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 word-
internal 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
!! enclose phonemic representations, e.g. !#$%&'!
() enclose phonetic representations, e.g. (*$%&')
!#
c
++,#-
rn
+
in
! denotes morphological tokenization; subscripts classify individual
morphs
+ morpheme boundary
. syllable boundary
/ denotes word-stress in IPA transcriptions (stress on the vowel to the
right)
!+,#-!
rn
denotes a single morpheme (classificatory subscript is outside
obliques)
o syllable
C the empty string
anter anterior
C any consonant
CFG context-free grammar
cons consonantal
cont continuant
coron coronal
DCG (Prolog) Definite Clause Grammar
del_rel delayed release
init initial
later lateral
OT Optimality Theory
PSG phrase structure grammar
sonor sonorant
SSG Sonority Sequencing Generalization
V any vowel
vfv vocal fold vibration
voc vocalic
Symbols used in morphological tokenization
r* root
s suffix
c* clitic
i inflectional ending
p pronominal
a adjectival
n substantival
v verbal
d durative process
r* resultative process
i iterative process
c* 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 Kahns (1976) syllable structure assignment rules .................................................... 45
2.2.3.2 Its (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 grammars 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 Halles theory on non-derived nouns................................................................ 292
4.4.2 Test of Halles theory on non-derived and derived nouns ............................................ 293
4.4.3 Test of Melvolds theory on non-derived and derived nouns ....................................... 294
4.4.4 Test of Melvolds theory on nouns, non-reflexive verbs and adjectives....................... 295
4.4.5 Test of Zaliznjaks theory on nominative singular derived nouns................................ 296
4.4.6 Test of Melvolds theory on nominative singular derived nouns.................................. 297
4.4.7 Analysis of errors in Melvolds and Zaliznjaks 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. Oliveriuss (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 !906
c
+&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!&%&: !$;0
ra
+%2
sn
+%
in3
!..................................... 322
Figure 25. The constraint pool.............................................................................................................. 324
Figure 26. Morphological/phonological structure of #8!&:% !$;0
ra
+%2
sn
+
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&:<!#&6- + /%! table (gen. sg.)...................................................... 256
Table 25. List of solutions for !#&/6-! (revised) ..................................................................................... 259
Table 26. Candidate accentuations before and after operation of the ALA.......................................... 269
Table 27. Differences between accentual and other categories posited by........................................... 275
Table 28. Procedural comparison of Melvold and Zaliznjak................................................................ 277
Table 29. Demonstration that Melvolds theory is problematic ........................................................... 279
Table 30. Demonstration that Zaliznjaks theory is problematic.......................................................... 280
Table 31. Number of candidate accentuations against.......................................................................... 282
Table 32. Ranking of underlying morpheme forms.............................................................................. 288
Table 33. Results of testing Halles theory on non-derived words....................................................... 293
Table 34. Results of testing Halles theory on non-derived and derived nouns.................................... 294
Table 35. Results of testing Melvolds theory on non-derived and derived nouns............................... 294
Table 36. Results of testing Melvolds theory...................................................................................... 295
Table 37. Results of testing Zaliznjaks theory .................................................................................... 297
Table 38. Results of testing Melvolds theory...................................................................................... 297
Table 39. Analysis of words incorrectly stressed by Melvolds theory................................................ 299
Table 40. Analysis of words incorrectly stressed by Zaliznjaks theory .............................................. 300
Table 41. Exceptions common to Zaliznjaks and Melvolds theories................................................. 302
Table 42. Prefixed nouns stressed incorrectly by Zaliznjak ................................................................. 303
Table 43. Words derived from prefixed stems ..................................................................................... 305
Table 44. Further words derived from prefixed stems.......................................................................... 305
Table 45. Results of testing the overall phonology for its ability to assign stress ................................ 340
Table 46. Results of testing Melvolds theory on 4,416 nouns............................................................. 341
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
Halles (1959) Sound Pattern of Russian. The following quotation (op. cit., p. 44)
serves to show that Halles 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 Chomskys 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 Chomskys
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 Redkin (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 computer-
interpretable 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 !*/%<0;2;! by the hand) and
are thus considered to function as a single phonological word, but two morphological words.
17
grammar that enumerates the forms of a language should also be able to assign them a
structural description (that is, parse them). These functions are clearly fulfilled by
phrase-structure grammars (PSGs), since in a PSG each rule can equivalently be
thought of as a partial structure, and each derivation can be represented as a directed
graph.
The ability of a grammar to parse (that is, provide some structural description
for the word) does not necessarily imply its ability to parse correctly. As Chomsky
and Miller (1963: 297) state, we have no interest, ultimately, in grammars that
generate a natural language correctly but fail to generate the correct set of structural
descriptions. A grammar which is able to assign structural descriptions to all relevant
well-formed utterances in a language is said to meet the condition of descriptive
adequacy, while a grammar which meets the more stringent requirement of assigning
correct structural descriptions to all well-formed utterances is said to meet the
condition of explanatory adequacy. In general, it is considerably harder to prove or
disprove a grammars explanatory adequacy than its descriptive adequacy, since the
former is a matter not just of linguistic data but of psychology as well (Chomsky
1965: 18-27). Moreover, it is important to realize that a parse should not necessarily
be considered incorrect just because it was unanticipated: such a parse may in fact be
a possible but unlikely parse. These factors all mean that establishing whether a given
grammar assigns correct structural descriptions is not always straightforward, and is
often a matter of judgement.

Conversely, English words of the form non-X (where X stands for an adjective) are a single
morphological word, but two phonological words.
18
Essentially, there are three good reasons for formulating a theory within the
framework of PSG. First, PSGs are the standard means of assigning hierarchical
constituent structure to strings, which is widely and uncontroversially regarded as an
important function of linguistics. The literature on phrase-structure grammar has been
developed over approximately 40 years, and owes much to Chomskys interest in
establishing a formal foundation for linguistics (e.g. Chomsky 1959, Chomsky and
Miller 1963, Chomsky 1963).
A second strength of the PSG formalism is that it has a straightforward
declarative interpretation. Phrase-structure grammar rules can equally validly be
seen as partial descriptions of surface representations or descriptions of information
structures, in Brown, Corbett, Fraser, Hippisley and Timberlakes (1996)
terminology. Specifically, context-free phrase-structure rules can be represented
graphically as tree structures (Coleman 1998: 99).
Third, there is a transparent relationship between PSGs and Definite Clause
Grammars (DCGs)
2
. This is perhaps the greatest advantage of using the PSG
formalism, because it means that a PSG can easily be implemented and tested
computationally. DCGs are a particular type of formalism available as part of the
programming language Prolog. For details of the workings of Prolog and DCGs, the
reader is invited to refer to a textbook on Prolog, such as Clocksin and Mellish
(1981). Here, it is sufficient to appreciate that DCGs can fulfil the functions of parsing
and generation, because Prolog is a declarative programming language. Thus, if a

2
DCGs are capable of defining recursively enumerable languages and CFGs are capable of defining
only context-free languages (which are a subset of the set of recursively enumerable languages). Thus,
to be more precise, the type of DCG used to implement the theory proposed in this dissertation is a
restricted type of DCG.
19
particular grammar is implemented as a DCG, it is possible to test the grammar
computationally to determine whether it describe[s] all and only the possible forms
of a language (Bird, Coleman, Pierrehumbert and Scobbie 1992). Throughout this
dissertation, I describe computational tests of this kind to determine whether the
different aspects of the grammar are accurate representations of the facts of the
language.
1.3.2 Context-free grammar
Having established in section 1.3.1 why I use the framework of PSG, I now
move on to explain the significance of my claim that nothing more powerful than a
context-free grammar (CFG) is necessary to describe the facts of Russian phonology.
The claim that CFG is sufficient is in contrast to McCarthy (1982), for example, who
claims that phonology is context-sensitive (p. 201). (Coleman 1998: 81 observes that
his phonology is an unrestricted rewriting system, since it is a context-sensitive
grammar with deletion: see McCarthys (1) on p. 201.) In other respects, however,
McCarthys aim is comparable to mine: McCarthy aims to provide a fair degree of
coverage, particularly in Hebrew phonology and Arabic morphology (p. 2), including
stress assignment.
CFGs are one of a number of types of grammar formalism in the Chomsky
Hierarchy (Chomsky 1959), represented in Figure 1. All of these grammar formalisms
are members of the family of PSGs. The place of a particular grammar within the
hierarchy is determined by the type of rules included in the grammar, as shown in
Table 1 (adapted from Coleman 1998: 79).
20
Figure 1. The Chomsky Hierarchy
Table 1. Types of rules permitted by grammars in the Chomsky Hierarchy
3
Type Grammar Rule types
allowed
Conditions on symbols
0 Unrestricted A B A e (V
T
V
N
)*
B e (V
T
V
N
)*
1 Context-sensitive A B/ C _ D
and
A
A e V
N
,
B e (V
T
V
N
)
+
,
C,D e (V
T
V
N
)*
2 Context-free A B A e V
N
,
B e (V
T
V
N
)*
3 Right linear A aB A e V
N
,
B e (V
N
{}),
a e V
T
3 Left linear A Ba A e V
N
,
B e (V
N
{}),
a e V
T

3
Note to Table 1: Following Chomsky (1959) and Coleman (1998), V
T
represents the set of terminal
symbols, V
N
the set of non-terminal symbols, X* a sequence of zero or more Xs, X
+
a sequence of one
or more Xs, and the empty string.
Linear grammar (type 3)
Context-free grammar (type 2)
Context-sensitive grammar (type 1)
Unrestricted grammar (type 0)
21
Because there has been a considerable amount of work carried out in phrase-structure
grammar, the properties of different types of PSG in the Chomsky Hierarchy are by
now well understood. These properties are important to consider when formulating a
theory, for reasons which will now be made clear.
On a very general level, the more restricted the grammar formalism, the better.
This follows, essentially, from the principle of Occams razor: as Coleman (1998: 80)
points out, the goal in developing a formal theory of natural-language syntax or
phonology, is to use a type of grammar which is as powerful as necessary, but as
restrictive as possible. It should be acknowledged, however, that context-free
grammars can in practice have a cost compared to more powerful types of grammars,
in that more powerful grammars may describe the same phenomena more simply
(with fewer features or more general rules, for example), and may even be able to
parse and generate more efficiently in some cases (Weinberg 1988).
However, there are other, perhaps more psychological, arguments in support
of choosing a grammar formalism no more powerful than context-free. Bresnan and
Kaplan (1982) set out a number of constraints that they suggest linguistic theory
should impose on the class of possible grammars, and CFGs adhere to all but one of
these constraints. The one constraint which CFG does not adhere to is the
universality constraint, which assumes that the procedure for grammatical
interpretation, m
G
, is the same for all natural language grammars G (Bresnan and
Kaplan 1982: xlvii). It is significant that Bresnan and Kaplans grounds for stating
that CFG does not adhere to this constraint come from syntax, not phonology:
Bresnan, Kaplan, Peters, and Zaenen 1982 have shown that there is no context-free
phrase-structure grammar that can correctly characterize the parse trees of Dutch.
The problem lies in Dutch cross-serial constructions, in which the verbs are
discontinuous from the verb phrases that contain their arguments The results of
22
Bresnan, Kaplan, Peters, and Zaenen 1982 show that context-free grammars cannot
provide a universal means of representing these phenomena. (p. xlix)
Of the other constraints, one is the creativity constraint. One of the claimed
contributions of generative grammar to linguistics was the observation that if a
grammar is to be an equally valid model both of linguistic perception and production,
it should be able not only to assign structure to strings, but also to generate strings
(hence the term generative grammar). This observation is, for example, one of the
foundational tenets of Chomsky (1957). As noted by Matthews (1974: 219),
generative linguistics was partly a reaction to structuralist linguistics, which (it was
claimed) emphasized assignment of structure at the expense of generation. Despite the
emphasis of generative linguists upon the generative, it is notable that context-
sensitive grammars and those more powerful are not necessarily reversible (Bear
1990). However, CFGs do always have the property of reversibility: that is, they can
be used either for generation or recognition.
Another constraint which CFGs satisfy is Bresnan and Kaplans reliability
constraint: that is, they can always accept or reject strings in a finite amount of time.
One of the properties of context-free (and more restricted) languages is that of
decidability (alternatively known as computability, Turing-decidability or
recursiveness). A language L is decidable if there is an algorithm for determining
membership in L; in other words, L is decidable if there is a grammar which can
decide whether a string is well- or ill-formed (a member of L or not) in a finite
amount of time. Languages of type m, where m s 1, are not necessarily decidable, but
those of type n, where n > 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
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
P: All words
Q: Words assigned some
structural description
R: Words assigned the correct
structural description
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 Melvolds
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 Oliveriuss 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 Zaliznjaks 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 stem-
inflection 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!&%) (nom./acc. pl.)
#8!&%& (gen. sg.) #8!&%,( (gen. pl.)
#8!&%8 (dat. sg.) #8!&%&+ (dat. pl.)
#8!&%,+ (instr. sg.) #8!&%&+) (instr. pl.)
#8!&%" (loc. sg.) #8!&%&: (loc. pl.)
29
(c) Pair the headword with its morphological tokenization, which is known
from the information in Oliverius (1976) (for example, !$;0%2! would
be paired with the tokenization !$;0
ra
+%2
sn
+
in1
!
4
), and deduce the
noun-stem by removing the inflectional ending (in this case, zero). The
noun-stem of !$;0%2! would thus be !$;0
ra
+%2
sn
!.
(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 !$;0
ra
+%2
sn
+%
in
!, !$;0
ra
+%2
sn
+;
in
!,
!$;0
ra
+%2
sn
+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
Number of
words
Non-nouns Head-words
1,525
Nouns Head-words
132
Nouns Head-words 835
Nouns Inflected forms

