explicit in the discourse between them. The paper consists of three parts: first it shows why OT "constraints" and NL "preferences" might be regarded as functional counterparts of each other and therefore as freely exchangeable. Second, it shows that they get radically different interpretations if viewed against the wider theoretical frameworks within which OT and NL approach the study of language. And finally, it points out possible ways in which the two approaches might profit from one another or how their concepts and methods might be integrated. Philippe Segeral and Tobias Scheer: Abstractness in phonology: the case of virtual geminates This paper contributes to the old debate of Abstractness in Phonology. More specifically, the notion of abstract segments is examined. In the relevant literature of the 70s, this name was given to objects which occur in the underlying structure of a morpheme, but appear in none of its phonetic manifestations, or, worse, are even absent from the segmental inventory of the language. The authors suggest that abstract segments fall into two categories, 1) those that are abstract because of their melodic properties and 2) those that are abstract because of their association between two autosegments. This distinction was not made in the relevant literature because autosegmental structures were not given birth by that time, or were just emerging. The authors further submit that compelling (internal) evidence for the existence of abstract segments mostly (always?) concerns those of the second kind. That is, a segment that is known in the segmental inventory of a language may be doubly associated to a given autosegment, without that double-association yielding the expected phonetic result. In the cases they examine, a melodic unit is doubly associated to the timing tier, and thus phonologically long, but enjoys an invariably short pronunciation. The vocalic instantiation of this kind of object has already been identified for the purpose of the analysis of Afro-Asiatic languages, cf. Larsen (1994), Lowenstamm (1996). Its consonantal counterpart is the object of the present empirical in-

Introduction

XVÜ

quiry: consonants that are underlying geminates, but which are always realised as phonetically simplex consonants. Such an object, which the authors call a virtual geminate, never betrays its geminate identity by a phonetic clue related to length, but by other properties that can be read off the phonetic environment. Two detailed case studies are presented that ought to show that the respective languages, Cologne German and Somali, may not be analysed properly unless the existence of virtual geminates is recognised. The discussion about virtual geminates relates to the topic of the book in the following way. The old problem of abstractness referring to the dissimilarity that grammar ought to allow between underlying and surface representations is mirrored in more recent constraintbased frames by the so-called family of faithfulness constraints that are designed to minimise the contrast between deep and surface structure. Violation of faithfulness is viewed as a consequence of the action of another set of constraints. In principle, the balance of both "forces" is not predictable, it may go way to one side, as well as far on the other. In this paper, the authors intend to show how this problem may be handled in a principled way. Instead of allowing for a principle to be violated in undetermined and a priori unlimited fashion, they submit that the area in which it is contradicted is well defined and predictable. Circumscribing the violation in place, nature and function allows to capture the phenomenon by setting a parameter, rather than by dispensing with the principle, or allowing for its violation. Moreover, it adds a generalisation to the grammar that does not emerge when the violation of principles is thought of as the unmarked case. Rajendra Singh: Constraints, preferences, and context-sensitivity in morphology Whereas both constraints and preferences seem reasonably wellmotivated on both physiological and distributional grounds in phonology, it is not entirely clear if morphology needs anything other than context sensitivity. In this paper, the author argues that once

XVI11

Katarzyna

Dziubalska-Kolaczyk

context-sensitivity is appropriately taken into account, the need to invoke constraints, of whatever sort, disappears from morphology. He in fact argues that both constraint-based accounts of morphology, such as Optimality Morphology, and constraint-rich accounts, such as Lexical Morphology, are inadequate, and that it is so for the reason behind Jespersen's (1924: 52) observation, repeatedly cited by Chomsky (though routinely ignored by his followers) that nobody ever dreamt of doing universal morphology. Aware of and impressed by the variety of forms and functions in morphological systems, Jespersen believed that it was not possible to construct any very tightly restricted system of universal morphology. This paper will in fact show that he was fundamentally right. Jerzy Wojcik: Old English fricatives: lenition and licensing The paper deals with the phonology of Old English fricatives and is primarily concerned with the mechanism responsible for the intervocalic voicing of Old English anterior spirants. Adopting the Government Phonology approach, it is demonstrated that the intervocalic voicing of Old English anterior spirants s,f and (9 as well as the deletion of χ are processes of lenition and are best analysed as a unitary phenomenon involving the delinking of the element h. The process of lenition is seen as a prosodic effect following from the application of the notion of Licensing Inheritance. The focus of the second part of the paper is the behaviour of the voiced velar fricative γ and the palatal fricative f . Assuming the phonological representation of Old English voiced fricatives as defined solely by the presence of an element responsible for the place of articulation, the author presents an interpretation of γ-j alliteration in early Old English and an account of γ-j alternations. Finally, he considers the phonological properties of the palatal fricative / and tries to account for the unavailability of its voiced counterpart j. As shown by the summaries, the present volume is both interdisciplinary and integrative. Divergent theoretical backgrounds of the authors do not preclude them from converging on the discussion of a

Introduction

XIX

central issue to phonological (as well as morphological) theory, the issue of the meaning, usability and compatibility of the notions: constraint and preference. It is treated either directly (i.e. constraint vs. preference) or via the consideration of highly related notions and tools such as: preference vs. default, rule vs. constraint, or process vs. constraint as well as violability of a constraint. It remains to be hoped that the reader will find this collective effort worthwhile. Katarzyna Dziubalska-Kolaczyk [email protected] Poznari, May 2001 Notes 1. I'd like to thank the authors for providing the abstracts of their contributions. The abstracts appear here only slightly modified.

Parameters and scales in syllable markedness: the right edge of the word in Malayalam Eugeniusz Cyran

1. Introduction Cross-linguistic preferences in the form of asymmetries known as markedness effects occupy a central position in most currently entertained phonological frameworks. The convergence of markedness phenomena observed in language acquisition, typology, historical change, as well as synchronically attested patterns of segment systems and their phonological behaviour ensures that linguists cannot afford to lose sight of the significance of capturing these effects. Thus, the fundamental goal is similar to most approaches, even if the means may differ. Among the typically discussed asymmetries are those involving subsegmental structures. For example, sonorants are generally found to be voiced, therefore, a combination [+nasal, -voice] must be treated as marked. In underspecification models the feature [-voice] must be present in the underlying representation of sonorants, while [+voice], which is predictable, may be supplied by redundancy rules. Markedness effects are also observed in the distribution of segments within a word. Word-final devoicing of obstruents in Polish is an example of positional neutralisation of contrast whereby a given position, say, the coda, is unable to maintain the marked laryngeal specification in this language. The direction of this neutralisation is towards the universally unmarked opposition among obstruents, that is, [-voice]. It should be noted, however, that models, which employ binary features, must resort to quite arbitrary statements concerning the marked values of particular features. These statements reflect the observable tendencies but do not explain why the opposite designation is ruled out. In this respect, systems with privative features seem to fare better.

2

Eugeniusz Cyran

Binary feature systems also suggest that the markedness oppositions are of a bilateral nature, however, multilateral oppositions involving markedness scales are frequently found in languages and deserve equal attention (e.g., Trubetzkoy 1939). One example of such a scale is the relative sonority of segments which at least when applied to the coda position predicts quite well the preferences for the types of segments that may occur in this context. Another preference scale may be observed concerning relative syllabic complexity. It is generally accepted to treat CV as the optimal syllable structure. It is the only syllable type that is present in all languages regardless of the maximal complexity of its onsets and rhymes. The increasing complexity of syllabic structure goes hand in hand with its relative markedness. Thus, a language with branching onsets (CCV) is not only more marked than one with only CV, but it also must have the less marked structure in its inventory. Somewhere between the two types of syllabic complexity, of which CCV is not at all the maximal one, there is also a structure with an empty onset (V) as well as CVC, that is a syllable closed by a coda consonant. Kaye and Lowenstamm (1981) are the first to have pointed out the implicational relationship between complex onsets and branching rhymes which says that a language with branching onsets must also have branching rhymes, while the reverse implication is not possible. This paper attempts to capture the implications concerning the relative markedness of syllable structures within the model of Government Phonology by referring to the interaction between the licensing potential of different types of nuclei and the increasing licensing demand made by syllabic structures of growing complexity. The licensing of syllabic complexities will be shown to correspond to licensing of segmental complexity, which is discussed in, for example, Harris (1997). The paper is organised as follows. In section 2, we discuss the markedness scale proposed in Kaye and Lowenstamm (1981). Section 3 illustrates how the segmental and syllabic markedness may be expressed in Government Phonology and points to some necessary refinements of the model in order to capture the observations made by Kaye and Lowenstamm. Section 4 contains a discussion of the

Parameters and scales in syllable markedness

3

syllable structure in Malayalam, which is an apparent counterexample to the markedness scale of syllabic complexity. In 5, Malayalam is placed in a broader typology of syllable structures permitted at phonological level by the model presented in 3. Particular attention is paid to the distribution of segmental and syllabic structures at the right edge of the word in this language. It is proposed that a single mechanism, that is, the interaction between the complexity of structure and licensing potential of nuclei may be able to explain the cross-linguistic preferences that characterise the distribution of segments and syllable types.

2. Syllable structure markedness Kaye and Lowenstamm (1981) propose that generally we are dealing with three groups of languages as illustrated below. (1)

I CV II CVC, CV III CCV, CVC,CV

Languages which belong to group I, like Desano or Zulu, have only the simplest syllable structure CV. Group II, with branching rhymes but still simplex onsets, e.g., Hungarian, constitutes a next step on the markedness scale. The third group allows for branching onsets, branching rhymes and simplex syllables, e.g., Polish. In order to account formally for the implications shown in (1), Kaye and Lowenstamm propose to index the markedness scale in the following way (1981: 292). (2)

Onset C 0 CC CCC Ci...Cn

Rhyme V 0 VC VCC VCi...Cn-i

Markedness 0 1 2 3 η

4

Eugeniusz Cyran

The markedness values are established separately for the onset and for the rhyme. For this reason branching onsets end up having the same markedness value as the branching rhyme contrary to the classification in (1) which suggests that the two structures must constitute separate levels. For this reason, Kaye and Lowenstamm postulate that the implication CCV z) CVC may be handled by a separate condition stipulating that the maximum markedness value for the onsets m may be equal but should not exceed that for the rhyme η (m < n). A more fundamental problem for the proposal at the time when it was made was the necessary assumption that syllable structure be present in the lexical representation rather than supplied by rules. The following section demonstrates that the basic insight of Kaye and Lowenstamm (1981), summarised in (1) above, may receive a fairly non-arbitrary description within the model of Government Phonology, and that there is no need for a separate condition differentiating branching onsets and rhymes, because they are not of the same markedness value. The model also gets round the problem of lexical presence of syllable structure by assuming that all syllabification is due to governing relations contracted between consonants at the lexical level. The nature of the underlying representation is such that it may receive interpretation without resorting to derivation to a surface form. This is achieved by assuming autonomous interpretability of phonological elements (Harris 1990; Harris and Lindsey 1995).

3. Markedness scales in Government Phonology Let us consider only one way to deal with markedness effects in Government Phonology, which involves making reference to the relative complexity of segments and syllables (Harris 1994, 1997).1 The purpose of this section is to demonstrate that the same basic mechanism, that is, the interaction between the relative strength of licensers and the relative complexity of the structures requiring licensing may capture markedness and typological tendencies.

Parameters and scales in syllable markedness

5

3.1. Segmental complexity Segments in this model are composed of privative elements. Thus, vocalic systems are defined in terms of three main resonance elements I, A, U corresponding to the corner vowels [i, a, u], which are most commonly found across languages. The mid vowels [e] and [o] are combinations (A, I) and (A, U) respectively. They are more complex segments and are assumed to be more costly in terms of licensing, where the actual cost is calculated straightforwardly from the number of elements involved. It is to be expected that in prosodically weak contexts, the less complex segments should have better chances of survival than the compounds. This prediction is borne out by phonological phenomena such as lowering or raising of mid vowels in unstressed nuclei as in, for example, Bulgarian and Catalan (Harris 1994). Since each element is autonomously pronounceable, the reduction in complexity does not hinder interpretability of the remaining material. Obstruent devoicing, as in Polish or German, is captured in exactly the same way as vowel reduction. Simply, the element defining the laryngeal activity is unlicensed in prosodically weak positions. Thus, the general principle responsible for markedness phenomena in segmental structure in GP is viewed as a distribution of various complexities within a word in such a way that the amount of phonological material tends to be greater in strong positions and reduced in weak ones. Harris (1997) proposes a coherent theory of neutralisation, which unifies the intimate relationship between the distribution of prosodic licensing within a word and the allocation of melodic contrasts. Later in this paper, we will see how the complexity of consonantal segments may account for cross-linguistic patterns of occurrence in word-final position. Below, we illustrate how syllable typology and markedness can be captured in GP by referring to the same concepts as in the case of segmental markedness, that is, complexity and licensing.

6

Eugeniusz Cyran

3.2. Syllabic complexity Syllabification in GP is government driven (Kaye, Lowenstamm and Vergnaud 1990). Adjacent positions enter into asymmetrical relations where one of the participants acts as the governor (T) and the other as the governee (R). The assignment of these functions in a given string is determined by the elemental complexity of the two adjacent segments (Harris 1990). Thus, the direction of governing relations is resolved on the basis of substantive factors (complexity slope). Additional, formal conditions on government, such as adjacency and directionality, ensure that all possible syllabic constituents recognised in GP are maximally binary left-headed governing domains. If the order of segments happens to be that of a less complex consonant followed by a more complex one, that is, RT, then the direction of the governing relation must be from right to left, in which case we are dealing with an interconstituent relation between an onset and the preceding rhymal complement ("coda"). The two types of relations are illustrated below.3 (3)

a.

Ο


(3a) illustrates constituent government, that is, a branching onset, as in the Polish word mokry 'wet', while in (3b) the rhymal complement (X2) is governed by the following onset as in Marta 'name'. Charette (1990, 1991) proposes that both types of governing relations must be licensed by the following nucleus, and distinguishes between direct (3b) and indirect (3a) government licensing as sepa-

Parameters and scales in syllable markedness

7

rate parameters defining licensing properties of nuclei. The positive setting of the two parameters is assumed to condition the presence of governing relations of the RT and TR type in a given language, and, in effect, of branching rhymes and branching onsets. It must be noted, however, that while the parameter on indirect government licensing may indeed be assumed to have direct influence on the presence of branching onsets, because the branching onset is nothing else but a left-headed governing relation which is licensed by the following nucleus, the same cannot be easily said about branching rhymes. Direct government licensing is responsible only for the governing relation between the onset and the "coda" consonant in the preceding syllable. It is not clear how this should evoke a branching rhyme structure, which itself is defined, like any other branching constituent, by a left-headed relation between the nucleus (xi) and the rhymal complement (X2). This is probably the reason why the government licensing parameters never fully replaced the parameters on branching constituents. The standard way of capturing syllable typology in GP is to refer to a set of parameters, which may allow the constituents to branch if set in the ON, or not if set in the OFF. (4) Onset Rhyme Nucleus

Branching ON/OFF ON/OFF ON/OFF

It follows that the model employs two overlapping though independent mechanisms to deal with the same aspect of phonological structure, that is, the presence of branching constituents. However, neither separately, nor in conjunction, are these mechanisms able to capture the observation made in Kaye and Lowenstamm (1981) concerning the implicational relationship between branching onsets and branching rhymes. To see this clearly, let us leave aside the parameter on the branching nature of nuclei and assume for the moment that either set of the parameters discussed above could suffice on its own.

8

Eugeniusz Cyran

The possible choices concerning the parameters on branching onsets and rhymes are given below. (5) parameters TR

b.

c.

d.

ON

OFF

OFF

ON

ON

OFF

ON

OFF

English

Zulu

Hungarian

???

branching onset = indirect gov. licensing

RT

a.

branching rhyme = direct gov. licensing

The problem lies in the nature of parameters in general, or rather in their independent status. Since each parameter is set separately, the only way to preclude (5d) above is to resort to arbitrary designation of such settings as marked or downright impossible.4 A more serious problem for GP is that as long as the parameters on government licensing properties of nuclei and parameters on branching constituents are allowed to coexist in the model, we cannot exclude conflicts between these separate types of parameters. For example, we must assume that the presence of branching onsets is due to two theoretically unconnected parameters - one which allows onsets to branch, and refers to the structure of the constituent, and the other, which is defining the licensing properties of the nuclei in a given language. (6) parameters

a.

b.

c.

d.

branching onset

ON

OFF

OFF

ON

indirect gov. licensing

OFF

ON

OFF

ON

???

???

Zulu

Polish

What (6) illustrates is that it is not clear what the potentially possible conflicting settings of the two parameters would yield. They must be assumed, therefore, to be switched ON or OFF in conjunction to account for the observable facts, which suggests that either the two parameters require additional justification to be maintained

Parameters and scales in syllable markedness

9

in the grammar, or some external mechanism must be evoked to link them. Below, we will pursue yet another option. Since, syllabification in GP is indeed a reflection of governing and licensing relations, let us assume that we can do without parameters on branching constituents and derive the syllable typology only by reference to licensing properties of nuclei. The latter will not be defined in terms of separate parameters but rather as a scale on which the cut-off points are defined by the complexity of syllabic structure to be licensed. Note that the elimination of parameters on branching constituents from the model does not affect such fundamental notions as, for example, the binary theorem. The maximally binary nature of constituents is guaranteed by the way governing relations are contracted and need not be doubly secured. The table below recapitulates the hierarchy proposed in Kaye and Lowenstamm (1981) and illustrates the proposed interaction between syllabic complexity and relative licensing strength of the licenser, that is, the following phonetically realised nucleus. (7) example

language

I

Ca

city

Desano

II

C.Ca (RT-cluster)

winter

Hungarian

III

CCa

trap

English

(TR-cluster)

ψ \/ \/

C = any consonant, Τ = governor, R = governee, a = any vowel

The general proposal is that the complexity of syllable structure in a given language is due to the licensing properties of its nuclei rather than to some extraneous parameters or constraints. A selection of the actual strength of nuclei in a given language is arbitrary, that is, either of the three choices is available, but the scale itself is by no means arbitrary. The three steps, or "quantal regions", to borrow a term from phonetic theory, along the scale of syllabic complexity are non-reversible or re-rankable. That the licensing invariably goes from the nucleus to the preceding onset, which imposes varying demands on its licenser depending on its function within a string. Either the onset is simplex, or

10

Eugeniusz Cyran

simplex but governing a "coda", or complex, that is, branching. While the complexity difference between a simplex intervocalic onset and a complex, branching onset is pretty obvious and requires no further justification, the rigid placement of RT clusters, that is, rhyme-onset relations, in the middle of the scale may need to be further substantiated.5 The fixed nature of the complexity scale is not its only advantage. Notably, the simplex onset is at last treated as a genuine part of syllabic markedness rather than an implied structure in the presence of more complex ones. It is where the scale begins and thus it plays a crucial reference point on the scale. The scale also offers a fresh look at the concept of markedness itself. More complex structures need not be viewed as violations of any universal conditions or constraints, but rather, as utilisation of all logically possible structural configurations, except that some of them happen to be more costly to license than others. Additionally, the model of Government Phonology imposes limits on the structural possibilities themselves. These follow from the nature of government. By eliminating parameters on branching constituents and modifying slightly the understanding of existing mechanisms in GP we are able to offer a fairly constrained theory of syllabification, which captures the observed tendencies across languages and, thanks to its simplicity, is not unviable in terms of, for example, learnability. However, given the fixed nature of the complexity scale it is very easy to falsify the proposal. Potentially detrimental to the model could be the existence of languages, which possess branching onsets (group III), but lack branching rhymes, that is, "codas" (group II). Below we will consider an example of such a language and further extend the proposal by taking into account different types of nuclei and their place in licensing gradation.6

4. The syllable structure of Malayalam The data from Malayalam are of interest to us for a number of reasons. First of all, this language has been claimed to have branching

Parameters and scales in syllable markedness

11

η

onsets but no codas, i.e. branching rhymes. In this respect it may constitute a counterexample to the predictions which follow from the markedness scale introduced in 3.2. Secondly, the discussion of the issues concerning the syllable structure of this language will allow us to further clarify what is understood by "coda" in our model as opposed to other current models of phonological description. Thirdly, Malayalam will be tentatively placed in a more general typology of syllabic organisation in terms of the licensing abilities of its nuclei. And finally, an attempt will be made to capture in a static kind of way the distribution of segments and syllabic types in Malayalam. This will be done mainly with respect to the right edge of the word. The main conclusion will be that Malayalam is not problematic for our typology. However, it appears that this language points to a need for a slight extension of the markedness scale to allow for a tripartite distinction between nuclei in terms of their licensing abilities, that is, a) a full vowel, b) a schwa, and c) an empty nucleus. Initially, we will concentrate only on one problem, which is directly relevant to the markedness scale of government licensing introduced in 3.2. One of its predictions is that it is unlikely for a language to go as far as having branching οonsets (V.TRV) without also having rhyme-onset relations (VR.TV). The discussion is based on Steriade (1981), K.P. Mohanan (1982, 1986), and T. Mohanan (1989).9 The data come from these sources.

4.1. The no-coda hypothesis There are three main arguments that led K.P. Mohanan (1982) to postulate the no-coda analysis for Malayalam. The first two are based on what we can call word level phenomena and refer to native speakers' intuitions in surface, phonetic syllabification and in language games. The third argument is strictly phonological in nature and for this reason it will receive more attention than the other two. The first argument is connected with the way speakers of this language syllabify words in experimental situations. For example [ampalam] 'temple', [b akti] 'devotion' and [dabba] 'tin can' are broken

12

Eugeniusz Cyran V»

1Λ

up as [a-mpa-lam], [b a-kti] and [da-bba] respectively. K.P. Mohanan (1982) claims that this indicates that the syllable structure in Malayalam has no codas and that for this reason all such medial clusters, regardless of their sonority make-up, must be linked to the onset. This argument is based on native speakers' intuitions concerning syllable divisions and its strength depends strictly on the degree to which such intuitions are taken to be valid for the discovery of the syllable structure of a given language. One reason why phonetic syllabification should be treated with caution is that on many occasions such experiments bring contradictory results concerning the syllable structure of a given language.11 Secondly, we cannot be sure that there is a direct connection between the principles underlying the intuitive divisions made by native speakers and the actual phonological principles of speech organisation. This view finds unexpected support in Malayalam itself, as we will see below, in that phonological facts from this language suggest something contrary to the native speakers' intuitions. The second argument adduced by K.P. Mohanan (1982) in favour of the no-coda hypothesis is based on the effects observed in a language game in which a nonsense syllable 'pa' is inserted before each syllable. Compare the effects of the same game in different, though related, languages (T. Mohanan 1989: 592). (8)

Malayalam a. banctanam paba-pandha-panam 'imprisonment' b. darsanam pada-parsa-panam 'vision' π

π

Λ.

Λ

π

π

Λ

Hindi paban-padha-panam padar-pasa-panam

Note that while in Malayalam the sonorant-obstruent cluster remains intact, it is broken up in Hindi. The 'pa' insertion could be given a uniform formulation: "insert 'pa' before every syllable" for both languages only if the clusters in Malayalam are treated as complex onsets rather than coda-onset sequences. This, in a nutshell, is

Parameters and scales in syllable markedness

13

the argument for treating these clusters as onsets, and in effect, in favour of the no-coda hypothesis. One must at least voice some reservations as to whether this necessarily means that any cluster, which is not broken up by 'pa', must be viewed as an onset. First of all, the formulation of the rules of the game could be justifiably different in the two languages. There is nothing in principle that enforces uniformity in how the nonsense syllable is inserted. Secondly, the different effects could easily derive from some constraint in the language, which disallows breaking up certain clusters. What we are talking about here are, for example, sonorant - obstruent sequences and geminates of which at least the latter structure consistently exhibits integrity and immunity to any 12 process that would break them up (e.g., Hayes 1986). In support of this view, we may mention the fact that Malayalam word-final rhyme-onset sequences exhibit similar integrity in that the [o]epenthesis, to be discussed below, takes place after the cluster rather than within it. It appears that the two arguments supporting the no-coda hypothesis are of a particular kind. T. Mohanan (1989) refers to these phenomena as occurring at "the less abstract levels of representation". Without completely discarding these arguments, we voiced some reservations as to their validity for the discovery of the syllable structure of the language. While the native speakers' intuitions concerning surface syllable divisions may need to be treated with more caution, the language game argument seems to be open to alternative analyses. The third argument in favour of the no-coda hypothesis is based on the process of obligatory [sj-epenthesis that inserts a schwa vowel after every word-final consonant other than [m] and [n] (9a), unless the following morpheme or word begins with a vowel (9b). (9) a. maram 'tree' awan 'he' b. /kaat/ > kaato but kaatooka 'ear' 'ornament for the ear' /maas/ > maas9 but maasewite 'teacher' 'where's the teacher'

14

Eugeniusz Cyran

In fact, this argument combines the observation that the final position is highly restricted in terms of what segmental types it allows, and the regular occurrence of [a] after the remaining consonant types, in the absence of a vowel that might be provided by the following word or morpheme. In K.P. Mohanan's view, the motivation for the epenthesis is precisely not to allow coda consonants. Thus the existence of words ending in [m] and [n] must be viewed as problematic for this account, especially that the group of the licit word-final sonorants can be extended in formal speech to, for example, [1, r, n]. The data below (from K.P. Mohanan 1986: 74) illustrate this point.

a. b. c. d. e. f.

formal maram awan aan awal paal wayar

colloquial maram awan aarid awaid paals wayaro

gloss 'tree' 'he' 'male' 'she' 'milk' 'stomach'

One may wonder if the extended number of segment types found word-finally in formal speech should constitute enough, or better evidence that codas exist in Malayalam. It is not entirely clear from Mohanan's approach at what stage a subset of segmental contrasts occurring in a given context ceases, or begins to be exceptional, and can be ignored for the benefit of a more attractive generalisation. It seems that in the case at hand, the problem may lie in the nature of the generalisation itself. Note that the decision concerning the (non)exceptionality of the subset of segments occurring word-finally tips the scale either in favour of coda or against it and thus involves a major decision concerning the syllable structure, which may impede its learnability. In this way of viewing things we could expect that the different registers of the same language may have quite disparate syllable structures depending on how many final consonants manage to convince us, or the learner, that the language has codas.

Parameters and scales in syllable markedness

15

No such problem arises if we take a different view on the syllabification of final consonants, that is, one in which final consonants are always onsets (Kaye 1990; Harris and Gussmann 1998). In such a model, the various cut-off points for the number of acceptable final segment types do not affect the overall syllabic design of the language. The larger or smaller subset of final onsets is due to different settings of the licensing potential of the following empty nucleus. This approach allows for doing away with the cumbersome problem of exceptionality in the word-final context. What is most important is that the word-final consonants, whether restricted melodically or not, have nothing to do with the question of there being codas in the language.13 To conclude, in contradistinction to many currently held views, the word-final context is not where we should be looking for clues concerning codas. In this respect the argument in favour of the nocoda hypothesis, which refers to the restricted nature of the wordfinal consonants can be diffused. This new situation calls for a different explanation of what motivates the [9]-epenthesis. However, details of a possible analysis of this problem must be postponed until later. In short, it is not impossible to view the schwa epenthesis as resulting from the pressure exerted on the final empty nucleus by the consonants, which are more complex melodically than [m] and [n]. The realisation of the empty nucleus provides the necessary melodic licensing but does not involve resyllabification, because the final consonants were onsets from the start. In Malayalam the major cut-off point for segmental licensing seems to be placed between sonorants and obstruents.14 Any distinction along the traditional sonority scale can be easily and nonarbitrarily translated into the elemental complexity of segments in GP. In general, all we need to say is that different complexities require different strength from the nuclei that license them. Since obstruents are inherently more complex than sonorants it is not impossible to assume that perhaps [s]-epenthesis occurs after those segments which require more licensing than a mere empty nucleus can discharge. For more details see section 5.3.

16

Eugeniusz Cyran

As for the distinction among sonorants - we will defer any definitive claims until later sections. However, we must mention some possible reasons why nasals may be easier to license than other sonorants, even the less complex ones. It can be observed crosslinguistically that nasals feature quite readily in restricted contexts, e.g., the Japanese nasal + obstruent is the only non-geminate coda consonant (cf. also the so called Prince languages), and [n] is the only consonant word-finally. More examples can be found in Bell (1971). Clearly we are not dealing here with an isolated phenomenon, typical of only Malayalam, and this phenomenon requires a principled explanation. Another line of inquiry seeking to explain the behaviour of nasals can also require looking at possible restrictions concerning other sonorants, especially the mono-elemental ones, that is, glides, which tend to display distributional restrictions of their own across-languages. To conclude the discussion of K.P. Mohanan's phonological argument for the no-coda hypothesis we must say it creates more problems than it provides answers. Firstly, it raises the problem of melodic exceptionality as the outcome of particular syllabic generalisations and forces us to make arbitrary decisions as to what is exceptional and what is not in quite an unrestricted way. Secondly, it is possible that the word-final context, on the basis of which the decisions concerning codahood are made in K.P. Mohanan's model, need not be indicative of the presence of codas at all. Such problems disappear in a model in which the word-final consonants are onsets. Their restricted character is irrelevant to the overall syllabic pattern of the language, but it does not mean that the non-arbitrary patterns of word-final restrictions should not receive a principled account. In what follows, we will demonstrate that such an account is possible. But first, let us look at a couple of phonological facts, which make a rather compelling case for recognising coda as part of the syllable structure of Malayalam. Of course, the context where it appears is word-medial.

Parameters and scales in syllable markedness

17

4.2. Basic facts of clustering in Malayalam Convincing arguments against the no-coda hypothesis for Malayalam, which are based on the distribution of consonant clusters in this language, can be found in Steriade (1981) and T. Mohanan (1989). (11) shows major patterns which will be commented on below. (11)

initial #TRV

Ρh dr,ty,tw

medial VTRV VRTV VRTRV dw, ty, pp, st, nt, ndr, ktr

final VRT#

pp, 11, st, mp

There are good reasons to believe that the clusters appearing word-initially are branching onsets. First of all, only two-consonant clusters are allowed there, and they are severely restricted in that sequences of falling sonority are ruled out just like in English.15 More details of the initial clusters are given below. (12)

a.

obstruent + liquid

b.

obstruent + glide

c. nasal + liquid d. sonorant + glide

plaawam drumam tyaagam pvisa mleecham nyaayam

'flood' 'tree' ' sacrifice' 'luster' 'copper' 'justice'

Word-medial clusters, on the other hand, are not subject to any particular restrictions. The sonority slope can tilt in either direction and, additionally, three-consonant clusters are also found in this context. (13)

a.

tonta ampalam apsara b. indram waktram

'throat' 'temple' 'nymph' 'night' 'face'

18

Eugeniusz Cyran

As observed in Steriade (1981), the fact that word-medial clusters do not exhibit similar restrictions to the word-initial ones is quite surprising if we follow K.P. Mohanan's proposal that they are also complex onsets. One would have to assume that any conditions used to derive the initial restrictions, e.g., those connected with the Sonority Sequencing Generalisation, must be allowed to operate only word-initially, and be ignored word-medially. That this can not be the correct solution is proved by further facts from Malayalam consonant cluster distribution and their phonological behaviour. Let us look at the data from the right edge of the word. 16 The word-final situation in Malayalam is most interesting. In colloquial speech, this language allows only [m] and [n] to occur finally. No other consonants or clusters appear in this context, except that, in formal speech, more sonorants become acceptable, as we saw in (10). The [9]-epenthesis, however, is not limited to the sonorants other than [m] and [n]. It is proposed (see e.g., T. Mohanan 1989) that both obstruents and certain clusters can also be assumed to be underlyingly morpheme-final, in which case, the [9]-epenthesis always occurs regardless of register. A representative set of facts concerning the word-final situation is given below. payarQ paals paato

'beans' 'milk' 'mark'

atuppo antasss bhrast9 panto

'stove' 'status' ' excommunication' 'ball'

In (14a) we have words ending in a single consonant, either sonorant or obstruent, while (14b) shows the licit final clusters. The pre[o J clusters, however, are interestingly limited only to those of falling sonority, i.e. RT#. There are no pre-fa] clusters of rising sonority in the lexicon (T. Mohanan 1989). The clear cut-off point between possible underlying RT# and impossible *TR# suggests that some prin-

Parameters and scales in syllable markedness

19

cipled way of explaining this fact should be sought. Note that the [q]epenthesis creates a context identical to the word-medial position in which no restrictions on the possible clusters are found. It appears then, that the absence of pre-[a] branching onsets in this language provides a strong argument in favour of the view that the licit RT# clusters are indeed morpheme-final, and that [s]-epenthesis is a live process, because only by referring to the final position can the distinction between RT# and *TR# be explained, for example, by referring to sonority sequencing. Note that if the schwa were treated as an underlying vowel then the clusters would be medial rather than final, ι -τ

and medially, no such restrictions would be expected to hold. Given that T. Mohanan's proposal concerning the underlying representations of the forms in (14) above is correct, it appears that stringent sonority restrictions hold not only word-initially but also wordfinally, thus the facts from the right edge of the word in Malayalam further weaken the no-coda hypothesis, and 1 8 the claim that wordmedial clusters are all syllabified as onsets. In addition to the distributional arguments against the no-coda hypothesis, Steriade (1981) also adduces facts from poetic meter, which require that the medial clusters be treated as hetero-syllabic, i.e. coda-onset sequences, for the purpose of distinguishing between light and heavy syllables. Though, admittedly, this type of argument should perhaps be classified together with the phonetic syllabification and the language game arguments used by K.P. Mohanan in support of an opposite proposal, its contradictory import must not leave our faith in word level phenomena unshaken. It is clear that the phonological behaviour of the clusters and the patterns of their distribution point to the fact that Malayalam observes universal well-formedness conditions on syllabification. Thus, the three major types of sequences occurring in word-medial position should be syllabified as V.TRV, VR.TV and VR.TRV respectively. In fact, the universal assumptions concerning phonological syllabification, whether we use the Sonority Sequencing Generalisation principle (SSG), or the theory of government in GP, force us to say from the start that these are the necessary syllabifications. Since the complexity scale presented in 3.2. deals with a typology of syllable that

20

Eugeniusz Cyran

holds at the phonological level rather than at the surface, these arguments should suffice to salvage our proposal. One must remember that both Steriade (1981) and T. Mohanan (1989) assume that, although at the underlying levels Malayalam has the coda and observes SSG, there must be a switch at later stages of derivation to accommodate the facts that led K.P. Mohanan (1982) to the no-coda hypothesis. This is an attempt to reconcile the mutually exclusive facts of this language, which means that at some stage of derivation some readjustment rules and resyllabification must take place by means of which all medial clusters will be relinked to the onset. At this level SSG is assumed not to be operative. We should remain sceptical of these procedures. For one thing, the model of Government Phonology, or any non-derivational model for that matter, must eschew anything akin to resyllabification or levels of derivation. Surface forms are basically what we find underlyingly in terms of structural relations. Differences, such as, for example, filling in of empty nuclei with melody in order to fulfil conditions on segmental and structural licensing may occur, however, they do not involve structural alterations. To put it differently, the theory of government must view the word level phenomena as lying outside the domain of phonology proper. Whether this is an advantage of the model or its doom is an empirical issue. To conclude this part of discussion, we saw that the no-coda hypothesis has little grounding in the phonology of Malayalam. More importantly, Malayalam is not a language which lacks branching rhymes while having branching onsets, and does not constitute a problem for the typological patterns observed in Kaye and Lowenstamm (1981). The complexity scale in GP, which attempts to capture these patterns, is not making wrong predictions about this language. What is more, one of the merits of the scalar approach is that it predicts quite straightforwardly what situation should not be attested. It is not impossible that languages can be found with branching onsets but no rhyme-onset clusters.19 We have stressed the fact that if the final consonants are not identified with codas, a few analytical problems can be avoided, for instance, highly restricted strings instantiating a particular syllabic con-

Parameters and scales in syllable markedness

21

figuration no longer threaten our generalisations about the syllable structure itself. On the other hand, it was crucial for us to prove that Malayalam is not entirely deprived of codas, which, defined in GP as the rhymal complement followed by an onset, are found in this language, just as in any other, only word-medially. Let us now turn to two quite related issues that stem from this discussion of Malayalam. First, we will try to understand why only [m] and [n] are allowed word-finally and how this can be explained in terms of the licensing potential of empty nuclei as distinct from that which characterises phonetically realised nuclei. Note that in this language full vowels license all possible structures. Therefore, the discrepancy between the licensing potential of the two types of nuclei merits a discussion. Second, we will look at the lexical design of phonological strings at the right edge of the word, which becomes more transparent under the GP assumptions of syllabification, and can be captured by referring solely to the licensing properties that different types of vowels in Malayalam possess.

5. Licensing properties of nuclei 5.1. Full vowels vs. empty nuclei Languages make, to some extent, arbitrary choices as to how much they allow their vowels to license. The possibilities are selected along a scale, where the licensing potential is commensurate with the complexity of the licensed structure. The scale itself, therefore, is not arbitrary because it is rooted in the highly restricted phonological representation. In this section, we will consider the relationship between the differences in licensing potential exhibited by full vowels and empty nuclei with respect to syllabic structures, and discuss some limitations on the available choices. Then we will move on to segmental licensing. Recall that the full vowels in Malayalam appear to license all types of structures, while the empty nucleus seems to be limited in its abilities to licensing simplex onsets only. This discrepancy may look

22

Eugeniusz

Cyran

odd or unprincipled at best, our goal is to demonstrate that there is order in this apparent chaos. Consider the full table of structural complexities, with the appropriate settings for Malayalam colloquial speech. (15) Malayalam I c II

c.c

III

cc_

[a]

[0] restricted

[a] = any full vowel, [0] = empty nucleus

The full vowels license all possible syllabic structures, that is, simplex onsets, as well as rhyme-onset relations and the most demanding branching onsets. The empty nucleus, on the other hand licenses only a subset of melodies in the least demanding structural configuration: the simplex onset. At no point is it able to license fricatives or stops, not to mention clusters of consonants. By comparison, in Polish, for example, the licensing abilities of the two types of nuclei go almost hand in hand. For convenience, a table with Polish facts is given below. (16)

„ Polish I

C

11

C.C_

III

cc_

[a]

[01

restricted

In Polish, realised nuclei can license all types of governing relations and similar properties are displayed by the empty nuclei. The difference, however, is that at each level of licensing, be it a simplex onset, a rhyme-onset relation, or a branching onset, the number of possible segments allowed before an empty nucleus constitutes a subset of those with a following full vowel. For example, [0]'s cannot license voicing in obstruents and also some contrasts such as

Parameters and scales in syllable markedness

23

secondary place of articulation, for example, palatalisation of noncoronals. However, the greatest restrictions are found within the highest level of licensing, i.e. the branching onset. Here we find [tr] as the best candidate in Polish, and very little else (Cyran, in press). We may state our first hypothesis concerning the melodic restrictions in the following way: the most stringent melodic restrictions on onsets (C_), and their relations (C.C_, CC_) will occur at the highest levels of licensing that the empty nucleus reaches. This is in perfect accordance with the scalar system we are trying to develop here. Of course, the reverse situation in which Polish would exhibit a whole range of possible final branching onsets, but at the same time would have severe restrictions for simplex final onsets, say, similar to those in Malayalam, is ruled out by the model, and should not occur. Thus, in Polish, the restrictions are found at the level of branching onset, while in Malayalam they occur at the lowest level. Probably the most striking language illustrating our hypothesis is Japanese (Yoshida 1996), which displays melodic restrictions at different levels of licensing for both full vowels and empty nuclei. In brief, it allows its full vowels to license simplex onsets and only two types of rhyme-onset relations, that is, geminates and nasal + obstruent, while its empty nucleus can only license one consonant finally: [n]. (17) Japanese I

c

II

C.CL

III

cc_

[a]

101 restricted

restricted

Thus it seems that we are very likely to find restrictions in the most complex structures that a language allows for and such facts are easily derivable from the licensing scale. Our hypothesis pinpoints the locus of potential melodic restrictions as the highest level of licensing allowed in a given language. Additionally, the comparison between Malayalam, Polish and Japanese seems to suggest that the

24

Eugeniusz Cyran

licensing potential of full vowels and empty nuclei is set independently of one another. In the following section we will attempt to present a typology of licensing discrepancies between the full vowels and empty nuclei in terms of which configurations are possible and which should be ruled out. Then we will return to the question of cut-off points for segmental licensing.

