A Computational Phonology of Russian

A Computational Phonology of Russian by Peter A. Chew

ISBN: 1-58112-178-4 DISSERTATION.COM

Parkland, FL • USA • 2003

A Computational Phonology of Russian Copyright © 2000 Peter A. Chew All rights reserved. Dissertation.com USA • 2003 ISBN: 1-58112-178-4 www.Dissertation.com/library/1121784a.htm

A Computational Phonology of Russian Peter Chew Jesus College, University of Oxford D. Phil. dissertation, Michaelmas 1999 Abstract This dissertation provides a coherent, synchronic, broad-coverage, generative phonology of Russian. I test the grammar empirically in a number of ways to determine its goodness of fit to Russian. In taking this approach, I aim to avoid making untested (or even incoherent) generalizations based on only a handful of examples. In most cases, the tests show that there are exceptions to the theory, but at least we know what the exceptions are, a baseline is set against which future theories can be measured, and in most cases the percentage of exceptional cases is reduced to below 5%. The principal theoretical outcomes of the work are as follows. First, I show that all of the phonological or morphophonological processes reviewed can be described by a grammar no more powerful than context-free. Secondly, I exploit probabilistic constraints in the syllable structure grammar to explain why constraints on word-marginal onsets and codas are weaker than on wordinternal onsets and codas. I argue that features such as [!initial] and [!final], and extraprosodicity, are unnecessary for this purpose. Third, I claim that !"! should be lexically unspecified for the feature [!sonorant], and that the syllable structure grammar should fill in the relevant specification based on its distribution. This allows a neat explanation of the voicing assimilation properties of !"!, driven by phonotactics. Fourth, I argue that jers in Russian should be regarded as morphological objects, not segments in the phonological inventory. Testing the grammar suggests that while epenthesis cannot be regarded as a major factor in explaining vowel-zero alternations, it might be used to explain a significant minority of cases. Fifth, I suggest that stress assignment in Russian is essentially context-free, resulting from the intersection of morphological and syllable structure constraints. I show that my account of stress assignment is simpler than, but just as general as, the best of the three existing theories tested. Finally, this dissertation provides new insight into the nature and structure of the Russian morphological lexicon. An appendix of 1,094 morphemes and 1,509 allomorphs is provided, with accentual and jer-related morphological information systematically included.

_______________________________

A Computational Phonology of Russian by Peter Chew

University of Oxford Jesus College Michaelmas 1999

_______________________________ Thesis submitted for the degree of Doctor of Philosophy at the University of Oxford

Acknowledgements I would like to thank my supervisor, John Coleman, for his help. Without his encouragement and support even before I embarked upon this research, I would doubtless now be a well-paid but bored chartered accountant. Auditing linguistic theories has proved to be more rewarding in many ways than auditing financial statements, and I am confident that the choice of leaving my previous job to pursue this research was the right one. It would not have been possible to complete this D. Phil. without the support of my wife, Lynn. She has always been there to give practical suggestions, as a sounding board for ideas, and simply as a partner in life, sharing encouraging and discouraging times together. God could not have given me a better wife. My parents have also been a great practical help, babysitting almost weekly, having us round for meals, and generally helping reduce the stress in our lives. Although Jonathan, who was born 15 months before I submitted this thesis, has taken time from my studies, we are very grateful for his arrival. I cannot think of a better way of spending my time, and I cannot imagine a better son. A number of people have read drafts of my work or listened to me, giving helpful advice which enabled me to sharpen my thoughts and improve the way in which I expressed them. Thanks (in alphabetical order) to Dunstan Brown, Bob Carpenter, Andrew Hippisley, Mary MacRobert, Stephen Parkinson, Burton Rosner, Irina Sekerina, Andrew Slater, and Ian Watson. Andrew Slater has also provided invaluable technical support. I often feel that he puts the rest of us to shame with his good humour, helpfulness, and a constant willingness to go the extra mile. My friends at the Cherwell Vineyard Christian Fellowship have provided a dependable support network which has kept Lynn and me going through not always easy times. First and foremost, they have encouraged us to keep looking towards the one without whom we can do nothing. However, I know I will also look back on the laughs we have had with Richard and Janet Remmington, Evan and Eowyn Robertson, Judy Irving, and others, on Thursday evenings with fond memories. Finally, I would like to thank my college, Jesus College, for providing very generous financial support throughout my time at Oxford. And without the financial support of the Arts and Humanities Research Board (formerly the British Academy), I would not have undertaken this research project in the first place.

List of abbreviations and symbols General symbols !! () !#c++,#-rn+in! morphs + . / right) !+,#-!rn obliques) " # anter C CFG cons cont coron DCG del_rel init later OT PSG sonor SSG V vfv voc

enclose phonemic representations, e.g. !#$%&'! enclose phonetic representations, e.g. (*$%&') denotes morphological tokenization; subscripts classify individual morpheme boundary syllable boundary denotes word-stress in IPA transcriptions (stress on the vowel to the denotes a single morpheme (classificatory subscript is outside syllable the empty string anterior any consonant context-free grammar consonantal continuant coronal (Prolog) Definite Clause Grammar delayed release initial lateral Optimality Theory phrase structure grammar sonorant Sonority Sequencing Generalization any vowel vocal fold vibration vocalic

Symbols used in morphological tokenization r* s c* i p a n v

root suffix clitic inflectional ending pronominal adjectival substantival verbal

d r* i c*

durative process resultative process iterative process completed process

*No ambiguity arises with respect to the use of non-unique symbols, because the meaning of each symbol is also dependent on its position; full details are given in section 3.2.1.2.

5 Table of contents Acknowledgements ..................................................................................................................................3 List of abbreviations and symbols ............................................................................................................4 Table of contents ...................................................................................................................................... 5 Table of figures.........................................................................................................................................7 List of tables .............................................................................................................................................8 Chapter 1: Introduction.............................................................................................................................9 1.1 Introduction ................................................................................................................................ 9 1.2 Why ‘computational’ linguistics? ............................................................................................13 1.3 The framework .........................................................................................................................16 1.3.1 Phrase-structure grammar ...............................................................................................16 1.3.2 Context-free grammar.....................................................................................................19 1.4 The methodology .....................................................................................................................24 1.5 The dataset used for the tests....................................................................................................26 1.6 Summary ..................................................................................................................................30 Chapter 2: Syllable structure ..................................................................................................................32 2.1 Overview and aims...................................................................................................................32 2.2 The syllable in phonological theory .........................................................................................34 2.2.1 Sonority and syllable structure .......................................................................................37 2.2.2 Morpheme structure constraints or syllable structure constraints? .................................40 2.2.3 Syllable structure assignment ......................................................................................... 43 2.2.3.1 Kahn’s (1976) syllable structure assignment rules .................................................... 45 2.2.3.2 Itô’s (1986) method of syllable structure assignment ................................................49 2.2.3.3 Syllable structure assignment in Optimality Theory..................................................51 2.2.3.4 Phrase-structure analysis of syllable structure ...........................................................54 2.2.3.5 Syllable structure assignment: conclusions................................................................56 2.3 A linear grammar of Russian syllable structure .......................................................................58 2.3.1 The phonological inventory of Russian .......................................................................... 59 2.3.1.1 Preliminaries: controversial issues.............................................................................59 2.3.1.2 The classification system ........................................................................................... 68 2.3.2 The syllable structure rules ............................................................................................. 72 2.4 A heuristic for deciding between multiple syllabifications ......................................................95 2.5 Extensions to the grammar ....................................................................................................... 99 2.5.1 Further phonological features .......................................................................................102 2.5.2 Four phonological processes in Russian .......................................................................105 2.5.2.1 Consonant-vowel interdependencies........................................................................ 105 2.5.2.2 Reduction of unstressed vowels ...............................................................................114 2.5.2.3 Word-final devoicing ............................................................................................... 120 2.5.2.4 Voicing assimilation ................................................................................................127 2.5.3 A test of the extensions to the grammar........................................................................141 2.6 Summary ................................................................................................................................146 Chapter 3: Morphological structure......................................................................................................149 3.1 Introduction and aims.............................................................................................................149 3.1.1 Generative approaches to word-formation.................................................................... 152 3.1.2 Morphology and context-free grammar ........................................................................158 3.2 A linear grammar of Russian word-formation .......................................................................161 3.2.1 The morphological inventory of Russian......................................................................161 3.2.1.1 Preliminaries: controversial issues...........................................................................164 3.2.1.2 The classification system .........................................................................................165 3.2.2 The word-formation rules .............................................................................................170 3.2.2.1 Words with no internal structure..............................................................................171 3.2.2.2 Nouns.......................................................................................................................172 3.2.2.3 Verbs........................................................................................................................178 3.2.2.4 Prefixation................................................................................................................180 3.3 Vowel-zero alternations in context-free grammar.................................................................. 185 3.4 A heuristic for deciding between multiple morphological analyses....................................... 202 3.4.1 Assigning costs to competing analyses.........................................................................205 3.4.2 Should the cost mechanism be based on hapax legomena?..........................................209

6 3.5 Tests of the word-formation grammar....................................................................................214 3.5.1 Test of coverage of the word-formation grammar ........................................................215 3.5.2 Test of the grammar’s treatment of vowel-zero alternations ........................................218 3.6 Conclusion .............................................................................................................................222 Chapter 4: Stress assignment: three existing theories...........................................................................224 4.1 Introduction ............................................................................................................................ 224 4.1.1 Two approaches to stress in Russian: the Slavist and the generative approaches.........224 4.1.2 Aims of this chapter......................................................................................................232 4.2 Three theories of stress assignment........................................................................................ 233 4.2.1 Halle (1997)..................................................................................................................233 4.2.2 Melvold (1989).............................................................................................................237 4.2.3 Zaliznjak (1985) ...........................................................................................................244 4.3 Derivational theories and underdeterminacy..........................................................................248 4.3.1 Computing underlying accentuations by ‘brute force’..................................................251 4.3.2 Backwards phonology and the Accent Learning Algorithm.........................................252 4.3.2.1 A concise encoding of ‘solutions’............................................................................257 4.3.2.2 Formalization of the Accent Learning Algorithm.................................................... 259 4.3.2.3 A small-scale demonstration of the ALA on a non-problem combination...............261 4.3.2.4 Problem words .........................................................................................................271 4.3.2.5 Modifications to the ALA to allow for different theories ........................................274 4.3.2.6 Conclusions from the ALA......................................................................................278 4.3.3 Unique specification of the morpheme inventory by defaults ......................................283 4.4 Tests to ascertain the coverage of the three theories ..............................................................291 4.4.1 Test of Halle’s theory on non-derived nouns................................................................292 4.4.2 Test of Halle’s theory on non-derived and derived nouns ............................................293 4.4.3 Test of Melvold’s theory on non-derived and derived nouns .......................................294 4.4.4 Test of Melvold’s theory on nouns, non-reflexive verbs and adjectives.......................295 4.4.5 Test of Zaliznjak’s theory on nominative singular derived nouns ................................296 4.4.6 Test of Melvold’s theory on nominative singular derived nouns..................................297 4.4.7 Analysis of errors in Melvold’s and Zaliznjak’s theories .............................................298 4.5 Summary ................................................................................................................................307 Chapter 5: Stress assignment: a new analysis.......................................................................................309 5.1 Introduction ............................................................................................................................ 309 5.2 Context-free phonology and stress in Russian .......................................................................311 5.2.1 Encoding which morpheme determines stress ..............................................................312 5.2.2 Polysyllabic morphemes...............................................................................................318 5.2.3 Post-accentuation..........................................................................................................319 5.2.4 Jer stress retraction ....................................................................................................... 325 5.2.5 Plural stress retraction...................................................................................................329 5.2.6 Dominant unaccented morphemes................................................................................333 5.2.7 Concluding comments about the context-free phonology ............................................336 5.3 A test of the entire grammar...................................................................................................338 5.4 Conclusions ............................................................................................................................343 Appendix 1: Russian syllable structure grammar ................................................................................. 346 Appendix 2: Russian word-formation grammar ...................................................................................355 Appendix 4: Morphological inventory .................................................................................................358 Appendix 5: The computational phonology as a Prolog Definite Clause Grammar.............................392 References ............................................................................................................................................413

7 Table of figures Figure 1. The Chomsky Hierarchy .........................................................................................................20 Figure 2. Classification of analyses of an imperfect grammar ...............................................................25 Figure 3. Tree-structure for !+%+%! .......................................................................................................75 Figure 4. Lattice showing the hierarchy of Russian phoneme classes..................................................110 Figure 5. The Russian vowel system ....................................................................................................115 Figure 6. The Russian vowel system in unstressed positions ...............................................................116 Figure 7. Partial syllabic structure of pretonic !%! after a [$back] consonant .......................................119 Figure 8. Tree-structure for !"#$%& !0'1$'2%! (0'1&'.23)..........................................................................138 Figure 9. Parse tree for '"#()*)+,-.$ !4'1$"',5,+6#&'! ......................................................................158 Figure 10. Examples of subtrees from Figure 9....................................................................................159 Figure 11. Morphological tokenization of '"#()*)+,-.$ !4'1$"',5,+6#&'!.........................................160 Figure 12. Parse tree for '"#()*)+/0 !4'1$"',5,+,7!..........................................................................161 Figure 13. Oliverius’s (1976) tokenization of *"'1)'& !5148&8',4%! ‘woman’....................................175 Figure 14. Parse tree for *"'1)'& !5148&8',4%! ‘woman’..................................................................... 175 Figure 15. Three alternative representations of !906c+&8',&rv+%svi+&'sa!...................................................181 Figure 16. Representation of the morpheme !-%#2!%!-%#62! ‘weasel’ ................................................... 190 Figure 17. Structure of #"'23"% ........................................................................................................... 199 Figure 18. Structure of 4,#,*56& ........................................................................................................200 Figure 19. Structure of 4,#*,5 ............................................................................................................201 Figure 20. Parse tree for -4"7)&6$',-.$ (with log probabilities) ........................................................208 Figure 21. Rank-frequency graph .........................................................................................................213 Figure 22. Analysis of coverage of morphology grammar ...................................................................217 Figure 23. Parse tree for -4"7)&6$',-.$..............................................................................................314 Figure 24. Morphological/phonological structure of #8!&%&: !$;0ra+%2sn+%in3!..................................... 322 Figure 25. The constraint pool.............................................................................................................. 324 Figure 26. Morphological/phonological structure of #8!&:% !$;0ra+%2sn+in1!.........................................327 Figure 27. Morphological/phonological structure of (/-,:./ !",#ra+/6&sn1+,in!..................................... 332 Figure 28. Morphological/phonological structure of 8:!,("'$ //;c+06"'ra+14'sn+in! .............................. 335

8 List of tables Table 1. Types of rules permitted by grammars in the Chomsky Hierarchy ..........................................20 Table 2. Analysis of words in on-line corpus .........................................................................................30 Table 3. Russian morpheme structure constraints on consonant clusters ...............................................41 Table 4. Reanalysis of morpheme-medial clusters using syllable structure ...........................................42 Table 5. Phonological inventories of different scholars .........................................................................65 Table 6. The phonemic inventory of Russian .........................................................................................67 Table 7. Classification of Russian phonemic inventory ......................................................................... 69 Table 8. Distribution of word-initial onsets by type ...............................................................................77 Table 9. Distribution of word-final codas by type..................................................................................88 Table 10. Further coda rules ...................................................................................................................90 Table 11. Exhaustive list of initial clusters not accounted for ................................................................91 Table 12. Exhaustive list of final clusters not accounted for ..................................................................92 Table 13. The twelve most frequently applying onset, nucleus and coda rules......................................97 Table 14. Feature matrix to show classification of Russian phonemes and allophones with respect to all features........................................................................................................................................103 Table 15. Allophonic relationships in consonant-vowel sequences .....................................................107 Table 16. Allophones of !%! and !6! ...................................................................................................... 117 Table 17. Results of phoneme-to-allophone transcription test .............................................................145 Table 18. Classification system for substantival inflectional morphs...................................................169 Table 19. Further categories of morphological tokenization ................................................................173 Table 20. Summary of results of parsing 11,290 words .......................................................................217 Table 21. Derivations of six Russian words in accordance with Halle (1997) .....................................237 Table 22. Derivations of five Russian words in accordance with Melvold (1989)...............................242 Table 23. Possible solutions for -.,6 !#&/6-! ‘table’ (nom. sg.).............................................................254 Table 24. Possible solutions for -.,6&:
9

Chapter 1: Introduction 1.1

Introduction This dissertation provides a coherent, synchronic, broad-coverage, generative

account of Russian phonology. By ‘broad-coverage’, I mean that it will cover a number of phonological phenomena (stress assignment, syllabification, vowel-zero alternations, word-final devoicing, voicing assimilation, vowel reduction, and consonant-vowel interdependencies) within a single constrained grammar. While I have not attempted to deal exhaustively with all the phonological problems of interest in Russian (for example, I do not attempt to account for all morphophonological alternations), the current work covers those areas which have attracted the most attention in the literature on Russian phonology. While all these aspects of Russian phonology have been richly documented, generally they have been dealt with in isolation; the one notable exception to this is Halle’s (1959) Sound Pattern of Russian. The following quotation (op. cit., p. 44) serves to show that Halle’s account of Russian phonology is also intended to be broad-coverage in the sense just outlined: When a phonological analysis is presented, the question always arises as to what extent the proposed analysis covers the pertinent data. It is clearly impossible in a description to account for all phonological manifestations in the speech of even a single speaker, since the latter may (and commonly does) use features that are characteristic of different dialects and even foreign languages. (E.g., a speaker of Russian may distinguish between nasalized and nonnasalized vowels in certain [French] phrases which form an integral part of his habitual conversational repertoire.) If such facts were to be included, all hopes for a systematic description would have to be abandoned. It is, therefore, better to regard such instances as deviations to be treated in a separate section and to restrict the main body of the grammar to those manifestations which can be systematically described.

10 The aim of the current work is thus substantially the same as that of Halle (1959). However, in the forty years since then there have been a number of advances, both linguistic and technological, which allow us to take a fresh (and perhaps more rigorous) look at some of the same phenomena which Halle and others attempted to describe. In the late 1950s and early 1960s Chomsky and co-workers pioneered work in developing a formal theory of language (Chomsky 1959, 1963, 1965); this work established clearly-defined links between linguistics, logic and mathematics, and was also foundational in computer science in the sense that the principles it established have also been applied in understanding computer programming languages. These advances make it possible to formulate a theory of Russian phonology, just as Halle did, but to test it empirically by implementing the theory as a computer program and using it to process very large numbers of words. Moreover, since the technological advances which make it possible to do this owe a great deal to Chomsky’s work, the transition from generative grammar to computational grammar can be a comparatively straightforward one. One of the defining features of generative grammar is the emphasis on searching for cross-linguistic patterns. Without denying the value of language-specific grammar, Chomsky and Halle (1968) (to many the canonical work of generative phonology) illustrates this thinking: ...we are not, in this work, concerned exclusively or even primarily with the facts of English as such. We are interested in these facts for the light they shed on linguistic theory (on what, in an earlier period, would have been called “universal grammar”) and for what they suggest about the nature of mental processes in general… We intend no value judgment here; we are not asserting that one should be primarily concerned with universal grammar and take an interest in the particular grammar of English only insofar as it provides insight into universal grammar and psychological theory. We merely want to make it clear that this is our point of departure in the present work; these are the considerations that have determined our choice of topics and the relative importance given to various phenomena. (p. viii)

11

The emphasis on cross-linguistic generalization, characteristic of Chomsky’s work, has characterized generative linguistics ever since: indeed, there is a considerable branch of linguistics (Zwicky 1992 is an example) which abstracts completely away from language-specific data. (This branch deals in what Zwicky 1992: 328 refers to as ‘frameworks’ as opposed to ‘theories’.) While frameworks have their place (indeed, a theory cannot exist without a framework), the difficulty is always that frameworks cannot be verified without theories. In this light, Chomsky and Halle (1968) claimed to establish both a cross-linguistic framework and a theory about English phonology. The focus of this description is on ensuring that the phonology of Russian proposed is both internally consistent and descriptively adequate & that is, that it makes empirically correct predictions about Russian & rather than on attempting to develop any particular linguistic framework. Exciting possibilities are open in this line of research thanks to the existence of computer technology. It is possible to state grammatical rules in a form which has the rigour required of a computer program, and once a program is in place, large corpora can be quickly processed. Thus the phonology of Russian presented here is ‘computational’ simply because of the advantages in speed and coverage that this approach presents. Establishing that a linguistic theory can be implemented as a computer program and verifying its internal consistency in this way is a valuable exercise in itself, but non-computational linguists may be sceptical: some may argue that this kind of approach does not contribute anything to linguistics per se. Whether or not this is criticism is well-founded (and I believe it is not), I hope that this dissertation

12 will satisfy even the more stringent critics by making a number of key contributions to linguistic knowledge. These are as follows. First, I propose that both the distribution of !"! and its behaviour with respect to voicing assimilation can be explained if !"!, unlike all other segments in the phonological inventory of Russian, is lexically unspecified for the feature [!sonorant]. The syllable structure rules determine whether !"! is [+sonorant] or [$sonorant], and this in turn determines how !"! assimilates in voice to adjacent segments. Second, I suggest that the greater latitude allowed in word-marginal onsets and codas, which is a feature of Russian and other languages (cf. Rubach and Booij 1990), can be explained naturally by a probabilistic syllable structure grammar. This approach allows features such as [!initial] and [!final] (cf. Dirksen 1993) to be dispensed with. Third, I show that vowel-zero alternations in Russian cannot fully be explained by a Lexical-Phonology-style account (such as that proposed by Pesetsky ms 1979) alone, nor can they be the result of epenthesis alone. I show empirically that a combination of factors, including (1) the morphophonological principles discovered by Pesetsky, (2) epenthesis, and (3) etymology, governs vowel-zero alternations. Fourth, I show that Russian stress can be accounted for with a high rate of accuracy by existing generative theories such as that of Melvold (1989), but I suggest a simpler theory which accounts for the same data with as good a rate of accuracy. The theory which I propose regards stress assignment as resulting from the interaction of morphological and syllable structure: existing generative theories do not acknowledge syllable structure as playing any role in Russian stress assignment. An integral part of my theory is a comprehensive inventory of morphemes together with

13 the accentual information which is lexically specified for each morpheme. The inventory which I propose, which is arrived at by computational inference, includes 1,094 morphemes and 1,509 allomorphs, while the longest existing list of this type, as far as I am aware, is the index of approximately 250 suffixes in Red’kin (1971). The structure of this dissertation is as follows. In this chapter, I set out in detail the concepts which are foundational to the whole work: the role which computation plays in my work (1.2), the framework which I use (1.3), and the methodology which underlies my work (1.4). Then, I discuss in detail aspects of the syllable structure and morphological structure of Russian in Chapters 2 and 3 respectively, in each case developing a formally explicit grammar module which can be shown to be equivalent to a finite state grammar. Chapter 4 describes in detail three theories of stress assignment in Russian. These are tested computationally to ascertain which is the most promising. Each of Chapters 2-4 begins with a section reviewing the relevant literature. Finally, in Chapter 5, I describe how the principal features of the preferred theory from Chapter 4 can be incorporated into a synthesis of the grammars developed in Chapters 2 and 3. The result is an integrated, internally consistent, empirically well-grounded grammar, which accounts for a variety of different aspects of Russian phonology. 1.2

Why ‘computational’ linguistics? In this dissertation, computation is used as a tool. Any tool has limitations, of

course: a large building cannot be built with a power drill alone, and, to be sure, there are problems in linguistics which computation is ill-suited to solve. On the other hand, anyone who has a power drill will try to find appropriate uses for it. Likewise, I aim to use computation for the purposes for which it is best suited. This, then, is not a

14 dissertation ‘about’ computational linguistics; it is a dissertation that uses computation as a tool in linguistics. What, then, are the strengths of computational tools in linguistics? Shieber (1985: 190-193), noting that the usefulness of computers is often taken for granted by computational linguists, lists three roles that the computer can play in the evaluation of linguistic analyses: the roles of straitjacket (forcing rigorous consistency and explicitness, and clearly delineating the envelope of a theory), touchstone (‘indicating the correctness and completeness of an analysis’), and mirror (‘objectively reflecting everything in its purview’). In short, the process of implementing a grammar computationally forces one to understand in detail the mechanisms by which a grammar assigns structure. Shieber states, for example, that …we have found that among those who have actually attempted to write a computerinterpretable grammar, the experience has been invaluable in revealing real errors that had not been anticipated by the Gedanken-processing typically used by linguists to evaluate their grammars & errors usually due to unforeseen interactions of various rules or principles. (p. 192)

This has also been my experience in developing the current phonology of Russian. In particular, areas such as stress assignment involve the interaction of a number of different grammar modules, and, as Shieber states, ‘decisions in one part of the grammar, while internally consistent, may not cohere with interacting decisions in another part’ (Shieber 1985: 190). Problems of this kind cannot always feasibly be foreseen without actually implementing and testing a theory on a corpus of data. Another perhaps self-evident strength of computers is their ability to process large volumes of data quickly: once a grammar has been implemented, the processing can take place without intensive effort on the part of the researcher. While in principle

15 generative theories can be implemented and tested by hand, the volume of data that typically has to be processed to achieve significant results means that this is an extremely tedious and time-consuming, if not impracticable, task. Clearly, computational techniques shift the burden for the researcher from data processing to the more interesting task of developing theories, identifying exceptions quickly, and debugging the theory as appropriate. Because the discipline of computational linguistics is still relatively young, it is perhaps understandable that many existing theories have neither been implemented nor tested computationally, but now that the means to validate theories are widely available, it is less justifiable for new theories still to be proposed in linguistics without being empirically tested: ‘the widespread practice of testing a few interesting cases is unreliable and is no substitute for an exhaustive check’ (Bird 1995: 14). It seems that at this stage in linguistic research, the efforts of linguists would be better directed towards implementing and testing existing theories rather than proposing new alternatives, since otherwise it cannot be demonstrated that the new alternatives measure up any better to the criteria of coverage, constrainedness and ability to integrate than the theories which they replace. It is also worth noting the limitations of computational analysis (which I set as the limits for this dissertation). Ultimately, computers follow instructions rather than making judgements, and while they are very good at evaluating grammars for consistency and descriptive adequacy, they cannot test for explanatory adequacy unless the programmer supplies the necessary information (that is, a standard against which to measure the accuracy of structures assigned by a grammar to strings). The judgement about the nature of the correct structures is a question of psychology, and

16 therefore I do not claim that the current phrase-structure context-free phonology of Russian is a psychological model. In this, my approach is exactly the same as that of Gazdar, Klein, Pullum and Sag (1985): We make no claims, naturally enough, that our grammatical theory is eo ipso a psychological theory. Our grammar of English is not a theory of how speakers think up things to say and put them into words. Our general linguistic theory is not a theory of how a child abstracts from the surrounding hubbub of linguistic and nonlinguistic noises enough evidence to gain a mental grasp of the structure of a natural language. Nor is it a biological theory of the structure of an as-yet-unidentified mental organ. It is irresponsible to claim otherwise for theories of this general sort… Thus we feel it is possible, and arguably proper, for a linguist (qua linguist) to ignore matters of psychology. But it is hardly possible for a psycholinguist to ignore language… If linguistics is truly a branch of psychology (or even biology), as is often unilaterally asserted by linguists, it is so far the branch with the greatest pretensions and the fewest reliable results… So far, linguistics has not fulfilled its own side of the interdisciplinary bargain. (p. 5)

1.3

The framework

1.3.1

Phrase-structure grammar In this dissertation, phonology and morphology, as modules of grammar, have

the function of enumerating or generating (the words of a) language. This view of grammatical modules is entirely in accordance with traditional generative linguistics (e.g. Chomsky and Miller 1963: 283-285). More precisely, a phonological grammar should be able to generate all and only the phonological words of a natural language; similarly, a word-formation grammar should enumerate all the morphological words (p-forms, in the terminology of Zwicky 1992: 334) of a natural language.1 The same

1

As noted by Booij and Rubach (1984), there may well not be a one-to-one mapping between ‘morphological words’ and ‘phonological words’ & well-known examples from Russian are preposition-noun phrases, all of which have a single stress (e.g. 9&: !8%8 !*/% 1, are always decidable. Bresnan and Kaplan argue that natural languages must be decidable, since:

23 It is plausible to suppose that the ideal speaker can decide grammaticality by evaluating whether a candidate string is assigned (well-formed) grammatical relations or not. The syntactic mapping can thus be thought of as reliably computing whether or not any string is a well-formed sentence of a natural language. This motivates the reliability constraint that the syntactic mapping must provide an effectively computable characteristic function for each natural language. (p. xl)

The principal objection which has been raised to this assumption, and one which is noted by Bresnan and Kaplan, is that native speakers often do not do well at parsing ‘garden path’ constructions such as The canoe floated down the river sank and The editor the authors the newspaper hired liked laughed. However, they suggest, plausibly, that these constructions do not disprove their hypothesis. After all, they argue, speaker-hearers can disambiguate these sentences and ‘recover from the garden paths’, given more (but not infinite) time, and possibly a pencil and paper. A third reason for choosing the formalism of CFG is that the ordering of the rules of CFGs will not affect the way in which they function or their end result (although the ordering of application of rules may have an effect on the outcome). All forms and constraints in CFGs are partial descriptions of surface representations, no rules do not ultimately constrain surface forms, all constraints must be compatible and apply equally, and any ordering of constraints will describe the same surface form (Scobbie, Coleman and Bird 1996). The motivation for this Order-free Composition Constraint, as Bresnan and Kaplan (1982: xlv) call it, is ‘the fact that complete representations of local grammatical relations are effortlessly, fluently, and reliably constructed for arbitrary segments of sentences’ (Bresnan and Kaplan 1982: xlv). Again, this does not hold for all types of grammar. There are thus a number of reasons why it is desirable to restrict a grammar so that it is no more powerful than context-free. To summarize, these are as follows:

24 ,

CFGs are a relatively restricted class of grammar, and we would like to choose the most restricted theory which will account for the facts;

,

CFGs have a generative as well as a recognitive interpretation;

,

CFGs are Turing-decidable;

,

the rules of CFGs need not be ordered in any particular way;

,

although CFGs have been shown to be unable to cope with all aspects of syntax, there is no evidence to suggest that they are insufficient as far as phonology is concerned.