8,798
Total 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
Bondarkos 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, Bondarkos
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 sonority-
based 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; Bloomfields
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 Halles 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
Halles 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 Kahns
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 Kahns 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: 186-
187, 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 Selkirks (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 Clementss and Selkirks
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 latter
6
. 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 Halles specific constraints on consonant clusters.
7
Halles
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 cross-
linguistic 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
Example in
morpheme-initial
position
Example in
morpheme-medial
position
Example in
morpheme-final
position
4 segments:
CCCR
{fstret,i} to meet not attested not attested
RCCC
not attested not attested {;orstv} stale
3 segments:
CCR
{skr,ip} squeak {kastr,ul,} saucepan {zatxl} musty
CCC
{stvol} gun barrel not attested {op;} common
RCR
not attested {v,irbl,ud} camel not attested
RCC
not attested not attested {tolst} fat
2 segments:
CC
{svet} light {asp,id} slate {kost,} bone
RC
{rtut,} mercury {almaz} diamond {smer;} cedar
JC
not attested {bajbak} marmot {ajv} quince
CR
{slep} blind {utrob} womb {ezl} staff
RR
not attested {jArlik} label {gorl} throat
CJ
{djak} clerk {ab,izjan} monkey not attested
RJ
{rjan} zealous {burjan} tall weeds not attested
Key to Halles symbols used in Table 3
R stands for any liquid, C stands for any consonant, J stands for the glide !7!, A (in {jArlik} 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,} Initial CCR, cf. {skr,ip}
{v,ir . bl,ud} Final R, initial CR, cf. {slep}
{a . sp,id} Initial CC, cf. {svet}
{al . maz} Final R, initial C
{baj . bak} Final J, initial C
{u . trob} Initial CR, cf. {slep}
{jAr . lik} Final R, initial R
{a . b,i . zjan} Initial CJ, cf. {djak}
{bu . rjan} 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 {zatxl} (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 {ezl},
43
and initial RC as in {rtut,}.
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 Booijs 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 template-
matching or templatic approaches (Selkirk 1984, It 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 It
(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 Kahns (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 !
o o o o
(where o 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 segments 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 !,=<D=<;=<E!.
Sonority and Kahns approach are not opposed in principle.
46
Kahn then proposes that word-initial consonant clusters should be syllabified
with the first syllable of the word, and word-final consonant clusters with the last
syllable. The syllabification of word-initial and word-final clusters is more
straightforward than that of word-medial clusters, because there is an inherent
ambiguity with any word-medial cluster as to where one syllable ends and the next
begins. As Kahn observes, the split will be determined in part by the rules governing
what are possible onsets and codas. He suggests that by studying what types of
cluster occur initially and finally in words in English, we can discover at least a subset
of the set of possible syllable-initial and -final clusters (p. 40). For reasons that will
not be recapitulated here, he goes on to conclude that the set of possible syllable-
initial (-final) clusters in English is identical to the set of possible word-initial (-final)
clusters (p. 41). For this reason, Kahn argues, one can deduce the rules about all
onsets (or codas), word-initial (-final) or otherwise, by inspection of the set of word-
initial (-final) clusters alone. This assumes, of course, that the list of word-initial
(-final) clusters is not a superset of the set of possible onsets (codas). However,
Rubach and Booij (1990) suggest, on the basis of facts from Polish, that the
constraints on clusters at the edges of words are weaker than those on word-medial
clusters. For example, !0&! is a possible word-initial cluster in Polish (as attested in
rt!" mercury), but this does not warrant the syllabification of karty cards as ka-
rty (Rubach and Booij 1990: 122). If this is true, then Kahns suggestion is too
unconstrained, since Kahn would conclude from rt!" that any !0&! should be
syllabified with an immediately following vowel.
Even with Kahns assumption, there is still ambiguity in cases such as CVCV:
since C is both a possible onset and a possible coda, both CV.CV and CVC.V would
47
be possible syllabifications. Kahn then suggests that in such cases there is a strong
tendency to syllabify in such a way that initial clusters are of maximal length,
consistent with the general constraints on word-initial consonant clusters (the so-
called Maximal Onset Principle), so the preferred syllabification in this case would be
CV.CV. Kahns approach is to encode the Maximal Onset Principle as an integral part
of the syllabification mechanism, as is clear from his statement quoted on p. 43 above.
As Steriade (1982: 75) puts it, Kahns algorithm is partly a specification of certain
aspects of the English syllable and partly a parsing device.
The second rule that Kahn proposes for syllable structure assignment is as in
(5):
(5) Rule II
a. C
1
.C
a
V C
1
.C
c
C
c+1
C
a
V
o o
where C
c+1
C
a
V is a permissible initial cluster
but C
c
C
c+1
C
a
V is not.
b. V C
1
.C
a
VC
1
C
j
C
j+1
C
a
xx
o o
where C
1
C
j
is a permissible final cluster but C
1
C
j
C
j+1
is not; x standing
below a segment indicates that the segment is not associated with any syllable.
Together, Rules I and II are sufficient to ensure that all segments in a well-formed
string, vowels and consonants, are associated with one syllable or another; and as
Kahn points out, the formal correlate of the fact that ill-formed strings will be rejected
is that not all segments will be associated. Thus from the derivation in (6) it can be
48
seen that !#9D4!, !+D#&! and !AF#&34! are all well-formed because all segments are
syllabified, but !2#9D4! and !+D#&2! are ill-formed because, in both cases, no rule acts
to assign !2! to any syllable.
(6) Rule I !#<9<D<4! !+<D<#<&! !<A<F<#<&<3<4! !2<#<9<D<4! !+<D<#<&<2!
o o o o o o
Rule IIa !#<9<D<4! !+<D<#<&! !<A<F<#<&<3<4! !2<#<9<D<4! !+<D<#<&<2!
o o o o o o
Rule IIb !#<9<D<4! !+<D<#<&! !<A<F<#<&<3<4! !2<#<9<D<4! !+<D<#<&<2!
o o o o o o
Kahns syllabification algorithm is a bottom-up algorithm, in the sense that it
starts with segments and builds up successively higher levels of syllable structure on
top of these segments. Indeed, bottom-up syllabification is a general feature of rule-
based approaches to syllabification (e.g. Steriade 1982, Rubach and Booij 1990).
There is no question that this approach to syllabification works, but it is easy to be left
with the impression that bottom-up syllabification is the only possible rule-based
method of syllabification. As will be shown in 2.3-2.4, it is quite possible to
implement Kahns hypotheses about onsets and codas within a framework which
assigns syllable structure in a manner which is not bottom-up. This is because Kahns
syllabification algorithm is in reality a framework in which a particular theory about
syllable structure is implemented. The questions of what are possible onsets and
codas, whether the sets of word-marginal onsets and codas are, respectively, identical
to the sets of all word-medial onsets and codas, and whether syllables really do
49
maximize their onsets, are questions of syllable structure rather than syllable
structure assignment.
2.2.3.2 Its (1986) method of syllable structure assignment
Its (1986) discussion of how syllable structure is assigned is essentially
aimed at showing that a template-approach, coupled with a directionality parameter, is
a valid and preferable alternative to the rule-based approach advocated earlier in Kahn
(1976) and Steriade (1982): valid, because it too fulfils the condition of descriptive
adequacy, and preferable, because, according to It (1986: 168), the principle of
directionality is independently motivated in a number of other areas of Prosodic
Phonology such as metrical phonology (Liberman 1975). In other words, It views the
fact that templatic syllabification is based on directionality as an advantage in itself.
It (1986: 164) notes that a single intervocalic consonant is always analyzed
as an onset. It correctly states that this principle can be encoded in a number of
ways: for example, the implementation of this principle in a rule-based approach was
shown in (5) above. The rule as stated in the template-approach, on the other hand, is
as follows: the sequence CV must belong to a single syllable (op. cit., p. 165).
Turning then to the question of the syllabification of longer intervocalic
consonant clusters, It observes that both VC
1
.C
2
V and V.C
1
C
2
V parsings occur in
natural languages. This can be seen if one compares Icelandic !#,B.-%! (It 1986: 173)
with English al.ge.bra (Kahn 1976: 138): both examples contain a sequence
consisting of a voiced stop followed by a liquid, but the English example syllabifies
this sequence as an onset and the Icelandic one syllabifies it as a coda followed by an
onset. As It states, these differences are accounted for in rule-based approaches by
different orderings of the onset and coda rules (ordering the onset rule first as in (5),
50
for example, yields the V.C
1
C
2
V parsing). In a template-approach, however, the
differences are explained by the prosodic principle of Directionality (It 1986: 167).
Thus the V.C
1
C
2
V parsing is a result of right-to-left template matching (providing
C
1
C
2
is a possible onset), while the VC
1
.C
2
V parsing results from left-to-right
matching (even if C
1
C
2
is a possible onset), where the directionality of mapping is
specified on a language-by-language basis. Its implication is that because the choice
of left-to-right or right-to-left parsing is relevant to templatic syllabification, and
because this choice is relevant in a similar way to, say, metrical theory, this is
therefore evidence that syllable structure is assigned templatically.
It also adduces other pieces of evidence for templatic syllabification. One of
these is epenthesis: she argues that templatic syllabification can provide a simpler
explanation of this phenomenon in certain languages than rule-based analyses. For
example, a difference between Cairene and Iraqi Arabic is that in the former,
triconsonantal clusters are broken up by an epenthetic vowel between the second and
third consonant, and in the latter, the epenthetic vowel surfaces between the first and
second consonants. In both dialects, four-consonantal clusters are broken by an
epenthetic vowel between the second and third consonants. It demonstrates that the
difference in triconsonantal clusters can be accounted for if both dialects are assumed
to have CVC templates, but if the directionality of mapping is left-to-right for Cairene
and right-to-left for Iraqi, and that the same model also correctly predicts the location
of the epenthetic vowel in four-consonant clusters in both dialects. As noted by
Kenstowicz (1994: 273), these facts present more of a problem for rule-driven
epenthesis.
51
Thus the advantages which It claims for templatic syllabification over rule-
based syllabification are that templatic syllabification is more in line with
directionality, and that it explains various phonological processes more simply. The
facts discussed by It do not appear to bear directly upon Russian syllable structure or
phonological processes. However, in due course I shall return to one of Its key
arguments, questioning whether directionality is really necessary or desirable in an
account of syllable structure.
2.2.3.3 Syllable structure assignment in Optimality Theory
McCarthy and Prince (ms 1993) present a third method of generating syllable
structure within the general framework of Optimality Theory (OT). OT does not fit
clearly into either the rule-based or templatic categories suggested by Blevins
(1995). However, Blevins (1995) largely predates formal publications in OT.
The following four principles are set out in McCarthy and Prince (ms 1993: 5)
as hallmark properties of OT:
(i) Violability. Constraints are violable; but violation is minimal.
(ii) Ranking. Constraints are ranked on a language-particular basis; the notion
of minimal violation (or best-satisfaction) is defined in terms of this ranking.
(iii) Inclusiveness. The candidate analyses, which are evaluated by the
constraint hierarchy, are admitted by very general considerations of structural well-
formedness; there are no specific rules or repair strategies with specific structural
descriptions or structural changes or with connections to specific constraints.
(iv) Parallelism. Best-satisfaction of the constraint hierarchy is computed over
the whole hierarchy and the whole candidate set.
The application of these principles to syllable structure assignment is outlined
in McCarthy and Prince (ms 1993: 13) as follows:
52
Syllable structure is generated under Optimality Theory in the same way as any other
grammatical property. The function Gen produces a candidate set of syllabic parses
for each unsyllabified input. The output of Gen accords with the most fundamental
structural principles, those that define what structures are to be contemplated as
possible, enumerating the vocabulary of categories and ensuring, for example, that o
dominates and not vice versa. Under these broad conditions, there will be a large
variety of candidate analyses for any given input; Universal Grammar gives a set of
well-formedness conditions, which, ranked into a grammar, will select the optimal
candidate from among the set of possibilities.
In other words, in OT as in Kahns syllable structure assignment rules,
syllabification is conceived of as a procedure which takes an unsyllabified string of
segments as input, and generates a syllable structure as output. I now turn to some
specific examples of OT syllabification.
As stated above, languages have a tendency to maximize onsets and minimize
codas: this accounts for the fact that the syllabification of CVCV is CV.CV, not
CVC.V. A CVCV schema is also used by McCarthy and Prince as the basis for
demonstrating syllabification in OT. Of the two forms in (7) (given by McCarthy and
Prince as candidate syllabifications of CVCV), (b) will be selected as optimal because
(a) violates two constraints, neither of which is violated by (b). The first of these is
NOCODA, which requires that syllables end on vowels, and the second is ONSET,
the constraint that syllables should begin with consonants.
(7) (a) CVC.V
(b) CV.CV
These are, however, not the only candidate syllabifications which McCarthy
and Prince consider for this schema. Further possibilities are listed in (8):
(8) (a) C.V.C.V
(b) <CVCV>
53
In (8a), each of the segments is in its own syllable; (8b), on the other hand, is wholly
unparsed (indicated by the angle brackets). In other words, it has no syllable structure
whatsoever. The former violates NOCODA twice, however, and also has two syllabic
consonants, and is therefore rejected; the latter violates the constraint PARSE, which,
as McCarthy and Prince state, is nothing other than Its (1986) Prosodic Licensing
constraint. Candidate (8b) is therefore also rejected.
Different rankings of the syllabification constraints will yield different syllable
structures. As can be seen from McCarthy and Princes Condition (ii) above, this fact
is used in OT to account for discrepancies in syllabification between one language
and another. For example, if the constraint PARSE is ranked higher than NOCODA (as
is presumably the case in English and Russian), then the preferred syllabification of,
say, !9%4! will be !.9%4.!, even though this violates the coda constraint; but if, for
example, NOCODA were ranked higher than PARSE (as might be the case, for instance,
in Mazateco, where codas are unattested, according to Blevins 1995: 217), then the
preferred syllabification will be !.9%.<4>!. In other words, !4! remains unsyllabified.
McCarthy and Prince state that in such cases unparsed elements are erased upon exit
from the level, in accordance with the principle of Stray Erasure, discussed earlier in
Steriade (1982) and It (1986). The principal attraction of OT is that there is believed
to be a single universal set of constraints, with language-specific differences
explained by different rankings of the constraints. The ordering of constraints in OT
fulfils the same function as the ordering of rules in Kahn (1976), or the directionality
of mapping in It (1986). However, the violability of constraints, one of the hallmark
54
features of OT, is the aspect of the framework to which I shall return in section
2.2.3.5.
2.2.3.4 Phrase-structure analysis of syllable structure
An alternative approach to syllable structure is in terms of phrase structure (or
in pre-generative terms, constituent structure). This approach is taken, for example,
by Harris (1951: 278f.), Hockett (1955: 52, 176), Haugen (1956), Cheng (1966),
Fudge (1969), Selkirk (1982a), Church (1983), Coleman (1990, 1996, 1998), Dirksen
(1993), Walther (1993), and Bird (1995). Phrase-structure grammar has been more
widely used in branches of linguistics other than phonology (see for example Gazdar,
Klein, Pullum and Sag 1985 and Pollard and Sag 1987, both phrase-structure analyses
of syntax and semantics).
The following quotation from Haugen (1956: 217) serves to illustrate the
general approach as applied to syllable structure:
Analysis of the syllable usually permits us to distinguish an irreducible minimum
which we may call the nucleus and an optional remainder which we may call the
margin Each of the constituents of the syllable consists of one or more phonemes,
with the vowels usually occupying the peak [read nucleus PC], the consonants
the margins; but consonants may also be included in the peak if vowels fail to occur
initially or finally. For each constituent one may then establish the number of
positions. For example, in Swedish the number of positions is anywhere from zero to
three. This we may represent by the formula (CCC).
The rules which Haugen suggests (for Swedish, in this instance) amount to
phrase structure rule schemata as follows:
(9) syllable (margin) nucleus (margin)
(10) margin C
(11) margin CC
(12) margin CCC
(13) nucleus V
55
A phrase-structure approach to syllabification is independent of, but not
necessarily opposed to, an approach based on sonority, since it is possible to
incorporate sonority into phrase structure rules: for an example, see Walther (1993).
For example, Haugens approach, which does not differentiate between the sonority
of different consonants, can only stipulate the maximum size of syllable margins (and
by extension syllables). For instance, in Haugens example, we saw that Swedish
onsets consist of up to three segments. This example, however, does not capture the
generalization that !90! is a permissible onset whereas !09! is not. However, Haugens
rule in (11), for example, could be rewritten as the rules in (14)-(15):
(14) onset obstruent liquid
(15) onset obstruent glide
By changing margin to onset, and C to obstruent, liquid, and glide,
then adding appropriate definitions to show what the possible instantiations of
obstruents, liquids and glides are, it is possible to capture generalizations about
sonority within phrase-structure grammar.
Phrase-structure analysis of syllable structure has the advantage that it
abstracts away from the issue of syllable structure assignment. Phrase-structure
analysis essentially involves making statements about the possible shapes of syllables,
rather than specifying how syllable structure is built up. The advantage of separating
the two is that the nature of syllable structure is a linguistic issue, which linguists
should properly focus their attention on, while the problem of how to assign structure
on the basis of a phrase-structure grammar is a computational problem which has
56
already been solved (in a number of different ways) by computer scientists. There is
the danger, therefore, that linguists who attempt to devise algorithms for assigning
syllable (or other linguistic) structure are re-inventing the wheel. This is a point to
which I shall return in section 2.2.3.5.
As mentioned in section 1.3.1, PSGs also have the advantage over all the
theories of syllable structure assignment reviewed in 2.2.3 that they have a transparent
relationship to DCGs. This means that a phrase-structure theory can straightforwardly
be implemented computationally, tested for coherence and coverage, and debugged if
necessary.
2.2.3.5 Syllable structure assignment: conclusions
As we have seen in sections 2.2.3.1-2.2.3.3, the critical factor which
differentiates alternative theories of syllable structure assignment is the way in which
each decides upon the correct syllabification in cases where there are multiple
possibilities. Kahn (1976) is a bottom-up, left-to-right algorithm; It (1986) is a top-
down algorithm (which is left-to-right or right-to-left, depending on the language);
and McCarthy and Prince (ms 1993) relies on ordering constraints and satisfying first
those which are highest-ranked, and then those which are ranked lower if possible. All
three theories rely in some way on ordering of rules. The arguments typically
advanced for one or another theory rest on their descriptive adequacy for example,
their ability to assign structures to well-formed strings, to place syllable boundaries in
appropriate positions, to account correctly for epenthesis rather than the
metalinguistic considerations discussed in section 1.3 (for example, the psychological
plausibility or power of one or another syllable structure assignment theory).
57
If we confront these more fundamental considerations, it is apparent that there
are problems with each of the syllable structure assignment mechanisms. One might
argue that Kahns algorithm is psychologically implausible because it involves as
many passes through the string as there are steps in the algorithm: it is unrealistic to
suppose that syllabification would take place in this way in the brain. We have seen
that directionality is a cornerstone of Its theory, yet the very principle of
directionality violates the order-free composition constraint discussed in 1.3.2, again
calling into question the theorys psychological plausibility. Finally, two aspects of
OT are problematic. The first is that because constraints are violable, it follows that
not all constraints are surface-true, and thus the question arises as to whether OT
constraints really are constraints. The second aspect of OT which is problematic is
the ordering of constraints, which again is counter to the order-free composition
constraint.
The more important point, however, is that it is not clear that there is any
advantage in arguing for bottom-up syllabification (as in Kahn 1976) versus top-down
syllabification (as in It 1986), when there a variety of reliable parsing techniques
9
which can assign syllable structure in different directions. It is important to separate
the issue of directionality of parsing (or, for that matter, any issue relating to how
syllable structure is assigned) from the grammatical constraints themselves. In this
chapter, the context-free syllabification grammar proposed for Russian can be applied
left-to-right or right-to-left with the same effect, and the same applies to all other

9
There is no need to list all the different available algorithms here, but a number are discussed in
Grishman (1986: 27-34). The algorithm used in DCG parsing is a top-down, left-to-right algorithm
with backtracking.
58
grammar modules described in this dissertation. It is the grammatical constraints
which account for phonological phenomena, not the implementation of the
grammatical constraints as a particular procedure.
To summarize, the debate on syllable structure assignment misses the main
phonological point, when, as is still frequently remarked upon (as in Clements 1990
and Bethin 1998), basic questions about syllable structure (for example, whether the
SSG is an adequate theory about syllable structure, and the extent of its adequacy)
still remain to be answered. Accordingly, I shall not outline a separate syllable
structure assignment algorithm to accompany the linear grammar presented in section
2.3, because this would be unnecessary. The algorithm used to assign syllable
structure computationally is simply the built-in DCG parsing algorithm. The grammar
itself does not depend on any particular ordering of rules. Because it is a linear
grammar, it can be used both to generate legal syllable structures and for parsing.
Furthermore, because the linear grammar is not deterministic, it is able to return
multiple syllabifications in cases like CVCCVC (which might be syllabified either as
CV.CCVC, or as CVC.CVC.) In particular, the deterministic algorithms of Kahn and
It can return only one solution for each string, which might be regarded as a
drawback. The insight that CV.CCVC is, in many cases, preferred to CVC.CVC, can
still be modelled within a linear grammar which returns multiple solutions, by
incorporating a probabilistic component into the grammar; I shall elaborate upon this
in section 2.4.
2.3 A linear grammar of Russian syllable structure
This section presents a specific proposal to describe the syllable structure of
Russian in terms of linear grammar. The purpose of this grammar is to parse Russian
59
word-forms exhaustively into syllables
10
, and the grammar will be subjected to
computational testing.
In order to build up a grammar of syllabification, two basic elements are
needed: the first of these is a lexicon of pre-terminal rules, specifying the alphabet
of segments which may occur in Russian together with all the segments relevant
features, and the second is the list of rules which assigns hierarchical structure to
these segments. These elements of the grammar will be considered in 2.3.1 and 2.3.2
respectively.
2.3.1 The phonological inventory of Russian
2.3.1.1 Preliminaries: controversial issues
In order to decide the alphabet of segments available for building up syllable
structure in Russian, we need first to take a position on three well-known
controversies in Russian phonological classification. These issues continue to be
debated: it is not my intention here to provide detailed arguments or computational
analysis substantiating one or another position, but it is necessary all the same to
consider the ramifications a decision one way or the other will have. The aim of this
section is to review the relevant literature and clarify where the current phonology
stands in relation to the rest of the field.

10
In the following discussion, I make the assumption that there is no ambisyllabicity in Russian. This
assumption may be an oversimplification, although see Selkirk (1982a) for arguments supporting my
position. In any case, the issue is not of major significance in what follows, and it would be possible to
reformulate the grammar to accommodate ambisyllabicity: Walther (1993) and Bird (1995: 62-63)
provide examples of how this can be done in context-free grammar.
60
The first controversy is whether [,) and [,] are allophones of a single phoneme.
This has been a matter of long-standing debate between the so-called Moscow and
Leningrad phonological schools: a detailed survey of this debate is given in
Reformatskij (1970). The Moscow line (or the allophonic view, as I shall call it),
which probably represents the majority view in phonological theory, is that [,] is an
allophone of !,!; the alternative (the independent phoneme view) is to include both
!,! and !,! as independent phonemes.
Advocates of the allophonic view (for example, Jakobson 1929, Avanesov
1956, Halle 1959
11
, Oliverius 1974, Panov 1979, the most recent Academy Grammar
[vedova et al. 1980, 1: 70], and Rubach 1995) have pointed mainly to the
complementary distribution of (,) and (,): the former occurs word-initially, after a
palatalized consonant or after a vowel, but (,) generally occurs only after non-
palatalized consonants. Examples are given in (16):
(16) </.$ (A,&') to be
<).$ (A',&') to hit
The independent phoneme view, which is held, for example, by ;erba (1912),
Gvozdev (1949), Larin (1959), Bulanin (1970), Lightner (1972), Matusevi; (1976)
and Kuhn Plapp (1996), has appealed in part to the articulatory and acoustic
differences between [,) and [,]: ;erba, for example, noted that Russians can easily
articulate [,] in isolation, whereas they have more trouble articulating some of the

11
But not apparently Halle (1997), to judge by the entry for the nominative plural in Halles Table 1
(op. cit., p. 281): this entry distinguishes the inflectional endings -y !,! and -i !,!.
61
uncontroversial positional variants of vowels such as (G) (the variant of !%! which
occurs between palatalized consonants). The independent phoneme view also hinges
in part on fringe phonological phenomena in Russian. For example, it is argued (e.g.
in Bulanin 1970: 80) that forms such as those in (17) are counter-evidence to the
proposition that (,)<never occurs word-initially, and indeed that the minimal pair in
(17a) is irrefutable evidence of the phonemicity of !,!.
(17) (a) )%&.$ (,2%&') to produce the sound (,)
/%&.$ (,2%&') to produce the sound (,)
(b) =0-,' (,7#64) [place-name]
='#)' (,4$',4) [place-name]
It has to be said, however, that such oppositions are rare: in particular, the examples in
(17b) are non-native place-names which sound strange to the Russian ear. Forms
where (,) appears to take on phonemic status might therefore be explained as
exceptions, without incurring a great empirical cost.
A more convincing phonological argument for the independent phoneme
view is put forward by Gvozdev (1949) and Kuhn Plapp (1996). This argument
focuses on the alternation of [,] and (,) at the edge of morphemes. Kuhn Plapp notes
that while an analysis which classifies [,] as an allophone of !,! is unproblematic
morpheme-internally, it is hard to explain, for example, why (B6&6") alternates with
(B6&6"') in (18) unless there are two distinct phonemes !,! and !,!. However, the
segments (,) and (,) are nonetheless in complementary distribution, a fact which the
phonology should explain.
62
(18) 5,.,( [B6&6") ready (predicate, masc.)
5,.,(/0 [B6&6"+,7) ready (attributive, masc. sg.)
5,.,().$ [B6&6"'+,&') to get ready (transitive)
I will return to take a position on this point below.
The second issue is whether the Russian velars (2
7
, B
7
, C
7
) are allophones of !2,
B, C! respectively, or whether !2
7
, 2, B
7
, B, C
7
,<C! are all phonemically distinct. Again,
the Moscow school takes the allophonic position, and the Leningrad school, the
independent phoneme position.
Essentially, the Moscow school arguments hinge on the fact that (2
7
, B
7
, C
7
) are
generally in complementary distribution with (2, B, C) (Avanesov 1956: 135-136,
Press 1986: 156): the former appear before the front non-low vowels !,=<1!, and the
latter elsewhere (before !%=<6=<;!, before consonants, and word-finally). This rule is
reflected in the fact that there are no orthographic sequences XY, where X stands for
any of %, 5, and : and Y stands for any of >, 2, ? and $. There are no exceptions to the
complementary distribution of (2
7
, B
7
, C
7
) and (2, B, C)<morpheme-internally in the
native vocabulary of Russian. Generally, the rule also holds across morpheme
boundaries, as shown by the examples in (19):
(19) #,-%+& ($6#2+%) blackboard (nom. sg.)
@!&9+& (?0%*+%) sentence (nom. sg.)
#,-%+) ($6#2'+,) blackboards (nom. pl.)
@!&9+/ (?0%*+,) sentences (nom. pl.)
The exceptions which can occur across morpheme boundaries, however, in some
cases result in minimal pairs. These are used to argue for the independent phonemicity
63
of !2
7
, 2, B
7
, B, C
7
,<C! (Gvozdev 1949: 64-65, Panov 1968: 80, Oliverius 1974: 104-
105). The minimal pairs in (20) are commonly cited as examples
12
.
(20) (a) A.,. %,. (1&3&<26&) this cat
A., .%+2. (1&3<&2'+6&) this weaves
(b) -&#%,+ (#H&26+) animal nursery (instr. sg.)
-,.%+2+ (#H&2'+6+) we shall weave it together
(c) <"!"5+& (A'D0'DB+%) banks, shores
<"!"5+> (A'D0'DB'+%) keeping (gerund)
Halle (1959) appears to be alone in not treating all the velars alike: his inventory (p.
45) assigns independent phonemic status to !2'! but not !B'=<C'!. His discussion of the
subject is brief, but it appears that he considers !2'! an independent phoneme on the
basis of the examples in (20b) (op. cit., p. 48), whereas he finds no similar evidence
for !B'=<C'!. Evidently, Halle did not consider examples such as (20c), which would
presumably have led him to postulate !B'! in addition.
The third controversial issue is the status of the long palatalized fricatives
(8'I=<5'I), as in the examples in (21).
(21) (a) 1). (8'I,&) shield
(b) !"93)% (0'18'I,2) engraver
(c) #!,**) ($065'I,) yeast
Although length is not generally phonemic in Russian, Avanesov (1956: 134) treats
!8'I=<5'I! as independent phonemes. Thelins (1974, 1983) position is that the phoneme
!8'I! exists morpheme-internally (as in (21a)), but that (8'I) represents a combination of