5.2. A typology of licensing discrepancies The relevant facts here concern only the licensing of syllabic structures, i.e. the three possible levels. We have seen that the licensing properties of vowels and empty nuclei may be identical, as in Polish, or they may differ markedly, as in Malayalam. This is part and parcel of the model and this is where most surface typological variation between languages originates. Thus, despite the restrictiveness of the model as to the maximal size of syllabic constituents, i.e. binary constituents, a fair amount of variety follows from the fact that languages select the licensing properties of their empty nuclei independently of the properties displayed by full vowels. What is more, languages may choose not to have empty nuclei word-finally altogether. One language, typically mentioned in this context is Italian, which has no empty nuclei word-finally, at least in content words, but whose full vowels allow for a whole range of structures, i.e. I, II, and III. Polish seems to take an opposite position to Italian in allowing for identical licensing possibilities for both filled and empty nuclei. In this context Malayalam represents one of the intermediate states/grammars, with its full vowels displaying similar properties to those of Polish and Italian, and empty nuclei only barely reaching level I. One can mention another option in this system, for example, Turkish, in which the full vowels license all structures, but the empty nucleus reaches level II. That is, word-finally, the language has both simplex onsets and rhyme-onset relations ([sart] 'rope'), but it does not have branching onsets in contradistinction to Polish, French or

Parameters and scales in syllable markedness

25

Icelandic. The table below illustrates the points we have just made. The full vowel box refers to all four languages. (18) Pol,Tur,Mal,Ital

Polish

Turkish

Malayalam

Italian

[a]

[01

10]

10]

[01

I II

c.c_

III

CC

It appears that there is no dependency between the selections made for full vowels and those for empty nuclei with respect to licensing abilities. To add to the variation we illustrate what would appear to be licensing scales for Hungarian, Japanese, Yucatec Maya and Zulu below. Hungarian allows both types of nuclei to reach level II, i.e. to license rhyme-onset sequences and, naturally, the lower, more simplex structures, but it excludes branching onsets (level III). Japanese allows for rhyme-onset sequences only before a full vowel, and simplex onsets before [0]. Yucatec Maya has no adjacent clusters but allows for a single consonant word-finally, while Zulu only has full vowels and simplex onsets. (19) Hungarian [a] I II III

101

Japanese [a]

[0]

Yucatec Maya [a]

[0]

Zulu [a]

f0l

c c.c_ cc_

It transpires that, contrary to our earlier assumptions, the relation between the settings for full vowels and empty nuclei are not entirely free. We have not seen a case where [0] would license more material than a full vowel in a given language. In fact, two restrictions must be mentioned that seem to hold in connection to the settings of licensing strength between full vowels

26

Eugeniusz Cyran

and empty nuclei. First, what we do not expect in this model is a discontinuity of licensing potential reflected in the absence of intermediate structures as illustrated in (20a). (20) *a. I II III

[a] c_ c.c_ cc

[01 ?

?

*b. J II III

[a] c.c_ cc

[01

?

The discontinuity of structures in which branching onsets are present but rhyme-onset relations are absent is exactly the problem that induced us to look at Malayalam. Likewise, no language should have complex onsets only, or have rhyme-onset clusters but not simplex onsets. These limitations refer to both types of licensers in (20a). A second restriction that could be proposed concerns the possibility that a language may select higher licensing potential for its empty nuclei than for its full vowels (20b). This excludes a number of impossible systems. For example, one in which RT sequences are found word-finally but not word medially. This restriction also excludes languages in which full vowels do not license anything, i.e. systems with only an arbitrary repetition of onsets and empty nuclei. Thus, problematic for our model would be a language which can only end its words with a vowel and complex clusters, as in (20b), or in which more complex structures are attested in the absence of less complex ones (20a). But even less extreme situations, like the hypothetical one in which we would find branching onsets word-finally but not initially or medially, would constitute a problem. What we expect is the opposite, that is, cases with initial and medial branching onsets but not final (e.g., Italian). The first restriction expresses the simple implication that simpler structures must be present in systems which have more complex structures. A possible explanation for this implication, which would be compatible with the markedness scale, is not difficult to think of.

Parameters and scales in syllable markedness

27

As for the second restriction, one of the reasons why the model disallows systems like (20b) is that empty nuclei constitute, by definition, weaker versions of their melodically sounded counterparts. This restriction may also be connected with the origin of empty nuclei, which presuppose full vowels in the system. It may be assumed that learners can only postulate [0]'s if they can do it on the basis of an existing "non-defective" structure. Hence the implications illustrated in (21) below require no further grammatically driven justification. (21)

[0] implies [a] (becase [0] originates from [a], say, historically, and in language acquisition.) C0 D Ca (C = any consonant, a = any full vowel) VC.C0 3 VC.Ca (C.C = R.T) V.CC0 => V.CCa (CC = TR)

Though we will not elaborate much on this point here, it is possible to derive similar implications for empty and filled onsets. What we have briefly illustrated above are fairly tentative approximations of what different nuclei can license in terms of syllabic structure and how these abilities may differ for individual languages thus producing a variety of surface syllable types. This is done in a model, which restricts the size of syllabic constituents to maximally branching ones, and possible syllable structures to three levels of complexity. In this context the Malayalam facts cease to look exceptional and in fact are welcome in that they fill in the factorial typology of syllabic types. Below, we deal with two final though quite interesting points concerning Malayalam. The first one can be informally dubbed as the "fine tuning" of licensing properties of nuclei, that is, scalar effects in licensing of melodies within the licit syllabic structures. The second point gathers together the possible syllabic and melodic licensing properties of nuclei in Malayalam in order to demonstrate that the distribution of strings and segments at the right edge of the word in this language forms a pattern which directly reflects the relative strength of nuclei, and hence, it can be coherently described without reference to any additional constraints.

28

Eugen iusz Cyran

5.3. Melodie restrictions on final onsets So far we have looked closely at one type of complexity requiring different degrees of licensing strength from the following nucleus. Each time the recipient of licensing is the preceding onset which can be simplex as in Polish tata [tata] 'father', a simplex one with a preceding coda consonant to license as in Marta [marta] 'name', or a complex one (branching onset), e.g., mokry [mokri] 'wet'. The required licensing strength is different for each structure, and the syllabic complexity has cut-off points between levels I and II and between II and III. The type of complexity we are turning to now is measured in terms of the number of elements present in a segment. This type of complexity has been discussed in the GP literature by, for example, Harris (1990, 1994, 1997), Cyran (1996), Scheer (1996), Gussmann (1998), and Rennison (1998). The model that we will follow is that of standard GP (Harris 1990). Syllabification can be understood as stemming from the interaction between different amounts of syllabic structure and the strength of the licensers, the nuclei, and there is asymmetry in the degree of structure that can be licensed by full vowels and empty nuclei. Similar asymmetry is to be expected in licensing of melodic material, which also exhibits complexity slopes. In this sense, the same principle governs the way in which licensing properties of nuclei can be applied both to syllable types and segmental material. This is a desirable situation given that governing relations, it will be recalled, are conditioned by segmental complexity of the consonants. Segments in GP are composed of one or more elements, and in general, more complex segments should be more difficult to license than the less complex ones. For example, it should be easier to license mono-elemental or bi-elemental segments than those containing three or four elements. Below we illustrate what such a scale translates to in terms of segment types in GP. The model makes use of the following elements: A, I, U, ?, h, N, H, L.21

Parameters and scales in syllable markedness

(22) sonorants fricatives stops

29

(I) = j,(A) = r, (U) = w, (A,?) = 1, (U,N) = m, (Α,Ν) = η (A,h,H) = s, (U,h,H) = f (U,h,?,H) = p, (A,h,?,H) = t

The representations above are not of any particular language. They are rather rough schemes corresponding to major classes and will have to be refined when a language is carefully analysed. Since simplex onsets typically exhibit a full range of segmental contrasts before a realised vowel, the complexity effects, logically, will be observed when the licenser is an empty nucleus, i.e. in the case of word-final consonants, but the whole picture does not exclude full vowels, as we will see below. Let us look at the facts concerning the final simplex onsets in Malayalam. Initially, we will follow the assumption made in T. Mohanan (1989) that underlyingly both sonorants and obstruents can be found in word-final position. For the purpose of completeness we add a group of simplex onsets - simplex in the syllabic, not segmental sense which are borrowings from Sanskrit where they appear as final. However, in Malayalam, they may not appear even underlyingly in this position. Such stems, marked as [+Sanskrit] are treated by the language as bound morphemes and always appear affixed, i.e. a full vowel must follow these consonants (T. Mohanan 1989: 622). (23) Sonorants

informal

formal

[01

[01

awa[n] 'he' mara[m] 'tree'

[m], [n]

Dl· Μ, Μ

example

bl

[0]

avra[ta] 'she' waya[ rs]' stomach'

Obstruents

Μ

[3j

~3 elements Sanskrit obstruents

katalaa[ss]

'paper'

wira[ka] 'firewood'

[a]

[a]

laa[bha]m 'profit' />aa[tha]/w 'lesson' ma[da]m 'intoxication'

30

Eugeniusz Cyran

Necessary details of segmental representations in (23) need to be worked out. However, a general pattern seems to transpire even if such details are left aside. First, let us observe that the growing complexity of segmental types dictates a growing (corresponding) licensing strength that is required from the following nucleus. Given the schemes in (22), there is no real problem to distinguish between sonorants on the one hand and obstruents on the other. All we need to say is that the cut-off point is more or less between two and three elements, in that more complex segments than bi-elemental cannot be licensed by [0], even in formal speech. The problem for us lies in the description of the situation in colloquial speech where mono-elemental segments like I, A, U in [j,r,w] are worse-off than bi-elemental nasals [m] (U,N) and [η] (A,N). We mentioned this problem earlier and suggested some ways of dealing with it. Let us, however, not distract ourselves from the general pattern, which seems to be true. It is interesting to note how Sanskrit words seem to be borrowed into Malayalam. Note that none of the Sanskrit consonants in (23) can be allowed finally in the underlying representation even though Malayalam possesses the process of epenthesis to deal with final consonants. One cannot fail to notice that these consonants might be a bit more complex than all those followed by schwa. In other words, the clues as to their class membership ([+Sanskrit]) are present in the segmental make-up (see also K.P. Mohanan 1986: 82). For example, some of the consonants are voiced, e.g., [d], which is not allowed finally in Malayalam. Clearly, if we wanted to apply the universal sonority scale to account for the cut-off points in (23) we would be hard put to explain why a voiced obstruent as in [madam] is separated from sonorants by a voiceless, that is, less sonorous obstruent [wiraks]. In other words, for the sonority scale the facts in (23) make no sense. In element theory, on the other hand, the actual factor responsible for this gradation of segments, can be expressed by the relatively more complex nature of [d], when compared with, e.g., [k]. This can be done by assuming that the voiced segment has an additional element specifying the laryngeal activity, i.e. has the element L, while [k] is neutral, or un-

Parameters and scales in syllable markedness

31

specified in this respect. Though it is only an assumption, and we will not attempt to verify it here, it is supported by the fact that the other consonants in (23) also bear obvious additional laryngeal specification. The word for 'lesson' has aspiration, that is, the element Η is present. All in all, it seems non-accidental that the mark [+Sanskrit] appears to go hand in hand with a greater complexity of segments involved. The lexical condition on what these consonants are followed by has a clear phonological basis. It appears then, that the distribution of [0], [a] and the full vowel which follow the word-final consonants can be neatly captured by the scale of licensing required by the different types of segmental complexity. The empty nucleus can only license selected sonorants in colloquial speech, and a larger set of sonorants in formal speech.23 The obstruents require a vowel with melody, hence they are always followed by schwa, and whether this schwa is lexical or epenthetic does not really matter because the even more complex segments of Sanskrit origin must be followed by a vowel of a different category, that is, a full vowel. Interestingly enough, the licensing requirements of the underlying final onsets are reflected in the licensing requirements of syllabic types. The full picture of the right edge of the word in Malayalam is discussed in the following section.

5.4. The right edge of the word in Malayalam In 5.1, we established that full vowels in Malayalam are able to license all three types of syllabic structure, while [0] only sanctions a restricted number of simplex onsets. (24) Malayalam I II III

c c.c cc

a

[01

32

Eugeniusz Cyran

We also saw in the previous section that in surface terms the schwa vowel seems to occupy an intermediate position between the full vowel and the empty nucleus. It licenses more structure than the [0] but less than the full vowel (a > 9 > 0). This is particularly evident in the case of Sanskrit borrowings in which the final consonants appear to require a full vowel as opposed to a mere schwa, not to mention [0]. Some support for the way we view these facts, that is, as stemming from the relationship between complexity and the licensing that the different complexities demand from the nucleus, can be found in the distribution of word-final syllabic types, which is basically a replica of the segmental scale in (23). (25) medial

word-final

simplex onset C_

[a]

[01

rhyme-onset C.C_

[a]

example flwfl[n] 'he' mara[m] 'tree'

bJ

atu[pps] 'stove' pa[lb]

'tooth'

paa[mpa] 'snake' branching onset CC_

[a]

[a]

/?a[tra]w? 'letter' t'fl[dra]wz 'night'

Some melodic details will be supplied below in the final table depicting the distribution of segments and strings at the right edge of the word in Malayalam. Let us concentrate on the word-final context. What the table illustrates is that there may be an underlying simplex word-final onset which will also surface as such, i.e. not followed by any type of overt vowel, e.g., [maram]. Some word-final clusters are allowed at the underlying level but they must be of a particular type, i.e. RT#, which includes sonorant + obstruent, s + obstruent, or geminates. These sequences are followed by a schwa on the surface. TR sequences never appear before a schwa.24 For example, words ending in TR in Sanskrit are, like some of the segments discussed previously, treated as bound morphemes and must be affixed. In ef-

Parameters and scales in syllable markedness

33

feet, then, they are followed by a full vowel. Again, let us look at some of the examples provided by T. Mohanan (1989: 622). (26) can[dia]n wak[tm]m in[

[fis]> f. b.

*V Ν [-nas]

*NT

DEP

[+nas]

MAX

/+nas/ *!

*v[+nasl

MAX

/a

5 4 Patricia Donegan Tableau 6b. Perception of 'dent': [det] given c. d. ^e.c/dent/, [dst]>

*V Ν [~nas]

*NT

DEP

[+nas] *

•I

MAX

*V[+aas]

/+nas/

MAX

/c/ *

**

**t

*

*

*

4. The role of Segmental Well-Formedness constraints It is generally recognized in OT that SegWF constraints may dominate Faithfulness constraints or be dominated by them, and that the appearance of a segment in a language depends on this domination. In the interpretation suggested here, SegWF constraints evaluate URs, so that URs are composed of optimal segments. Where SegWF constraints dominate relevant Faithfulness constraints, they universally eliminate certain potential segments from the underlying representations of the language - i.e. they constrain the segment inventory. For example, if *[+Voi, -SON] dominates MAX [VOI], then voiced obstruents will be eliminated from URs. Since WF constraints are presumed to dominate Faithfulness constraints in the initial ranking, no additional mechanisms are required to achieve this limitation on the input and its corresponding characterization of perception. Note that this does not mean that the SegWF constraints necessarily eliminate segments from the language altogether. A SeqWF constraint may dominate a SegWF constraint and thus require the occurrence of segments that are ruled out of the UR inventory. (This mirrors NP, where a Lenition that nasalizes vowels before nasals may apply after, and thus override, the process that eliminates nasalized vowels from the inventory). This interaction of Segmental WF constraints with Faithfulness constraints thus limits, to some extent, the "freedom of the input" or the "richness of the base". However, it retains the principle that the "locus of explanatory action" is still the ranked constraint hierarchy (Prince and Smolensky 1993: 3). Systematic gaps and omissions in a speaker's outputs are not explained merely by gaps and omissions in

Constraints and processes in phonological perception

55

the input. They are still explained by the phonology itself. (This is, of course, exactly what one would expect of any phonological theory which attempts to encompass phonological acquisition and change compare, e.g., Stampe 1973, 1987; Andersen 1973, etc.). But there are gaps and omissions in a language's lexicon. They reflect the surface patterns (not the other way around), and they also reflect principles of segmental well-formedness. This limitation on "possible inputs" corresponds to speakers' frequent perception of equivalence of sounds that differ only by nondistinctive features. For example, speakers of English judge bed and bend - and, for that matter, debt and dent - to have the same vowel. Similarly, English learners of Khmer often have trouble perceiving the difference between /kaa/ 'work' and /koo/ 'neck', or between /koo/ 'neck' and /koo/ 'mute'. In OT, the SegWF constraints might be taken to perform the function that the Fortitions perform in Natural Phonology - that of constraining speakers' perceptions and thereby constraining, or in fact, creating, the phoneme inventory. To illustrate the role of SegWF constraints, we can compare their ranking and its results for the input inventory (or phoneme inventory) in three languages: English, French, and Hindi. Consider the following constraints: VN

Vowels

are nasalized

before nasal

consonants

(SEQWF). *V[+NAS] M A X /nas/ DEP [nas]

Vowels are not nasalized ( S E G W F ) . Nasality in U R corresponds to nasality in S R Nasality in S R corresponds to nasality in U R

(FAITH). (FAITH).

These constraints, ranked differently, account for vowel nasalization: (1) In English: VN » *V » DEP[nas], MAX /nas/. As a result, nasalized vowels occur before nasal consonants in SRs, but the S E G W F constraint, *V, ensures that vowels in UR are all nonnasal. Nasalized surface vowels are perceived as plain. (But note that they are perceived - rightly, as in bend [bend], or

56

Patricia

Donegan

"wrongly", as in bent [bet] - as occurring before nasal consonants.) (2) In French, the faithfulness constraints MAX/nas/ and DEP[nas] dominate VN and *V. With the faithfulness constraints most highly ranked, both nasalized and nonnasalized vowels occur in URs, and they appear as nasalized and nonnasalized vowels, respectively, in SRs. The nasalized SR vowels are perceived as themselves. (3) In Hindi, VN » DEP[nas], M A X / n a s / » *V. As a result of this ranking, both nasalized and nonnasalized vowels appear in URs, and some nonnasalized UR vowels are nasalized in SRs. These vowels can be perceived as themselves, since the faithfulness constraints dominate *V. See Tableaux 7a-b. Tableaux 7a-b. Vowel Nasality in Hindi: Production and Perception. VN: Vowels are nasalized before nasal consonants. » sponds to nasality in the SR. »

MAX [nas]: Nasality in the UR corre-

DEP[nas]: Nasality in the SR corresponds to na-

sality in the UR. *V : Vowels are non-nasalized. Perception of [jä:na:] as /jo:na:/ depends on morphological identification with other forms, like /j a:o:/ 'go away!'. 7a. Production

vN /ha:/ 'yes' [ha:] [hä:] /kya:/ 'what' [ky&] [kya:] /jaa-na/ 'to go' [jä:na:] [ja:na:]

MAX

/nas/

DEP

[nas]

*v

*! *

*

*!

*

Constraints and processes in phonological perception

57

7b. Perception

vN [hää] 'yes' /haa / w /hää/ [kya:] 'what' /kyä:/ /kya:/ [ja:na:] 'to go' /jama:/ /ja:na:/

MAX

/nas/

DEP

[nas]

*v

*T *

*

*

*!

This corresponds to Prince and Smolensky's notion of "Stampean occupation", the idea that the UR is, as far as possible, identical to the SR; this means that an output is perceived faithfully in the absence of constraint rankings that require otherwise. This is modified by "lexical optimization", which favors URs from which alternative SRs can be predicted, and which is sometimes thought to favor absence of markings in URs.

5. Perception based on nondistinctive features: other cases Although the usual situation is for speakers to ignore feature differences in SR that are not distinctive, the example of vowel nasalization in English (as in 5a-b) is by no means an isolated instance of perception of a distinction based on a non-distinctive phonetic feature. There are many cases in which nondistinctive features provide surface marking of a distinction which has itself been neutralized. Here are a few examples: (1) Vowel lengthening or diphthongization marks final obstruent voice in some dialects of American English that have final obstruent devoicing. Thus [bs:t] or [beat] for bed is distinguished from [bet] bet. There is a similar situation in some German dialects. (2) Mortlockese speakers perceive English medial voiced and voiceless stops as distinct even though voicing is nondistinctive in

5 8 Patricia

Donegan

Mortlockese, because these medial voiced and voiceless stops are perceived as the surface forms of single stops vs. geminates, respectively (Paul Lassettre, p.c.). (3) In conservative dialects of Japanese, [t] is not distinguished from [ts] nor from [tj], yet when final vowel deletion occurs in /matu/, giving [mats], or in /matji/, giving [matj], the two are not perceived as equivalent. The English word match is perceived as /mati/, the name Mats is perceived as /matu/ {mat is perceived as /matto/) (see Ohso 1972; Vance 1987b). (4) In most American English dialects, pairs like kitty ['kiti] and kiddie ['kidi] are homophonous as ['kiri] when the coronal stop is flapped and voiced. But in some dialects, pairs like writer and rider are distinct even when Flapping occurs, because the stressed vowels differ: ['rajrr] is perceived as rider and f r s i r r ] is perceived as writer, although /ai/ and /si/, being in complementary or nearly complementary distribution, are usually regarded as non-distinct. (But see Vance 1987a; Donegan 1993.) (5) Although Korean has only one liquid, the English medial /l/ vs. /r/ difference is perceived by Korean speakers - as a gemination difference. A word like English belly is perceived as /beLLi/, and berry is perceived as /beLi/ in Korean. In each of these cases, the phonetic distinction ([ε:] vs [ε], [ρ] vs [b], [ts] vs [tj"], [-i] vs [1] is produced (by native speaker or by foreigner) with a feature difference that is not ordinarily recognized as a distinctive feature in the language. It is perceived as a different difference - one that is distinctive in the perceiver's system. Maximal faithfulness does not offer a satisfactory solution. If, as in the Smolensky proposal, hearers perceive with maximal faithfulness and lexical retrieval is the cause of their confusions, then hearers ought not to perceive the nondistinctive differences cited here at all. Pater's account of asymmetrical production and perception is not relevant here, since these cases do not show such asymmetry.

Constraints and processes in phonological perception

59

6. Opaque Surface Representations in OT These cases involve opaque SRs, which are generally problematic for OT. As shown above (Tableaux 5a-b), it is necessary for OT to admit SegWF constraints as part of the perceptual evaluation. And yet the admission of SegWF constraints is not sufficient to account for all such cases. The Japanese [ts] vs [tj] case is an example. As shown in Tableau 8 on page 60, which shows evaluation for production, incorrect SRs are predicted when final (devoiced) vowels are deleted. indicates the winning candidate. Correct SRs which are not predicted by any constraint ranking are indicated by For /mati/, [matj] is selected by MAX [+pal]; [tj] preserves in the SR a feature of the UR palatal vowel I'll. But this requires a somewhat non-standard interpretation in which MAX constraints are satisfied when a feature of one segment in the UR appears on a different segment in the SR. (This was also required by the nasalization case in Tableaux 5a-b.) But for /matu/, this won't work, since [ts] does not directly preserve in the SR any feature of the deleted vowel. There is no ranking of the relevant constraints which would favor [kats] over *[kat] for/katu/. Tableau 8. Japanese production of affricated stops before deleted high vowels. Assumes a style in which *V[hi]/[-voi] is ranked above MAX V. (Devoiced vowels are deleted.) Incorrect SR is predicted when final (devoiced) vowels are deleted. Correct, but non-predicted SR is indicated by < . STOP:

Stops are unaffricated (SegWF). C A F F H I : Coronal stops are affricated be-

fore high vowels (SeqWF). CPAL: Coronals are palatal before high palatal vowels (SeqWF). *V[hi] /[-voi]: High vowels in devoicing environments are not produced (SeqWF). 7 MAX V : A vowel in input corresponds to a vowel in the output (Faith). ID 1-0 C: Consonants are identical in input and output (Faith). MAX [pal]: Palatality in the input corresponds to palatality in the output (Faith). DEP [pal]: Palatality in the output corresponds to palatality in the input (Faith).

lDl-0 c MaxV Stop

* *

*

CS

οe ο Ό Ο kl Oh

06 a

δ

*'

*

* *

*

CPAL

*

*

* *'

*

CAf Hi

ω (Λ (υ C ccx a

* *

*

*

*

* *

*

*

*

*

*

Dep [pal]

*

*

Max [pal]

* *

No Coda

Donegan

*V[hi] / [-voi]

6 0 Patricia

*

*

JÜ 0

£

"

ί^Γ ι

1

X


*

*

*

*

*

*

*

CPAL/PAL

*

Ω m Ω E CTF

ex 03 G

*

JXJ Ü-


r is to enlarge the candidate set in a very limited way, stipulating the phonologically unnatural phenomenon of r epenthesis." However, we have seen that an output-oriented functional explanation like the one defended in the preceding section can provide us with a very good reason why this change, far from being unnatural, straightforwardly descends from speakers' natural behavior. As Donegan (1993: 118) puts it: In these dialects [i.e., those with r-intrusion, LG] the lenition applies obligatorily. Therefore speakers find final r's difficult or impossible to pronounce ... The 'intrusive r' does not, then, intrude because the speaker makes up an r-insertion rule. Instead, the r appears by analysis, when speakers assume that, because some final schwas represent /r/'s, other final schwas do so as well.

As can be seen, the OT-account is rather controversial.6 However, let us proceed further by discussing another case that highlights an intrinsic weakness of OT.

5. Extrinsic rule ordering and two-levelled phonologies Delving into the English r-phenomena discussed in the preceding section, we face another fact that apparently points to an intrinsic weakness of OT. In some varieties of English (cf. Wells 1982: 238ff.) there is a smoothing rule, according to which off-gliding

118

Livio Gaeta

diphthongs get a monophthongal realization when occurring in a prevocalic environment: (16)

chaos /'keios/ —» ['khe:Ds] coward /'kaued/ —> ['kha:9d] /'gsüiq/ ['g3iiq]

R-insertion is not a possibility for these words. We thus conclude that the smoothing rule must be ordered after r-insertion: (17)

a.

/'g^uiq/ Smoothing [g3:iq] R-Insertion *['g3:nq]

b.

/'gauiq/ R-Insertion Smoothing

['g3:iq]

What we have here is a classical case of counterfeeding order, since the reverse order of the rules would incorrectly produce a form with r-insertion (cf. (17a)). In this respect, Booij (1996: 277) claims that "counterfeeding order is a form of extrinsic rule ordering that is clearly at odds with non-derivational phonology". He mentions a case from Dutch, in which non-native words ending in a syllable with VC rhyme that does not bear main stress exhibit vowel lengthening. The vowel of the last syllable is lengthened before non-native suffixes, which are all vowel-initial, as in the following examples: (18)

kdn[o]n 'canon' organisät[o]r 'organizer' sdt[a]n dlfab[e]t 'alphabet' profess[o]r

kan[o:]niek 'canonical' organisat[o\]risch 'organizational' sat[a:]nisch 'satanical' alfab[e:]tisch 'alphabetical' profess[o:]räal 'professorial'

Words that do not meet the non-stressed syllable condition do not display vowel lengthening:

Striving for optimality

(19)

ton kanon 'gun' libretto

119

t[o]nndge 'number of tons' kan[o]nnier 'gun man' librettist

In a two-levelled phonology like the model of Lexical Phonology advocated by Booij, there are two basic ingredients to account for this alternation: cyclic assignment of stress, and extrinsic ordering of vowel lengthening before stress assignment. The derivation for kanonnier is as follows: (20)

1st cycle Main Stress Rule 2nd cycle Vowel Lengthening Stress Erasure Main Stress Rule Secondary Stress output

[ka:non]yv ο [[ka:n5n]Ni:r]N

ι: [kanoni:r]

We have to put Vowel Lengthening before Stress Erasure, because otherwise the latter would feed Vowel Lengthening producing incorrect results (lengthening of the second vowel). This is clearly excluded for a non-derivational approach like OT. However, I am not convinced that counterfeeding order represents such a dramatic problem for OT. For instance, we can easily account for the nonappearing of r-intrusion in the case seen in (17) above by assuming a constraint that prohibits a glide from occurring intervocalically (*VGV), provided that our generator contains the rule enlarging the candidate set seen in the preceding section:7 'gsuiq 'gauir) -» 'gsnrj 'gsinq

*VGV

MAX-ΙΟ

DEP-IO

**

*

*!

**

120

Livio Gaeta

A much more difficult case for a non-derivational approach to phonology such as OT is represented by the specular case of counterbleeding order. I will discuss this phenomenon on the basis of data taken from Algares, a variety of Catalan spoken in Alghero (cf. Loporcaro 1997). In Algares, there is a rule of i-epenthesis in words ending in obstruent (possibly followed by a liquid) when a consonantal onset follows (cf. Loporcaro 1997 for details): (22)

[poti 'vara] 's/he can see' ['fetsi 'be] 's/he did well' ['tegki 'fret] Ί am in a hurry'

/pot/ 's/he can' /fets/ 's/he did' /teqk/ Ί keep'

The rule only applies between consonants in external sandhi, and does not take place in prepausal (cf. (23 a)) or prevocalic position (cf. (23b)): (23)

a. _ ### [ke a'rip(*i)] 'that I'm coming' [ηο sa 'pot(*i)] 'one cannot' b.

##V[a'rip(*i) a'vuj] 'I'm coming today' ['pot(*i) ari'ba] 's/he may arrive'

According to Loporcaro, the following rule accounts for the epenthesis: (24)

0

[i] / *C]PW _ PW[C [where *C cannot be syllabified]

Since syllabification is directly called for in the rule, Loporcaro assumes the following syllabification algorithm: (25)

a. Nucleus: any vowel b. Head: i) any intervocalic consonant e.g., [bi.'za.ju] 'great-grandfather'

Striving for optimality

121

ii) two intervocalic consonants, iff Ci = stop/fricative, C2 = rhotic/glide e.g., [u.'bri] 'to open', [as.ta'sjo] 'station' c. Coda: any postvocalic C, not syllabified as a head e.g., ['mas.kra] 'male' In this perspective, i-epenthesis allows final obstruents to be syllabified. We can represent this alternation in OT-terms by making use of very general constraints that enable us to dispense with the syllabification algorithm (cf. Gaeta in press): (26)

Input segments must have output correspondents NO-CODA(OBSTR): Obstruents (but s) may not occur in a coda D E P - I O : Every segment in the output must have a correspondent in the input MAX-ΙΟ:

Notice first that PARSE must dominate NO-CODA since the final consonant is realized in absolute final position as shown in (23a): (27) pot

NO-CODA

MAX-IO

*

—» pot po

*!

Moreover, NO-CODA must dominate D E P - I O to allow for iepenthesis: (28) pot vera

MAX-IO

pot 'vera

DEP-IO

*!

—» poti 'vera po 'vera

NO-CODA

*

*!

122

Livio

Gaeta

Loporcaro points out that extrinsic rule ordering is necessary to prevent resyllabification destroying the context for i-epenthesis in the following case, where resyllabification is theoretically possible according to the algorithm given above: (29) a. [as'tiki ru'miqt] 'I'm sleeping' b. *[as'ti.k ru'miqt] ['poti ru'mi] 's/he may sleep' *[po.t ru'mi] ['sepi Faq'ta] Ί know how to wash' *['se.p raq'ta] Syllabification across word boundaries is rather widespread in Algares since, as we have seen above in (23b), it blocks i-epenthesis when a vowel-beginning word follows a final unsyllabified consonant: (30)

[a.'ri.p a.'vuj] 'I'm coming today' ['po.t a.ri.'ba] 's/he may arrive'

Thus, in the rule-based approach adopted by Loporcaro we must assume the counterbleeding rule ordering in (31a) to obtain the correct output, since in the reverse order syllabification bleeds epenthesis (cf. (31b)): (31)

a. /'pot ru'mi/ 'poti ru'mi 'po.ti. ru.'mi

b. /'pot ru'mi/ epenthesis po.t ru.'mi resyllabif.

—

resyllabif. epenthesis

In OT-terms we can easily account for this effect by assuming two further general constraints like the following ones (cf. McCarthy and Prince 1993): (32)

ONSET: Every syllable must have an onset ALIGN-L: Align (Stem, Left, PrWd, Left)

Now, this handful of constraints is able to select the correct output quite straightforwardly:

123

Striving for optimality

(33) pot fumi

ONSET

MAX-IO

NO-CODA

ALIGN-L

*J

'pot fu'mi 'poti fu'mi

•

'po.t | fu'mi

*!

'po fu'mi

pot afiba 'pot afi'ba

DEP-IO

*!

ONSET

MAX-IO

NO-CODA

ALIGN-L

DEP-IO

*

*!

'po.t | afi'ba

*

'poti afi'ba

*!

'po afiba

*!

*

*

However, this neat account collapses, if we try to generate the optimal candidate for words ending in obstruent in a prepausal context: 134) pot

MAX-IO

pot

DEP-IO

*!

poti po

NO-CODA

*

*!

For the candidate harmonic with this constraint ranking (indicated by the apple in the tableau in (34)) shows i-epenthesis, which is incorrect (cf. (23a) above). To select the correct output, we are forced to assume two different rankings for the lexical and for the postlexical level. At the lexical level, constraints must be ranked in the following way, where D E P - I O dominates NO-CODA:

124

Livio Gaeta

(35) pot

DEP-IO

MAX-IO

NO-CODA

—> pot

*

poti

*!

po

*!

Booij (1996, 1997) also discusses problematic cases of counterbleeding order, such as a Dutch devoicing rule that affects obstruents in syllable coda. He observes that two different levels (and accordingly different constraint rankings) must be assumed to account for the fact that the final consonant of a verb followed by a clitic beginning with schwa is first devoiced and then resyllabified as onset of the following syllable, since Dutch does not allow for schwabeginning syllables: (36)

(ik) heb 't Ί have it'

/hebst/

(hij) had't 'he had it'

/ h a d st/

((ik) heb 'r Ί have her'

/heb sr/

(Πε)σ(ρ9ί)σ (ha) σ

(ΐθί)σ

(he)a(p3r)a)

A similar problem, dependent on rules applied at the post-lexical level, is also posed by Algares, since there are optional rules of consonant deletion as in the following example (cf. Loporcaro 1997: 216): (37)

/a'kap de traba'Xa/ [a'kapi de traba'Äa] [a'kapi e traba'Aa]

Ί finish working' i-epenthesis /d/ 0

In this case, the surface form exhibits the i-epenthesis seen above, which can only be due to the underlying presence of an obstruent subsequently deleted in the derivation. This is however expressly denied by OT, which makes it impossible to refer to an abstract intermediate level of derivation. Booij (1996: 72) stresses that the necessary reintroduction of a certain degree of serialism is not in princi-

Striving for optimality

125

pie incompatible with OT. In my opinion, however, the assumption of any kind of serialism represents a very unsatisfactory result, since the main goal of the theory is to avoid serial derivation and to allow for parallel computing of possible output candidates. Under this view, the assumption of an abstract intermediate level would constitute an aporia for the theory.

6. Constraints and natural processes After having pointed out the weakness of an OT-approach to language change, let us relax the perspective and look at OT from the viewpoint of a functionalist. In this section I will pursue this aim by discussing another recent review in OT-terms of a very well-known language change: the West-Germanic Gemination (= WGG). WGG is a language change that affected Old English, Old Saxon and Old High German, as is shown by the following examples: (38)

a. b. c. d.

Go. sat.jan -skap.jan Go. kun.jis hal.ja Go. naqaps aqizi Go. akrs ON ep.le

OS

settian OE settan skeppian scieppan kunnies OS hellia OHG nackot ackus akkar OS ceppel OE

'to set' 'to create' 'race GEN' 'hell' 'nacked' 'axe' 'acre' 'apple'

With respect to Gothic and Old Norse, West Germanic languages display gemination of a consonant when the latter happened to occur intervocalically before a glide or a liquid. Notice that only voiceless obstruents geminated before a liquid (cf. (38d)). According to Murray and Vennemann (1983), this language change is explained by the so-called Preference Laws for Syllable Structure (cf. Vennemann 1988). Murray and Vennemann put forward the idea that this change was determined by the rise of marked syllable structures as a consequence of the Germanic preference for bimoraic stems. In

126

Livio Gaeta

stems containing a short vowel such as the ones reported in (38), the following consonant was syllabified in the coda of the first syllable in order to provide the stem with the preferred moraic structure. As a consequence, the glides and the liquids were in the syllable onset: ο

(39)

+

sat.jan skap.jan + kun.jis + hal.ja + nak.wat + ak.wizi + ak.rs + ap.li

+

In this way, bad syllable contacts occurred, since the consonant closing the preceding syllable was much stronger than the following one in terms of the Scale of Consonantal Strength reproduced below: (40) glides

1

r

1

nasals

voiced fricatives

2

3

4

5

voiced stops 6

voiceless voiceless fricatives stops > 7

8

Gemination took place in order to improve the syllable contact, while at the same time preserving the bimoraicity of the stem: (41)

+sat.jan + nak.wat + hal.ja + ap.la

+

sat.tjan nak.kwat + > hal.lja > +ap.pla > >

+

Now, this is a typical functionalist approach. The how of the language change consists in the doubling of a consonant to make it become the onset of the following syllable. The why is a local markedness reduction, i.e. the improvement of the syllable contact. Notice that in the case of liquids and nasals gemination only occurs when

Striving for optimality

127

the preceding consonant is a very strong one, namely a voiceless stop (cf. Go. akrs, OS akkar vs. Go. ligrs, OHG legar 'bed'). Moreover, in the Scale of Consonantal Strength given above a distinction is made between laterals and vibrants, where the latter are weaker than the former. Apart from cross-linguistic considerations, this distinction is motivated in Germanic by the fact that WGG does not occur when a vibrant precedes the glide: (42)

Go. far.jan 'to walk' OE, OS ferian

OHG ferien,fergen, feren9 Go. nas.jan 'to save' OE, OS nerian OHG nerien, nerigen, nergen

Thus, given the weakness of vibrants with respect to glides, the syllable contact was not bad in this case, and gemination did not take place. Recently, Ham (1998) has suggested treating WGG in a OT framework. He first observes that this last point concerning the nonoccurrence of gemination after Μ is stipulatory. Thus, he prefers to derive it from a "robust cross-linguistic tendency to avoid geminate r" (1998: 233). This point however is marginal, since Ham admits that cross-linguistic sequencing restrictions show that /r/ and III should not be treated as being equal in sonority (cf. Eng. snarl vs. *snalr, Swed. karl 'fellow' vs. *kalr, It. tarlo 'woodworm' vs. *talro, etc.). A second observation is more relevant: Murray and Vennemann's explanation is completely unable to account for gemination occurring when glides are in a syllable contact position, as in the following cases: (43)

Gmc. +frawjö- > WGmc. +frawwjö- > OHG frouwe 'to enjoy' + awjö- > +awwjö- > OHG auwia 'meadow'

Since it is claimed that glides possess the same consonantal strength, gemination is in this case unexpected, since the syllable

128

Livio Gaeta

contact should not be bad. Moreover, even assigning a stronger consonantal value to the bilabial glide does not help, because we are forced to assume that /w/ is stronger than /r/, and at the same time essentially weaker than the other ones, since it gave rise to WGG when preceded by other consonants (cf. OHG nackot, ackus in (38c) above). Thus, Ham suggests dividing the change into two parts. On the one hand, we find gemination motivated by Vennemann's (1988) contact laws. This concerns the cases (38c-d) above, namely when a bilabial glide or other sonorants were in the onset of a syllable following an obstruent-ending one. On the other hand, there are the cases where the palatal glide /j/ occurred in the onset position of the second syllable (cf. (38a-b) above). Ham claims that this change can be accounted for by a constraint forbidding the palatal glide in onsetinitial position: (44)

G [j*:

onset-initial j is dispreferred

Evidence for this constraint is provided by several Germanic languages: (45)

a.