1.4

The methodology Generative linguists often claim that linguistics is a ‘science’. This claim is

made for phonology, for example, in Halle (1959: 24). What is meant by this? Sommerstein (1977: 9) answers this question with respect to phonology as follows: In science we frame and test hypotheses. It does not matter in the least how these hypotheses are arrived at in the first place; it is the exception rather than the rule for an interesting hypothesis to be reached by a mechanical procedure, such as phonemic analysis essentially is. Rather, what makes a hypothesis scientific or unscientific is whether it can be stated what kind of empirical evidence will tend to disconfirm it, and what kind will definitely refute it. And there is no reason why this general scientific principle should not be valid for phonological analysis.

Thus any grammar we propose has the status of a scientific theory that attempts to account for observed linguistic data. On a philosophical level, the data exist independent of any grammar; in other words, the existence of sentences, words, etc., in a language does not depend on our ability to formulate grammar rules to account for them. The only way of determining how well a grammar really does fit the data is to test it empirically. One way in which scientific methodology can work is ‘incrementally’: we look at the cases where a theory does not fit the data and modify

25 the theory accordingly. One would hope that the coverage of each successive theory advanced using this kind of methodology would eventually approach 100%. I shall now elucidate what is meant here by the ‘coverage’ of a linguistic theory. As we saw in 1.3.1, a given grammar may be descriptively but not explanatorily adequate, but the converse is not possible. It may also be neither descriptively nor explanatorily adequate, which means that it fails altogether to assign a structural description to some utterances. For an imperfect grammar of this type, the set of correctly parsed utterances will be a subset of the set of parsed utterances, which in turn will be a subset of the set of all utterances, as Figure 2 illustrates. Figure 2. Classification of analyses of an imperfect grammar P: All words

Q: Words assigned some structural description R: Words assigned the correct structural description

There are three measures that we shall be interested in. The first of these is coverage, the number of utterances in Q as a percentage of the number of words in P. The second is correctness or structural coherence, the number of utterances in R as a percentage of the number in P. The third is the number of utterances in R as a percentage of the number in Q. Arguably, the second of these is the best overall

26 measure, but as we do not always have access to data which tell us what the ‘correct’ structures are, in some cases we have to use the first instead. The third measure will be most relevant in Chapter 5, where we need to separate the issues of morphological structure and stress assignment in order to be able to do a like-for-like comparison between the phrase-structure phonology proposed in this dissertation and Melvold’s theory. The methodology that underlies the current work is also ‘incremental’. In subsequent chapters I advance theories about the syllable structure and morphological structure of Russian words which are arrived at by trial and error: see, for example, (91) in 3.2. The process of actually checking the descriptive adequacy of a grammar is straightforward and well-suited to computational processing, since the latter is fast and reliable. 1.5

The dataset used for the tests In order to test a grammar computationally, it is necessary to have some kind

of lexical database which one can use as the dataset for the tests. As a minimum, the database used in the research described here has to give the following information for every word therein: ,

A phonological transcription

,

The position of the word-stress

,

The position of all morpheme boundaries within the word

,

The part of speech Additional information which would have been desirable for each word in the

corpus, but was unobtainable on a systematic basis, was as follows:

27 ,

A phonetic transcription

,

The position of all syllable boundaries within the word Although there are many existing electronic corpora for different languages

(including Russian), the requirements of the research described in this dissertation were such that no existing electronic corpus was adequate for the purpose. Thus part of the preliminary work necessary was to compile a purpose-made lexical database. In this section, I discuss how I did this. Oliverius (1976) contains a list of 2,493 morphologically tokenized words. However, these words are all headwords. There are two major reasons why it is desirable to extend Oliverius’s list to include inflected forms. First, if the dataset is restricted to the words in Oliverius (1976), a large portion of the vocabulary of Russian (all the inflected forms) is missed. This is unacceptable because the current dissertation explicitly deals with the application of phonological theories of Russian to the output of both derivation and inflection. Secondly, the larger the dataset used as the basis for testing theories, the greater the level of significance the results will have. One way of computationally extending the list to include inflected forms would be to compute the inflected forms (with stress) from the head-words and information about their stress patterns. This information can all be found in Zaliznjak (1977), which is available in an electronic version. A program could be written to go through the list of words in Oliverius (1976), matching each to the relevant entry in Zaliznjak (1977), and generating the appropriate inflected forms. Although it could be automated, even this approach would be a large undertaking, primarily because of the thoroughness of Zaliznjak’s description: the key in Zaliznjak which explains the meanings of the tags to each entry takes up a significant amount of space in the

28 dictionary (132 pages). This information is not included in the electronic version, and it would all have somehow to be input manually if the inflected forms of all entries in the dictionary were to be generated computationally. Fortunately, however, this was unnecessary. One of the products of the research carried out by Brown and his colleagues at the University of Surrey (Brown, Corbett, Fraser, Hippisley and Timberlake 1996) is a ‘theorem dump’ listing the inflected forms of 1,536 nouns. This file includes comprehensive information about word-stress, but the words are only partly morphologically tokenized (since steminflection but not stem-internal morpheme junctions are given). In order to ensure that all possible forms from the University of Surrey theorem dump were fully morphologically tokenized, each headword from the theorem dump was matched to headwords from Oliverius (1976) and the morphological tokenization of inflected forms was extrapolated from the morphological tokenization of the headword, by the procedure outlined in (1):

(1)

(a)

For each headword (e.g. #8!&% !$;0%2! ‘fool’) in Oliverius (1976), find whether it is a noun by searching through the on-line version of Zaliznjak (1977), which provides part-of-speech information.

(b)

For each (headword) noun identified by (a), search for all related inflected forms in the theorem dump. For #8!&% !$;0%2! ‘fool’, these would be as follows: #8!&%& (gen. sg.) #8!&%8 (dat. sg.) #8!&%,+ (instr. sg.) #8!&%" (loc. sg.)

#8!&%) (nom./acc. pl.) #8!&%,( (gen. pl.) #8!&%&+ (dat. pl.) #8!&%&+) (instr. pl.) #8!&%&: (loc. pl.)

29 (c)

Pair the headword with its morphological tokenization, which is known from the information in Oliverius (1976) (for example, !$;0%2! would be paired with the tokenization !$;0ra+%2sn+in1! 4), and deduce the noun-stem by removing the inflectional ending (in this case, zero). The noun-stem of !$;0%2! would thus be !$;0ra+%2sn!.

(d)

Morphologically parse the inflected forms using the parsing information about the stem from (c), and parsing whatever is to the right of the stem as the inflectional ending. In this example, the inflected forms would be parsed !$;0ra+%2sn+%in!, !$;0ra+%2sn+;in!, !$;0ra+%2sn+6+in!, etc. More detailed information on how inflectional endings are categorized and distinguished is given in section 3.2.1.2.

The procedure in (1) was automated, except in the case of nouns which exhibit a vowel-zero alternation within the stem (such as ,%', !62/46! ‘window’ [nom. sg.], ,%,' !/6264! ‘windows’ [gen. pl.]). The morphological tokenization for these forms was input manually. As it turned out, 967 of the 2,493 words in Oliverius (1976) were nouns; 835 of these were included in the theorem dump. Some of these nouns are identified by the theorem dump as having incomplete paradigms, so the number of inflected forms including head-words identified by step (b) of (1) was 9,633 (slightly less than 12 835 = 10,020).

4

The notation is explained fully in section 3.2.1.2.

30 The morphologically parsed inflected forms were combined with the rest of the morphologically parsed head-words in Oliverius (1976), giving a sample of fully morphologically parsed words as in Table 2. Table 2. Analysis of words in on-line corpus Category

Head-words or inflected forms

In Oliverius (1976)

In theorem dump

Non-nouns Nouns Nouns Nouns

Head-words Head-words Head-words Inflected forms

   -

 

Number of words

Total

1,525 132 835 8,798 11,290

Regrettably, the on-line corpus of 11,290 word-forms does not include any inflected forms for non-nouns, which means that the results presented in this dissertation will have greatest weight in their applicability to nouns. But it would not be fair to say that this dissertation is limited in its scope to nouns, because, as can be seen from Table 2, the number of non-nouns is great enough that statistically significant results can still be achieved. When more comprehensive electronic corpora of Russian become available, it will certainly be interesting to see whether re-running some of my tests on these corpora gives results in line with those I report here; presumably, the null hypothesis would be that this will be the case. 1.6

Summary In this chapter, I have established the approach which I employ in developing

a computational phonology of Russian, and dealt with various issues relating to my perspective. To summarize, the aim in subsequent chapters is to formulate a broad-

31 coverage phonology, which is generative, context-free, coherent, and makes predictions that can be shown empirically to be correct, or at least a good first approximation at correctness. To the extent that this aim succeeds, this work will fill an important gap in the literature to date, as no other work of which I am aware meets all these criteria simultaneously.

32

Chapter 2: Syllable structure 2.1

Overview and aims This chapter presents a sonority-based syllable structure grammar of Russian.

As well as aiming to advance a specific proposal about Russian, I also aim in this chapter to contribute to the general debate on syllabification in two ways. First, because the grammar is implemented as a Prolog DCG and tested for its coverage of a corpus of Russian words, I am able to identify a list of exceptions to the Sonority Sequencing Generalization (SSG), which is widely accepted in one form or another as the standard means of accounting for phonotactic constraints. The list of exceptions is comprehensive with respect to the dataset tested, so the test allows us to quantify precisely how problematic Russian is for the SSG. Secondly, we shall see further evidence that it is worthwhile to include a formal definition of the term ‘syllable’ in a phonology, as Fudge (1969) suggests: it is not enough to refer to the syllable without explicitly defining it, as in Chomsky and Halle (1968). The syllabification grammar outlined here is put to work in a variety of areas of Russian phonology: it takes on a role as a structure in which to apply phonotactic constraints, a role familiar from Kahn (1976); it is also the structure for the implementation of other phonological constraints, such as assimilation, word-final devoicing, consonant-vowel interdependencies and vowel reduction; and, as will become apparent in Chapter 5, it takes on a novel role in stress assignment (novel, because no other treatment of Russian stress hinges on syllable structure in the way which I suggest).

33 To my knowledge, there are no comprehensive treatments of Russian syllable structure comparable to the one proposed in this chapter. Bondarko (1969) is a proposal, based on experimental measurements of relative durations of consonants and vowels in the speech chain, that all consonants (and consonant clusters) in Russian (except for cluster-initial !7!) syllabify together with the following vowel, meaning that almost all Russian syllables are open. If this is true, this would amount to a comprehensive proposal on Russian syllable structure, but the problem with Bondarko’s proposal is that it says nothing about the kinds of clusters that cannot occur syllable-initially. In other words, the evidence that Bondarko examines excludes evidence about the phonotactic constraints of Russian: for example, Bondarko’s theory does not explain why no Russian words begin with !42!. This kind of consideration is the starting-point of this dissertation; after all, a generative grammar must be able not only to assign syllable structure, but also to generate legal structures and rule out illegal ones. Thus the grammar I propose, contrary to Bondarko (1969), suggests that a number of different types of closed syllable can occur in Russian. The remainder of this chapter is organized as follows. Section 2.2 reviews the literature on syllable theory. Sections 2.3-2.4 describe a phrase-structure sonoritybased theory about Russian syllable structure. This theory is a linear (i.e. Type 3) grammar, with all the advantages this brings (see section 1.3.2). However, the nature of the constraints needed to account for Russian syllable structure is far from obvious. The primary aim of the discussion in sections 2.3-2.4 is to establish what these constraints are, rather than debating the issue of how syllable structure is assigned. I then move on in section 2.5 to select four key aspects of Russian phonology which have attracted attention in the literature: the constraints on the consonant clusters

34 permissible in Russian, assimilation in voicing and palatalization, word-final devoicing and reduction of unstressed vowels. For each of these problem areas, I set out what appear to be the facts as generally accepted: the aim of this section is to show that these facts need not be seen as divorced from syllabification, but an account of them can be integrated into the existing PSG of Russian syllable structure. Indeed, in some cases, there is a clear advantage in this kind of integration. For example, the properties of !"! with respect to voicing assimilation are most simply explained by taking into account the features which the syllable structure grammar assigns to !"!. The result is that a single grammar fulfils a variety of functions, assigning syllable structure, mapping phonemic representations to phonetic representations, and, as we shall see in Chapter 5, acting as an indispensable component in a theory about stress assignment in Russian. 2.2

The syllable in phonological theory The syllable is by no means a recent construct. It was discussed as a unit of

linguistic organization in, for example, Whitney (1865), Sievers (1881), Jespersen (1904), de Saussure (1916), Grammont (1933), Bloomfield (1933) and Hockett (1955). Bloomfield, for example, states that ‘the ups and downs of syllabication play an important part in the phonetic structure of all languages’ (p. 121; Bloomfield’s emphasis). It was in the 1950s and 1960s that the status of the syllable was both implicitly and explicitly questioned in generative phonology: implicitly, by its notable

35 absence in Halle (1959)5 and Chomsky and Halle (1968), and explicitly, in Kohler (1966: 346-348). As Fudge (1969: 261-262) points out: Chomsky and Halle (1968) continually invoke syllables, monosyllables, disyllables, etc. in their less formal discussions (in the text frequently, but sometimes also within the systems of rules proposed), and even postulate a feature Syllabic “which would characterize all segments constituting a syllable peak” (354). Unfortunately, none of these terms are made explicit in the text or in the rules… The term “syllable” does not even figure in the index of Chomsky and Halle (1968). In fact, we may state that it is not satisfactory to deal with the structure of one element in terms of statements designed to deal with the structure of an essentially different and only indirectly related element. If we want to state syllable-structure, we must explicitly introduce the element syllable into our linguistic description, and state its relations to the other elements of the phonological hierarchy; it is precisely this which Chomsky and Halle (1968) fail to do.

From that time, partly as a reaction to Chomsky and Halle’s work, phonological theory has swung back towards endorsing the syllable. Indeed, even before Halle (1959), Haugen (1956: 215-216) writes of the syllable that ‘one would be tempted to deny its existence, or at least its linguistic status, as some have done, were it not for its wide persistence as a feature of most linguistic descriptions… those who attempt to avoid the syllable in their distributional statements are generally left with unmanageable or awkward masses of material’. This shortcoming of Chomsky and Halle’s theory is pointed out not only by Fudge (1969), who argues that the element ‘syllable’ should be made explicit, but also by Hooper (1972) and Vennemann (1972); the latter uses evidence from languages other than English to ‘advocate… the incorporation of syllable boundaries and syllables in phonological descriptions’ (p. 2). Perhaps the best-known work pointing out the inadequacies of Chomsky and Halle (1968), though, is Kahn (1976): Kahn states that in describing productive 5

For further discussion of the absence of the syllable in Halle (1959), see section 2.2.2.

36 phonological processes he was ‘hampered by the absence of a generative theory of syllabification’ (p. 17). Kahn observed, in particular, that the phonotactic constraints of English could be accounted for indirectly but simply by considering syllable structure (pp. 40-41, 57-58). Clements and Keyser (1983), endorsing Kahn’s hierarchical analysis of the syllable, argued however that syllabicity was not a property of segments per se as Kahn suggested (Kahn 1976: 39), ‘but rather involves the relationship between a segment and its neighbors on either side’ (Clements and Keyser 1983: 5): to account for this, they proposed analyzing syllables in terms of three tiers, the syllable tier and segmental tier (as in Kahn 1976) and an additional CV tier. Selkirk (1984) follows Clements and Keyser in rejecting [!syllabic] as a feature of segments. Despite the criticisms of certain aspects of Kahn’s approach, it has generally been acknowledged since Kahn (1976) that the syllable is an indispensable unit of linguistic organization. For example, a contemporary description of Slavic prosody, Bethin (1998), makes the following statement: We find that many prosodic features are restricted to or expressed on syllables, that certain restrictions on permissible consonant and vowel sequences are best described as holding within a syllable, that there are phonological and morphological processes which seem to be conditioned by the syllable, and that many of these processes count syllables but do not, as a rule, count phonemes or segments. (p. 192)

It seems, therefore, that the syllable is here to stay in linguistic theory, and in particular that an account of syllable structure is an essential part of a generative phonology of Russian. One aim of this chapter, therefore, is to put forward a specific grammar of Russian syllable structure as part of the overall phonology proposed in this dissertation. This grammar is explicit about what Russian syllables are; it does

37 ‘state its relations to the other elements of the phonological grammar’, as Fudge puts it, and because the theory is implemented computationally and tested for its coverage, a standard is set against which future proposals can be measured. 2.2.1 Sonority and syllable structure The notion of the syllable is inextricably linked to that of ‘sonority’, which has for more than a century been believed by phonologists to be an important factor in the structure of syllables (Whitney 1865, Sievers 1881: 159-160, Jespersen 1904: 186187, de Saussure 1916: 71ff. and Grammont 1933: 98-104). Essentially, the idea is that segments can be categorized with respect to sonority: those that are more sonorous tend to stand closer to the centre of the syllable, and those that are less sonorous closer to the margin. Clements (1990: 284) notes that ‘this principle… expresses a strong cross-linguistic tendency, and represents one of the highest-order explanatory principles of modern phonological theory’. However, there are a number of questions about sonority which to date have not been answered. Essentially, these have to do with (a) how sonority is defined, and (b) at what linguistic level sonority holds (Clements 1990: 287, Bethin 1998: 19-21). As far as the first of these is concerned, there have been various attempts at defining sonority. Bloomfield (1933: 120-121) equated sonority with the ‘loudness’ of segments (the extent to which some sounds ‘strike the ear more forcibly than others’); another proposal is that sonority can be derived from basic binary categories, identical to the major class features of standard phonological theory (Selkirk 1984, Clements 1990); and some have suggested that sonority does not have any absolute or consistent phonetic properties (e.g. Hooper 1976: 198, 205-206). Even ignoring the question of how sonority is defined phonetically, there is disagreement on what the

38 sonority hierarchy is; and until this issue is resolved, as Selkirk points out, discovering the phonetic correlates of sonority will be difficult. For example, Clements (1990: 292-296) proposes a hierarchy where obstruents are less sonorous than nasals, nasals less sonorous than liquids, and liquids less sonorous than glides. Glides, in turn, are seen as non-syllabic vowels. On the other hand, Selkirk (1984: 112) sets out a more detailed hierarchy, as follows: (2)

Sounds (in order of decreasing sonority) % 1=6 ,=; 0 +=4 # "=*=> ?=@ A=$=B 9=&=2

Whatever the exact classification of sounds by sonority, it seems to be a general rule that for each peak in sonority in a string of phonemes, there will be a syllable (Bloomfield 1933, Selkirk 1984, Clements 1990). Perhaps the best-known formulation of the sonority principle is Selkirk’s (1984: 116) ‘Sonority Sequencing Generalization’ (SSG): In any syllable, there is a segment constituting a sonority peak that is preceded and/or followed by a sequence of segments with progressively decreasing sonority values.

39 In this formulation, the syllabicity of segments depends on their position, rather than on any inherent phonological property of their own (Clements and Keyser 1983: 4-5, Selkirk 1984: 108, Blevins 1995, Bethin 1998): sonority peaks simply align with syllable peaks. This offers an explanation in terms of syllable structure for the fact that glides and approximants can function as either consonants or vowels, a fact that was noted as early as Sievers (1881: 157). Clements (1990: 287) points out, however, that this principle is not without exception: for example, in US English yearn, the liquid (0) is the syllable peak: however, as it is preceded by a glide, it does not constitute a sonority peak. The sonority principle is also used to account for the phonotactic constraints that apply in onsets and codas: sonority must increase during syllable onsets, and decrease during syllable rimes. Thus, according to Clements’s and Selkirk’s formulations of the SSG, !+9! would be ruled out as a syllable onset, but accepted as a coda, because !+! is a nasal and !9! is a (less sonorous) occlusive. Presumably, therefore, languages in which !+9! does occur as an onset are considered exceptional to the SSG. A further refinement to the SSG specifies a ‘minimum sonority difference’ that must obtain between two adjacent positions in a syllable: Harris (1983: 21) suggests that this minimum might vary from language to language. For instance, Spanish requires adjacent consonants in the same syllable to be non-adjacent on the sonority scale Obstruent < Nasal < Liquid < Glide < Vowel; thus, Spanish allows only a subset of the rising-sonority clusters as onsets. The second question, regarding the level at which sonority holds, is addressed by Clements (1990: 323). Clements asks whether the SSG has the status of a violable

40 cross-linguistic default, or whether it is a firm principle at some abstract level. He suggests the latter6. The SSG would certainly be of more phonological interest if it turned out that Clements was right, and a major challenge for a computational phonology of Russian is to find out whether the SSG can be implemented as a ‘firm’ principle. It has been my experience, however, that explaining away all the ‘exceptions’ is extremely hard. One approach may be to acknowledge, as Clements (1990: 290) does, that ‘a hard core of irreducible exceptions will [always] remain’, but this clearly weakens the force of the SSG. I shall return to this point at the end of section 2.3.2. 2.2.2

Morpheme structure constraints or syllable structure constraints? We now take a detailed look at one of the proposals of Halle (1959) in the

light of the SSG: as mentioned earlier, neither Halle (1959) nor Chomsky and Halle (1968) explicitly define the syllable, let alone sonority. According to Halle (1959), the constraints on sequences of segments (phonotactic constraints) apply instead within the domain of the morpheme. (Halle was not the only researcher to suggest the morpheme as a relevant domain for phonotactic constraints: for example, Isa;enko 1970 follows a similar approach, as do philologists: see for example Benveniste 1935). Table 3 shows Halle’s specific constraints on consonant clusters.7 Halle’s analysis could have been simplified by considering syllable structure as well as morpheme structure, as we shall now see.

6

Despite his adherence to sonority as a ‘firm’ principle, Clements is still forced to admit some crosslinguistic variation in the periphery of the syllabification domain. 7 Information on fully-specified versus incompletely-specified morphonemes, which is included by Halle, is omitted in the table as it is not relevant to the present discussion.

41 Table 3. Russian morpheme structure constraints on consonant clusters (adapted from Halle 1959: 57) Type of cluster 4 segments: CCCR RCCC 3 segments: CCR CCC RCR RCC 2 segments: CC RC JC CR RR CJ RJ

Example in morpheme-initial position

Example in morpheme-medial position

Example in morpheme-final position

{fstr’et,i} ‘to meet’ not attested

not attested not attested

not attested {;’orstv} ‘stale’

{skr,ip} ‘squeak’ {stvol} ‘gun barrel’ not attested not attested

{kastr,’ul,} ‘saucepan’ not attested {v,irbl,’ud} ‘camel’ not attested

{z’atxl} ‘musty’ {’opš;} ‘common’ not attested {tolst} ‘fat’

{svet} ‘light’ {rt’ut,} ‘mercury’ not attested {slep} ‘blind’ not attested {djak} ‘clerk’ {rjan} ‘zealous’

{’asp,id} ‘slate’ {alm’az} ‘diamond’ {bajb’ak} ‘marmot’ {utr’ob} ‘womb’ {jArl’ik} ‘label’ {ab,izj’an} ‘monkey’ {burj’an} ‘tall weeds’

{kost,} ‘bone’ {sm’er;} ‘cedar’ {ajv} ‘quince’ {žezl} ‘staff’ {gorl} ‘throat’ not attested not attested

Key to Halle’s symbols used in Table 3 R stands for any liquid, C stands for any consonant, J stands for the glide !7!, A (in {jArl’ik} stands for a ‘further unspecified non-diffuse vowel’, i.e. the archiphoneme representing the set {e,o,a} (ibid., p. 75; the statement on p. 57 that A stands for a diffuse vowel is obviously an error).