12
Only relevant morpheme boundaries are shown.
64
two phonemes elsewhere (see below). Thelins argument is that the phonology of
Russian is currently being restructured, so that while the grapheme 1 stood for two
phonemes previously in Russian, this is not the case in contemporary Russian, since,
as noted by Gvozdev (1949: 67), (8'I=<5'I) in Russian are acoustically indivisible. It has
also been suggested (e.g. by Bulanin 1970: 83) that the issue of whether !8'I=<5'I!
should be included in the inventory has been confused by the existence of the Cyrillic
grapheme 1, which stands for (8'I).
The more widely held viewpoint is that (8'I) is the phonetic realization of the
fricative-affricate sequences !#&8'!, !8&8'!, !5&8'!, and (5'I) the phonetic realization of !55!.
This means that !8'I=<5'I! can be excluded from the phonemic inventory. This is the line
followed by Gvozdev (1949: 42), Halle (1959: 45), Bulanin (1970: 51), Oliverius
(1974: 103) and Coats (1997). Gvozdev and Oliverius point out that if !8'I=<5'I! are
phonemes, then cases arise where a single phoneme would have to be split across
morpheme boundaries, as in the example in (22).
(22) - + 3"+ (8'I1+) with what
This reason for treating !8'I=<5'I! as sequences of phonemes, it might be countered, is
more an artefact of the limitations of segmentally-based phonology than of the
properties of !8'I=<5'I! in Russian. However, Coats adduces further evidence to suggest
that (8'I) is a sequence of two phonemes. He observes that certain morphological
phenomena in Russian are sensitive to the number of consonants in certain positions:
an example is the two consonant rule for the formation of imperatives of i-stem
verbs with stressed roots. This rule states that the imperative forms end in !,! if the
65
stem ends in two or more consonants (as in (23a)), but in zero if the stem ends in a
single consonant (as in (23b)).
(23) (a) %,:'3) !264&8'+,! finish!
,.(/:%') !6&",24+,! get out of the habit!
(b) <!,:-$ !A06#'+! throw!
(-.&: '$ !"#&%4'+! 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!
No. of
consonantal
phonemes in
inventory
No. of
vowel
phonemes in
inventory
Avanesov (1956) 34 5
Halle (1959)

33 10
Bulanin (1970)
35 6
Gvozdev (1949)
35 6
Kuhn Plapp (1996)

* * * *
Coats (1997) * * * *

* *
Oliverius (1974) 35 5
Panov (1979) 34 5
Key to Table 5
Phoneme is included in given inventory
Phoneme is excluded from given inventory
* No explicit evidence available from given work
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 !,= 2
7
= B
7
= C
7
=<8'I=
5'I! from the phonological inventory, and thus to opt for the more minimal approach in
each of the three areas discussed above. This approach is followed not just in this
chapter, but also in Chapter 3. My strategy in accounting for Russian morphology is
to deal with cases like (B6&6") ~ (B6&6"') by listing both alternants as separate items
in the morpheme inventory rather than by rule. In other words, I accept an
enlargement of the morphological inventory for the sake of a reduction in size of the
phonological inventory. This is not to say that it would be impossible to formulate an
67
analysis whereby the phonological and morphological inventories are both minimized,
and all alternations, phonological and morphological, are accounted for by rule.
However, an attempt to deal with this aspect of Russian morphophonology
comprehensively and consistently would be the subject of another dissertation.
My assumptions reduce the size of the inventory to 37 phonemes (32
consonants and 5 vowels), as follows:
Table 6. The phonemic inventory of Russian
Vowels:
Front Back
High: !,! !;!
Mid: !1! !6!
Low: !%!
Consonants:
Labial
Bilabial Labio-
dental
Dental Alveolar Palatal Velar
Hard Soft Hard Soft Hard Soft Hard Soft Hard Soft Hard Soft
unvoiced
9 9
7
& &
7
2
stop
voiced
A A
7
$ $
7
B
unvoiced
&# &8
7
affricate
voiced
unvoiced
? ?
7
# #
7
8 C
fricative
voiced
" "
7
* *
7
5
unvoiced approximant
voiced
7
unvoiced nasal
voiced
+ +
7
4 4
7
unvoiced lateral
voiced
- -
7
unvoiced trill
voiced
0 0
7
68
2.3.1.2 The classification system
In the syllable structure grammar which I develop below, the following 9
features are also used. The features and their meanings are substantially the same as in
Chomsky and Halle (1968: 298ff).
(24) Features
[del(ayed)_rel(ease)] [coron(al)] [voc(alic)]
[anter(ior)] [cont(inuant)] [later(al)]
[cons(onantal)] [nasal] [sonor(ant)]
The classification of the Russian phonemes with respect to these features is shown in
the feature matrix in Table 7 on page 69. (This table also shows the relative sonority
of the phonemes, following Selkirks (1984) schema in (2) above.)
T
a
b
l
e

7
.

C
l
a
s
s
i
f
i
c
a
t
i
o
n

o
f

R
u
s
s
i
a
n

p
h
o
n
e
m
i
c

i
n
v
e
n
t
o
r
y
w
i
t
h

r
e
s
p
e
c
t

t
o

C
h
o
m
s
k
y

a
n
d

H
a
l
l
e

s

f
e
a
t
u
r
e
s
T
h
i
s

t
a
b
l
e

l
i
s
t
s

o
n
l
y

t
h
o
s
e

f
e
a
t
u
r
e
s

r
e
l
e
v
a
n
t

t
o

t
h
e

b
a
s
i
c

s
y
l
l
a
b
l
e

s
t
r
u
c
t
u
r
e

g
r
a
m
m
a
r
;

T
a
b
l
e

1
4

o
n

p
a
g
e

1
0
3

l
i
s
t
s

f
u
r
t
h
e
r

f
e
a
t
u
r
e
s

w
h
i
c
h

w
i
l
l

b
e

i
n
t
r
o
d
u
c
e
d

t
o

a
c
c
o
u
n
t

f
o
r

p
h
o
n
o
l
o
g
i
c
a
l

p
r
o
c
e
s
s
e
s

i
n

R
u
s
s
i
a
n
.
70
Some comments about this classification system are in order.
First, the classification of !7! as [+consonantal, +vocalic] departs from
Chomsky and Halle (1968), who classify the nearest English segment y as
[consonantal, vocalic] (op. cit., p. 176). However, under the analysis followed here,
!7! and !,! are distinguished only by the feature [consonantal]. This is in line with the
idea that the syllabicity of segments depends solely on their position (see page 39
above), because the function of the feature [consonantal] here is simply to
distinguish those segments which can associate only with nucleus slots (the
[consonantal] segments) from those which may or may not be able to associate with
nucleus slots (the [+consonantal] segments). However, the possibility of
[+consonantal] segments associating with nucleus slots is never exploited in Russian,
unlike in English (cf. bottle (AF&-J)).
Secondly, where in Table 7 a phoneme is specified + or with respect to any
feature, this specification is, in effect, lexical: no specification of this type can later be
changed, since feature-changing is not allowed in declarative grammar. This means,
for example, that if a phonological rule applies to [+continuant] segments, then we
know for certain that the rule will at no stage in the grammar apply to !9!, which is
lexically specified [continuant].
Third, any feature matrix of the type in Table 7 implies the existence of certain
phonological natural classes. Natural class is used here to mean a set of phonemes
which unifies with a particular structural description, and a structural description is
any feature or combination of features. A phoneme unifies with a structural
description if, for every feature in the structural description, it meets one of two
conditions: either it has a feature with the same specification, or it is unspecified for
71
that feature. An example of the first type is !&#!, which is the only phoneme which
unifies with [+consonantal, vocalic, sonorant, +delayed_release, +anterior]. The
second type is exemplified by !"! and !"'!, which unify with both [+sonorant] and
[sonorant]. (This underspecification of !"! and !"'! will be exploited in the analysis
which follows.) In cases of the second type, one may think of the phonological rules
as filling in the relevant features, together with the + or specification.
The fourth point is an observation about the notation which will be used. The
features listed in Table 7 are used to enrich the linear grammar: each constituent may
be associated with a feature, or combination of features. The notation that will be used
subscripts the feature or feature-combination beneath the name of the node. (This
approach is equivalent to the one followed in Coleman and Pierrehumbert 1997.)
Thus in (25) [+consonantal, sonorant] is subscripted beneath the first segment node
as a way of denoting any segment which unifies with [+consonantal, sonorant].
(25) onset segment segment
[+continuant, [+consonantal,
sonorant] continuant]
In addition to the 9 features just listed, a further two binary features will be
referred to: [init(ial)] and [final]. These features do not, however, classify
individual phonemes; instead, they classify the constituents weakfoot, strongfoot,
syllable, onset and rime. Any of these units can be any combination of [+init], [init],
[+final] or [final], providing that no unit is both [+init] and [init], or both [+final]
and [final]. The features [init] and [final] are used in essentially the same way in
Dirksen (1993).
72
Further features will be introduced in section 2.5.1 to deal with phonological
processes other than syllabification.
2.3.2 The syllable structure rules
With the classification system in place, the formulation of the pre-terminal
rules is straightforward. There will be one rule for each phoneme, and each rule will
state the information which is lexically specified for the given phoneme. For example:
(26) segment !7!
[+consonantal,
+vocalic,
+sonorant,
+continuant,
coronal,
anterior,
lateral,
nasal,
delayed_release]
segment !0!
[+consonantal,
+vocalic,
+sonorant,
+continuant,
+coronal,
anterior,
lateral,
nasal,
delayed_release]
The remaining rules are where the phonotactic constraints of Russian are
stated. In what follows, I shall assume the following basic syllable structure:
(27) syllable (onset) rime
rime nucleus (coda)
73
The constituents onset, rime, nucleus and coda are adopted as conforming to
standard phonological theory (Kenstowicz 1994: 252-253). This kind of syllable
structure is defended in Selkirk (1982a), and is also in line with recent computational
analyses such as Walther (1993) and Coleman and Pierrehumbert (1997), although
Pierrehumbert and Nair (1995) have suggested that syllable structure may be flatter
than (27) suggests. However, the position one takes on this matter is not of major
significance in what follows.
Using the classification system outlined in the previous section, the grammar
may be enlarged to define the constituents onset, nucleus and coda as follows:
(28) onset segment
[+cons]
nucleus segment
[cons]
coda segment
[+cons]
Russian, in common with nearly all
13
languages, permits CV syllables. Unlike a
number of languages (e.g. Fijian, according to Blevins 1995: 219), however, it also
permits rimes with codas. One of the grounds for assuming this is that there are many
Russian words which end in a consonant. One might, therefore, assume that the rules
in (29) are rules of Universal Grammar, while (30) is specific to those languages
(including Russian) which permit codas. (Alternatively, one could view (30) as

13
The idea that VCV is cross-linguistically syllabified V.CV, is (in standard phonological theory) held
to be an inviolable principle of Universal Grammar (see for example Rubach and Booij 1990: 124, 125,
144). However, see Breen and Pensalfini (1999) for discussion of Arrernte, which (they argue) does not
permit CV syllables.
74
universal but parametrical, and activated for Russian but not Fijian, in the style of the
analysis proposed in Coleman, Dirksen, Hussain and Waals 1996.)
(29) (a) syllable onset rime
(b) rime nucleus
(30) rime nucleus coda
Russian also permits onsetless syllables, as evidenced by the fact that many Russian
words begin with a vowel. Again, this is not a linguistic universal: Blevins (1995:
219) cites Cairene Arabic and Dakota as examples of languages in which all syllables
have an obligatory onset. Thus we may assume that (31), like (30), is activated only
for certain languages; but since Russian is one of these languages, (31) must be
included in the current grammar.
(31) syllable rime
Now, a realistic syllable structure grammar must be able to define words
consisting of any number of syllables. For the current grammar to be able to do this,
constituents must be added above the syllable node. The rule schemata in (32)
represent a first approximation to what might be required.
(32) prosodic_word (weakfoot) strongfoot
weakfoot syllable (weakfoot)
[stress]
strongfoot syllable (weakfoot)
[+stress]
75
These rules state that the minimal prosodic_word must consist at least of a single
stressed syllable (dominated by the intermediate strongfoot node). The fact that the
second of these three rules is recursive (refers to itself) means that unstressed
syllables can be added to the stressed syllable, theoretically ad infinitum. The feature
[stress] is adopted without discussion here, although the way in which the grammar
governs its distribution is examined in detail in Chapter 5. On the basis of these rules,
the tree-structure associated with, say, !+%+%! would be as in Figure 3:
Figure 3. Tree-structure for /mama/
The rules up to this point form the basis of the syllable structure grammar. To
give some idea of the coverage of this skeleton grammar, the rules so far assign
prosodic_word
strongfoot
onset rime
nucleus
!%!
weakfoot
syllable
[stress]
syllable
[+stress]
!+!
onset rime
nucleus
!%! !+!
76
structure to 7,553 word-forms of a total sample of 11,290 tested (in other words, to
around 669% of the forms tested).
Part of the reason that the coverage of the grammar so far is not higher is that
Russian also allows complex onsets and codas. Again, this is in contrast to various
other languages (e.g. Finnish and Cairene Arabic, according to Blevins 1995: 219).
Although one should be cautious about Kahns (1976) suggestion that the sets of
onsets and codas are identical to the sets of word-marginal onsets and codas
respectively, there is no doubt that one can learn a great deal about onsets and codas
in general by looking at word-marginal clusters. First of all, therefore, let us consider
onsets.
Table 8 on page 77 shows the relative frequencies of different types of two-
segment onset based on entries in Zaliznjak (1977). (This table abstracts away from
the voiced/voiceless and palatalized/unpalatalized distinctions, so for example !"!
stands for !"!, !?!, !"'! and !?'!. The total number of word-initial onsets in the sample
(19,978) is the number of entries in Zaliznjaks (1977) dictionary which have initial
onsets of exactly two consonants.)
T
a
b
l
e

8
.