ON ungr vs. OE geong vs. Got .juggs, OHG jung, OS jung 'young' b. ON άr vs. OE ger vs. Got.jer, OHG jär, OS jär 'year' c. OE twe.ga, twe.gra (< Gmc. *twajje)'two GEN' d. OHG zwe.io, zwei.io 'two GEN' (cf. Got. twad.dje, ON tveg.gr)

The first two examples concern the general deletion of /j/ in word-initial position in North Germanic. In this connection, Ham (1998: 247) mentions that "in OE initial *j appears to have fortified to a voiced fricative, represented orthographically as " whereas Old High German, Old Saxon, and Gothic preserved the glide. This however is a misunderstanding of OE spelling, since Old English presented a general process of lenition of voiced velar obstruents that came to merge with the palatal glide (cf. Lass 1994: 53ff.), as can be seen in examples like gieldan 'yield', geard 'yard' (cf. OHG geltan

Striving for optimality

129

'to be valid', garto 'garden'), compared to geoc 'yoke' (cf. Lat. jugum, Go. juk). The second group of examples (cf. (45c-d)) represents the phenomenon of the so-called Germanic Verschärfung ('hardening'), whereby Gmc. + /j/ was geminated when preceded by a short vowel and followed by a vowel (either long or short). Subsequently, the geminated glide was fortified in Gothic and Old Norse, whereby a sequence /ddj/ and /ggw/ resulted in Gothic, and /ggj/ and /ggw/ in Old Norse (cf. Suzuki 1991). According to Ham, WGG must be seen as the result of the reranking of the constraint avoiding onset-inital palatal glides ( σ []*) over a constraint forbidding geminates (NO-GEMINATES). Thus, we have the following two stages:10 (46)

a. West-Germanic prior to gemination before /j/ ONSET, Ρ Α ΐ Τ Η - Ι Ο ( μ ) , N O - G E M I N A T E S

σ-μμ *σίϋ b. West-Germanic after gemination before /j/ ONSET, ΡΑΠΉ-ΙΟ(μ), NO-GEMINATE r σ - μ μ , * G [j I NO-GEMINATES

The demotion of the constraint forbidding geminates joined with the promotion of the constraint requiring a syllable onset not being filled by a palatal glide account for the WGG. Once more, we see that for OT the how of language change coincides with the why. Unmarked structures emerge against other possible configurations. Now, let us try to speculate on this approach from the viewpoint of a functionalist. In Ham's approach, one question remains unexplained, namely why a constraint such as * σ [] is needed. He neatly shows that both glides are rather dispreferred as syllable onsets throughout all

130

Livio

Gaeta

Germanic languages. For example, looking at the Germanic languages, we find the following data: f

(47)

a. b. c. d. e.

ON Openn 'Odin', OE uton, wuton, OHG wöXan ON όρα 'to scream', Go. -wopjan Go. triggws 'true', ON tryggr, OE treowe, OHG gitriuwi Go. bliggwan 'strike', OHG bliuwan, OE bleowan Go. skuggwa 'mirror', ON skuggi, OHG scuwo, OE scuwa

As is shown by these examples, /w/ glide comparatively displays the same range of phenomena as 1)1, although they are differently distributed across the Germanic languages. On the one hand, we find that /w/, too, underwent deletion in North Germanic (cf. (47a-b)); on the other, it also underwent Germanic Verschärfung, giving rise to gemination and fortition (cf. (47c-e)). Moreover, West-Germanic languages show a number of phenomena that involve /w/, which is fortified (/w/ > /g/ or /k/) in onset position (cf. Ramat 1969: 52): (48)

a. b. c. d.

Gmc. *juwunpi, Go.junds 'young girl' vs. OS iugud, OFr jogethe, OE geogod, OHG jugund Gmc. *newun, Go., OHG niun 'nine' vs. OS, OFr nigun, OE nigon OS naco 'ship', OE naca, OHG nahho (cf. Lat. nävis) Go. speiwan 'to spit' vs. OS spekaldra, OHG speichaltra

Besides, we find several cases where /w/ underwent a process of deletion. It did so first of all when occurring in syllable onset before or after another consonant:11 (49)

OHG rehhan 'to chase', antluzzi 'face', ON reka, lita vs. Go. wrikan, wlits, OS wrecan, wliti OHG aha 'river', llhan 'to lend' Go. ahwa, leihwan OS sulian 'to dirty' (< +sulwiari), selitha 'habitation' (Go. salipwos), frata(h)a 'ornament' (< +fratwa, OE frcetwe)

Striving for optimal ity

131

A second favored context for deletion is when the glide occurs as onset of a second member of a compound (often with vowel backrounding): OHG heimort 'to home', OHG wurzala 'root' (cf. OE wyrtwalu), etc. Moreover, these general processes of fortitions and lenitions also accompany the later developments of West-Germanic languages: (50)

a. early MHG (ca. 1100): /w/ > N! (MHG beib 'wife' < weib) b. OE/wr/> ModE/r/ (OE wrecca > ModE wretch)

As is shown by this evidence, Ham's assumption of a constraint forbidding /j/ as syllable-onset is corrrect, but limited. The other glide /w/ also displays a number of phenomena that lead to a similar conclusion. However, Germanic usually has suffixes beginning with /j/. Suffixes beginning with /w/ are rare. Therefore, /j/ is more likely to give rise to dispreferred consonant clusters, even though clusters containing /w/ of a following suffix are not excluded. Moreover, there are several strategies to avoid both glides as syllable onsets. As mentioned above at the end of section 3, a common issue for functionalist approaches to language is the notion of multicausality (cf. Dressier 1985b: 265ff.), both in the sense that one function can be served by several operations (multiple strategies) and that one operation may serve several functions simultaneously (multifunctionality). In the case under discussion here, the first strategy is fortitive, i.e. a glide is promoted to the corresponding obstruent. This holds both for syllable-initial /j/ (cf. (45b) and (48) above to which OHG gunste 'youngest', gungiron 'younger' < jüngste, jungiron may be added) and /w/ (cf. (50) above). A further fortitive strategy is provided by gemination, which makes the glides perceptually and articulatorily stronger. This is shown by the cases of Germanic Verschärfung shown in (45c-d) and (47c-e) above. The second strategy is lenitive, i.e. deletion (cf. (49) above), or (and this is more relevant for our purposes) gemination of the preceding consonant that "covers" the syllable onset. Since this gemi-

132

Livio

Gaeta

nation is a context-sensitive process, it can be viewed as a lenitive process from the point of view of the glide because the latter turns out to occupy a secondary (perceptually and articulatorily weaker) position within the syllable. In fact, this process is accompanied in several Germanic languages by deletion of the glide and incidentally by vowel reduction: (51)

a. OHG sezzan, OS settian, OFr setta, OE settan 'to set' b. OHG hello,, OS hellia, OFr helle, OE hell 'hell' c. OHG frauwen 'to enjoy' (< +frawwjan) d. OHG auwia, ouwa 'meadow' (< +awjö~)

As one would expect, fortitions apply rather in strong contexts (e.g., word-initially), and lenitions apply in weak contexts. Of course, gemination only applies word-internally. Moreover, all this evidence showing the unstable status of both glides as syllable onsets suggests that the dispreference for syllableinitial glides may be related to the same (functional) difficulty. In fact, syllables are preferred to have a good onset that sharply drops towards the syllable nucleus (cf. the Head Law in Vennemann 1988: 13). Morever, especially when glides occur as word-internal syllable onsets, they give rise to bad syllable contacts. Hence we observe a strong tendency towards avoiding them. This does not exclude the possibility that the general process of avoiding sonorants in syllableinitial position might also concern other sonorants (i.e. nasals and liquids), starting with the lenitive positions, i.e. word-internally after consonants.12 More generally, there are numerous cues that point to the instability of glides across the Germanic languages, especially when these occur in onset position. For Gothic, Vennemann (1985) has shown that glides underwent a general process of strengthening + [i] > [j] and+[u]>[u]: (52)

jiukan wlits ga-aggwjan

[jiu.kan] [ulits] [-aggu.jan]

'to fight' 'face' 'to press hard'

Striving for optimality

waurstwja prafstjan bandwa waurstwa

[uorsty.ja] [f>rafst.jan] [band.ua] [uorts.ua]

133

'worker' 'to comfort' 'sign' 'worker'

Moreover, North Germanic also displays cases where language change has affected both glides: besides the Germanic Verschärfung seen in (45c-d) and in (47c-e) above, processes of fortition can be mentioned, as in ON var 'was', vinna 'work', where strengthening of word-initial /w/ > /v/ is observed. Thus, what we more generally observe is the uncertain status of glides in Germanic, reflected in dispreferred syllable structures that are tendentially eliminated by means of multiple strategies. Under this wide perspective, the possibility of having gemination of /w/ in front of /j/ (cf. (43) above) is not suprising: it is immediately related to the particular status of glides within (or rather at the margin of) the syllable structure. The inverse case (gemination of /j/ before /w/) however is not attested. In this respect, a weaker consonantal strength of /j'/ with respect to /w/ cannot be excluded: It should be emphasized that not only may [i] be weaker than [u], as shown by the occurrence of initial + [ u ] but not + [i] before liquids in ProtoGermanic, but the same may be true of their fricative counterparts [in Gothic, LG], shown by their distributional properties in contemporary Standard German; in particular, German [j] is less offensive in hiatus than [v], as shown by ?[mi]i.jön, ä.ri.ja] vs *[vir.tu.vös, pä.du.va]. The palatal fricative [j] remains much closer phonetically to palatal [i] and [i] than the labial fricatives [v] and [υ] remain to primarily velar [u] and [u] (Vennemann 1985: 216-217).

Similar arguments can be also made valid for West-Germanic languages. Thus, Ham's approach has proved to be helpful, since it forced us to look for a new substantive base of data, that otherwise would have gone unnoticed. As it stands, however, his explanation is not sufficient, since it must be on the one hand extended to the whole set of glides, and on the other viewed in the context of two general 11

134

Livio Gaeta

questions: the unstable status of glides in Germanic, and their effect on possible syllable structures.

7. Conclusion To sum up, I have reviewed several examples of grammatically initiated changes highlighting the methodology and the epistemological background that underly different theoretical approaches. In particular, functionalist approaches have been compared to the recent model of OT, with the aim of showing possible weakness and common properties. As for the explanatory power of the theory, we have seen that OT shares with generative linguistics the assumption that explanation equals derivation. Despite the simple format of the model, language change is made directly dependent on a purely formal reranking of the constraints, leaving the expression of markedness to the intrinsic content of the constraints. However, outputoriented constraints can shed light on other hitherto neglected possible factors of phonological change. Moreover, it is not clear if a constraint-based approach such as OT admits of multicausality, multiple outcomes of language change, markedness conflicts among constraints, and conflicts among components of grammar, which are typical accomplishments of functionalist approaches. In principle, it should at least be possible to express all these notions through the formal apparatus of the theory, although the alleged emergence of the unmarked should be carefully inspected on the basis of substantive evidence. In this respect, I have not discussed another important aspect, namely the necessity, repeatedly underscored in functionalist approaches like Natural Phonology, of distinguishing between morphophonemic and phonological alternations, usually blurred in generative approaches. Nothing in OT, however, prevents us from making this difference clear. As a way of conclusion, let me end with two quotations which are very revealing of scholars' different, and contrasting, attitudes towards OT. Aronoff (1998: 411-^112) stresses the counterintuitive formal apparatus of the theory, which appears to him quite far from

Striving for optimality

135

what we usually consider as phonology, and refers to the so-called "Kripkean problem": [W]e call many notions syntactic or phonological because they were developed in the study of syntax or phonology, and not necessarily because they have inherent ties to that phenomenological domain, in the physicist's sense of phenomenology, and not the philospher's sense ... I think that similar things can happen for example with optimality theory, where there's nothing about optimality as a formalism that has anything to do with phonology, and yet, because it was developed for the study of phonology, we tend to think that it's phonology.

On the other hand, Nathan (1999: 324-325) thinks that OT can be considered an interesting formal apparatus for functionalist theories like Natural Phonology that try to investigate the conflict between speech production and perception: I believe that well over a century of scholarship on the mental organization compels functional phonologists to retain the fundamental insight known as the phoneme, and the concomitant idea that we need two levels coupled with a set of principled relations between them. At the minimun, this is what the preceding century of both functional and formal linguistics has to offer us as modern functionalists. But in addition, I suggest that the currently developing formal theory, OT, can offer us a better metaphor for relating the two levels - not so much a transduction device as a mechanism for reconciling demands of speech and perception.

In conclusion, I think that these constrasting feelings are a good starting point for developing common investigation programs for phonologists of different persuasions. In this sense, language change is an optimal challenge for everybody. Notes 1. Forms with intrusive Irl like draw[r]ing may be considered a later, and sociophonetically still strongly dispreferred, extension of the phenomenon (cf. Wells 1982: 222ff.).

136

Livio Gaeta

2. According to Cho (1998), the process of constraint promotion or demotion is mediated by stages of undoing the ranking and creating the new ranking, as for instance in the following successive stages: stage: FAITHFULNESS » N O - C O D A » FINAL C II stage: FAITHFULNESS , N O - C O D A » FINAL C III stage: N O - C O D A » FAITHFULNESS » FINAL C I V stage: N O - C O D A » FAITHFULNESS , FINAL C V stage: N O - C O D A » FINAL C » FAITHFULNESS I

3.

4.

5.

6.

7.

The intermediate stages II and IV also allow for a certain degree of variation, since there is no dominance order between respectively FAITHFULNESS and N O - C O D A in stage II, and FAITHFULNESS and FINAL C in stage IV. Since there is no agreement within OT upon this gradual view of constraints promotion (e.g., cf. Ham 1998: 242ff. for criticism), I will leave it aside. Notice, moreover, that the alleged constraints are in a conflictual relation, since CODA minimizes the number of consonants disallowed in a syllable coda, while FINAL C explicitly requires that a final coda must if possible be filled by a consonant. However, the theory does not make any prediction as to how the conflict should be solved, except by stating an alleged emergence of the unmarked. In this respect, notice that Wells (1982: 224) observes that "intrusive Irl ... is widely regarded as incorrect or slovenly ..., so that the speech-conscious may make some effort to avoid it. ... One widespread tactic is the use of a glottal stop instead of /r/". Moreover, Wells (1982: 226) reports that r-intrusion is also found after a long final vowel: Degas ['deigair] and Sickert, the social milieu [mi:'lj3:r] of Alexander Pope, etc. Other problems regarding serialism are caused by the schwa-insertion rule seen in (4a) above. As pointed out by Rubach (1996), this rule, adding a schwa in words like prepare [es] when Irl turns out to be in coda position (cf. prepa.rative [se], without schwa-insertion), requires an intermediate step in the derivation, in which Irl in the coda is visible before getting deleted. In other words, prior to the deletion, the Irl must be syllabified into the coda in order to trigger schwa-insertion. As Rubach (1996: 228) observes, "The difficulty with this analysis is that Optimality Theory prohibits serialism that would not be correlated with morphological constituency, as would be the case here". Following the so-called Correspondence Theory (cf. McCarthy and Prince 1994), the general constraint FAITHFULNESS has been in this tableau splitted into the two constraints M A X - Ι Ο and D E P - Ι Ο , that require respectively faithfulness of the input with respect to the output (i.e., no deletions), and vice versa (i.e., no insertions). See below for a discussion.

Striving for optimality

137

8. To distinguish them from ungrammatical forms, marked by *, reconstructed forms are marked by 9. Notice that in Alemannian and Franconian r-gemination is attested: ferren, werren (cf. Braune/Eggers 1987: 113). 10. The other constraints required are rather general, or are the translation in OT terms of what we discussed above. They are defined as follows: every syllable has an onset σ - μ μ : stressed syllables are bimoraic ΡΑΙΤΗ-Ιθ(μ): the addition or subtraction of moras from the input is dispreferred ONSET:

Their status is immaterial to the present purposes. Moreover, notice that a specific constraint forbidding the gemination of /r/ ( N O - G E M I N A T E r) is assumed, which dominates the constraint forbidding the palatal glide in onset-initial position (see above). 11. I omit from this discussion cases of w-deletion when a back vowel follows, as in OHG huosto 'cough' < *hwösto, OHG suozi 'sweet' < *swuozi, etc. (cf. Braune/Eggers 1987: 103). 12. Notice that in the case of bad syllable contacts between glides and other sonorants the conflict was usually resolved (besides WGG, cf. (38b) above) either by vowel deletion (cf. (49) above) or by epenthesis: OHG Ν. melo 'flour', G. melwes/melawes, OS N. melo, G. melas. 13. Moreover, as mentioned above, because of the low number of w-initial suffixes, a cluster /j.w/ had comparatively fewer opportunities of occurring.

References Anttila, Raimo 1988 Historical and Comparative Linguistics. (2nd edition.) Amsterdam/Philadelphia: John Benjamins. Aronoff, Mark 1998 Isomorphism and monotonicity: or the disease model of morphology. In: Stephen Lapointe, Diane K. Brentari and P.M. Farrell (eds.), Morphology and its Relation to Phonology and Syntax, 411— 418. Stanford: Center for the Study of Language and Information. Blevins, Juliette 1997 Rules in Optimality Theory: two case studies. In: Iggy Roca (ed.), Derivations and Constraints in Phonology, 227-260. Oxford: Clarendon Press.

138

Livio Gaeta

Booij, Geert 1996 Lexical Phonology and the derivational residue. In: Jacques Durand - Bernard Laks (eds.), Currents Trends in Phonology. Models and Methods, 65-92. Salford: Salford University Press. 1997 Non-derivational phonology meets lexical phonology. In: Iggy Roca (ed.), Derivations and Constraints in Phonology, 261-288. Oxford: Clarendon Press. Braune, Wilhelm 1987 Althochdeutsche Grammatik. (14th edition, completely revised by Hans Eggers.) Tübingen: Niemeyer. Cho, Yong-Mee Yu 1998 Language change as reranking of constraints. In: Richard M. Hogg Linda van den Bergen (eds.), Historical Linguistics 1995. Vol. 2: Germanic Linguistics, 45-62. Amsterdam/Philadelphia: John Benjamins. Donegan, Patricia J. 1993 On the phonetic basis of phonological change. In: Charles Jones (ed.), Historical Linguistics: Problems and Perspectives, 98-130. London: Longman. Donegan, Patricia J. and David Stampe 1979 The study of Natural Phonology. In: Daniel Dinnsen (ed.), Current Approaches to Phonological Theory, 126-173. Bloomington, Ind.: Indiana University Press. Dressier, Wolfgang U. 1980 A semiotic model of diachronic process phonology. Wiener Linguistische Gazette 22-23: 31-94. 1985a Explaining Natural Phonology. Phonology Yearbook 1: 29-51. 1985b Morphonology. Ann Arbor: Karoma Press. Gaeta, Livio in press Cambiamento linguistico e modelli linguistici orientati verso l'uscita: falsificabilitä teorica e verificabilitä empirica. In: R. Sornicola (ed.), Teoria e Dati empirici. Atti del XXXIII Congresso Internazionale di Studi della Societä di Linguistica Italiana (Napoli, 2830/10/1999). Rome: Bulzoni. Ham, William H. 1998 A new approach to an old problem: Gemination and constraint reranking in West Germanic. Journal of Comparative Germanic Linguistics 1: 225-262. Hurch, Bernhard 1998 Optimalität und Natürlichkeit. ZAS Papers in Linguistics 13: 115139.

Striving for optimality

139

Jespersen, Otto 1909 A Modern English Grammar on Historical Principles. Part I: Sounds and Spellings. Heidelberg: Carl Winter. Johansson, Stig 1973 Linking and intrusive /r/ in English: a case for a more concrete phonology. Studia Linguistica 27: 53-68. Keller, Rudi 1994 Sprachwandel. Von der unsichtbaren Hand in der Sprache. (2nd edition.) Tübingen/Basel: Francke. King, Robert 1969 Historical Linguistics and Generative Grammar. Englewood: Prentice-Hall. Kiparsky, Paul 1982 Explanation in Phonology. Dordrecht: Foris. Lass, Roger 1994 Old English. A Historical Linguistic Companion. Cambridge: Cambridge University Press. Loporcaro, Michele 1997 On vowel epenthesis in Alguer Catalan. In: Pier Marco Bertinetto, Livio Gaeta, Georgi Jetchev and David Michaels (eds.), Certamen Phonologicum III, 205-227. Turin: Rosenberg & Sellier. McCarthy, John J. 1991 Synchronic rule inversion. In: Laurel A. Sutton, Christopher Johnson and Ruth Shields (eds.), Proceedings of the 17th Annual Meeting of the Berkeley Linguistic Society, 192-207. Berkeley: Berkeley Linguistic Society. 1993 A case of surface constraint violation. Canadian Journal of Linguistics38:169-195. McCarthy, John J. and Alan S. Prince 1994 The emergence of the unmarked. Optimality in prosodic morphology. In: Merce Gonzalez (ed.), NELS 24. Proceedings of the North Eastern Linguistic Society, 333-379. Amherst, Mass.: University of Massachusetts. McMahon, April, Paul Foulkes and Laura Tollfree 1994 Gestural representation and Lexical Phonology. Phonology 11: 277316. Murray, Robert W. and Theo Vennemann 1983 Sound change and syllable structure in Germanic phonology. Language 59: 514-528. Nathan, Geoffrey S. 1999 What functionalists can learn from formalists in phonology. In: Michael Darnell, Edith Moravcsik, Michael Noonan and Kathleen

140

Livio Gaeta

Wheatley (eds.), Formalism and Functionalism in Linguistics. Vol. 1: General Papers, 305-327. Amsterdam/Philadelphia: John Benjamins. Ortmann, Albert 1998 Consonant epenthesis: its distribution and phonological specification. In: Wolfgang Kehrein and Richard Wiese (eds.), Phonology and Morphology of the Germanic Languages, 51-76. Tübingen: Niemeyer. Paul, Hermann 1920 Prinzipien der Sprachgeschichte. (5th edition.) Tübingen: Niemeyer. Pullum, Geoffrey Κ. 1976 The Duke of York gambit. Journal of Linguistics 12: 83-102. Raffelsiefen, Renate 1998 Semantic stability in derivationally related words. In: Richard M. Hogg and Linda van den Bergen (eds.), Historical Linguistics 1995. Vol. 2: Germanic Linguistics, 247-267. Amsterdam/Philadelphia: John Benjamins. Ramat, Paolo 1969 Grammatica dell'antico sassone. Milan: Mursia. Rubach, Jerzy 1996 Shortening and ambisyllabicity in English. Phonology 13: 197-237. Stampe, David 1973 On chapter nine. In: Michael Kenstowicz and Charles Kisseberth (eds.), Issues in Phonological Theory, 44-52. The Hague: Mouton. Suzuki, Seiichi 1991 The Germanic Verschärfung·, a syllabic perspective. Journal of Indoeuropean Studies 19,2: 163-190. Vennemann, Theo 1972 Rule inversion. Lingua 29: 209-242. 1985 Phonologically conditioned morphological change. Exceptions to Sievers' law in Gothic. In: Edmund Gussmann (ed.), Phonomorphology. Studies in the Interaction of Phonology and Morphology, 193-219. Lublin: The University Press of the Catholic University of Lublin. 1988 Preference Laws for Syllable Structure and the Explanation of Sound Change. Berlin/New York/Amsterdam: Mouton de Gruyter. 1993 Language change as language improvement. In: Charles Jones (ed.), Historical Linguistics: Problems and Perspectives, 319-344. London: Longman. Wells, John C. 1982 Accents of English. Cambridge: Cambridge University Press.

Striving for optimality

141

Wetzels, Leo 1981 Analogie et lexique. Le probleme de l'opacite en phonologie generative. Ph.D. dissertation, University of Nijmegen. Wurzel, Wolfgang U. 1994 Grammatisch initiierter Wandel. Bochum: Brockmeyer. 1998 On markedness. Theoretical Linguistics 24: 53-71.

Preferences as defaults in computational phonology Dafydd Gibbon

1. Preferences and defaults - typical or untypical? A preference is intuitively an element of a set which is valued more highly than other elements; the preferred element can be interpreted as an object or a property, a state, an event or a course of action. Similarly, a default is a typical state or typical choice which is valued more highly than other choices, other things being equal (other meanings of "default" are of no concern here). Neither a preference nor a default is a fixed value: a preference may not be favoured under all conditions, and a default may be overridden by other considerations in specific situations. The concept of once-for-all fixed rules is gradually being replaced in linguistics by a notion of preference, starting perhaps with Trubetzkoy's notion of privative opposition, though various notions of markedness to Vennemann' s Preference Laws and Optimality Theory, as well as the generalisation and irregularity hierarchies of computational lexicon theory. Quantitative expressions of markedness range from Labov's variable rules to probabilistic automata as used in corpus tagging and in automatic speech recognition. The mechanisms used to express preference valuations range from marked/unmarked values in feature structures through default inheritance hierarchies with orthogonal and prioritised inheritance, default automata, to fully fledged default logics and non-monotonic reasoning. Perhaps the simplest gloss for default is typical, a word which deserves more extensive empirical and phenomenological analysis than it can be given here. In everyday life, it is a vehicle for stereotypes and many forms of prejudice, the Occam's razor wielded in the daily struggle to avoid as far as possible the chore of thinking rationally. In this pragmatic context, the allies of typical are adverbs of subjective

144

Dafydd Gibbon

frequency, such as usually, mostly, probably, which are used not as quantifiable statements of frequency and probability, but as adverbs of degree which qualify stereotypic convictions. The main goal of the present study is to not to argue for the use of default mechanisms for explicating preferences in phonology and prosody, since this is quite widely accepted. Rather, I will try to collect a wide range of different aspects of phonology and prosody, and show, sometimes in novel ways, the relevance of defaults to them, with the longer term goal of developing a unified theory of defaults and preferences in phonology and prosody. I will look at the simple example of English plurals to start with. Then I will move to default-override relations in phonetics, prosody, and the lexicon, followed by a detailed discussion of phonotactic (sequential) defaults, referring to a number of types of default automaton. This will be followed by a discussion of linguistic approaches to preferences, including Vennemann's Preference Law theory, and Optimality Theory. Finally I will take a brief look at the Integrated Lexicon approach to modelling default-override relations in the lexicon. It will be a difficult tight-rope act along the line which joins descriptive and computational linguistics, and I ask the reader's indulgence if the presentation wobbles a little too much for comfort to the one or the other side.

1.1. Preferences, defaults, ignorance and bliss In a sense, preferences were invented to help us to avoid complicated decisions, and defaults to help us to live in ignorance. Sometimes ignorance is bliss; defaults define prejudices in everyday life which help to avoid and generate misunderstanding and strife - though as often as not, prejudices in everyday life unfortunately often turn out to be indefeasible. Since ignorance is perhaps the most natural and the most widespread of the mental states, defaults may arguably be taken as an essential component of natural reasoning in general, and of natural linguistic theorising in particular. In fact, defaults contribute towards elegant explanations for a wide range of linguistic facts,

Preferences as defaults in computational phonology

145

and are "typically" concerned with the following kinds of reasoning in theoretical phonology: 1. Generalisations about "normal properties" of phonological objects (markedness of features specifications, i.e. roughly: weighting of attribute values). 2. Orderings over sets of generalisations with respect to their specificity (rule ordering by specificity of context, and some uses of the "elsewhere condition". 3. Rule application principles in the absence of information about other contexts, overridable by interaction with other principles (one interpretation of Optimality Theory). 4. Heuristic abbreviations over a finite vocabulary (some uses of the "elsewhere condition".

1.2. Defaults, non-monotonic reasoning, and IF-THEN-ELSE chies

hierar-

A simple example of non-monotonic ordering over sets of generalisations is the English plural affixation rule. This rule is conveniently exemplified by predicates referring to farmyard animals (cats, dogs, horses, oxen, sheep - /kaets dogz hosiz ükson Jip/), and is traditionally formulated as a set of ordered morphophonological rules, with shared rule parts indicated by braces; one rule style (without going into details, and using features to abbreviate natural classes of phonemes) is:

{N-PLURAL}

-»

iz / ζ / s /

[sibilant] [voiced] [voiceless]

Why is this rule set non-monotonic? A rule set is monotonic if adding a rule leads, as one might expect, to an increase in the number of outputs for a given set of inputs. On the other hand, a rule set is non-monotonic (as with non-monotonic functions in algebra) if add-

146

Dafydd

Gibbon

ing more rules produces fewer outputs for a given set of inputs. If the suffix /-z/ could be added freely to all English nouns, this would be a monotonic rule system: add a rule, and the number of possible output forms increases. But in English, this is not really the case, for instance, if the noun ends in a voiceless consonant: add this rule, and the number of possible combinations decreases. And this is not all, because even when the noun ends in a voiced consonant, if this consonant happens to be a sibilant the rule does not apply either: so add another rule, and the number of possible combinations decreases again. Avoiding discipline-specific notation, a non-monotonic rule set is easily formulated in a familiar IF-THEN-ELSE form (cf. also Kiparsky's elsewhere condition in phonological theory): IF an English word is a lexical exception THEN its plural must be looked up in the lexicon (e.g., ox-oxen', sheep) ELSE it takes a suffix A N D

sheep-

IF it ends in a sibilant THEN the suffix is /iz/ (e.g., horse-horses) ELSE IF it ends in a voiceless segment THEN the suffix is /s/ (e.g., cat-cats) ELSE the suffix is /z/ (e.g., dog-dogs)

An alternative but equivalent formulation is in terms of defaults and overrides with "UNLESS", which means "EXCEPT IF" or "IF NOT", and uses the inverse ordering: An English word takes the plural ending /z/ UNLESS it ends in a voiceless segment, in which case it ends in /s/ UNLESS it ends in a sibilant, in which case it ends in /iz/ UNLESS it is a lexical exception, in which case it must be looked up in the lexicon.

Preferences as defaults in computational phonology

147

In the present example, the most general default case only covers one context, namely final voiced segments. In fact, in this example the voiceless/voicing conditions could be reversed, indicating some arbitrariness in the decision and a need for independent motivation. The independent motivation could lie in the fact that the voiceless condition applies to far fewer segments, namely only the voiceless obstruents (except /tf s J/, i.e. /p t k f Θ/), than the voicing condition, while the voicing condition applies not only to the voiced obstruents (except /d3 ζ 3/) but also to the nasals /m η η/, the glides /j w/ and all the vowels. A hierarchy of constraints like this indicates that the notion of default is itself relative: there is not just one elsewhere condition, but a series of more and more general elsewhere conditions. The ranked constraints of Optimality Theory are related in similar fashion. However, we are not concerned here with discussing linguistic motivation for a particular style of default ordering among constraints, or for a particular analysis of the treatment of plural in English, but with illustrating the default ordering principle itself. Defaults in the most general sense of the term can simply be a waste bin to cover up for ignorance of the facts, as already noted. Defaults and overrides in this open-ended semantic or pragmatic sense are widely considered too complex to handle in a formal system. But in describing linguistic forms, as in the present case, the situation is different: the defaults cover a closed world which is completely known, and are not used as ignorance flags but to express interesting generalisations about irregularities or markedness inside this closed world which could not otherwise be expressed neatly. Defaults and overrides in a closed world can be spelled out exhaustively in completely monotonic fashion by conjoining positive and negative constraints, the length of the conjuncts being a function of the most specific condition in the default hierarchy. The exhaustive formulation means that IFF {if and only i f ) can be used instead of IF, and that the rules can be considered in any order: 1.

The plural of an English word ends in /z/ IFF it does NOT end in a sibilant AND it does NOT end in a voiceless segment AND it is NOT idiosyncratic.

148

2.

3. 4.

Dafydd Gibbon

The plural of an English word ends in /s/ IFF it ends in a voiceless segment AND it does NOT end in a sibilant AND it is NOT idiosyncratic. The plural of an English word ends in /iz/ IFF it ends in a sibilant AND it is NOT idiosyncratic. The plural of an English word must be looked up in the lexicon IFF it is idiosyncratic.

The terms monotonic and nonmonotonic are used to refer to the fully explicit and the default-based types of constraint. Think of the four cases listed here as axioms from which, given an English word stem, the full plural form can be inferred. Any of these four cases describes a subdomain of English plural forms independently of the others, and collecting any of them simply monotonically increases the overall domain described. The UNLESS and IF-THEN-ELSE axiom sets are different. The first axiom purports to describe all English plurals, but adding more axioms non-monotonically decreases the overall domain described by this first rule.

2. The domain of speech sounds 2.1. Phonology: irregularity defaults In a well-known paper, Morris Halle and Sylvain Schlumberger (1989) discussed "why phonology is different". Without going into the details of their discussion, two obvious answers lie in blocking (defaults and overrides) in time-based features of the phonological domain: 1.

History: language, and in particular the sound pattern of a language, develops on the one hand by grammaticalisation and lexicalisation, and complementarily by accretion and productivity, with conflicts between internal (system inherent) and external (social) constraints leading to inconsistencies, i.e. irregularities.

Preferences as defaults in computational phonology

2.

149

Example: prestige pronunciations from other dialects, or loan words, which may introduce extrinsic sound patterns. Performance: sound patterns are realised fairly directly by physical systems with partially independent components which are only loosely synchronised with each other and are arranged in physical sequences in which the effects of one component may overshadow the effects of another. Example: in Bielefeld may be pronounced in English or German as /imbibfelt/, with the alveolar /n/ overshadowed (covered, blocked) by the bilabial /m/ (at later stages of grammaticalisation, not being articulated at all).

But in recent research into the syntax of speech, driven by spoken language engineering requirements, and by research into the formalisation of prosodic phonologies by means of event logics, shows that phonology is not all that different, for instance from syntax, or indeed semantics. Part of semantics is also concerned with events and temporal relations like the phonetic and phonological events and temporal relations into which phonological units are mapped; part of syntax is different from the way in which it is traditionally conceived in logic and linguistics, in being essentially dependent on processing and therefore temporal constraints, as suggested in work by John Hawkins (1994) on performance theories of order and constituency. So it may be that phonology is no different after all, at least not on these grounds.

2.2. Phonetics: mapping defaults Taking current wisdom about phonetic interpretation at face value, I will explicate phonetics as a formal model for phonological representations, INTphon = f τ rel

153

7 • llp^' ρ < r
Pa.o.lo > {Pa.vo.lo, Pa.go.lo} And consonant deletions if > 1: OHG: loss of /h/: hwiz > wiz 'white' English: cf. knee, gnome, gnat

4.5. Optimality Theory (OT) Optimality Theory has been one of the most productive methodologies for phonological description for a number of years. The approach would merit detailed explication on its own in terms of nonmonotonic reasoning, but the present context is not the place for this. Optimality Theory (OT, Prince and Smolensky 1993) has two main descriptive goals: 1. Assignment to some linguistic input of a structural description (representation) by selecting from a (possibly infinite) set of candidate structural descriptions the one which best satisfies a ranked set of universal constraints. 2. The establishment of different rankings of the universal constraints in order to account for the difference of representations in different languages.

Preferences as defaults in computational phonology

191

The issue is frequently formulated procedurally as a search problem: 1. To select competing representations from a large search space of possible representations. 2. To select competing theories (constraint rankings) from a large search space of possible rankings. Optimality Theory, like Vennemann's Preference Law approach, seeks the optimal, best representation; unlike Vennemann's approach it does not deal with the diachronic dimension (but see Löhken 1997). This means that there is no homogeneous definition of wellformedness; well-formedness is a function of the interaction of constraints. In a sense, this is not new. The traditional style of defining phonological well-formedness in generative phonology, going back 40 years, was by separate redundancy rules operating over lexical representations. The application of all redundancy rules yielded wellformed structures. However, no principles were defined for the assignment of lexical representations; these were conceived of as idiosyncrasies. The overall effect was oddly atomistic, in particular over domains which are quite well-defined, as the discussion of syllable structure in the context of automata and of Vennemann's Preference Theory shows. The atomistic effect is enhanced by use of the conventional style of argumentation: rather than dealing with whole systems, and attempting to capture their interactions, results are presented for selected isolated problems in fragments of widely different languages. This is not to say that the work is not based on careful intensive and extensive description; the opposite is true, in general. But it is noteworthy that on the empirical side the bible of generative phonology (Chomsky and Halle 1998), for example, takes quite the opposite tack both in describing English phonology in intimate detail and in presenting a coherent theoretical framework. It is also noteworthy that on the theoretical side, Vennemann starts with a general theory of syllable structure before following up with detailed analyses of interesting fragments: structural categories such as HEAD, BODY, NUCLEUS, CODA are coherently defined as a frame of reference for constraints.

192

Dafydd Gibbon

Optimality Theory has little to say about the nature of the search problem, or about how to get beyond the implication of a simplistic "generate and test" strategy (the GEN function) of constraint satisfaction, though work has been done on this (Tesar 1995), and a number of partial implementations are available on the web. One of the favourite topics of OT descriptions is syllable boundary assignment (syllabification), because syllables provide domains for many phonological rules and, being a parsing problem, syllabification is a classic variety of search problem; see also Jurafsky and Martin (2000). An example of tableau oriented OT argumentation is given in Table 1. The lowest number of violations is selected. In this case: 1. The columns contain the constraints (partially) ordered left-right in priority ranking order from strictest to least strict constraint. 2. The rows contain potentially infinitely many output candidates (freely generated). 3. The tableau is filled from left to right with evaluations by matching candidates with constraints: nothing is entered if the constraint is not violated, and an asterisk is entered (and the rest of the row is shaded) if the candidate violates the constraint. 4. The number of asterisks in a tableau cell is its value. 5. A simple algorithm for filling the tableau is given below. 6. Evaluation: Throw out the rows containing cells with exclamation marks and shadings (the shadings are actually redundant), and the remaining candidate or candidates win. Table 1. An Optimality Theory tableau (Output2 wins)

Input: χΐ.,.χη Constraint 1 Constraint_2 Constraint_3 Constraint_4 Output 1 *! ** Output2 Output3 *!

...