First, if every consonant is associated with one or another syllable, then we need no longer consider ‘medial’ clusters, those represented in the second column of Table 3. These clusters will either become syllable-initial or will split so that the first part of the cluster is syllable-final, and the rest of the cluster is syllable-initial (within the following syllable). Table 4 shows how the syllabified versions of the morphemes in the ‘medial cluster’ column would look, assuming that consonants always syllabify with the following nucleus where possible in accordance with the Maximal Onset Principle (see e.g. Kahn 1976, discussed in 2.2.3.1 below). Table 4 also shows how each of the examples can then be related to existing initial and final clusters within

42 Table 3. Once this has been done, there is effectively no need for the middle column in Table 3. Table 4. Reanalysis of morpheme-medial clusters using syllable structure Example from Table 3

Reanalysis after syllabification

{ka . str,’ul,} {v,ir . bl,’ud} {’a . sp,id} {al . m’az} {baj . b’ak} {u . tr’ob} {jAr . l’ik} {a . b,i . zj’an} {bu . rj’an}

Initial CCR, cf. {skr,ip} Final R, initial CR, cf. {slep} Initial CC, cf. {svet} Final R, initial C Final J, initial C Initial CR, cf. {slep} Final R, initial R Initial CJ, cf. {djak} Initial CJ, cf. {djak}

Now consider clusters of the types CJ, RJ, and JC. Table 3 shows that CJ and RJ occur initially but not finally, and JC occurs finally but not initially. The SSG, of course, has an explanation for this: the segment-types C and R are lower in sonority than segment-type J, thus CJ and RJ rise in sonority, while JC falls in sonority. Of course, this is not to suggest that all the examples in Table 3 are explained by the SSG. For example, taking just the liquid R into account, it is not surprising that CCCR should be attested only initially, while RCCC and RCC are attested only finally, but this does not help account for the adjacent C segments. Even more challenging for the SSG are the instances of final CCR as in {z’atxl} (although note that Table 3 lists morphemes, not word-forms, and it might be questioned whether 9&.:6 !*%&C-! ‘musty’ (attributive) really exists as a word-form in Russian), final CR as in {žezl},

43 and initial RC as in {rt’ut,}.8 Although it is not clear that sonority and syllable structure account for all the material presented by Halle, it does seem that they allow some significant simplifications to be made to his analysis. Finally, it should be said that this need not be taken as a denial that some phonotactic constraints are morphologically determined: for example, there is a clear principle in Russian and other Slavic languages that nominal roots always end in a consonant (Isa;enko 1970: 88, Townsend 1975: 15). It probably is the case that there are both morphological and syllabic constraints on phonotactics. However, there is clearly a useful role for the syllable, which Halle (1959) ignores. 2.2.3 Syllable structure assignment With the renewed interest in the syllable in the 1970s came a debate in phonological theory over the mechanism by which syllable structure is assigned. ‘Syllable structure assignment’ does not equate to ‘syllable structure constraints’, although the two are usually intermingled in the debate on syllable structure assignment. For example, Kahn (1976) makes the following statement: The system of rules assigning syllable structure to strings of segments, as envisioned here, does not refer back to some general set of constraints on possible word-initial and -final clusters which is pervasive throughout the phonology. It is rather in the syllable-structure assignment rules themselves that these constraints are found. (p. 45)

8

Although see Rubach and Booij (1990) for arguments that Polish onsets and codas, which contains many clusters similar to those in Russian, are all constrained by the SSG. The application of Rubach and Booij’s analysis to Russian within the current framework is drawn out in more detail below: see pp. 79ff.

44 Blevins (1995: 222) argues that the most basic division between the different ideas on syllable structure assignment is between ‘rule-based’ approaches (Kahn 1976, Steriade 1982, Clements 1990 and Rubach and Booij 1990), and ‘templatematching’ or ‘templatic’ approaches (Selkirk 1984, Itô 1986, Archangeli 1991). Blevins explains the distinction as follows: Template-matching algorithms for syllabification scan the segmental string in a fixed, language-particular direction (left to right, right to left), assigning successive segments to positions in a syllable template, always mapping to as many positions inside a given syllable template as possible. Rule-based algorithms posit an ordered set of structure-building rules which have similar status to that of other phonological rules: such rules may or may not apply directionally and do not require that syllable structure be maximalized in any sense from the start.

In the following two sections, I take a closer look at the mechanisms employed by rule-based and templatic approaches to syllabification, taking Kahn (1976) and Itô (1986) respectively as representative of such approaches. In section 2.2.3.3 I also review the alternative method of syllabification presented in Optimality Theory. Each method of syllable structure assignment claims that the order of application of rules (left-to-right, right-to-left) is crucial; however, this claim not only runs counter to the principles of declarative grammar, but also misses the point of syllable structure. In fact, these methods of syllable structure assignment represent only a small subset of possible methods: as stated in section 1.3.2, context-free rules may be applied in many different orders and ‘directions’ to yield the same structures. This also applies, of course, to rules in more restricted grammar formalisms. For this reason, I shall argue that it is better for phonology to focus on syllable structure itself, and leave syllable structure assignment to be dealt with by one of the various parsing algorithms which have already been developed for this purpose.

45 2.2.3.1 Kahn’s (1976) syllable structure assignment rules Kahn (1976: 39) states that he is ‘concerned with the question of how phonetic syllable structure is derived in English, i.e. with those rules of English phonology which assign syllable membership to segments’. He assumes that these rules take a string of segments as ‘input’, and yield a fully-syllabified structure as output. What follows is a brief recapitulation of the rules which he proposes: the first of these is given in (3): (3)

Rule I: With each [+syllabic] segment of the input string associate one syllable. (Kahn 1976: 39)

Thus, for example, if !+D#D#D9D! is the input string, Rule I will yield (4):

(4)

!

+

D "

#

D "

#

D "

9

D

!

"

(where " represents a syllable node; Kahn uses the less standard symbol $)

The feature [+syllabic] is inherited from Chomsky and Halle (1968). As we have seen, this feature, in its role as an inherent phonological property of segments, was later rejected (e.g. in Selkirk 1984). Selkirk considered [+syllabic] unnecessary on the grounds that syllabicity is better seen as a function of a segment’s environment. Furthermore, Kahn makes no reference at all in these rules to sonority. However, this is perhaps inconsequential, as the wording ‘each [+syllabic] segment’ in (3) could be replaced with, say, ‘each segment whose sonority is at least that of !,=
!264&8'+,! !6&",24+,! !A06#'+! !"#&%4'+!

‘finish!’ ‘get out of the habit!’ ‘throw!’ ‘get up!’

As Coats observes, Zaliznjak (1977: 84) ‘gives a version of the “two consonant” rule which explicitly mentions that 1 and two consonants function the same way in the operation of the rule’ (Coats 1997: 159). The fact that 1 patterns with the forms in (23a) seems to be conclusive evidence that 1 stands for a sequence of two phonemes. A summary of the differences between the phonological inventories advocated in different works is given in Table 5. Table 5. Phonological inventories of different scholars Work

!,!

!2'!

!B'!

!C'!

!8'I! !5'I!

Avanesov (1956) Halle (1959) Bulanin (1970) Gvozdev (1949) Kuhn Plapp (1996) Coats (1997) Oliverius (1974) Panov (1979)

   * -

   *  -

  *  -

  *  -

 * 

 * 

Key to Table 5  *

Phoneme is included in given inventory Phoneme is excluded from given inventory No explicit evidence available from given work

No. of No. of consonantal vowel phonemes in phonemes in inventory inventory 34 5 33 10 35 6 35 6 * * * * 35 5 34 5

66 Most of the arguments set out above boil down to the question of how to reduce the phonological inventory to a minimum, such that the phonological analysis can remain consistent. For each of the three areas mentioned, it is significant that the more minimal view of the phonological inventory generally runs into problems only with word-formation (as Kuhn Plapp makes clear in the case of [i] and [i]). Apparent ‘problems’ for the minimal-inventory view may therefore simply indicate that the rules governing morpheme-internal phonology are not quite the same as those which apply at morpheme boundaries, a fact which Pesetsky observed some time ago, albeit in slightly different terms: A number of rules, like the rule of yer-lowering, have exceptions. Crucially, these rules… can be shown to interact with other rules in a way that shows they must, or can be cyclic. Exceptions to such rules appear both in lexicalized words and in spontaneous coinages… By contrast, rules like palatalization before front vowels, fronting of i to i after velars, or the change of underlying w to v in most dialects show no unmotivated exceptions. These rules do not have to be ordered before any cyclic rules, and, moreover, violate recent formulations of the strict cycle condition… by applying to segments in environments wholly contained by the first cycle of a word. This suggests that the appropriate distinction may be between cyclic and post-cyclic rules. Only cyclic rules, apparently, admit unmotivated exceptions. (Pesetsky ms 1979: 14-15).

The approach followed in this dissertation is to exclude all of !,= 27= B7= C7=
segment [$sonor, +cont, +anter]

segment [$cont, $del_rel]

segment [+cons, +sonor, $nasal]

86 Finally, we need to consider four-consonant onsets. All such onsets in Russian begin with !"#! or !"*!, and the overwhelming majority have a stop and a liquid as the third and fourth segment respectively. Examples are given below: (46)

(9# (95!8-.'8.$ (9#!,5'8.$ (-%6,%,3)(&.$ (-%!/.$ (-46"-% (-4!/-%)(&.$ (-.!,).$

!"*A06#! !"*B-'%$! !"*B0;#&4;&'! !"*$06B4;&'! !"#2-6-6&8',"%&'! !"#20,&'! !"#9-'1#2! !"#90,#2,"%&'! !"#&067,&'!

‘upthrust’ ‘glance’ ‘to feel sad’ ‘to give a start’ ‘to tousle’ ‘to unseal’ ‘splash’ ‘to sprinkle’ ‘to build (into)’

These onsets can be accounted for using the following rule: (47)

onset [+init]

'

segment [$sonor, +cont, $coron, +anter]

segment [$sonor, +cont, +coron, +anter]

segment [$cont, $del_rel]

segment [+cons, +voc]

The onset rules proposed above account for virtually all occurring initial clusters; I will return to state what the exceptions are below. I now turn to codas. As with onsets, I start by considering word-marginal cases. Table 9 on page 88 shows what codas occur word-finally and with what frequencies (as Table 8 showed for word-initial onsets). Again, the frequencies are based on the entries in Zaliznjak (1977) and the figure abstracts away from voiced/voiceless and palatalized/unpalatalized distinctions. The total number of codas

87 represented in this table is 6,648, which is the number of entries in Zaliznjak’s (1977) dictionary which have final codas of exactly two consonants.14

14

This is significantly lower than the equivalent figure for word-initial onsets (19,978), suggesting that in general Russian is more permissive of complex onsets than complex codas.

Following segment

89

It is clear from Table 9 that the picture is far from straightforward. Although a greater proportion of word-final codas than word-initial onsets conform to the SSG (5,295 of the total 6,648, or 7926%), the SSG clearly needs to be supplemented. Again, the emphasis in the following discussion will be on attaining a good level of coverage (over 95%) rather than on accounting for every fact. First of all, Table 9 shows that most segments less sonorous than !-! can follow the liquids in word-final codas (completely in accordance with the SSG). To account for these codas, I propose the following rule. (48)

coda

'

segment [+cons, +voc]

segment [+cons, $voc]

Table 9 also shows that almost any segment can precede !$! or !&! in a coda; the only exception is !&#!. The following rule can be used to account for these codas. (49)

coda

'

segment [$voc, $del_rel]

segment [$cont, +coron, +anter, $del_rel]

In two-segment codas, a number of segments can occur after !4!. The majority of these are of one of two types: either !*=
‘shriek’ ‘wild boar’ ‘metre’ ‘cycle’ ‘rouble’ ‘beaver’ ‘sequence’ ‘body-size’ ‘tiger’

segment [$sonor, +cont, +coron, +anter] segment [$cont, $coron, $anter] segment [+coron, $cont, $anter] segment [+cons, $voc, +anter] segment [+cons, +voc, +coron]

91

This now brings us to the syllable structure grammar in its final version. Although not all onset and coda types have been accounted for, the vast majority have: the grammar has a coverage of 11,122 word-forms out of 11,290 (98251%). The full grammar is given in Appendix 1, with cross-references to the text above. A full list of the clusters not accounted for by the syllable structure grammar I propose is given in Table 11 and Table 12 below. The only medial cluster which failed to be parsed by the grammar was !A8&8'! (as in ,0-.( '&3&6$-.( 4!,)9(,#-.( 4!,-.!&'-.( 5,-8#&!-.( )-%8--.( 6? (,6'&

‘power’ ‘to please’ ‘wave’

(3,%#'&')< (-.%"'D) ("H,-%)
131 Undoubtedly, this is a case where further experimental evidence is called for. Evidence of this kind would show whether syllabification does have an effect on Russian voicing assimilation. On the other hand, Halle and Avanesov do not claim that the cases in (76) are instances where the devoicing of the sonorants is obligatory, so I shall proceed on the assumption that sonorants are voiced by default (i.e. their voicing is unaffected by Russian voicing assimilation), and that the option of devoicing is governed by some minor rule which, following Halle (1959), I shall not describe in detail. Once again, the behaviour of !"! is idiosyncratic (see 2.3.2 and 2.5.2.3 above). Followed by a sonorant, !"! behaves as a sonorant (Avanesov 1956: 162, Halle 1959: 63, Coats and Harshenin 1971, Jakobson 1978, Halle and Vergnaud 1981: 17); in other words, it neither triggers nor undergoes voicing assimilation. As noted above, there are a number of reasons to substantiate the idea that !"! is phonologically a glide in such circumstances (Andersen 1969, Coats and Harshenin 1971). An example is given below:

(77)

F("!$

‘Tver’ (place name) ()37M0')

!&"'10'!

However, when !"! precedes an obstruent, it counts as an obstruent (Jakobson 1956: 199, Andersen 1969: 121-122, Halle 1959: 63, Jakobson 1978: 107, Shapiro 1993: 2). This means that it both assimilates in voice to the following obstruent (78a), and triggers assimilation in the previous obstruent if there is one (78b).

132 (78)

(a)

(b)

(-.!"3& (#!85 4,(.,!).$ % (#,(" )9 (:,#&

‘meeting’ ‘suddenly’ ‘to repeat’ ‘to the widow’ ‘out of the entrance’

(2+).71&8'3)< (3*.;2) (932)H0',&')< (53*H"'1)< [D+286$3)


&"

(+)7

159 Inspection of the parse tree allows us to identify the subtrees such as the ones listed in Figure 10. Figure 9 is effectively ‘the join (unification) of [these] local graphs, each of which represents a rule’ (Coleman 1998: 165). Each of these subtrees can then be converted in a straightforward manner into the rules in (89). A fully-specified representation such as Figure 9 therefore actually embodies the grammatical rules: there is no distinction between rules and representation.

Figure 10. Examples of subtrees from Figure 9

verb-stem

prefix (89)

subst. suffix

verb root

(+)7

Exhaustive list of CFG rules inferred from parse tree in Figure 9 word-form noun noun-stem adj-stem verb-stem prefix verb-root adj. suffix subst. suffix infl. ending

' ' ' ' ' ' ' ' ' '

noun noun-stem adj-stem verb-stem prefix !4'1! !$"',5! !,+! !6#&'! !!

noun-inflection subst. suffix adj. suffix verb-root

It should be emphasized that the phrase-structure approach to morphology is not new. This approach is followed, for example, by Selkirk (1982b) and Lieber (1992).

160 Predictably, a context-free word-formation grammar can fulfil the same parsing and generation functions as a context-free (or linear) syllable structure grammar: just as the syllable structure CFG identifies the positions of syllable boundaries within a word, a morphology CFG identifies morpheme boundaries. As we shall see in Chapter 4, this is the minimum requirement for morpheme-based stress assignment in Russian. Clearly, though, a CFG does more than just morphological tokenization (segmenting a word into a string of morphemes), as in Figure 11. It also assigns hierarchical structures to these strings. Figure 11. Morphological tokenization of '"#()*)+,-.$ !-7$*37&>&"(+)7!

-7$

*37&>

&"

(+)7

Morphological tokenization underlies Oliverius (1976) and many other analyses of Russian, including those of Melvold (1989) and Halle (1997). However, tokenization is only part of the picture: some important generalizations are lost if one ignores hierarchical morphological structure, and CFG provides precisely the link needed to make explicit the relationship between tokenizations and parse trees. The type of generalization which can be made by CFG rules but not by tokenization becomes very clear if we compare the parse tree for '"#()*)+,-.$ !4'1$"',5,+6#&'! ‘immovable property’ (already given in Figure 9) with that for '"#()*)+/0 !4'1$"',5,+,7! ‘immovable’, given in Figure 12. The words have a very similar structure: in fact, the only structural difference is that the noun contains the substantival suffix !6#&'!. In order to explain '"#()*)+/0, therefore, we need add only two rules to the CFG in (89): these are given in (90).

161 (90)

adjective inflectional ending

adj-stem !,7!

' '

infl. ending

The rules in (89) and (90) will, moreover, be of use in describing many other words of similar structure even if the words consist of different morphemes, since the higher-level CFG rules describe word-structure in terms of morpheme types, not in terms of the morphemes themselves. Thus, if the non-terminal symbols of the CFG are chosen felicitously, it may be possible to reduce the overall size of the grammar significantly. Figure 12. Parse tree for '"#()*)+/0 !-7$*37&>&"&?!

word-form adjective adj-stem verb-stem

3.2

adj-infl adj. suffix

prefix

verb root

-7$

*37&>

&"

&?

A linear grammar of Russian word-formation

3.2.1 The morphological inventory of Russian In section 1.4 I set out the methodology by which I evaluate different wordformation grammars by reference to criteria of coverage, coherence and scope. As far

162 as coverage is concerned, different grammars (or different versions of grammars) are compared by measuring the percentage of cases where the observed form matches the form predicted by the theory. In this chapter, we are primarily interested in the ability of different versions of a word-formation grammar to assign correct structural morphological descriptions to strings. Some kind of standard against which to measure the output of a word-formation grammar is therefore needed; we need to know what the correct morphological descriptions are in order to tell whether the predicted forms are correct. In the case of morphology, this is not always an easy question to answer, as both the kinds of morphological structure and the morphological inventory we posit will depend on our theory of morphology. There is a risk, therefore, that arguments about which word-formation grammar covers the data better may be circular, as they cannot be resolved by reference to extrinsic data. But this is true only when we consider morphology in isolation: in reality, morphology is interconnected with other aspects of grammar. This means that any position we take about morphology may have implications in other areas; for example, theory X which assumes that the morpheme inventory is reduced to a minimum may yield different stress assignment results to theory Y, in which there is a large morpheme inventory and a minor role for morphological rules. What is important overall is that a generative linguistic theory should be coherent, as well as being able to assign structural descriptions to strings; and coherence is harder (but not necessarily impossible) to achieve, the more a theory relies on rules and the less it relies on the lexicon. The relationship of morphology to other areas of grammar (such as stress assignment) will not be considered until section 5.3. In the meantime, however, we

163 need some kind of foundation upon which to build a specific grammar of Russian morphology. For this, I used the purpose-made lexical database described in 1.5, which is based on Oliverius’s (1976) morphological tokenization. The structures assigned to strings by each version of the word-formation grammar were compared against the structures given in this database, using the procedure in (91). (91)

Procedure to check the output of word-formation grammars (a) Unparse the words from the lexical database, and keep a record of the associated morphological tokenization (as per Oliverius 1976). (b) Use CFG rules in the given version of the grammar to map the unparsed words from (a) to parse trees. (c) Map the parse trees to morphological tokenizations by removing information about hierarchical structure. (d) Compare the CFG-generated morphological tokenization from (c) with that of Oliverius, from (a).

This move means, of course, that I commit myself both to Oliverius’s view of how Russian words are segmented, and to his morpheme inventory. In anticipation of the criticism that this move places undue reliance on a particular view of morphology, I would give the following responses. First, formally explicit, detailed and systematic lexicographical work such as Oliverius (1976) is hard to come by. As Jefremova (1971: 117) states in a review of Worth, Kozak and Johnson (1970), ‘it is always the case that lexicographical works are easy to criticize and very hard to compile’ [my translation]. Secondly, one has to take some kind of position about the morphological structure of a large number of Russian words in order to proceed at all, and as the only such set of judgements which lends itself readily to the type of analysis advanced here, Oliverius’s dataset seemed the only feasible point from which to start.

164 (Tixonov’s 1985-1990 derivational dictionary of Russian is less amenable to use for morphological parsing, because although it has systematic information on the morphological structure of a larger number of Russian words, it does not include a classification of individual morphemes or morphs.) Finally, it should be pointed out that the commitment to Oliverius’s analysis is only provisional; it is after all central to my methodology that any decision can later be retracted on the basis of empirical evidence (see section 1.4). 3.2.1.1 Preliminaries: controversial issues Because Oliverius’s morphological inventory is explicitly formulated, it is easy to identify where it stands on a number of controversial issues. The points of contention fall into two broad areas. The first relates to the phonological inventory implicit in the morpheme inventory. This aspect of Oliverius’s theory will be only briefly considered, as a more detailed discussion has already been given in 2.3.1.1. Oliverius’s analysis of Russian morphology assumes the same phonological inventory as in Oliverius (1974); that is, !,! and !8'I! are not included in the inventory (being represented as !,! and !8'&8'! respectively), but !2'!, !B'! and !C'! are. An important consequence of this is that Oliverius is forced to list many morphemes of which the final segment is a velar as alternating, e.g. !&',C!%!&',C'!. For reasons discussed in 2.3.1.1, the current grammar treats !2'!, !B'! and !C'! as allophones of !2!, !B! and !C! respectively; the tangible benefit of this now is that the morphological inventory can also be simplified, collapsing any alternating ‘allomorphs’ of this type into a single allomorph. However, pairs of allomorphs are still separately listed where one member of the pair ends in a palatalized non-velar consonant and the other in the corresponding non-palatalized non-velar consonant.

165 The second area which is open to discussion also relates to the treatment of allomorphic alternations, but in this case the alternations in question are between forms which differ phonemically, in the terms of the phonological inventory established in Table 7. For example, Oliverius lists !$'14'! as alternating with !$4'!, and these two forms differ most obviously in the number of phonemes each contains: !$'14'! consists of three phonemes, and !$4'! of two. In the light of the fact that certain analyses have attempted to treat alternations such as !$'14'!%!$4'! as phonological alternations, it might appear that the distinction drawn here between !&',C!%!&',C'! and !$'14'!%!$4'! is arbitrary or even wrong, but I shall argue that the distinction is valid and indeed unavoidable. I shall return to discuss this further in section 3.3. 3.2.1.2 The classification system For the grammar proposed here, I based the terminal rules on the morpheme inventory in Oliverius (1976), which consists of 1,759 entries. This approach means that the question ‘what is a morpheme?’ is answered in the current framework simply by stating that it is one of the entries in Appendix 3. (In most cases, these entries correspond in a one-to-one fashion with the entries in Oliverius (1976); the exceptions are those already discussed, such as !&',C!%!&',C'!.) This section gives details of the classification system used for the morpheme inventory (this system essentially follows that of Oliverius 1976), and how the morpheme inventory is represented in context-free grammar. Oliverius labels each of the morphs which he identifies with a classificatory code consisting of one or more characters: each of these characters specifies one of the attributes (or features) of the morph. What follows in this section is essentially a recapitulation of Oliverius’s key (ibid, pp. 65-70), although the characters used have

166 been changed to suggest linguistic terminology in English rather than in Latin (as used by Oliverius). The first character, which is always a letter of the alphabet, specifies whether the morpheme in question is a root, suffix, clitic, or inflectional ending, according to the key in (92): (92)

Key to classification of morpheme types (1) Code

Meaning

r s c

root suffix non-final morpheme which can stand in isolation (i.e. clitic): prefix, preposition, conjunction, interjection or particle inflectional ending

i

If the first character is c, then there are no further letters, implying that clitics need not be classified further. However, if the first character is r, s, or i, there are more letters, each of which specifies further attributes of the morpheme. The next character specifies the morpheme’s ‘fundamental semantic property’ (Oliverius 1976: 65) & whether the morpheme is pronominal, adjectival, substantival or verbal & according to the key in (93): (93)

Key to classification of morpheme types (2) Code

Meaning

p a n v

pronominal adjectival substantival verbal

167 Thus a morpheme labelled rn is a substantival root, a morpheme labelled sv is a verbal suffix, a morpheme labelled ip is a pronominal inflection, etc. If the morpheme is a root or any suffix other than a verbal suffix, Oliverius specifies no further attributes. If the morpheme is a verbal suffix, then there may be one further letter which specifies whether the morpheme connotes durativity, repeated action, or a single completed action (corresponding to various meanings of the imperfective and perfective aspects in Russian), according to the key in (94): (94)

Key to classification of morpheme types (3) Code

Meaning

d

durative process (#8!&.)(',-.$, #6)."6$',-.$, 4!,.>*"'',-.$ 4!,7"--&) resultative process (A@@"%.)(',-.$, #"0-.("'',-.$, !"986$.&.)(',-.$ 4!,7"--&) iterative process ()."!&.)(',-.$, 4,(.,!>"+,-.$, +',5,%!&.',-.$ 4!,7"--&) completed process (4!":,#)+,-.$ 4!,7"--&)

r i c

Finally, the last character in the code may be a digit, irrespective of the morpheme’s attributes specified up to this point. This digit serves to distinguish two ‘homophonemic morphemes’ (morphemes which have the same phonological composition and the same attributes, but nevertheless have distinct meanings). An example where distinction needs to be made in this way is between !$'1-! (as in ,.#"6 !6&$'1-! ‘department’) and !$'1-! (as in #"6&.$ !$'1-%&'! ‘to do’): the former is assigned the label rv1 and has the connotation ‘divide’, the latter is assigned the label rv2 and has the connotation ‘do’. This means of distinguishing homophonemic morphemes is not relevant to the type of morphological analysis described in this chapter, but it

168 becomes relevant in stress assignment, where two morphemes may differ not just in their meaning, but also in their accentuation (see 4.3 and following). For example, !,4!rn1 is specified as a morpheme which does not trigger stress retraction, but !,4!rn2 does trigger stress retraction in the plural: for further discussion, see section 4.3.2.6. Table 18 shows the serial-number classification system which I have adopted to distinguish substantival inflectional morphs. In certain cases, there is homonymy between different cases, and I have followed the analysis of Carstairs (1987: 142143)26 in terms of which features have been collapsed together. (This table is not intended to include all inflectional morphs; it includes only those which occur in the word-list from Oliverius 1976. Thus, for example, the locative singular inflectional morph !;! is excluded.)

26

All the instances of homonymy in Table 18 are described by Carstairs as instances of ‘systematic’ (as opposed to ‘accidental’) homonymy.

169 Table 18. Classification system for substantival inflectional morphs Code in1 in2 in3 in4 in5 in6 in7 in8 in9 in10 in11 in12 in13 in14 in15 in16 in17 in18 in19 in20 in21 in22 in23

Morpheme !! !! !%! !%! !%+',! !%+! !%C! !1! !17! !,! !67! !6"! !7;! !+',! !%! !%! !%! !6! !6+! !,! !;! !;! !1!