D
i
s
t
r
i
b
u
t
i
o
n

o
f

w
o
r
d
-
i
n
i
t
i
a
l

o
n
s
e
t
s

b
y

t
y
p
e
(
b
a
s
e
d

o
n

d
a
t
a

f
r
o
m

Z
a
l
i
z
n
j
a
k

1
9
7
7
)
c
o
n
s
v
o
c
s
o
n
o
r
c
o
n
t
c
o
r
o
n
a
n
t
e
r
l
a
t
e
r
n
a
s
a
l
d
e
l
_
r
e
l
S
e
g
m
e
n
t
S
o
n
o
r
-
i
t
y
P r e c e d i n g s e g m e n t
78
As can be seen from Table 8, the structure of complex onsets in Russian is rather
chaotic. If all word-initial onsets in Russian obeyed the SSG, all the numerical counts
in the table would be to the left of the bold lines which run from top-left to bottom-
right. However, this is not the case: one can calculate from Table 8 that 13,520 of the
19,978 onsets (only 677%) strictly conform to the SSG. Clearly, the SSG is a good
start in accounting for the word-initial onsets of Russian, but something more is
needed.
Ideally, we would like to formulate a grammar which generates all the attested
onsets, and no others. However, the scattered nature of the entries in Table 8 means
that any such grammar would have to incorporate a large number of rules.
Nevertheless, there are a few quite simple generalizations which can be drawn out of
Table 8, to provide a good level of coverage (over 95%) of the data.
A quick inspection of Table 8 reveals first that most segments, except !&#!,
whose sonority does not exceed that of !+! and !4!, can precede !7!, !-! and !0! in initial
onsets. (Moreover, there is some indication that the lower the sonority of the initial
segment, the more likely the liquids are to be preceded by them; for example, !A0! and
!A-! occur with greater frequency than !C0! and !C-!. However, I shall not attempt to
model this observation here.) The facts point to the need in the grammar for a rule like
the following:
(33) onset segment segment
[+cons, [+cons,
voc, +voc]
del_rel]
79
The vast majority of the exceptions to the SSG are fricative-stop onsets (e.g.
!*$!, !#&!, !#2!). In particular, !"=<*! appear to be able to precede virtually any segment;
!8!, on the other hand, tends to avoid preceding other fricatives. Complex onsets which
are exceptional to the SSG but do not begin with either !"!, !*! or !8! account for only
just over 2% of the total number of word-initial onsets.
Goldsmith (1990) and Blevins (1995: 211) both note that similar apparently
exceptional onsets also occur systematically in English (e.g. stop, skin). One should
not automatically assume (as Kahn does) that the existence of any given word-initial
onset necessarily implies the existence of the same onset word-medially, because as
Blevins and a number of other researchers (for example Rubach and Booij 1990)
point out, many such violations of the SSG are often confined to the edge of the
syllabification domain. Thus the approach taken here (for the time being) is to
distinguish between word-marginal and word-medial onsets and codas, allowing
greater latitude in the word-marginal rules. The way in which this is done is to replace
the rule schemata in (34) with the rules in (35).
(34) prosodic_word (weakfoot) strongfoot
weakfoot syllable (weakfoot)
[stress]
strongfoot syllable (weakfoot)
[+stress]
syllable onset rime
80
(35) prosodic_word strongfoot
[+init]
prosodic_word weakfoot strongfoot
[+init, final] [init]
weakfoot syllable
[o init, | final] [o init, | final, stress]
weakfoot syllable weakfoot
[o init, | final] [o init, [| final]
stress]
strongfoot syllable weakfoot
[o init] [o init, [+final]
+stress]
syllable onset rime
[o init, | final] [o init] [| final]
rime nucleus coda
[o final] [o final]
To express the fact that certain onsets exist only word-initially, we specify [+init] in
the rules for those onsets. In all other onset rules, no specification is given for [init],
which effectively means that any word-medial onset can also be a word-initial onset,
but not vice versa. Later (in section 2.4), we shall see that the syllable structure rules
can be set up in such a way that the features [init] and [final] are unnecessary, but
such that the rules still model the improbability of finding certain clusters in anything
other than word-marginal position.
Now, in Russian it seems reasonable to suppose that fricative-stop clusters
beginning with !*=<#! may occur not only as initial onsets but as medial ones as well,
on the basis of forms such as ,.-.&.$ !6&#&%&'! to lag. Thus the following rule is
warranted.
81
(36) onset segment segment
[sonor, [cont]
+cont,
+coron,
+anter]
However, it is less clear that fricative-stop clusters beginning with !8! should be
allowed to occur other than as an initial onset: conservatism dictates that one should
assume that these clusters are only word-initial, unless there is evidence to the
contrary. To license such clusters as word-initial (but not word-medial) onsets, the
rule in (37) can be added to the grammar.
(37) onset segment segment
[+init] [sonor, [cont,
+cont, del_rel]
+coron,
anter]
(37) excludes the possibility of !&#! occurring as the stop, explaining the absence of
!8&#! onsets.
As can be seen from Table 8, !"! and !*! (together with their unvoiced
counterparts) can precede any segment. However, it seems that these segments can
precede only sonorants word-medially; there are forms such as ,.(!&1&.$
!6&"0%8&8'%&'! to avert, ,.(6"%&.$ !6&"-'12%&'! to divert, 4,#-+,.!".$ !96$#+6&0'1&'!
to spy on, 4,#-'"*')% !96$#4'154',2! snowdrop, ,.-(". !6&#"'1&! reflection,
,.9(8% !6&*";2! echo, but no forms beginning in 4,#(--, 4,#(9-, 4,#(B-, 4,#(*-,
4,#-:-, 4,#(:-. Thus the following rules appear to be justified:
82
(38) (a) onset segment segment
[+init] [sonor, [+cons]
+cont,
+anter]
(b) onset segment segment
[sonor, [+cons,
+cont, +sonor]
+anter]
The rules in (38) exploit the fact that !"! is lexically unspecified for the feature
[sonorant] in the phonological inventory; whether !"! is [+sonorant] or [sonorant] is
determined solely by the syllable structure rules. (38) does no more than express
formally a fact which has already been noted by a number of researchers, namely that
!"! behaves like an obstruent when it precedes an obstruent, and like a sonorant when
it precedes a sonorant (Jakobson 1956: 199, Andersen 1969, Coats and Harshenin
1971, Halle and Vergnaud 1981). Both Andersen (1969) and Halle and Vergnaud
(1981) link the behaviour of !"! to its putative phonological status as a glide !K!; for
evidence supporting this, see Andersen (1969) and Coats and Harshenin (1971).
However, the formal expression of the dual status of !"! by underspecification in the
phonological inventory, coupled with the idea that phonotactics determines whether
!"! is [+sonorant] or [sonorant], has not, to my knowledge, been suggested in the
literature to date. This analysis also pays a dividend in accounting for voicing
assimilation, because the behaviour of !"! with respect to voicing assimilation
depends precisely on whether !"! is [+sonorant] or [sonorant]. If [+sonorant], it
patterns with !-=<0=<+=<4!; if [sonorant], it patterns with other obstruents. This will be
discussed in more detail in section 2.5.2.4.
83
The rules in (38) overlap not only with each other, but also with (36). No
conflation is possible, however, because (36) and (38b) refer to onsets in general,
while (38a) applies to the more specific class of [+init] onsets.
Table 8 also shows that a rule is needed to account for nonsonorant-sonorant
onsets, which may occur in almost any combination. We have already accounted for
those in which the sonorant is also a liquid or !7! (see (33)). Thus we are left with the
nasals and !"!. Here, it seems that nonsonorant-nasal onsets (with the exception of
those dealt with by (38b)) are confined to word-initial position, whereas nonsonorant-
!"! onsets may occur in any position, as shown by the following examples:
(39) ,.B(/!)(&.$ !6&8",0',"%&'! to fling away
4,#:(&.).$ !96$C"%&',&'! to pick up
4,#.("!#).$ !96$&"'10$',&'! to confirm
,.%().&.$ !6&2"',&%&'! to give as good as one gets
,.7("-.) !6&&#"'1#&',! to finish blossoming
Additionally, there are no onsets (word-initial or otherwise) in which !&#! is the initial
segment and a nasal is the second. Thus the rules proposed here are as follows:
(40) (a) onset segment segment
[+init] [sonor, [+cons,
del_rel] voc,
+sonor,
+nasal]
(b) onset segment segment
[sonor] [+cons,
+sonor,
nasal]
84
Again, there is some overlap between these rules and previous ones, but no rule is
otiose.
Now, we move on to consider onsets consisting of three segments. The
segment most commonly found at the beginning of such onsets is !"!. A large number
of these then have !#! or !*! as the second segment, with any sonorant possible as the
third segment, as the examples in (41) show.
(41) (9C"9*&.$ !"*71*5%&'! to drive up
(9!/( !"*0,"! explosion
(96,+ !"*-6+! breaking in
(9+&: !"*+%C! flapping
(9',- !"*46#! payment
(9(&6).$ !"*"%-',&'! to load
I suggest the following rule to account for these:
(42) onset segment segment segment
[+init] [sonor, [sonor, [+cons,
+cont, +cont, +sonor]
coron, +coron,
+anter] +anter]
The other major category of three-segment word-initial onset which has !"! as
the first segment and !#! or !*! as the second, has one of the stops as the third segment.
The following rule can account for these onsets:
(43) onset segment segment segment
[sonor, [sonor, [cont,
+cont, +cont, del_rel]
coron, +coron,
+anter] +anter]
85
Also common are three-segment word-initial onsets where the second segment
is a stop and the third a liquid or !"!. In this case, either !"! or !#! can be the first
segment. For example:
(44) (<6)9) !"A-',*',! nearby
(<!,-).$ !"A06#',&'! to throw in
(568<$ !"B-;A'! deep down
(5!/9&.$-> !"B0,*%&'#'%! to sink ones teeth into
(#()5&.$ !"$"',B%&'! to move into
(#!85 !"$0;B! suddenly
(%6&# !"2-%$! contribution
(%!&.7" !"20%&&#1! in short
(46&($ !"9-%"'! swimming
(4!&(, !"90%"6! to the right
(.!&().$ !"&0%"',&'! to inveigle
-<!,-).$ !#A06#',&'! to throw down
-56&*)(&.$ !#B-%5,"%&'! to smoothe out
-5!8*&.$ !#B0;5%&'! to unload
-#(,).$ !#$"67,&'! to double
-#!8*).$ !#$0;5,&'! to bring together
-%6&# !#2-%$! store house
-%!/.$ !#20,&'! to hide
-46&(',0 !#9-%"467! floatable
-4!&(%& !#90%"2%! information
-.!&#&.$ !#&0%$%&'! to suffer
These onsets can be accounted for by the following rule:
(45) onset segment segment segment
[sonor, [cont, [+cons,
+cont, del_rel] +sonor,
+anter] nasal]
86
Finally, we need to consider four-consonant onsets. All such onsets in Russian
begin with !"#! or !"*!, and the overwhelming majority have a stop and a liquid as the
third and fourth segment respectively. Examples are given below:
(46) (9<!,- !"*A06#! upthrust
(956># !"*B-'%$! glance
(95!8-.'8.$ !"*B0;#&4;&'! to feel sad
(9#!,5'8.$ !"*$06B4;&'! to give a start
(-%6,%,3)(&.$ !"#2-6-6&8',"%&'! to tousle
(-%!/.$ !"#20,&'! to unseal
(-46"-% !"#9-'1#2! splash
(-4!/-%)(&.$ !"#90,#2,"%&'! to sprinkle
(-.!,).$ !"#&067,&'! to build (into)
These onsets can be accounted for using the following rule:
(47) onset segment segment segment segment
[+init] [sonor, [sonor, [cont, [+cons,
+cont, +cont, del_rel] +voc]
coron, +coron,
+anter] +anter]
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 Zaliznjaks (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.
F o l l o w i n g s e g m e n t
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 796%), 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 segment
[+cons, [+cons,
+voc] voc]
Table 9 also shows that almost any segment can precede !$! or !&! in a coda;
the only exception is !&#!. The following rule can be used to account for these codas.
(49) coda segment segment
[voc, [cont,
del_rel] +coron,
+anter,
del_rel]
In two-segment codas, a number of segments can occur after !4!. The majority
of these are of one of two types: either !*=<#=<8! or noncontinuants. Many such forms
are recent loan-words: examples are -"&'- !#'1%4#! show, 48'B !9;48! punch,
()'. !"',4&! screw, and <&'% !A%42! bank. These can be described by the following
rules.
90
(50) coda consonant consonant
[+cons, [sonor,
voc, +cont,
+sonor, +coron]
+nasal]
coda consonant consonant
[+cons, [cont]
voc,
+sonor,
+nasal]
I propose five further rules to cover most of the miscellaneous codas still
unaccounted for. A number of these codas also occur in comparatively recent loan-
words. The rules are shown in Table 10.
Table 10. Further coda rules
Coda type Examples of forms with this coda type Rule
!9#!, !2#! )'#"%- !,4$'12#! postcode
5)4- !B,9#! plaster
coda segment segment
[cont, [sonor,
coron] +cont,
+coron,
+anter]
!*B!, !#2! +,95 !+6*B! brain
,.48-% !6&9;#2! leave
coda segment segment
[+cont, [cont,
+coron, coron,
+anter, anter]
sonor]
!8&8'! ("1$ !"'18&8'! thing coda segment segment
[sonor, [+coron,
+cont, cont,
+coron, anter]
anter]
!*+!, !*4!,
!*"!, !##!
-,7)&6)9+ !#6&#,7%-',*+! socialism
*)9'$ !5,*4'! life
.!"9( !&0'1*"! sober
%6&-- !2-%##! class
coda segment segment
[sonor, [+cons,
+cont, voc,
+coron, +anter]
+anter]
!9=&=2=A=$=B!
+ !-=0!
(,46$ !"69-'! shriek
("4!$ !"'190'! wild boar
+".! !+'1&0! metre
7)%6 !&#,2-! cycle
!8<6$ !0;A-'! rouble
<,<! !A6A0! beaver
%&#! !2%$0! sequence
%"56$ !21B-'! body-size
.)5! !&',B0! tiger
coda segment segment
[cont, [+cons,
del_rel] +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 (9851%). 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 ,<1"-.(, !6A8&8'1#&"6! society).
Table 11. Exhaustive list of initial clusters not accounted for
by the syllable structure grammar
Initial clusters Example Phonological
transcription
Gloss
!&8'&!
3.,
!&8'&6!
what
!-'$!
6$#&
!-'$%!
ice (gen. sg.)
!B$'!
5#"
!B$'1!
where
!2&!
%.,
!2&6!
who
!-A!
6<&
!-A%!
forehead (gen. sg.)
!+&8'!
+3&.$->
!+&8'%&'#'%!
rush
!+4'!
+'"')"
!+4'14',71!
opinion
!+4!
+',5,
!+46B6!
many
!9&'!
4.)7&
!9&',&#%!
bird
!0&!
!.&
!0&%!
mouth (gen. sg.)
!#C"!
-:(&.).$
!#C"%&',&'!
seize
!&8&8'!
.1&."6$',
!&8&8'%&'1-'46!
painstakingly
!8&8'!
1"%&
!8&8'62%!
cheek
92
Table 12. Exhaustive list of final clusters not accounted for
by the syllable structure grammar
Final clusters Example Phonological
transcription
Gloss
!#-'!
+/-6$
!+,#-'!
thought
!&"'!
(".($
!"'1&"'!
branch
!5A!
#!8*<
!$0;5A!
friendships (gen. pl.)
!-4&#!
-,6'7
!#6-4&#!
suns (gen. pl.)
!--!
+".&66
!+'1&%--!
metal
!0-!
5,!6
!B60-!
throats (gen. pl.)
!#-!
-+/-6
!#+,#-!
sense
!++!
."6"5!&++
!&'1-'1B0%++!
telegram
!44!
.,''
!&644!
tonnes (gen. pl.)
!99!
5!844
!B0;99!
groups (gen. pl.)
!#&'0!
+)')-.!
!+',4',#&'0!
minister
!?0!
7)@!
!&#,?0!
figure
!4&0!
7"'.!
!&#14&0!
centre
!#&0!
,!%"-.!
!6021#&0!
orchestra
!42&!
48'%.
!9;42&!
point
!&8'"!
4,3(
!96&8'"!
soils (gen. pl.)
!2"!
<8%(
!A;2"!
letters (gen. pl.)
!#&"!
-81"-.(
!#;8&8'1#&"!
beings (gen. pl.)
!7#&"!
:,9>0-.(
!C6*'%7#&"!
economies (gen. pl.)
!-'#&"!
'&3&6$-.(
!4%&8'%-'#&"!
commands (gen. pl.)
!$#&"!
4!,)9(,#-.(
!906,*"6$#&"!
productions (gen. pl.)
!4#&"!
4!,-.!&'-.(
!906#&0%4#&"!
spaces (gen. pl.)
!0#&"!
5,-8#&!-.(
!B6#;$%0#&"!
states (gen. pl.)
!##&"!
)-%8--.(
!,#2;##&"!
arts (gen. pl.)
!&#&"!
6?<,4/.-.(
!-';A69,&#&"!
curiosities (gen. pl.)
!"#&"!
38(-.(
!&8';"#&"!
feelings (gen. pl.)
Many of the lexemes in question are of high frequency, and the noun-forming suffix
!#&"! is productive. However, it is also true that it would be unusual to find many of
these same lexemes in the plural (e.g. economies, curiosities), and it is notable that
a high proportion of the final clusters unaccounted for arise only in the genitive plural
93
forms. Since the figure of 9851% quoted above as the coverage of the syllable
structure grammar does not take token frequency into account, it is likely that a fairer
estimate of the grammars coverage, weighted by the frequency with which forms
actually occur in the language, would be even greater than 9851%.
Clearly, some of the rules proposed do not conform to the SSG (for example,
the rule in (36)). Now, suppose we add a constraint to the grammar to the effect that
all clusters must strictly conform to the SSG, and call the new grammar G
1
. (This is
not to say that G
1
contains all possible rules which adhere to the SSG, because some
rules of this type were never included. This, in turn, was because the rules in question
did not appear to be motivated by the data of Russian.) Suppose also that in G
1
we do
not allow for edge effects, as Blevins refers to them (that is, we do not allow
violations of the SSG even at the margins of words). It turns out that the coverage of
G
1
is 10,147 word-forms of the total 11,290 (that is, 8986%). Clearly, this is
significantly lower than the coverage of the grammar which I propose. One is forced
to conclude that the SSG without edge effects, although perhaps a good first
approximation, is too strong a constraint for Russian.
It is possible, however, to contemplate an alternative grammar (call it G
2
)
which adheres to the SSG but does allow edge effects. The segments which occur at
the left edge of initial onsets and violate the SSG are !"!, !*! and their unvoiced
counterparts (see (41)-(46) above). One might postulate two rules in G
2
, one which
allows a single extraprosodic [+consonantal, +continuant, sonorant, +anterior]
segment to be appended to the left edge of a prosodic word, and the second of which
allows a pair of extraprosodic segments, both of which are [+consonantal,
+continuant, sonorant, +anterior], with the additional constraints that the former is
94
[coronal] (!?! or !"!) and the latter [+coronal] (!#! or !*!) (cf. (46) and (47)). A rule to
license extraprosodic [+consonantal, +continuant] segments at the right edge of a
prosodic word (as in +".! !+'1&0! metre, (".($ !"'1&"'! branch, %,!&<6$ !260%A-'!
vessel, %,++8')9+ !26++;4',*+! communism) could also be included in G
2
. The
coverage of G
2
without the rule to license extraprosodic segments at the right edge of
the word is 10,839 forms out of 11,290 (9601%), and with the rule is 10,889 forms
(9645%). Although significantly better than the coverage of G
1
, both of these are still
lower than the figure of 9851% quoted above, and the difference is highly significant
(with p < 334 10
24
, in the case of G
2
with the rule for extraprosodic segments at
the right edge).
The main reason that the coverage of G
2
is significantly lower is that a
grammar which allows edge effects can never account for clusters which violate the
SSG and occur word-medially. Rules (42), (43), (45) and (47), while very close to
being extraprosodic licensing rules of the type suggested, are not always limited in
their applicability to word-initial clusters. The limitation of sonority violations to
word edges is also too strong a constraint for Russian, as can be seen from forms such
as ,<-.&',(%& !6A#&%46"2%! situation, 4!"#-.&()."6$ !90'1$#&%"',&'1-'!
representative, and ,.-.!&').$ !6&#&0%4',&'! to set aside, where !#&! and !#&0! appear
to be word-medial onsets. At least 226 (2.00% of the total 11,290) of the words
unparsed by G
1
are of this type: they clearly do not involve SSG violations at the
word margins.
Another problem with allowing edge effects is that it weakens the force of the
SSG. If edge effects can be allowed, then one must state cross-linguistically first
whether any type of segment can be extraprosodic, and secondly, whether is there a
95
limit to the number of extraprosodic segments allowed. Even with these restrictions,
there is a danger that the range of consonantal clusters allowed at word edges for any
given language will be too wide (in other words, the grammar will over-generate).
There is no doubt that the SSG is helpful as a general consideration in
formulating syllable structure rules. Perhaps the best way of viewing sonority is as a
linguistic universal or default which provides basic syllabification rules, some of
which are more core and apply to all languages (which one might think of as
Universal Grammar rules), and others of which are in activated only in certain
languages. However, it is difficult to see how the SSG can be incorporated formally
into a grammar of Russian without sacrificing some descriptive adequacy: the ideal of
a few very general rules to account comprehensively for Russian syllabification is
undoubtedly hard to attain. In any case, I would not want to claim that the boundary
between Universal Grammar and language-specific grammar is necessarily a very
clear one, and perhaps the generality of different syllable structure rules is more a
matter of degree than of absolutes.
2.4 A heuristic for deciding between multiple syllabifications
In 2.2.3, I reviewed three theories of syllable structure assignment, and
showed that a key distinguishing factor between them is the approach each takes in
cases of possible ambiguity: for example, should VCCV be syllabified V.CCV or
VC.CV, assuming that either is possible? Because the tests of coverage described in
2.3.2 were tests of the grammars descriptive rather than explanatory adequacy, we
were able to ignore this question: we were interested only in whether the grammar
assigned any structure to strings.
96
However, the importance of assigning the right structure, rather than just any
structure, becomes particularly clear in the theory of stress assignment which I
propose in Chapter 5: whether a string is syllabified VC.CV or V.CCV may, in some
cases, be critical in determining which syllable of the form is stressed. Thus it is
important to have some mechanism for choosing one syllabification over other
possible syllabifications.
Coleman and Pierrehumbert (1997) propose just such a mechanism, a
probabilistic syllabification grammar, which lists all possible analyses but assigns to
each a probability. The probability of a given analysis applying is established by
tabulating the analyses over all parsed words in a training set. The theoretical
advantage of this solution is that it adheres to the order-free composition constraint
(see 1.3.2), unlike any of the syllable structure assignment theories reviewed in 2.2.3.
In this section, a mechanism similar to that used by Coleman and
Pierrehumbert is outlined. The training set was the dataset described in 1.5, and the
training process followed that of Coleman and Pierrehumbert in two respects. First,
every single form in the dictionary was included no matter how foreign or anomalous
it might appear to be, because we have the working hypothesis that low probabilities
can explain the poor productivity of anomalous patterns (Coleman and Pierrehumbert
1997: 54). Secondly, word-medial consonants were syllabified together with the
following vowel wherever the grammar allowed, in accordance with standard
phonological theory.
Table 13 compares the twelve most frequently applying immediately pre-
terminal rules (each rule whose left-hand side is onset, nucleus or coda).
97
Table 13. The twelve most frequently applying onset, nucleus and coda rules
Cross-ref. to
text above
Rule Times applied % of total
(28)
nucleus segment
[cons, +voc]
32,666 44.76%
(28)
onset segment
[+cons]
23,295 31.92%
(28) coda segment
[+cons]
7,797 10.68%
(36) onset segment segment
[+cont, [cont]
sonor,
+coron,
+anter]
1,643 2.25%
(33) onset segment segment
[+init] [+cons, [+cons,
voc, +voc]
del_rel]
1,252 1.72%
(38b) onset segment segment
[+cont, [+cons,
sonor, +sonor]
+anter]
995 1.36%
(40b) onset segment segment
[sonor] [+cons,
+sonor,
nasal]
885 1.21%
(45) onset segment segment segment
[+cont, [cont, [+cons,
sonor, del_rel] +sonor,
+anter] nasal]
717 098%
(38a) onset segment segment
[+init] [+cont, [+cons]
sonor,
+anter]
205 028%
(49) coda segment segment
[voc, [cont, +coron,
del_rel] +anter, del_rel]
200 027%
(48) coda segment segment
[+cons, +voc] [+cons, voc]
93 013%
(37) onset segment segment
[+init] [+cont, [cont,
sonor, del_rel]
+coron,
anter]
74 010%
Others 3,156 432%
Total 72,978 10000%
Using these results, each of the rules is paired with the figure representing its
probability of occurrence: for example, the first rule in Table 13 is assigned the figure
04476, the second is assigned the figure 03192, and so on. The overall figure used as
98
a heuristic to determine the overall likelihood of a given syllabification is the product
of the probabilities of occurrence of each relevant constituent in the phonological
word. For example, although !+%.+%! and !+%+.%! are both possible analyses, the
former will be favoured over the latter because its overall likelihood, measured in the
way just described, will be greater than that of !+%+.%!, as shown.
(51) Analysis Measure of likelihood
!+%.+%! 03192 04476 03192 04476 = 002041
!+%+.%! 03192 04476 01068 04476 = 000683
This overall measure of the likelihood of a given syllabification favours onsets over
codas principally because the probability of any given segment occurring as an onset
is greater than the probability of the same segment being a coda: for example, the
probability of a [+consonantal] segment occurring as an onset is some three times
greater than the corresponding probability for codas. This guarantees that a given
consonantal segment will syllabify with the following vowel. However, it is not
guaranteed that a given consonantal cluster will be exhaustively syllabified as an
onset, even if it is possible for that cluster to syllabify as an onset. For example, the
grammar licenses !"#&! as both an initial and a non-initial onset. The probability of the
occurrence of this cluster as an onset is 00001233 (this is not shown in Table 13). On
the other hand, the probability of the occurrence of !"! as a coda is 01068, and the
probability of the occurrence of !#&! as an onset is 002251. Thus the overall measure
for the probability of word-medial !"#&! being syllabified !".#&! is 0002405, which
means that the preferred syllabification for word-medial !"#&! will always be !".#&!.
99
In the grammar proposed, some clusters are licensed by the grammar as onsets
or codas that is, such onsets and codas have non-zero probabilities but only in
word-marginal position. It is interesting that in every case, these clusters have a
higher probability when analyzed as a coda plus an onset (which, of course, can
happen only when the clusters are word-medial). For example, the probability
associated with the rule which licenses !"#&0! as a word-initial onset (see (47)) is
00004385, and the overall measure of the probability of !".#&0! word-medially is
03192 0009825 = 0003136. This means that the distinction between word-
marginal and word-medial onsets and codas can be dispensed with in the probabilistic
grammar, because those rules which are supposedly limited to application in word-
marginal positions will always be disfavoured: there is always a more probable
alternative available. I take this as a further advantage of a probabilistic approach to
syllabification.
2.5 Extensions to the grammar
In this section, I show how four well-known phonological processes in
Russian can be accounted for within the existing syllabification grammar. This is
done by enriching the grammar with further features: no separate rule system is
needed. Furthermore, while the modifications undoubtedly increase the intricacy of
the grammar, they do not increase its formal power: the grammar remains at most
linear, and therefore strings can still be generated and recognized by it in a finite
amount of time.
The four phonological processes I have chosen to examine are the devoicing
of word-final consonants and consonant clusters, consonant-vowel interdependencies,
assimilation in voice within consonant clusters, and reduction of unstressed vowels.
100
All of these topics have been extensively discussed both in work by Slavists and by
Western generative linguists. For example, Avanesov (1956), Bulanin (1970) and
Jakobson (1978) and Halle (1959) all include some account of these topics. While the
focus of Slavists has been on attaining a high level of descriptive precision, Halle
(1959), in common with other generativists, emphasizes the systemic nature of
phonology, attempting to show how different phonological processes interrelate.
Despite the difference in methodology between the two traditions, there is a
striking amount of agreement between them as to the facts. This is largely because
Halle drew heavily upon Slavist sources.
As mentioned in section 1.1, the emphasis of this dissertation is similar to that
of Halle (1959): to show how different phonological facts cohere within a single
framework. Although Halle postulated a set of phrase-structure rules within his
analysis, Halles framework is different from the phrase-structure framework used
here in that the former explicitly relies on the ordering of rules. Furthermore, although
Halles conception of grammar is generative (Halle 1959: 24), it is not clear whether
his rules can always be used for both generation and parsing, since his phonological
theory includes a transformational component (p. 26), and, as mentioned in 1.3.2,
transformational grammars are not necessarily reversible. But even if Halles
grammar is reversible, Occams razor dictates that an entirely linear phrase-structure
phonology is preferable to one including a transformational component, since Halle
explicitly states that the transformational level opens the way to effecting changes
in the representation that are beyond the power of the Phrase Structure rules (p. 26).
In other words, it is a weakness of Halles account that part of the grammar is
101
expressed in a more powerful formalism if it can be shown (as I do here) that the
more restricted formalism of linear grammar is adequate.
The aim of this section, therefore, is not just to re-implement old ideas in a
new framework; this in itself would not represent any advance in knowledge.
However, it is of course important to show that a new framework PSG in this case
can handle the linguistic facts. Instead, the aim is to show that neither a
transformational level nor rule-ordering are needed in a formally explicit account of
the four phonological processes which were also dealt with in Halle (1959). Once
again, since the phonological processes are implemented within the existing Prolog
DCG, the grammars predictions can be tested empirically. The results show that the
linear grammar can explain exactly the same facts as Halles analysis, but in a simpler
way, and therefore the current analysis is to be preferred.
As stated above, there is broad agreement in the literature on the facts of the
four phonological processes under consideration. In each case, I state the areas of
agreement, including appropriate references to the literature. The PSG rules
developed in 2.5.2.1-2.5.2.4 are generally based on Halle (1959), as this work seems
representative in terms of the phonological statements it makes. The aim of each
section is to develop linear grammar rules which replicate the predictions of the
standard theory. The focus is on phonological rules which are reported to apply
obligatorily, rather than those (such as sonorant devoicing) which are optional.
Finally, in section 2.5.3, I describe a test to determine whether the complete grammar
models the phonological processes under discussion as expected.
102
2.5.1 Further phonological features
In section 2.3.1.2, 11 features, 9 of which were used to classify the phonemes
of Russian, were introduced. These features were used in the syllable structure
grammar outlined in 2.3.
I now introduce a further 10 features. Three of these are binary features which
are lexically specified for certain segments, as follows:
(52) [high]
[back]
[voice]
All of these are also features used in Chomsky and Halle (1968). The feature [sharp]
was proposed in Jakobson, Fant and Halle (1952) as one of the twelve binary
oppositions sufficient to distinguish the sounds of the worlds languages; many of the
examples which they give to illustrate the distinction between [+sharp] and [sharp]
come from Russian. The feature [sharped] is used with the same meaning in Halle
(1959). However, this feature is replaced here by [back] (the two features are the
converse of one another), since, as I shall show, it is unnecessary to include both
features in the phonology in order to account for the relevant phenomena.
Table 14 is a feature matrix which shows how each phoneme is specified for
all relevant lexical features. A blank means that the phoneme is neither specified +
nor for the relevant feature.
[ - b a c k ]
[ + b a c k ]
[ + b a c k ]
[ - b a c k ]
[ + b a c k ]
[ + b a c k ]
[ + h i g h ]
[ - h i g h ]
[ + h i g h ]
[ - h i g h ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k , - h i g h ]
[ + h i g h ]
[ + b a c k ]
[ - b a c k ]
[ + b a c k ]
104
The remaining 7 features are non-lexical. This is because all are initially unspecified
for all phonemes; it is the grammar that assigns the + or values. The first three of
the features are as follows:
(53) [vocal_fold_vibration] (abbreviated [vfv])
[peripheral]
[stress]
The feature [stress] is self-explanatory: it applies to syllables, [consonantal]
segments dominated by them by association, and the intermediate nodes in the
grammar (rime and nucleus). The feature [peripheral] is used to specify the
quality of vowels: a centralized vowel is [peripheral], and a less centralized vowel is
[+peripheral]. [peripheral] applies to the same constituents of the grammar as
[stress]. The feature [voice] is distinct from [vfv] in that the former is lexically
specified for segments and the latter is assigned by the grammar. Further discussion of
these three features is postponed to sections 2.5.2.2-2.5.2.4.
Finally, the remaining four features, which are also non-lexical, are complex
features. These features are as follows.
(54) [onset:A]
[rime:B]
[left:C]
[right:D]
In (54), A and B both stand for complex feature matrices (e.g. [+high] or [+back,
high]). The feature [onset:A] applies to [consonantal] segments, the constituent
rime, and the intervening constituent nucleus. Conversely, the feature [rime:B]
applies to [+consonantal] segments and the constituent onset. These features are
105
used to encode context in processes of harmony between adjacent consonants and
vowels, and will be explained in detail in section 2.5.2.1.
The features [left:C] and [right:D] are used to encode information about the
composition of the left and right edges of all pre-terminal prosodic constituents.
Again, C and D stand for complex feature matrices. For example, the left edge of a
syllable is the leftmost segment in its onset. Thus, if [left:X] is the specification of the
syllable, X will stand for the feature matrix specifying this segment. This information
is needed to enforce word-final devoicing and voicing assimilation across syllable
boundaries. Again, the use of these features will be explained more fully in sections
2.5.2.3-2.5.2.4.
2.5.2 Four phonological processes in Russian
2.5.2.1 Consonant-vowel interdependencies
There are a number of well-known constraints in Russian on consonant-vowel
sequences. The features which are relevant to this discussion are [high], [back],
[onset:X] and [rime:X]. There are two sets of facts which need to be considered.
The first is the behaviour of consonantal phonemes under the influence of
following vowels. Velar phonemes (those which are [+consonantal, +back, +high])
are unpalatalized before [+back] vowels and palatalized before the [back] vowels !,!
and !1!
15
. The [+consonantal] segments which are [back] are always palatalized,
regardless of what vowel follows them. The [+consonantal, +back, high] phonemes