Preferences as defaults in computational phonology

193

In fact the shading operation is not quite correct since only the loser rows are shaded; the winner is also shaded in the tableaux in the literature. The mnemonically named lower level procedures also require definition. Partial OT tableau filling algorithm DEFINE constraints!!,...,numberOfConstraints]; DEFINE candidates[l,...,numberOfCandidates]; // Too bad if the set is infinite. i=0; WHILE (i < numberOfConstraints) { j = 0; WHILE (j < numberOfCandidates) { IF (unshaded(cell[i,j] AND VIOLATE(cand[j],constraint[i])) THEN addAsManyAsterisksAsViolations(cell[i,j]); ++ j l j=0; WHILE(j < numberOfCandidates) { k=0; WHILE (k < numberOfCandidates) { IF (moreViolations(cell[i,j],cell[i,k])) { THEN addExclamationPointIfNoneThereAlready(cell[i,j]); 1=1-4-1; WHILE(1 'reclining', [xi]try 'cunning', [xi]mn 'hymn', [xi]drant 'hydrant', [xi]dra 'hydra' [xc]tpliwy 'boastful', [xe]mia 'chemistry', [xe]bel 'plain', [xc]rbata 'tea', [xc]roiczny 'heroic'

These examples show that the velar spirant differs from the velar plosives domain-internally. There are other examples, however, where frontness assimilation is observed by all velar consonants; consider the alternations involving two native suffixes, namely the feminine noun forming suffix -ini and the derived imperfective suffix -iw. The addition of these suffixes to a base ending in a velar results in Frontness Sharing being observed. (12) a. proro[k] 'prophet'

bo[g]a 'god GEN.SG'

monar[x]a 'monarch' b. klas[k]ac 'applaud' ska[k]ac 'jump' przyle[g]ac 'adhere' urq[g]ac 'deride' zako[x]ac 'fall in love' pods!u[x]ac 'overhear' rozdmu[x]ac 'blow out'

-

proro[ci]ni FEM bo[ji]ni 'goddess' monar[qi]ni FEM oklas[ci\wac (pod)ska[ci]wac przyle\^i\wac urq[ji]wac zako[qi\wac podsIu[gi]wac rozdmu[q\]wac

Similarly morpheme-internally both [51] and [9ε] are perfectly possible, e.g.,

Hidden identity, or the double life of segments

241

(13) [51 ]chot 'giggle', [gi]storia 'history', we[qi]kul 'vehicle', [91]nina 'quinine' [9c]na 'hyena', [9e]rarchia 'hierarchy', [9e]ronim 'personal name' The importance of such examples cannot be overemphasised. The complete failure of frontness assimilation before the front vowels of the endings in (9) or domain-internally in (1 lb) cannot be explained as resulting, for example, from the absence of the palatal congener of the velar spirant in the Polish system. As (12) and (13) show the sound [9] is regularly found in native and foreign words both before front vowel endings and domain-internally. If we assume that frontness assimilation holds for velar consonants, then the conclusion must be that some of the segments we have called the velar spirant [x] are simply not velar. Note that there is no reason why frontness assimilation should fail in certain words but not in others. Likewise there is no reason why [x] - unlike the two velar plosives - should allow a following [i] in some cases (9, 1 la), whereas in others it behaves like the velar plosives. Thus the segment [x] covers two types of phonological objects, of which only one is a velar and accordingly behaves like the other velars. In other words, [x] leads a double phonological life even if phonetically it emerges as one segment. There can be little doubt that when [x] patterns with the velar plosives, it is also a velar consonant. These are the cases illustrated in (12) and (13) when velars disallow a following retracted vowel and when they display frontness assimilation. In words where these conditions are not met we are dealing with a spirant which is not velar. An obvious question which suggests itself is: if they are not velar, then what are they? Two tentative answers can be suggested: it is possible to have phonological segments without full specification, i.e. without some properties associated with the pronunciation of the segment. If we follow this line of thinking we can claim that in Polish there is a voiceless velar spirant and a voiceless spirant, both of which are rendered phonetically as [x]. Alternatively, we could speculate about the non-velar spirant being for example the glottal [h], an interpretation which has some phonetic support: an increasing

242

Edmund

Gussmann

tendency has been observed for the glottal spirant to replace the velar one. Jassem (1954: 98) notes that the initial spirants in words such as chata [xata] 'cottage' or hymn [ximn] 'hymn' tend to be pronounced with the consonants starting the English words harm, home, i.e. as [hata] and [himn]. In the latter case we would posit (h_) and (h) as representations for the velar and glottal spirants respectively. No matter which of the two analyses we may want to select, the main point remains unaffected: the phonological properties of segments are determined not only by an inspection of their phonetic characteristics but also by considering the way they behave in the system of a language. If there is a conflict between the phonetic and phonological clues, it is the latter that get the upper hand. 4. Modern Welsh vowels6 The dialect of North Wales contains three high vowels [i, i, u], three mid ones [e, a, o] and one low vowel [a]. With the exception of [a] they can all be long, but Welsh quantity is not our concern here; it has been argued elsewhere (see Buczek-Zawila 1998) that the length of nuclei can be determined by considering stress and the consonantal neighbours of the vowel in question. In (14) we illustrate the basic vowels of North Welsh. (14) cig [ki:g] 'meat' sych [si:%] 'dry' crwn [krun] 'round MASC' is elder [i'selder] 'lowness' sydyn ['sadin] 'sudden' cron [krön] 'round FEM' caniad [kanjad] 'song' A characteristic property of Welsh is the fact that the schwa vowel [θ] cannot appear as the last vowel of a word, apart from a few monosyllabic function words. Thus * [ssQ] or * [ turns] are not possi-

Hidden identity, or the double life of segments

243

ble words in Welsh; schwa is regularly found in the non-final position, both stressed and unstressed, e.g., (15) cysgu ['kaski] 'sleep VB' cymeriad [ks'merjad] 'reputation'

ysgafn [sskavn] 'light ADJ' Cymraeg [kam'raig] 'Welsh'

Additionally, schwa alternates with the high vowels [i, u] in a number of words by a regularity that Welsh grammarians have termed Vowel Mutation. Its operation can be seen in the following sets of alternations: (16) a.

bryn [brin] 'hill' tyn [tin] 'tight' co/[kri:v] 'strong' sych [si:χ] 'dry' b. cwm [kum] 'valley' bwrdd [burö] 'table' trwm [trum] 'heavy' crwn [krun] 'round

-

bryniau ['branje] PL

-

tynnu fteni] 'pull VB' cryfder ['krevder] 'strength' sychder ['saxder] 'drought'

-

cymoedd ['ksmoö] PL

-

byrddau [tardc] PL

-

trymaidd

-

crynion [krenjon] PL

[tromeö] 'sultry'

MASC'

The left-hand column words all contain a single pronounced vowel, either [i] or [u]. If an inflectional or derivational suffix is attached to them, the root nuclei find themselves in pre-final position where these vowels are not acceptable. Instead the vowel [s] appears in the stressed nucleus. Similarly, if a suffix is attached to longer words whose last vowel is [i], schwa appears in its place before the suffix, e.g., (17) cybydd ['ksbiö] 'miser' mynydd ['msniö] 'mountain' sydyn [ssdin] 'sudden' -

cybyddion [ka'bsöjon] PL mynyddoedd [ms'noÖoiÖ] PL sydynrwydd [ss'dsnruiö] 'suddenness'

244

Edmund

Gussmann

We can see then that schwa, which is barred from word-final nuclei, replaces the two high vowels in pre-flnal position. The process of vowel replacement lies at the heart of Welsh Vowel Mutation which, in synchronic terms, constitutes the Welsh equivalent of vowel the reduction found in other languages: some vowels are barred from specific positions, and other vowels take their place there. What these vowels are and how the positions are to be defined are problems for the phonology of individual languages. Our reasoning would lead us to expect that in Welsh the vowels [i, u] are not found outside the final position, but this conclusion is not true in absolute terms. The back rounded vowel [u] appears pre-finally in a handful of borrowings like swper [super] 'supper' and, more interestingly, when the final nucleus also contains [u]; if a suffix is attached to such a form, schwa replaces both instances of [u], e.g., (18) cwmwl ['kumul] 'cloud' cwmwd [kumud] 'neighbourhood' bwrlwm [burlum] 'gurgling'

-

cymylau [ko'malai] PL

-

cymydog [ks'msdog] 'neighbour'

-

byrlymu [bar'bmi] 'bubble over'

An inspection of forms like these reveals that pre-final [u] is possible only when directly followed or supported by an identical vowel in the final nucleus. Once the final source of support is removed, [u] cannot appear in pre-final position, hence *[ku'mudog] is not a possible form. Examples like (18), although interesting in themselves, do not invalidate the reality of Vowel Mutation in Welsh: they show that phonological subregularities may interact with or impinge upon dominant generalisations. Keeping these subregularities in mind we can rightly conclude that in Welsh [u] is only possible in the last (pronounced) nucleus of a word. When we turn to the other vowel involved in Vowel Mutation, namely [i] we have to reach a different conclusion. As the examples

Hidden identity, or the double life of segments

245

in (19) show, this vowel may genuinely resist Vowel Mutation and remain unaffected in pre-final position. (19) pur [pi:r] 'pure' hurt [hin] 'self llun [li:n] 'picture' papur ['papir] 'paper'

-

puro ['pi:ro] 'purify' hunan ['hinan] PL lluniau [linje] PL papurau [pa'pire] PL

These examples of forms resisting Vowel Mutation are very different from those in (18): there is nothing that could be said to unite them or set them apart from forms conforming to Vowel Mutation. The tentative conclusion seems to be then that [i] may but does not have to conform to Vowel Mutation. Alternatively we can say that [i] displays two different types of phonological behaviour, as exemplified in (16a, 17) on the one hand and (19) on the other. If we believe in the synchronic reality of Vowel Mutation in Welsh, then the fact that only some instances of [i] are well-behaved with respect to it must mean that the vowel [i] masks two phonologically different objects. In the terminology adopted in this paper the Welsh vowel [i] must have a double identity. We seem to have identified enough phonological evidence to justify the setting up of two different phonological units. As an additional bit of evidence let us note the effect the addition of a suffix containing [i] has on the preceding vowel [a]: (20) a.

mab [ma:b] 'son' aeth [ai0] 'he went' plant [plant] 'children' b. glas [gla:s] 'green can [ka:n] 'song'

- mebyd ['me:bid] 'youth' - euthum ['aiOim] Ί went' - plentyn [plentin] 'child' - glasu [glasi] 'become green' - canu ['ka:ni] 'sing'

This seems to indicate that of the two vowels pronounced uniformly as [i] one is front and hence fronts the preceding vowel in (20a); the other would accordingly have to be back. What this means for the phonological structure of Welsh high vowels would need to

246

Edmund Gussmann

be worked out in a more comprehensive account of the phonology of the language.

5.

Conclusion

The extent to which phonetics determines phonology has been a moot question in the discipline, with answers varying as much as the approaches themselves. The question has received no generally accepted answer and is likely to loom large in phonological discussion in the future. Attempts made in some quarters to reduce phonology to phonetically motivated (or phonetically grounded) regularities are unlikely to prove fruitful. Phonetics does seem to play a role in phonological descriptions - nobody is likely to suggest that phonological [oe] should be realised as phonetic [g] for example but the extent of the influence is limited. That phonology controls mechanisms which are not reducible to those of other domains such as phonetics has often been stressed in the past (e.g., Hellberg 1978, 1980; Anderson 1981; Gussmann 1984). The basic reasoning is explicitly presented in Hale and Reiss (2000: 162): "Phonology is not and should not be grounded in phonetics since the facts that phonetic grounding is meant to explain can be derived without reference to phonology". In this paper we have attempted to show that the phonological identity of segments must be determined by phonological criteria which in certain cases go against the evidence of phonetics. Phonetics merely provides a garb for the more basic phonological or cognitive elements, a garb which is often deceptive and ambiguous.

Notes 1. It should be added that within Optimality Theory there have also been attempts to eliminate underlying representations altogether and thus embrace the monostratal view of phonology (see Kager 1999: 413^420 for some discussion).

Hidden identity, or the double life of segments

247

2. See Brockhaus (1995) for an extensive discussion of German terminal devoicing within the framework adopted in this study and Harris (1997) for a more general theory of neutralisation. It is worth pointing out that Optimality Theory also entertains the possibility of input features being abstractly present in the output (Kager 1999: 378). 3. Since word-finally we find the expression (Α·Ι) it is not clear what decides when this expression corresponds to phonetic [e] and when it corresponds to phonetic [ε]. 4. For analyses of the labio-dentals in somewhat different terms see Andersen (1969), Hayes (1984). 5 For more discussion of the elemental structure of these expressions see Cyran and Nilsson (1998). 6. The discussion of Welsh data is based on Thomas (1984), Buczek (1998) and Buczek-Zawila (1998).

References Andersen, Henning 1969 The phonological status of the Russian labial fricatives. Journal of Linguistics 5: 121-127. Anderson, Stephen, R. 1981 Why phonology isn't 'natural'. Linguistic Inquiry 12: 493-539. Brockhaus, Wiebke 1995 Final Devoicing in the Phonology of German. (Linguistische Arbeiten 336.) Tübingen: Max Niemeyer Verlag. Buczek, Anita 1998 The vowel that cannot be long: The story of the Welsh central vowel [a]. In: Eugeniusz Cyran (ed.), Resonance Elements in Phonology A Study in Munster Irish, 55-64. (PÄSE Studies and Monographs 3.) Lublin: Wydawnictwo Folium. Buczek-Zawila, Anita 1998 Principles and parameters of modern Welsh phonology. Ph.D. dissertation, Catholic University of Lublin. Charette, Monik 1991 Conditions on Phonological Government. (Cambridge Studies in Linguistics 58.) Cambridge: Cambridge University Press.

248

Edmund Gussmann

Charette, Monik and Asli Göksel 1996 Switching and vowel harmony in Turkic languages. In: Henryk Kardela and Bogdan Szymanek (eds.), Α Festschrift for Edmund Gussmann from his Friends and Colleagues, 29-55. Lublin: The University Press of the Catholic University of Lublin. Cyran, Eugeniusz 1997 Resonance Elements in Phonology - A Study in Munster Irish. (PÄSE Studies and Monographs 3.) Lublin: Wydawnictwo Folium. Cyran, Eugeniusz (ed.) 1998 Structure and Interpretation. Studies in Phonology. (PÄSE Studies and Monographs 4.) Lublin: Wydawnictwo Folium. Cyran, Eugeniusz and Morgan Nilsson 1998 The Slavic [w>v] shift: A case for phonological strength. In: Eugeniusz Cyran (ed.), 89-100. Gussmann, Edmund 1984 Naturalness, morphonology, and the Icelandic velar palatalisation. Nordic Journal of Linguistics 7: 145-163. Hale, Mark and Charles Reiss 2000 'Substance abuse' and 'dysfunctionalism': Current trends in phonology. Linguistic Inquiry 31: 157-169. Harris, John 1990 Segmental complexity and phonological government. Phonology 7: 255-300. 1994 English Sound Structure. Oxford: Blackwell. 1997 Licensing inheritance: An integrated theory of neutralisation. Phonology 14: 315-370. Harris, John and Geoff Lindsey 1995 The elements of phonological representation. In: Jacques Durand and Francis Katamba (eds.), Frontiers of Phonology: Atoms, Structures, Derivations, 34-79. London: Longman. Hayes, Bruce 1984 The phonetics and phonology of Russian voicing assimilation. In: Mark Aronoff and Richard T. Oehrle (eds.), Language Sound Structure. Studies in Phonology Presented to Morris Halle by his Teacher and Students, 318-328. Cambridge, Mass.: The MIT Press. Hellberg, Stefan 1978 Unnatural phonology. Journal of Linguistics 14: 157-177. 1980 Apparent naturalness in Faroese phonology. Nordic Journal of Linguistics 3: 1-24.

Hidden identity, or the double life of segments

249

Jassem, Wiktor 1954 Fonetyka j^zyka angielskiego [Phonetics of English], Warszawa: Panstwowe Wydawnictwo Naukowe. Kager, Rene 1999 Optimality Theory. (Cambridge Textbooks in Linguistics.) Cambridge: Cambridge University Press. Kaye, Jonathan, Jean Lowenstamm and Jean-Roger Vergnaud 1985 The internal structure of phonological elements: A theory of charm and government. Phonology Yearbook 2: 305-328. Thomas, Alan, R. 1984 A lowering rule for vowels and its ramifications, in a dialect of North Welsh. In: Martin J. Ball and Glyn E. Jones (eds.), Welsh Phonology. Selected Readings, 105-124. Cardiff: University of Wales Press.

Working with licensing constraints Jonathan Kaye

1. Introduction In this paper I will discuss the central role of licensing constraints (henceforth "LC's") in phonological systems and how they may be viewed as one of the principal engines of phonological events. LC's were originally designed to explain restrictions on the combinatorial properties of elements. Given a theory of phonological expressions (to be given below), the underlying assumption is that any syntactically well-formed combination of elements should be present in a phonological system unless explicitly excluded. Since, as far as we know, no language expresses the full range of theoretically possible combinations of elements, LC's were proposed as language-specific constraints on such possibilities. A subset of a small set of possible LC's is sufficient to define the lexical set of, say, nuclear expressions of a given linguistic system. Recent work has shown that the usefulness of LC's extend far beyond their original raison d'etre. In particular it is a pleasure to recognise the two seminal articles of Monik Charette and Ash Göksel (Charette and Göksel 1996, 1998) which have provided the leadership in this field and the inspiration for this present work. I will briefly review part of their work in a later section. In the following section, I give a succinct summary of the element theory of phonological representations.1 This will make clear to the reader the types of representations to which LC's are applicable.

252

Jonathan Kaye

2. The element theory: a summary 2.1. Elements E, the set of elements is defined below. (1)

Ε = {A, I, U, H, L, ?}

and an identity element, usually represented as "_" in phonological expressions. Each element is a monovalent, (potentially) interpretable phonological expression. Its actual interpretation depends on (i.) what phonological constituent (see below) dominates it and (ii.) whether it occupies a head or operator position within a phonological expression (see below).

2.2. Phonological expressions All speech sounds are phonological expressions. A phonological expression is defined as an ordered pair Phonological Expression = (0,H) such that: i. ii. iii.

Ο · Ε (Ο possibly empty) Η . Ε (possibly the identity element) Η. Ο

By convention, the first member of the ordered pair is called the operator(s), the second, the head of the phonological expression. Expressions headed by the identity operator are called headless. All other expressions are called headed. The head of an expression is said to license its (set of) operators. By way of illustration consider the (Southern British) English set of stressed branching nuclei. They are represented in the following table:

Working with licensing constraints

({} A )

for

((},i)

fee

((},U)

too

({A},I)

pay

({A},U)

foe

({U},A)

saw

253

LC's involve the combinatorial possibilities of the elements A, I and U. We shall see in a later section that the elements Η and L have their own licensing constraints, at least when they are manifestations of tone. The element ? is problematic and indeed may be spurious.2 Limiting ourselves to A, I and U and following the definition of a phonological expression given above, we have the following possible (nuclear) expressions. (3) (OA)

(0,U)

({A},I)

({A,I},U)

({A}, J

({A},U)

({A,U},I)

({I},J

({I}, A)

({I,U}A)

({U},J

({I},U)

({A,I}, J

({U},A)

((A,U},_)

((U},I)

((I,U},_)

({A,I,U},J

We can see that there are 20 possible expressions involving the elements A, I and U. Columns 1, 2 and 3 contain headed expressions consisting of 1, 2 and 3 elements, respectively. Column 4 and 5 contain the headless expressions. The final expression in column 5 is the "empty" expression, containing no head and the empty set of operators. Table (2) contains far less than 20 expressions (6, to be exact)

254

Jonathan Kaye

so LC's for English must be postulated in order to generate exactly those 6 expressions. (4)

English LC's (stressed branching nuclei) i. All expressions are headed. ii. I and U may not combine. iii. I must be a head.

LC (4i) states that all expressions must be headed. This eliminates all of the expressions in columns 4 and 5 of (3). (5)

(U,A)

({A},I)

({A,I},U)

-OA^J

KM)

({A},U)

({A,U},I)

({I}, J

((},U)

({I}, A)

({I,U},A)

({U},J

({I},U)

((A,T},_)

((U},A)

({A,TT},J

({U},I)

({TJT},J

({A,T,U},_) ({}.J

LC (4ii) states that all expressions containing I and U are to be excluded. Note that some expressions have already been ruled out by (4i).

Working with licensing constraints

255

(6) (IIA)

({A},I)

i w d —

(0,1)

((A},U)

- e t i ö — -ttte)

(0,u)

({I}A)

-tm^— ({A,I},J

((U},A)

({A,U},_)

Finally, LC (4iii) eliminates all expressions where I is not a head as seen below. (7)

((},A)

({A},I)

OA

((},i)

((A},U)

ΟΑ,μμ^

am

-

m

^

—

-KW

Tg^Ar; • W d — tgy^cC (|U},A)

((A,U),__)

The order in which these LC's are applied is, of course, irrelevant. Furthermore, as stated above, some phonological expression may be excluded by more than one LC. In sum, a small set of LC's can generate the set of permissible (nuclear) expressions in a given language. But is delimiting the set of permissible expressions the only work that LC's do? This is the question to which we turn in the next section.

256

3.

Jonathan Kaye

LC's and Turkish vowel harmony: a discussion of Charette and Göksel (1996,1998)

Given the form in which LC's are expressed, it is not unnatural to suppose that a licensing relation exists between head and operator of a phonological expression. Specifically, it has been suggested that heads of expressions may license their operators. For example, in many languages we note that the element A may occur as a head as in ({},A) but all expressions of the form (X,A) are illicit (where X represents some non-null subset of E, the set of elements). This constraint may be expressed as follows: (8)

A may not license operators.

Assuming this "intra-expression" licensing does exist, then the term paradigmatic licensing would be an appropriate expression to designate such licensing. Following this line, we can also suppose that inter-nuclear interactions such as vowel harmony involve a similar sort of licensing. We can term this type of licensing, syntagmatic licensing. With these terms in mind and with the understanding of how LC's work in the definition of expression-inventories as illustrated above, we can now formulate the contribution of Charette and Göksel as follows: (9)

The Charette-Göksel (CG) Hypothesis Paradigmatic LC's are recapitulated syntagmatically.

Put simply, (9) means that the same LC's that define, say, the nuclear inventory of a language also define inter-nuclear relations as in vowel harmony. In their articles (see above) Charette and Göksel present detailed analyses of several Turkic languages. For illustrative purposes I will limit my discussion to their analysis of Turkish. Charette and Göksel posit the following LC's for Turkish (1998: 71).

Working with licensing constraints

(10)

257

i. Operators must be licensed ii. A is not a licenser iii. U must be a head

The LC's in (10) generate the following phonological expressions for Turkish. Note that they allow for the "empty" expression shown in the final row of this table. (11)

Turkish Nuclear Expressions Expression

Turkish Letter

(OA) (OJ)

ι

({},U)

u

({i},u)

ü

({A},I)

e

(|A},U)

ο

({A,I},U)

ö

(U,J

ι

Now we can apply the Turkish nuclear LC's to the case of Turkish vowel harmony. The central claim of (9) is that the facts of Turkish vowel harmony can be derived from these LC's. Let us see how Charette and Göksel do this. I reproduce their table (Charette and Göksel 1998: 69) which displays the complete harmony facts of Turkish.

258

Jonathan

Kaye

(12) Stem

Gloss

Plural

2nd Pers. Poss.

kil

'clay'

kil-ler

kil-in

kül

'ash'

kül-ler

kül-ün

kul

'subject'

kul-lar

kul-un

kel

'bald patch'

kel-ler

kel-in

köy

'village'

köy-ler

köy-ün

kol

'arm'

kol-lar

kol-un

kas

'muscle'

kas-lar

kas-in

kil

'hair

kil-lar

kil-in

I briefly summarise the facts shown in (12) below. (13)

Turkish vowel harmony i. All nuclear expressions may appear in the Ni4 position. ii. Only ({},A) and ({},_) may appear in other positions. iii. I spreads from Ni to any position. iv. U spreads only to a headless position. v. A does not spread.

Charette and Göksel proceed to derive the harmony facts in (13) from the Turkish LC's in (10). A is not a licenser (lOii) paradigmatically so A cannot license anything syntagmatically. Harmonic spreading is viewed by Charette and Göksel as a form of licensing. Thus, we can derive (13v). U must be a head (lOiii) so U can only spread into a headless expression (13iv). U cannot spread to a position containing ({},A) since that expression is headed and, as (lOiii) states, U must be a head. There are no LC's involving I and so I is expected to spread freely. Thus, we can derive (13iii).

Working with licensing constraints

259

Charette and GöksePs analysis gives us some rather fine-grained predictions. When / spreads to a recessive harmonic position containing A as a head (as in kel-ler < kel-lar) then / must appear as an operator. / can either be a head or an operator in an expression. It will happily occupy the operator position in the suffix vowel of keller. Notice that A must not switch from head to operator in this situation, otherwise there would be no reason why the element U could not spread in like circumstances (e.g., kul-lar would become *kul-lor). But this means that the 2 "e's" of kel-ler must be different: the first contains the expression ({A},I) whilst the second contains ({I},A). This means that the 2 "e's" of kel-ler must be pronounced differently since they do not have the same representation. As Charette and Göksel note, this is entirely correct: the 2 "e's" of keller are indeed pronounced differently. One may wonder if the appearance of the expression ({I},A) is not a violation of LC (lOii). In fact, it is not A that is licensing / in the operator position, but rather the / in the N] positions that licenses itself in the suffix vowel of kel-ler. In sum, Charette and Göksel show that starting from a set of LC's and their hypothesis (9), the properties of Turkish vowel harmony can be derived. They need not be stipulated. In the next section I will show other properties of vowel harmony that are natural consequences of LC's. To illustrate this point I turn to Finnish vowel harmony.5

4. LC's and Finnish vowel harmony Finnish vowel harmony can be succinctly expressed by viewing the following table.

260

Jonathan Kaye

(14) Harmony

No Harmony

"Ü"

"u"

"Ö"

"a"

"ä"

"o"

Neutral

"e"

Each column of (14) contains a set of nuclei which behave alike with respect to vowel harmony. Nuclei from column 1 may not cooccur with nuclei from column 2. The nuclei from column 3, the socalled "neutral vowels" may occur with nuclei from either column 1 or column 2 (but of course not both simultaneously). Is there any formal property shared by the vowels in each of these 3 columns that would allow for a natural expression of these harmony facts? The answer is provided by the licensing constraints necessary to generate the set of Finnish nuclear expressions, to wit (15)

Finnish LC's i. All expressions are headed ii. U must be a head

The LC's in (15) generate the nuclear expressions organised into the harmonic table (14) above. (16) Harmony

No Harmony

Neutral

ÜI},U) "ü"

({},U) "u"

({}J) "i"

({A,I},U) "Ö"

({},A) "a"

({A},I) "e"

({I},A) "ä"

({A},U) "o"

Working with licensing constraints

261

Comparing (16) with (14) is quite revealing. In (16) the formal properties of each subset of nuclear expressions stands out clearly. The "harmony" column is characterised as the set of all nuclear expressions containing / as an operator. The "no harmony" column is the set of all nuclear expressions not containing the element I. Finally, the "neutral" column is the set of all nuclear expressions containing / as a head. The expression of Finnish vowel harmony follows quite naturally from the LC-generated set of nuclear expressions. (17)

Finnish Vowel Harmony If a nuclear expression in a phonological domain contains / as an operator, the element / must be present (as head or operator) in every nuclear expression in the phonological domain.

From (17) the facts of Finnish vowel harmony described above can be easily derived. No expression in the "harmony" column can co-occur with an expression from the "no harmony" column. All "harmony" expressions have / as an operator, and according to (17) every nuclear expression in the same domain must contain an I. But the expressions in the "no harmony" column are precisely those that do not contain I. Thus, mixing "harmony" and "no harmony" expressions violates (17). The "neutral" expressions, i.e. those containing/ as a head, may freely co-occur with "harmony" expression. They satisfy the requirement that every expression of the domain must contain the element I. They do contain the element I. "Neutral" expressions may also co-occur with "no harmony" expressions. The requirement that every expression in a domain must contain the element / is limited to those expressions having / in the operator position. This is not the case for the "neutral" expressions. There is a pleasing rigidity found in the relation between the Finnish LC's and Finnish vowel harmony. Given the inventory of Finnish nuclear expressions, we are obliged to postulate the set of LC's in (15). These LC's generate a set of expressions shown in (16) which provide the formal basis for distinguishing the 3 behavioural classes of nuclear expressions. Simply put, if Finnish didn't have the

262

Jonathan Kaye

set of vowels that it has, it wouldn't have the vowel harmony that it has. Suppose for example Finnish had the following hypothetical vowel system: (18) Harmony

No Harmony

Neutral

({U},I)"Ü"

({},U)"u"

ÜU) "i"

({A,U},I) "Ö"

({},A)"a"

({A},I) "e"

({A},U) "o"

({U},A) "0"

We have simply replaced the expression ({I},A), with the expression ({U},A). Notice, however, that the LC's must change to reflect this new nuclear inventory. In fact they become identical to those of (4) without (4ii). A glance at (18) shows that there are no no longer formal properties that distinguish the columns. The element / is not an operator in any expression. This property cannot distinguish column 1 from column 3. Likewise, /-headed expressions are found in both column 1 and column 3. We have lost the formal characterisation of the "neutral" set of nuclear expressions. This example clearly illustrates the link between LC's and the possible harmonic systems associated with them. An English nuclear system could not coexist with a Finnish harmonic system.

5. LC's and tonal systems To this point we have considered LC's and their role in harmonic systems. The LC's involved in defining nuclear inventories were a subset of (1), viz. A, I and U. I will now consider two of the remaining elements, Η and L, and their role in LC's that define certain tonal systems. To begin this discussion, I will present a very short theory of tones.6

Working with licensing constraints

263

Discussion of tonal systems is often quite confusing. For example, Mandarin (Putonghua) is said to have "4 tones". I will show that the use of tone in this context is inaccurate. What Mandarin has are 4 tonal patterns. This notion becomes crucial when I proceed to the elaboration of a theory of tones. I present the following definitions: (19)

A tone is part of nuclear expression associated to one or more skeletal positions.

Taking the simplest case, and assuming that the elements Η and L mark "high tone" and "low tone" respectively, when they appear in the operator position of a nuclear expression, a single skeletal position may have at most 3 (level) tonal contrasts: (20)

Tonal expressions Ν X Η

Ν X

Ν X

L

In (20) the elements Η and L are simply abbreviations for any nuclear expressions ({H,X},Y) and ({L,X},Y), respectively. They represent "high tone" and "low tone", respectively. The third structure of (20), covers the case where neither Η nor L are present, i.e. any phonological expression (X,Y) where Η ·Χ & L ·Χ. This is the socalled "toneless" or "mid tone" expression. (21)

Tonal Inventories i. Simple: Η or Toneless ii. Complex: H, L or Toneless

264

Jonathan Kaye

I claim that tonal inventories are limited to two types: those that have 2 level tonal contrasts, the simple case, and those that have 3 level tonal contrasts, the complex case. I claim 2 tone systems are never ambiguous. In theory, we could have 3 possible 2 tone systems given the above definitions: Η and Toneless, L and Toneless and Η and L. My claim is that all 2 tone systems are Η and Toneless. Furthermore, I must make the outrageous claim that all tonal systems purported to have more than 3 level tonal contrasts on a single posiη tion, are either misanalysed or mistranscribed. (22)

Tonal Patterns i. The scope of tonal patterns is a phonological domain. ii. Tonal patterns are defined using LC's referring to the elements Η and L. iii. A tonal pattern must have a head which can be intrinsic or positional. iv. Η is an intrinsic head. v. Tonal patterns are left-headed; Ni is a positional head.

Tonal patterns are further subject to the following constraints. I would like to suggest that these constraints are universal. (23)

Constraints and behaviour of tonal LC's i. Extended OCP: A tonal element my only appear once in a tonal pattern. ii. A tone spreads rightward from a head position to a toneless position.

Taken together (22) and (23) define the universal properties of tonal patterns. The extended OCP (23i) means that not only are successive identical tones excluded from a tonal pattern (the traditional OCP) but even non-adjacent identical tones are excluded. Thus tonal patterns, HH_ or H_H are both ruled out by (23i). (23ii) states that tonal patterns H_ and L_ (where Η and L occupy the head position of the pattern, typically Ni) are interpreted as follows.

Working with licensing constraints

265

(24) Tone spreading

Ν

Ν

Ν

Ν

(24) shows that the tones occupying the left (head) position of the pattern spread rightward to the toneless position being interpreted as "level high" and "level low", respectively. Note that without (22) and (23) we have the following logical possibilities for tonal patterns. tn

(25)

where t represents the number of tonal contrasts and n, the number of nuclei in the pattern. Concretely, for a system containing 3 tonal contrasts and a pattern consisting of 2 nuclei, there are 3 = 9 logical possibilities. These are given below: (26)

ΗΗ H_ HL

_H _L

LH L_ LL

The set of possible tonal patterns in (26) is immediately reduced by (23i) to (27). (27)

iHi H_ HL

_Η L

LH L_ fc-L

266

Jonathan Kaye

The interpretation required by (23ii) yields the final result for tonal patterns consisting of 2 tonal markings. (28)

ΗΉ H_ = "Η Η" HL

_Η

LH L _ = "L L" Li

_L

With these preliminaries out of the way, I can now proceed to a discussion of tonal licensing constraints in Mandarin. Mandarin is said to have 4 "tones". This turns out to be mistaken. What Mandarin has, is 4 tonal patterns. The scope of these tonal patterns is phonological word. In the case of all Han languages studied to date, this pattern takes the form shown in (29) below.8 (29)

The Han Template Ο,

N,

02

N2

X

X

X

X

The Han template consists of 4 positions: 2 onset-nucleus pairs. The Mandarin tonal patterns are expressed on the nuclear projection of (29), to wit, Ni, and N2. Traditionally, Mandarin "tones" are described as follows:9 (30)

Mandarin "tones" i. High Level (tone 1 ) ii. High Rising (tone 2) iii. Low Rising (tone 3) iv. Falling (tone 4)

In sum, Mandarin tonal patterns consist of one level tone, 2 rising tones and 1 falling tone My assumption is that this is not a random

Working with licensing constraints

267

distribution of possible tonal patterns but that it follows directly from tonal LC's. I propose the following LC's for Mandarin: (31)

Mandarin Tonal LC's i. Tonal inventory = complex (H, L, _) ii. Mandarin tonal patterns have an intrinsic head.

(31 i) stipulates that both Η and L are available for tonal marking, along with toneless positions. From (3Iii) and (22iii, iv) it follows that any Mandarin tonal pattern is licit provided that it contains the element H. A tonal pattern must have a head. Mandarin tone patterns have an intrinsic head. Η is the intrinsic head. From the 7 possible tonal patterns given in (27), only 4 meet the requirements of (31). They are shown below. (32)

i. Η _

ii. _ Η

iii. L Η

iv. Η L

The patterns of (32) are all and only the tonal patterns containing the Η element. Recall that according to (23ii), Η _ must be interpreted as "Η H". This leaves us with one level tonal pattern (32i), 2 rising patterns (32ii, iii) and one falling pattern (32iv). Of course this matches actual Mandarin tonal patterns (30) perfectly. In sum, acquiring the Mandarin tonal system involves only learning the content of (31). The actual tonal patterns follow from this theory of tones.

6. Conclusion In this article I have tried to show that the use of LC's in phonology can be extended far beyond a mere catalogue of nuclear expressions. Work in this area is still in its infancy. The extension of LC's to nonnuclear systems such as onsets lies largely in the future. I believe that the examples given here serve as a strong indication that this direction of research will be extremely fruitful.

268

Jonathan Kaye

Notes 1. 2. 3. 4. 5. 6. 7.

This section is drawn largely from (Kaye in prep. a). For discussion of this point see (Jensen 1994). This allows for headed expressions plus the "empty nuclear expression". The N] position refers to the first nucleus in a phonological domain. I am grateful to Lorna Gibb for discussion of Finnish vowel harmony. More detailed discussion of this theory is to be found in (Kaye in prep. b). This includes much of my work on Vata and other Kru languages where my sin was misanalysis. 8. See Chiu (1994) and Goh (1997) for discussion. 9. Frequently these so-called tones are described in terms of integers ranging from 0 to 5.1 regard numbers as inappropriate for the analysis of tonal systems.

References Charette, Moruk and Asli Göksel 1996 Switching and vowel harmony in Turkic languages. In: Henryk Kardela and Bogdan Szymanek (eds.), A Festschrift for Edmund Gussmann, 29-56. Lublin: The University Press of the Catolic University of Lublin. 1998 Licensing constraints and vowel harmony in Turkic languages. In: Eugeniusz Cyran (ed.), Structure and Interpretation: Studies in Phonology, 65-88. (PÄSE Studies and Monographs 4.) Lublin: Wydawnictwo Folium. Chiu, Andy 1994 Syllable structure of Taiwanese in the framework of government phonology. M.A. dissertation, Department of Linguistics, SOAS, University of London. Goh, Yeng-seng 1997 The Segmental Phonology of Beijing Mandarin. Taipei: The Crane Publishing Co. Jensen, Sean 1994 Is ? an element? Towards a Non-segmental phonology. In: Margaret Cobb and Sean Jensen (eds.), Working Papers in Linguistic and Phonetics, vol. 4, 71-78. London: SOAS, University of London. Kaye, Jonathan in prep, a A Users' Guide to Government Phonology. in prep, b A Short Theory about Tone.

Rules vs. constraints in modeling phonological change: the case of Raddoppiamento Fonosintattico Michele Loporcaro

0. Introduction For those trained as linguists during the second half of the twentieth century, the notion of rule has come to be considered as a matter of course, as something intrinsic to our understanding of linguistic structure. Things have been changing recently, though. While the term "rule" continues to be used, there is now an increasing shift towards non-dynamic (i.e. static, or "declarative") models of linguistic description. In these models, whose most successful representative is nowadays Optimality Theory (henceforth OT), the grammar of any specific language is described as the product of a ranking of universal violable constraints selecting among candidate outputs. Thus, the rule component is dispensed with. In this paper, I will tackle the question from the vantage point of phonological change. Many instances of change which were previously described as changes in the rule component have been recently reanalyzed within OT as the product of constraint re-ranking. Consider for instance Löhken's (1997) treatment of vowel lengthening in Middle High German. No lengthening rule is assumed (see e.g., the rule in Vennemann 1972: 191). At stage (la) a constraint RLL-μ, preventing lengthening, outranks STRESS-μμ, imposing that all stressed syllables be bimoraic; the reverse is true in (lb), and this reversal represents the change. (1)

Löhken (1997): vowel lengthening in Middle High German (e.g., [tu.gent] > [tü.gent])

270 a.

Michele Loporcaro input: /tugent/

undominated constraints

FILL -μ

i-ir 1. [tu.gent]

input: /tugent/ 1. [tu.gent] 2. [tuu.gent]

non-selective constraints

**

*! undominated constraints

STRESS -μμ *!

FOOTΜΑΧ(μ) *

*

2. [tuu.gent] b.

STRESS -μμ

FILL-μ

non-selective constraints

FOOTΜΑΧ(μ) *

*

**

One argument often mentioned when comparing the relative merits of static vs. dynamic descriptive models, is that the former directly allow for the expression of what is universal in language, whereas rule-based descriptions, it is claimed, are confined to the language-specific dimension. From the point of view of the historical linguist, I must say that I cannot see anything inconvenient about being language-specific. It is, of course, part of the task of historical linguistics to try to detect constant factors, found to be at work crosslinguistically in different instances of change. But there is also a highly language-specific aspect about doing historical linguistics, one which is rooted in the object itself: history is by definition "specific", not general or universal. In what follows, I will present some considerations as to how we should best describe phonological change, whether with rules or with constraints. To do this, I will take into consideration an example drawn from the diachronic phonology of Romance languages, viz. the historical development of Italian Raddoppiamento Fonosintattico (henceforth RF). The paper is organized as follows. In § 1 I first sketch briefly the synchronic working of RF in modern standard Italian, and state the diachronic problem which we will have to deal with. §2 reviews the alternative diachronic explanations which have been put forward for the rise of RF. Since it turns out that reanalysis was crucially involved in this historical development, §3 briefly discusses the notion, arguing that constraint-based models bring us back to lack of distinction between reanalysis and analogy. Finally, §4 discusses a re-

Rules vs. constraints

in modeling phonological

change

271

cent OT account of RF and addresses some general problems concerning the modeling of postlexical sandhi processes within OT, compared with more traditional, rule-based models.