Features nom., acc. sg. gen., acc. pl. nom. sg. nom., acc. pl. (neuter nouns) instr. pl. dat. pl. loc. pl. dat., loc. sg. gen., acc. pl. nom., acc. pl. instr. sg. gen., acc. pl. instr. sg. instr. pl. gen., acc. sg. nom. pl. (masc. nouns) 27 nom. sg. (neuter nouns)28 nom., acc. sg. instr. sg. dat., gen., loc. sg. acc. sg. dat. sg. nom. pl.29

Cross-reference to Carstairs (1987)

page 142, no. 21b page 142, no. 22a n/a page 142, no. 20 n/a n/a n/a page 142, no. 23 page 142, no. 22a page 142, no. 21a n/a page 142, no. 22a n/a n/a page 142, no. 22b n/a n/a page 142, no. 20 n/a page 143, no. 25 n/a n/a n/a

Throughout this dissertation, where words are morphologically tokenized (segmented into morphs), morpheme boundaries are shown by the + symbol. The phonological content of each morph is followed by a subscript showing the morpheme type. Thus !#c++,#-rn+in1! denotes a word consisting of three morphs: the prefix !#!, the noun-root !+,#-!, and a noun-inflection with zero phonological content. Individual

27

For example, in 4!,@"--,!& !906c2+?'1#rv+#sa+60sn+%in16! ‘professors’. For example, in )+> !,+'rn+%in17! ‘name’. 29 For example, in 5!&*#&'" !B0%5$rv+%4'sa1+sn1+1in23! ‘citizens’. 28

170 morphs are represented with the subscript outside the obliques, as follows: !#!c, !+,#-!rn, !!in1, etc. The translation of Oliverius’s morpheme inventory into context-free grammar is straightforward: there will be one terminal rule for each entry in the morpheme inventory. (Since the number of morphs in a language is likely to be much greater than the number of ways in which different types of morph may combine together, the morpheme inventory is likely to represent the bulk of the grammar.) For example, where Oliverius lists !26-A%#! as an ‘rn’ morph (a noun-root), we have the linear rule in (88):

(95)

rn

'

!26-A%#!

Clearly, the translation itself raises no theoretical issues, but note again that in using Oliverius’s morpheme inventory I am making a (provisional) commitment to his particular morphological analysis. 3.2.2

The word-formation rules Having established a set of pre-terminal rules, the element of the CFG which

is still lacking is the set of rules which specify the hierarchical structure of morphologically complex words. In the following sections, I set out specific proposals for what these rules should be.

171 3.2.2.1 Words with no internal structure Inspection of Oliverius’s list of morphological tokenizations reveals that there are a number of words consisting of a single morpheme: three examples are given in (96) below. (96)

4"!"# ( 4,%& ,:

!9'10'1$!c !"!c !962%!c !6C!c

‘in front of’ ‘in’ ‘while’ ‘oh!’

These words, which are all prepositions, conjunctions and interjections, can be subsumed under the general term ‘particle’. The first CFG rules which we need are therefore ones to license such words. These rules will be as in (97): (97)

word particle

' '

particle c

The terminal rules of the CFG have already made explicit which morphemes are ‘c’ morphemes. For example, we already have the rules in (98). (98)

c c c c

' ' ' '

!9'10'1$! !"! !962%! !6C!

The CFG rules in (97) and (98) will account for all the 67 single-morpheme words in Oliverius (1976).

172 3.2.2.2 Nouns Having dealt with monomorphemic words, we now turn to nouns. First, we shall consider non-derived nouns (those which do not contain a derivational suffix; cf. Townsend 1975: 149). There are 325 non-derived nouns in Oliverius (1976): three examples are given in (99). (99)

!8%& 3"+,#&' 4,6"

!0;2rn+%in3! !&8'1+6$%4rn+in1! !96-'rn+6in18!

‘hand’ ‘suitcase’ ‘field’

These words can all be accounted for by the context-free rule in (100).

(100) noun

'

rn

in30

To make it clear that nouns are examples of words, the rule in (101) is needed:

(101) word

'

noun

So far, then, the CFG of Russian morphology is as in (102):

30

This rule, of course, incorporates a non-arbitrary relationship between ‘rn’ (noun-root) and ‘in’ (noun-inflection), which could be further generalized (e.g. in terms of agreement of features attaching to each constituent morpheme).

173 (102) word word noun c rn

etc. etc.

in

etc.

' ' ' '

noun c rn !9'10'1$!

'

!0;2!

'

!%!

in

We next turn to derived nouns with a single suffix. Some examples are shown in Table 19. Table 19. Further categories of morphological tokenization Type of tokenization

rv+sn+in (verb-root + noun-suffix + infl.) ra+sn+in (adj-root + noun-suffix + infl.)

Example

Number of words of this type in Oliverius (1976)

*)9'$ !5,rv+*4'sn+in! ‘life’ !&#,-.$ !0%$ra+6#&'sn+in! ‘joy’

100

Total

46

146

The CFG rules in (103) can account for these words, but it is equally possible to license them using the more articulated rules in (104). (103) noun noun

' '

rv ra

sn sn

(104) noun noun-stem noun-stem noun-stem verb-stem adj-stem

' ' ' ' ' '

noun-stem rn verb-stem adj-stem rv ra

in in in sn sn

174

The grammar fragments in (103) and (104) are weakly but not strongly equivalent, as they generate the same set of strings (in Table 19) but do not assign the same structural descriptions (parse trees) to those strings. It is of theoretical importance which grammar we choose, since ‘we have no interest, ultimately, in grammars that generate a natural language correctly but fail to generate the correct set of structural descriptions’ (Chomsky and Miller 1963: 297). Although (103) may appear to be more economical since it embodies fewer rules than (104), (104) has a number of key advantages from a linguistic point of view. First, in (104), the concepts ‘verb-stem’, ‘noun-stem’ and ‘adjective-stem’ have now been formalized and can be used to support further definitions, such as that in (105), which states that an adjective stem can be formed from a noun-stem and an adjectival suffix.

(105) adj-stem

'

noun-stem

sa

More important, however, (104) treats affixation as the appending of a suffix to a unitary morphological object (the base) (cf. Townsend 1975: 26). The rules in (104) cannot look inside a base to determine any facts about its composition other than that it is a noun-stem, adjective-stem, etc. As a result, (104) stems from a constrained but very general hypothesis about affixation. (103), on the other hand, assigns equal importance in the parsing hierarchy to each morpheme in the string: the rules in (103) lack generality, because their applicability is limited to the 146 words with the specific structures listed in Table 19, and a separate rule would be required to account for, say, nouns with the structure rn+sn+in.

175 To give a concrete example which demonstrates the generality of (104), consider the word *"'1)'& !5148&8',4%! ‘woman’, to which Oliverius (1976) attributes the morphological tokenization in Figure 13): Figure 13. Oliverius’s (1976) tokenization of *"'1)'& !>$-@)@7&-#! ‘woman’

>$-rn

@)@7sa

&-sn

#in

Using (104) and (105), this word can be described by the parse tree in Figure 14 (as is apparent, this parse involves nesting of the term ‘noun-stem’). Figure 14. Parse tree for *"'1)'& !>$-@)@7&-#! ‘woman’

word noun noun-stem adj-stem noun-stem

in sn

sa

rn

>$-

@)@7

&-

#

Affixation rules of the type in (104)-(105) allow the categories ‘noun-stem’ and ‘adjective-stem’ to be nested ad infinitum: ‘adjective-stem’ can be defined in terms of ‘noun-stem’, and vice versa. The definition of a constituent in terms of itself

176 is known in the theory of computation as recursion (Chomsky and Miller 1963: 290). Without recursion, only a finite number of strings can be parsed; but if a grammar contains recursive rules, there is no theoretical limit to the length or number of words which can be described by the grammar. The rules in (104)-(105) are formally very neat and do not violate the requirement of context-freeness. But the main objection which can be raised against them is on a linguistic, not a formal level: the rules are too general. Essentially, there is no evidence in Russian to suggest that suffixation really is recursive. If suffixation were truly recursive, that would imply that a given suffix, not just a type of suffix, could appear a limitless number of times in a single word. Stated another way, if the rules in (104)-(105) (plus appropriate pre-terminal rules) can license (106a), then they must also be able to license the words in (106b), (106c), etc: (106) (a) (b) (c)

!#9'1&#rv+,7sn+%-'sa+4sa+6#&'sn+in! !#9'1&#rv+,7sn+%-'sa+,7sn+%-'sa+4sa+6#&'sn+in! !#9'1&#rv+,7sn+%-'sa+,7sn+%-'sa+,7sn+%-'sa+4sa+6#&'sn+in!

Thus a grammar of morphology based on recursive rules is likely both to accept as well-formed, and to generate, a large class of words whose well-formedness is questionable. This, of course, is perhaps a problem only because until this point we have glossed over the semantics of word-formation. It may be, for example, that a truer representation of Russian word-formation rules avoids the problem of recursion because of the way in which semantic features are distributed. For example, (108), which allocates each constituent a multivalued semantic feature, retains some of the generality of (107), but is not a recursive grammar fragment:

177

(107) noun-stem noun-stem noun-stem verb-stem adj-stem adj-stem

' ' ' ' ' '

(108) noun-stem ' [semantics:concrete]

rn verb-stem adj-stem rv ra noun-stem

sn sn sa

rn

noun-stem ' [semantics:deverbal]

verb-stem [semantics:concrete]

sn

noun-stem [semantics:action]

adj-stem [semantics:past_participle]

sn

'

verb-stem ' [semantics:concrete]

rv

adj-stem ' [semantics:concrete]

ra

adj-stem ' [semantics:desubstantival]

noun-stem [semantics:concrete]

sa

adj-stem ' [semantics:desubstantival]

noun-stem [semantics:deverbal]

sa

Again, these rules are probably a first approximation: after all, there is no reason in principle why semantic features should not be combined in any quantity or data structure, including empty data structures, in the style of Unification Grammar. I should emphasize that the nature and identity of semantic features are peripheral to my argument, so I shall not dwell on this point here: it is sufficient just to assume that semantics does play a role in word-formation, and that the distribution of semantic features is such that the grammar is non-recursive. Returning to the formal properties of the grammar so far, it is worth noting that all the rules in the grammar are left-embedding (Chomsky and Miller 1963: 290).

178 In other words, the rightmost branch of each expansion is a pre-terminal symbol, and consequently the parse tree as a whole can acquire ‘depth’ only through the leftmost branches of each expansion, as can be seen graphically in Figure 14.31 The import of this is that the grammar is left-linear, and the language defined by the grammar is a regular language (Chomsky 1963: 394). This means that in their restrictiveness the rules of Russian word-formation set out so far32 surpass the original goal (using a grammar no more powerful than context-free). 3.2.2.3 Verbs There are a number of factors which conspire to make the verbal morphology of Russian more complex than its nominal morphology. First, according to Melvold (1989: 79), the number of non-derived (traditionally referred to as athematic) verbs in Russian is only around 90, which is much less than the number of non-derived nouns, even from the sample of 2,493 words in Oliverius (1976). Thus derivation plays a much more important role in verbal word-formation. Secondly, verbal inflections are not monomorphemic, since they consist of a tense and an agreement suffix. Furthermore, these suffixes are often phonologically ‘invisible’, because suffixation often results synchronically in adjacent vowels which are elided (Townsend 1975: 81). All the verb infinitives in Oliverius have one of the following sequences as their final morphs:

31

This format is what we would expect in a grammar of Russian morphology, given the fact that most affixation in Russian (suffixation and inflection) takes place on the right of the word (i.e. the right is where the pre-terminal symbols are added). The treatment of prefixation proposed in section 3.2.2.4 preserves the left-linearity of the grammar. 32 However, the rules which will be formulated in section 3.3 are context-free, not linear.

179

(109) !+sa+ia! !&'sa+ia! !&',sa+ia!

(example: +,3$ !+6&8'rv+sa+ia! ‘to be able’) (example: 3).&.$ !&8',&rv+%svi+&'sa+ia! ‘to read’) (example: )#.) !,$rv+&',sa+ia! ‘to walk’)

Since all infinitives are assigned a zero-inflection by Oliverius, it is unclear whether this inflection is really a necessary element in the string. Not only does its inclusion mean that the morphological tokenization contains more elements than necessary, but to assert that the structure of )#.) ‘to go’ is !,$rv+&',sa+ia! implies that !,$rv+&',sa+ia! and !,$rv+&',sa! are distinct structural units in the grammar of Russian. Oliverius concedes that the zero-morph has only ‘residual’ meaning which originated in the Slavic opposition between the infinitive and supine (Oliverius 1976: 42). In view of the lack of any strong synchronic justification for this analysis, it seems simpler to disregard the final zero morpheme on all infinitives, and call the final !!, !&'! or !&',!<morpheme in the infinitive a verbal inflectional ending, i.e. !!iv, !&'!iv or !&',!iv. The non-derived infinitives +,3$ ‘to be able’ and )#.) ‘to go’ would thus be described by the rules in (110): (110) word verb

' '

verb verb-stem

iv

To account for derived verbs, we use exactly the same technique as for derived nouns. Thus 3).&.$ !&8',&rv+%svi+&'iv! ‘to read’ could be parsed by the additional rules in (111):

180 (111) verb verb-stem [+derived]

'

verb-stem

iv

'

verb-stem [$derived]

sv

The ‘sv’ element in the second rule of (111) is actually the ‘thematic vowel’, so that this rule essentially says that a verb-stem is obtained by adding a thematic vowel to a verb-root. The addition of just one rule to this grammar will account for reflexive verbs, which in Russian look just like non-reflexive verbs, but have the final morpheme !#'!c or !#'%!c (e.g. 3).&.$-> !&8',&rv+%svi+&'iv+#'%c! ‘to be read’): (112) verb [+reflexive]

'

verb [$reflexive]

!#'(%)!c

3.2.2.4 Prefixation As pointed out in Townsend (1975: 149), ‘prefixation plays a very minor role in formation of nouns’. This is confirmed in Oliverius’s (1976) analysis, where 837 of the 967 prefixed forms include either a verb root or suffix. Most Russian prefixation can thus be dealt with under the heading of verbal word-formation, and it is this which I shall concentrate on in this section. Prefixation in Russian verbal forms is very closely intertwined with verbal aspect. As Townsend states: With certain limited exceptions the addition of a prefix to a simplex imperfective stem perfectivizes that stem. In traditional analysis and most grammar books a prefix is either “nonsemantic” or “semantic”: if the former, it merely perfectivizes the stem and is regarded as the perfective “partner” of the imperfective verb (Townsend 1975: 116).

181 Thus imperfective 3).&.$ !&8',&rv+%svi+&'sa! ‘to read’ can be perfectivized by adding the prefix !906!c (which in this case is non-semantic), forming !906c+&8',&rv+%svi+&'sa!. The question is: what is the correct structural description of !906c+&8',&rv+%svi+&'sa!? The candidates which one might consider are shown in Figure 15.

Figure 15. Three alternative representations of !6.(c+)@7&)rv+#svi+)7sa!

infin

906c

&8',&rv

infin

906c

&8',&rv

infin

906c

&8',&rv

The question of the correct structural analysis of Russian prefixed verbal forms was considered at some length in Pesetsky (ms 1979). Pesetsky suggested that the phonological behaviour of prefixed verbal forms (particularly with respect to vowel-zero alternations; we will return to this in 3.3) could be more easily explained within a framework which allows cyclic rules if the structure of these forms were as shown in the third representation in Figure 15. Yet, as he noted, there is a problem with this structure: While the bracketing [i.e. structure & PC] shown… yields the correct result when the phonological rule of yer-lowering is applied to it, it is at odds with one’s intuitions about the semantic correlates of morphological structure. In most cases, in

182 Russian as in English, the combination of a prefix and a verb root acts semantically like a verb in its own right, often bearing an idiosyncratic meaning not compositional of the usual meanings of the verb and prefix. So, while podH- usually denotes something like “up to” as a prefix, and žIg, unprefixed, is glossed as “burn”, the combination of podH- and žIg acquires the related, but unpredictable meaning “set on fire”. Even less predictable meanings are found in other prefix-verb combinations. For example, we have seen that s-;el/so-;la shows behavior similar to pod-žëg/podožgla, suggesting the same constituent structure. The prefix sH- in other contexts is glossable as “down from”, while the somewhat archaic ;It means “read”. The combination, however, means “consider (x as y)”, or sometimes “tally”. Such idiosyncratic meanings are preserved under the addition of inflectional and derivational suffixes. Thus, while the phonology of words like podožgla and podžëg suggests the bracketing of (27a) below, the semantics seems to suggest the bracketing of (27b): (27)

a. [ prefix b. [[prefix

[[root] [root]]

inflection]] inflection] (Pesetsky ms 1979: 10)

This is therefore a classic example of a bracketing paradox. A better-known example from English, also mentioned in Pesetsky (ms 1979), is the word unhappier. Pesetsky notes that one of the stress rules of English (‘attach the comparative suffix -er only to monosyllabic adjectives and to disyllabic adjectives in orthographic -y’) seems to imply the structure [un[[happy]er]], where un- is added ‘after’ this rule has applied, while the semantics appear to call for the structure [[un[happy]]er], since unhappier means ‘more not happy’ instead of ‘not more happy’. Pesetsky concludes: The problem is formally identical to our problem in Russian, except that the phonological difficulties which would be induced by accepting the semantic bracketing as morphologically real are of a different nature. The fact that we can factor out the same sort of semantic difficulty from the two examples suggests that the semantic problem is the real one, and not the phonological problem, and that the phonologically motivated bracketings of (27a) and (30) should be considered morphologically real. We will need a special rule of semantic interpretation, with some sort of universal status, to yield the correct readings for such forms. (Pesetsky ms 1979: 12)

183 Paradoxes of a very similar type33 in Polish were also considered in Booij and Rubach (1984), who also noted that prefixes have a special status with respect to phonological rules, and concluded that prefixes are best treated as independent phonological words. The contribution of Pesetsky (ms 1979) and Booij and Rubach (1984) is clear: prefixation must be treated with caution in formulating rules (like jer-lowering and deletion) which affect the phonological form of words. However, it is significant that both Pesetsky (ms 1979) and Rubach and Booij (1984) were working within the framework of Lexical Phonology. In Lexical Phonology, cyclic rules are ordered with respect to one another and with respect to postcyclic rules, but in context-free or linear grammar, rule-ordering is not an option. In the next section, I shall show, in any case, that the vowel-zero alternations described by Pesetsky can be modelled in context-free grammar irrespective of the morphological structure of words (and still capturing the generalization about the special status of prefixation). Thus there is no reason to avoid the morphologically-motivated bracketings like [[un[happy]]er], etc. The structure which captures the regularities of word-formation best is the first of the representations in Figure 15. Moreover, this is the only one of the structures suggested which avoids centre-embedding. There are at least two reasons to prefer this structure: not only would this preserve the idiosyncratic meanings of certain prefixed forms under suffixation and inflection, as noted by Pesetsky, but it also allows a straightforward analysis of deverbal nouns such as ,.48-% !6&c+9;#2rv+sn+in! ‘leave’, related to ,.48-%&.$ !6&c+9;#2rv+%svi+&'sa! ‘to let go’. Instead of deriving 33

Booij and Rubach’s data also relate to prefixed forms and the interaction of word-formation rules with vowel-zero alternations. Like Pesetsky, they claim that these are governed by cyclic rules.

184 !6&c+9;#2rv+sn+in! from !6&c+9;#2rv+%svi+&'sa! by some kind of subtractive operation (which presents problems for phrase-structure analysis), both !6&c+9;#2rv+sn+in! and !6&c+9;#2rv+%svi+&'sa! would be derived from the base !6&c+9;#2rv!. The corresponding linear rule schemata would therefore have to be those in (113). (113) (a)

(b)

verb-stem [$derived]

'

(c)

rv

verb-stem [+derived]

'

verb-stem [$derived]

sv

noun-stem

'

verb-stem

sn

(The rule in (113b) accounts for forms such as !6&c+9;#2rv+sn+in!.) Just as prefixes are closely related to verbal roots, explaining the fact that idiosyncratic meanings are preserved under suffixation and inflection, it is presumably also the case that prefixes are closely related to nominal roots: an inflected noun form is almost always compositional of the meaning of the stem (including prefix) and inflection. This suggests the need for the rule schemata in (114): (114) noun-stem [$derived]

'

(c)

rn

noun-stem

'

noun-stem

sn

[+derived]

[$derived]

Because prefixes tend to attach predominantly to verbal structures, the rules proposed above are probably the most important prefixation rules in Russian. These rules are those which Pesetsky (ms 1979) identified as semantically most intuitive.

185 However, Pesetsky did not adopt these rules; instead, both he and Rubach and Booij (1984) suggested a semantically counter-intuitive analysis of prefixation, namely that prefixes must be immediately dominated by the word node in the phrase-structure hierarchy. To understand fully why Pesetsky, Rubach and Booij followed the analysis they did, and why this analysis is rejected here despite their arguments, we will have to look in more detail at vowel-zero alternations in Russian. 3.3

Vowel-zero alternations in context-free grammar Vowel-zero alternations such as that embodied in !$'14'!%!$4'! have been

discussed extensively in the phonological literature under the general heading of ‘jers’. It is generally believed that in Common Slavic jers were non-alternating short vowels !I! and !H! (e.g. Bethin 1998: 104). These jers, it is supposed, underwent a change in Late Common Slavic such that word-final jers and jers in a syllable before a syllable with a non-jer vowel (both known as ‘weak’ jers) were deleted. All other jers (the ‘strong’ jers) developed into full vowels (in Russian, into !1! and !6! respectively). These processes are commonly referred to in Slavic linguistics as Havlík’s Law. As Bethin (1998: 206) states, a key problem is how the jers should be represented synchronically in Slavic languages. The traditional linear generative phonology approach (exemplified by Lightner 1972 and Pesetsky ms 1979 for Russian; Gussmann 1980, Booij and Rubach 1984, Rubach 1984 for Polish) has been to maintain that the jers are a type of vowel. This approach claims that jers are distinct entities in the phonological inventory, distinguished from non-alternating vowels in that the former are [+high, $tense] (as opposed to !,!, !;! which are [+high, +tense] and !1!, !6!, !%! which are [$high,

186 $tense]). This approach was used not only to explain vowel-zero alternations, but also: This representation of jers allowed Lightner to maintain that consonantal palatalization could be derived from the front or back specification on vowels, i.e., [stol] “table” and [stol’e], prep sg; [glupij] “stupid” and [glup’et’] “to become stupid,” just as [d’en’] “day” and [dn’a], gen sg, or [bol’en] “sick,” masc sg and [bol’na], fem sg... have derived palatalized or non-palatalized consonants from the front or back feature of the following jers. This predictability of palatalization before front vowels was relegated to a palatalization rule and it was seen as an advantage for that analysis. (Bethin 1998: 206)

The linear generative phonology approach entails certain problems, however. One of these, on a perhaps trivial level, is that it involves augmenting the phonological inventory. More important, however, it involves (a) postulating phonological entities which never surface in their underlying forms, and (b) absolute neutralization involving these entities (for example, when jers surface as !1! and !6! in Russian, they are phonetically indistinguishable from non-alternating !1! and !6! respectively). This aspect of traditional generative phonology was criticized, notably in Postal (1968), Kiparsky (1973) and in Natural Generative Phonology (Hooper 1976), on the grounds that it allows the range of possible phonological theories to become excessively unconstrained. However, it has since been argued, specifically with reference to vowel-zero alternations (in Polish, although of the same type that occur in Russian), that the most natural analysis in some cases does involve postulating abstract entities which never surface (Gussmann 1980: 83, Szpyra 1992). An alternative approach views jer representation as a function of syllabification, deriving the vowel by epenthesis where consonants cannot otherwise be syllabified (e.g. Piotrowski 1992 for Polish). For example, according to the

187 epenthesis approach, the alternation of !A6A0! with !A6A10! in (115) is explained by the fact that !A0! violates the SSG, and !1! is inserted to remedy this. (115)
noun -162810

noun-stem

infl. ending

-152445

-12365

adj-stem

sn

-122109

-32336

adj-stem

sa

-102110

noun-stem -62575

rv

sa

-32535

sn

-42009

-22566

+67$)+

&?

-42009

-12999

-22566

#,7

-32535

-

-12999

(+)7

-32336

-12365

209

3.4.2 Should the cost mechanism be based on hapax legomena? Baayen and Sproat (1996) discuss morphological disambiguation of a type similar to that described above: For a common form such as [Dutch] lopen ‘walk’ a reasonable estimate of the lexical prior probabilities is the MLE36, computed over all occurrences of the form. So, in the UdB [Uit den Boogaart & PC] corpus, lopen occurs 92 times as an infinitive and 43 times as a finite plural, so the MLE estimate of the probability of the infinitive is 0.68. For low-frequency forms such as aanlokken or bedraden, one might consider basing the MLE on the aggregate counts of all ambiguous forms in the corpus. In the UdB corpus, there are 21,703 infinitive tokens, and 9,922 finite plural tokens, so the MLE for aanlokken being an infinitive would be 0.69. Note, however, that the application of this overall MLE presupposes that the relative frequencies of the various functions of a particular form are independent of the frequency of the form itself [my emphasis & PC]. For the Dutch example at hand, this presupposition predicts that if we were to classify -en forms according to their frequency, and then for each frequency class thus defined, plot the relative frequency of infinitives and finite plurals, the regression line should have a slope of approximately zero (pp. 156157)

Baayen and Sproat go on to show that this is not the case: analysis of the data from the UdB corpus shows that given a Dutch form in –en, the more frequent the word, the more likely the form is to be a finite plural. Further, they show that a disambiguation system designed to predict whether a previously unseen –en form is an infinitive or a finite plural gives considerably more accurate results if it is based on tagging just the forms of lowest frequency (specifically, the forms which occur just once, or the hapax legomena) in a corpus, than if the entire corpus is tagged to give an overall measure of the MLE. The reason for this is that ‘idiosyncratic lexical

36

Maximum Likelihood Estimate: this is computed by ‘counting the number of times a form occurs in each of its various functions and dividing by the total number of instances of the form (in any function)’ (Baayen and Sproat, 1996: 156).