15
Avanesov (1956: 103); Halle (1959: 67; Rule P 6d).
106
are always unpalatalized, except that those other than !5=<8=<&#!
16
are palatalized before
the [back, high] vowel !1!
17
. There are very few other exceptions to this rule
18
.
The second set of facts is the behaviour of vowel phonemes under the
influence of preceding consonants. The facts here are simpler, because in stressed
positions, only the [back, +high] vowel !,! behaves unusually: it is pronounced as the
apparently back vowel (,) (in both stressed and unstressed positions) when preceded
by a [+back, high] consonant
19
, but as the front vowel (,) otherwise. The other
vowels !%!, !1!, !6! and !;! are always pronounced as (%), (1), (6) and (;) respectively
when stressed; thus, stressed [back] !1! is always pronounced as a front vowel, and
stressed [+back] !%!, !6! and !;! are always pronounced as back vowels. In unstressed
positions, the [+back, high] vowels !%! and !6! are pronounced as an apparently front
(D) when they follow a [back] consonant
20
, but as back (D) in all other instances.
Unstressed !1! and !;! (which are [back, high] and [+back, +high] respectively) are
always pronounced as front !D! and back !;! respectively, no matter what consonant
precedes them.
Traditional analyses of Russian phonology (e.g. Avanesov 1956) have treated
these facts as allophonic alternations, as shown in Table 15. In the table, each tick
mark shows a possible allophonic combination ([+back] versus [back]), and in each

16
Avanesov (1956: 103); Halle (1959: 66; Rule P 5b).
17
Halle (1959: 63; Rule P 1a). Avanesov (1956: 103) regards sequences of such consonants and !1! as
impossible, which implies that he considers the consonant in e.g. !&'1! to be underlyingly [+palatalized].
18
As Halle notes, the exceptions are generally confined to recent loan-words in Russian, e.g. ."+4
(&1+9) tempo.
19
Avanesov (1956: 103); Halle (1959: 72; Rule P 12). Halle comments that this rule obviates the
necessity for postulating (,) as a separate phonological entity distinct from {*i}.
20
Avanesov (1956: 110, 118); Halle (1959: 70; Rule P 9a).
107
case the principal or default allophone is shaded. I gloss over the difference between
the treatment of stressed and unstressed !%=6!.
Table 15. Allophonic relationships in consonant-vowel sequences
!1! !;! !%=<6! !,!
|back
high]
[+back
+high]
[+back
high]
|back
+high]
Vowels
Consonants
|back]
allophone
[+back]
allophone
[+back]
allophone
|back]
allophone
|back]
allophone
[+back]
allophone
!5=8=&#!
[+back]
allophone

[back,
+cons]
phonemes
[back]
allophone

[+back]
allophone

[+back,
high,
+cons]
phonemes
except
!5=8=&#!
[back]
allophone

[+back]
allophone

[+back,
+high,
+cons]
phonemes
[back]
allophone

Traditional generative phonology, on the other hand, describes these
relationships in terms of ordered and feature-changing rules. For example, Halle
(1959: 70) contains the following rule, which clearly involves a [diffuse] segment
becoming [+diffuse]:
Rule P 7g. In position after {c} {} and {} unaccented noncompact nondiffuse
vowels become diffuse and of high tonality.
The challenge here is to encode these facts in the existing grammar, so that (1)
there is no feature-changing or rule-ordering, and (2) additional complication is
minimized. I shall now describe how this is done.
108
First of all, the domain of any phonological rules to describe these phenomena
must be the syllable. This is because the facts described all hold for CV sequences in
Russian, and these in turn must always be tautosyllabic if the Maximal Onset
Principle holds for Russian. Thus all the relevant constraints can be stated as syllable
structure constraints. The constraints appear to involve a kind of harmony between
consonants and following vowels, where [back] consonants tend to prefer to be
adjacent to [back] vowels, and [+back] consonants prefer to be adjacent to [+back]
vowels. One might suppose, therefore, that (55) would account for the facts.
(55) syllable onset rime
[o back] [o back]
However, this is not the case. (55) would rule out syllables such as !&#1!, where !&#!
does not have a [back] allophone and !1! does not have a [+back] allophone. It is not
possible to resolve this problem by underspecifying one or other of !&#! and !1!: !1!
must be underlyingly [back], because it elicits the non-principal [back] allophone
of underlyingly [+back] preceding consonants such as !$!, and !&#! must be
underlyingly [+back], because it elicits the non-principal [+back] allophone of
underlyingly [back] following !,!. We know that !$! is underlyingly [+back], because
it elicits the [+back] allophone of following !,!; we know that !,! is underlyingly
[back] because it elicits the [back] allophone of a preceding velar; and we know
that the velars are lexically [+back] because they elicit the [+back] (principal)
allophone of following (unstressed) !%! and !6!.
The fact that some segments appear to be more susceptible than others to
conditioning by neighbouring segments suggests a precedence hierarchy, where
109
phonemes higher in the precedence hierarchy are more assertive than those lower
down: the more assertive phonemes will show a greater tendency to affect
neighbouring phonemes, and a lesser tendency to be affected by them. Figure 4
(which is not a subsumption lattice) is a graphical representation of these
relationships.
110
Figure 4. Lattice showing the hierarchy of Russian phoneme classes
with respect to sequence structure constraints
Notes to Figure 4
D
stands for the highest-precedence category
E
stands for the lowest-precedence category
Key to classes in Figure 4
No. Features (where applicable) Segment(s) included
1. [back, high, consonantal] !1!
2. !5=8=&#!
3. [back, +consonantal] !7=0'=-'=+'=4'=#'="'=*'=?'=A'=$'=9'=&'=&8'!
4. [+back, +high, consonantal] !;!
5. !0=-=+=4=#=8="=*=5=?=A=$=9=&=&#!
6. [back, +high, consonantal] !,!
7. [high, +back, consonantal] !%=6!
8. [+back, +high, +consonantal] !2=B=C!
4. [+back, +high,
consonantal]
8. [+back, +high,
+consonantal]
1. [back,
high, cons]
6. [back, +high,
consonantal]
7. [high, +back,
consonantal]
3. [back,
+consonantal]
D
2.
5.
E
111
For example, the fact that in the lattice [back, high, consonantal] dominates
[+back, +high, +consonantal] (that is, a path from the former can be traced directly
down to the latter) correlates to the fact that the non-principal [back] allophones of
the velar consonants are elicited by a following !1!. In contrast, the node labelled
[back, +consonantal] neither dominates, nor is dominated by, the node labelled
[+back, +high, consonantal]. This correlates to the fact that [back] consonants do
not elicit the non-principal allophone of following !;!, and neither does !;! elicit the
non-principal allophones of [back] consonants.
These relationships can be represented in linear grammar by using a rule like
(56), where the onset and the rime have the features [rime:Y] and [onset:X]
respectively. These features are a means of encoding the context of the syllables
constituents: for example, the specification of a rime as onset:[+high, back] means
that the rime follows a [+high, back] onset. The general rule, then, will be as
follows:
(56) syllable onset rime
[rime:[ back, o high], [onset:[o high, | back],
o high, back,
| back] o high]
The values of o and | are lexically specified for the onset segment, and and o are
lexically specified for the nucleus segment. When the values are filled in for the
features [onset:X] and [rime:Y], what is obtained is effectively the featural
specification of an allophone. The vowels which dominate a given consonant in
Figure 4 are specified as the value of Y for the non-principal allophone of that
consonant, and (likewise) the consonants which dominate a given vowel are specified
112
as the value of Y for the non-principal allophone of that vowel. Consider for example
the sequence !A1!, which is realized (A'1). The fact that the non-principal allophone
(A') of !A! is elicited only by !1! is represented in Figure 4 by the fact that the node
where !1! is located, [back, high], dominates the node where !A! is located (node 5).
Where !A! is followed by !1!, it therefore ends up with the specification in (57), which
is a kind of phonotactic subcategorization similar to that used in Coleman and
Pierrehumbert (1997).
(57) [+consonantal,
vocalic,
sonorant,
continuant,
coronal,
anterior,
rime:[back, high]]
This representation maps to a sound (or range of sounds) which would traditionally be
represented by the symbol (A'). It is important, however, to appreciate that the
representation in (57) and (A') are not equivalent to one another, as we shall see
below.
Similarly, where !,! is pronounced as (,), it would be represented as shown in
(58):
(58) [consonantal,
+vocalic,
+sonorant,
back,
+high,
onset:[high]]
113
Whether the onset preceding !,! is [+back] or [back] does not need to be specified in
(58), since the redundancy relationship [high, +consonantal] [back] holds.
The cases where the principal allophone (,)<surfaces are simply the
complement of those where the non-principal allophone surfaces, since a binary
opposition is involved. Thus the representation for the [back] allophone of !,! will be
as in (59):
(59) [consonantal,
+vocalic,
+sonorant,
back,
+high,
onset:[+high]]
Now let us consider the representation of a phoneme which is unaffected by
neighbouring segments, taking !;! as an example. Here, the value of X in onset:X is []
(the empty feature which stands for all segments), meaning that (;) can follow any
onset, as shown in (60).
(60) [consonantal,
+vocalic,
+sonorant,
+back,
+high,
onset:[]]
The implications of this analysis are as follows. First, where traditional
frameworks have postulated allophones versus phonemes, the current framework
treats allophones and phonemes as essentially the same objects. Secondly, instead
of feature-changing, rule-ordering, or context-sensitive rules, we have accounted for
114
the facts using a linear grammar. It may seem that by specifying a context in the rule,
the power of the grammar has been increased to context-sensitivity. However, this is
not the case: each feature-enriched rule in the grammar can still be rewritten in terms
of a number of linear rules without feature enrichment. It must be acknowledged that
the addition of extra features to ensure that the grammar remains context-free is a
good example of the kind of loss of generality which can result when more powerful
grammar formalisms are avoided (see the discussion on page 21 above). Nevertheless,
the fact that the grammar is no more powerful than context-free is a guarantee of its
ability to parse and generate strings within a finite amount of time, and this is a major
computational advantage, as well as a psychological advantage, for all the reasons
outlined in 1.3.2.
What has been proposed, therefore, is a grammar which can account for the
consonant-vowel interdependencies of Russian to yield the same results as traditional
analyses. The results of testing this proposal will be discussed in 2.5.3.
2.5.2.2 Reduction of unstressed vowels
There is little controversy that Russian unstressed vowels differ from their
stressed counterparts in two ways: first, their duration is considerably less, and
secondly, they are reduced or centralized (Avanesov 1956: 105-106, Bulanin 1970:
102, Bondarko 1977: 110-111). The Russian vowel system is a five-term system, as in
Figure 5.
115
Figure 5. The Russian vowel system
However, in unstressed positions the distinctions between !%! and !6! on the one hand
and !,! and !1! on the other collapse (Avanesov 1956: 106-125): the former are
usually represented as (H) or (3)
21
, while the latter are usually represented as (D)
22
.
Figure 6 demonstrates these neutralizations.