1. Raddoppiamento Fonosintattico: the synchronic facts and the diachronic questions to be asked RF consists of lengthening of a word-initial consonant, triggered by a preceding word with stress on the final vowel, be it a polysyllable or a stressed monosyllable (2a-b). (2)

a. faro [b:]ene b. sto [b:]ene

'I'll do well' (oxytones) 'I'm fine' (stressed monosyllables)

The examples in (2a-b) instance phonologically regular RF: all words ending in a stressed vowel regularly trigger it (including loans or nonce forms). This is informally stated in rule (3): (3)

C ->C:/V # _

[where # = $]'

In work grounded in Generative Phonology over the past three decades (cf. Saltarelli 1970, 1983; Vogel 1978, 1982: 66 ff.; Chierchia 1986; Kaye et al. 1990: 206; Sluyters 1990: 92; Bullock 1991: 115; Nikiema 1992: 11), it has been proposed that there is no separate rule of RF, but rather, that RF reduces to an epiphenomenon of a general constraint on syllabic quantity, a constraint which is seen at work in (4): (4)

a. cane 'dog' CV:CV (V: = ^200ms) canne 'reeds' CVC:V (V = ^100ms) b. farä 'will do' . . . C V # (V = ^100ms) c. farä bene (*-V C-) —»farä [b:]ene 'it'll do well'

In Italian, stressed vowels are short in closed syllables and long in open syllables ((4a)), except word-finally ((4b)), where they are also

272

Michele

Loporcaro

short. Thus it has been proposed that RF is simply a repair strategy to prevent the creation of illegal patterns such as the one which would arise in (4c) (with a stressed short vowel followed by a short initial consonant), if RF did not apply. Of course, this explanation cannot be directly equated with analyses making use of violable constraints a la OT. The basic idea underlying the approach in (4c) is of the type which has now found its fullest expression within OT: the attempt to reduce the weight of the rule component in linguistic description, and to shift the 'work' performed by a grammar from rules to representations (and constraints on output representations).2 I have already shown elsewhere why an account of RF as a wellformedness constraint has to be rejected (cf. Loporcaro 1988: 349352; 1997a: 11; I will review the topic briefly in §4.1 below). In what follows, I will simply talk about "stress conditioned RF", assuming rule (3). Before we consider the diachrony of RF, a further piece of information needs to be introduced along with the regular RF of (2), there is also a phonologically non-regular RF, triggered by the items listed in (5). This verson of RF applies after some non-stressed monosyllables (5 a) and after a handful of polysyllabic words which are not stressed on the final vowel (5b): (5)

a. e [t:]w

'and you' (a 'to', da 'from', e 'and', ο 'or', ma 'but', ne 'nor', tra, fra 'between') b. come [t:]e 'like you' {come 'like', dove 'where', quälche 'some')

RF in (5a-b), unlike in (2a-b), is not amenable to a phonologically transparent condition: there is no final stress preceding the target consonant here, yet this does undergo gemination. The morphemes belonging to the lists in (5a-b) must simply be marked in the lexicon by means of a rule feature [+RF], as shown in (6): Λ

Rules vs. constraints in modeling phonological change

(6)

C - » C:/ [_]wd # _ I [+RF]

273

[where # = $]

This synchronic irregularity, though, has a straightforward historical explanation. The Latin antecedents of the words in (5) all ended in a consonant, which were assimilated to the following wordinitial consonant in sentence phonetics, thereby yielding gemination. This is shown in (7): (7)

ET VIDES

>

»QUOMODO+ET ME

>

e [v:]edi'and come[m:]e

(you) see'

'like me'

Then the final consonants were lost from phonemic representation; and the process became lexicalized, as expressed by the rule feature in (6). Now, the question the historical linguist has to ask is what exactly is the historical relationship between irregular RF (in (5)) and regular RF (in (2)), triggered by final stress. In fact, as shown in (8) and (9), not all words with stress on the final vowel have lost a final consonant: RF is found after final stress, both in (8a) che [\:]edi, where a final consonant was in the etymon, and in (8b) tu [v:]edi, where no consonant was ever present to assimilate historically. (8)

(9)

a.

QUID VIDES

>

b.

TU VIDES

>

assimilation + ψ '

stress

·'

—

+ +·1ί.

che [vi\edi 'what do you see' tu[v(\edi6you see'

(7) (8a) (8b)

examples: ET VIDES OUID VIDES TU VIDES

The diachronic issue can in turn be analyzed into two subproblems. Once we take for granted that RF in (5) was triggered by assimilation, as shown in (7), a) what exactly happened in (8a)? Was assimilation responsible for the genesis of RF in this case as well?

274

Michele Loporcaro

And b) what happened in (8b), where no final consonant was ever there? How exacly did final stress begin to act as a trigger of RF?

2. The rise of Raddoppiamento Fonosintattico: two alternative hypotheses As I mentioned at the outset (§1), it is the contention of recent work in generative phonology that regular RF is the product of a general output constraint on syllabic quantity (as seen in (4a-b)). This explanation has been extendend from synchrony to diachrony, to account for the rise of regular RF in the history of Italian. The authors mentioned in (10) have proposed that regular RF (2) arose in Proto-Romance, as a by-product of the collapse of Latin contrastive vowel quantities. (10) a. b.

Korzen (1980), Vincent (1988), Repetti (1991) QUID VIDES TU VIDES

> *ke vedi > *tü vedi

final C loss V-shortening

> ke [v:]edi > tu [\\]edi

RF

As shown in (10), a conspiracy is assumed: on the one hand, final consonants were dropped (10a), and on the other hand final long vowels were shortened (10b). The two processes converged to create an illegal pattern in which a stressed short vowel was followed by a short consonant. Hence RF was born, as a sandhi application of the same constraint seen in (2a-b), which bars such strings within words. This view is summarized in (11): (11)

Regular RF as the product of a constraint on syllable quantities a. regular (stress conditioned) RF is Proto-Romance; b. it arose as a by-product of the collapse of Latin contrastive vowel length.

Rules vs. constraints in modeling phonological change

275

In Loporcaro (1997a) I have argued for an alternative view, which is summarized in (12): (12)

Regular RF as a sandhi rule, born through reanalysis a. regular (stress conditioned) RF is not Proto-Romance; b. it arose later, only in some Romance varieties including Tuscan (on which standard Italian is based), from the reanalysis of assimilatory gemination at word boundary, of the kind illustrated in (7). This alternative view is presented more analytically in (13)-

(14): b. a. /et vides/ /dat pane(m)/

(13) 1st stage 2nd stage 3rd stage

gray white

(14)

sandhi assimilation (spoken Latin) irregular RF (whole Romance) regular RF (Tuscan & Gallo-Romance) final stress: final consonant: = gemination - no gemination

c. /tu vides/

[VI]

m m

[VI]

[PO

[VI]

+

+ +

+

[VI]

Μ [V]

a.

stage 1 > 2: Loss of final consonants from the underlying representation. Its input being lost, gemination is reanalyzed as triggered by an idiosyncratic feature of the lexical items involved (= irregular RF, cf. (5)) b. stage 2 > 3: Gemination in (13b) is reanalyzed as triggered by a preceding final stress (= regular RF, now extended to (13c)).

There is a wealth of evidence in favour of the account in (12-14) that can be gleaned from inspection of ancient texts as well as from comparison of modern Romance languages and dialects. This evi-

276

Michele Loporcaro

dence is discussed in detail in Loporcaro (1997: 41-117). A very brief summary of the crucial facts allowing us to posit the three subsequent stages (13a-c) would include the following. In the first stage in (14) - say, in spoken (late) Latin - final consonants were assimilated to following initial consonants, yielding a geminate in both (13a) and (13b): that is, regardless of stress. Evidence for this fact, which largely goes unnoticed in Latin handbooks, is found in Latin inscriptions. A few examples of recorded assimilations are given in (15a), whereas (15b), a distich from a Pompei inscription, would be metrically corrupted unless we assume that the string ama valia peria qui ... was actually pronounced ama [b:]alia [pi\eria [k:]ui ... (cf. Fanciullo 1997): (15)

a.

sud die (= sub die) CIL V 8280, at tuos (= ad tuos) CIL VI 31066, emmimoriam (= in memoriam) CIL III 14014; b. quisquis ama valia, peria qui nosci amare bis [t]anti peria, quisquis amare vota (CIL IV 1173)

No gemination is found in (13c) TU VIDES, at this stage, as in tu [v]ides there is no final consonant, and hence no possible input to gemination. The development from stage 1 to stage 2 is due to an independent change (cf. (14a)): final consonants are lost prepausally and before vowels and, consequently, they disappear from underlying representation, a very well known change from Latin to Romance. But for our sandhi clusters in (13a-b) this loss had no effect at the surface. The sandhi geminate in e.g., [e 'v:e:di] (13a) is preserved. Evidence for this preservation comes from the fact that this geminate has been carried over all the way to modern standard Italian, as seen in (5). However, this geminate underwent reanalysis: at stage 2, it was no longer the product of a sandhi assimilation, since final consonants had disappeared from the input, but it was reanalysed as depending on a lexical idiosyncrasy of words like those in (13a-b), that is, of all the words which had lost a final consonant. Conclusive evidence for this stage 2 comes from comparative data. Sardinian, as well as all the dialects spoken in southern Italy

Rules vs. constraints in modeling phonological

change

111

never developed a regular, stress-conditioned RF. In other words, all of these varieties still are at stage 2, today. In (16) this situation is exemplified with data from Neapolitan: (16)

a. b.

['ka 'diiifa] 'what are you saying' < QUID DICIS (=(13b)) ['ki 'ri:jb] 'whom are you saying' < QUI DICIS (= (13c))

RF occurs in (16a) after ['ka], a stressed monosyllable which has lost a final consonant, whereas it does not take place in (16b) after ['ki], which never ended in a consonant, in spite of the fact that it carries stress on the final vowel. Now we finally come to stage 3 in (13), which was reached by Tuscan only. A second reanalysis took place, as argued in (14b). Wherever the sandhi geminate was directly preceded by stress (as in 10b), stress was a natural candidate for a phonologically transparent trigger of the gemination. Then, the geminate in (13b) was reanalysed as the effect of a rule such as (3). This is the point in time when regular stress-conditioned RF was born, which consequently has nothing to do with the collapse of Latin contrastive vowel quantities. This final reanalysis, of course, did not concern (13a): unstressed monosyllables went on triggering RF due to the preservation of an idiosyncratic lexical feature. But the important consequence of this second reanalysis - at the surface was the extension of RF to (13c): it is at this moment (presumably, not later than the 10th century A.D., whereas the loss of distinctive vowel quantity was completed by the 5th century; cf. Loporcaro 1997a: 55-70, 133-141) that we can locate the origin of the pronunciation ['tu 'vieidi], which was our explanandum. This is in my view the most reasonable explanation of how RF came to be applied after words like tu (< TU), which are stressed on the final vowel, but did not contain a final consonant in Latin.

278

Michele

Loporcaro

3. Rules, constraints and the distinction between reanalysis and analogy What can we conclude from this discussion relevant to our topic, "rules vs. constraints"? Obviously, my account of the rise of RF would have to be radically reformulated in models which do not admit phonological rules. Some aspects of this reformulation will be touched upon in §4. In the present section I will develop some brief remarks concerning one specific point, viz. the distinction between reanalysis and analogy and the way this distinction can be represented in rule-based vs. constraint-based models. A significant advance in our understanding of the topic was the identification of the rise of stress-conditioned RF as an instance of reanalysis. Schuchardt (1874), the first to sketch an explanation along the lines we have followed here, used a misnomer for the change implying the extension of RF from (13b) ['ke 'vieidi] to (13c) ['tu 'v:e:di]: he called it analogical extension. The ideological background behind this move of Schuchardt's is clear. He thought he had succeeded in showing that a change which is entirely lautgesetzlich in its consequences (as we have seen in (2)) was not the product of a sound law but had arisen instead through analogy: that is, through what counted as the chief disturbing factor interfering with Lautgesetzlichkeit in the Neogrammarian world. Now - after one more century of research in historical linguistics - we can tell the difference between analogy and reanalysis. The latter, unlike the former, is firmly grounded within the realm of regular sound change.4 This distinction has become part of the linguists' common sense, and is totally independent of specific formalisms. Consider for instance the following quotation: Reanalysis modifies underlying representations ... and brings about rule change. Analogy, strictly speaking, modifies surface manifestations and in itself does not effect rule change, although it does effect rule spread either within the linguistic system itself or within the community (Hopper and Traugott 1993: 32).

Rules vs. constraints in modeling phonological

change

279

The concept of rule has a central role to play, in the establishment of the distinction. We can observe in passing that Hopper and Traugott (1993) are surely not writing in the Sound-Pattem-ofEnglish generative orthodoxy of the seventies, although if one goes through recent literature in the rules-vs.-constraints debate (e.g., the papers collected in Roca 1997), one might easily get the impression that rules were used exclusively by generative phonologists. (The recent paper by Hurch 1999 is illuminating on this point.) To illustrate why the change at stage 3 (in (13)) cannot be legitimately termed "analogical",5 let us contrast it with a genuine case of analogical extension of RF, found in central and northern Calabrian dialects (cf. Loporcaro 1997a: 116 fn. 24, 1997b: 49). In these varieties, third person verb forms cause RF since they once ended in a dental stop, which assimilated ((17a), (17c)).6 In the imperfect, the [+RF] feature has been extended to first singular forms by analogy (17b), since the 3rd and 1st persons coincide phonetically. Note that this analogical extension is not observed in the present tense (17d), where the phonetic coincidence which provided ground for analogy in the imperfect is lacking: η

a.

< [kan'taißa 's:empre] '(s/he) always sang' < b. [kan'taißa 's:empre] '(I) always sang' < c. ['kanta 's:empre] '(s/he) always sing' d. ['kantu ls£mpre/*'siempre] < '(I) always sing'

CANTABAT SEMPER CANTABAM SEMPER CANTAT SEMPER CANTO SEMPER

In the first person, the rise of RF rests on no historical ground, as final -M deleted early in spoken Latin and consequently never acts as a trigger for sandhi assimilation. Obviously, no new rule has arisen here, unlike what happened in the third stage of our reconstruction in (13). Simply, one lexically specified feature characteristic for the phonological behaviour of one form has been extended to another form. This is analogy in its clearest manifestation: an isolated

280

Michele Loporcaro

change, independent from the rest of the system, which does not at all affect the rule component. Now, how can we handle these facts if we no longer have a rule component? How can we cope with the difference between (17) and (13), between analogy and reanalysis (or rule-based changes in general). One of the basic features of constraint-based, output oriented models like OT is precisely the fact that they tend to obscure this difference. Consider for instance Kenstowicz's (1996: §3.1) recent analysis of Standard Italian intervocalic /s/ voicing. The empirical fact to be explained is that voicing applies, in prefixed words, when the prefix boundary follows /s/ (as in di[z]onesto 'dishonest') but not when the boundary precedes /s/ (as in a[s\ociale 'antisocial'). This is a very simple description of the facts: /s/ voices in a context which is phonologically defined (intervocalic position), but the rule is blocked by a morphological condition (cf. e.g., Bertinetto 1999; Loporcaro 2000 with specific reference to intervocalic /s/ voicing, and Pensado 1999, on / > h in Spanish). Instead of this, Kenstowicz proposes an account based on Correspondence Theory, in the form of the tableau reproduced in ( 1 8 ) . ( V = constraint satisfaction, * = constraint violation, *! = fatal violation.) (18) BT

ES3

Base-Identity di[z]-onesto di[s]-onesto a-[s]ociale a-[z]ociale & sociale

V V

*VsV

V *! *

*!

The lower-ranked constraint in (18) substitutes for the /s/-voicing rule, while the higher-ranked correspondence constraint (BaseIdentity) is responsible for the blocking of voicing whenever the form to which voicing should apply happens to occur elsewhere in the language, as an independent word, with voiceless [s]. This is the case for asociale (cf. sociale) not for disonesto (cf. *dis). We can

Rules vs. constraints

in modeling phonological

change

281

here disregard the fact that the proposal suffers from empirical inadequacy, as for instance the prefixes bis- 'twice', tris- 'three times' do undergo voicing (e.g., bi[z]-dvolo 'grand-grandfather', tri[z]άνοΐο 'grand-grand-grandfather') in spite of the fact that the same morphemes also occur as independent words, with voiceless [s]: il 6i[s] 'the encore', un /n'[s] 'a group of three' (e.g., tris d'assi 'three aces').8 Apart from this, the crucial point for our present concern is that the difference between analogy (here expressed by Base-Identity) and regular sound change (here /s/ voicing) evaporates in this model. The two are not formally distinct. Rather, they both receive the same expression, under the form of constraints within one and the same hierarchy.9

4. Conclusion It is not my contention that the account of the rise of RF advocated here cannot be rephrased within output oriented models such as OT. In fact, accounts in that vein have been produced recently. For completeness, I will devote the next two subsections to a brief discussion of actual (§4.1) or potential (§4.2) no-rule approaches to RF.

4.1. RF in OT

Absalom and Hajek (1997: 170-175) conclude their brilliant paper on RF by sketching an OT analysis of RF in standard Italian. While in-depth discussion of their proposal would exceed by far the scope of the present paper, I will discuss only one specific point which seems of central relevance to the discussion at hand. In previous work on RF (cf. Papa 1981; Loporcaro 1988: 349352, 369ff.; Agostiniani 1989: 36ff.) it was pointed out that nonapplication of RF in the context of word-final glide deletion (e.g., Florentine /fa'rai 'bene/ —» [fa'ra 'be:ne]/*[fa'ra 'bieine] '(you) will do well') provides a crucial argument against constraint-based ac-

282

Michele Loporcaro

counts of RF. Absalom and Hajek (1997: 171) address the issue and formulate a constraint «*ΡΐΝΑίμ (glide): Underparse second mora of a word-final diphthong (optional), i.e.farai may surface as fara'». In tableau 4 (on p. 175) the constraint is seen to be crucially involved in barring both *[fa'rai 'beine] and *[fa'ra 'b:e:ne] at the same time as candidate outputs for the input /fa'rai 'bene/, since both incur a fatal mark (*!) by violating it. However, this solution poses descriptive problems. While the latter candidate is totally ungrammatical, the former is optionally possible at a low speech-rate. Thus, the authors are right in qualifying glide-deletion as optional, but precisely for that reason they cannot be right in claiming that violation of one and the same constraint (viz. *ΡΐΝΑίμ (glide)) rules out, by the same token, an ungrammatical string (*[fa'ra 'b:e:ne]) as well as an optionally occurring Lentoform ([fa'rai 'beine], to be written without an asterisk).10 Thus, the account seems in need of further refinement. (And it is indeed presented as a tentative analysis, in appendix to a paper whose main thrust consists in the entirely convincing disproof of claims on RF put forth in recent work on the topic in Generative Phonology: e.g., Bullock 1991; Repetti 1991; Vogel 1994.) Conversely, it is easy to realize that a rule-ordering analysis, by which application of RF precedes (optional) glide-deletion in a counter-feeding order, derives the observed facts without any difficulty (cf. Loporcaro 1997a: ll). 11

4.2. Further prospects for constraint-based

analyses

ofRF?

There are two facts that must be regarded as crucial for any analysis of RF in both synchrony and diachrony: the first is that it is a sandhi phenomenon, in which phonology and syntax by definition interact (in an interesting way, one might add); the second is that it consists of a regular and of an irregular part (viz. (2) and (5)), and this is what the diachronic account expounded here crucially capitalizes on, by explaining the rise of rule (6) as an instance of reanalysis. If we now concentrate on the second point, it is easy to see along which lines a full-fledged OT account of the rise of RF could be de-

Rules vs. constraints in modeling phonological

change

283

veloped. Irregular RF could be handled by a theory of lexical exceptions such as that proposed by Kraska-Szlenk (1999), where each irregular morpheme is viewed as a specific constraint. The rise of regular RF out of stage 2 ((13)), on the other hand, could be handled as an instance of lexicon optimization (Prince and Smolensky 1993 and much subsequent work: e.g., Ito, Mester, and Padgett 1995; Yip 1996; Löhken 1997 etc.). Of course, intellectual exercise is always good per se. But the point is whether there is any aspect of our concrete understanding of the history of RF - as I have illustrated it here, by elaborating on the concept of reanalysis within a simple rule-based model - that can be substantially improved by switching to a constraint-based model. Consider finally the other basic fact about RF mentioned above, viz. the circumstance that it is a sandhi rule. Some problems may arise concerning the input. There are two quite different ideas of input of RF, which circulate in the OT literature. According to one, reflected in the "richness-of-the-base principle",12 there exists one (infinite) set of inputs which is the same cross-linguistically. As a consequence, literally anything goes as an input for any language, including e.g., [kf:::rz:::] for, say, the competence of English or Italian speakers.13 It is then constraint ranking which does the job of excluding [kf:::rz:::] from the range of winning candidates that surface as grammatical forms in English or Italian. There is, however, another view of the input in the same literature. When the label "lexicon optimization" is used to designate the choice of the most harmonic input among possible alternatives (by means of the "tableaux de tableaux" procedure; cf. e.g., Ito, Mester, and Padgett 1995; Yip 1996), then we are to conclude that inputs are in the lexicon. (Otherwise, what on earth would "lexicon optimization" mean?)14 And if they are in the lexicon, then, by definition they must be stored, which might pose some problems, with respect to the above mentioned [kf:::rz:::] and surely poses problems for a sandhi process such as RF. This is a postlexical process, whose inputs, to be matched with candidate outputs, consist of phrases rather than lexical units, as exemplified in (19): (Constraints are not specified but indicated generically as x, y, z.)15

284

Michele Loporcaro

(19) comp^TO comp[lTO comp[*TO: comp^vy

X

y

ζ

kiar ]ciofi

V

kar ]ciofi kzr]ciofi k:SLT]ciofi

V V V

V

*t

V V

*!

* *!

V

'(s/he) bought artichokes' We can ask where the corresponding inputs (which must be phrases) are concretely retrieved from, in order to be matched with candidate outputs (also phrases, of course). If the answer is "from the lexicon" - since inputs are in the lexicon - then this implies that syntax must be in the lexicon. Concretely, in our case compro carciofi '(s/he) bought artichokes' must be stored in the lexicon, as well as compro mele/macchine etc. '(s/he) bought apples/cars' or any other of the infinite set of theorically possible pairings of two Italian words, such that RF applies at the boundary between them. Clearly, what we need here to state these facts in a sensible way is not an infinite list of concrete inputs. It is rather a postlexical phonological rule referring to phonological segments, like the one in (3) that we have been elaborating in order to propose our account of the rise of stress-conditioned RF in the history of standard Italian.16 Notes This paper was presented orally at the DGfS conference in Konstanz (Feburary 1999): I thank the audience for helpful discussion. I am also grateful to Tom Cravens for comments on a previous draft. Usual disclaimers apply. 1. The condition in square brackets in (3) bars the application of RF before initial geminates and heterosyllabic clusters (e.g.,sto [s]tudiando 'I'm studying'). For more details on the synchronic working of RF cf. Loporcaro (1997a: ch. 1) and the further references mentioned there. 2. OT accounts of RF begin to be available: see the valuable paper by Absalom and Hajek (1997: 170-175) (cf. §4.1 below). 3. In spite of recent criticism of rule features (cf. Inkelas et al. 1997), I think no other solution is at hand in the present case. Specifically, the account of RF in (5a-b) proposed in Loporcaro (1988: 364-365), positing an empty consonant in word-final position, rests on no independent surface evidence other than the application of RF itself. It is consequently tantamount to assuming a rule fea-

Rules vs. constraints in modeling phonological change

285

ture, since the presence of a final empty consonant must be specified for each of the lexical items in (5a-b). 4. I am aware that this is a controversial point. I would not subscribe to the notion of analogy advocated in e.g., Anttila (1977), where the concept is stretched to include many instances of regular sound change. (A consistent development of such a theoretical stance is represented by more recent proposals such as Skousen 1995). With respect to Schuchardt, specifically, there is a challenging attempt (Vennemann 1972) to argue that precisely he, rather than the Neogrammarians, was the real forerunner of the 20th century conception of rulebased phonological change. 5. It is not idle to make this point explicit, since the rise of RF in tu [d:]ici 'you say' is still sometimes labeled "analogical extension" in the literature in theoretical phonology (cf. e.g., Napoli and Nespor 1979: 877 ff., fn. 12; Absalom and Hajek 1997: 176). 6. This is RF of the irregular kind: the verb forms involved are not oxytones, and the dialects at issue lack stress-conditioned RF altogether, as exemplified in (16) with Neapolitan. 7. Note that phonetic coincidence, while it has favoured the analogical change in (17b), is in itself not a sufficient condition. In the dialect of San Giovanni in Fiore, spoken within the Calabrian area under discussion, third plural present verb forms end in [-u] (< Latin -UNT) and bring about RF: e.g., [kintu 's:empre] '(they) always sing'. These forms are homophonous with the 1st singular, which however did not acquire the feature [+RF]: ['kintu 'zempre] '(I) always sing' (cf. Loporcaro 1995: 546-547). 8. The relation between the two is transparent, both formally and semantically (both are decoded as 'twice'). Cf. Bertinetto (1999: 275 ff.) for more comments on the empirical shortcomings of Kenstowicz's analysis. 9. This seems to be a recurrent pattern which is often encountered during the establishing phase of "revolutionary" models. As Anttila (1977: 76) puts it: "Generative historical linguistics has confused the issue [viz. the distinction of 'sound change' and 'analogy', M.L.] in very much the same way as structuralists did in the beginning." 10. The ungrammaticality only concerns the relation of the quoted output to the input meaning '(you) will do well', since the same phonetic string is the normal output, in both Florentine and standard Italian, of underlying /fa'ra 'bene/ '(s/he) will do well', where RF regularly applies. 11. Nespor (1990: 251) raises a data question, meant to deny the empirical foundation of the argument summarized here. She quotes the examples ventitre [bi]arche 'twenty-three boats' vs. ventidu(e) [b]arche 'twenty-two boats' (respectively, (16a-b) in Nespor 1990), noting that RF is present in the former, not in the latter, and then goes on to observe:

286

Michele

Loporcaro

The example in (16b) contrasts with (17) [i.e. ventidue [b]arche —> ventidü [bi]arche, M.L.], a form found in certain dialects of Italian, where the final syllable of ventidue (i.e. e) is deleted. In this case RS [i.e. RF, M.L.] does apply. This is unfortunately false. No dialect anywhere in Italy applies RF in such a context; in particular, this is not the case in Florentine, in which *ventidu [bi\arche is ungrammatical. True, the string as such occurs in e.g., Romanesco as well as in all other Central-Southern Italian dialects. But the length of initial [b:] is not due to RF there, since in these varieties [b:] is categorically long in any intervocalic position, both word-internally and at word boundary: e.g ru[b:]ä 'to steal', le [b-^arche 'the boats' (never *ru\b\a, *le [b]arche vs. standard Italian rubare, le barche). This stands out clearly as soon as another consonant is substituted for [b:] in the example mentioned: ventidü [1 ]inguisti 'twenty-two linguists' (never *ventidü [1:]inguisti). In sum., the fact that RF is categorically inhibited by the application of glide-deletion suffers no exception, contrary to what is claimed by Nespor (1990). 12. The principle (cf. Prince and Smolensky 1993: 191; Smolensky 1996: 3) goes as follows: "The source of all systematic cross-linguistic variation is constraint reranking. In particular, the set of inputs to the grammars of all languages is the same." 13. Actually, even non-linguistic material should be included in this infinite set. This is a seldom discussed implication which follows straightforwardly, however, from the abolition of distinctiveness as a criterial property for input elements carried out within OT (cf. e.g., Kirchner 1997). 14. Kager (1999: 34) addresses the problem explicitly, and argues that richness of the base and lexical optimization are not in contradiction. Still, I think there is a conceptual problem here, if one does not want to give up the idea that the lexicon is a model of linguistic information stored in the speaker's mind. 15. Tableau (19) applies to RF the OT treatment proposed for another well-known external sandhi phenomenon, viz. French liaison, in Perlmutter (1998) and Steriade (1997). In those papers, competing candidates are phrases, just as in (19). 16. Of course, there is still the alternative of expressing candidates in terms of /

/

unspecified phonological strings (say, V#C:, V:#C etc.) rather than as lexical items, a procedure followed e.g., by Prince and Smolensky (1993) in their discussion of structural constraints defining syllable structure. But in that case, in my view, such a constraint-based representation ultimately boils down to a mere notational variant of a phonological rule.

Rules vs. constraints in modeling phonological

change

287

References Absalom, Matthew and John Hajek 1997 Raddoppiamento Sintattico: What happens when the theory is on too tight? In: Pier Marco Bertinetto, Livio Gaeta, Georgi Jetchev and David Michaels (eds.), Certamen phonologicum III, Papers from the Third Cortona Phonology Meeting, April 1996, 159-179. Turin: Rosenberg & Sellier. Agostiniani, Luciano 1989 Fenomenologia dell'elisione nel parlato in Toscana. Rivista italiana di dialettologia 13: 7^46. Anttila, Raimo 1977 Analogy. The Hague/Paris/New York: Mouton. Bertinetto, Pier Marco 1999 Boundary strength and linguistic ecology (mostly exemplified on intervocalic /s/-voicing in Italian). Folia Linguistica 33: 267-286. Bullock, Barbara E. 1991 Mora-bearing consonants in coda position and related quantity effect. In: Pier Marco Bertinetto, Michael Kenstowicz and Michele Loporcaro (eds.), Certamen phonologicum II, Papers from the 1990 Cortona Phonology Meeting, 105-120. Turin: Rosenberg & Sellier. Chierchia, Gennaro 1986 Length, syllabification and the phonological cycle in Italian. Journal of Italian Linguistics 8: 5-33. CIL = Corpus Inscriptionum Latinarum. Berlin: Preussische Akademie der Wissenschaften, since 1863. Fanciullo, Franco 1997 Anticipazioni romanze nel latino pompeiano. Archivio Glottologico Italiano 82: 186-198. Hopper, Paul J. and Elizabeth C. Traugott 1993 Grammaticalization. Cambridge: Cambridge University Press. Hurch, Bernhard 1999 Optimalität unf Natürlichkeit. Ms., University of Graz. Inkelas, Sharon, Ophan Orgun and Cheryl Zoll 1997 The implications of lexical exceptions for the nature of grammar. In: Iggy Roca (ed.), Derivations and Constraints in Phonology, 393 418. Oxford: Clarendon Press. Ito, Jungko, Armin Mester and Jaye Padgett 1995 Licensing and underspecification in Optimality Theory. Linguistic Inquiry 26: 571-613. Kager, Rene 1999 Optimality Theory. Cambridge: Cambridge University Press.

288

Michele

Loporcaro

Kaye, Jonathan, Jean Lowenstamm and Jean-Roger Vergnaud 1990 Constituent structure and government in phonology. Phonology 7: 193-231. Kenstowicz, Michael 1996 Base-identity and uniform exponence: Alternatives to cyclicity. In: Jacques Durand and Bernard Laks (eds). Current Trends in Phonology: Models and Methods, vol. 1, 363-393. Salford: ESRI. Kirchner, Robert 1997 Contrastiveness and faithfulness. Phonology 14: 83-111. Korzen, lorn 1980 II raddoppiamento sintattico e la geminata nella Variante toscana dell'italiano-standard. Risultati di un'indagine sperimentale. Studi Italiani di Linguistica Teorica e Applicata 9: 333-366. Kraska-Szlenk, Iwona 1999 Syllable structure constraints in exceptions. In: J. R. Rennison and K. Kühnhammer (eds.), Phonologica 1996. Syllables!? Proceedings of the 8th International Phonology Meeting, Vienna, 1 -3 November 1996, 113-131. The Hague: Thesus. Löhken, Sylvia C. 1997 Deutsche Wortprosodie. Abschwächungsund Tilgungsvorgänge. Tübingen: Stauffenburg Verlag. Loporcaro, Michele 1988 History and geography of raddoppiamento fonosintattico: remarks on the evolution of a phonological rule. In: Pier Marco Bertinetto and Michele Loporcaro (eds.), Certamen phonologicum, Papers from the 1987 Cortona Phonology Meeting, 341-387. Turin: Rosenberg & Sellier. 1995 Raddoppiamento fonosintattico dopo III persone plurali del verbo nei dialetti di Conflenti (CZ) e di San Giovanni in Fiore (CS) Rendiconti dell 'Accademia dei Lincei 392: 543-553. 1997a L'origine del raddoppiamento fonosintattico. Saggio di fonologia diacronica romanza. Basel/ Tübingen: Francke. 1997b Lengthening and Raddoppiamento fonosintattico. In: Martin Maiden and Mair Parry (eds.), The Dialects of Italy, 41-51. London: Routledge. 2000 Stress stability under cliticization and the prosodic status of Romance clitics. In: Lori Repetti (ed.), Phonological Theory and the Dialects of Italy, 137-168. Amsterdam/Philadelphia: John Benjamins. Marotta, Giovanna 1995 Apocope nel parlato di Toscana. Studi Italiani di Linguistica Teorica e Applicata 24: 297-322.

Rules vs. constraints in modeling phonological change

289

Napoli, Donna Jo and Marina Nespor 1979 The syntax of word-initial consonant gemination in Italian. Language 55:812-842. Nespor, Marina 1990 On the separation of prosodic and rhythmic phonology. In: Sharon Inkelas and Draga Zee (eds.), The Phonology-Syntax Connection, 243-258. Chicago/London: University of Chicago Press. Nikiema, Emmanuel 1992 More than Coda Conditions in Italian phonology. In: Christiane Laeufer and A. Morgan Terrell (eds.), Theoretical Analyses in Romance Linguistics, 3-18. Amsterdam: Benjamins. Papa, Eugene 1981 Is raddoppiamento resyllabification? In: Heles Contreras and Jürgen Klausenburger (eds.), Proceedings of the Tenth Anniversary Symposium on Romance Linguistics (Papers in Romance, III, suppl. 2), 249-260. Pensado, Carmen 1999 Frontera de prefijo, aspiration de "f" y procesos de nasalization en la historia del espanol. Romance Philology 52: 89 112. Perlmutter, David 1998 Interfaces: explanation of allomorphy and the architecture of grammars. In: Stephen Lapointe, Diane K. Brentari and P.M. Farrell (eds.), 307-338. Morphology and its Relation to Phonology and Syntax. Stanford: Center for the Study of Language and Information. Prince, Alan S. and Paul Smolensky 1993 Optimality Theory: Constraint interaction in generative grammar. Ms., Rutgers University, New Brunswick. Repetti, Lori 1991 A moraic analysis of raddoppiamento fonosintattico. Rivista di Linguistica 3: 307-330. Roca, Iggy (ed.) 1997 Derivations and Constraints in Phonology. Oxford: Clarendon Press. Rohlfs, Gerhard 1982 Ein archaischer phonetischer Latinismus im nördlichen ("lateinischen") Kalabrien. Zeitschrift für Romanische Philologie 98: 547-549. Saltarelli, Mario 1970 A Phonology of Italian in a Generative Grammar. L'Aja: Mouton. 1983 The mora unit in Italian phonology. Folia Linguistica 17: 7-24.

290

Michele Loporcaro

Schuchardt, Hugo 1874 De quelques modifications de la consonne initiale dans les dialectes de la Sardaigne, du Centre et du Sud de l'Italie. Romania 3: 1-30. Skousen, Royal 1995 Analogy: A non-rule alternative to neutral networks. Rivista di Unguis tica 7: 213-231. Sluyters, Willebrord 1990 Length and stress revisited: a metrical account of diphthongization, vowel lengthening, consonant gemination and word-final vowel epenthesis in modern Italian. Probus 2: 65-102. Smolensky, Paul 1996 The Initial State and 'Richness of the Base' in Optimality Theory. Johns Hopkins University Technical Report, JHU-CogSci-96-4. Steriade, Donca 1997 Similarity and lexical conservatism in surface analogy. Ms., University of California at Los Angeles. Vennemann, Theo 1972 Phonetic analogy and conceptual analogy. In: Theo Vennemann and Τ. Η. Wilbur (eds.), Schuchardt, the Neogrammarians, and the Transformational Theory of Phonological Change: Four Essays, 181-204. Frankfurt am Main: Athenäum. Vincent, Nigel 1988 Non-linear phonology in diachronic perspective: stress and wordstructure in Latin and Italian. In: Pier Marco Bertinetto and Michele Loporcaro (eds.), Certamen Phonologicum. Papers from the 1987 Cortona Phonology Meeting, 421^-32. Torino: Rosenberg & Sellier. Vogel, Irene 1978 Raddoppiamento as a resyllabification rule. Journal of Italian Linguistics!: 15-28. 1982 La sillaba come unitä fonologica. Bologna: Zanichelli. 1994 Phonological Interfaces in Italian. In: M. L. Mazzola (ed.), Issues and Theory in Romance Linguistics: Selected Papers from the 23rd Linguistic Symposium on Romance Languages (LSRL XXIII), April 1-4, 1993, 109-126. Washington, D.C.: Georgetown University Press. Yip, Moira 1996 Lexicon optimization in languages without alternations. In: Jacques Durand and Bernard Laks (eds.), Current Trends in Phonology: Models and Methods, vol. 2, 757-788. Salford: ESRI.

Are optimality theoretical "constraints" the same as natural linguistic "preferences"? Nikolaus Ritt

0. Abstract This paper discusses the meanings which the terms "constraint" and "preference" have in Optimality Theory (henceforth OT) on the one hand and Natural Linguistics (henceforth NL) on the other. It attempts to show that both terms, which at first sight might appear to play comparable roles in the two frameworks, are used to refer to very different things within each of them. The reason why these differences in usage are seldom made explicit and tend to be pasted over and overlooked is that they reflect fundamental tenets of OT on the one hand and NL on the other. These are implicitly taken for granted by representatives of each of the two camps but often fail to be come explicit in the discourse between them. Thus, rather than comparing and evaluating the two theoretical approaches and the concepts they employ, this paper intends to help avoid potential misunderstandings and to facilitate the integration of ideas from the two schools. It consists of three parts: first it shows why OT "constraints" and NL "preferences" might be regarded as functional counterparts of each other and therefore as freely exchangeable. Second, it shows that they get radically different interpretations if viewed against the wider theoretical frameworks within which OT and NL approach the study of language. And finally, it points out possible ways in which the two approaches might profit from one another or how their concepts and methods might be profitably integrated.

292

1.

Nikolaus

Ritt

Introduction

First, however, a few words on the motivation of this paper. The assumption on which it rests is that linguistic theories should help to understand and/or explain empirical properties of languages as they manifest themselves in language use, textual output, the linguistic competence of speakers, the acquisition of languages, their social implementation and distribution, their historical evolution and so on. Of course, specific linguists will be interested in specific subsets of the large variety of language related problems that there are. To give my own case as an example,1 I usually deal with phonological and morphological phenomena from various historical stages of English, and expect linguistic theories to help me understand and explain them. Now, as is well known, the international linguistic community as it presents itself today does not share anything like a generally accepted theoretical framework for the study of language on which individual linguists could safely draw when addressing specific language related problems. Instead it hosts a variety of different and often rivalling paradigms. Among them are OT and NL. For linguistic practice such variety represents a challenge, because it forces individual researchers to make rather difficult decisions. Thus, one may either choose to approach a particular problem within the orthodoxy of a particular framework, in which case one has first to decide for which of them to go and then to determine what exactly counts as orthodox within it. On the other hand, one may decide to pick some concepts and methods from one paradigm and some from others and assemble them into a kind of personal conceptual and methodological toolkit or "theory". In that case, one has to be extremely careful once again, because elements that appear to make perfect sense within the frameworks they are derived from may easily lose their usefulness, if they are removed from their usual environments and transferred into new ones. Finally, one may try and find some compromise between options one and two, and start from a (necessarily personal) version of one of the frameworks that are "out there" and cautiously alter and enrich it as the need (or the opportunity) arises.

Are OT "constraints " the same as NL "preferences "?

293

As far as I am concerned - both to make matters clear and to provide an example - I have often found functional and particularly naturalist approaches very helpful and illuminating, and, being based at Vienna, "Dresslerian" Natural Linguistics, if it may thus be called, has always been a particularly handy source of inspiration for my work. However, I could neither help nor ever wanted to avoid confrontation with other theoretical approaches, and I have never been exceedingly peculiar about where I might get suitable building blocks for the edifice that I have come to regard as my very personal linguistic world view. In other words, I have taken the third of the options outlined above. Starting from the safety of a particular theoretical perspective, I have ventured into foreign territory from time to time and brought back one or the other heretical idea. In going for that kind of mild eclecticism, I suspect that I may be a typical representative of historical linguists in general, although this is no more than an intuition, I hasten to add. Anyway, eclecticism - mild or otherwise - has its dangers, as is well known. As already indicated above, when one adopts a couple of ideas from here and others from there, it may easily turn out to that they do not fit together well, and if they don't, incoherence is likely to follow. To make things worse, when one doesn't belong to the very core of a particular camp (as very few people actually tend to do), there's always the additional danger that one might give an inappropriate reading to some of the concepts developed there, because one lacks the necessary background information, which is often shared only by true initiates of the respective school of thought. Being aware of these dangers, I would like this contribution to be understood as a public exercise, in which I assume the role of a historical linguist with a naturalist background trying to find out in what way he might profit from the methods developed and the insights gained in OT. Trying to outline what (I think that) "constraints" and "preferences" mean when they are used by Optimality Theorists on the one hand, and Naturalist Linguists on the other, I hope to point out some of the dangers that may await one when one tries to integrate concepts of different theoretical origins. In the exercise, I will find myself obliged to sketch, in very few words, my personal inter-

294

Nikolaus

Ritt

pretation of what OT and NL "are about". Needless to say, those interpretations are going to be subjective and over-simple, and colleagues more knowledgeable about either OT or NL, or both, are likely to find the theoretical frameworks to be distorted to the point of caricature. To them I extend my apologies, hoping at the same time that they might derive some profit from learning how their views may happen to be received and interpreted within the wider linguistic community.