210 properties of high-frequency words dominate the statistical properties of the highfrequency ranges, thus making the overall MLE a less reliable predictor of the properties of the low-frequency and unseen cases’ (p. 163). Baayen and Sproat’s conclusion is well-supported both theoretically and empirically. Although the disambiguation problems described above are not entirely analogous to the disambiguation problems discussed by Baayen and Sproat (mainly because they are not systematic ambiguities in the lexicon like that between finite plurals and infinitives in Dutch), their findings appear to suggest that P(Mx) may not be the best basis on which to compute the cost of a given parse. This is because P(Mx) is computed from the entire lexicon, not just low-frequency words. In other words, Baayen and Sproat’s results might lead us to believe that a parser based on P(Mx) would give wrong analyses for low-frequency words because it has a tendency to use too many unproductive, high-frequency morphs. This effect would be exacerbated by the fact that most higher-frequency morphs consist of few phonemes, meaning that the combinatorial possibilities of using them in parsing would increase. To take an example, in Oliverius (1976) A.,. !1&6&! ‘this’ is tokenized as !1c+&rp+6&ip!, and A.,. !1&6&! has a word token frequency of 4,554. (The total corpus lists 313,564 word tokens, a figure which is arrived at by summating the frequencies of all words in Oliverius 1976.) 4,554 is a high word token frequency: in fact, A.,. is ranked the ninth most frequent word in Oliverius (1976). The morph !1!c does not appear in any other word in the corpus, and therefore its morpheme token frequency is also 4,554: log P(!1!c) is thus log (4,554 / 1,318,181) ; J22462. This is also comparatively high (as stated above, the closer the log value is to zero, the higher the LPP); for example, log P(!1!svr) is J22554. (!1!svr is a highly productive morph

211 occurring in 40 of the words in Oliverius 1976, many of which have a low frequency: for example, ()#'".$-> !"',$rv+4'sa+1svr+&'sa+ia+#'%c! ‘to become visible’ has a frequency of 15, and the least frequent word listed in Oliverius has a frequency of 14, both also out of 313,564. Because !1!c consists of a single phoneme, this morph can be used to parse any word where there is an !1! (subject to the appropriate phrase structure constraints being met), and the high value of log P(!1!c) will tend to make it more profligate in such circumstances. Surprisingly, perhaps, inspection of the words parsed incorrectly by the grammar using P(Mx) to measure the LPPs of competing parses does not appear to bear out Baayen and Sproat’s results: I found no clear examples of wrong parses of low-frequency words which were attributable to the inclusion of morphs with high P(Mx) but low productivity. In fact, the opposite appeared to be the case: highfrequency words were parsed incorrectly using productive affixation rules which generally apply to low-frequency words. For example, +/ !+,! ‘we’ is the 16th most frequent word in Oliverius. It is parsed by Oliverius as !+rp+,ip!, but by my grammar as !+,rv+sn+in! (cf. #"6 !$'1-rv+sn+in! ‘affairs’ [gen. pl.] and +/.$ !+,rv+&'sa+ia! ‘wash’ [infin.]). There is in fact a straightforward explanation for the apparent lack of agreement between Baayen and Sproat’s results and my own. My approach to Russian word-formation has been underpinned by the use of ‘number of correctly parsed words’ as one of the performance indicators by which a grammar should be evaluated. Consequently, I have focused on determining those rules which apply productively (i.e. to large numbers of words). As all the preceding discussion has shown, this has effectively meant focusing on low-frequency words, instead of attempting to write

212 rules to explain the morphological structure of high-frequency words. Thus I have factored out the effect of high-frequency rules, but in a different way to that proposed by Baayen and Sproat. I believe that my approach & listing morphological exceptions, which tend to be high-frequency words & can be justified both cognitively and computationally. In cognitive terms, the more frequent and familiar a word is, the more likely it is to be stored as an integral, unparsed, object in the mental lexicon (in line with the suggestions of Aronoff and Anshen 1998). On this approach, the ‘irregularities’ characteristic of high-frequency words do not need to be analyzed; they are simply listed. Moreover, in computational terms, the memory cost of storing the ‘idiosyncratic’ high-frequency lexical items as ‘exceptions’ is likely to be immaterial. This is because the proportion of the lexicon which is high-frequency is very low, as can be seen from Figure 21 (on page 213), where the frequencies of the words in Oliverius (1976) are plotted in rank order. (Distributions of this shape are very characteristic of lexical frequency distributions: see Miller and Chomsky 1963: 456ff and Baayen 1992 for further discussion.)

Frequency

(The x-axis of the graph shows only every 19th lexeme.)

214

To summarize, the reason that the CFG based on P(Mx) is able to give generally satisfactory results is that the use of lexical prior probabilities to evaluate competing analyses is not the only way of blocking spurious parses; it is also possible to do so by writing the CFG in a way such that it includes only the more productive rules within the language. This is, in fact, in the spirit of Church’s (1986: 156) suggestion that ‘although the cost mechanism handles a large number of cases, it would be better to aim toward a tighter grammar of morphology which did not overgenerate so badly’. The only disadvantage of this approach is that no analysis can be found for high-frequency words whose idiosyncratic morphological structures are not described by the grammar. However, I have questioned whether it is really necessary to analyze such words, suggesting that a more appropriate approach is simply to list these words as integral items in the lexicon, without highly articulated structure. This concludes the discussion of multiple parses and their evaluation. In the next section, I shall present and discuss empirical results demonstrating the coverage and accuracy of the grammar developed so far. 3.5

Tests of the word-formation grammar In the text up to this point, I have presented and discussed fragments of the

overall context-free grammar. The full context-free grammar which I have developed is given in Appendix 2. This section describes two tests which were carried out on the word-formation grammar. The first of these is a test of the grammar’s coverage in general (that is, an evaluation of its ability to assign correct morphological structure to strings of

215 phonemes), and the second is a test specifically of the grammar’s ability to vocalize and delete jers in the appropriate places. 3.5.1

Test of coverage of the word-formation grammar The grammar, excluding the modifications proposed in 3.3, was implemented

as a Prolog DCG and used to process the 11,290 forms in Table 2 on page 30. The DCG was used to compute all possible parses (and associated morphological tokenizations) for each word in this list, then a separate procedure selected the parse with the lowest cost using the method outlined in 3.4.1. The corresponding tokenization was then compared to that given by Oliverius. For each word processed, there were four possible results: (129) (a) (b) (c) (d)

The grammar was unable to assign any parse to the word The grammar’s morphological tokenization did not match Oliverius’s morphological tokenization The grammar’s morphological tokenization was not significantly different from Oliverius’s morphological tokenization The grammar’s morphological tokenization matched Oliverius’s morphological tokenization

Two morphological tokenizations were deemed to be ‘not significantly different’ if the only differences between the two were either morphs with zero phonological content, or homophonous morphs of the same type, or both. Thus !0;2rn+,in10! does not significantly differ from !0;2rn+,in20!, and !B6#96$rn+sn+%+in6! does not significantly differ from !B6#96$rn+%+in6!. Of the first pair just mentioned, the two members are, respectively, the genitive singular and nominative plural of !8%& !0;2%! ‘hand’. Moreover, the two forms are distinguished in terms of their stress placement. There is no way in which

216 the morphological parser can be expected to disambiguate these without any contextual information (such as information about syntactic function in the absence of stress placement, or vice versa), and thus we should not be at all concerned if the parser fails to find a morphological tokenization exactly matching the original. Note also that extra or missing zero morphs (as in the second pair) do not adversely affect the grammar’s descriptive adequacy unless so-called zero morphs can have phonological content which is invisible in certain surface forms. In this case, the inclusion or exclusion of these morphs by the grammar could potentially be of some consequence. This is relevant here because the morphology proposed in this chapter will form the basis for a stress assignment grammar (see Chapter 5). One could conceive of a situation, for example, where the morphology assigns morphological structure but omits a zero morpheme with some distinctive accentual property, and on this basis an incorrect stress assignment is predicted. However, none of the aspects of the phonology of Russian which I look at in this dissertation suggests that ‘invisible’ morphs are phonologically necessary. In particular, the stress assignment analysis which I propose in Chapter 5 excludes the possibility that zero morphemes may have any effect on stress assignment, so it seems that as far as the phonology of Russian is concerned there is no difference between !B6#96$rn+sn+%+in6! and !B6#96$rn+%+in6!. To summarize the results of parsing the 11,290 words, the percentage of words which fell into each of the four categories in (129) was calculated. The summary of these results is given in tabular format in Table 20, and graphically in Figure 22.

217 Table 20. Summary of results of parsing 11,290 words using grammar in Appendix 2 No of words

%

Parsed words Parse was the same as Oliverius's Parse was not significantly different from Oliverius's Subtotal Parse did not match Oliverius's Parsed words: total

5,216 2,291 7,507 1,978 9,485

46220 20229 66249 17252 84201

Unparsed words

1,805

15299

11,290

100200

TOTAL

Figure 22. Analysis of coverage of morphology grammar

Unparsed 16%

Parse did not match Oliverius's 18%

Parse was not significantly different from Oliverius's 20%

Parse was the same as Oliverius's 46%

218 The performance of the parser could be improved by redefining the morpheme inventory, reinterpreting complex stems as single morphemes, particularly for highfrequency stems. A consequence of this move would be an enlargement of the morpheme inventory. However, it is worth noting that even the existing wordformation grammar has significantly reduced the amount of memory required to store the 11,290 words. The size of the computer file listing the parsed forms was 540,837 bytes; this compares to the size of the morpheme inventory, plus the size of the grammar, plus the size of the list of exceptions (wrongly parsed and unparsed words) (181,231 bytes, 17,097 bytes and 158,125 bytes respectively), totalling 356,453 bytes. This represents a 34% reduction in storage space. Moreover, the reduction in storage space will be greater if one takes into account the fact that the grammar and morpheme inventory will generalize to out-of-vocabulary words (i.e. those not included in the 11,290-word dataset). This gives some indication of the benefits of adopting a morpheme-based approach to word-formation. 3.5.2

Test of the grammar’s treatment of vowel-zero alternations The Prolog DCG was then modified in accordance with the proposals in 3.3 to

account for vowel-zero alternations. This, of course, also involved modifying the morpheme inventory appropriately. Oliverius’s morpheme inventory simply lists allomorphs, including pairs which exhibit a vowel-zero alternation. With a ‘jer theory’ in place, each of these pairs can be collapsed into a single entry in the morpheme inventory (I shall refer to such entries as ‘jer-morphemes’), resulting in an overall reduction in the amount of information which has to be listed lexically. It has to be said, however, that this reduction was not great: the size of the inventory was reduced by only about 4%.

219 Since the objective of this test was to determine the grammar’s ability to vocalize and delete jers correctly, the sample for the test included only word-forms containing at least one jer-morpheme. Of the 11,290 words in Table 2, 3,403 words met this description. Although the vocalization and deletion of jers is presented in traditional generative phonology and Lexical Phonology as a process with ordered rules, it in fact turns into a parsing and generation problem in the current framework. The grammar should generate only those forms in which the jers vocalize and delete in the correct places, if the theory which it embodies is correct. At the same time, the grammar should parse (or accept) only those forms which have vocalized and deleted jers in the correct places, and reject all other forms. Thus one way of testing the grammar is to present it with a list of forms grouped as follows: each entry in the list consists of both the well-formed and ill-formed versions of a given word (e.g. #'> ‘day’ !$4'%! (gen. sg.) versus *#"'> !$'14'%!; #"'$ !$'14'! ‘day’ (nom. sg.) versus *#'$ !$4'! and *#'" !$4'1!). (Here, the latter examples in each case are ‘ill-formed’ only in that jers are vocalized or deleted in the wrong places.) If the grammar is accurate, it should accept all words in the former category and reject all words in the latter category. In designing the experiment, we wish to disregard cases where well-formed words are rejected not because the vocalization or deletion of the jers is incorrect, but because the grammar cannot assign structure to these forms anyway (for example, because they were within the 15299% of unparsed words in Table 20). We also wish to disregard cases where well- or ill-formed words are accepted, but assigned the wrong morphological tokenization. Thus we count only those entries in the list where either a well-formed or ill-formed alternative was accepted and assigned the correct

220 morphological tokenization. Of the 3,403 entries in the list, only 2,618 met this criterion. The results of the test were as follows. In 2,381 cases (90295% of the total), the form with jers correctly vocalized and deleted was accepted and all ill-formed alternatives were rejected. In the remaining 237 cases (9205%), the correct form was rejected and one of the ill-formed alternatives was accepted. In no cases was more than one alternative of the same form accepted, because in any given position (according to the grammar as formulated in 3.3) there is no optionality as to whether a jer deletes or vocalizes. These results show that the theory advanced by Lexical Phonology makes broadly the correct predictions (since the grammar tested does no more than implement the existing theory proposed by Pesetsky in context-free grammar). It is interesting to look more closely at the 9205% of cases that represent exceptions to Pesetsky’s theory, to see if there are any common factors which might enable these exceptions to be explained. Of these 237 forms, 88 have jers which have been incorrectly deleted in prefixes. For example: (130) Actual form Predicted form -,("-.$ -, -,(". -,+) ‘youth’ (instr. pl.) !+6-6$+6#&'+/%+',! !+/6-6$+6#&'! n/a n/a n/a !+/6-6$6#&'+/%+',! !+/6-6$6#&'+%+',! n/a n/a n/a

,:
‘door’ (genitive plural) ‘door’ (genitive singular) ‘town’ (genitive plural) ‘town’ (locative singular) ‘king’ (nominative singular) ‘king’ (dative singular) ‘game’ (genitive singular) ‘game’ (accusative singular) ‘game’ (nominative plural)

As each word is considered by the ALA, it is possible to record the reasoning by which certain candidate accentuations are eliminated in the form of an output log. The output log records the surface form of each word under consideration, counts the word’s possible underlying solutions and records this number. The underlying solutions are then listed in columnar format as in (139) above (with an extra column for the constraints on stem markedness; this is relevant only to plural forms, since according to Halle stress retraction occurs only in the plural for marked stems). As the ALA processes the word #("!":0 !$"'10'/17! ‘doors’ (gen. pl.), therefore, it gives the output log shown in (142): (142) !$"'10'rn+/17 in9! 2 Accentuations !$"'10'!rn, !/17!in9 !$"'10'/!rn

45

Jers !17!in9: no jer

Stem unmarked unmarked

It is assumed in (141) that the genitive singular inflectional morpheme is identical in #(":!) !$"'/10',! ‘door’ and )5!/: !,B0/,! ‘game’, with the )// orthographic alternation reflecting the allophonic alternation between [,] and [,] conditioned by the palatalization or lack thereof of the preceding consonant, and therefore that this morpheme must be assigned the same underlying form !,! in both #(":!) !$"'/10',! and )5!/: !,B0/,!.

263

In other words, there are two ways in which the surface stress of !$"'10'/17! could be explained: either (a) !17!in9 is underlyingly accented (shown in column 1 of solution 1) in which case !$"'10'!rn must be unaccented (this constraint is also shown in column 1 of solution 1) and nothing can be deduced about the placement of jers (shown by the fact that column 2 is empty), or (b) !$"'10'!rn must be post-accented (shown in column 1 of solution 2), nothing is known about the accentuation of !17!in9 (this lack of a constraint is reflected by the absence of !17!in9 in column 1 of solution 2), and !17!in9 must have no initial jer. Both solutions rely on the constraint that the stem !$"'10'rn! must be unmarked, since, for stress to have been retracted, there would have to be a syllable to the right of the one which received surface stress. Since in this case the final syllable of the word is surface stressed, stress retraction cannot have taken place. Having listed all the possible solutions, the ALA now deduces anything that can be deduced about the underlying forms of the relevant morphemes and stems. For example, without any information about how !$"'10'!rn influences word-stress, we know that there are three candidate accentuations for the morpheme: !$"'/10'!, !$"'10'/! and !$"'10'!. However, from (142) we know that !$"'/10'! must now be eliminated as a possibility, because this accentuation would not be consistent with either solution in (142). As already mentioned, we also know that !$"'10'rn! cannot be a marked stem. The output log records these inferences as in (143): (143) . !$"'/10'!rn is not a possible accentuation. . !$"'10'!rn is not a marked stem

264 Now the ALA moves on to consider #(":!) !$"'/10',! ‘door’ (gen. sg.). As the continuation of the output log shows in (144), two solutions are listed: both of these rely upon the operation of the rule which stresses the word’s initial syllable when no other rules can operate. If no other information were available about how !$"'10'!rn influences word-stress, one might suppose that a possible explanation of the surface stress is that !$"'10'!rn is underlyingly accented. None of the solutions listed for !$"'/10',! includes this possibility, however, as this accentuation has already been ruled out in (143). Effectively, the ALA has learnt from (143) that it need not consider underlying solutions for !$"'/10',! which are already known to be impossible. Now, further inferences can be drawn: first, that !$"'10'!rn must not be post-accented, and secondly, that !,!in20 must not be accented. (144) !$"'/10'rn+,in20! 2 Accentuations

Jers

!$"'10'!rn, !,!in20 !$"'10'!rn, !,/!in20

Stem

. !$"'10'/!rn is not a possible accentuation. . !/,!in20 is not a possible accentuation.

The output log continues in a similar fashion for the rest of the words under consideration, making inferences as appropriate, as in (145). (145) !B606$rv+sn1+/6"in12! Accentuations

!B606$!rv, !!sn1, !/6"!in12 !B606$!rv, !/!sn1 !B606$/!rv

3

Jers

Stem

unmarked !6"!in12: no jer unmarked !!sn1: no jer; !6"!in12: no jer unmarked

. !B/606$!rv is not a possible accentuation. . !B60/6$!rv is not a possible accentuation. . !B606$rv+sn1! is not a marked stem

265 !B/606$rv+sn1+1in8! 2 Accentuations !B606$!rv, !!sn1, !1!in8 !B606$!rv, !!sn1, !1/!in8

Jers

Stem

. !B606$/!rv is not a possible accentuation. . !/!sn1 is not a possible accentuation. . !/1!in8 is not a possible accentuation. !260/6-'rn+in1! 2 Accentuations !260/6-'!rn !2606-'/!rn

Jers

Stem

!O!in146

. !2/606-' rn! is not a possible accentuation. . !2606-' rn! is not a possible accentuation. !2606-'rn+/;in22! 1 Accentuations !2606-'/!rn

Jers

!;!in22: no jer

Stem

. !260/6-'!rn is not a possible accentuation. !,B0rv+sn1+/,in20! 1 Accentuations !,B0/!rv

Jers

!!sn1: no jer; !,!in20: no jer

Stem

. !/,B0!rv is not a possible accentuation. . !,B0!rv is not a possible accentuation. . !!sn1 does not contain a jer. !,B0rv+sn1+/;in21! 1 Accentuations !,B0/!rv

!/,B0rv+sn1+,in10! 1 Accentuations !,B0/!rv

. !,B0rv+sn1! is a marked stem

46

Jers are represented by the symbol O.

Jers

Stem

Jers

Stem

!!sn1: no jer; !;!in21: no jer

!!sn1: no jer; !,!in10: no jer

marked

266

This is the end of the first iteration of the ALA. In accordance with (140) (c), the ALA now returns to the beginning of the dataset for the second iteration, the output log for which is shown in (146). Note in particular that only one solution is listed in (146) for !$"'10'/17!, compared with two in the first iteration. This is because the ALA deduced in (144) that !$"'10'!rn could not be post-accented, but this inference was made only after !$"'10'/17! was last considered. (146) !$"'10'rn+/17in9! 1 Accentuations

Jers

!$"'10'!rn, !/17in9!

Stem

unmarked

. !17!in9 is not a possible accentuation. . !17/!in9 is not a possible accentuation. !$"'/10'rn+,in20! 2 Accentuations !$"'10'!rn, !,!in20 !$"'10'!rn, !,/!in20

!B606$rv+sn1+/6"in12! Accentuations

!B606$!rv, !!sn1, !/6"!in12

1

Jers

Stem

Jers

Stem

unmarked

. !6"!in12 is not a possible accentuation. . !6"/!in12 is not a possible accentuation. !B/606$rv+sn1+1in8! 2 Accentuations !B606$!rv, !!sn1, !1!in8 !B606$!rv, !!sn1, !1/!in8

!260/6-'rn+in1! 1 Accentuations !2606-'/!rn

. !O!in1 contains a jer.

Jers

Stem

Jers

Stem

!O!in1

267 !2606-'rn+/;in22! 1 Accentuations !2606-'/!rn

Jers

Stem

Jers

Stem

Jers

Stem

!;in22!: no jer

!,B0rv+sn1+/,in20! 1 Accentuations !,B0/!rv

!!sn1: no jer; !,!in20: no jer

!,B0rv+sn1+/;in21! 1 Accentuations !,B0/!rv

!!sn1: no jer; !;!in21: no jer

!/,B0rv+sn1+,in10! 1 Accentuations

Jers

!,B0/!rv

!!sn1: no jer; !,!in10: no jer

Stem

marked

As can be seen, the ALA was able to make a total of 5 further inferences on the second iteration. In (147), on the third iteration, however, no further inferences are made, which means that the learning is complete and the ALA is terminated. (147) !$"'10'rn+/17in9! 1 Accentuations

Jers

!$"'/10'rn+,in20! 2 Accentuations

Jers

!$"'10'!rn, !/17!in9

!$"'10'!rn, !,!in20 !$"'10'!rn, !,/!in20

!B606$rv+sn1+/6"in12! Accentuations

!B606$!rv, !!sn1, !/6"!in12

!B/606$rv+sn1+1in8! 2 Accentuations !B606$!rv, !!sn1, !1!in8 !B606$!rv, !!sn1, !1/!in8

!260/6-'rn+in1! 1 Accentuations !2606-'/!rn

1

Jers

Stem

unmarked

Stem

Stem

unmarked

Jers

Stem

Jers

Stem

!O!in1

268 !2606-'rn+/;in22! 1 Accentuations !2606-'/!rn

!,B0rv+sn1+/,in20! 1 Accentuations !,B0/!rv

!,B0rv+sn1+/;in21! 1 Accentuations !,B0/!rv

!/,B0rv+sn1+,in10! 1 Accentuations !,B0/!rv

Jers

Stem

Jers

Stem

Jers

Stem

!;!in22: no jer

!!sn1: no jer; !,!in20: no jer

!!sn1: no jer; !;!in21: no jer

Jers

!!sn1: no jer; !,!in10: no jer

Stem

marked

What, then, has the ALA learnt from this dataset? It has managed to eliminate some of the candidate underlying forms; for some, but not all, morphemes it has succeeded in eliminating all but one underlying form; and for some morphemes, none of the candidate underlying forms have been eliminated. The specific results for candidate accentuations are shown in Table 26.

269 Table 26. Candidate accentuations before and after operation of the ALA Morpheme

Candidate accentuations before ALA operation

Candidate accentuations after ALA operation

!$"'10'!rn

!$"'10'! !$"'/10'! !$"'10'/! !17! !/17! !17/! !,! !/,! !,/! !B606$! !B/606$! !B60/6$! !B606$/! !! !/! !6"! !/6"! !6"/! !1! !/1! !1/! !2606-'! !2/606-'! !260/6-'! !2606-'/! !! !/!

!$"'10'!

!17!in9

!,!in20

!B606$!rv

!!sn1 !6"!in12

!1!in8

!2606-'!rn

!!in1 !;!in22

!,B0!rv

!;!in21

!,!in10

!;! !/;! !;/! !,B0! !/,B0! !,B0/! !;! !/;! !;/! !,! !/,! !,/!

!/17! !,! !,/! !B606$!

!! !/6"! !1! !1/!

!2606-'/! !! !/! !;! !/;! !;/!

!,B0/! !;! !/;! !;/! !,! !/,! !,/!

270 This simple example suffices to show that even with logical inference, Halle’s theory is underdetermined: after all constraints have been considered, it is not the case that all morphemes have exactly one underlying form. In section 4.3.1, the LSS was defined as the number of morpheme inventories which are possible within the constraints of a given theory, and 3x (where x is the number of entries in the morpheme inventory) was quoted as an approximation to the LSS. Thus, for the list of morphemes in Table 26, the approximate size of the LSS should be 313, or 1259 - 106. More precisely, for a morpheme inventory where the number of morphemes with x underlying forms is nx, the exact size S of the overall LSS will be given by (148):

(148)

S

>

0 n0

-

1n1

-

...

-

x nx

Clearly S will be zero when n0 > 0: this is a situation which arises when the theory is invalidated by problem words (see 4.3.2.4), and some surface forms can no longer be paired with any underlying forms. It also follows that for the ‘ideal’ theory where every morpheme in the inventory has one and only one underlying form, S will be 1. Using the information from Table 26, we can now compute the exact size of the LSS for Halle’s theory and the dataset in (141), before applying logical inference and without taking into account whether stems are marked or not. The actual figure, which is not too far from the original estimate of 1259 - 106, is given in (149):

(149) 00 - 10 - 22 - 39 - 42 ;

1226 - 106

271 Likewise, the size of the LSS for the same theory and dataset after applying logical inference, again without taking into account whether stems are marked or not, is given in (150):

(150) 00 - 17 - 23 - 33

=

2216 - 102

The fact that logical inference reduces the LSS size from 1,259,712 to 216 (which is a reduction by a power of approximately 226) is indicative that logical inference can be an extremely powerful heuristic for determining the underlying forms for a derivational theory. The measure of LSS after application of logical inference, rather than before, is also useful to give an idea of the true scale of underdeterminacy of a given theory. Logical inference, after all, does not represent the addition of extra constraints to a theory: it simply involves taking a theory to its natural conclusions. 4.3.2.4 Problem words In some cases the constraints of a theory may never be satisfiable. I illustrate this, again using Halle’s (1997) theory, with the following words: (151) -,7)&:6$'/0

!#6&# + ,7 + /%-' + 4 + ,7!

(152) -,7)&6)-.):3"-%)0

!#6&# + ,7 + %-' + ,#&' + /,&8' + 1#2 + ,7! ‘socialist’

For both of these words the three initial morphemes are the same:

‘social’

272

(153) ! #6&# + ,7 + %-'
On the basis of (151), a backwards application of Halle’s theory would lead one to conclude that at least one of the following propositions must be true: (154) (a) (b) (c)

!%-'! is accented !%-'! is post-accented, and there is a following jer !,7! is post-accented

However, none of the solutions for (152) is consistent with any of the possibilities in (154). In other words, once (152) comes under consideration by the ALA, one is forced to eliminate all the possible solutions for (151), leading to the paradoxical situation where a surface form cannot be paired with any underlying forms. Henceforth, I shall refer to the words which have zero underlying forms as ‘problem words’. The question which then arises is: what conclusions should be drawn from problem words? Clearly, in a naïve sense, the existence of problem words for a given theory disproves the theory: it reduces the LSS for a theory to zero, and shows that the rules of the theory are not sufficient to account for 100% of the data, no matter what accentual specification is chosen for the morpheme inventory. Yet there would be problem words even for a theory with a coverage of 9929%, and one would surely not

47

This morphological parse is in accordance with Oliverius (1976). It is debatable whether !#6&#,7%-! might better be considered to be a single morpheme, but either possibility serves to illustrate the point being made above.

273 wish to reject such a theory. If it could be shown that a given theory accounted for all but 021% of the words in Russian, that would clearly be a result of great importance in itself. This point is made in Chomsky and Halle (1968: ix): We see no reason to give up rules of great generality because they are not of even greater generality, to sacrifice generality where it can be attained. It seems hardly necessary to stress that if we are faced with the choice between a grammar G1 that contains a general rule along with certain special rules governing exceptions and a grammar G2 that gives up the general rule and lists everything as an exception, then we will prefer G1. For this reason, citation of exceptions is in itself of very little interest. Counterexamples to a grammatical rule are of interest only if they lead to the construction of a new grammar of even greater generality or if they show some underlying principle is fallacious or misformulated. Otherwise, citation of counterexamples is beside the point.

In other words, a theory can be of interest even if it does not account for all cases, and it would therefore be ill-advised to dismiss Halle’s (or any other) theory & or even to add extra rules to account for problem words & before determining exactly what the exceptions to the theory are. There are also other issues which problem words raise. The first of these is as follows. Suppose we have some word W for which the last remaining solutions have just been eliminated through the operation of step (b) (ii) of (140). The output of (b) (ii) is now fed into step (b) (iii). Since (b) (iii) states that we ‘eliminate all relevant underlying morpheme representations not included at least once in the output of (b) (ii)’, if follows that on the basis of W, all the morphemes included in W will have their remaining underlying forms eliminated. At this point, of course, the size of the LSS is reduced to zero, as we saw in the previous section. Furthermore, when other words containing any of these morphemes undergo step (b) (ii), they too will have all their solutions eliminated, and so on. The cumulative effects may well eventually cause all the underlying forms for virtually the whole morpheme inventory to be eliminated.