21
Halle (1959: 73; Rule P 14a). In pretonic and in absolute word initial position unaccented (%)
becomes more lax. In all other positions (%) is, moreover, heavily reduced, i.e. (3).
22
Halle (1959: 70; Rule P 9a).
,
1
%
6
;
116
Figure 6. The Russian vowel system in unstressed positions
The following examples (from Gvozdev 1949: 13) demonstrate these neutralizations:
(61) !"#$%& I shall illuminate !"#$%!&'&%()*! +,#$%"&'&%()*-
!"#'%& I shall boo !"#$%#&'&%()*! +,#$%"&'&%()*-
()"*) she grazed !.$#/()0! +.%#/()0-
(!"*) ambassador (gen. sg.) !.&#/()0! +.%#/()0-
In the literature it is generally agreed that the degree of centralization (or conversely,
peripherality) of vowels is greater in some positions than in others. This principle is
relevant to the choice between +,- and +1-. Effectively, the factors which govern this
choice are (a) whether !0! or !"! is word-initial, and (b) whether !0! or !"! are in the
pretonic syllable (the syllable immediately preceding the stressed syllable) (Avanesov
1956: 106, 115, 120; Bulanin 1970: 103; Bondarko 1977: 112). Table 16 shows that if
!0! or !"! are either word-initial or pretonic, they are pronounced more peripherally
than +1-, i.e. as +,-.
(2
1
,(
3
4
5
0
"
*
117
Table 16. Allophones of /a/ and /o/
Pretonic syllable Not pretonic syllable
Word-initial syllable
(H) (H)
Not word-initial syllable
(H) (3)
The rules outlined in 2.5.2.1 have a bearing on the rules in this section: as
stated in 2.5.2.1, if !%! and !6! are unstressed and follow a [back] consonant, they are
pronounced as front vowels, or (D) (Avanesov 1956: 110, 118). The way in which
Halle (1959) deals with this is to stipulate two ordered rules: the first (P 9a) causes !%!
and !6! to become !,! after a [back] consonant, and the second (P 14b) which states
that unaccented diffuse vowels are reduced and lax. In other words, in Halles
system the phonetic representation (6#"'D8&8';) for ,-(>18 !6#"'%8&8';! I shall
illuminate, which is homophonous with the first two examples in (61), would be
derived as in (62):
(62) Phonological (underlying) representation !6#"'%8&8';!
Rule 9a [accent, [+diffuse, / [+sharp] ___ "',8&8'
+compact] +high tonality]
Rule 14b [accent, [+reduced, "'D8&8'
+diffuse] +lax]
Phonetic representation (after operation of other rules) (H#'"'D8'8';)
The challenge in declarative grammar is to model this phonological process
(i.e. combine vowel reduction and consonant-vowel interdependencies) without the
rule-ordering which (62) relies upon. The problem can be solved by using the features
118
[stress] and [peripheral] for all vowels. These features are shared by nucleus and
syllable nodes dominating the relevant segment. The latter feature, like [onset:X] and
[rime:Y], is used as a means of encoding the context of a syllable: a syllable is
[+peripheral] if it immediately precedes the [+stress] syllable. Given the facts
presented in Table 16, [peripheral] is redundant except in cases where !%=<6! follow a
[+back] consonant.
According to the traditional accounts, therefore, it is necessary to know all of
the following to determine the quality of a vowel: whether it is stressed, and if not
whether it is pretonic or word-initial; and what kind of consonant it follows. This
implies that all three features [onset:X], [stress], and [peripheral], will be needed at
some point in the fully-specified representations for vowels. For example, the
representation for a pretonic !%! preceded by a [back] consonant would be as in (63).
In this case, because the preceding onset is [back], the feature [peripheral] is
redundant.
(63) [consonantal,
+vocalic,
+sonorant,
+back,
high,
stress,
onset:[back]]
The partial syllable structure associated with this segment (excluding features not
directly relevant to this discussion) would be as in Figure 7:
119
Figure 7. Partial syllabic structure of pretonic /a/ after a [back] consonant
Again, the vowel in Figure 7 (which is also represented by the feature matrix in (63))
maps to the sound denoted by (D) in traditional Russian phonology textbooks such as
Avanesov (1956). Thus, one can think of (D) as an alias for the representation in
Figure 7. The notation (D) is unimportant in itself; it is simply a convenient way of
representing a phonetic event
23
. Now, where phonological neutralization occurs, it
must be the case that two or more distinct fully-specified feature matrices map to the
same phonetic event, and correspondingly, two or more feature matrices must map to
the single alias denoting that phonetic event. For example, if pretonic !1! and !%! are
not distinguished when they follow a [back] consonant, then this must be expressed

23
More precisely, (D) and other such aliases should probably be thought of as standing for a range of
phonetic events, owing to facultative variability.
!%!
onset:[back]
[stress]
nucleus
stress, +peripheral
onset:[back]
[high, front]
rime
stress, +peripheral
onset:[back]
[high, front]
onset
[back]
syllable
[stress,
+peripheral]
120
by the mapping of the feature matrices for both segments to (D). The alias for each
fully-specified feature matrix must be specified lexically (as in Table 14), not by
context-free rules, since two distinct constituents cannot be merged into a non-distinct
constituent without feature-changing or destructive processes. Thus, the use of aliases
in this way is an ingenious method of analyzing neutralization without feature-
changing. One must acknowledge, however, that the cost of this method is that a new
type of phonological object is created.
Clearly, the rules of vowel reduction interface directly with those of stress
assignment. In section 5.2, we shall see how the grammar supplies the features
[stress] and [peripheral], based only on the segmental content of a word.
In this section, then, we have seen how the judicious use of the features
[stress] and [peripheral] can be used in conjunction with the consonant-vowel
interdependency rules from the previous section to model reported phonetic facts
about Russian. This is also done in such a way as to avoid rule-ordering and feature-
changing, and ensuring that the grammar remains no more powerful than linear.
2.5.2.3 Word-final devoicing
Russian, like German and Polish, undergoes a phonological process whereby
word-final obstruents and obstruent clusters are phonetically devoiced (Avanesov
1956: 163-164, Halle 1959: 64). For example, the devoicing of word-final !$! in (64)
leads to a situation of homophony between %,. cat and %,# code.
(64) (a) %,.& cat (gen. sg.) (26t3)< !26&%!
%,. cat (nom. sg.) (26t) !26&!
(b) %,#& code (gen. sg.) (26d3)< !26$%!
%,# code (nom. sg.) (26t)< !26$!
121
An example of the application of the word-final devoicing rule to an obstruent cluster
can be seen in (65).
(65) -C"9# conference (#71st)< !#71*$!
-C"-. he will eat (#71st) !#71#&!
The rule does not apply, however, to sonorants (nasals, liquids, and glides; that
is, all non-obstruent consonants) or clusters of sonorants in Russian: in all instances in
(66), for example, (-) and (4) (where applicable) are pronounced with vocal fold
vibration.
(66) 9&6& hall (gen. sg.) (*%l3) !*%-%!
9&6 hall (nom. sg.) (*%l) !*%-!
(,6'& wave (nom. sg.) ("Hln%) !"6-4%!
(,6' waves (gen. pl.) ("6ln) !"6-4!
However, in words containing a final consonant cluster consisting of
obstruent(s) and a final sonorant, the obstruent(s) are reported to retain their
underlying voicedness (a term which I shall explain further below), while the sonorant
may assimilate (but not obligatorily) in voice to the obstruent to its left (Gvozdev
1949: 37). These assertions have been corroborated experimentally by Barry (1989).
(Barry also suggests that the obstruent may completely, or more likely partially,
devoice.) Some examples are shown in (67):
(67) %&#! sequence (2%dr) !2%$0!
6).! litre (-',tr ) !-',&0!
!).+ rhythm (0',tm) !0',&+!
122
In a word-final cluster including sonorant(s) where the final segment is an
obstruent, the sonorant(s) may assimilate in voice to the (devoiced) obstruent
(Avanesov 1956: 169-170). Examples are given in (68):
(68) 5,!# proud (B6rt) !B60$!
4,!. port (96rt) !960&!
Reformatskij (1975: 129-133) suggests that !"! devoices when word-final in
exactly the same way as an obstruent. This is uncontroversial, but according to
Reformatskij, the behaviour of word-final !"! is unlike that of other obstruents in that
the preceding obstruent in the same cluster retains its underlying voicedness or
voicelessness (cf (67) above). For example:
(69) .!"9( sober (predicative) (&0'1zI) !&0'1*"!
.!"9(/0 sober (attributive) (&0'1zvD7) !&0'1*",7!
+2!.( dead (predicative) (+'6rtI) !+'60&"!
+2!.(/0 dead (attributive) (+'6rtvD7) !+'60&",7!
Reformatskijs claim is cited (and perpetuated) in a number of subsequent
works which attempt to accommodate it within a coherent phonological system
(Jakobson 1978, Halle and Vergnaud 1981: 16, Berendsen 1983, Hayes 1984: 319).
Yet the claim is not corroborated by Barrys (1989) experimental study. This study
suggests that the obstruent preceding word-final !"! obligatorily devoices, i.e. (&0'1sI).
If this is true, !"! behaves in every respect like an obstruent when word-final, as can
be seen by comparison with the forms in (70).
123
(70) +,95 brain (nom. sg.) (+6sk) !+6*B!
+,95& brain (gen. sg.) (+'6zq3) !+6*B%!
Some phonologists (e.g. Lightner 1972: 331-334) have argued for a parallel
between word-final devoicing and voicing assimilation (discussed in the next section).
The suggestion is that word-final devoicing is really assimilation to the voicelessness
of the following word-boundary. As Houlihan (1979) points out, this type of analysis
presupposes that phonological boundaries are intrinsically specified for phonological
features. Houlihan argues that the reduction of voicing assimilation and word-final
devoicing to a single phonological process obscures a key distinction between
assimilatory and non-assimilatory phonological rules: the former may apply across
phonological boundaries, but the more major the boundary, the less likely they are to
apply, whereas exactly the opposite is true for the latter. The approach followed here
is to separate word-final devoicing and voicing assimilation: phonological boundaries
are represented in the current framework as divisions between phrase-structure
constituents rather than as constituents in their own right. Boundaries and constituents
are thus fundamentally different, and the idea of assimilation of a constituent to a
boundary is nonsensical because boundaries, unlike constituents, cannot be specified
for features.
Let us now summarize what appear to be the generally accepted facts about
word-final devoicing in Russian:
124
(71) (a) Sonorants are always voiced
24
.
(b) Obstruents are always unvoiced in word-final clusters of which the last
phoneme is an obstruent. Obstruent here includes !"!.
(c) Obstruents retain their underlying specification for voice in word-final
clusters of which the last phoneme is a sonorant.
It is comparatively easy to encode these facts in the existing grammar by recourse to
the feature [right:X]. Effectively, word-final devoicing means that the right edge of a
phonological word must be specified [vfv]. But since strongfoot is always the
rightmost constituent of prosodic_word (see (35) on page 80), it is sufficient just to
state that the rightmost constituent of every strongfoot is [vocal_fold_vibration], as
follows:
(72) strongfoot syllable
[+stress, right:[vfv]]
strongfoot syllable weakfoot
[+stress] [right:[vfv]]
Each constituent dominated by strongfoot needs then to inherit and pass on the
relevant specification to its subconstituents: some examples are given in (73).