2.

Why "constraints" and "preferences" might come across as near synonyms

As is well known, the term "constraint" is extremely common in the discourse of a large number of academic disciplines. As far as the term "preference" is concerned, it is well represented even in everyday language. At the same time, however, most linguists working in the nineteen nineties are likely to associate the terms with two particular schools of linguistic thought, namely Optimality Theory on the one hand, and Natural Linguistics on the other. In these, the terms "constraint" and "preference" form part of the technical vocabulary and carry specific, "theory-laden" meanings. To make matters slightly more complicated, it can be difficult to tell for individual occurrences of the two words, if they are supposed to convey the meanings they have in general usage, or rather their theory specific ones. This is because the two types of meaning overlap to a considerable degree. Thus, the two innocent looking lexical items qualify perfectly as what the philosopher of science David Hull has dubbed "weasel words" (1988: 6). These represent "key words in ... scientific disputes [that have] ... a certain plasticity to them" and "expand ... and contract... in systematic ways to fulfill the needs of the scientists using them." Sometimes, they are used in theory-specific senses, at others they are used as if they were nontheoretical terms. Since, in the cases of "constraint" and "prerefence" a nontheoretical reading of the two items makes sense more often than not when they occur in both OT and NL texts, theoretical individualists and eclectics, as

Are OT "constraints " the same as NL "preferences "?

295

well as students consulting literature in search for solutions to specific language related questions, may easily get the impression that OT "constraints" and NL "preferences" refer essentially to very similar things. Examples of scholars from camps who use the terms in ways that seem to support this conclusion are not difficult to find. For purposes of illustration, I shall quote from three authors. In all quotations the italicised occurrences of the terms "constraint" and "preference" appear to be more or less co-referential. There is a ... tradition in ... linguistics which holds that structural patterns of grammar are determined by highly general preferences or constraints [my italics, NR] that may come into conflict with each other (Haspelmath 1999: 1). These partial assimilations can be viewed as a compromise between achieving the F2 targets for the consonant and vowel, and a preference [my italics, NR] to avoid fast movements between the two (hence a preference to minimize the difference between the two) (Flemming 1999: §9). We have seen that mutual assimilation between consonants and vowels in CV sequences is one instance of phonetic compromise that can be perspicuously analyzed in terms of optimization with respect to conflicting constraints (Flemming 1999: §47). There is a universal preference for two beats per foot. ... Three aspects of the above statement require some discussion: (a) the preference [my italics, NR] for a trochaic foot pattern, (b) the preference for ... (DziubalskaKolaczyk 1995: 58). Dogil (1980) suggests the Trochaic Projection Constraint [my italics, NR] (1980: 92) to account for elementary systems of the world's languages due to the existing internal evidence (1980: 89ff.) (Dziubalska 1995: 58).

Also, if one looks at the ways in which the two approaches might talk about specific phenomena, such as for example the phenomenon that Anglo Norman loans into Middle English underwent what has sometimes be referred to as a right to left stress shift, one might find additional support for the idea that "constraints" and "preferences" are just slightly different ways of talking about the same thing. Thus, arguing within an optimality theoretic framework, Donka Minkova attributes that stress shift to the relatively high ranking of an "Initial

296

Nikolaus Ritt

Prominence" constraint (which basically demands that word initial syllables should be footheads) in Middle English. At this point, the details of OT need not concern us. Experts will be familiar with them, and for all others3 it is sufficient to know for now that OT assumes that there exist a so-far unspecified but finite number of universal constraints on linguistic output, each of which, however, may be violated. What is grammatical in any particular language is then determined by how universal constraints are "ranked" within the grammar of that language. Since "Initial Prominence" is assumed to have been high ranking in Middle English, the fact that a potential form /prin.ceps/ would have violated it, while its rival /prin.ceps/ clearly did not, explains why the latter and not the former was selected as the grammatical form of the word in Middle English. Following the formal conventions regarded as most appropriate to the task within OT, Minkova (1997: 141) captures this explanation in a tableau of the following kind: •^prin.ceps Root Stress Initial Prominence Non-finality 4

WSP Rightward Main Stress

prin.ceps !*

V

!*

V

From an NL perspective, on the other hand, the argument might look roughly like this: there exists a universal preference for trochees (which are initially prominent, of course). Since it is typologically adequate for stress timing languages to respond to this preference, Middle English (which apparently was stress timing) must have been likely to implement it as well, so that, not surprisingly, /prin.ceps/ came to be "preferred" over /prin.ceps/. Although NL does not provide a formalism comparable to OT tableaus, it thus seems that the universal preference for trochees assumed by Naturalist Linguists plays a role in their framework which is similar to that which constraints like "Initial Prominence" play within OT. Therefore, the impression might arise once again that

Are OT "constraints " the same as NL "preferences "?

297

"preferences" and "constraints" do similar jobs, although within slightly different frameworks. Since the discourse of both Optimality Theorists and Natural Linguists seems to suggest that "constraints" and "preferences" have similar functions, one might further conclude that OT "constraints" could be readily imported into the NL framework, and vice versa. One a general level one might be tempted to ask, for example, whether OT tableaux would not represent a suitable enrichment of the formal machinery of NL. Alternatively, one might want to have a more detailed look at some of the individual "constraints" that OT has come up with, or at some of the "preferences" that are assumed in "NL", in order to see whether some of them might not be profitably traded among the two camps.

3. Theory specific meanings of "constraints" and "preferences": an exercise in inter-terminological translation Now, everybody with only a slight inkling of the vast differences that may exist between "formalist" approaches to language (such as OT) on the one hand and "functional" ones (such as NL) on the other will know that things are not likely to be as easy. Although most practising linguists are aware of the problem, however, they do not normally address it explicitly, or in the most adequate manner. Once we have been socialised within a particular scholarly sub-community we tend to adopt and use its specific language or jargon as if it were the most natural thing to do so. And indeed, in some way it is: after all, we need to get on with our work, rather than get involved in possibly endless meta-reflection. At some point, when we approach any particular problem we are interested in, we have to have our conceptual and methodological tolls ready and concentrate on wielding them. Also, those among our colleagues with whom we prefer to communicate tend to share our language anyway, and about the rest of the world we often prefer not to care too much. This may have advantages for the specific theories we happen to be attached to, because it keeps them safe from outside interference, but whether it is

298

Nikolaus Ritt

at the same time good for us and our understanding of the subjects we profess, is another matter entirely. Without explicitly wanting to do so, we create language barriers for ourselves and thus prevent ideas from flowing as freely within the greater linguistic community. No matter how successful or marginal each specific theoretical camp may eventually find itself to be, we all tend - each in their own ways - to settle into a kind of comfortable parochialism, which becomes awkward only at international conferences, when nobody can think of good questions to ask speakers. The linguistic community being organised as it is, then, the transfer of concepts among individual schools is not likely to be straightforward, superficial impressions notwithstanding. Thus, even when we intend to relate concepts to each other that appear to be so straightforwardly comparable as OT "constraints" and NL "preferences" and that appear to be used as if they were near synonyms, we are likely to be challenged first of all as translators, and only afterwards as representatives of our respective theoretical persuasions. We need to treat the discourses of OT and NL as if they were in different, albeit historically related languages and ask ourselves how the OT term "constraint" translates into NL and vice versa. The assumption that underlies our belief in the possibility of translation in the first place, of course, is that languages represent reality rather than constructing it in the radical sense. Simply speaking, when one wants to translate a word from language A into language B, one needs to find out which aspect of reality it represents first, and then one has to find the most adequate way of representing that particular aspect of reality in language B. Look then, in a very crude manner, at the aspects of reality that OT and NL try to capture with their respective special purpose languages and see how the terms "constraint" and "preference" map onto them. What both schools have in common, and what makes them comparable in the first place, of course, is that both are concerned with the question what natural human languages are like and why. Unfortunately, this is not to say much. The word "language" represents a very complex and at the same time multi-facetted set of

Are OT "constraints " the same as NL "preferences "?

299

phenomena. Thus, in a particularly nasty way, "language", in its everyday sense, seems to inhabit various different ontological domains, schematically represented in figure (4) below: first, language is a mental and therefore biological phenomenon. As such it is referred to as knowledge of a particular language, as I-language, as a brain state, as competence, grammar or whatever. Second, language is an externally physical phenomenon that comes in the form of sound waves or visual patterns. These are referred to as texts, utterances, pieces of discourse, Ε-language or by similar terms. Third, language is a behavioural phenomenon, stretching over the mental, physiological and physical domains. As such, it comes in the form of (mind governed) activities of the sensori-motor system, i.e. articulatory gestures, perceptual processes and so on. These processes are often called performance, speech acts, utterances or discourse. Fourth, language has also often been regarded as a social or even abstract phenomenon, inhabiting a transpersonal domain of unclear ontological status, reminiscent of what Popper referred to as "world III". To make things worse, the interrelationships among languagerelated phenomena from the various domains just mentioned are complex and poorly understood, and much the same holds true for the relations between language related phenomena on the one hand and certain potentially relevant aspects of the non-linguistic universe, such as the make-up of human organisms, the working of human minds in general, the structure and the dynamics of human societies and social groups, the properties of culture in the widest sense, and so on and so forth. To make this rather general issue slightly more tangible, some of the possibilities are outlined in the following graph.

300

Nikolaus Ritt

(2)

Language and its place in the universe (simplified)

Individual grammars

Grammar Pool

Communicative Behaviour Individual texts

Biosphere

Physical Universe

Clearly, when one is confronted with such an intricate network of complexly related phenomena and tries to understand any aspect of it, it becomes essential to think of a suitable strategy in order not to get completely lost. One certainly cannot dedicate the same amount of attention and care to all potentially relevant aspects at once. One has to distribute one's energies economically. Now, although both OT-like approaches to language and Naturalist ones are primarily interested in language as a mental phenomenon, i.e. langue, competence, or I-language, they differ crucially in how they approach and attempt to deal with the complex array of phenomena that the language-related parts of the universe represent. Rooted in - while attempting to revolutionise - the formalist Chomskian tradition of American Generative Linguistics (see, for

Are OT "constraints" the same as NL "preferences"?

301

example, Chomsky: 1995 for a survey of crucial asumptions), OT pursues a straightforward divide-and-conquer strategy: competences, I-languages or "grammars" are deliberately isolated and abstracted from the rest of the universe, so to speak. They are thought of as brains states representing computational systems whose most essential property is that they "generate", "derive" or "attribute structure" to all well-formed expressions in the languages which they happen to be competences of, and that they thus distinguish such well-formed textual expressions from those that are not well-formed or grammatical. Scientific attention is then focussed directly on modelling such idealised systems. At the same time the possible shapes which such systems can assume are believed to be constrained - rather tightly by the properties of a mental language module, regarded as a species specific genetic endowment of human minds. Thus, the primary goal of the research programme of which OT is a recent variant is to develop explicit, coherent and rigidly formalised models of grammatical knowledge. The secondary (but not less important) goal is to relate such models of individual grammars to a more general model of the assumed genetically determined human language module, referred to as "Universal Grammar", "Language Acquisition Device" or "Initial State". Individual models of specific grammars, models of "Universal Grammar" and models of the relation between them must be licensed by a comprehensive "Theory of Grammar", about which there should (at least in principle) be consensus within the scholarly community. Representing deliberate abstractions, models of "grammar" have to be externally adequate (i.e. comply with empirically accessible manifestations of language) only in the sense of being compatible with grammaticality judgements on actual or potential texts. Such judgements are normally derived through introspection and/or elicitation. On the other hand, generative competence models also need to be internally adequate. Criteria for internal adequacy are typically derived from considerations of descriptive economy and/or aesthetics, and to some extent they appear to be a matter of finding majorities within the academic community. - The relation between "I-language", "grammar" or "competence" on the one hand and other language related aspects of the world on the other is deliber-

302

Nikolaus Ritt

ately left for later consideration, and sometimes regarded as uninteresting and/or inaccessible to systematic study. In the graph below, which schematically represents the sections of the universe on which American Generative Linguists focus their attention, they figure as a grey background area imprinted with question marks. (3)

Language and the universe from a Generative point of view Universe

grammaticality judgements

grammaticality judgements

nmar

Naturalist linguists, on the other hand, being rooted in functionalist traditions, pursue a rather different strategy. Approaching the task of modelling I-languages relatively casually and with comparably little concern for explicit formalisation, they prefer to scan more or less systematically - sometimes even rather crudely - a much wider area of language related phenomena in search for potentially telling

Are OT "constraints " the same as NL "preferences "?

303

correlations. The idea behind this is that grammars might have the properties they seem to have because of the particular ways in which they are (functionally) related to their more or less language related context, i.e. to the makeup of human bodies, the working of cognition, communicative needs, social dynamics, cultural transmission and change, etc. A schematic representation of the aspects of the world that Naturalists investigate in their search for the nature of grammar is given in figure (4) below. Of course, while the functionalist approach suggests a possibly broad search for correlations between properties of grammars on the one hand, and properties of basically all aspects of the universe to which grammars might be functionally related on the other, it forces Naturalist linguists to approach language not only from various different ends but in practice also with differing methodologies specifically tailored to the needs of particular perspectives. Therefore, the unification of their findings into explicit and formally coherent theories tends to be slightly problematic. It creates a constant need for so-called "bridge theories" and possibly comprehensive meta-frameworks, such as semiotics, for example. Sometimes also, emerging inconsistencies or apparent contradictions, if they are at all noticed, are simply left unresolved for the moment in order to be considered more carefully at a later stage. Thus, quite apart from the question which of the two strategies is more reasonable and which will eventually turn out to be more successful, it is obvious that there is a vast difference between the research programmes to which OT and NL subscribe with regard to the sections of language related reality they deal with. This, in turn, is likely to have consequences for the roles which the concepts we are here dealing with, i.e. "constraint" and "preference" will assume within the discourse of the two linguistic sub-communities.

304

Nikolaus Ritt

Language and the universe from a functionalist perspective Universe

acquisition language disorders

attitudes

neurobiology

poetry

Grammar social structures

semiotics speech errors

language change

physiology universal grammar

Look at "constraints" first. Within OT, they refer to descriptive devices within a novel way of modelling grammars that is striving for dominance within the generative research community. It is, essentially, a new descriptive formalism. Instead of generating surface structures of well-formed expressions in a language from more basic underlying structures by means of intrinsically ordered sequences of re-write rules of the general format A —» Β /C D, OT derives them by having an evaluative component ("EVAL") select the "optimal" or most "harmonic" output from a large set of competing potential outputs, assumedly produced in parallel by a powerful generative component ("GEN"). Which of the competing outputs gets selected depends on the way in which constraints, in turn provided by universal grammar, are ranked within the grammar of a specific language.

Are OT "constraints " the same as NL "preferences "?

305

Since it is in principle legal for outputs to violate constraints, it is the language specific ranking of constraints that determines which of the competing outputs gets selected as the most harmonic one. Basically, an output gets selected if the most high-ranking of the constraints it violates still ranks lower than the constraints violated by competing outputs. Allegedly, the grammar design proposed by OT can distribute "the explanatory burden" better than "re-write rule theory", with which there was a growing dissatisfaction and which Prince and Smolensky, the fathers of OT, explicitly called "loose and uninformative" (1993: 3). It would clearly be pointless to discuss here to what degree claims to the superiority of OT over re-write rule theory are justified or not, nor would it make sense to launch a general discussion of the merits of either OT or the generative research programme in general. Instead, what needs to be pointed out is within OT, "constraints" refer primarily to theory internal descriptive/explanatory devices, designed to facilitate the modelling of idealised "grammars" or "competences". Their interpretability in terms of the material world is not a crucial issue as long as they work effectively enough within generative competence models, because such models are primarily required to achieve - or possibly simulate - what actual neuronally embedded competences of real speakers do, without however necessarily achieving or doing it by identical, or even similar means. Therefore, such questions as where constraints might reside in human brains and how or why they might get there are not felt to be particularly urgent, though it would of course be nice if they happened to be answerable. In NL, on the other hand, the term "constraint" is - or so it seems to me - typically used to refer to language external conditions, such as the make-up of human bodies or the principles that govern information processing in human brains. As external "constraints", these conditions are in turn assumed to impose restrictions on the probable designs of functional competences. However, the mechanics of the relation are in many cases not fully understood. Thus, the respective referents of OT vs. NL "constraints" belong to completely different ontological domains, as illustrated in the following graph.

306

Nikolaus Ritt

(5)

The ontologies of OT vs. NL "constraints" NL constraints

OT c o n s t r a i n t s Individual grammars

Grammar Pool

jmunicative behaviour Individual texts

Biosphere

Physical Universe

Consider "preferences" next. NL faces the problem - often polemically held against the whole functionalist approach (see, for instance, Lass 1980 and 1997) - that the generalisations which it derives about the structures of specific competences or I-languages and which it derives often from language external constraints take the form of statistical, or probabilistic laws. Like OT "constraints", these probabilistic generalisations may sometimes conflict with one another, but unlike OT, NL provides as yet no equally explicit, principled and straightforward mechanism for describing how such conflicts get resolved, although considerable progress in that respect is certainly being made. Yet, remember NL is facing a considerably more daunting task than OT, which - possibly wisely but possibly also over-cautiously - tends to restrict itself to grammar internal con-

Are OT "constraints " the same as NL "preferences "?

307

siderations. The way from external constraints to language internal design leads right though the realm of human minds, which are under the influence of allegedly autonomous, free-willed human selves. Since reliable or even widely accepted models of the human mind are still a long way from being developed - in spite of Pinker (1997), Dennett (1991), Hofstadter (1979), Crick (1994) and others - Naturalist linguists face the problem of deciding what to do in the meantime. It is perfectly legitimate, it seems to me, that they should adhere to idea of hypothetically relating competence design to contextual, language external constraints even though they might not be able to chart the detailed mechanics involved in the interaction. Thus, in the absence of a theory about the mechanics of minds and selves, NL needs a concept in which to couch all that is unknowable for the time being, and this is the task which the concept of "preferences" is designed to perform, it seems to me. And it performs it exceedingly well. Reminding one of the impenetrable mysteries underlying personal likes and dislikes with regard to such trivial matters as tea or coffee, milk or cream, cats or dogs, Mozart or Bach it represents a disarmingly honest acknowledgement of where NL reaches - at least for the time being - its explanatory limit. Now, as far as OT is concerned, it is clear that it should not be particularly interested in "preferences" of this kind. Focussing on Ilanguage, it never even needs to face the problems associated with accounting for idiosyncratic speaker behaviour. From the American generative perspective, behaviour represents part of performance and is thus deliberately excluded from linguistic investigation proper. Not surprisingly, therefore, in the introduction to Prince and Smolensky (1993), for example, the term "preference" is conspicuously avoided and where it does occur in OT literature, it is typically in the more informal parts - the more typically so the more orthodox the contribution.

308

Nikolaus Ritt

4. Conclusion So, what I hope to have made clear is that the relationship between the usages of the terms "preference" and "constraint" within OT on the one hand and NL on the other is anything but straightforward and that they certainly do not simply translate into each other. Trying to enrich the set of "preferences" acknowledged within NL through importing OT "constraints", or vice versa, would thus represent an endeavour too risky and complicated, in my view, to be worth the effort. If at all, it seems to me, the two approaches might profit from each other in less direct ways. They might try, for example, to broaden their horizons by trading concerns rather than products. Thus, OT tempts linguists to neglect the question how grammar internal constraints and their rankings should be justified. Instead, they tend to defend specific constraints and their rankings by arguing that "they work" within descriptive models of particular languages. Thus, there is always the danger of circularity and explanatory emptiness. NL, on the other hand, could do well with a good measure of explicitness. Although it might work sufficiently well as a research strategy, it may not be enough in the long run defend the probabilistic nature of some of NL's generalisations by vague reference to goal conflicts and the inherent unpredictability of human individuals. At least, one will have to try and investigate where exactly unpredictability lurks and why. Those are just two obvious examples, of course. Definitely, there are further areas in which ideas might be profitably exchanged. As I hope to have shown, however, the success of all such trade seems to require serious attempts to overcome the language barriers that have been formed between the linguistic subcommunities.

Notes 1. It is only because my attitude towards and my expectations in general linguistic theory may be representative of a larger section of the linguistic community, that I am bringing them up here. 2. Particularly Dressier (1985, 1989) and Dziubalska-Kolaczyk (1995).

Are OT "constraints " the same as NL "preferences "?

309

3. Particularly those who want to read on before familiarising themselves with OT by reading for instance Archangeli and Langendoen (1997) or Prince and Smolensky (1993). 4. WSP is short for "Weight-to-Stress Principle".

References Archangeli, Diana and D. Terence Langendoen (eds.) 1997 Optimality Theory: an Overview. Oxford: Blackwell. Chomsky, Noam 1995 The Minimalist Program. Cambridge, Mass.: MIT Press. Crick, Francis 1994 The Astonishing Hypothesis: the Scientific Search for the Soul. London: Simon and Schuster. Dennett, Daniel C. 1991 Consciousness Explained. London: Penguin. Dogil, Grzegorz 1981 Elementary accent systems. In: Wolfgang U. Dressler, Oskar E. Pfeiffer and John R. Rennison (eds.), Phonologica 1980, 89-99. Innsbruck: Innsbrucker Beiträge zur Sprachwissenschaft. Dressler, Wolfgang U. 1985 Morphonolgy: the Dynamics of Derivation. Ann Arbor: Karoma. 1989 Semiotische Parameter in einer textlinguistischen Natürlichkeitstheorie. Wien: Verlag der österreichischen Akademie der Wissenschaften. Dziubalska-Kolaczyk, Katarzyna 1995 Phonology without the Syllable. A Study in the Natural Framework. Poznari: Motivex. Flemming, Edward 1999 Phonetic optimization: compromise in speech production. Stanford University Online Paper, [www.stanford.edu/~flemming/paper/ot. pdf, Date of retrieval: 24/04/1999.] Haspelmath, Martin 1999 Optimality and diachronic adaptation. Zeitschrift für Sprachwissenschaft 18:180-206. Hofstadter, Doulas R. 1979 Gödel, Escher, Bach. An Eternal Golden Braid. New York: Basic Books.

310

Nikolaus Ritt

Hull, David L. 1988 Science as a Process. Chicago and London: University of Chicago Press. Lass, Roger 1980 On Explaining Language Change. Cambridge: Cambridge University Press. 1997 Historical Linguistics and Language Change. Cambridge: Cambridge University Press. Minkova, Donka 1997 Constraint ranking in Middle English stress shifting. English Language and Linguistics 1: 135-177. Pinker, Steven 1997 How the Mind Works. New York/London: Norton. Prince, Alan and Paul Smolensky 1993 Optimality Theory: Constraint Interaction in Generative Grammar. RuCCS Technical Report 2. Rutgers Center for Cognitive Science, Rutgers University and Department of Computer Science, University of Colorado at Boulder, [to appear, Linguistic Inquiry Monographs. Cambridge, Mass.: MIT Press.]

Abstractness in phonology: the case of virtual geminates Philippe Segeral and Tobias Scheer

1. Introduction Paul Kiparsky's (1968) question how abstract is phonology? initiated a discussion about the dissimilarity that phonological theory should allow for between phonetic reality and phonological representations. One aspect of the debate concerned abstract segments. These are objects which occur in the underlying structure of a morpheme, but appear in none of its phonetic manifestations, or, worse, are even absent from the segmental inventory of the language.1 In Generative Phonology, the discussion on abstractness led to the view that abstract structures should be dispensed with in the presence of solutions that are bound to the surface in a more direct fashion. This position is expressed in subsequent work by Kiparsky (Kiparsky 1982 contains a number of relevant papers) and, among others, by Tranel (1981). It was the cornerstone of Natural Generative Phonology (Hooper 1976; Vennemann 1974a, b) and Natural Phonology (e.g., Stampe 1973; Dressier 1974). However, the debate never came to a conclusive end. In particular, no satisfactory definition of what is natural, simple, elegant, phonetically plausible, psychologically real or typologically invariant could be achieved (Hellberg 1978; Koutsoudas 1980; Dinnsen 1980). Neither could an evaluation metric be derived that would be able to measure different degrees of abstractness (Kiparsky 1974; Campbell 1981; Goyvaerts 1981). Proponents of abstract analyses with no other limitation than the learnability of rules (cf. Skousen 1981) are, among others, Schane (1968, 1974), Dell (1973), Selkirk and Vergnaud (1973), Dinnsen (1980) and, taking abstractness to an extreme point, Lightner (1981). Kenstowicz and Kisseberth (1977: 1-62)

312

Philippe Segerai and Tobias Scheer

provide an extensive discussion of the topic. They argue in Kenstowicz and Kisseberth (1979: 204ff.) that in some cases, there is no alternative to abstract analyses that are motivated by compelling internal evidence. Abstractness as a topic of phonological debate has played little or no role in the 80's. Focus was switched on the internal structure of segments, as well as on issues regarding autosegmental representations. In the 90's, theories that are concerned with constraints are involved in neither of the former issues; abstractness, the internal structure of segments and phonological representations play little or no role in the discussion. However, we believe that the questions related to abstractness, though raised in the 70' s, are still relevant for phonological theory. As a fact in point, another typical topic of the rule-ordering period has come back into the spotlight recently: Optimality Theory is seeking to cope with opacity. In this paper, we discuss a type of abstract segment that seems to be recurrently required for the purpose of various analyses in genetically unrelated languages. Attention is drawn on underlying geminates that are always realised as simple consonants. Such an object, which we call a virtual geminate, never betrays its geminate identity by a phonetic clue related to length, but by other properties that can be read off the phonetic environment. A situation of this kind has been identified for example in Danish, cf. Larsen (1994), Lowenstamm (1996). Virtual geminates are not exactly abstract segments in the sense of the above discussion. Assuming autosegmental representations, their melodic structure is identical to the one of the corresponding simplex consonant, which is part of the phonemic inventory of the language in all cases we are aware of. Their virtual character is due to the double association they enjoy in regard of the skeleton, and which does not lead to a long pronunciation. The very same situation is actually encountered for vowels, too. In certain Afro-Asiatic languages, peripheral vowels behave as long, even though they are phonetically as short as central vowels, which behave like true simplex segments. Lowenstamm (1991) has

Abstractness in phonology: the case of virtual geminates

313

formalised this observation, and a number of generalisations in various languages crucially rely on this view (Segeral 1995, 1996; Bendjaballah 1999). If there are reasons to believe that vowels may enjoy a double association and yet be phonetically simplex, not only is it possible, but expected to also find virtual geminates. For there is no difference in representing virtual long vowels and virtual geminates in autosegmental structures. (1) below illustrates this situation for a virtual /tt/ and a virtual /aa/.

Hence, autosegmental representations, that were not available (or not well enough established) when the debate on abstractness took place, invite to differentiate two kinds of abstract segments. Phonemes whose underlying identity is never visible on the surface (or not even present in the inventory of the language) on the one hand, and an underlying relationship between two or more autosegmental tiers that has not the expected surface-effect on the other. Although it is not advisable to make a formal prediction on the grounds of the empirical evidence of one article, it may turn out that "abstract segments" in the pre-autosegmental sense are always (or mostly) of the kind that we are dealing with in the next few sections, that is as under (1). In any event, we submit that the two case studies which we present conform to the minimal requirement imposed on abstract underlying forms: they must be retrievable by children during the process of acquisition. Objects that do not occur on the surface may be incorporated into underlying representations as long as their identity can be recovered on the grounds of phonetic information provided by the context.

314

Philippe Segerai and Tobias Scheer

We adduce evidence for the existence of virtual geminates in Cologne German (section 2) and Somali (section 3), arguing that a proper analysis of these languages cannot be carried out unless the existence of virtual geminates is recognised.

2. Cologne German Cologne German (henceforth CG) is peculiar in that it features a velarisation of dental consonants. The reverse movement illustrated by palatalisations is extremely common, and usually triggered by an adjacent palatal segment. The few velarisations on record are either of dissimilatory nature or obey syllabic configurations, cf. English "1-darkening" in Codas. The Cologne German velarisation is worthwhile because it does not instantiate a dissimilation, nor is it conditioned by any syllabic pattern. Compare the Cologne forms below with their Standard New High German (henceforth NHG) cognates. (2) a.

b.

correspondence

NHG

Cologne German

spelling

gloss

n-o

btfawn tsawn Kajn lajns nojn bifawt Kajtan hojta Jawderli? Jnajdsn lajdan bunt pints Kundan bindsn byndsl

bKuq 'tsug κι η lir,

braun Zaun Rhein Leine neun Braut reiten heute schauderlich schneiden leiden bunt Pinte runden binden Bündel

brown fence Rhine rope nine bride ride today frightening cut suffer colorful pint round tie bundle

t-k

d-g

c.

nt-qk nd - qg 2

nyr) btruk ifiga hyk JugiJ jnigs ligs bui]k pigk Kyijs biga pyqal

Abstractness in phonology: the case of virtual geminates

315

The correspondence shown is regular, and is present in the entire Cologne lexicon. A complete description is given in Scheer (forthc.); the phenomenon is also discussed in the traditional comparative and dialectological literature, e.g., Hönig (1877), Münch (1904), Lessiak (1933), Frings and Schmitt (1942), Bertram (1935), Müller (1942). Synchronically, there is no way to predict the context in which this velarisation occurs. The only generalisation that may be drawn is of comparative nature: New High German simple dentals as under (2a,b) occur as velars in Cologne German words when preceded by a diphthong [aj, aw, oj] in New High German, and only in this environment. However, there are numerous Cologne German words that do not possess velars where New High German shows diphthongs, e.g., both NHG and CG [ftajn, bajn, bKajt, lawt, fxojd(s)] Stein, Bein, breit, laut, Freude 'stone', 'leg', 'large', 'noisy', 'joy', CG *[ftiq, big, bKik, luk, fxyk]. Diachronic evidence clears up the situation: the common ancestor of both New High German and Cologne German is Middle-High German (henceforth MHG, about 1050-1350 A.D.). As shown under (3), New High German diphthongs have two different MHG sources (e.g., Paul et al. 1989: 68ff., 105ff.). (3) mein neues Haus NHG [aj] [oj] [aw] Bein, säugen, Auge

'my new house' MHG min niuves hüs —"" [ii] [yy] [uu] ^ T ^^

[ej] [oj] [ow] MHG bein, säugen, ouge 'leg, breastfeed, eye'

Cologne German contains 97 velar roots that have a New High German dental cognate, 42 of which are simplex consonants as under (2a, b).3 All of those 42 words show a root-vowel that goes back to a MHG long high vowel, none originates in a MHG diphthong. We may thus identify Cologne German velarisation as the following diachronic process: MHG simplex dentals become Cologne German velars when preceded by MHG long high vowels.

316

Philippe Segerai and Tobias Scheer

(4) MHG [n,t,d] > CG [rj,k,g] / MHG {[ii,uu,yy]}_ The velarisation discussed is to be regarded as a synchronically inactive process that has occurred in the evolution from MHG to Cologne German. It cannot be part of the synchronic phonology of Cologne German since the triggering context "a preceding long high vowel" has been altered since MHG. Another piece of evidence concerns the triggering context: the vowel preceding velarised consonants has encountered an abnormal evolution. In non-velarised words, MHG long high vowels come out unaltered in CG, e.g., MHG hüs, wip, tiuvel [huus, viip, tyyval] > CG [huus, viif, dyyval] 'house', 'woman', 'devil'. 4 Contrary to this regular evolution, all velar Cologne German forms that possess a New High German dental cognate come along with a short vowel, e.g., MHD brün, hiute, sniden [bKuun, hyyta, sniidsn] > CG [bKug, hyk, jniga] braun, heute, schneiden etc., cf. (2). Vowel-shortening in velarised words is thus to be interpreted as a consequence of velarisation. When looking at the part of the corpus that involves dental clusters becoming velar in Cologne German (2c), it turns out that the generalisation made under (4) also holds true as far as vowel-quality is concerned. MHG [nt, nd] become CG [qk, qg] if preceded by a high vowel. The dental cluster of all items on record in the 97-root list mentioned earlier is preceded by a high vowel. Only is this vowel short, and not long as before. We are thus facing a general process whereby velarisation is triggered for all candidates by preceding high vowels. This condition is sufficient in case of dental clusters. It is only necessary for the velarisation of simplex dental consonants, however. The sufficient condition here is the length of the preceding vowel. Hence, the quality of the triggering vowel is uniform for all kinds of velarised items. Its quantity depends on the quantity of the target: if the candidate is short, the trigger must be long. If it is long, the trigger is short. For the sake of clarity, all generalisations that have been made so far are summed up below.

Abstractness in phonology: the case of virtual geminates

(5)

317

Generalisations a. only high vowels that precede dentals trigger velarisation b. simplex dentals velarise only if preceded by a long vowel c. when velarisation of simplex dentals takes place, the preceding long vowel shortens d. dental clusters [nt,nd] velarise after short vowels5

The length of the trigger and the target is in complementary distribution. This is a strong indication to the effect that the overall volume involved in the process is constant. Namely, it must be of three timing units: in all cases, trigger and target occupy three skeletal slots. This is illustrated in (6) below (T=dentals, K=velars). (6)

VNT > VrjK VVT > VKK

MHG bunt [bunt] > CG [buqk] MHG hiute [hyyta] > CG [hyk]

The crucial feature of (6) is the gemination of simplex dentals when velarisation occurs: VVT > VKK. Indeed, this is the only way of accounting for vowel shortening (5 c) while maintaining the overall constant volume of three skeletal slots. Indeed, if simplex dentals are not only affected by velarisation, but also by gemination, the overall picture makes sense. Only under this proviso can the Cologne German velarisation be described as a uniform and unique process. The intricate alternations concerning the quantity of the segments involved may then be explained as under (7). (7) a.

b.

velarisation of simplex consonants takes place only after long vowels because only long vowels may provide the additional skeletal slot which is required for the gemination of the velar. velarisation of dental clusters [nt,nd] is triggered by short vowels because clusters do not need to acquire an additional skeletal position in order to be geminate: they are already double.

318

Philippe Segerai and Tobias Scheer

c.

long vowels preceding velarised consonants undergo shortening because their second skeletal position is occupied by the velar. If velarised simplex dentals are to be regarded as geminates at the end of the diachronic process at stake, are they still underlyingly geminate in present-day Cologne German? On the surface, they never appear as such, as was stated before. The following diachronic analysis could be argued for: MHG simplex dentals were velarised and geminated in the mentioned context. The stage of the language where this result occurred in is unrecorded. In further development, an independent process of degemination took place. These three steps appear in (8) below. (8)

a. MHG b. intermediate stage: after velarisation and gemination c. CG: after degemination

[VVT] *[VKK] [VK]

There is, however, evidence to the effect that phonetically simplex velars which come from dentals are still underlyingly geminate in actual Cologne German. Consider the distribution of Cologne German underlying /g/, which enjoys six different surface manifestations (so-called spirantische Kontaktregel, see Frings 1955 on this issue). MHG [g] appears in Cologne German as shown under (9).

Abstractness in phonology: the case of virtual geminates

319

(9) CG

gloss favourable, big

ejaal

egal

doesn't matter

Wfron,_V

nasaejal

Nägel

nails

c

balje

balgen

scuffle

v_v ÜJ /

spelling

jynstij", JKOOS günstig, groß

[ γ ] / V V n o n front 0 /

V z a a y s , naaysl sagen, Nagel

C

zaeaet, asesen

say INF, nail

er sagt, regnen he says, rain

/g/ Lf] /

V front

[χ]/

Y

[k] ^

#

cevKiJ, f u f t s i j übrig, fünfzig

be left, fifty

zaax !

sage !

say ! 2SG

_#C

zaaks

sag es

_-(C)

zaak

sagt!

say it! 2SG say! 2PL

non front

#

This alternation is synchronically active, as witnessed by the fact that the entire Cologne German lexicon is affected, including contemporary loans.6 For instance, words such as Gage ('fee', from French), Galerie ('galery', from French), Gig ('concert of a band', from English) and Glacehandschuhe ('transparent gloves', from French) are pronounced [jaaj, jabKii, jik, jlasants]. If a Cologne German speaker is asked to repeat a nonsense word such as Guntil, the result is [juntil]. As may be seen under (9), none of the allophones is identical to the underlying identity of the phoneme. Hence, if the allophony is part of the synchronic grammar of Cologne German, one should expect that there is no [g] at all in this language. This indeed is a correct prediction - except for voiced velar stops that are the result of velarisation: [g]s that are phonetically present in Cologne German have come into being through velarisation. Lexical items containing [g] are given under (10) below.7

320

Philippe Segerai and Tobias Scheer NHG

CG

NHG

CG

NHG

CG

krauten

kKugs

reiten

Kiga

streiten

lauten

luga

schneiden

Beutel

schauderlich

Jugil

weiten

itfiga viga

leiden

ligs

schreiten

^tsigij dyga

Trude

d^ygaljs

läuten

jtfiga lygs

deuten häuten

hygs

zeitig

Jtxigs bygsl

If the synchronic status of velar voiced stops which come from dentals were /g/ as suggested in (8), they would of course be expected to behave in the same way as regular /g/. Since velars that have a dental source are not affected by the allophony described under (9), their underlying identity must be different from /g/. The only solution that is consistent with both the velarisation and the allophony grants a synchronic geminate identity /gg/ to velars that originate in MHG dentals. These virtual velar geminates may be easily identified through the process of acquisition. Assuming that learners have successfully incorporated into their grammar the process shown under (9), they know that any time they come across a [g], its underlying representation must be different from /g/. If one does not wish to introduce an otherwise non-existing phoneme into underlying representations, and if absolute neutralisations are disqualified (cf. Kiparsky 1968), then the geminate identity is the only straightforward way for the learner to implement [g] into his/her grammar. In fact, learnability may be used as an argument to prefer non-phonemic abstract segments in the sense of (1) over abstract phonemes: virtual length is the only way to achieve a picture where the structure arrived at in the process of acquisition is consistent with diachronic and synchronic analysis. In sum, we are thus left with an underlying geminate that does never appear on the surface as such. Only its degeminated version [g] is observable. Note that there are no geminate consonants on the surface in Cologne German at all. It could thus be argued that degemination is a feature that is enforced upon /gg/ by the general structure of the language. In the next section, however, we discuss a parallel case of virtual geminates in Somali, a language where phonetic geminate consonants do occur.

Abstractness in phonology: the case of virtual geminates

321

3. Somali In Somali (Cushitic), intervocalic [t] such as in [daqnataa] 'you/she SG feel(s) pain' must be analysed as a phonological geminate. That is, the melodic identity of "t" is associated to two consonantal positions, while being invariably simplex on the surface. This interpretation is enforced by the interplay of three facts, that we present hereafter.

3.1. Morphological analysis In order to identify the morphological structure of the quoted [daqnataa], let us consider the conjugational paradigm of an unmarked verb such as keen 'bring', which represents the class of verbs that is traditionally termed "weak". ο

1 SG 2 SG 3 SG MASC 3 SG FEM 1 PL 2 PL 3 PL

root

SUBJ. A G R .

Tense marker

surface

keen keen keen keen keen keen keen

0

aa aa aa aa aa aa aa

keenaa keentaa keenaa keentaa keennaa keentaan keenaan

t 0

t η t 0

η η

Table (11) shows that the subject agreement marker [-t-] characterises 2 SG, 3SG FEM and 2PL forms, while [-n-] indicates IPL. All other persons enjoy a zero marker. Subject agreement is located between the root and the tense marker. The latter is [-aa] in the above examples illustrating the "Habitual Present" paradigm, as opposed to "Past" forms that are achieved by [-ay]: [keen-ay] '1/ he/ they have brought', [keen-t-ay] 'you SG, she has brought'. In addition, a final [-n] marks 2PL and 3PL. The set and the distribution of Subject Agreement markers shown is constant throughout all weak verbal paradigms in Somali.9

322

Philippe Segerai and Tobias Scheer

Somali weak verbs fall into different classes, two of which bear on the present demonstration. The above keen illustrates the unmarked verb class (so-called first conjugation). The verb [daqnataa] quoted earlier, on the other hand, belongs to the third conjugation, which contains autobenefactive and reflexive items. The characteristic marker of verbs belonging to this class is the suffix /-(a)t-/ that is directly attached to the stem, and whose vowel alternates with zero. For instance, the first person singular of the first conjugation verb fir 'comb TRANSITIVE' is [fir-aa] Ί comb TRANSITIVE', with the Subject Agreement marker [-0-] evidenced under (11). The autobenefactive version of the same verbs is [fir-t-aa] Ί comb my hair', where the class-marker /-(a)t-/ occurs in zero grade. Hence, the morphological composition of the autobenefactive form [daqnataa] 'you SG/she feel(s) pain' must be as under (12a). (12b) shows the related non-autobenefactive counterpart [daqantaa] 'you SG/She hurt(s)'.10 (12)

stem

a. daqn b. daqan

AUTOBEN

2SG HAB. PRES.