274 Thus, as soon as we determine that a theory encounters problems, we find that it is no longer possible to use logical inference to obtain information about the underlying forms of entries in the morpheme inventory. The second issue is that we cannot isolate a list of exceptions to a theory by logical inference (i.e. by using the ALA): identifying exceptions is not simply a matter of listing all problem words. The reason for is that in reality it is ‘problem combinations’ (i.e. combinations of words such as !#6&#,7/%-'4,7, #6&#,7%-',#&'/,&8'1#2,7!, for which no consistent underlying representations can be found), rather than problem words, which challenge a theory. When a ‘problem word’ is encountered during the running of the ALA, it might be that the word in question is an exception to the theory. On the other hand, it might be that the word is not exceptional, but simply could not be analyzed given previous inferences. This could happen if there were a morpheme common to the word in question and a previously considered ‘exceptional’ word. Exceptional words should presumably not be used to infer anything about the underlying forms of their constituent morphemes, as, by definition, an exceptional word would lead one to draw the wrong conclusions. However, there is no means of knowing whether a given word is exceptional, and therefore no means of tracking where unfounded inferences might have previously been made. For this reason, logical inference is of little use in identifying a comprehensive list of exceptions to a given derivational theory. 4.3.2.5 Modifications to the ALA to allow for different theories The descriptions of the ALA, and the processes of logical inference which it incorporates, have up to this point been on as general a level as possible: although illustrations primarily made reference to Halle (1997), the observations in these

275 sections are true of the whole family of derivational theories which includes Halle (1997), Melvold (1989) and Zaliznjak (1985). On the other hand, the initial set of candidate underlying forms will vary from theory to theory, as will the specific backwards grammar rules which the ALA uses to rule candidate forms out. This is made clear in (140), where we stated that the ALA computes what is allowed by the given theory. In section 4.2, we have already seen that Halle (1997), Melvold (1989) and Zaliznjak (1985) all operate with different accentual categories. These differences are recapitulated in Table 27. Table 27. Differences between accentual and other categories posited by Halle (1997), Melvold (1989) and Zaliznjak (1985) Accented morphemes are possible Post-accented morphemes are possible Unaccented morphemes are possible Pre-accented morphemes are possible Dominant morphemes are possible Jers can affect stress Features of stem can affect stress

Halle (1997)

Melvold (1989)

Zaliznjak (1985)

yes

yes

yes

yes

yes

yes

yes

yes

no

no

no

yes

no (all morphemes are effectively recessive) yes

yes (suffixes only)

yes

yes

no

yes

yes

no

Since Melvold (1989) incorporates all the accentual categories and rules that Halle (1997) does, as well as additional ones, it must be the case that for any given word, the number of ways in which the surface stress can be explained by Halle’s theory is smaller than the number of ways it can be explained by Melvold’s theory. This is because the addition of rules to a theory implies that there are additional ways

276 in which a surface form may be explained by underlying forms. Thus, in a sense, Halle’s theory can be said to be more constrained than Melvold’s theory (although, as we shall see in 4.4, the price which is paid for this is that Halle’s theory is not as well equipped to deal with derived words). For instance, with Halle’s theory, we know that any morphemes to the left of the stressed syllable must be unaccented (the only exception being if the morpheme immediately to the left is post-accented), and in general we know nothing about the accentuation of morphemes to the right of the stress. But in Melvold’s theory, we cannot be sure that morphemes to the left of the stress are unaccented: according to her theory, it could be that the stressed syllable is part of a dominant accented morpheme, in which case the morphemes to the left might or might not be accented, and furthermore, they might be either dominant or recessive. In Melvold’s theory, we can be sure only that none of the morphemes to the right of the stressed syllable (if it is not word-initial) is dominant, although these morphemes could still be accented. A word-initial stress could be explained in the same ways that Halle explains it (i.e. by the presence of an accented morpheme, or by the absence of any accented morphemes48), but it could also be explained by the presence of a dominant unaccented suffix (Melvold 1989: 75-78), as long as all the following suffixes are recessive.

48

Halle could explain an initial stress by stipulating that all morphemes in the word are unaccented. Because Melvold states that accentuation is cyclic, however, not all morphemes in a word need to be unaccented for there to be initial stress: just the prefix (if any), root, and first suffix to the right of the root need to be unaccented, as we saw with +,:6,#,-.$, +,:6,#,-.>+ (see section 4.2.2). In Melvold’s theory all suffixes would nevertheless need to be recessive in such cases.

277 Comparison between Halle (1997) or Melvold (1989) on the one hand, and Zaliznjak (1985) on the other, is less straightforward. As indicated in 4.2.3, the mechanisms of Zaliznjak’s theory are somewhat different from those of the other theories: in order to assign stress in Zaliznjak’s theory, it is necessary to consider which morpheme stress ‘attaches’ to (see (134) and (135)). Table 28 compares the procedures which must be followed in order to assign stress under the two theories. Table 28. Procedural comparison of Melvold and Zaliznjak Melvold

Zaliznjak

Find which morpheme determines stress

Find which morpheme determines stress Find which morpheme stress is attached to Find which vowel is stressed

Find which vowel is stressed

The extra stage in derivation required by Zaliznjak’s theory has the consequence that surface stress in a given word will, on average, have many more possible explanations in Zaliznjak’s theory than in Melvold’s theory, even though the number of accentual oppositions in the two theories is comparable. This can be illustrated using the same example as in (135), 3"!.2*)% !&8'10&'/65,2! ‘small sketch’: if we assume that the morphological tokenization of this word is !&8'10&'rn+/65sn+,2sn+in1!, it turns out that in Zaliznjak’s theory there are 324 possible accentual specifications for the word, all of which would yield the correct stress assignment, compared with the 63 possible accentual specifications which Melvold’s theory implies49. Thus two related facts about Zaliznjak’s theory become clear: first, that Zaliznjak’s final rule will in general

49

In both cases, the ‘number of accentual specifications’ is computed without taking into account the abbreviatory notation introduced in 4.3.2.1. Computed on the same basis, therefore, the ‘number of accentual specifications’ implied by Halle’s theory for !#&6-! would be 30 (see Table 23).

278 reduce the extent to which logical inferences can be made, and secondly, that Zaliznjak’s theory is considerably less constrained than Melvold’s. 4.3.2.6 Conclusions from the ALA We have already seen in 4.3.2.4 that the ALA is able to show whether a given theory will encounter problems: however, if it is found that a theory is problematic, then this is all that the ALA will indicate. Thus the ALA will give no indication of the scale of the problems associated with a theory, nor will it give reliable results regarding the optimal specification of the morpheme inventory. Since Halle’s theory was used to illustrate this, we clearly also have the specific result from 4.3.2.4 that Halle’s theory is problematic, as it was unable to account for both -,7)&:6$'/0 !#6&#,7%-'4,7! ‘social’ and -,7)&6)-.):3"-%)0 !#6&#,7%-',#&',&8'1#2,7! ‘socialist’. In other words, we know that the coverage of Halle’s theory must be less than 100%. When I applied the ALA to Melvold’s theory using a dataset of 9,512 nouns as the basis for drawing logical inferences50, I obtained the result that this theory is also problematic. The following example will suffice to demonstrate this. If we assume Melvold’s theory holds, we are led to make the following chain of inferences:

50

For information on the compilation of the large datasets used for computational tests, see 1.5.

279 Table 29. Demonstration that Melvold’s theory is problematic Word Morphological Inferences tokenization, following Oliverius (1976) !-!sn1 is not dominant 51 '&3&:6, ‘beginning’ !4%c+&8/%rv+-sn1+6in18! If !,!svr is dominant, it must also 4!&:()6, ‘rule’ !90/%"'rv+,svr+-sn1+6in18! !,#):."6$ ‘parent’

!06$'rv+/,svr+&'1-'sn+in1!

!,:#)'& ‘homeland’

!0/6$'rv+,4sn1+%in3!

568rn1+>/in3# #9%+5>%0'rn+)*>rn1+/$rv+D%,>rn1+>/in3# #4/>c+*/>c+-'>2+rv1+>rn1+c+rn1+rn1+>/in3# #7')*'>)+ra+,'>21rn++,->rn+8;rn+,('>ra3+)*>rn1+>/in3# #*'2>c+0%>c++,/>rv1+,>ra1+>%3rn1+c1+&>/3rv+>rn1+c1+&>/+rv+,'>rn1+c1+%1>80rn+%->ra+D>/rvc+c+>/4rv+D>/rvc+*>ra1+*>ra1+>%+,'rn+c+5),>rv+>rn1+c+0)6>rv+>rn1+c+5>8+3rv+>rn1+c1+,('21'>20'ra+>rn+c+->%0rv+>rn1+c+rn1+>/in3# #5%>c+rn2+c+rn1+/-'rv+D>)rvr+Drn1+>/in3# #1'>2$'rn+D>%*rn1+D%3>sn+c+-'>20'rv1+D>2rvs+rn1+>/in3# #+%>c+-'>2+rv3+,'>rn1+c+5%>c++>%$rp+>rn2+2$rv+%->ra+D>/rvc+2*'rn3+8,('rv+>rvr+D>2*'ra+>)3sn+c+$>;Crv+> sn1+sn1+< in1! !6A>c+0>%*rv+>sn1+c+9,&>rv+>sn1+c+$,C>rv+> sn1+c+9>;#2rv+> sn1+c1+&8'10'>1$'ra+>sn+c+">6$rv+>sn1+c+sn1+c+c+#>6Arp+>sn2+c+5>%#rv+>sn1+c+0>6"'ra+>14'sn3+‘belt’ 4,>-',:0 !967%#467! ‘pertaining to a belt’ -4,:-,< ‘means’ -4,-,::.$, the suffix !14'!sn3 is still present in the form of the allomorph !4'!, but here the following morpheme (the verbal suffix !%!) is dominant: since it is to the right of !4'!, it is the morpheme which determines the word-stress instead of !4'!. I conclude this section by stating that Zaliznjak’s ‘basic rules’ appear to be inferior to Melvold’s theory in three ways: first, their scope is more limited than that of Melvold’s theory, since they apply only to derived nouns (and even then, Zaliznjak states that they do not apply in the inflection of derived nouns). In the sections above, I showed that it is entirely reasonable to suppose that ‘morpheme-based’ rules can account for the stress of non-derived words. Secondly, as we saw in 4.2.3, careful examination of Zaliznjak’s theory shows that it can objectively be said to be less elegant than Melvold’s, since it contains more rules. These two reasons are the principal reasons for rating Zaliznjak’s theory less highly. The third reason is that the coverage of Melvold’s theory appears to be (marginally) better, for which we saw evidence in 4.4.5-4.4.6. 4.5

Summary The results which have been obtained in this chapter can be summarized as

follows. Essentially, the rule ‘stress the rightmost dominant suffix if there is one, otherwise stress the leftmost accented morpheme if there is one, otherwise stress the initial syllable of the word’, which I have referred to as Melvold’s theory, is sufficient to account for 95234% of a corpus of 9,512 nouns, derived and non-derived, given the

308 right accentual specification of the morpheme inventory. However, the ‘right accentual specification’ needs to be supplied in addition to the theory: all three theories, including Melvold’s, were under-determined in that they did not include enough information for them to be tested as they were. Aside from the limitations in coverage already discussed, I have also shown that Melvold’s rules are less good at predicting the stress of verbal infinitives and adjectives. It therefore seems to me that this rule is limited in its applicability to nouns, and I believe that an important avenue for future research will be to formulate and test better rules for parts of speech other than nouns.

309

Chapter 5: Stress assignment: a new analysis 5.1

Introduction In Chapter 2 and Chapter 3, I developed detailed context-free grammars of

Russian syllable structure and word-formation. Chapter 4 compared three alternative theories of Russian stress assignment: the evidence presented in that chapter suggested that of the three theories, Melvold’s provides the most promising startingpoint in formulating a broad-coverage account of stress in Russian words. In this chapter, I show how the two grammars already proposed can be integrated into a single ‘phonology’ which is able to solve diverse grammatical problems. I also take the principal ideas of Melvold’s theory and show how they can be implemented within this phonology of Russian. This re-implementation demonstrates that although Melvold’s theory is couched in ‘procedural’ terms, its most important features also have a declarative implementation. This is no surprise: Coleman (1998: 9) demonstrates that a typical ‘imperative’ phonological analysis can be re-expressed declaratively, and shows (Coleman 1998: 97ff.) that any phrasestructure grammar (even one more powerful than context-free) can be viewed nonderivationally. The grammar in this chapter formalizes the salient aspects of Melvold’s theory, but is no more powerful than context-free. The result of this is that the theory is shown to have all the advantages outlined in section 1.3.2. We shall see in this chapter how stress assignment fits in with the rest of the grammar (the morphological and syllable structure modules). This has not always been clear in mainstream generative phonology, which has often treated stress assignment and syllabification as somewhat separate mechanisms. Throughout this

310 dissertation, I have emphasized the importance of ensuring that a complex grammar is coherent. It will become clear in this chapter not only that a morpheme-based account of Russian stress can cohere with other generally-accepted principles of linguistic theory (such as the analysis of morphological and syllable structure), but that coherence is desirable: without it, the theory is more complicated than it need be. The analysis which will be proposed is not a straight re-implementation of Melvold’s theory. In certain respects, I depart from, or add to, the theory. For one thing, Melvold’s theory does not tell us how to parse strings morphologically; it assumes that this information is given. The omission of a formal description of Russian morphology in effect leaves Melvold’s theory under-determined, as we saw in section 4.3. It is possible to conceive of a different grammar of morphology which would yield different stress assignments in conjunction with the same (Melvold’s) theory about Russian stress assignment. It is important to appreciate, however, that Melvold’s theory cannot assign stress at all without some kind of word-formation grammar. Thus I would argue that the addition of a word-formation grammar is not a ‘departure’ from Melvold’s theory; it is better seen as the unavoidable filling in of a gap. But there are other tangible differences between the theory I propose and Melvold’s theory, in the sense that the two theories can (at least, potentially) make different predictions about the stress assignment of particular words. I shall argue that post-accentuation and ‘jer stress retraction’ are more simply viewed as functions of syllabification in modern Russian, rather than (for example) looking at postaccentuation as the ‘projection of a left bracket’ (Idsardi 1992, Halle 1997). In other words, Melvold’s rules are more powerful than they need be: a simpler and more

311 general theory can account for virtually all of the cases explained by Melvold’s theory. But this chapter does more than just provide a new account of stress in Russian. It also presents a computational phonology of Russian which synthesizes the proposals advanced up to this point; using this phonology, one can find answers to any of the following questions:

,

What is a word’s syllable structure?

,

What is a word’s morphological structure?

,

When do jers vocalize, and when do they delete?

,

Which syllable of a word is stressed?

,

Is a word well-formed or ill-formed?

,

What are the possible words of the language?

It goes without saying that imperfections in the model may give rise to incorrect solutions to these problems, but at least these can be identified, and the model then improved upon. In other words, once again a standard is set against which future proposals can be measured. 5.2

Context-free phonology and stress in Russian In Chapter 4, I showed that Melvold’s theory cannot be implemented and

tested without two elements which are missing from her account: (a) a grammar which shows the location of morpheme boundaries within words (a word-formation grammar) and (b) an extensive morpheme inventory (a list showing how morphemes are ‘underlyingly’ specified for accentuation and dominance).

312 The word-formation grammar presented in Chapter 3 fills the first of these two ‘gaps’ in the theory. Because this grammar can be used for parsing, it is able, given only the sequence of phonemes in a word, to deduce that word’s morphological structure (not just where the morpheme boundaries are within the word, but also its hierarchical structure). The second of the gaps, the absence of a complete morpheme inventory, is filled by Appendix 3: the basic entries in this appendix come from Oliverius (1976), and the accentual tagging of the entries was obtained by using the ALA and trial and error as described in 4.3.3. With all the necessary elements in place, I show in sections 5.2.1-5.2.6 how the principal ideas from Melvold’s theory can be implemented as a simple extension to the word-formation grammar (including the morpheme inventory) by enriching the word-formation grammar with appropriate features. Although the word-formation grammar is context-free, and therefore the enriched grammar is also context-free, it is possible to ensure that the enriched grammar makes exactly the same predictions as the equivalent parts of Melvold’s ‘procedural’ grammar. If one chooses to view the overall grammar in a declarative light, then, for each structure presented to it, this grammar is able to ‘fill in’ all unspecified features (such as stress), or reject the structure if there is no possible way of filling in the underspecified features consistent with those that are pre-specified. 5.2.1 Encoding which morpheme determines stress In order to encode which morpheme determines stress, I suggest that the feature [!stress] should be included in the word-formation grammar as well as in the syllable structure grammar. A morpheme which ‘determines’ the stress of a word is [+stress]; any other morpheme is [$stress]. This means that one and only one

313 morpheme in each word can be [+stress]. Thus, for example, a word’s [+stress] morpheme would be its rightmost dominant morpheme, or its leftmost (post-)accented morpheme if the word contains no dominant morphemes. The feature [+stress] (in the word-formation component of the grammar) corresponds to the leftmost asterisk in Melvold (1989), and to the left parenthesis of Idsardi (1992) and Halle (1997) (to be precise, the left parenthesis is to the left of the [+stress] morpheme). However, a morpheme which is [+stress] in one word need not be [+stress] in another word, so clearly [+stress] is not a feature which is specified as part of the lexical entry of each morpheme in the inventory. Thus the constraints which ensure that [+stress] and [$stress] are distributed as they should be, with one and only one [+stress] morpheme per word, must be part of the word-formation grammar. According to Melvold’s theory, the distribution of [+stress] and [$stress] will be determined solely by the lexical accentual features of morphemes in a word and their positions. We saw in 4.2.2 that these features are (a) dominant versus recessive (which can alternatively be thought of as [!dominant], with [$dominant] being simply another name for recessive), and (b) (post-)accented versus unaccented. I shall leave aside post-accentuation for now, and start by considering the simple case where morphemes are [!accented]. As observed in section 4.2.2, stem ‘markedness’ is also a lexically-specified accentual feature; I return to this later as well. Suppose now that we stipulate that the feature [!stress] is possessed not just by the terminal symbols in a word’s morphological structure (that is, by the morphemes themselves), but also by all the non-terminal symbols in the tree. We shall say that any symbol which dominates exactly one [+stress] terminal symbol is also [+stress], and any symbol which dominates no [+stress] terminal symbols is [$stress].

314 (The stipulation above that one and only one morpheme can be [+stress] is equivalent to stipulating that no symbol may dominate more than one [+stress] terminal symbol.) For example, the constituent structure of -4"7)&6$',-.$ !#9'1&#+,7+%-'+4+6#&'+! ‘speciality’ (excluding the information on parse probability given in Figure 20 in Chapter 3), in which we assume that !%-'! is [+dominant] and that no [+dominant] morphemes stand to its right, could be represented as in Figure 23. Figure 23. Parse tree for -4"7)&6$',-.$

noun [+stress]

noun-stem

infl. ending

[+stress]

[$stress]

adj-stem

sn

[+stress]

[$stress]

adj-stem

sa

[+stress]

[$stress]

noun-stem

sa

[+stress]

[$stress]

rv

sn

[$stress]

[$stress]

+67$)+

&?

[$stress]

[$stress]

#,7

[+stress]

-

[$stress]

(+)7

[$stress]

315 In this grammar, it is equally valid to think of the value of [!stress] as being inherited by small morphological units from larger morphological units, or the reverse. It is, of course, one of the key features of a declarative grammar that the ordering of processing does not matter. It also follows from the stipulations above that the start symbol of the grammar (representing a whole word) must always be [+stress], since it will always dominate exactly one terminal symbol (morpheme) which is [+stress]. As we have seen, the word-formation grammar (apart from the jer deletion and vocalization rules) consists entirely of left-embedding context-free rules. Each rule in the grammar is of one of the following four forms: (160) (a) (b) (c) (d)

X stem stem stem

' ' ' '

stem stem root prefix

inflection suffix root

(where X stands for noun, verb, adjective, adverb etc)

There is, in fact, a general way of enriching each of these four rule schemata so as to capture the four basic ideas from Melvold’s theory, namely: (161) a) b) c) d)

A dominant suffix always ‘wins’ the stress from a stem to its left. An accented left-hand constituent always ‘wins’ the stress from its [$dominant] sister. A [+accented] right-hand constituent always ‘wins’ the stress from its [$accented] sister. If no stress is assigned (e.g. because no morphemes in the word are accented), the word is stressed on its leftmost syllable by default.

316 We shall consider these in order. First, to capture the idea that a dominant suffix always determines the stress, regardless of what is to its left, we need to map the following feature structures onto all four of the rule schemata in (160): (162) [+accented, 3 stress]

'

[$stress]

[+dominant, 3 stress]

Thus, for example, (160b) becomes (163): (163) stem [X] {X Y Z

stem suffix [Y] [Z] [+accented, 3 stress], [$stress], [+dominant, 3 stress]} '

9 9 9

The reason that the stem in (163) is specified as [+accented] is that, as Melvold puts it, the newly formed stem enters the next ‘cycle’ accented. If it were not specified in this way, the new stem would lose the stress to a [+accented, $dominant] suffix added to its right, which would be incorrect. Now, according to Melvold’s theory, only suffixes can be [+dominant]. But if non-suffixes are left lexically unspecified for the feature [!dominant], the wrong predictions will be made. For example, in a ‘simple’ stem consisting of a [+accented] prefix and a [$accented] root, the root will unify with [+dominant, 3 stress], causing stress to be incorrectly assigned to the root. Thus the principle that only suffixes can be [+dominant] does need to be encoded in the grammar somehow. One possible way of doing this is to specify every non-suffix lexically as [$dominant]; however, this clearly misses an important generalization. A preferable alternative is to modify (160d) to read as follows:

317

(164) stem ' [X] {Y, Z 9

prefix root [Y] [Z] [$dominant]}

We now move on to consider how (161b) can be encoded in the grammar. This can be done by adding the constraints in (165). These state that a [+accented] left-hand constituent inherits the value of [!stress] of the mother node, providing its sister is [$dominant]. It does not matter whether the left-hand constituent is assigned the feature [+accented] lexically, or by the grammar. Again, the mother node is [+accented] for the same reason as in (162)-(163). (165) [+accented, 3 stress]

'

[+accented, 3 stress]

[$dominant]

This rule will apply if the left-hand constituent is a prefix and the right-hand constituent a root. In this case, it does not matter, because we want a [+accented] prefix to be assigned the stress, regardless of the accentual properties of the root to its right. Third, we consider the constraints necessary to ensure that a [+accented] righthand constituent inherits the mother node’s specification for [!stress] if its sister is [$accented]. The necessary constraints are as follows: (166) [+accented, 3 stress]

'

[$accented]

[+accented, 3 stress]

Finally, we need constraints to allow for the possibility that both of the subconstituents are [$accented]. These constraints are as follows:

318

(167) []

'

[$accented]

[$accented, $dominant]

This set of constraints will apply to words consisting entirely of unaccented morphemes (e.g. 5,:!,#8 !B606$rn+;in22! ‘town’ [dat. sg.]; cf. Table 21 on page 237), or those consisting of a stem, the root and innermost suffix of which are both unaccented (e.g. +,:6,#,-.>+) !+6-6$+6#&'+/%+',!
!
[$stress]

$

;

0

%

nucl.

onset

rime

2

[+stress]

syllable

rime

[?stress]

nucl.

nucl.

onset

%

[+stress]

onset

[+stress]

syllable

weakfoot

rime

[+stress]

syllable [+stress]

strongfoot phonological word

323 In certain identifiable respects, the ‘syllabification’ interpretation of postaccentuation differs in its predictions from Melvold’s theory. First, it is conceivable in Melvold’s theory that a morpheme with a final vowel could be post-accented, but such a morpheme could not exist in the proposal which I make. Secondly, it is also conceivable in Melvold’s theory that there might be a case of the type !…V1C1+C2V2…!, where the morpheme ending in C1 is post-accented, but C1C2 is not a possible onset. According to Melvold’s theory, V2 would be stressed, but according to the ‘syllabification’ view V1 would be stressed. Where two theories differ in their predictions as here, it clearly cannot be the case that both theories are right. The question therefore arises as to whether the ‘syllabification’ interpretation of post-accentuation is really justified as a simplification to Melvold’s theory: perhaps the syllabification interpretation fails to cover certain words in Russian which are covered by Melvold’s theory. This is an empirical question which can be resolved only by comparing the overall coverage of the two theories; I shall return to address this question in 5.3. If the ‘syllabification’ interpretation of Russian stress assignment is correct, however, we see clearly from Figure 24 how it is possible for stress to be borne both by syllables, and by segments, and, indeed, by morphological constituents. Each of the grammar modules, word-formation and syllable structure, simultaneously places constraints on the surface form, filling in the specification for [!stress] as appropriate. This is entirely in accordance with the standard constraint-based view of language (cf. Figure 25, modified from Bird 1995: 30):

324 Figure 25. The constraint pool syntax

phonology

utterance

semantics

morphology

The force of Figure 25 is essentially that each grammatical module constrains utterances, and an utterance which is well-formed must not violate any of these constraints. Because the modules operate in parallel, not in series, the output of one module cannot be said to be the ‘input’ for any other: instead, the utterance is the simultaneous output of all the modules. Now, if a language L is constrained by two grammars G1 and G2, which define languages L1 and L2 respectively, then L must be at the intersection of L1 and L2: the strings of L must comply with the constraints of both grammars in order to be wellformed. Thus stress assignment in Russian can be seen as a function of the intersection of morphology and syllable structure. Chomsky (1963: 380) shows that the union or intersection of a context-free and linear grammar is a context-free grammar; and since the current word-formation grammar is context-free and the syllable structure grammar is linear, it follows that Russian stress assignment itself is context-free.

325 5.2.4

Jer stress retraction The next issue which needs to be examined is that of ‘accented jer’ stress

retraction as described in Melvold’s theory. Just like post-accentuation, this type of stress retraction can be seen as a function of syllabification. Indeed, I shall argue that the theory of stress assignment need make no explicit reference to jers per se. This is because if jers are morphological objects which manifest themselves phonologically sometimes as a vowel and sometimes as zero (as argued in 3.3), then, as we shall see, ‘jer stress retraction’ will fall out naturally as a result of syllabification. The majority, if not all, the cases of ‘accented jer’ stress retraction cited by Halle (1997) and Melvold (1989) involve inflectional endings consisting of a jer (nominative singular or genitive plural) added to nouns with post-accented stems. For example, !2606-'rn+/;in! ‘king’ (dat. sg.) is stressed on the inflection because !2606-'!rn is post-accented; in !260/6-'rn+in!, the stress is supposedly retracted to the first available syllable before the end of !2606-'!rn because the inflectional ending contains a jer. However, this kind of stress retraction can also be viewed as a function of how the final consonant of !2606-'!rn syllabifies, just as was the case in 5.2.3. After all, the feature which !2606-'rn+/;in! and !260/6-'rn+in! have in common is that in both cases the !-'! in stem-final position is in the stressed syllable. The rule of jer stress retraction can be formulated as follows. We stipulate, using the rules in the syllable structure grammar, that a syllable is [+stress] if its coda is [+stress]. (Using previous rules, a syllable could be [+stress] only if either its nucleus or its onset were [+stress].) We are now stating, then, that [+stress] spreads throughout a syllable to all of its constituents, onset, nucleus and coda. In other words, the rule of stress ‘retraction’ is not retraction

326 at all, but is no different in kind from the post-accentuation rule formulated in (170) above. (171) rime [X] {

'

nucleus [Y]

X 9 [3 stress], Y 9 [3 stress], Z 9 [3 stress] }

coda [Z]

To see how this works, consider the constituent structure of #8!&:% !$;0ra+%2sn+in1! ‘fool’ (nom. sg.), shown in Figure 26.