24
Since the evidence about devoicing of sonorants suggests that devoicing is only facultative, I take it
that devoicing is governed by a rule whose application is only optional. As such, I consider the
inclusion of such a rule in the present grammar not to be a high priority.
125
(73) weakfoot syllable
[right:X] [stress, right:X]
weakfoot syllable weakfoot
[right:X] [stress] [right:X]
syllable onset rime
[right:X] [right:X]
rime nucleus coda
[right:X] [right:X]
Clearly, the feature [vfv], as used in (72), is not the same as the feature [voice]: for
example, !*! is [+voice], but could be [vfv] depending on the context. [voice] is a
classificatory property of phonemes, whereas [vfv] denotes whether the vocal folds
actually vibrate in production of the segment. We can now be more precise about the
meaning of underlying voicedness. In cases where a segment retains its underlying
voicedness (as in (71c), for example), the segment is specified as [o voice, o vfv]. In
other words, in such cases there is agreement between the features [voice] and
[vfv]. To give a concrete example, !$! and !&! are [+voice] and [voice] respectively;
in !2%$0! and !-',&0! they are also [+vfv] and [vfv] respectively, but in !26$! and !26&!
they are both [vfv]. Unless one allows feature-changing, a move which would
compromise the declarativeness of the grammar, there is no alternative but to
distinguish [voice] and [vfv] (although the actual names of these features are of
course arbitrary as far as the grammar is concerned).
The rules in (72) will assign the feature [vfv] to any consonantal segment at
the right edge of a word, including sonorants. However, the inertness of sonorants
(according to the standard theory) with respect to word-final devoicing (and voicing
assimilation) is expressed in the fact that the aliases for sonorants are the same
126
whether they are [+vfv] or [vfv]. In other words, no matter what specification for
[vfv] is assigned to a sonorant, the sonorant is unaffected.
We have not yet considered the voicing constraints in obstruent clusters,
including those of the type mentioned in (71c). These will be considered in section
2.5.2.4, since it is more convenient to think of the devoicing of a word-final cluster as
the interaction of final devoicing and voicing assimilation. This approach allows the
correct distinction to be made between the examples given in (65) and (67), where in
the former case obstruents devoice before other obstruents, but retain their
underlying voice before sonorants in the latter case.
The analysis in this section has followed the claims of textbooks, namely that
word-final devoicing results in neutralization: [+voiced] obstruents become [voiced]
word-finally, merging with their counterparts which are lexically [voiced]. As
mentioned in section 2.5.2.2, neutralization is expressed here in the fact that two
feature matrices map to the same alias. However, the idea of word-final devoicing
resulting in neutralization has been called into question by both Slowiaczek and
Dinnsen (1985) and Burton and Robblee (1997). These studies, which are based on
acoustic measurements, suggest that word-final voicing neutralization of the same
kind in Polish is not complete. Nevertheless, as Blumstein (1991: 116) points out,
such evidence at least confirms a number of critical assumptions made by
phonological theory and importantly confirms the reality of a process such as
neutralization (op. cit., p. 116). In other words, it certainly is the case that word-final
voiced obstruents are articulated with less vocal fold vibration than the same
obstruents in pre-sonorant position, and the standard theory, that word-final devoicing
involves neutralization, is at the very least a good first phonological approximation.
127
And even if it were no more than this, it is only the system of aliases mentioned
above which would need to be revised: the substance of the grammar would be
unaffected.
In this section, then, we have considered the voicing constraints on rightmost
constituents in Russian. I have shown that the general rule is quite simple: all
segments in the rightmost coda of a word are [vfv]. The feature [voice] is not
referred to in the word-final devoicing rules: reference to [vfv] is sufficient to make
the correct predictions. In the next section, we shall see instances where [voice] does
need to be referred to in the phonological rules, since there is no way of accounting
for voicing assimilation without access to information about the lexical voicedness
of segments.
2.5.2.4 Voicing assimilation
In autosegmental theory (Goldsmith 1976), assimilation is seen as the sharing
of a feature among adjacent segments. In this section, I present an analysis of Russian
voicing assimilation which is in line with this view. In the terms of the current
analysis, adjacent segments which undergo voicing assimilation (and this does not
include all consonants, let alone segments) are required to agree in the feature
[vocal_fold_vibration], or [vfv].
Although it represents just one small corner of the phonology of Russian,
voicing assimilation has attracted a disproportionate amount of attention in
phonological literature. Much of this literature stems from the discussions in Jakobson
(1956 and 1978) and Halle (1959). Jakobsons (1978) article was controversial
because some of the facts he reported appear to be characteristic of a non-standard
dialect of Russian, although his description purports to be of standard Russian. Halles
128
treatment of voicing assimilation, on the other hand, was controversial in its claim
that an intermediate classical phonemic level of representation was unnecessary.
Both of these controversies will be discussed in detail below.
The main claim of Halles which I shall take issue with, however, has been
less frequently discussed in the literature. This is the claim that the grammar need be
more powerful than linear (as mentioned on page 100, Halle refers to a
transformational component in the phonology). Since a grammars power is a
property which can be objectively established, this claim of Halles can easily be
refuted if it can be shown that the facts which he accounts for can also be accounted
for by a less powerful grammar. The rules which I suggest can account for voicing
assimilation (like all other rules in this chapter) are linear rules, and I would claim that
nothing more powerful is needed to account for the facts observed in Russian
phonology.
Generally, existing accounts are inexplicit about the domain of voicing
assimilation in Russian, assuming that it occurs in clusters (in a manner reminiscent
of the way in which Chomsky and Halle 1968 deal with English stress). This implies
that voicing assimilation is unaffected by syllable boundaries (cf Halle 1959);
however, overall the evidence seems to be conflicting. Some experimental evidence
bearing on the subject is available (for example, see Bondarko 1977 for
spectrographic analysis, and Burton and Robblee 1997 for evidence that Russian
voicing assimilation does occur across prepositional boundaries). Burton and
Robblees evidence also appears to show indirectly that assimilation appears to take
place across syllable boundaries (e.g. in the cluster !$*!), while Halle (1959) and
Avanesov (1956) describe a situation in which the voicing of sonorants appears to
129
differ (albeit facultatively) depending on the location of syllable boundaries; for more
details, see below. What is clearly needed is a systematic study of the effects of
phonological boundaries upon Russian voicing assimilation. Meanwhile, the
phonological rules I propose here are tentatively based on the assumption that the
domain of Russian voicing assimilation is the phonological word. It would not be
surprising if future experimental work pointed to flaws in this assumption, given the
existing evidence in linguistics that syllable boundaries have an effect on the
operation of similar phonological rules: see for example Kenstowicz (1994: 251).
However, it is possible that Russian voicing assimilation is just not as constrained as
other assimilation rules in the domain of its operation (cf. Houlihans 1979: 21
suggested constraint on assimilatory phonological rules: assimilatory phonological
rules originate within the weakest possible phonological boundaries and later spread
to apply across progressively stronger phonological boundaries).
The rules of Russian voicing assimilation are as follows. The behaviour of
segments with regard to voicing assimilation depends on whether they are obstruents
or sonorants (Halle 1959: 63), as is the case in word-final devoicing (see 2.5.2.3).
Halle expresses the rule of voicing assimilation as follows:
A single obstruent or several consecutive obstruents occurring in sequence regardless
of intervening preposition or word boundaries will be termed an obstruent cluster
If an obstruent cluster is followed by a ! (dash) boundary or by a sonorant, then with
regard to voicing the cluster conforms to the last segment; if it is voiced, so are all
other segments in the cluster; if it is voiceless, so is the entire cluster. (p. 64)
Avanesov (1956: 163-166) gives essentially the same account. Halle explicitly
states that this rule applies across morpheme boundaries (op. cit., p. 61), which, as
130
mentioned earlier, is corroborated in part by Burton and Robblees (1997) findings.
Examples illustrating the rule of voicing assimilation are given in (74).
(74) (a) 48-% setting in motion (9;#2)< !9;#2!
-48-% flush (sp;#2) !#9;#2!
(b) #&.$ to give ($%&')< !$%&'!
-#&.$ to give away (zd%&') < !#$%&'!
Clusters composed entirely of sonorants (including !"!, when it functions as a
sonorant; see p. 82 above) are voiced. The examples in (75) illustrate this:
(75) (6&-.$ power (vl%#'&')< !"-%#&'!
'!&().$-> to please (nr%"'D&#3) !40%"',&'#'%!
(,6'& wave ("Hln%)< !"6-4%!
In fact, the voicing of sonorants is more or less a general rule (Halle 1959: 44).
According to Halle (1959: 44) and Avanesov (1956: 169-170), the only cases where
sonorants are (optionally) unvoiced (other than those already mentioned in 2.5.2.3)
involve instances where the sonorant precedes [vfv] obstruents in a cluster, as in the
examples in (76):
(76) !.8.$ mercury !0&;&'! (0L &;&')
,%.><!$-%,0 pertaining to October !62&'%A0'#267! (62&'%90L '#267)
If syllabification has no effect on voicing assimilation, then it should logically
follow that !0! in 4,!.& !960&%! port (gen. sg.) may also be optionally devoiced,
despite the fact that the syllable boundary in !960&%! falls clearly between !0! and !&!
(unlike in !0&;&'!). To my knowledge, this has not been discussed in the literature.
131
Undoubtedly, this is a case where further experimental evidence is called for.
Evidence of this kind would show whether syllabification does have an effect on
Russian voicing assimilation. On the other hand, Halle and Avanesov do not claim
that the cases in (76) are instances where the devoicing of the sonorants is obligatory,
so I shall proceed on the assumption that sonorants are voiced by default (i.e. their
voicing is unaffected by Russian voicing assimilation), and that the option of
devoicing is governed by some minor rule which, following Halle (1959), I shall not
describe in detail.
Once again, the behaviour of !"! is idiosyncratic (see 2.3.2 and 2.5.2.3 above).
Followed by a sonorant, !"! behaves as a sonorant (Avanesov 1956: 162, Halle 1959:
63, Coats and Harshenin 1971, Jakobson 1978, Halle and Vergnaud 1981: 17); in
other words, it neither triggers nor undergoes voicing assimilation. As noted above,
there are a number of reasons to substantiate the idea that !"! is phonologically a glide
in such circumstances (Andersen 1969, Coats and Harshenin 1971). An example is
given below:
(77) F("!$ Tver (place name) (tv'M0') !&"'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) (-.!"3& meeting (Istr'1&8'3)< !"#&0'1&8'%!
(#!85 suddenly (vdr;2) !"$0;B!
4,(.,!).$ to repeat (93ItH0',&')< !96"&60',&'!
(b) % (#,(" to the widow (qvdH"'1)< !2"$6"'1!
)9 (:,#& out of the entrance [DsIx6$3)< !,*"C6$%!
As we saw on page 82 above, the specification of !"! for the feature
[sonorant] is determined in the current theory by phonotactics. As a result, modelling
the behaviour of !"! with regard to voicing assimilation presents no problem in
principle; effectively, the rule !"! is an obstruent when followed by an obstruent and
a sonorant when followed by an obstruent has already been incorporated into the
grammar.
Now, there has been a long-standing debate about the issue of so-called
sonorant voicing transparency in Russian. This debate originated with Jakobson
(1978), which claimed (based on his own auditory judgements) that while sonorants
do not trigger or undergo voicing assimilation, obstruents in CLC clusters (where C
represents obstruents and L represents a sonorant) assimilate to one another across
the intervening sonorant: for example, )9 G7"'-%& !,z mts14#2%! (-#+&#-) out of
Mtsensk. Jakobsons claims are controversial for a number of reasons. The first of
these is that they are at odds with most other analyses of Russian, including
Jakobsons own earlier work (Jakobson 1956), Andersen (1969) and Shapiro (1993).
The second is that they imply that Russian voicing assimilation is unusual as a
phonological process, in that it occurs between non-adjacent segments (although, as
Berendsen 1983 points out, there is a similarity between the process which Jakobson
describes and that of vowel harmony in Turkish [Halle and Vergnaud 1981], and such
processes do not present a problem for non-linear phonology). An experimental study,
133
Robblee and Burton (1997), which attempts to resolve discrepancies in the various
descriptions of Russian phonological voice (p. 407), is rather inconclusive, although
Robblee and Burton conclude by speculating that if voicing assimilation occurs across
sonorants, then this may be with the sonorant as an active participant in voicing
assimilation (p. 428). In other words, obstruents assimilate in voice across a sonorant
only when the sonorant itself assimilates in voice to the following obstruent.
Although Jakobsons article has been, and continues to be influential in
mainstream phonology (cf. Hayes 1984: 318, Kenstowicz 1994: 70), it appears on
balance that his claims are to be taken with some caution. Given the fact that his
observations have not generally been corroborated by other studies, it seems that
Shapiros (1993) assessment that they may be true for Jakobsons own idiolect, but
not for contemporary standard Russian, is well-founded. If this is so, the facts about
standard Russian can be summarized as follows:
(79) (a) !"! and !"'! are sonorants if followed by a sonorant, and obstruents if
followed by an obstruent.
(b) Sonorants are voiced no matter what precedes or follows them.
(c) Obstruents assimilate in voice to following obstruents. Before sonorants,
voiced and voiceless obstruents are distinguished.
These facts present us with a twofold challenge. The first is to express the
constraints monotonically in linear grammar; the second is to express the voicing
assimilation constraints in terms of syllable structure.
I start by considering the implementation of voicing assimilation within
onsets. Instead of positing two levels of representation, it is necessary to distinguish
between underlying voice [voice] and surface voice [vfv], otherwise it is not
134
possible to maintain monotonicity in the grammar. The distinction between
underlying and surface voice may be reminiscent of the distinction considered (but
rejected) in Chomsky and Halle (1968: 169) between phonological and phonetic
features:
It might be proposed, in the light of the distinction between phonological and
phonetic distinctive features, that the two sets be absolutely unrelated that the
rows be labeled entirely differently in the phonological and phonetic matrices Only
the phonetic features would now be substantive; the phonological features would be
without physical content and would provide an arbitrary categorization.
since all phonological rules would operate on the empty categories A, B, etc.,
gradually filling them in and revising their entries, the grammar would now have to
be supplemented with a set of rules operating at the point at which all matrices are
fully specified and providing that phonetic features be associated with the categories;
for example we would have rules providing that [o A] [o vocalic], [o B] [o
consonantal] But every grammar will have to have exactly these rules; hence they
do not contribute in any way to the choice among grammars and can just as well be
eliminated from all grammars.
However, there is a crucial difference between my distinction and Chomsky
and Halles. Under Chomsky and Halles suggestion, [voice] would at some stage be
rewritten as [vfv], if [voice] were a phonological feature and [vfv] a phonetic one.
In the grammar which I propose, [voice] does not change into [vfv]. Neither
category is empty, and it would not be true to say that [vfv] and [voice] are
absolutely unrelated. Instead, [vfv] is unspecified in the lexical entry of the
segments of Russian. It is perfectly legitimate to view both features as simultaneously
phonological and phonetic, since both features are present at all stages in the
derivation. In view of this, Chomsky and Halles arguments against distinguishing
phonological and phonetic features do not carry over to the distinction between [vfv]
and [voice].
Now we consider how voicing assimilation is enforced in the grammar. This is
done by unification of feature structures (see Shieber 1986 for further discussion, and
135
Boyer 1988 for details of how feature structure unification can be performed
computationally). To each existing rule, a voicing assimilation constraint is added, as
shown in (80). (The voicing assimilation constraint is enclosed in braces, and X and Y
stand for any feature structures licensed under the existing onset rules.)
(80) onset segment segment
X Y
{voicing_assim(Y,X)}
The addition of unification-based constraints in this way does not increase the power
of the grammar, because each rule can still be compiled out into linear rules without
feature structures.
There are three voicing assimilation constraints, depending on whether X and
Y are [+sonorant] or [sonorant]. These are as follows:
(81) (a) voicing_assim(Y,X):-
Y [+sonorant],
X [+sonorant, +vfv].
(b) voicing_assim(Y,X):-
Y [+sonorant],
X [sonorant, o voice, o vfv].
(c) voicing_assim(Y,X):-
Y [sonorant, o vfv],
X [o vfv].
(where P Q denotes P unifies with Q)
This meaning of these constraints is as follows. (81a) states that if Y (the segment to
the right) unifies with [+sonorant] and X (the segment to the left) unifies with
[+sonorant] (i.e., X and Y are both sonorants), then X is [+vfv]. (81b) states that if Y is
[+sonorant] but X is [sonorant], then the specification of X for [vfv] is the same as
the specification of X for [voice] (in other words, a nonsonorant X retains its
136
underlying voicedness before a sonorant). Finally, (81a) states that if Y is [sonorant],
the specification of X for [vfv] is the same as the specification of Y for [vfv],
regardless of whether X is a sonorant or not.
These constraints can be added to all the onset and coda rules, with the result
that the specification of segments for [vfv] is determined by the grammar rather than
lexically: for all consonants this feature is determined by the distribution of the
consonant.
It is worth considering what changes would need to be made to the voicing
assimilation constraints in order to model the facts of Jakobsons idiolect as described
in Jakobson (1978). All that would be required is to replace all three voicing
assimilation constraints in (81) with the single constraint in (82):
(82) voicing_assim(Y,X):-
Y [o vfv],
X [o vfv].
Now we move on to consider the domain of the voicing assimilation rules.
Proceeding on the assumption that syllable structure does not affect voicing
assimilation, we see that it is not enough just to apply the constraints in (81) within
onsets and codas, because voicing assimilation could occur across the boundary
between the coda of one syllable and the onset of the next, as in the following
example:
137
(83) &'"%#,. !%4'1k.d6&! (H4'Dq.d6&) anecdote
25
The domain of voicing assimilation must therefore be some constituent higher in the
phrase-structure hierarchy than syllable. In the current PSG, the voicing assimilation
constraints can be enforced by stipulating that the right margin of any constituent
must agree in voice with the left margin of the next constituent. This is where the
features [left:X] and [right:Y] are needed. Take for example, the third rule from (34),
which is repeated here (without the feature [+stress]):
(84) strongfoot syllable weakfoot
The following modification to this rule, where the rule in (85) replaces that in (84),
will ensure that the voicing of one syllables coda always agrees with the voicing of
the following syllables onset in the relevant cases.
(85) strongfoot syllable weakfoot
[left:A, [left:A, [left:C,
right:D] right:B] right:D]
{voicing_assim(C,B)}
(where A, B, C, D represent feature matrices as discussed in 2.5.1)
To exemplify these rules, the structure associated with !"#$%& is shown in Figure 8. It
is assumed here that the phonetic representation of !"#$%& !0'1$'2%! radish is (0'1&'23)