-at

-t

-aa

[daqnataa]

-t

-aa

[daqantaa]

'you (SG)/ she feel(s) pain' 'you (SG)/ she hurt(s)'

In other words, the simplex [-t-] in [daqnataa] is the phonetic reflex of two underlying /t/s that belong to two different morphemes.

3.2. Resistance to voicing The simplex [t] in [daqnataa] is also unexpected because dental stops voice and spirantise in intervocalic position in Somali. This process may be observed for the feminine definite article /-ta/, which surfaces as [-öa] if preceded by a vowel-final stem. Compare /laan-ta/ = [laanta] 'the branch', /deg-ta/ = [deg-ta] 'the peak' vs. /mindi-ta/ = [mindiöa] 'the knife', /hoojo-ta/ = [hoojaöa] 'the mother'.11 Illustration of the same process is also given by the autobenefactive /-(a)t-/ at stake in [daqnataa]. Its /-t-/ appears in voiced and spirantised form on the surface when it comes to stand in intervocalic

Abstractness in phonology: the case of virtual geminates

323

position. For instance, the first person singular of the autobenefactive verb kabbo 'sip' is [kabbaöaa] from /kabb-at-aa/, just as the first per1Ο son signular of daqno 'feel pain' [daqnaöaa] from /daqn-at-aa/. In autobenefactive verbs, LSG that bears a zero Subject Agreement marker and 2SG where agreement is achieved through l-X-l are thus opposed by the sole contrast [-Ö-] vs. [-t-]. (13)

daqno 'feel pain' [daqnaöaa] [daqnataa]

kabbo 'sip' [kabbaöaa] [kabbataa]

tjoogso 'stop o.s.' [tjoogsaöaa] [tjoogsataa]

1SG = STEM + (a)t + 0 + aa 2SG = STEM + (a)t + 1 + aa

It is a striking fact that intervocalic [t] resists voicing precisely in the cases where the morphological analysis reveals the presence of two dentals. On the other hand, whenever morphology provides one single l-X-l, voicing applies and a [-Ö-] is observed on the surface.

3.3. Vowel - zero alternations The attentive reader has noticed that vowel-zero alternations are frequent in the examples mentioned. They occur in two different sites. In a CViCV2C-stem, V2 may alternate with zero (daqn- / daqan'feel pain', cjerg- / cfereg- 'feel full', kidf- / kidif- 'shop', boqr- / boqor- 'make king', gudb- / gudub- 'cross'), in which case its vocalic manifestation is always a copy of Vi. The other instance concerns the vowel of the autobenefactive suffix, which appears as either [-0t-] (/fir-(a)t-aa/ = [fir-t-aa]) or [-aö-] (/daqn-(a)t-aa/ = [daqn-aö-aa]). In this case, the alternating vowel is invariably [a]. These alternations are controlled by the phonotactics of the consonantal environment: iff the alternation-site is preceded or followed (or both preceded and followed) by more than one consonant, the 13 vowel appears. In any other configuration, no vowel may be observed. Somali thereby instantiates the cross-linguistically stable pattern of vowel-zero alternations found in many other languages

324

Philippe Segerai and Tobias Scheer

such as Slavic, French, Moroccan Arabic or Hungarian (Scheer 1997, 1999). Another way of describing the situation is to say that three consonants in a row (or CC#) are prohibited: *CiC2C3. If the alternationsite is stem-internal, a copy of the first stem-vowel breaks up the cluster between Ci and C2 like in first conjugation verbs /daqn-t 2SG -aa/ = [daqantaa] 'you SG hurt'. In case the alternation-site is located in the autobenefactive suffix, potential *QC2C3 clusters are vocalised between C2 and C3, e.g., /daqn-(a)t-0 LSG-aa/ = [daqnaöaa] Ί feel pain'. Accordingly, no vowel surfaces in first person singular forms of first conjugation verbs. In /gudb-0-aa/ = [gudbaa] Ί cross', /kidf-0aa/ = [kidfaa] Ί shop', no *CCC cluster arises since the ISG marker is zero. No *CCC cluster is encountered in ISG forms of autobenefactive verbs either despite the presence of the suffixal -t- if the stem of these verbs ends in one single consonant: /fir-(a)t-0-aa/ = [firtaa] Ί comb my hair', /tuur-(a)t-0-aa/ = [tuurtaa] Ί pile up'. Contrastingly in second person forms of the first conjugation where the Subject Agreement marker is -t-, a three-membered cluster arises, which is broken up by a copy of the first stem-vowel, cf. /gudb-t-aa/ = [gudubtaa] 'you SG cross'. In the same way, first person singular forms of autobenefactive verbs whose stems end in two consonants are vocalised since the autobenefactive /-(a)t-/ adds a third consonant: /kabb-(a)t-0-aa/ = [kabbaöaa], /tJoogs-(a)t-0-aa/ = [tjoogsaöaa], /daqn-(a)t-0-aa/ = [daqnaöaa]. In second person forms of autobenefactive verbs, it may come as a surprise that the suffix does vocalise in [firataa, tuurataa, qabataa] even though no *CCC seems to occur. If the above analysis is correct, only two reasons may be invoked for this inordinate behaviour: either the alternating vowel is preceded, or it is followed by a consonant cluster. In other words, either the stem-final [-r,-b] in the above examples are to be viewed as clusters, or the following [-t-] is a geminate. As a point in fact, Somali has overt voiced geminates, among which [rr] and [bb], while voiceless consonants never occur phonetically geminated. Hence, there is no way to analyse the stemfinal consonants in the examples mentioned as underlying geminates,

Abstractness in phonology: the case of virtual geminates

325

since they would appear as such on the surface. The [-t-] following the alternating vowel, however, would not betray its identity if it were an underlying geminate. It is hard to believe in coincidence when considering that the [-t-] which triggers the abnormal presence of the alternating vowel is precisely the same [-t-] that irregularly resists voicing in intervocalic position. Its status as a phonological geminate would explain both issues. Moreover, its geminate identity is consistent with the morphological structure of the forms at stake: [firataa] is /fir-(a)t-t-aa/. In sum, we face phonetically simplex [t]s in intervocalic position that 1) resist voicing and 2) trigger the vocalisation of a preceding alternation-site in the same way as CC clusters do. In other words, these [t]s behave like geminates. Furthermore, morphological analysis shows that they do indeed represent two distinct /t/s. We believe that this set of facts cannot be accounted for unless it is admitted that intervocalic [t]s in Somali are phonological geminates /tt/. In absence of this analysis, the autobenefactive paradigm remains mysteriously irregular, no rule or intervocalic voicing may be formulated, nor do vowel-zero alternations obey any recoverable pattern.

3.4. Benefits The assumption of virtual geminates sheds also light on other intriguing situations elsewhere in the grammar. In some first conjugation verbs such as matag 'vomit', hitiq 'walk slowly' or mutuh 'speak without reserve', a medial intervocalic voiceless [t] occurs. Curiously, these verbs also disregard the regular vowel-zero alternations that are normally found in CViCVjC-verbs (Orwin 1995: 75; Saeed 1998: 23). Compare the ISG form [gudbaa] from /gudb/ with ISG forms [matayaa, hitiqaa, mutuhaa], *[matgaa, hitqaa, muthaa], from matag, hitiq, mutuh respectively. If [t] in these verbs is a geminate, both its resistance to voicing and its triggering quality of the stem-vocalisation are regular. These verbs do not need to be regarded as "exceptions" any longer.

326

Philippe Segerai and Tobias Scheer

In Somali compounds, the initial consonant of the second member is regularly geminated if the first member ends in a vowel. Compare for instance φγο 'ears' attached to beel 'lack o f , yielding [cjeyabbeel] 'deafness'. However, when the same word φγο 'ears' precedes tir 'cancel', the result is [cjeyatir] 'ignore, refuse to listen', bearing a phonetically simplex [t]. This [t] obviously instantiates a geminate. Another piece of evidence comes from the free variation between geminates and homorganic nasal-stop clusters that is found in Somali (Keenadiid 1976:XXII). Consider for example the doublets Ϋαψφ / Ϋαηφ 'scabies', middi / mindi 'knife', higgo / higgo 'hiccough'. Other cases of this free variation with homorganic nasal-stop clusters, however, involve simplex [t]s: mitid/mintid 'persevere', ?utub / iuntub 'small group', sitaaiso /sintaa?so 'put ornamental trappings on a horse', matag / mantag 'vomit'. Again here, the phonetically simplex [t] clearly has the phonological value of a geminate. Finally, the autobenefactive paradigm of verbs whose stem ends in -ad adduces striking evidence for the geminated character of intervocalic [t]. The second person singular forms of gado 'buy', qodo 'dig' and quuöo 'eat, feed on' are [gaöataa], [qoöataa] and [quuöataa], respectively. The presence of the boldfaced [-a-] does not come as a surprise anymore since the [t] on its righthand side is an underyling geminate (/STEM + (a)t (autob.)+1 (2SG) + aa/). When looking at the corresponding first person singular forms, a geminate [dd] could be expected. Indeed, the stem-final /-d/ together with the autobenefactive suffix /-(a)t-/ (which occurs in zero grade because the personal marker is zero) form an underlying /-d-0t-/ cluster, which could assimilate to /-dd-/ = [-dd-]. The actual result, however, is a simplex [-t-], as shown under (14). (14) LSG 2SG

stem /...Vt /...Vt

AUTOBEN.

PERS.

HAB. PRES.

-(a)t -(a)t

-0 -t

-aa/ -aa/

[gataa, qotaa, quutaa] [gaöataa, qoöataa, quuöataa]

This realisation could be the result of a regressive assimilation /dt/ /tt/ = [t]. Or rather, the stems could be underlyingly voiceless and

Abstractness in phonology: the case of virtual geminates

327

undergo intervocalic voicing in quotation forms /gat-o/ = [gaöo]. In this case, the underlying cluster in first person singular forms is /tt/ = [t] without assimilation. Whatever the correct analysis, under (14), in yet another morphological environment, we encounter one more case of simplex [t]s that necessarily enjoy a geminate phonological identity. In sum, a very easy pattern, that learners may discover without any pain, holds true for all instances of dental stops in Somali: any intervocalic [t] is an underlying geminate /tt/, while all voiced interdental fricatives [Ö] represent an underlying simplex /t/ or /d/.

3.5. Velar virtual geminates The demonstration so far concerned only dentals. However, other segments behave alike in Somali. We shall examine the case of the velar stop hereafter in some more detail, place restrictions do not allow for the discussion of other segments.14 Underlyingly simplex ikJ is voiced and spirantised in intervocalic position just as /t/. This is evidenced by the masculine definite article -ka: /dab-ka/ = [dabka] 'the fire', /beer-ka/ = 'the liver' vs. /guri-ka/ = [guriya] 'the house', /qoraa-ka/ = [qoraaya] 'the writer'. In the same way, clusters of two preverbal prepositions (/u+ka/ = [uya] 'to+from') as well as sequences of object pronouns and prepositions (/i+ku/ = [iyu] 'me+in') yield the same result. There are, however, intervocalic [k]s in Somali. Their behaviour is strictly parallel to the one that has been reviewed for [t], and a geminate underlying identity /kk/ must be concluded on. Just as [t], [k] is in free variation with homorganic nasal-stop clusters, e.g., sokor / sogkor 'sugar', maakaal / maarjkaal 'residue left after the pressing of sesame'. The geminated character of intervocalic [k] is also witnessed by the same apparently deviant behaviour with respect to vowel-zero alternations. For instance, first conjugation C V j C V j C verbs with medial [k] behave like those with medial [t], i.e. their second vowel never alternates with zero. Examples are lSG [fekeraa], *[fekraa]

328

Philippe Segerai and Tobias Scheer

from feker 'think', lSG [makalaa], *[maklaa] from makal 'put in danger', lSG [hukumaa], *[hukmaa] from hukun 'govern, rule'. In these forms, the simplex [k] has the same effect as a consonant cluster in that it prohibits absence of the second stem vowel (again, if the stem-final consonant were a geminate, it would surface as such). Finally, stems that end in [-Vk] should never tolerate the autobenefactive suffix /-(a)t-/ to occur in zero-grade if [k] is an underlying geminate. In first person singular forms, the morphological structure is /...Vkk-(a)t-0-aa/. The corresponding verbs confirm this prediction: buko 'become sick', c[ako 'hide oneself and tuko 'pray' seem to display the same structure as fir ο 'comb (autoben)', but in fact pattern with kabbo 'sip', cf. lSG [tukaöaa] from tuko as [kabbaöaa] from kabbo, against [firtaa] from firto. The existence of a voicing process in intervocalic position in Somali, as well as the perfect regularity of vowel-zero alternations allow phonological geminates /tt, kk/ to enjoy a simplex phonetic interpretation.15 The recoverability of the geminate status of intervocalic [t, k] is guaranteed by these external clues. As before in Cologne German, learners may unambiguously reconstruct the geminate whose identity is not betrayed by its length, but by the "irregular" behaviour of its environment. A parallel situation obtains for voiced stops. Underlying voiced geminates /bb, dd, gg/ are usually realised as such [bb, dd, gg]. Optionally, however, a Somali speaker may pronounce simplex [b, d, g] only. This possibility is due to that fact that underlying simplex voiced stops, as was mentioned earlier, are spirantised in intervocalic position: /b, d, g/ surface as [β, ö, γ], respectively. The nonspirantisation of a simplex phonetic voiced stop in intervocalic position therefore is sufficient in order to unambiguously betray its geminate identity. The difference between geminate and nongeminate voiced stops is that the latter always spirantise, while the former never do.

Abstractness in phonology: the case of virtual geminates

329

4. Conclusion In the preceding pages, we have argued that the phonology of two particular languages, Cologne German and Somali, may not be understood unless the existence of underlying geminates that have a simplex phonetic value is admitted. The geminate identity of these segments is contained in the acoustic signal. But instead of being encoded in the length or melody of the segment itself, it may be read off its environment. As stated in the introduction, the need of assuming virtual geminates seems to be recurrent in various analyses of different languages. Moreover, the exact vocalic counterpart, virtual long vowels, have also been identified (Lowenstamm 1991). The discussion on abstractness in the 70s, and namely the debate concerning abstract segments, relates in the following way to the issues discussed in this paper. According to current autosegmental representations, geminates and long vowels are melodically identical to their short counterparts. The only difference is the number of timing-units this melodic content is attached to. It is interesting to note that virtual geminates and virtual long vowels are virtual precisely because of the relation that two autosegments (the segmental and the timing tier) contract, and because of nothing else. Abstract segments in the sense of the early generative debate thus fall into two different species if autosegmental representations are assumed: segments that are abstract because 1) their melodic identity or a part of it does never occur on the surface, or 2) their association to another autosegment has not the usual phonetic consequences. By the time the discussion on abstractness took place, autosegmental representations were not part of phonological theory, or were just emerging. Hence, the formal distinction between different abstract segments could not be made, and indeed is absent from the literature. We would like to suggest that the majority, if not all cases where analyses recurring to abstract segments are compelling involve an association between two autosegments which is not phonetically executed.

330

Philippe Segerai and Tobias Scheer

In epistemological terms, this analysis is cumulative in the sense of Encreve (1997) and Durand and Laks (1996). When new theories arise and science moves on, open questions that are related to and were debated in the old framework are often left behind. This would not be harmful if people expected the new theory to provide tools that allow to view the unresolved questions in a new light after a while. New theories are cumulative if they do not forget about old problems that still beg the question. They are non-cumulative if they ignore old insights and questions that have been raised prior to their birth and with no relation to them. Unfortunately, generative phonology in the past 30 years bears quite a bit of a non-cumulative character. Rule-based theories such as SPE were mainly concerned with procedural issues, i.e. rule-ordering. They had their typical set of data, which was different from the one that gave rise to autosegmental representations. These emerged through the analysis of tone and syllabic phenomena. In the 80s, mainstream phonology investigated the internal structure of segments, leaving behind the questions that were of interest before. The relevant data were e.g., vowel harmony and all kinds of adjacent and distant assimilation that could reveal natural classes of sounds. In a subsequent step, syntagmatic relations among segments such as Licensing were focused on, leading to more restrictive views of syllable structure. Relevant data involved for example lenition. And finally, the 90s are back to procedural issues and the general organisation of grammar (Optimality Theory, Declarative Phonology). It is true that part of the moves are of cumulative nature, cf. the recent debate on counter-bleeding within OT. Most frequently, however, new theories do not consider old questions and old data, cf. for instance the absence of issues related to phonological representations in OT. Our contribution to the old question of abstractness is of cumulative nature. It takes advantage on a formalism that did not or poorly exist at the time abstractness was discussed (autosegmental representations) in order to propose a generalisation that may make the old question appear in a different light: abstract segments are abstract because of an association between two autosegments that is not phonetically executed.

Abstractness in phonology: the case of virtual geminates

331

Finally, the old problem of abstractness referring to the dissimilarity that grammar ought to allow between underlying and surface representations is mirrored in more recent discussion by the so-called family of faithfulness constraints that are designed to minimise the contrast between deep and surface structure. Violation of faithfulness is viewed as a consequence of the action of another set of constraints. In principle, the balance of both "forces" is not predictable, it may go way to one side, as well as far on the other. In this paper, we have intended to show how this problem may be handled in a principled way. Instead of allowing for a principle to be violated in undetermined and a priori unlimited fashion, we submit that the area in which it is contradicted is well defined and predictable. Circumscribing the violation in place, nature and function allows to capture the phenomenon by setting a parameter, rather than by dispensing with the principle, or allowing for its violation. Moreover, it adds a generalisation to the grammar that does not emerge when the violation of principles is thought of as the unmarked case.

Notes 1. The discussion on abstractness in synchronic grammar has a parallel in diachronics. Reconstructed forms that indo-europeanists arrived at by the end of the last century as displayed in Brugmann and Delbrück (1886-1990) did not take into account the fact that indo-european was a real language pronounced by real speakers. Highly abstract constructions bearing impressive consonant clusters remained unquestioned. In later times, these were cast doubt on as for instance by Benveniste (1935:Iff). More recently, typological considerations of living languages have substantially transformed the classic Brugmannian picture through Gamkrelidze and Ivanov's (1995) glottalic theory. However, Saussure's (1879) Laryngeal theory has become a consensual grounding of indo-european reconstruction. Laryngeals are abstract segments in the sense that their existence was inferred by Saussure on the basis of an empirical record they were absent from. In this view, the relation holding between the two levels of synchronic analysis (surface - underlying) has the same status as the one relating earlier and subsequent forms in diachronic analysis. In both cases, the existence of abstract segments may be evidenced by advances in

332

Philippe Segerai and Tobias Scheer

"technical" sophistication: Anatolian brought Laryngeals to light, and the reality of underlying abstract segments may be expected to appear when we will be able to identify neuronal structures. 2. As in Standard German and a number of English varieties, Cologne Ajg/ is realised [η], explaining the absence of the alleged /g/ in CG [Kyqa, bigs, pygsl]. The underlying identity of [q] is established by the same arguments as for New High German and English such as, for instance, the absence of wordinitial [η]. In addition, it may be observed on the surface when devoiced by a following voiceless obstruent as e.g., in the second person singular: /biqg-s/ = [biqks]. 3. The exhaustive list is available in Scheer (forthc.). 4. See Hönig (1877: 15) for more illustration. 5. The behaviour of dental clusters following long vowels may not be tested since configurations such as [VVNC] do not exist in MHG. The few apparent counter-examples that may be come by in New High German are the result of recent evolutions: Mond 'moon' [moond] < Monat 'month' [moonat], fahnden [faandsn] is an intensive o f f i n d e n 'find' and was short until MHG: vanden. 6. See Wrede (1958, 1: 265) and Hönig (1877: 17 ff.) on this issue. 7. There are two other sources of [g]. Loans from Low Germanic such as NHG Roggen [KOgsn] 'rye' that show a geminated -gg- in German spelling appear in Cologne German with a simplex [g], i.e. CG Rögge [bfcegs]. Words that contain a -ck- in German spelling, which is the graphic indication of its former geminated character, may also be represented by [g] in Cologne German. In NHG wackeln, einnicken, Rücken [vaksln, ajniksn, Kyksn] = CG waggele, nügge, Rögge [vagste, nyga, lioegs] 'be shaky, nod off, back (part of the body)', an underlying /k/, or even /kk/, is voiced intervocalically. In any event, no CG [g] is a manifestation of /g/. 8. As opposed to "strong" verbs whose personal markers are prefixed. The five verbs at stake areyilil 'be in a place', yimi(d) 'come', yiqiin 'know', yidhi 'say' and yahay 'be'. 9. Except for imperatives and reduced forms. 10. The stem-internal vowel-zero alternation that is observed (daqn- / daqan-) will be discussed in a subsequent section. 11. In the latter example, an independent process controls the stem-final o-a alternation, cf. Saeed(1998: 26). 12. Somali verbs are traditionally given in their quotation form, that is the imperative 2SG. For autobenefactive verbs, IMP 2SG is [stem-o]. Strangely enough, this is the only form in the entire paradigm that does not show the autobenefactive suffix.

Abstractness in phonology: the case of virtual geminates

333

13. The right margin of the word behaves like a consonant: alternation-sites followed by a word-final consonant are vocalised, e.g., imperative 2SG /kidf/ = [kidif], *[kidf] 'shop!'. 14. Parallel behaviour is observed for [tj, J", q, w, j] (there is no [p] in Somali). See Barillot and Segeral (forthc. a, b) for a complete record of the facts. 15. A similar case can be made for Modern Hebrew. Berman (1997: 315) points out that the otherwise regular syncope of non-high vowels in nouns and adjectives before stressed suffixes "is blocked by historical medial geminates". Compare for instance SG [matos] PL [mtos-im] 'air-plane' vs. SG [ganav] - PL [ganav-im] 'thief. Clearly, the "historical" geminate in [ganav-im] < Tiberian Hebrew [gannabiim] is still present in the underlying form of modern Hebrew. It is not phonetically expressed since Modern Hebrew does not geminate anymore, but the resistance to the otherwise regular (and synchronically active) syncope guarantees its underlying presence, and makes it recoverable. If no virtual geminate is assumed here, there is no way of viewing the Modern Hebrew syncope as a regular phonological process.

References Barillot, Xavier and Philippe Segeral forthc. a On phonological processes in the '3rd conjugation' in Somali. In: Denise Perett (ed.), Afro-Asiatic Grammar 3. forthc. b La syncope en somali. Linguistique Africaine. Bendjaballah, Sabrina 1999 Trois figures de la structure interne des gabarits. Ph.D. dissertation, Universite Paris 7. Benveniste, Emile 1935 Origine de la formation des noms en indo-europeen. Paris: Maisonneuve. Berman, Ruth 1997 Modern Hebrew. In: Robert Hetzron (ed.), The Semitic Languages, 312-333. London: Routledge. Bertram, Otto 1935 Der Wandel nd zu ng am Oberrhein. Zeitschrift fur Mundartforschung 11: 6-12. Brugmann, Karl and Berthold Delbrück 1886-1900 Grundriss der vergleichenden Grammatik der indogermanischen Sprachen. Straßburg: Trübner.

334

Philippe Segerai and Tobias Scheer

Campbell, Lyle 1981 Generative phonology vs. Finnish phonology: retrospect and prospect. In: Didier Goyvaerts (ed.), Phonology in the 80's, 147-182. Ghent: Story-Scientia. Dell, Frangois 1973 Les regies et les sons. Paris: Hermann. Dinnsen, Daniel 1980 Phonological rules and phonetic explanation. Journal of Linguistics 16: 171-191. Dressier, Wolfgang 1974 Diachronic puzzles for Natural Phonology. In: Anthony Bruck, Robert A. Fox and Michael W. La Galy (eds.), Papers from the parasession on Natural Phonology, 95-102. Chicago: Chicago Linguistic Society. Durand, Jacques and Bernard Laks 1996 Why phonology is one. In: Jacques Durand and Bernard Laks (eds.), Current Trends in Phonology: Models and Methods, 3 13. Salford: ESRI. Encreve, Pierre 1997 L'ancien et le nouveau. Quelques remarques sur la phonologie et son histoire. Langages 125: 100-123. Frings, Theodor 1955 Vom g, von seinen Lautwerten und von germanischen Sprachlandschaften. Rheinische Vierteljahresblätter 20: 170-191. Frings, Theodor and L.E. Schmitt 1942 Gutturalisierung. Zeitschrift für Mundartforschung 18: 49-5 8. Gamkrelidze, Thomas and Vjaceslav Ivanov 1995 Indo-European and Indo-Europeans. Berlin/New York: Mouton de Gruyter. Goyvaerts, Didier 1981 Introduction. In: Didier Goyvaerts (ed.), Phonology in the 80's, 1 26. Ghent: Story-Scientia. Hellberg, Stefan 1978 Unnatural phonology. Journal of Linguistics 14: 157-177. Hönig, Fritz 1877 Wörterbuch der Kölner Mundart. Köln: Bachem. [Reprinted, Köln 1952.] Hooper, Joan 1976 An Introduction to Natural Generative Phonology. New York: Academic Press. Keenadiid, Ciismaan Yaasin 1976 Qaamuuska af-Soomaaliga. Florence: Le Monnier.

Abstractness in phonology: the case of virtual geminates

335

Kenstowicz, Michael and Charles Kisseberth 1977 Topics in Phonological Theory. New York: Academic Press. 1979 Generative Phonology. Description and Theory. San Diego: Academic Press. Kiparsky, Paul 1968 How abstract is phonology? Bloomington, Ind.: Indiana University Linguistics Club. 1974 On the Evaluation Measure. In: Anthony Bruck, Robert A. Fox and Michael W. La Galy (eds.), Papers from the parasession on Natural Phonology, 328-337. Chicago: Chicago Linguistic Society. 1982 Explanation in Phonology. Dordrecht: Foris. Koutsoudas, Andreas 1980 The question of rule-ordering: some common fallacies. Journal of Linguistics 16: 19-35. Larsen, Bergeton 1994 Some aspects of vowel length and sted in Modern Danish. MA thesis, Universite Paris 7. Lessiak, Primus 1933 Beiträge zur Geschichte des deutschen Konsonantismus. Brünn/Prag/Leipzig/Wien: Rohrer. Lightner, Theodore 1981 New explorations into Derivational Morphology. In: Didier Goyvaerts (ed.), Phonology in the 80's, 93-99. Ghent: Story-Scientia. Lowenstamm, Jean 1991 Vocalic length and syllable structure in Semitic. In: Alan S. Kaye (ed.), Semitic Studies in Honor of Wolf Leslau on the Occasion of his 85th Birthday, 949-965. Wiesbaden: Harrassowitz. 1996 CV as the only syllable type. In: Jacques Durand and Bernard Laks (eds.), Current Trends in Phonology: Models and Methods, 419-441. Salford: ESRI. Müller, Josef 1942 Einige Bemerkungen zur rheinischen Gutturalisierung, bei der Wörterbucharbeit aufgelesen. Zeitschrift für Mundartforschung 18: 58-59. Münch, Ferdinand 1904 Grammatik der ripuarisch-fränkischen Mundart. Bonn: Cohen. Orwin, Martin 1995 Colloquial Somali. A Complete Language Course. New York/London: Routledge. Paul, Hermann, Peter Wiehl and Siegfried Grosse 1989 Mittelhochdeutsche Grammatik. (23rd edition.) Tübingen: Niemeyer.

336

Philippe Segerai and Tobias Scheer

Saeed, John Ibrahim 1998 Somali. Dublin: John Benjamins. Saussure, Ferdinand de 1879 Memoire sur le systeme primitif des voyelles dans les langues indoeuropeennes. Leipzig: Teubner. Schane, Sanford 1968 French Phonology and Morphology. Cambridge, Mass.: MIT Press. 1974 How abstract is abstract? In: Anthony Bruck, Robert A. Fox and Michael W. La Galy (eds.), Papers from the parasession on Natural Phonology, 297-317. Chicago: Chicago Linguistic Society. Scheer, Tobias 1997 Vowel-zero alternations and their support for a theory of consonantal interaction. In: Pier Marco Bertinetto, Livio Gaeta, Georgi Jetchev and David Michaels (eds.), Certamen Phonologicum III, 67-88. Torino: Rosenberg and Sellier. 1999 Aspects de l'alternance schwa-zero ä la lumiere de "CVCV". Recherches Linguistiques de Vincennes 28: 87-114. forthc. Von kölscher Gutturalisierung und verborgener Doppelkonsonanz. Ms., University of Nice. Segeral, Philippe 1995 line theorie generalisee de l'apophonie. Ph.D. dissertation, Universite Paris 7. 1996 L'apophonie en Ge'ez. In: Jacqueline Lecarme, Ur Shlonsky and Jean Lowenstamm (eds.), Studies in Afro-Asiatic Grammar, 360391. The Hague: Holland Academic Graphics. Selkirk, Elisabeth and Jean-Roger Vergnaud 1973 How abstract is French phonology? Foundations of Language 10: 249-254. Skousen, Royal 1981 Analogical sources of abstractness. In: Didier Goyvaerts (ed.), Phonology in the 80's, 55-92. Ghent: Story-Scientia. Stampe, David 1973 A dissertation on Natural Phonology. Ph.D. dissertation, University of Chicago. Tranel, Bernard 1981 Concreteness in Generative Phonology. Evidence from French. Berkeley, Ca.: University of California Press. Vennemann, Theo 1974a Phonological concreteness in natural generative grammar. In: Roger W. Shuy and Charles Bailey (eds.), Toward Tomorrow's Linguistics, 202-219. Washington, DC: Georgetown University Press.

Abstractness in phonology: the case of virtual geminates 1974b

337

Words and syllables in natural generative grammar. In: Anthony Bruck, Robert A. Fox and Michael W. La Galy (eds.), Papers from the parasession on Natural Phonology, 346-374. Chicago: Chicago Linguistic Society.

Wrede, Adam 1958 Neuer Kölnischer Sprachschatz. 3 Vols. Köln: Greven.

Constraints, preferences, and context-sensitivity in morphology* Rajendra Singh

For all the relatives, near and distant, of Leopald Kupelwieser

(1796-1862)

No one ever dreamed of doing universal morphology. Otto Jespersen (1924), cited repeatedly in Chomsky (1996) One might view Optimality Theory as the band-aids getting together, realizing their own power, and deciding that they could get along quite nicely without the patient. Kevin Russell (1997: 110)

1. Introduction Whereas both constraints and preferences seem reasonably wellmotivated on both physiological and distributional grounds in phonology, it is not entirely clear if morphology needs anything other than context sensitivity. In this paper, I shall argue that once contextsensitivity is appropriately taken into account, the need to invoke constraints, of whatever sort, disappears from morphology. I shall in fact argue that both constraint-based accounts of morphology, such as Optimality Morphology, and constraint-rich accounts, such as Lexical Morphology, are inadequate, and that it is so for the reason behind Jespersen's (1924: 52) observation, repeatedly cited by Chomsky (though routinely ignored by his followers) that nobody ever dreamt of doing universal morphology. Aware of and impressed by the variety of forms and functions in morphological systems, Jespersen believed that it was not possible to construct any very tightly restricted system of universal morphology. This paper will in fact show that he was fundamentally right.

340

Rajendra Singh

2. On constraint-based and constraint-rich accounts Perhaps I should begin in reverse chronological order, that is, with Optimality Theory (OT). Just to get a taste of OT morphology (OTM), try (1) below as a statement of what OT-morphologists say can and must be said by OT-morphology in order not to say that the pluralizing suffix -z of English is a suffix: (1)

= KR4. 38 PL-AFTER-N: ALIGN (Plural, left; Noun-stem, Right)

Or, consider the rather unusual intuition of Russell's (1997: 124) that "both suffixes the pluralizing -z and the possessive -z are somehow present in the surface form" of words like cats'. Obviously, what he means is that /kaetz/ can be interpreted as the possessive of the plural cats, provided, of course, it appears in the appropriate syntactic position. The only intuition speakers of English seem willing to swear by is that in such cases the plural is identical with the singular possessive and the plural possessive. Correspondence Theory, invoked by Russell, is, of course, a possible way to describe the intuition that speakers have, but only an OTist can, it seems to me, have the intuition Russell claims to have. The truth of the matter leaks through such revealing statements as the following from Russell (1997: 126), who must be given credit for his honesty: (2)

"To solve this problem, we might be tempted to dive straight in and create an alignment constraint analogous to PLAFTER-N"

The constraint he creates is (3)

POSS-AFTERN

(2), he tells us, is "not far from correct" (1997: 127). It needs to be revised a wee bit to get the right one, which is given in (4):

Constraints, preferences, and context-sensitivity

(4)

341

POSS-AFTER-STEM.

He also gives the game away when he "simply assumes" that the cross-morphemic restriction in Paamese that prohibits sequences with i and u on opposite sides of a morpheme boundary is "a language particular constraint against the undesired situation" (Russell 1997: 111). Presumably the number of such constraints is equal to the number of cross-morphemic constraints used in the languages of the world. And then there are, of course, morpheme particular constraints (Russell 1997: 121-130): (5)

ALIGN {-er Left; Trochee, Right) English

(6)

If nauta, then first declension Latin

The story does not really quite stop here. Although enthusiastic about restating blocking in OT terms, when faced with coexistent forms, Russell is seriously willing to consider the possibility of entertaining unranked constraints (1997: 132), and Pesetsky (1997: 160) is, of course, not alone in entertaining constraints that are tied (cf. Archangeli and Langendoen 1997: 212, for instance). These examples, to borrow from Russell, illustrate "how powerful OT can be" (1997: 127). This much power, some of us might still remember from our student days, used to be seen as the opposite of "interesting" and "restricted". An unconstrained constraint-based theory has the ring of a paradox, if not of irony. And there is, of course, our old friend "explanation". Russell claims that Bybee's (1985) universal ordering of verbal affixes of tense, aspect, and mood can be explained by alignment constraints. As I see matters in morphology, the problem is not that such creation is sometimes necessary to describe the facts of the matter but that there is really nothing else except such creations in OTmorphology, again not because OT-morphology is wrong but because there are no possible anchors available for such ad hoc constraints in morphology - all categorization theories are likely to meet their Waterloo in morphology.

342

Rajendra Singh

But perhaps I ought to look farther than Russell's essentially pedagogical presentation. I shall proceed to do so, but must first mention, without dwelling on them, what seem to me to be the obvious problems associated with each of the following: (i)

(ii) (iii)

This sort of model is proud of the fact that the sort of theory its advocates were trained in is unable to provide a satisfying account of the fact that a reduplicant has an important and continuing relationship with its base, and that they have the answer, faithfulness, modulated, of course, by parametric prosodic choices. It is, unfortunately, quite insensitive to the unfaithfulness of reduplication seen in dozens of South Asian and Southeast Asian languages, whose ample reduplication patterns not only violate the violable faithfulness but insist on unfaithfulness (cf. Vijaykrishnan 1998). Bits of morphology associated with language-particular categories are necessarily non-universal. Then there is the family of problems of the sort illustrated by IruLa (cf. Wiltshire 1994) and Mohawk (cf. Potter 1994). These involve re-rankings at the drop of a morphophonemic hat. IruLa presents a challenge to the view that when the same suffix is added to stems of the same phonological shape outputs cannot be different, and the facts of Mohawk lead Potter to not only accept distinct levels of representation, each requiring a different ranking, but also to propose that Parse family of constraints is sensitive to the semantic content of morphological units.

I should perhaps leave the contradictions and problems associated with formulations and observations of the sort illustrated above (in i to iii) alone and concentrate on how what OTM is trying to say is said better. I shall return to such a demonstration, but first I want to take up what I am calling constraint-rich morphology, of which LM is a good example. As OT-morphology and LM share what they see as the serious problem of order, I shall concentrate on it, unfortunately neglecting other things. One of the basic principles of LM is a

Constraints, preferences, and context-sensitivity

343

contemporary paraphrase of the old Paninian principle of krt before taddhita. The idea is that the non-occurrence of certain sequences of affixes, attested in all languages, should be exploited to the hilt and promoted as a general constraint. The non-existence of words terminating in *ness-ic in English, for example, is seen as a warrant for saying that -ic is krt or level 1 suffix and -ness is a taddhita or level 2 affix and that the combination is bad because 2 must follow 1. Counter-examples abound in the literature, and various ad hoc moves have been suggested or actually made to take care of cases where the suggested ordering rules out combinations, such as gandhistic, nationalistic, developmental, transformational grammarian, and ungrammaticality in English, that are abundantly attested, and I shall not dwell upon them. What I would like to draw your attention to is what I shall refer to as Fabb's Problem. Fabb (1988) concentrates on combinations that are neither attested nor ruled out by contemporary analogues of the krt-taddhita principle. He demonstrates that out of a total of 1849 pairs of suffixes that could be obtained from combining 43 suffixes that attach to words in English, only 1186 can be eliminated on the basis of parts of speech selection and only another 49 can be eliminated on the basis of what he calls local restrictions on suffixation, such as -ful being attachable only to nouns with final stress. This leaves a total of 614, from which another 155 can be eliminated on the basis of level-ordering of the sort LM-morphologists have proposed, leaving us with 459 pairs, as against the 50 pairs actually attested in English. The remaining unattested 409 pairs cannot be ruled out by any known version of level-ordering. Fabb (1988: 530) concludes: "It can be seen that level ordering achieves relatively little in predicting which suffix pairs exist and which do not." Notice that Fabb's problem will disappear if it turns out that morphological operations in English are such that they in fact will never have the occasion to produce things Fabb and those he is critiquing are interested in ruling out. A proper formulation of morphological processes, as we shall see, almost automatically yields that result.

344

Rajendra Singh

3. Preferences, typology, and context-sensitivity Hopefully having convinced you that constraints can neither serve the role of being the material which morphologies can plausibly be built with nor can they restrict matters morphological, I shall now attempt to show that (a) word-formation is really a very local matter, which is best accounted for with the help of Word Formation Strategies (WFS) that do not appeal to anything other than the variables and constants they can be said to contain, and (b) this sort of formulation, fortunately, allows one to a) state the valid part of OTM, b) resolve Fabb's Problem, and c) the meta-generalizations that Vennemann (1983) and Dressier (1987, 1988, and 1999) refer to as preferences and rightly see as functioning as intermediate-level elucidatory principles. Begging to differ from the prevalent, quintessentially Paninian view of morphology, which sees morphology as a combinatorics of units smaller than the word, involving word-internal syntax in some versions, and allows morphological operations on those units, Ford and I have argued that all that needs to be said about word structure in any language (of any type whatsoever) can and must be said by instantiations of the schema in (7) below. We refer to these instantiations as W(ord) F(ormation) S(trategies) because as generalizations drawn from known particular facts, they can be activated in the production and understanding of new words (cf. Ford and Singh 1991 and Ford, Singh, and Martohardjono 1997): WFS's must be formulated as generally as possible, but, and this is crucial, only as generally as the facts of the matter permit. (7)

/Χ/α

/Χ'/β

where: a.

/X/ a and /Χ'/β are words and X and X' are abbreviations of the forms of classes of words belonging to categories α and β (with which specific words belonging to the right category can be unified or on to which they can be mapped).

Constraints, preferences, and context-sensitivity

b. c. d. e. f.