327 Figure 26. Morphological/phonological structure of #8!&C% !*D.ra+#4sn+in1!

noun [+stress]

noun-stem

infl. ending

[+stress]

[$stress]

adj-stem

sn

[+stress]

[$stress]

$

;

0

%

[+stress]

[+stress]

rime

[+stress]

nucl.

nucl.

onset

2

coda

[+stress]

onset

[+stress]

rime

[+stress]

[?stress]

syllable

syllable

weakfoot

strongfoot

[+stress]

phonological word

328 As in the previous section, it is important to clarify exactly how the ‘syllabification’ interpretation of ‘accented jer’ stress retraction differs in its predictions from Melvold’s theory, if at all. The answer, again, is that there are cases where the predictions could potentially differ. One such case is as follows. In Melvold’s theory it is possible to conceive of morphemes with an inherently accented jer. A morpheme of this type, if it were [+stress], would always cast stress onto the previous syllable; effectively, it would be a pre-accented morpheme. In the context-free account which I am proposing, it is also possible to conceive of a morpheme where [+stress] is lexically required to associate with the jer (cf. (169)). However, if the morpheme were [+stress] and the jer were deleted, a situation would arise where no terminal symbols would be able to associate with the feature [+stress], resulting in default stress assignment to the initial syllable (the mechanism for which will be described in 5.2.6). Of course, the initial syllable might not be the same as the syllable immediately preceding the morpheme with the deleted jer. The only other case where the two theories differ in their predictions is in words of the type !…V1C1+JC2V2…!, where …V1C1 is a post-accented morpheme, C1C2 is a possible onset, J is a jer and V1, V2 are vowels. In this case, Melvold’s theory would predict that stress surfaces on V1, whereas the ‘syllabification’ interpretation predicts that stress will be on V2. Since the two alternative theories make different predictions, we have to decide once again which of the two sets of predictions matches the observed data better. Just as in 5.2.3, this question can be resolved only by empirical analysis, and again I leave discussion of this until section 5.3.

329 5.2.5

Plural stress retraction We now turn to the rule of stress retraction in certain plural forms. According

to Halle and Melvold, in plural nouns with a marked stem, stress surfaces one syllable to the left of where it would have surfaced had the stem been unmarked. The reason that implementation of this rule is at first sight not straightforward is that the grammar somehow needs to be able to recognize marked stems so that stress can be retracted in those words and no others. As stated in section 4.3.2.1, a stem is not necessarily the same as a morpheme, but may consist of more than one morpheme. This suggests that ‘markedness’ would have to be included as a feature of pre-terminal, not terminal, symbols in the morphological representation. The reason that this is problematic is that morphemes, the terminal symbols of the word-formation grammar, are the only level at which markedness can be ‘lexical’, since the morpheme inventory is effectively the list of terminal symbols in the word-formation grammar. However, as noted by Lagerberg (1992: 11), ‘derived words in Russian in the overwhelming majority of cases have fixed stress’. It follows from this that stress retraction should occur only in stems where the root morpheme immediately precedes the inflectional ending. This suggests that ‘stem markedness’ arises from a lexical property of the last morpheme in a stem (which will be the root in the case of a nonderived word).58 If this is true, then each morpheme in the inventory can be lexically assigned a feature which we might call [!pl(ural)_retr(action)]. Thus, for example, !*',+!rn is [+plural_retraction] (cf. 9)+&: !*',+rn+/%in3! ‘winter’ (nom. sg.), 9):+&+)

58

This proposal would, moreover, be consistent with Siegel’s (1977) Adjacency Constraint. See Pesetsky (1979) for detailed arguments for the applicability of the Adjacency Constraint to Russian.

330 !*'/,+rn+%+',in5! ‘winters’ (instr. pl.)). The fact that all derived words purportedly have fixed stress is then explained by a lexical principle that all suffixal morphemes are [$plural_retraction]. In fact, not all derived words do have fixed stress, since deadjectival abstract nouns whose stem end in -)'- !,4sn1! and -,.- !6&sn1! (e.g. 568 weakfoot( Z1,[left:_,right:V2],W), strongfoot( Z2,[left:V1,right:R],S), {unif(V2,[peripheral:1|_]), Z is Z1 + Z2, voicing_assim(V1,V2)}. pwd(Z,R,pwd(S)) --> strongfoot( Z1,[left:_,right:R],S), {Z is Z1 + 1}. weakfoot(Z, [left:L,right:R],weakfoot(W1,W2)) --> syllable( Z1, [left:L, right:V2], W1), weakfoot( Z2,[left:V1,right:R], W2), {unif(V2,[stress_det:0,stress:0,peripheral:0|_]), Z is Z1 + Z2, voicing_assim(V1,V2)}.

393 weakfoot( Z, syllable(

[left:L,right:R],weakfoot(W1)) --> Z, [left:L, right:R], W1), {unif(R,[stress_det:0,stress:0|_])}.

strongfoot( Z, syllable(

[left:L,right:R],strongfoot(S)) --> Z, [left:L, right:R], S), {unif(R,[stress_det:1,stress:1,pl_retr:0,vfv:0|_])}.

strongfoot( Z, syllable(

[left:L,right:R],strongfoot(S,W)) --> Z1, [left:L, right:V2], S), weakfoot( Z2,[left:V1,right:R],W), {unif(V2,[stress_det:1,stress:1|_]), unif(R,[pl_retr:0,peripheral:0,vfv:0|_]), Z is Z1 + Z2, voicing_assim(V1,V2)}.

strongfoot( Z, syllable(

[left:L,right:R],strongfoot(SA,SB)) --> Z1, [left:L, right:V2], SA), syllable( Z2, [left:V1, right:R], SB), {unif(V2,[stress_det:0,stress:1|_]), unif(R,[stress_det:1,stress:0,pl_retr:1,vfv:0|_]), Z is Z1 + Z2, voicing_assim(V1,V2)}.

strongfoot( Z, syllable(

[left:L,right:R],strongfoot(SA,SB,W)) --> Z1, [left:L, right:V4], SA), syllable( Z2, [left:V3, right:V2], SB), weakfoot( Z3, [left:V1,right:R],W), {unif(V4,[stress_det:0,stress:1|_]), unif(V2,[stress_det:1,stress:0|_]), unif(R,[peripheral:0,pl_retr:1,vfv:0|_]), Z is Z1 + Z2 + Z3, voicing_assim(V1,V2), voicing_assim(V3,V4)}.

394 syllable( onset( rime(

Z, [left:LM, right:RM], syllable(O,R)) --> Z1, [left:LM, right:LRM], O), Z2, [left:RLM, right:RM], R),

{ Z is Z2 + log(Z1/71637), cv_interdependency(RLM,LRM), unif([stress_det:H1,stress:H2,peripheral:H3|_],LRM), unif([stress_det:H1,stress:H2,peripheral:H3|_],RLM) }. syllable(

Z, [left:LM, right:RM], syllable(R)) --> rime( Z, [left:LM, right:RM], R), {unif(LM,[onset:[]|_])}.

onset(

32, [left:P1, right:P4], onset(S1,S2,S3,S4)) --> segment(P1,S1), segment(P2,S2), segment(P3,S3), segment(P4,S4), {unif( [sonor:0, cont:1, coron:0, anter:1 |_], P1), unif( [cont:1, sonor:0, coron:1, anter:1 |_], P2), unif( [cont:0, del_rel:0 |_], P3), unif( [cons:1, voc:1 |_], P4), voicing_assim([sonor:1|_],P4), voicing_assim(P4,P3), voicing_assim(P3,P2), voicing_assim(P2,P1), cv_interdependency(P4,P3), cv_interdependency(P3,P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2), unif([stress_det:H|_],P3), unif([stress_det:H|_],P4)}.

395 onset(

2, [left:P1, right:P3], onset(S1,S2,S3)) --> segment(P1,S1), segment(P2,S2), segment(P3,S3), {unif( [cont:1, sonor:0, coron:0, anter:1 |_], P1), unif( [cont:1, sonor:0, anter:1 |_], P2), unif( [cons:1, sonor:1 |_], P3), voicing_assim([sonor:1|_],P3), voicing_assim(P3,P2), voicing_assim(P2,P1), cv_interdependency(P3,P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2), unif([stress_det:H|_],P3)}.

onset(

9, [left:P1, right:P3], onset(S1,S2,S3)) --> segment(P1,S1), segment(P2,S2), segment(P3,S3), {unif( [cont:1, sonor:0, coron:0, anter:1 |_], P1), unif( [cont:1, sonor:0, anter:1, coron:1 |_], P2), unif( [cont:0, del_rel:0 |_], P3), voicing_assim([sonor:1|_],P3), voicing_assim(P3,P2), voicing_assim(P2,P1), cv_interdependency(P3,P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2), unif([stress_det:H|_],P3)}.

396 onset(

717, [left:P1, right:P3], onset(S1,S2,S3)) --> segment(P1,S1), segment(P2,S2), segment(P3,S3), {unif( [cont:1, sonor:0, anter:1 |_], P1), unif( [cont:0, del_rel:0 |_], P2), unif( [cons:1, sonor:1, nasal:0 |_], P3), voicing_assim([sonor:1|_],P3), voicing_assim(P3,P2), voicing_assim(P2,P1), cv_interdependency(P3,P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2), unif([stress_det:H|_],P3)}.

onset(

1643, [left:P1, right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cont:1, sonor:0, coron:1, anter:1 |_], P1), unif( [cont:0 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

onset(

74, [left:P1, right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cont:1, sonor:0, coron:1, anter:0 |_], P1), unif( [cont:0, del_rel:0 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

397 onset(

1200, [left:P1, right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cont:1, sonor:0, anter:1 |_], P1), unif( [cons:1 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

onset(

1252, [left:P1, right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cons:1, voc:0, del_rel:0 |_], P1), unif( [cons:1, voc:1 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

onset(

53, [left:P1, right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [sonor:0, del_rel:0 |_], P1), unif( [cons:1, voc:0, sonor:1, nasal:1 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

398 onset(

885, [left:[sonor:0|P1], right:P2], onset(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [sonor:0 |_], P1), unif( [cons:1, sonor:1, nasal:0 |_], P2), voicing_assim([sonor:1|_],P2), voicing_assim(P2,P1), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

onset(

23295, [left:P1, right:P1], onset(S1)) --> segment(P1,S1), {unif( [cons:1|_], P1), voicing_assim([sonor:1|_],P1)}.

/* Rimes */ rime(

Z, [left:L, right:L], rime(N)) --> nucleus( NUC, [left:L], N), {Z is log(NUC/71637)}.

rime(

Z, [left:LM, right:RM], rime(N,C)) --> nucleus( NUC, [left:LM], N), coda( COD, [right:RM], C), {Z is log(NUC/71637) + log(COD/71637), unif([stress_det:H1,stress:H2,peripheral:H3|_],RM), unif([stress_det:H1,stress:H2,peripheral:H3|_],LM)}.

/* Single-consonant codas */ coda(

5535, [right:P], coda(C)) --> segment(P,C), {unif(P,[cons:1|_]), voicing_assim(P,[sonor:1|_]), cv_interdependency([cons:1|_],P)}.

399 coda(

64, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cons:1, voc:1|_],P1), unif( [cons:1, voc:0|_],P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

79, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [del_rel:0|_], P1), unif( [cont:0, coron:1, anter:1, del_rel:0|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

12, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cons:1, voc:0, sonor:1, nasal:1|_], P1), unif( [cont:1, sonor:0, coron:1|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

400 coda(

10, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cons:1, voc:0, sonor:1, nasal:1|_], P1), unif( [cont:0|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

1, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cont:0, coron:0|_], P1), unif( [cont:1, sonor:0, coron:1, anter:1|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

1, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [sonor:0, cont:1, coron:1, anter:1|_], P1), unif( [cont:0, coron:0, anter:0|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

401 coda(

15, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [sonor:0, cont:1, coron:1, anter:0|_], P1), unif( [cont:0, coron:1, anter:0|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

8, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [sonor:0, cont:1, coron:1, anter:1|_], P1), unif( [cons:1, voc:0, anter:1|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

coda(

16, [right:P2], coda(S1,S2)) --> segment(P1,S1), segment(P2,S2), {unif( [cont:0, del_rel:0|_], P1), unif( [cons:1, voc:1, coron:1|_], P2), voicing_assim(P2,P1), voicing_assim(P1,[sonor:1|_]), cv_interdependency([cons:1|_],P2), cv_interdependency(P2,P1), unif([stress_det:H|_],P1), unif([stress_det:H|_],P2)}.

nucleus(

32666, [left:L], nucleus(P)) --> segment(L,P), {unif( [cons:0, voc:1|_], L)}.

segment([],N,[F1|T],T).

402 segment(F,N,[F1|T],T):phoneme(F1,X,""), name(N,X), unif(F,F1). /*Definition of voicing assimilation constraints*/ voicing_assim(X,Y):unif(X,[sonor:1|_]), unif(Y,[sonor:1,vfv:1|_]). voicing_assim(X,Y):unif(X,[sonor:1|_]), unif(Y,[sonor:0,voice:V,vfv:V|_]). voicing_assim(X,Y):unif(X,[sonor:0,vfv:V|_]), unif(Y,[vfv:V|_]). /*Definition of consonant-vowel interdependency constraints*/ cv_interdependency(X,Y):unif( [cons:0, back:F1, high:F2, onset:[back:F4,high:F3|_]|_], unif( [cons:1, high:F3, back:F4, rime:[high:F2,back:F1|_]|_],

X),

Y).

cv_interdependency(X,Y):unif( [cons:1|_], X), unif( [cons:1, rime:[back:1|_]|_], Y). /*Word-formation grammar*/ mwd(T,M,morphological_word(N),A) --> noun(T,M,N,A). mwd(T,[pl_retr:0|_],morphological_word(V),A) --> verb(T,V,A). mwd(T,[pl_retr:0|_],morphological_word(R),A) --> refl_verb(T,R,A). mwd(T,[pl_retr:0|_],morphological_word(Adj),A) --> adj(T,Adj,A). mwd(T,[pl_retr:0|_],morphological_word(Adv),A) --> adv(T,Adv,A). mwd(T,[pl_retr:0|_],morphological_word(P),A) --> part(T,P,A). noun([T1|T2],F1,noun(NS,P),A) --> noun_stem(T2,F2,NS,Z1), in(T1,P,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3), pl_retr(F1,F2,F3) }.

403 verb([T1|T2],F1,verb(X,Y),A) --> verb_stem(T2,F2,X,Z1), iv(T1,Y,F3,'',_,_,Z2), { A is Z1 + log(Z2/1318181), unif([stress:1|_],F2), vowel_zero(F1,F2,F3) }. refl_verb([T1|T2],F1,refl_verb(X,Y),A) --> verb(T2,F2,X,Z1), c(T1,Y,[pos:particle|F3],'Sa',_,_,Z2), { A is Z1 + log(Z2/1318181), unif([refl:1|_],F1), unif([refl:0|_],F2), unif([refl:1|_],F3) }. adj([T1|T2],F1,adj(X,Y),A) --> adj_stem(T2,F2,X,Z1), ip(T1,Y,F3,_,_,_,Z2), { not(X = [("N",_)|_]), not(X = [("T",_)|_]), A is Z1 + log(Z2/1318181), unif([stress:1|_],F1), unif([stress:1|_],F2), vowel_zero(F1,F2,F3) }. adv([T1|T2],F1,adv(X,Y),A) --> adj_stem(T2,F2,X,Z1), ipu(T1,Y,F3,_,_,_,Z2), { not(X = [("N",_)|_]), not(X = [("T",_)|_]), A is Z1 + log(Z2/1318181), unif([stress:1|_],F1), unif([stress:1|_],F2), vowel_zero(F1,F2,F3) }. part([T],part(Y),A) --> c(T,Y,_,_,_,_,Z1), { A is log(Z1/1318181) }. noun_stem([T],F,noun_stem(Y),A) --> rn(T,Y,F,_,_,_,Z1), { A is log(Z1/1318181) }.

404 noun_stem([T1,T2],F1,noun_stem(X,Y),A) --> c(T2,X,[pos:prefix|F2],_,_,_,Z1), rn(T1,Y,F3,_,_,_,Z2), { A is log(Z1/1318181) + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[collective|F1],noun_stem(X,Y),A) --> noun_stem(T2,[deverbal|F2],X,Z1), sn(T1,Y,F3,'Estv',_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[collective|F1],noun_stem(X,Y),A) --> noun_stem(T2,F2,D1,Acc1,X,Z1), sn(T1,Y,F3,'Estv',_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[qualityname|F1],noun_stem(X,Y),A) --> adj_stem(T2,F2,X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[agent|F1],noun_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sn(T1,Y,F3,'TeL',_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[obj_w_attrib|F1],noun_stem(X,Y),A) --> adj_stem(T2,[concrete|F2],X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }.

405 noun_stem([T1|T2],[obj_w_attrib|F1],noun_stem(X,Y),A) --> noun_stem(T2,F2,X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[obj_w_attrib|F1],noun_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[deverbal|F1],noun_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[diminutive|F1],noun_stem(X,Y),A) --> noun_stem(T2,F2,X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[diminutive|F1],noun_stem(X,Y),A) --> noun_stem(T2,[deverbal|F2],X,Z1), sn(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1,T2|T3],[diminutive|F1],noun_stem(X,Y,Z),A) --> noun_stem(T3,F2,X,Z1), sn(T2,Y,F3,_,_,_,Z2), sn(T1,Z,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }.

406 noun_stem([T1,T2|T3],[diminutive|F1],noun_stem(X,Y,Z),A) --> adj_stem(T3,F2,X,Z1), sn(T2,Y,F3,c,_,_,Z2), sn(T1,Z,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. noun_stem([T1,T2,T3|T4],[diminutive|F1],D,noun_stem(X,Y,Z,P),A) --> adj_stem(T4,F2,X,Z1), sn(T3,Y,F3,c,_,_,Z2), sn(T2,Z,F4,_,_,_,Z3), sn(T1,P,F5,_,_,_,Z4), { A is Z1 + log(Z2/1318181) + log(Z3/1318181) + log(Z4/1318181), assign_stress(Fx,F2,F3), assign_stress(Fy,Fx,F4), assign_stress(F1,Fy,F5), vowel_zero(Fx,F2,F3), vowel_zero(Fy,Fx,F4), vowel_zero(F1,Fy,F5) }. noun_stem([T1|T2],[action|F1],noun_stem(X,Y),A) --> adj_stem(T2,[sfppart|F2],X,Z1), sn(T1,Y,F3,_,_,3823,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[institution|F1],noun_stem(X,Y),A) --> noun_stem(T2,F2,X,Z1), sn(T1,Y,F3,iJ,"2",1829,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. noun_stem([T1|T2],[institution|F1],noun_stem(X,Y),A) --> adj_stem(T2,[deverbal|F2],X,Z1), sn(T1,Y,F3,iJ,"2",1829,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }.

407 noun_stem([T1|T2],[place|F1],noun_stem(X,Y),A) --> noun_stem(T2,F2,X,Z1), sn(T1,Y,F3,ic,"",524,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T],F,adj_stem(Y),A) --> ra(T,Y,F,_,_,_,Z1), { A is log(Z1/1318181) }. adj_stem([T1,T2],F1,adj_stem(X,Y),A) --> c(T2,X,[pos:prefix|F2],_,_,_,Z1), ra(T1,Y,F3,_,_,_,Z2), { A is log(Z1/1318181) + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1,T2],F1,adj_stem(X,Y,Z),A) --> ra(T2,Y,F2,_,_,_,Z1), sa(T1,Z,F3,_,_,_,Z2), { A is log(Z1/1318181) + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1,T2,T3],F1,adj_stem(X,Y,Z),A) --> c(T3,X,[pos:prefix|F2],_,_,_,Z1), ra(T2,Y,F3,_,_,_,Z2), sa(T1,Z,F4,_,_,_,Z3), { A is log(Z1/1318181) + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. adj_stem([T1|T2],[desubst|F1],adj_stem(X,Y),A) --> noun_stem(T2,F2,X,Z1), sa(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }.

408 adj_stem([T1|T2],[desubst|F1],adj_stem(X,Y),A) --> noun_stem(T2,[agent|F2],X,Z1), sa(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1|T2],[desubst|F1],adj_stem(X,Y),A) --> noun_stem(T2,[deverbal|F2],X,Z1), sa(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1,T2|T3],[desubst|F1],adj_stem(X,Y,Z),A) --> noun_stem(T3,[deverbal|F2],X,Z1), sa(T2,Y,F3,_,_,_,Z2), sa(T1,Z,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. adj_stem([T1,T2|T3],[desubst|F1],adj_stem(X,Y,Z),A) --> noun_stem(T3,F2,X,Z1), sa(T2,Y,F3,ik,"2",1168,Z2), sa(T1,Z,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. adj_stem([T1|T2],[desubst|F1],adj_stem(X,Y),A) --> noun_stem(T2,[obj_w_attrib|F2],X,Z1), sa(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }.

409 adj_stem([T1,T2|T3],[desubst|F1],adj_stem(X,Y,Z),A) --> noun_stem(T3,[deverbal|F2],X,Z1), sa(T2,Y,F3,ik,"2",1168,Z2), sa(T1,Z,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. adj_stem([T1|T2],[comparative|F1],adj_stem(X,Y),A) --> adj_stem(T2,F2,X,Z1), sa(T1,Y,F3,’aJw’,"",47,47), { A is Z1 + log(47/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1|T2],[comparative|F1],adj_stem(X,Y),A) --> adj_stem(T2,F2,X,Z1), sa(T1,Y,F3,’eJ’,"",_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1|T2],[deverbal|F1],adj_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sa(T1,Y,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1,T2|T3],[ppart|F1],adj_stem(X,Y,Z),A) --> verb_stem(T3,F2,X,Z1), sa(T2,Y,F3,n,"2",_,Z2), sa(T1,Z,F4,n,"2",_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. adj_stem([T1|T2],[ppart|F1],adj_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sa(T1,Y,F3,’t’,"1",_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }.

410 adj_stem([T1|T2],[sfppart|F1],adj_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sa(T1,Y,F3,’n’,"2",_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. adj_stem([T1|T2],[sfppart|F1],adj_stem(X,Y),A) --> verb_stem(T2,F2,X,Z1), sa(T1,Y,F3,’t’,"1",_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. verb_stem([T],F,verb_stem(X),A) --> rv(T,X,F,_,_,_,Z1), { A is log(Z1/1318181) }. verb_stem([T1,T2],F1,verb_stem(X,Y),A) --> c(T2,X,[pos:prefix|F2],_,_,_,Z1), rv(T1,Y,F3,_,_,_,Z2), { A is log(Z1/1318181) + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. verb_stem([T1|T2],[thematic|F1],verb_stem(Y,Z),A) --> verb_stem(T2,F2,Y,Z1), sv(T1,Z,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. verb_stem([T1,T2|T3],[frequentative|F1],verb_stem(Q1,Q,R),A) --> verb_stem(T3,F2,Q1,Z1), sv(T2,Q,F3,_,_,_,Z2), sv(T1,R,F4,_,_,_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }.

411 verb_stem([T1|T2],[denominal|F1],verb_stem(Y,Z),A) --> adj_stem(T2,F2,Y,Z1), sv(T1,Z,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. verb_stem([T1|T2],[denominal|F1],verb_stem(Y,Z),A) --> noun_stem(T2,F2,Y,Z1), sv(T1,Z,F3,_,_,_,Z2), { A is Z1 + log(Z2/1318181), assign_stress(F1,F2,F3), vowel_zero(F1,F2,F3) }. verb_stem([T1,T2|T3],[denominal|F1],verb_stem(X,Y,rv(a)),A) --> noun_stem(T3,[deverbal|F2],X,Z1), sa(T2,Y,F3,’ov’,[],2328,Z2), sv(T1,rv(a),F4,_,"i",_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. verb_stem([T1,T2|T3],[denominal|F1],verb_stem(X,Y,rv(a)),A) --> noun_stem(T3,F2,X,Z1), sa(T2,Y,F3,’ov’,[],2328,Z2), sv(T1,rv(a),F4,_,"i",_,Z3), { A is Z1 + log(Z2/1318181) + log(Z3/1318181), assign_stress(Fx,F2,F3), assign_stress(F1,Fx,F4), vowel_zero(Fx,F2,F3), vowel_zero(F1,Fx,F4) }. /*Definition of stress assignment constraints*/ assign_stress(X,Y,Z):unif([acc:1,stress:S|_],X), unif([stress:0|_],Y), unif([dom:1,stress:S|_],Z). assign_stress(X,Y,Z):unif([acc:1,stress:S|_],X), unif([acc:1,stress:S|_],Y), unif([dom:0|_],Z). assign_stress(X,Y,Z):unif([acc:1,stress:S|_],X), unif([acc:0|_],Y), unif([acc:1,stress:S|_],Z).

412 assign_stress(_,Y,Z):unif([acc:0|_],Y), unif([acc:0,dom:0|_],Z). /*Definition of jer constraints*/ vowel_zero(X,Y,Z):unif([strong:P|_],X), unif([strong:Q|_],Y), unif([strong:P,left:[high:Q,tense:0|_]|_]). vowel_zero(X,Y,Z):unif([strong:P|_],X), unif([strong:0|_],Y), unif([strong:P,left:[tense:1|_]|_]). /*Definition of plural stress retraction constraints*/ pl_retr(X,Y,Z):unif([pl_retr:1|_],X), unif([num:pl,pl_retr:1|_],Y). pl_retr(X,Y,Z):unif([pl_retr:0|_],X), unif([pl_retr:0|_],Y), unif([pl_retr:0|_],Z). /*The general definition of the unification function*/ unif(X,X):- !. unif([F:V1|Rest1],[F:V2|Rest2]):- !, unif(V1,V2), unif(Rest1,Rest2). unif([F1|Rest1],[F2|Rest2]):unif([F2|Rest3],Rest1), unif(Rest2,[F1|Rest3]).