345

' represents (all the) form-related differences between /X/ and /X'/ α and β are categories that may be represented as featurebundles the represents a bidirectional implication (if X, then X' and if X', then X) X' is a semantic function of X ' can be null iff α Φ β

It should be obvious that we do not think that either intralinguistic (inflections vs. derivation, affixation vs. compounding etc.) or inter-linguistic (flectional, isolating etc.) morphological diversity can affect (7) in any fundamental way. It allows for, albeit at a slightly abstract level, a unified account of what have sometimes been seen as different types of morphologies. The diversity that exists can be, hopefully, read off the system of strategies that instantiate (7) above, but it does not need to be expressed as a difference in type: a difference in content does not, in our view, necessarily constitute a difference in form (of rules or strategies). (7) also denies any theoretical status to descriptive labels such as "concatenative", "nonconcatenative", "affixal", "non-affixal" etc. Again, multiplicity is there but resides in descriptive paraphrases of instantiations of (7). As all morphological relationships can be expressed by strategies instantiating (7), morphology has little or no architecture and, to change the metaphor, no non-negotiably binding traffic rules (such as krt before taddhita). Although there may well be constraints on what sorts of things or concepts can be morphologized (= constitute categories relevant for a morphological description), undoubtedly related to limitations on conceptualization. There seem to be no constraints on particular instances of (7), though all manifestations of (7) must, quite obviously, relate (single) words with (single) words.1 Representations of the speaker's knowledge of the patterns of morphological relatedness in her language, Morphological Strategies (= instantiations of (7)) are invoked only in moments of crisis, i.e., when the speaker needs to analyze or fashion a word she needs for the purpose at hand, often to meet a syntactically enforced require-

346

Rajendra Singh

ment. Their exploitation, of course, helps her to bridge the gap between the actual words she happens to know and the possible words she can be said to know - actually their existence makes the known merely a subset of the knowable. When they are invoked to produce what will become words, their "outputs" are seamless wholes, with no brackets, boundaries, or a-cyclic graph fragments in them. They are not there to be deleted; they are just not there. WFS's cannot supply these things because they do not have them. And neither the strategies nor their "outputs" have any syntactic constituency relationships marked in them in any fashion whatsoever. In both the active and the passive mode, they license the words a speaker has or may come up with (in the "on line" mode).

3.1. Some examples Below, I provide some examples of morphological strategies from a variety of languages (the parenthetical comment draws attention to what some would like attention drawn to). When not enclosed in phonemic bars, words from languages other than English are given in their standard transliterated form (in which D stands for retroflex d) and are provided with glosses. Although more than the minimum number of words required for the postulation of a WFS (= two pairs of words exhibiting the same difference) is available in each case, we provide only one pair in each case to save space. The category specifications below are also minimal (but sufficient to meet the needs at hand). (8)

English a. /X/N Marx b. /Xik/ADJ critic

/Xizm/,Ν Marxism /XisizmAΝ criticism

Constraints, preferences, and context-sensitivity

347

(The /s/ in criticism is a concomitant consequence and an integral part of the morphological operation that can be used to form nouns from adjectives terminating in /ik/, an operation or rule that is in competition with the general rule in (8a) above. It is, therefore, part of the representation of the word criticism, to which the phonology of English (= the phonological processes of English) will apply, as it must. Needless to add that the Paninian view of phonology sees the /s/ in criticism as "phonologically derived" from a fkl despite the fact that the allegedly phonological part of the operation in question is not generalizable beyond the morphological categories with which it is bound and in which it actually shows up.) (9)

Khasi /X/N step 'rising sun'

/myn X/N mynstep 'morning'

(There is, apparently, some debate about myn-words being affixed words or compounds (cf. Philip 1997), a debate which presupposes that the distinction is a viable one! (cf. Tirumalesh 1997).) (10)

Kashmiri a. /X/v

thag 'to cheat'

/X/N thag 'a cheat'

(The agent noun is generally said to be derived from the verb with the help of the famous zero-suffix, despite the fact that generally there is no evidence (cf. Sanders 1984) to support the putative deverbal character of the agent noun (for Kashmiri, cf. Wali and Koul (1997: 270): "The syntax of deverbal nouns is similar to that of nonderived nouns with respect to gender, number, and case-marking.") and despite diachronic derivations that follow the opposite path.)

348

Rajendra Singh

(10)

b.

/XVC/v.,NT.

mar 'to die'

/ Χ V : C/v. caus.

ma:r 'to kill'

(This vriddhi alternation is treated by some scholars as a part of Kashmiri phonology despite the fact that inter-consonantal vocalic lengthening is never required in Kashmiri and is in fact not only associated with causativization but is the only mark of it in words like /ma:r/ 'to beat', /ga:l/ 'to melt' (causative), /da:l/ 'to remove'.) (11)

Bangla (a. k. a. Bengali) /X/ v likhe write

/Xnewa:/V likhenewa: to first write down

(/newa:/ is one of the handful of "supporting verbs" or vectors that appear in structures traditionally described as "compound verbs" despite the fact that the freedom that label implies would generate far more combinations of verbs with verbs than are actually treated as "compound verbs" even by those who use this label. The grammatical subservience of elements like /newa:/ in such structures is demonstrated with remarkable clarity and elegance in Dasgupta (1989: 215-222), in particular.) (12)

Hindi a.

/Xa/n. NOM.SG.MASC.

/ΧΪ/Ν. SG.FEM.

laDaka: 'boy'

laDaki 'girl'

(As masculine nouns ending in /a/ have a straightforward feminine correspondent in /i/, there is absolutely no need to postulate an intermediate /laDak/, an entity whose postulation is forced by the rather peculiar architecture of Paninian morphology. The generalization speakers seem to use is (12a); cf. Singh and Agnihotri 1997.)

Constraints, preferences, and context-sensitivity

(12)

b. /Χ/η sava:r 'rider'

^

349

/ghuDX/n ghuDsava:r 'horse-rider'

(/ghuD/ can, obviously, be called the combining form of the word /gho:Da:/ 'horse' in this vikari "compound", but it is not clear what is gained by doing so, particularly because the form that combines is not always the combining form (see 13b below). It is hard to see why Paninians do not call it a "prefix"). (12)

c.

/X/N ghoDa: 'horse'

/Xga:Di/N ghoDa:ga:Di 'horse-carriage'

(Words like /gho:Da:ga:Di/ are typically presented as made up of two words despite the fact that whereas /gho:Da:/ is very freely commutable, /ga:Di/ is hard or impossible to find a substitute for (cf. Singh and Dasgupta 1999). It is quite clear that despite the fact that there is no vikaar in /ga:Di/, it has lost its word-hood - it only sounds like the word /ga:Di/, as Bhartrihari would have said.) (13)

Sanskrit a.

/Xa/V, IMP.2SG. bhava 'you be'

/Xa:mi/V, PRES.ISG. bhava:mi Ί am'

(Notice that there is no need to go through the famous "root" /bhu/ to capture this part of what is involved here. Nor is there any need to appeal to this fictitious construct to capture other such relationships.) (13)

b.

/ X / N , DAT.PL.

/ X / N , ABL.PL.

ka:mebhyas 'love'

ka:mebhyas 'love'

350

Rajendra Singh

(That this straightforward partial syncretism must, under some lights, be stated in a meta-grammar of Sanskrit or made to follow from some Paradigm Structure Condition(s) is surely a reflection on those lights.)

3.2. A recapitulation Even at the risk of being redundant, I wish to underline the fact that the WFS's presented in section 3.1. above do not appeal to or use any Paninian construct such as dhatu, anga, vibhakti, pratyaya etc. Nor do they use concepts such as inflection, derivation, and compounding etc. Yet these strategies seem to be able to say exactly what needs to be said about the bits of morphology they describe, and some of what they do say is hard, if not impossible, to say in Paninian terms. Whereas Paninian morphology sees what could be called morphological complexity as a matter of layers of morphological structure, our strategies invite one to think of words it would call complex words as made up of variables and constants that have been non-hierarchically put together, provided, of course, there are strategies that license such analyses. Thus, both English Marxism and Hindi /gho:Da:ga:Di/ can be analyzed as made up of substrings that correspond to what is varied and what is held constant in the relevant strategies (they are, it is important to underline, identified as such ONLY in the strategies). If these strategies are in fact invoked to create these words, they will not, we want to emphasize, supply any boundaries or brackets, only seamless wholes (marxism and /gho:Da:ga:Di/) that will show up as words after phonological processes have given them the phonetic shape they must have to count as words. The need to divide non-category bearing substrings into roots and stems etc. or what is held constant and can be seen as a string into prefixes, suffixes, and infixes etc. remains a mystery to us. As for the word, it is clearly indicated by a bald, unadorned and unsupported X, whose ability to bear a category does not depend on the presence of some other supporting material or by an X and the bound material whose support it needs before it can take on the burden of

Constraints, preferences, and context-sensitivity

351

bearing a category (cf. Hindi /gho:Da:/ as an instance of the left-hand pole of (12c) above and Hindi /laDak/ as an instance of the variable in the left or right hand pole of (12a) above). And it is, of course, forever nitya. What is held constant but cannot be neatly localized as a "morpheme" is, of course, the Achilles' heel of morphemology. Although the different types of substrings in deconcatenatable representations can be easily identified (and accessed for morphological parsing, should it be required), there seems to be no need to ontologize them. Although speakers certainly have the capability of accessing even discontinously spread constants, it is not clear if there is any need to ontologize even the continuously spread ones. Access to morphology, in other words, should not be, as it often is, interpreted as access to Paninian morphology (= any theory of morphology that permits morphological operations on units smaller than the word; most such contemporary theories also assume word-internal syntax, for which, it is useful to remember, there is virtually no psycholinguistic evidence). As for morphological typology, it is perhaps only a matter of the types of X's that dominate particular morphologies. Thus it is possible to refer to a morphological system or a part of a morphological system in which only bald, unadorned X's bear categories as wordbased and to systems or subsystems characterized by the absence of such bald X's as non-word based. This naming device does not, however, require giving up the assumption that morphology relates whole words with whole words, obviously contra Panini, for whom the word is quintessential^ a derived entity (cf. Deshpande 1997). There are, in effect, no constraints here, and those proposing some must show why they think they are necessary. It should be obvious that the formalism we propose denies all of the traditional Paninian or neo-Paninian distinctions such as inflection vs. derivation, affixation vs. compounding, concatenative vs. non-concatenative etc. A morphologically complex word is simply a word part of whose phonic content can be seen as a constant in some WFS, and whatever can be seen as a constant is the footprint of the relevant morphological operation. All I need to add is that we have also shown that these WFS's must be formulated only as generally as the facts of an I-

352

Rajendra Singh

Lexicon warrant because morphological operations are locally anchored in the sets of words that actually give rise to them. Ignoring this locality produces such archi-suffixes as Plenat's (1996) -Vche, "allo-forms" of which are seen in French words like bavoche /bavosh/. An account of the appearance of this fictitious suffix requires, according to Plenat, not only the entire machinery of OT morphology but also weighing and statistical ranking of the constraints. The fact is that if the relevant operations are formulated only as generally as the facts of the matter warrant, there is no need to do anything of this kind at all - words that end in lb, d, k/ take ash, those that end in /v, j, r, n/ take osh, etc. (cf. Ford and Singh 1999). Morphological operations are very, very local affairs, but, apparently, Plenat prefers contradiction to locality. We don't, for example, say that -en is attached to English adjectives to make verbs with a causative interpretation and that that rule has phonological constraints on it but simply that for most speakers of English the verbalization of those adjectives that end in obstruents leaves -en as a footprint. Similarly, we don't say, as a Paninian would, that English words like doghouse, nut-house, monkey-house etc. are made up of two words but simply that words like dog, monkey and nut appear as substitutes for variables in morphologically complex words in which the phonemic sequence house functions as a constant. Our formulation allows us to dispense with Paninian constructs such as roots, stems, prefixes, suffixes, augments, linking morphemes, state morphemes, zero morphemes and other such objects of wonder. It is important to repeat that we have never denied speakers the analytical abilities that could lead to the segmentation of these Paninian objects. What we have argued against is granting these things any status outside of the morphological operations in which they figure, even in languages in which such Paninian segmentation is easily possible and against the inclusion of any boundaries or the imposition of any hierarchical structure on the variables and constants that appear in or constitute WFS's. The non-independence of constants is, as we have argued, attested, for example, by the fact that speakers of French haven't, don't, and are not likely ever to confuse the nominalising -ment with the adverbialising -ment - heureusage is really

Constraints, preferences, and context-sensitivity

353

not likely to become a French adverb ever. The identity of the form of the so-called affix is never a source of confusion - its source strategy always accompanies it.

3.3. On solving traffic and typological problems Notice that our approach allows us to solve the problems generated by the imposition of the krt/taddhita Constraint in Lexical Morphology. There is nothing mysterious about the ordering restrictions that actually obtain in natural languages. They follow straightforwardly either from the local nature of morphological operations or from the fact that some morphological operations are really substitution operations. If -ity must attach to -able adjectives, it is because there is a WFS like (14)

/Xable/

/Xability/

And it may well be the case that the appearance of two so-called affixes in a particular order automatically rules out their cooccurrence in the other order not because there is an ordering constraint but because the constant that is seen as a combination of two affixes really does not have any parts that can be moved around. Equally clearly, if "suffix" -abc does not co-occur with "suffix" -def in any order, it is because there is a WFS like: (15)

/Xabc/

/Xdef/

Now, we understand why only a handful of possible combinations of so-called morphemes are actually attested in natural languages: only those that are licensed by the WFS's of a language are actually observed. Morphology is not a branch of combinatorics but a study of the form-related ways in which languages relate their words with their other words, ways speakers can and do exploit to enrich their lexica. We can safely conclude that if morphological rules are sensitively formulated, they correctly capture the locality the naive

354

Rajendra Singh

speaker bestows upon them, and that such context-sensitive formulations require no conditions, constraints or preferences. Finally, gleaning the meta-generalizations that lead to preferences, unanchored chez Vennemann and anchored in a theory of semiotics chez Dressier. The latter makes a very convincing case for abandoning absolute constraints, particularly in morphology, and replacing them with preferences (in order to explain what would otherwise go, as it has sometimes gone in the past, unexplained). These preferences are grounded in extra-linguistic bases (semiotics for morphology). Dressier, thus, derives the result that although root and stem based morphology is possible, word-based morphology is preferred for semiotic reasons. I agree with him, but would like to plead that differences of this sort are in fact gleanable from the unifying formalism we have proposed for morphology. These differences can be obtained from an inspection of the sorts of variables and constants that figure in the WFS's of a language, without giving up the idea that morphology is in fact word-based. The WFS's that contain noncategory bearing variables describe what can be called non-wordbased morphological operations and the WFS's that contain only category-bearing variables at one of the two poles describe what are ordinarily called word-based operations. Compounds can also be identified relatively easily without making a distinction between compounding and affixation - if the constant is a combining form or merely isophonic with a word, the WFS that contains it can be called a compounding WFS, though the validity of the label remains primarily diachronic. A compound is, after all, best defined as a word in the morphological company of forms that used to be full-fledged words till yesterday (those that were words till yesteryear have already become "affixes"). Synchronically, there is, of course, no real distinction between compounding and affixation, even in languages like German (cf. Becker 2000). In other words, the distinctions we need for Dressler's preferences in morphology are in fact eminently available in these strategies without postulating multiple morphologies. It is important to point out that Dressier (1987, in particular) insists, rightly in my view, on the derivability and not on the direct incorporation of preferences in

Constraints, preferences, and context-sensitivity

355

grammar (1987: 72). I would like to submit that the context-sensitive but constraint-free theory of the sort I have been advocating provides descriptions from which distinctions of the sort the preference theory has and explains can be easily deduced. These preferences, it is equally important to point out, are not weaker constraints but simply entries in Wang's (1968) "price book", leading, quite naturally, to Mayerthaler (1981)'s and Dressler's extra-grammatical evaluation. And that is the way it should be because tendencies are the result of subtle interplay amongst the various sub-theories of Natural Morphology. To conclude, if morphological rules are sensitively formulated, they correctly capture the locality the naive speaker bestows upon them, and they do so without invoking any conditions or constraints. As for the differences and distinctions explained by Preference Theory, they (= the distinctions and NOT the explanations) do seem extractable from such formulations.

Notes

1

2

This is a slightly revised version of a paper presented at the Poznan workshop on constraints and preferences (Adam Mickiewicz University, Poznan, May 1 2, 1999). I am grateful to Katarzyna Dziubalska for the invitation to participate in that workshop and to Wolfgang Dressier for several useful discussions and comments. This version owes a lot to the latter. 1 am also grateful to the other participants of the Poznan workshop. The usual disclaimers apply in full force. Throughout this paper, I use the term "constraint" in its pristine sense. The notion of "violable constraint" is a theory-dependent artefact of OT. The point of saying it this way is to make it clear that so-called compounds are single words and DO NOT contain two or more words (cf. Singh and Dasgupta 1999). As for the notion of default, Dressier (1999: 390) points out that it goes back to the "circular and surface descriptivism of distributional structuralism" and computer science, and can in fact be reanalyzed in terms of preference convergence.

356

Rajendra Singh

References Archangeli, Diane and D. Terence Langendoen (eds.) 1997 Optimality Theory: An Overview. Oxford: Blackwell. Aronoff, Mark 1976 Word Formation in Generative Grammar. Cambridge, Mass.: MIT Press. Becker, Thomas 2000 On the non-hierarchical structure of compounds. In: Rajendra Singh (ed.), The Yearbook of South Asian Languages and Linguistics, 283292. New Delhi: Sage Publications. Bybee, Joan L. 1985 Morphology. Amsterdam: Benjamins. Chomsky, Noam 1996 The Minimalist Program. Cambridge, Mass.: MIT Press. Dasgupta, Probal 1989 Projective Syntax: Theory and Applications. Pune: Deccan College. Deshpande, Madhava 1997 Building blocks or useful fictions: Changing views of morphology in ancient Indian grammatical thought. In: Dick van der Meij (ed.), India and Beyond, Aspects of Literature, Meaning, Ritual and Thought: Essays in Honour of Frits Staal, 71-127. London: Kegan Paul International (in association with the Institute for South Asian Studies, Leiden). Dressier, Wolfgang U. 1987 Morphological islands. In: Ross Steele (ed.), A Festschrift for M. A. K. Halliday, 71-79. London: Longman. 1988 Preferences vs. strict universals. In: Michael Hammond and Michael Noonan (eds.), Theoretical Morphology, 143-154. San Diego: Academic Press. 1999 On a semiotic theory of preferences in languages. In: Michael Haley and Michael Shapiro (eds.), The Peirce Seminar Papers, 390-415. New York: Berghahn Books. Fabb, Nigel 1988 English suffixation is constrained only by selectional restrictions. Natural Language and Linguistic Theory 6: 527-539. Ford, Alan and Rajendra Singh 1991 Propedeutique Morphologique. Folia Linguistica 25, 3-4: 549-575. 1999 On the locality of morphological operations. Ms., Universite de Montreal.

Constraints, preferences, and context-sensitivity

357

Ford, Alan, Rajendra Singh and Gita Martohardjono 1997 Pace Panini. New York: Peter Lang. Jespersen, Otto 1924 The Philosophy of Grammar. London: George Allen & Unwin. Mayerthaler, Willi 1981 Morphologische Natürlichkeit. Wiesbaden: Athenaion. Pesetsky, David 1997 OT and syntax. In: Diane Archangeli and D. Terence Langendoen (eds.), Optimality Theory: An Overview, 134-170. Oxford: Blackwell. Philip, Mayjee 1997 Compounding in Khasi. In: Anvita Abbi (ed.), Languages of Tribal and Indigenous Peoples of India, 361-372. Delhi: Motilal Banarsidass. Plenat, Marc 1996 Morphophonologie des derives en -Vche. Recherches Linguistiques de Vincennes 26: 113-150. Potter, Brian 1994 Serial optimality in Mohawk prosody. Chicago Linguistic Society 30: 347-361. Chicago: Chicago Linguistic Society. Russell, Kevin 1997 OT and Morphology. In Diane Archangeli and D. Terence Langendoen (eds.), Optimality Theory: An Overview, 102-133. Oxford: Blackwell. Sanders, Gerald 1988 Zero derivation and the overt analogue criterion. In: Michael Hammond and Michael Noonan (eds.), Theoretical Morphology, 143-154. San Diego: Academic Press. Singh, Rajendra and Rama Kant Agnihotri 1997 Modern Hindi Morphology. Delhi: Motilal Banarsidass Singh, Rajendra and Probal Dasgupta 1999 On so-called compounds. In: Rajendra Singh (ed.), The Yearbook of South Asian Languages and Linguistics, 265-278. New Delhi: Sage Publications. Tirumalesh, Kilingar Venkapp 1997 Kannada prefixation. In: Maramraj Hariprasad, Valluri Hemalatha Nagarajan, Punnappurath Madhavan, and Kumarlingam Gopalkrishnan Vijaykrishnan (eds.), Phases and Interfaces of Morphology, 7 4 84. Hyderabad: The Central Institute of English and Foreign Languages.

358

Rajendra Singh

Vennemann, Theo 1983 Theories of linguistic preferences as a basis for linguistic explanations. Folia Linguistica Historica 6: 5-26. Vijaykrishnan, Kumarlingam Gopalkrishnan 1998 Conflict resolution in Optimality Theory: The case of antifaithfulness requirement in reduplicative morphology. Paper presented at GLOW 1999, Central Institute of English and Foreign Languages, Hyderabad, India, 20-23 January 1999. Wali, Kashi and Omkar Koul 1997 Kashmiri. London: Routledge. Wiltshire, Caroline 1994 Feature constraints in IruLa. Chicago Linguistic Society 30, 1: 428442. Chicago: Chicago Linguistic Society. Wang, William S.-Y. 1968 Vowel features, paired variables, and the English Vowel Shift. Language 44: 695-708.

Old English fricatives: lenition and licensing* Jerzy Wojcik

1. Introduction The goal of this paper is to examine the phonological properties of Old English fricatives. The main focus here is on establishing the mechanism responsible for the voicing of Old English anterior fricatives, i.e. / , s and θ 'ιη intervocalic contexts as well as analysing the status of the voiced velar fricative γ and the palatal fricative / - the two fricative sounds whose phonological behaviour renders them highly exceptional in the synchronic phonology of Old English. The framework adopted in the present paper is that of Government Phonology (GP) as developed in the works of Kaye, Lowenstamm and Vergnaud (1985, 1990), Harris (1994, 1997), Cyran (1997), and others. The structure of the paper is as follows. In section 2, we discuss two phenomena - the voicing of the anterior spirants/ 5 and #and the deletion of the voiceless velar fricative x. This is followed, in section 3, by a discussion of the voiced velar fricative γ. Section 4 deals with the Old English voiceless palatal f . Finally, in section 5, we offer some conclusions.

2. The voicing of f, s and θ 2.1. The presentation of the data The Old English system of spirants has been the subject of numerous analyses carried out by scholars of different theoretical backgrounds. On the one hand, a number of works can be mentioned couched in structural terms whose authors, e.g., Moulton (1954, 1972), Kurath (1956), Hogg (1992a), generally agree that from the point of view of the phonemic analysis the Old English fricative phonemes /f, Θ, s/ had voiced allophones [ν], [ö] and [z] between vowels. Occasionally the

360

Jerzy Wöjcik

allophonic status of the voiced anterior spirants has been questioned, see for example Bammesberger (1988), who suggests that a strict application of the standard procedures of the phonemic analysis might indicate that the voiced fricatives enjoyed the phonemic status in Old English. In a similar fashion, scholars of other, sometimes quite disparate, theoretical backgrounds, e.g., Campbell (1959: 20), Lass and Anderson (1975), have formulated rules responsible for the voicing of fricatives in Old English which are basically statements of certain distributional facts to the effect that voiced fricatives are expected to be found in intervocalic position, whereas their voiceless counterparts initially and finally. The major part of this paper will be concerned with the nature of the mechanism necessary to describe the voicing of Old English anterior spirants. We start our discussion by presenting some examples showing a well-known property of Old English phonology - the appearance of voiceless vs. voiced anterior spirants is completely predictable from the phonetic context, so that voiced fricatives are only found in intervocalic position. Consider the examples below. ( 1 ) a.

voiceless

θ

initially

far an 'fare'

pencan 'think'

singan

finally

hläf' bread'

äp 'oath'

hüs 'house'

w«//"'wolf

heorö 'hearth'

fyrs 'furze'

next to another

arfter 'after'

sing

breost 'breast'

obstruent (including itself) b.

voiced

pyffan 'puff

ν

moppe 'moth'

Ö

ofer 'over'

cyssan 'kiss'

ζ

öper 'other'

hüses 'house GEN.SG'

wulfas

heordas

fyrsas

'wolf

'hearth

'furze

NOM.PL'

hraefn 'raven'

NOM.PL'

aSpm 'breath'

NOM.PL'

hüs I 'Eucharist'

Old English fricatives:

lenition and licensing

361

One interesting property emerging from (1) is that fricatives are voiced not only between vowels {ofer, öper, hüses) but also intervocalically when preceded by a liquid / or r as in wulfas, heordas, fyrsas. Lass (1994: 71) formulates the rule of fricative voicing in the following way: fricative f,6,s are voiced intervocalically when a stressed vowel precedes, optionally followed by a liquid, and an unaccented vowel follows. Another important aspect of the distribution of Old English anterior spirants is noted by, e.g., Anderson (1985: 218), who draws attention to the importance of morphological structure, which explains why words like äsettan 'set', Wilsum 'desirable' orprymful 'glorious' show the voiceless s and/respectively, even though the consonants are preceded and followed by voiced segments. As we can see, these words are morphologically complex and the position in which the fricatives appear can be viewed as the morpheme initial position, where Old English anterior fricatives are expected to be realised as voiceless sounds (as infaran 'fare',pencan 'think' and singan 'sing' in (la)). We believe that this sensitivity of fricative voicing to morphological structure strongly suggests that whatever mechanism is responsible for the voicing of fricatives it is phonological rather then purely phonetic in nature. Moreover, consider the examples in (2) below which illustrate an interesting, albeit frequently overlooked, aspect of Old English phonology; the voiceless fricative χ is deleted in precisely the same environment as that for the voicing o f f (9 and s.] (2)

a. feoh 'money NOM.SG' seolh 'seal NOM.SG' mearh 'horse NOM.SG' b. hlcehhan 'laugh' c. heorte 'heart NOM.SG'

-

ßos ( 0 can also be represented as h-loss, which is definitely what we should expect given our earlier observations concerning a unitary nature of both fricative voicing and x-deletion. In order to determine that we have to examine the phonological representation ofx. Harris (1994: 126) represents a traditional GP assumption, which regards velar consonants as defined by the presence of the neutral element @ (alternatively represented as v°), characterising velarity in consonants. His representation o f x consists then of two elements (h. More recently, however, it has been proposed that velar consonants should be represented as headless or empty-headed.6

Old English fricatives: lenition and licensing

365

Under this assumption the representation of χ could be (h. _). This representation allows us to express alternations between χ and 0 as a phonological process targeting (i.e. delinking) the element h. This is clearly desirable since accepting the proposed solution enables us to capture an important generalisation inherent in the intuition that both fricative voicing and deletion of χ are really the same; what the two processes involve is a simple operation consisting in the delinking of a single element h.7 To summarise the discussion so far: Old English intervocalic voicing o f f , 6 , s and the deletion of x, which operate in the same context (between voiced segments) can be analysed in a uniform fashion and represented as a lenition process consisting in the loss of the element h. In what follows we shall suggest that the loss of h results from the application of Licensing Inheritance, as developed in Harris (1992, 1994, 1997), which states that a position's ability to license phonological material is determined by its place in the prosodic structure.

2.3. h-loss as an instance of Licensing Inheritance Generally speaking, the amount of phonological material, i.e. the number of elements in a given segment, is expected to be depleted in prosodically weak positions. One of the predictions following from the application of Licensing Inheritance is that the onset licensed by the recessive nucleus should not be able to accommodate as many elements as the onset licensed by the dominant nucleus. This asymmetry in the behaviour of foot-initial and foot-internal onsets results from the fact that foot-initial onsets are licensed directly by the head of the domain (cf. 6a below) whereas foot-internal onsets inherit their licensing potential at two removes from the head of the domain as in (6b). The two configurations are represented in (6) below, after Harris (1997).

366 (6)

Jerzy Wöjcik

(a) foot-initial C

(b) foot-internal C (intervocalic/domain final)

[Xl]o [X2]n [X3]o [X4]n

ψ 'V [XJ]o[X2]N [*3]O [X4]N

C Returning to our discussion of the Old English data, we immediately notice that the word-initial onset in Old English will always be licensed by the dominant nucleus of the foot. This in turn means that initial onsets are prosodically strong enough to license the element h and hence emerge as voiceless. The representation we have introduced in (6b) identifies both an intervocalic consonant and a word-final consonant as prosodically weak. However, there is an important difference between the two contexts, namely they differ in the nature of the nucleus which follows the onset in question. In the case of an intervocalic onset the following nucleus is melodically filled, while a domain final consonant is followed by a nucleus which lacks melody. According to Harris (1997), this distinction is crucial as not every lenition process occurring in one context necessarily also occurs in the other. Harris mentions for example Spanish s-debuccalisation which is present word-finally but not intervocalically. More crucially for our analysis of Old English fricatives, however, there are also lenition processes confined only to foot-internal contexts. Harris (1994: 210) cites, for example, the intervocalic suppression of ö occurring in some dialects of Modern English (those spoken in the north of Ireland) and a historical process which affected d in Dutch; hence cognate English brother and Dutch broer. Harris (1997: 210) further claims that such processes can be formally expressed as the failure of a nucleus licensed on the foot-projection to confer enough licensing potential to the preceding onset. As a result an onset cannot a-license certain melodic elements. We believe that the same mechanism can be employed in the description of fricative voicing in Old English. The situation can be represented as in (7):

Old English fricatives:

(7)

a.

ONi

Ο N2

b.

lenition and licensing

Ο Ν, Ο N 2 Ο N 3

XXX XX

XXX X X

h

h

u

A h [s]

367

u

XX

A e s > ^ (delinking) ι

[z]

In (7a) the fricative s finds itself between two nuclei Ni and N2. Since N2 is empty, lenition does not occur and the segment is realised as voiceless. In (7b), on the other hand, the nucleus under N2 is melodically filled by the vowel of the inflectional ending. As indicated above, N 2 , being melodically filled, is not able to license all elements in the representation of s and consequently the element h is delinked from the representation. This results in a segment which is phonetically realised as the voiced fricative z. In sum then, employing the formal notion of Licensing Inheritance, we have presented the mechanism whereby the anterior spirants.?, Gmc. /γ/, with later hardening of γ to g after a nasal, in gemination and at some stage also initially. The process of hardening is obviously part of a general development affecting all voiced fricatives emerging either from IE voiced aspirates or through Verner's Law acting on voiceless fricatives (see, e.g., Prokosch 1939: 60; Lass 1994: 56). In the case of the hardening of γ it is especially the initial position which appears to be problematic as it is not certain when exactly the initial γ became a stop. On the one hand, there are scholars like Bülbring (1902: §487), Jordan (1974: §184), Campbell (1959: §50.4) who place the change in late Old English. On the other hand, one can mention an influential scholarly tradition represented by Luick (1914-1940: §633) or Brunner (1965: §211), suggesting that the occlusion of κ to g was prehistoric. Some modern scholars, for example Lass and Anderson (1975), also maintain that initially there was a stop throughout the Old English period, while others (see, e.g., Lass 1994: 57) adopt a late date of hardening. There seems to be a general agreement, however, that initially Kwas still present at the time of palatalisation, which explains why its palatalised reflex is always j. In our analysis we follow those scholars who place the word initial change of γ to g in late Old English. Let us begin our discussion by noting certain peculiarities in the phonological behaviour of γ The exceptionality of the voiced velar fricative is best visible when we consider its distribution. The first thing to notice is that it does not pattern with the remaining Old English voiced fricatives. We have seen above that voiced fricatives are expected to be found only in the intervocalic position. This is not true of the velar fricative. If we do not accept an early date for the initial

Old English fricatives:

lenition and licensing

373

occlusion of γ to g, then words like göd 'good', gän 'go', gutt 'gut' should be pronounced with the word-initial γ Secondly, according to Campbell (1959: §446-447), the voiced velar fricative, at least in early West Saxon, seemed to appear word-finally. Anderson (1985) lists words like mearg 'marrow', flog 'flayed' with the word final κ as opposed to χ in words like mearh 'horse', fleah 'flew'. Those final occurrences of γ appear to be eliminated in the course of the Old English period; Campbell (1959: §447) noted back spellings such as mearg for mearh. Finally, κ was also present word-medially between back vowels, for example lagu 'law', dragan 'draw', etc. Schematically the distribution of κ in Early and Late Old English can be represented as follows: (13) initially medially finally

Early Old English gän 'go' [γ] dragan 'draw' [γ] mearg 'marrow' [γ]

Late Old English [g] [Y]

[x]

Below we take up the problem of the phonological representation of γ which should allow us to account for all the peculiarities displayed by this segment.

3.1. The phonological representation of γ Let us start the discussion by looking at the possible ways of representing γ. A standard GP assumption concerning the representation of j'can be found in e.g., Bloch-Rozmej (1998), who proposes to analyse γ as (h._) in her account of Irish, where this sound arises due to the lenition of d and g. Under this account, κ is defined by the presence of h responsible for noise and an empty head defining the velar place of articulation. Quite clearly, however, if our analysis of the Old English facts presented above is correct, γ cannot possess h in its elemental make-up, otherwise it would be phonetically voiceless as is the case with the voiceless s,6,f andx. As indicated in the previous section, the exceptionality of γ best shows in its distributional freedom, which

374

Jerzy Wojcik

evidently follows from the fact that γ is the only Old English voiced fricative which does not result from the voicing of its voiceless counterpart in the intervocalic context. On the other hand, the phonological representation of γ should reflect the fact that it is a member of the same class of sounds as the remaining voiced fricatives. With this in mind, let us note that the representations of the voiced fricatives we have proposed above make use of the two resonance elements: U and A. The relevant representations are repeated in (14) below. (14)

ν (U), ζ (A), Ö ( A )

The only resonance element which is not employed in the representations in (14) is I. Let us hypothesise, then, that I is the element used in the phonological representation of γ. If this is correct, γ could be represented as (I), where the headedness of the element I is responsible for the phonetic effect of friction10 and additionally serves to distinguish γ from j, which is a headless object (I). It should be emphasised that analysing γ as (I) is beneficial since in this way we are able to present a uniform treatment of all voiced fricatives as defined by the presence of the relevant resonance element. Ideally, the proposed representation of κ should also be able to account for the distributional freedom of this segment - the fact which still remains unexplained.11 What may be important in this respect is the behaviour of γ throughout the Old English period (as represented in (13) above). Note that κ had a relatively free distribution only in Early Old English. The exclusion of γ from the word-initial and word-final positions, which sets apart Late Old English from Early Old English, could then be seen to follow from an attempt to regularise this highly exceptional segment - phonologically similar to the remaining voiced fricatives (in terms of its phonological representation) but at the same time displaying a considerable amount of distributional leeway. The situation which had been reached by Late Old English clearly aligns κ with the remaining voiced fricative sounds; γ is akin to the remaining Old English voiced fricatives both in terms of its phonological representation and its distribution. Therefore, adopting the representation of γ defined as (I) may bring us closer to understanding the differences in the distribution of γ

Old English fricatives:

lenition and licensing

375

between Early and Late Old English. One interesting consequence of this changed distribution of γ is that, in a fashion resembling the anterior spirants, γ starts to participate in contextually conditioned alternations. Consider some Late Old English examples in (15). (15)

plög [ploix] 'plough' dweorg 'dwarf -

plögas [plo:yas] N O M . P L dweorgas Ν Ο Μ . P L

The existence of such Late Old English examples indicates that the relation between χ and γ can now be analysed along the lines proposed for the anterior spirants. In such a case the appearance of γ in an intervocalic context would be a straightforward consequence of the delinking of h. Note that, if κ is to be represented as (I), the voiceless velar spirant ;t found in the examples in (15) should be analysed as (h.I). In fact, these Late Old English developments have interesting consequences for the phonological standing of JC. This is so because alongside the pattern presented in (15) (JC alternating with γ) Late Old English continues to preserve another type of χ (the one we have come across earlier), which still alternates with 0 , so that eolh gives in nom.pl. eolas rather then *eol[y]as. The existence of these two patterns of alternations suggests that phonologically we are dealing with two different kinds of χ ; one alternating with 0 and, as suggested in our discussion above, phonologically represented as (h._), and the other alternating with γ and phonologically most probably represented as (h.I). To sum up, representing Old English voiced velar fricative γ as (I) allows us to treat uniformly the class of Old English voiced fricative sounds as defined by the presence of the relevant resonance element. Additionally, we propose to view the change whereby by LOE γ is found exclusively in intervocalic position as following from an attempt to bring j^into line with the remaining Old English voiced spirants. In effect, the voicing of all Late Old English spirants results from h not being licensed intervocalically. In the following section we shall present additional evidence which further strengthens the proposal we have made concerning the phonological representation of OE γ

376

Jerzy Wöjcik

3.2. Old English alliterative practice One of the peculiarities of Old English alliteration is found in the early poetic practice where the voiced velars and palatal approximants from Proto-Germanic alliterate freely with each other. Minkova (in prep.) cites the following examples illustrating this point. (16) godes [γ] 13 'God'

alliterates with

GEN.SG

geomor [j] 'mournful', G. jammervoll

god [γ] 'God'

alliterates with

geong [j] 'young', G.jung

goldwine [γ] 'gener-

alliterates with

geara [j] 'year' GEN.PL, G.

ous lord'

Jahr

In the later poetic works, presumably after the change of γ —> g in the word-initial position, alliteration becomes 'regular' and words with j alliterate only with themselves. It is perhaps worth noting that traditionally the change in alliterative practice in Late Old English has been regarded as one of the arguments for placing the hardening of γ to g in Late Old English (e.g., Campbell 1959: §430). While the Late Old English situation is clear and the lack of alliteration between the postulated g and j is not surprising, the interpretation of the Early Old English situation has been the subject of much controversy.14 It has sometimes been suggested that the appeal to alliterative practice is not conclusive (e.g., Hogg 1992a: §2.56 n.2) since exceptional alliteration facts might be regarded as resulting from the so-called eye-alliteration. Under this interpretation it is graphs which alliterate (the relevant sounds were both spelled with ) and not sounds. A contrary view is expressed in Minkova (in prep.) who convincingly argues in favour of accepting linguistic identity as the basis for alliteration. If this is the case, our observations referring to alliteration facts can be summarised in the following way (adapted from Minkova in prep.).

Old English fricatives:

(17) Alliteration Reconstruction

lenition and licensing

377

Early Old English Late Old English : : [j ]: [γ] m-ui [g]:[g]

The importance of alliterative practice for our discussion of the phonological status of γ in Old English is quite clear. Representing γ as (I) offers a straightforward interpretation of Early Old English alliteration facts: γ and j alliterate because phonologically they are defined by the presence of the same element, namely I. It thus seems that alliteration lends further support to the representation of κ we have been advocating.15 In the next section we are going to investigate further implications following from the adopted interpretation of γ

3.3.

γ - j

alternations

The most direct piece of evidence showing that γ and j are phonologically related can be seen in morphological alternations. There are numerous cases in Old English when the voiced velar fricative appears in some forms of a morpheme while j appears in others. Historically, the alternations result from palatalisation whereby γ —> j in the context of a front vowel.16 Consider the examples below which show the most important cases of allomorphic variation caused by the palatalisation of Y(18)

a. dse[j] 'day', we[j] ea\j]e

das [j]es GEN.SG - da[y]as

' w a y ' - wefy^as

ΝΟΜ.PL

' e y e NOM.SG' - ea[\]an

b. SW[Y] 'sorrow' -

sor\}\e

NOM.PL

GEN.DAT.SG, NOM.PL

ACC.SG - sor[y]um

DAT.PL

The examples listed in (18a) show a pattern in which j is found in the word final position, e.g., dseg, or when followed by a front vowel e, e.g., dasge, eage. In those examples γ appears only when followed by a back vowel as in dagas. The proposal we would like to make is that the pattern of alternations as found in the examples in (18a) is triggered by the same mechanism which is responsible for the vocalic alterna-

378

Jerzy Wojcik

tions between ae and α in e.g.,