413 References Afaneseva, Alla. 1984. Udarenie v razli7nyx tipax razgovornoj re7i: imja suš7estvitel'noe. Voronež: VGU. Andersen, Henning. 1969. The phonological status of the Russian “labial fricatives”. Journal of Linguistics 5, 121-127. Anderson, Stephen R. 1988. Morphological theory. In Newmeyer (1988), 146-191. Archangeli, Diana. 1991. Syllabification and prosodic templates in Yawelmani. Natural Language and Linguistic Theory 9, 231-283. Aronoff, Mark. 1976. Word formation in generative grammar. Cambridge, MA: MIT Press. Aronoff, Mark and Frank Anshen. 1998. Morphology and the lexicon: lexicalization and productivity. In Spencer and Zwicky (1998), 237-247. Aronoff, Mark and Richard T. Oehrle (eds.). 1984. Language sound structure. Cambridge, MA: MIT Press. Avanesov, Ruben Ivanovi;. 1956. Fonetika sovremennogo literaturnogo russkogo jazyka. Moscow: Izdatel’stvo Moskovskogo Universiteta. Avanesov, Ruben Ivanovi; and S. I. Ožegov. 1959. Russkoe literaturnoe proiznošenie i udarenie: slovar’-spravo7nik. Moscow: Gosudarstvennoe izdatel’stvo inostrannyx i nacional’nyx slovarej. Baayen, Harald. 1992. Statistical models for word-frequency distributions: a linguistic evaluation. Computers and the Humanities 26, 347-363. Baayen, Harald and Richard Sproat. 1996. Estimating lexical priors for low-frequency morphologically ambiguous forms. Computational Linguistics 22, 155-166. Barry, Susan. 1989. Aspects of sonorant devoicing in Russian. Speech Hearing and Language: Work in Progress 3, 47-59. Department of Phonetics and Linguistics, University College London. Barxudarov, S. G. et al. (eds.). 1975. Slovar’ russkogo jazyka XI-XVII vv. Moscow: Nauka. Bear, John. 1990. Backwards phonology. In Karlgren (1990), volume 3, 13-20. Beard, Robert. 1998. Derivation. In Spencer and Zwicky (1998), 44-65. Benveniste, Émile. 1935. Origines de la formation des noms en Indo-Européen (second edition). Paris: Librairie Adrien-Maisonneuve.

414 Berendsen, Egon. 1983. Russian voicing assimilation and the behavior of Moscow V. Linguistic Analysis 12, 399-413. Bethin, Christina Y. 1998. Slavic prosody: language change and phonological theory. Cambridge: Cambridge University Press. Bird, Steven. 1995. Computational phonology: a constraint-based approach. Cambridge: Cambridge University Press. Bird, Steven, J. S. Coleman, J. B. Pierrehumbert and J. M. Scobbie. 1992. Declarative Phonology. In A. Crochetière, J.-C. Boulanger, and C. Ouellon (eds.). Proceedings of the Fifteenth International Conference of Linguists. Quebec: Presses de l’Université Laval. Blevins, Juliette. 1995. The syllable in phonological theory. In Goldsmith (1995), 206-244. Bloomfield, Leonard. 1933. Language. New York: H. Holt and Company. Blumstein, Sheila E. 1991. The relation between phonetics and phonology. Phonetica 48, 108-119. Bondarko, Lija Vasil’evna. 1969. The syllable structure of speech and distinctive features of phonemes. Phonetica 20, 1-40. Bondarko, Lija Vasil’evna. 1977. Zvukovoj stroj sovremennogo russkogo jazyka. Moscow: Prosveš;enie. Booij, Geert. 1983. Principles and parameters in prosodic phonology. Linguistics 21, 249-280. Booij, Geert and Jerzy Rubach. 1984. Morphological and prosodic domains in lexical phonology. Phonology Yearbook 1, 1-27. Boyer, Michel. 1988. Towards functional logic grammars. Natural Language Understanding and Logic Programming 2, 45-61. Breen, Gavan and Rob Pensalfini. 1999. Arrernte: a language with no onsets. Linguistic Inquiry 30, 1-25. Bresnan, Joan (ed.). 1982. The mental representation of grammatical relations. Cambridge, MA: MIT Press. Bresnan, Joan, and R. M. Kaplan. 1982. Introduction: grammars as mental representations of language. In Bresnan (1982), xvii-lii. Bresnan, Joan, Ronald M. Kaplan, Stanley Peters, and Annie Zaenen. 1982. Crossserial dependencies in Dutch. Linguistic Inquiry 13, 613-635. Bromberger, Sylvain and Morris Halle. 1989. Why phonology is different. Linguistic Inquiry 20, 51-70.

415 Brown, Dunstan, Greville Corbett, Norman Fraser, Andrew Hippisley and Alan Timberlake. 1996. Russian noun stress and network morphology. Linguistics 34, 53107. Bulanin, Lev L’vovi;. 1970. Fonetika sovremennogo russkogo jazyka. Moscow: Visšaja Škola. Burton, Martha W. and Karen E. Robblee. 1997. A phonetic analysis of voicing assimilation in Russian. Journal of Phonetics 25, 97-114. Butler, Christopher. 1985. Statistics in linguistics. Oxford: Blackwell. Bybee, Joan L. 1985. Morphology: a study of the relation between meaning and form. Amsterdam: John Benjamins. Bybee, Joan L. 1988. Morphology as lexical organization. In Hammond and Noonan (eds.). Theoretical morphology: approaches in modern linguistics. San Diego, CA: Academic Press, Inc. 119-141. Carstairs, Andrew. 1987. Allomorphy in inflexion. London: Croom Helm. Cheng, Robert. 1966. Mandarin phonological structure. Journal of Linguistics 2, 135158. Chomsky, Noam. 1957. Syntactic structures. The Hague: Mouton. Chomsky, Noam. 1959. On certain formal properties of grammars. Information and Control 2, 137-167. Chomsky, Noam. 1963. Formal properties of grammars. In Luce, Bush and Galanter (1963), 323-418. Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press. Chomsky, Noam. 1970. Remarks on nominalization. In Roderick A. Jacobs and Peter S. Rosenbaum (eds.). Readings in English transformational grammar. Waltham, MA: Ginn & Co. 184-221. Chomsky, Noam, and Morris Halle. 1968. The sound pattern of English. New York: Harper & Row. Chomsky, Noam, and George Miller. 1963. Introduction to the formal analysis of natural languages. In Luce, Bush and Galanter (1963), 269-321. Christie, William. 1976. Another look at classical phonemics. Language Sciences 39, 37-39. Church, Kenneth. 1983. Phrase-structure parsing: a method for taking advantage of allophonic constraints. Ph. D. dissertation, MIT. Reprinted by the Indiana University Linguistics Club, Bloomington, IN.

416 Church, Kenneth. 1986. Morphological decomposition and stress assignment for speech synthesis. In The Twenty-third Annual Meeting of the Association for Computational Linguistics, Proceedings. Morristown, NJ: Association for Computational Linguistics. 156-164. Clements, George N. 1990. The role of the sonority cycle in core syllabification. In John Kingston and Mary E. Beckman (eds.). Papers in laboratory phonology I: between the grammar and physics of speech. Cambridge: Cambridge University Press. 283-333. Clements, George N. and Samuel J. Keyser. 1983. CV phonology: a generative theory of the syllable. Cambridge, MA: MIT Press. Clocksin, William F., and Christopher S. Mellish. 1981. Programming in Prolog. Berlin: Springer-Verlag. Coats, Herbert S. 1997. On the phonemic status of Russian (8':). Russian Linguistics 21, 157-164. Coats, Herbert S. and Alex P. Harshenin. 1971. On the phonological properties of Russian v. The Slavic and East European Journal 15, 466-478. Cole, Jennifer S. 1995. The cycle in phonology. In Goldsmith (1995), 70-113. Coleman, John S. 1990. Unification phonology: another look at "synthesis-by-rule". In Karlgren (1990), volume 2, 79-84. Coleman, John S. 1993. English word-stress in unification-based grammar. In Ellison and Scobbie (1993), 97-106. Coleman, John S. 1996. Declarative syllabification in Tashlhit Berber. In Durand and Laks (1996), 177-218. Coleman, John S. 1998. Phonological representations: their names, forms and powers. Cambridge: Cambridge University Press. Coleman, John S., A. Dirksen, S. Hussain and J. Waals. 1996. Multilingual phonological analysis and speech synthesis. In Proceedings of computational phonology in speech technology: second meeting of the ACL special interest group in computational phonology. Somerset, NJ: Association for Computational Linguistics. 67-72. Coleman, John S. and Janet Pierrehumbert. 1997. Stochastic phonological grammars and acceptability. In Computational phonology: third meeting of the ACL special interest group in computational phonology. Somerset, NJ: Association for Computational Linguistics. 49-56. Corbett, Greville, and Norman Fraser. 1993. Network morphology: a DATR account of Russian nominal inflection. Journal of Linguistics 29, 113-142. Dirksen, Arthur. 1993. Phrase structure phonology. In Ellison and Scobbie (1993), 8196.

417 Dresher, B. Elan and Jonathan Kaye. 1990. A computational learning model for metrical phonology. Cognition 34, 137-195. Dressler, Wolfgang U. 1985. Morphonology : the dynamics of derivation. Ann Arbor: Karoma Publishers Inc. Durand, J. and B. Laks (eds.). 1996. Current trends in phonology: models and methods, volume 6. Salford: European Studies Research Institute, University of Salford. Dybo, Vladimir Antonovi;. 1981. Slavjanskaja akcentologija: opyt rekonstrukcii sistemy akcentnyx paradigm v praslavjanskom. Moscow: Nauka. Ellison, T. M. and J. M. Scobbie (eds.). Computational phonology. Edinburgh working papers in cognitive science 8. Edinburgh: Centre for Cognitive Science, University of Edinburgh. Evans, Roger, and Gerald Gazdar. 1989a. Inference in DATR. In Proceedings of the fourth conference of the European chapter of the Association for Computational Linguistics, 60-62 April 6989. Manchester: University of Manchester Institute of Science and Technology. 66-71. Evans, Roger, and Gerald Gazdar. 1989b. The semantics of DATR. In Anthony G. Cohn (ed.). Proceedings of the seventh conference of the society for the study of artificial intelligence and simulation of behaviour. London: Pitman. 79-87. Evans, Roger, and Gerald Gazdar. 1995. DATR: A language for lexical knowledge representation. Cognitive Science Research Paper CSRP 382. Brighton: University of Sussex. Fedjanina, Nina Arxipovna. 1976. Udarenie v sovremennom russkom jazyke. Moscow: Russkij jazyk. Fudge, Erik C. 1969. Syllables. Journal of Linguistics 5, 253-286. Garde, Paul. 1976. Histoire de l'accentuation slave. Paris: Institut d’Études Slaves. Garde, Paul. 1980. Grammaire russe: I. Phonologie. Morphologie. Paris: Institut d’Études Slaves. Gazdar, Gerald, Ewan H. Klein, Geoffrey K. Pullum and Ivan A. Sag. 1985. Generalized phrase structure grammar. Oxford: Blackwell. Gladney, Frank Y. 1995. The accent of Russian verbforms. Journal of Slavic Linguistics 3, 97-138. Goldsmith, John A. 1976. Autosegmental phonology. Ph. D. dissertation, MIT. Reprinted by the Indiana University Linguistics Club, Bloomington, IN. Goldsmith, John A. 1990. Autosegmental and metrical phonology. Oxford: Blackwell.

418 Goldsmith, John A. (ed.). 1995. The handbook of phonological theory. Oxford: Blackwell. Grammont, Maurice. 1933. Traité de phonétique. Paris: Librairie Delagrave. Grishman, Ralph. 1986. Computational linguistics: an introduction. Cambridge: Cambridge University Press. Gussmann, Edmund. 1980. Studies in abstract phonology. Cambridge, MA: MIT Press. Gvozdev, Aleksandr Nikolaevi;. 1949. O fonologi7eskix sredstvax russkogo jazyka: sbornik statej. Moscow: Izdatel’stvo Akademii Pedagogi;eskix Nauk RSFSR. Halle, Morris. 1959. The sound pattern of Russian: a linguistic and acoustical investigation. The Hague: Mouton. Halle, Morris. 1973. The accentuation of Russian words. Language 49, 312-348. Halle, Morris. 1975. On Russian accentuation. Slavonic and East European Journal 19, 104-111. Halle, Morris. 1997. On stress and accent in Indo-European. Language 73, 275-313. Halle, Morris, and Paul Kiparsky. 1979. Internal constituent structure and accent in Russian words. Folia Slavica 3, 128-153. Halle, Morris, and Paul Kiparsky. 1981. Review of Garde (1976). Language 57, 150181. Halle, Morris, and Jean-Roger Vergnaud. 1981. Harmony processes. In W. Klein and W. Levelt (eds.). Crossing the boundaries in linguistics. Dordrecht: Reidel. 1-23. Halle, Morris, and Jean-Roger Vergnaud. 1987. An essay on stress. Cambridge, MA: MIT Press. Harris, James W. 1983. Syllable structure and stress in Spanish: a nonlinear analysis. Cambridge, MA: MIT Press. Harris, James W. 1989. The stress erasure convention and cliticization in Spanish. Linguistic Inquiry 20, 339-363. Harris, Zellig S. 1951. Structural linguistics. Chicago: University of Chicago Press. Haugen, Einar. 1956. The syllable in linguistic description. In M. Halle, H.G. Lunt, H. McLean, and C.H. Van Schooneveld (eds.). For Roman Jakobson. The Hague: Mouton. 213-221. Hayes, Bruce. 1984. The phonetics and phonology of Russian voicing assimilation. In Aronoff and Oehrle (1984), 318-328.

419 Hayes, Bruce. 1995. Metrical stress theory: principles and case studies. Chicago: University of Chicago Press. Hockett, Charles F. 1955. A manual of phonology. Baltimore: Waverly Press. Hooper, Joan B. 1972. The syllable in linguistic theory. Language 48, 525-540. Hooper, Joan B. 1976. An introduction to natural generative phonology. New York: Academic Press. Houlihan, Kathleen. 1979. On assimilatory and non-assimilatory phonological rules. Glossa 13, 13-26. Idsardi, William J. 1992. The computation of prosody. Ph. D. dissertation, MIT. Illi;-Svity;, Vladislav Markovi;. 1963. Imennaja akcentuacija v baltijskom i slavjanskom: sud'ba akcentuacionnyx paradigm. Moscow: Izdatel’stvo Akademii Nauk SSSR. Isa;enko, A. V. 1970. East Slavic morphophonemics and the treatment of the jers in Russian: a revision of Havlik’s Law. International Journal of Slavic Linguistics and Poetics 13, 73-124. Itô, Junko. 1986. Syllable theory in prosodic phonology. Ph. D. dissertation, University of Massachusetts, Amherst. Jackendoff, Ray S. 1975. Morphological and semantic regularities in the lexicon. Language 51, 635-671. Jakobson, Roman. 1929. Remarques sur l’évolution phonologique du russe comparée à celle des autres langues slaves. Prague: Jednota Keskoslovenských MatematikL a FysikL. Jakobson, Roman. 1931. Die Betonung und ihre Rolle in der Wort- und Syntagmaphonologie. Travaux du Cercle Linguistique de Prague 4, 164-182. Jakobson, Roman. 1956. Die Verteilung der stimmhaften und stimmlosen Geräuschlaute im Russischen. In Margarete Woltner and Herbert Bräuer (eds.). Festschrift für Max Vasmer zum 70. Geburtstag am 28. Februar 6956. Wiesbaden: In Kommission bei Otto Harrassowitz. 199-202. Jakobson, Roman. 1978. Mutual assimilation of Russian voiced and voiceless consonants. Studia Linguistica 32, 107-110. Jakobson, Roman, C. Gunnar M. Fant, and Morris Halle. 1952. Preliminaries to speech analysis: the distinctive features and their correlates. Acoustics Laboratory Technical Report No. 13, January 1952. Cambridge, MA: MIT. Jefremova, Tatjana Fedorovna. 1971. Review of Worth, Kozak and Johnson (1970). Voprosy jazykoznanija 1971 (4), 112-117. Jespersen, Otto. 1904. Lehrbuch der Phonetik. Leipzig and Berlin: B. G. Teubner.

420 Kahn, Daniel. 1976. Syllable-based generalizations in English phonology. Ph. D. dissertation, MIT. Published 1980 in New York by Garland Press Karlgren, Hans (ed.). 1990. COLING-90: papers presented to the 63th international conference on computational linguistics. Helsinki: University of Helsinki. Kenstowicz, Michael J. 1994. Phonology in generative grammar. Oxford: Blackwell. Kiparsky, Paul. 1973. Phonological representations. In Osamu Fujimura et al. (eds.). Three dimensions of linguistic theory. Tokyo: TEC. 1-136. Kiparsky, Paul. 1982. Lexical morphology and phonology. In The Linguistic Society of Korea (ed.). Linguistics in the morning calm: selected papers from SICOL-6986. Seoul: Hanshin Publishing Company. 3-91. Kiparsky, Paul and Morris Halle. 1977. Towards a reconstruction of the IndoEuropean accent. In Larry M. Hyman (ed.). Studies in stress and accent. Los Angeles, CA: Department of Linguistics, University of Southern California. 209-238. Kohler, Klaus. 1966. Towards a phonological theory. Lingua 16, 337-351. Kuhn Plapp, Rosemary. 1996. Russian !i! and !i! as underlying segments. Journal of Slavic Linguistics 4, 76-108. Lagerberg, Robert. 1992. The relationship between stress and affixation in Russian. Ph. D. dissertation, School of Slavonic and East European Studies, University of London. Larin, Boris Aleksandrovi;. 1959. Russko-anglijskij slovar’-dnevnik Ri7arda Džemsa (6668-6669). Leningrad: Izdatel’stvo leningradskogo universiteta. Leben, William R., and Orrin W. Robinson. 1976. ‘Upside-down’ phonology. Bloomington, IN: Indiana University Linguistics Club. Liberman, Mark. 1975. The intonational system of English. Ph. D. dissertation, MIT. Lieber, Rochelle. 1981. On the organization of the lexicon. Bloomington, IN: Indiana University Linguistics Club. Lieber, Rochelle. 1992. Deconstructing morphology: word formation in syntactic theory. Chicago: University of Chicago Press. Lightner, Theodore M. 1972. Problems in the theory of phonology, volume 6. Edmonton: Linguistic Research. Luce, R. Duncan, Robert R. Bush and Eugene Galanter (eds.). 1963. Handbook of mathematical psychology, volume 2. New York: John Wiley and Sons. Matthews, Peter H. 1974. Morphology. Cambridge: Cambridge University Press. Matusevi;, Margarita Ivanovna. 1976. Sovremennyj russkij jazyk: fonetika. Moscow: Prosveš;enie.

421 McCarthy, John J. 1982. Formal problems in Semitic phonology and morphology. Ph. D. dissertation 1979, reprinted. Bloomington, IN: Indiana University Linguistics Club. McCarthy, John J., and Alan S. Prince. ms 1993. Prosodic morphology I: constraint interaction and satisfaction. University of Massachusetts, Amherst and Rutgers University. Melvold, Janis L. 1989. Structure and stress in the phonology of Russian. Ph. D. dissertation, MIT. Miller, George and Noam Chomsky. 1963. Finitary models of language users. In Luce, Bush and Galanter (1963), 419-491. Nesset, Tore. 1994. Russian stress: stress as an inflectional formative in Russian noun paradigms and Bybee’s cognitive morphology. Oslo: Novus Press. Newmeyer, Frederick J. (ed.). 1988. Linguistics: The Cambridge survey I. Cambridge: Cambridge University Press. Oliverius, ZdenMk F. 1974. Fonetika russkogo jazyka. Prague: Státní pedagogické nakladatelství. Oliverius, ZdenMk F. 1976. Morfemy russkogo jazyka: 7astotnyj slovar’. Prague: Univerzita Karlova. Panov, Mixail Viktorovi;. (ed.). 1968. Russkij jazyk i sovetskoe obš7estvo: sociologolingvisti7eskoe issledovanie. Moscow: Nauka. Panov, Mixail Viktorovi;. 1979. Sovremennyj russkij jazyk: Fonetika. Moscow: Visšaja Škola. Pesetsky, David. ms 1979. Russian morphology and lexical theory. Cambridge, MA, MIT. Pierrehumbert, Janet and Rami Nair. 1995. Word games and syllable structure. Language and Speech 38, 77-114. Pinker, Steven and Alan Prince. 1988. On language and connectionism: analysis of a parallel distributed processing model of language acquisition. Cognition 28, 73-193. Piotrowski, Marek. 1992. Polish yers in non-linear phonology. In Wolfgang U. Dressler, H. C. Luschützky, O. E. Pfeiffer, and J. R. Rennison (eds.). Phonologica 6988: Proceedings of the sixth international phonology meeting. Cambridge: Cambridge University Press. 215-227. Piotrowski, Marek, Iggy Roca & Andrew Spencer. 1992. Polish yers and lexical syllabicity. The Linguistic Review 9, 27-67. Pollard, Carl, and Ivan Sag. 1987. Information-based syntax and semantics, volume I: Fundamentals. Stanford, CA: Center for the Study of Language and Information.

422 Postal, Paul M. 1968. Aspects of phonological theory. New York: Harper & Row. Press, J. Ian. 1986. Aspects of the phonology of the Slavonic languages. Amsterdam: Rodopi. Red’kin, Vitalij Andrejevi;. 1971. Akcentologija sovremennogo russkogo literaturnogo jazyka: posobie dlja u7itelej. Moscow: Prosveš;enie. Reformatskij, Aleksandr Aleksandrovi;. (ed.). 1970. Iz istorii ote7estvennoj fonologii. O7erk. Xrestomatija. Moscow: Nauka. Reformatskij, Aleksandr Aleksandrovi;. 1975. Fonologi7eskie ètjudy. Moscow: Nauka. Robblee, Karen E. and Martha W. Burton. 1997. Sonorant voicing transparency in Russian. In Wayles Browne, E. Dornisch, N. Kondrashova and D. Zec (eds.). Annual workshop on formal approaches to Slavic linguistics: the Cornell meeting 6995. Ann Arbor: Michigan Slavic Publications. 407-434. Rubach, Jerzy. 1984. Cyclic and lexical phonology: the structure of Polish. Dordrecht: Foris. Rubach, Jerzy. 1986. Abstract vowels in three dimensional phonology: the yers. The Linguistic Review 5, 247-280. Rubach, Jerzy. 1995. Representations and the organization of rules in Slavic phonology. In Goldsmith (1995), 848-866. Rubach, Jerzy, and Geert Booij. 1990. Syllable structure assignment in Polish. Phonology 7, 121-158. Sadock, Jerrold M. 1991. Autolexical syntax: a theory of parallel grammatical representations. Chicago: University of Chicago Press. de Saussure, Ferdinand. 1916. Cours de linguistique générale. Paris: Librairie Payot. Scalise, Sergio. 1984. Generative morphology. Dordrecht: Foris. Š;erba, Lev Vladimirovi;. 1912. Russkie glasnye v ka7estvennom i koli7estvennom otnošenii. St. Petersburg: Tipografija Ju. N. Erlix. Schane, Sanford. 1971. The phoneme revisited. Language 47, 503-521. Scobbie, James, John Coleman and Steven Bird. 1996. Key aspects of declarative phonology. In Durand and Laks (1996), 685-709. Selkirk, Elisabeth O. 1982a. The syllable. In Harry van der Hulst and Norval Smith (eds.). The structure of phonological representations (part II). Dordrecht: Foris. 337383. Selkirk, Elisabeth O. 1982b. The syntax of words. Cambridge, MA: MIT Press.

423 Selkirk, Elisabeth O. 1984. On the major class features and syllable theory. In Aronoff and Oehrle (1984), 107-136. Shapiro, Michael. 1986. The Russian system of stress. Russian Linguistics 10, 183204. Shapiro, Michael. 1993. Russian non-distinctive voicing: a stocktaking. Russian Linguistics 17, 1-14. Shieber, Stuart M. 1985. Criteria for designing computer facilities for linguistic analysis. Linguistics 23, 189-211. Shieber, Stuart M. 1986. An introduction to unification-based approaches to grammar. Stanford: Center for the Study of Language and Information. Siegel, Dorothy. 1977. The adjacency condition and the theory of morphology. Proceedings of the North-Eastern Linguistics Society 8, 189-197. Sievers, Eduard. 1881. Grundzüge der Phonetik. Leipzig: Breitkopf und Härtel. Slowiaczek, Louise, and Daniel Dinnsen. 1985. On the neutralizing status of Polish word-final devoicing. Journal of Phonetics 13, 325-341. Sommerstein, Alan H. 1977. Modern phonology. London: Arnold. Spencer, Andrew. 1986. A nonlinear analysis of vowel zero alternations in Polish. Journal of Linguistics 22, 249-280. Spencer, Andrew. 1998. Morphophonological operations. In Spencer and Zwicky (1998), 123-143. Spencer, Andrew, and Arnold M. Zwicky (eds.). 1998. The handbook of morphology. Oxford: Blackwell. Sproat, Richard. 1985. On deriving the lexicon. Ph. D. dissertation, MIT. Abstract published in The Linguistic Review 5, 185-196. Stang, Christian S. 1957. Slavonic accentuation. Oslo: I kommisjon hos H. Aschehoug (W. Nygaard). Steriade, Donca. 1982. Greek prosodies and the nature of syllabification. Ph. D. dissertation, MIT. Švedova, Natalia Jul’evna et al. (eds.). 1980. Russkaja grammatika. Moscow: Nauka. Szpyra, Jolanta. 1992. Ghost segments in nonlinear phonology: Polish yers. Language 48, 277-312. Thelin, Nils B. 1974. On the phonological status of the Russian geminate palatals. Russian Linguistics 1, 163-178.

424 Thelin, Nils B. 1983. The Russian geminate palatals in a processual and more realistic functional perspective, or: why Michael Flier’s monkey-tail argument cannot convince me of there being 5 (or more) phonemes in 18%&. Russian Linguistics 7, 167-192. Tixonov, Aleksandr Nikolaevi;. 1985-90. Slovoobrazovatel’nyj slovar’ russkogo jazyka, volumes 1-2. Moscow: Russkij jazyk. Toman, JindNich. 1998. Word syntax. In Spencer and Zwicky (1998), 306-321. Tornow, Siegfried. 1984. Die häufigsten Akzenttypen in der russischen Flexion. Berlin: Osteuropa-Institut. Townsend, Charles E. 1975. Russian word-formation. Corrected reprint of original edition, 1968. Columbus, OH: Slavica. Trofimov, Mixail V. and Daniel Jones. 1923. The pronunciation of Russian. Cambridge: Cambridge University Press. Vennemann, Theo. 1972. On the theory of syllabic phonology. Linguistische Berichte 18, 1-18. Walther, Markus. 1993. Declarative syllabification with applications to German. In Ellison and Scobbie (1993), 55-79. Weinberg, Amy S. 1988. Mathematical properties of grammars. In Newmeyer (1988), 416-429. Whitney, William D. 1865. On the relation of vowels and consonants. Journal of the American Oriental Society 8, 357-373. Worth, Dean S., Andrew S. Kozak and Donald B. Johnson. 1970. Russian derivational dictionary. New York: American Elsevier Publishing Company. Zaliznjak, Andrej Anatol’evi;. 1967. Russkoe imennoe slovoizmenenie. Moscow: Nauka. Zaliznjak, Andrej Anatol’evi;. 1977. Grammati7eskij slovar’ russkogo jazyka: slovoizmenenie. Moscow: Russkij jazyk. Zaliznjak, Andrej Anatol’evi;. 1985. Ot pravoslavjanskoj akcentuacii k russkoj, Moscow: Nauka. Zemskaja, Elena Andreevna. 1973. Sovremennyj russkij jazyk. Slovoobrazovanije. Moscow. Zwicky, Arnold M. 1992. Some choices in the theory of morphology. In Robert Levine (ed.). Formal grammar: theory and implementation. New York and Oxford: Oxford University Press. 327-371.