25
If all Russian syllables are open, as suggested by Bondarko 1969, the problem of intersyllabic
voicing assimilation does not arise, because every consonant cluster must fall wholly within an onset.
Bondarkos claim that all Russian syllables are open was rejected for other reasons, however; see 2.3
for details.
138
and that a syllable boundary falls between (&') and (2), following Avanesov (1956:
164). (In this structure, only the relevant parts of the feature matrices are shown.)
Figure 8. Tree-structure for !"#$%& /r'cd'ka/ |r'ct'.ka]
Finally, let us consider where the proposed analysis stands in relation to
Halles claim (based on the facts of voicing assimilation) that a morphophonemic
level of representation, together with the phonological rules which derive the phonetic
prosodic_word
syllable
left:[+sonor],
right:[vfv,
sonor]
onset rime
[right:vfv,
sonor]
nucleus
!0!
[+voice
+vfv, +sonor]
!1! !2!
[voice
"?", sonor<)
weakfoot
left:[voice,
vfv, sonor]
syllable
left:[ voice,
vfv, sonor]
!%!
onset
[vfv,
sonor]
rime
nucleus
coda
[vfv, sonor]
!$'!
[+voice, vfv,
sonor]
strongfoot
139
representations, is sufficient to account for the observed facts of voicing assimilation
(Halle 1959: 22-23). Halles morphophonemic representations correspond to the
phonemic representations in obliques (i.e. !!) in this dissertation and in Chomsky
and Halle (1968), because the function of both is to account for ambiguities due to
homophony, as Halle puts it. Examples of morphophonemic representations from
Halle (1959: 23) are given below:
(86) Phrase Gloss Morphophonemic
representation
(a) +,% 6) was (he) getting wet? !+62<-',!
+,% </ were (he) getting wet !+62<A,!
(b) *"3$ 6) should one burn? !51&8'<-',!
*"3$ </ were one to burn !51&8'<A,!
Now, Halles discussion was controversial in its claim that there need be no
separate intermediate level at which surface contrasts were stated, a level referred to
by Christie (1976) as the classical phonemic level. According to Christie, the
classical phonemic representations are derived from morphophonemic ones by
morphophonemic rules, and allophonic rules then do the work of deriving phonetic
representations from classical phonemic ones (Christie 1976: 38).
For example, the classical phonemic representations derived from the
morphophonemic representations in (86a) above would be !+62<-',! and !+6B<A,!
respectively, with the alternation between !2! and !B! accounted for by
morphophonemic voicing assimilation rules. Although voicing assimilation does
apply in the case of *"3$ </, the classical phonemic representations of the phrases
in (86b), on the other hand, would be !51&8'<-',!, !51&8'<A,!. The representation !51$5'<A,!
140
is impossible because !&8'! does not have a voiced counterpart !&8'! in the phonemic
inventory of Russian (see Table 7 above). The alternation (&8')~($5') must therefore be
treated as allophonic.
Halle (1959) argued that recognizing the classical phonemic level thus meant
stating the voicing assimilation rules twice, once as morphophonemic rules and once
as allophonic rules, resulting in the loss of a generalization. After the publication of
Halles monograph, however, the level of surface contrast continued to be defended,
for example by Schane (1971) and Christie (1976). I would argue that Schanes view
of phonology is more unconstrained than is necessary, in that it involves
morphophonological rules convert[ing] one segment into another (Schane 1971:
517). Rules of this kind must be at least context-sensitive. Christie presented an
analysis of Russian voicing assimilation which retains a level of surface contrast but
manages to state the generalization about voicing assimilation only once: his solution
is to recognize a separate phoneme of voice in the phonological inventory, which
observes the phonotactic constraint that it can occur only in the position following an
obstruent and preceding a non-obstruent segment (Christie 1976: 38). The obvious
cost of Christies analysis, however, is an enlargement of the phonological inventory.
The current work stands closer to Halle (1959) than to Christie (1976) since I
do not recognize a separate phoneme of voice. The important lesson to be learnt
from Halle (1959) is that the attempt to distinguish levels of representation formally
often ends up muddying the waters rather than resulting in a simpler view of
phonology. In the current grammar, it is true that all features must become more fully-
specified as the grammar fills in information. It is also true that some features, such
as [voice], are redundant in the process of mapping fully-specified representations
141
(surface representations) to phonetic events, while others, such as [vfv], are always
unspecified in minimally-specified representations (underlying or phonemic
representations). This distinction between [voice] and [vfv] may tempt one to
establish a general distinction between underlying and surface features. However, in
most cases it does not make sense to assign features to an underlying or surface
level, since most features are needed both to distinguish one phoneme from another
and in the mapping of fully-specified segments to phonetic events. Thus, establishing
a general distinction between underlying versus surface features or levels of
representation is, in my view, unhelpful.
2.5.3 A test of the extensions to the grammar
The subject of the test described in this section is the full linear syllable
structure grammar including the enrichments proposed in 2.5.2.1-2.5.2.4. Testing the
grammar, in the case of the four phonological processes under consideration, involves
a check that the correct phonetic representations are derived from the underlying
phonological representation by the grammar. This raises the question of what the
correct phonetic representations actually are. Unfortunately, there are no dictionaries
available, as far as I am aware, which list phonetic representations of Russian words
systematically. There is an index in Trofimov and Jones (1923) containing full
phonetic transcriptions for a sizeable number of Russian words. Clearly, however, the
date of this publication means that the transcriptions may not consistently reflect
modern pronunciation. Avanesov and Oegov (1959) is a comprehensive phonetic
dictionary of Russian containing approximately 52,000 words, but it gives the
transcriptions of only those parts of each word which do not correspond
straightforwardly to orthography (for example, [9
$
<
$
] is given as the partial
142
transcription of )9<>',0, indicating the fact that !*! assimilates to !A! in
palatalization). Furthermore, as Halle suggests in his discussion of the physical
measurement of speech, there is the issue that correct phonetic representations can
ultimately be obtained only on the basis of acoustic analysis. And the very process of
attempting to formalize a grammar may raise unforeseen but real questions, such as
the effect of syllable boundaries on voicing assimilation, as discussed in 2.5.2.1-
2.5.2.4.
There are therefore a number of practical problems associated with conducting
a full-scale test of the grammar using the whole corpus described in 1.5. The approach
taken here is therefore perhaps more modest. Essentially, I aim to determine whether
the predictions of the grammar are as expected, i.e. that they are in line with the
predictions of traditional works on Russian phonology such as Avanesov (1956) and
Halle (1959).
We now turn to the details of how the test was conducted. As we have seen,
phonetic representations in the current grammar are fully-specified terminal
symbols, that is, symbols for which all features are specified as + or . Each fully-
specified terminal symbol, in turn, has an alias such as (D), (A'), etc, which stands for
the complex of features which each segment represents. Underspecification in the
current framework, on the other hand, is reflected in the fact that the pluses and
minuses are not given for all features applicable to a phoneme. For example, the
symbol in (87) represents the phoneme !%! in the most underspecified form possible:
!%! must have the feature [rime:[high]], since this feature is lexically specified for the
phoneme !%!.
143
(87) !%!
[rime:[high]]
The symbol in (88), on the other hand, represents one of the possible ways in
which !%! can be fully specified. The alias for (88) is the symbol (D), as shown in Table
14 on page 103.
(88) !%!
[stress,
+peripheral,
onset:[back]
rime:[high, +low]]
In this framework, therefore, the problem of phoneme-to-allophone
transcription reduces to the filling in of missing feature matrices as in Table 14.
(Pluses and minuses are themselves [albeit trivial] instances of feature matrices.) If
one wishes to represent allophones as phonetic symbols instead of as feature
complexes, one then needs to substitute the aliases in each case.
As mentioned in Chapter 1, there are a number of functions which grammars
should be able to fulfil: (a) parsing (assigning a structure to) well-formed strings; (b)
generation of well-formed strings; (c) rejecting (completely or incompletely specified)
ill-formed strings; and (d) filling in the information missing from incompletely
specified (well-formed) strings. Providing the grammar is decidable (see 1.2) these
functions will all be fulfilled. And, as we have seen, context-free languages are all
decidable. Thus the linear grammar described in this chapter will always be able to fill
in the missing information from an underspecified string, or convert phonemic
representations into phonetic ones (or vice versa).
144
The sample used for the test was a selection of 1,000 word-forms. These forms
were chosen such that all four of the phonological processes discussed above would
come into play. Almost all words (providing they contain an onset) activate the
consonant-vowel interdependency rules. Reduction of vowels is relevant only in
forms of two or more syllables, so monosyllabic words were excluded from the
sample. Word-final devoicing is relevant to words whose final segment is an
obstruent (including !"!): all words which did not meet this description were therefore
also excluded from the sample. Finally, voicing assimilation is relevant to words in
which there is an obstruent cluster. However, words which did not include an
obstruent cluster were not excluded from the sample: this was to ensure that sonorants
did not undergo or trigger voicing assimilation, in accordance with the standard
theory. The 1,000 words in the sample were selected randomly from the total set of
words which met these criteria. These 1,000 words were stored as entries in a
computer file (as underspecified phonological representations, with information on
stress placement included). Against each entry, a phonetic representation based on the
predictions of the standard theory, but deliberately excluding the effects of
phonological processes not discussed here (such as assimilation in palatalization and
raising of !1! before a palatalized consonant, as in "-.$ !71#&'! (7M#'&')), was entered
manually.
The grammar was then presented with each entry in the file, and required to
parse it. As a result of this, each word-form in the sample became a fully-specified
feature matrix. The fully-specified feature matrices were then automatically converted
into phonetic representations, based on the aliases given in Table 14. These were then
145
compared (again, automatically) to those which had been entered manually. The
results of this operation were as follows:
Table 17. Results of phoneme-to-allophone transcription test
No. of words % of total
Correctly transcribed 986 986
Unparsed 14 14
Total 1,000 1000
14 forms were unparsed because the grammar was unable to assign a syllable
structure to them, for the reasons already discussed at the end of section 2.3.2. The
failure of the grammar to assign structure to these forms thus had nothing to do with
the phonological processes under discussion. These 14 forms included, for example,
'&3&6$-.( !4%&8'%-'#&"! leaderships (gen. pl.). As can be seen, no forms were
incorrectly transcribed. In other words, whenever the grammar was able to assign
structure to a string of phonemes, it invariably filled in the correct missing
specifications, resulting in a correct phonetic representation.
It should be emphasized that this level of success was attained only after
several passes through the data, with modifications made to debug the grammar after
each pass. The rules given in sections 2.5.2.1-2.5.2.4, of course, reflect the final
debugged version of the grammar.
To summarize this section, I have shown by means of an empirical test that it
is entirely feasible to model four well-known phonological processes within a syllable
structure grammar, with essentially the same results as those predicted by classical,
rule-based, generative accounts which rely on rule-ordering, feature changing, and/or
transformational rules. The implementation of these phonological processes in a
146
monotonic framework, where structure is built up rather than modified destructively,
shows how underspecified versus fully-specified representations can be viewed in
linear grammar. In the course of the discussion, I have also raised some questions
about the relationship of syllable structure to voicing assimilation; the results of future
experimental research addressing these questions will doubtless have implications for
the phonological model proposed here.
2.6 Summary
This chapter has dealt with issues surrounding Russian syllabification and
segmental phonological processes. I have made a distinction between the constraints
and methods of syllabification, corresponding to the distinction between phrase-
structure rules and their procedural interpretation. The discussion of the linear
implementation of aspects of Russian syllabification served to show that it is helpful
to keep constraints and methods distinct, because the same grammar can be
implemented procedurally in many different ways. In phonology, however, what we
are primarily interested in is the constraints themselves. Three different approaches to
syllabification were explored in this chapter. None of these approaches achieve any
more than linear grammar does, given a procedure for interpreting the linear grammar.
Arguably, however, linear grammar has the advantage in that it keeps the grammar
and its procedural interpretation separate.
Developing the idea of syllabification within linear grammar, I showed in this
chapter that other phonological constraints word-final devoicing, consonant-vowel
interdependencies, voicing assimilation and unstressed vowel reduction can all be
incorporated into the syllabification grammar with minimal extra complication.
During this discussion, however, I argued that further acoustic tests should be carried
147
out to determine the validity of many of the phonological constraints which are
commonly claimed to hold: most significantly, perhaps, the issues surrounding the
interaction of syllable structure with voicing assimilation have not yet been addressed
systematically.
A key outcome of the discussion of phonological processes in Russian was the
establishment of a fundamental distinction between three different types of feature.
One type is exemplified by [voice]: this type of feature is lexically specified for
segments. One can think of these features as being phonological features,
underlyingly specified for each relevant segment in the phonological inventory. The
second type, exemplified by [onset:X], is non-lexical for all segments in the
phonological inventory: the specification of segments for [onset:X] and other such
features is determined by the properties of the segment in question in its particular
environment. Features of this type can be thought of as purely phonetic or surface
features. The third type, exemplified by [sonorant], is lexically specified for some
but not all segments in the phonological inventory, and features of this type have
some aspects in common with both phonological and phonetic features. It should
be stressed, however, that the phonological features do not change into phonetic
features at some stage in the derivation, as envisaged in Chomsky and Halle (1968).
Instead, all three types of feature are effectively present at all stages, except that the
purely phonetic features are redundant in underlying representations. Most
important, however, the current analysis has the advantage that it clearly remains
within the bounds of linear grammar, although this gain is offset by having to use
extra features in the grammar.
148
In Chapter 3, I shall examine the rules of Russian word-formation in a similar
manner, developing context-free rules to account for the structure of morphologically
complex word-forms. In Chapter 5, the word-formation grammar will be combined
with the syllabification grammar from this chapter; with a few modifications, this
combined grammar will be demonstrated to account for Russian word-stress, with a
good rate of accuracy and in a manner broadly consistent with that proposed in
metrical stress theory.
149
Chapter 3: Morphological structure
3.1 Introduction and aims
In this chapter, I present a context-free grammar of Russian word-
formation. A quick inspection of the contents of the Academy Grammar (vedova et
al. 1980) reveals that Russian word-formation is a complex area: over 300 pages in
the grammar are devoted to the subject. It is therefore hard to do the subject justice in
the space of a chapter. For this reason, the scope of this dissertation is necessarily
limited in its application to morphology: I do not make any claims about the
relationship of morphology to syntax or semantics, and therefore the morphemes
given in Appendix 3 as part of the grammar are not associated with morphosyntactic
features (although there is of course no reason in principle why such features could
not be added). However, the aim here is not so much to contribute to scholarly
knowledge about word-formation per se, as to lay the foundations for a formal
analysis of vowel-zero alternations (discussed later in this chapter) and stress
assignment (discussed in Chapter 5). As is generally acknowledged, morphological
analysis is crucial to both these areas (see Oliverius 1976: 60, Garde 1980, Halle
1973, 1997, and Melvold 1989 on morphology and stress assignment; see Pesetsky ms
1979, Gussmann 1980, Rubach 1984, and Booij and Rubach 1984 on morphology and
vowel-zero alternations in Russian and related languages). It is therefore essential to
make at least a first approximation at a Russian word-formation grammar, which can
then be brought together in a coherent manner with other grammatical modules, such
as syllable structure.
150
The chapter is organized as follows. Section 3.1.1 reviews the relevant
literature on morphological theory. In section 3.1.2, I review how context-free
grammar can be applied to word-formation, following (for example) Selkirk (1982b)
and Lieber (1992). In section 3.2 I then develop a specific context-free grammar of
Russian word-formation, showing first how Oliveriuss (1976) comprehensive
morphological tokenization of 2,493 Russian words can be utilized to establish a
morphological inventory, and then incorporating the observations made in existing
scholarly works to account for hierarchical structure in different parts of speech in
Russian. In 3.3 I show how an account of vowel-zero alternations can be incorporated
into the existing grammar. In section 3.4 I propose a heuristic for deciding between
multiple candidate morphological structures assigned by the same grammar, and in
3.5 I describe tests which show that the word-formation grammar has an overall
coverage of 6649%, and that it predicts the vocalization and deletion of jers correctly
in 9095% of cases. In the latter case, I suggest that the error rate of the theory could
be reduced if the theory were modified to take account of epenthesis in a relatively
small number of well-delineated cases.
The coverage of my grammar (6649%) is comparable to that of existing
computationally-implemented grammars (e.g. 73% in Worth, Kozak and Johnson
1970, 63% in Coleman 1993), but is clearly not as high a coverage as that of the
syllable structure grammar presented in Chapter 2. Although the higher failure rate of
the grammar does feed through to other areas of grammar (for example, if a string
cannot be assigned morphological structure, and stress assignment depends on
morphological analysis, the string will not be assigned a stress either), there is no
reason to believe that the percentage of correctly stressed words would not increase
151
proportionately with the coverage of the word-formation grammar. Thus the fact that
the word-formation grammar set out in this chapter is only a first approximation does
not affect the main argument in the concluding chapter of this dissertation, which is
that Russian stress assignment is context-free and can be reliably predicted on the
basis of morphological and syllable structure.
Despite the fact that the grammar in this chapter has a comparatively low
coverage, this chapter follows essentially the same approach as other chapters. I make
no proposals without testing them empirically for consistency and coverage; no rules
are more powerful than context-free; and the requirement that the grammar should be
able to be implemented computationally ensures that it remains formally explicit.
Owing to constraints of time and space, some specific areas are not
investigated as fully as might be desirable. For example, morphophonological
alternations such as !+6B!~!+65!, as in +,58 !+6B+;! I can versus +,*"B$
!+65+18'! you can, are dealt with by listing the relevant allomorphs (!+6B! and
!+65!, in this case) separately in the lexicon. This is not necessarily to imply that I
view !+6B!~!+65! as fundamentally different from, say, !$6#2!~!$6#62!, which I do
account for by rule. Ideally, one would like to formulate the theory which most simply
accounts for all the facts, and it is probably the case that the current theory could be
simplified further. A detailed theory of morphophonological alternations in Russian
might well be the subject of another dissertation. Nevertheless, I believe that the
theory I propose represents an advance on existing knowledge in that no existing
generative theories have been tested. The tests of coverage carried out on my own
theory supply a valuable measure of accountability, or set a standard, against which
any future theories will have to measure up. This is true not just for the
152
morphophonological theory proposed in this chapter, but also for the theories of
syllable structure, stress assignment, and so on, developed elsewhere in this thesis.
3.1.1 Generative approaches to word-formation
The areas of disagreement in the literature on morphology have principally
related to the division of labour between the lexicon and morphological rules. Some
(Jackendoff 1975, Aronoff 1976, Bybee 1988, Zwicky 1992) have played down the
importance of derivational rules, arguing that most information is stored in the lexicon
and that there is no derivation of lexemes from one another as such: on this
interpretation, words, not morphemes, are the basic signs of morphology, and
derivation is reduced to a set of static relations between items in the lexicon. This type
of approach is variously referred to as static-condition morphology (Zwicky 1992)
and word-based morphology (Anderson 1988: 162). Connectionist approaches (e.g.
Bybee 1985) and Network Morphology (Corbett and Fraser 1993, Brown, Corbett,
Fraser, Hippisley and Timberlake 1996), which will be discussed further in the next
chapter, are also essentially word-based. A psycholinguistic argument advanced for
word-based morphology is that it is psychologically more plausible than the
alternative view that lexemes are derived from one another on-line (Bybee 1988).
At the other end of the spectrum is what Anderson refers to as morpheme-
based morphology (Pesetsky ms 1979, Selkirk 1982b, Lieber 1981, 1992, Scalise
1984, Sproat 1985, and Toman 1998), which emphasizes word-internal hierarchical
structure and sees word-formation as the addition of affixes (which have their own
independent entries in the lexicon) to stems. It has been noted that there are certain
aspects of morphology which cause problems for morpheme-based approaches,
notably infixation, reduplication, and subtractive morphology (Anderson 1988: 153,
153
Sadock 1991: 26, Beard 1998: 48, Spencer 1998: 140); many of these issues are
directly relevant to Russian, and will be touched upon in this chapter.
Related to the issue of the lexicon versus morphology is that of how
morphophonological processes should be described. This is particularly relevant to
Russian and other Slavic languages: as Spencer (1998: 133) notes, there are a number
of Slavic suffixes which palatalize the final consonant of their bases, leading to
situations of allomorphy. One approach to this phenomenon is to list the different
phonological shapes a given morpheme may have as part of its lexical entry, as
suggested by Zwicky (1992: 336). The other approach, which tends to be espoused by
generative phonology (for example, Chomsky and Halle 1968, and more recently
Rubach 1984, Booij and Rubach 1984, Dressler 1985), is to posit a single underlying
phonological representation for each morpheme, and derive the appropriate surface
morph by quasi-phonological means, as Spencer puts it. Booij and Rubach (1984)
also include an analysis of vowel-zero alternations in these terms.
A third approach attempts to accommodate both word-based and morpheme-
based morphology under one umbrella. The argument is that the more productive a
morphological rule is, the more likely it is to be involved in on-line derivation, but the
use of existing items in the mental lexicon is always a possibility, particularly if those
items are of high frequency (Pinker and Prince 1988, Baayen and Sproat 1996,
Aronoff and Anshen 1998). As Aronoff and Anshen put it, the search for the proper
word can be viewed as a race between the mental lexicon and the morphology. Both
operate simultaneously, and the faster one wins (Aronoff and Anshen 1998: 240).
154
As can be seen, these areas of discussion reduce to decisions between
assigning information to the lexicon or encoding it by rules. As Chomsky (1970: 185)
notes, there is a trade-off between the two:
In general, it is to be expected that enrichment of one component of the grammar will
permit simplification in other parts. Thus certain descriptive problems can be handled
by enriching the lexicon and simplifying the categorial component of the base, or
conversely
It is important to appreciate that declarative grammar in the sense of
Scobbie, Coleman and Bird (1996) is not necessarily synonymous with maximal-
lexicon grammar: declarativeness does not in itself guarantee that derivational depth
(the reliance on the categorial component of the grammar, in Chomskys terms) will
be reduced. While the current theory of Russian phonology is certainly declarative, it
is also one which views language as a system for which we would like to find the
simplest possible description, following the general principle of descriptive economy.
No claims are made about psychological reality; it may even be that the principle of
descriptive economy is irreconcilable with that of psychological reality, and that
linguists must choose to adhere to one or the other principle depending upon their
individual interests and the way in which they view language. On the other hand, it
may be that following the principle of descriptive economy will lead to the discovery
that stronger linguistic theories are also psychologically the correct ones; this is
clearly what one would hope for. It is important to appreciate, however, that following
the descriptive economy route as opposed to the psychological route need not
involve sacrificing declarativeness.
Zwicky (1992) advances a number of arguments for word-based morphology;
however, some of these arguments are flawed. First, Zwicky argues for word-based
155
morphology, but still proposes a framework in which there are rules to account for
regularities (which he describes as default-setting, functions, and redundancy
predictors). There is no doubt that lexical redundancy rules must play some role in
grammar: Chomsky (1970) provides evidence of this. However, Zwicky claims that
derivational frameworks for [inflectional rules and derivational rules] are
inherently more powerful than static-condition frameworks. But it appears that he
measures the power of static-condition frameworks without taking regularity rules
into consideration. One cannot have it both ways: either the rules are part of the
formal theory, or they are not. If they are, whether one calls them default-setting,
functions, redundancy predictors, transformations, or anything else, is of no
importance: it is not legitimate to exclude them from consideration when establishing
the power of a theory or framework. If the rules are not part of the theory, then the
regularities remain unaccounted for.
Secondly, Zwicky claims that in derivational frameworks (for any
component[s] of grammar), interactions between rules are a matter of applicational
sequence, or ordering (Zwicky 1992: 344). This need not be true: in this chapter, I
propose a grammar of Russian morphology which can be seen as both derivational (in
that words are built up by on-line rules from morphemes, and the latter are the items
in the lexicon) and unordered (because it is no more powerful than context-free).
Finally, Zwicky (1992: 340) is dismissive of frameworks which aim to
minimize the lexicon:
I can find no good place in a morphological framework for a truly minimal list of any
sort. And every defense I have seen of a minimal morphicon, right up to Bromberger
and Halle (1989), rests not on theoretical observations but on claims about the mental
morphicon. Somehow, whenever theorists start talking about the morphicon they end
up worrying about what people can keep in their heads.
156
However, as Chomsky (1970: 185-186) states, this kind of argument misses the point,
because:
The proper balance between various components of the grammar is entirely an
empirical issue. We have no a priori insight into the trading relation between the
various parts. There are no general considerations that settle this matter It would be
pure dogmatism to maintain, without empirical evidence, that the categorial
component must be narrowly constrained by universal conditions, the variety and
complexity of language being attributed to the other components.
Thus the question of whether a maximal-lexicon theory is better than a minimal-
lexicon theory cannot be answered in the abstract: it is an empirical question which
can be answered only by looking at the specific proposals of the two theories and
deciding which proposals are right. The most important consideration, clearly, is that
strong linguistic generalizations should not be missed.
One issue on which there is general agreement among morphologists is that at
least the following information needs to be included for each lexical entry: a
phonological representation, a grammatical representation (used by the languages
syntax), and a semantic representation (Jackendoff 1975: 642, Sadock 1991, Zwicky
1992, Beard 1998). The morpheme inventory which I propose in Appendix 3 aims to
include phonological representations comprehensively. Some syntactic information
(such as information about grammatical number) is included where it appears to be
relevant to phonology. However, the claims which I make in this dissertation about
syntax or semantics are necessarily very limited in scope. It would certainly not be a
problem in principle to extend the morpheme inventory to include comprehensive
syntactic and semantic information, but all I would claim here is that syntax and
semantics can help alleviate problems of overgeneration in the word-formation
157
grammar (see 3.2.2.2), and that semantics can sometimes be helpful in resolving
problems connected with the morphological inventory (see 4.3.2.6 for an example).
Clearly, to formulate a specific theory about Russian morphology, some
position also needs to be taken on the issues on which there is no agreement,
otherwise it is impossible to proceed. In the present analysis, I aim in general to
reduce the size of the lexicon to a minimum, allowing the morphology to do most of
the work in deriving one form from another, for the sake of a maximally explanatory
theory. In other words, I follow a morpheme-based approach. While acknowledging
that some facts of morphology are hard to explain using this approach, I would argue
that the benefit outweighs the cost: the principal advantage of a morpheme-based
approach is that it substantially reduces the size of the lexicon. Morpheme-based
analysis also lends itself well to phrase-structure analysis, allowing consistency to be
maintained with the view of syllable structure advanced in Chapter 2. As we shall see
in Chapter 5, this becomes particularly important in stress assignment. A generative
context-free account of Russian stress, based on a parametric view of stress
assignment (Hayes 1995), can be expressed very neatly as the intersection of a phrase-
structure, morpheme-based word-formation grammar, and a phrase-structure
syllabification grammar.
One area in which I have not pursued minimization of the lexicon to its fullest
extent is in the treatment of allomorphs (resulting mainly from consonant
palatalizations). For example, the reader will note from Appendix 3 that !82/6-!
rn
and
!82/6-'!
rn
have independent entries (see page 383). This choice to exclude
morphophonological alternations of this type from the current account was made
mainly owing to limitations of time, although such an analysis does have the potential
158
to reduce the size of the morpheme inventory further. As we have seen, this approach
would be more in the spirit of generative grammar than word-based morphology. It
may prove, with future research, that such an analysis of allomorphy is fruitful.
3.1.2 Morphology and context-free grammar
As mentioned in section 1.3.1, context-free rules can be represented
graphically as tree structures (parse trees). An example of this was seen on page 75.
Naturally, this applies as much to morphological structure as it does to syllable
structure. For example, Figure 9 shows a morphological parse tree for '"#()*)+,-.$
!4'1$"',5,+6#&'! immovable property.
Figure 9. Parse tree for '"#()*)+,-.$ /n'cdv'ijimost'/
prefix
noun-stem
adj-stem subst. suffix
noun-infl.
verb-stem adj. suffix
verb root
n'c dv'ij im ost'
word-form
noun
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
(89) Exhaustive list of CFG rules inferred from parse tree in Figure 9
word-form noun
noun noun-stem noun-inflection
noun-stem adj-stem subst. suffix
adj-stem verb-stem adj. suffix
verb-stem prefix verb-root
prefix !4'1!
verb-root !$"',5!
adj. suffix !,+!
subst. suffix !6#&'!
infl. ending !!
It should be emphasized that the phrase-structure approach to morphology is not new.
This approach is followed, for example, by Selkirk (1982b) and Lieber (1992).
verb-stem
prefix verb root
subst. suffix
ost'
160
Predictably, a context-free word-formation grammar can fulfil the same
parsing and generation functions as a context-free (or linear) syllable structure
grammar: just as the syllable structure CFG identifies the positions of syllable
boundaries within a word, a morphology CFG identifies morpheme boundaries. As
we shall see in Chapter 4, this is the minimum requirement for morpheme-based stress
assignment in Russian. Clearly, though, a CFG does more than just morphological
tokenization (segmenting a word into a string of morphemes), as in Figure 11. It also
assigns hierarchical structures to these strings.
Figure 11. Morphological tokenization of '"#()*)+,-.$ /n'cdv'ijimost'/
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).
n'c dv'ij im ost'
161
(90) adjective adj-stem infl. ending
inflectional ending !,7!
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 /n'cdv'ijimij/
3.2 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 word-
formation grammars by reference to criteria of coverage, coherence and scope. As far
word-form
verb root
adjective
adj-stem adj-infl
verb-stem adj. suffix
prefix
n'c dv'ij im ij
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 Oliveriuss (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 Oliveriuss 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, Oliveriuss dataset seemed the only feasible point from which to start.
164
(Tixonovs 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 Oliveriuss 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 Oliveriuss 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 Oliveriuss theory will be only
briefly considered, as a more detailed discussion has already been given in 2.3.1.1.
Oliveriuss 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 Oliveriuss 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 root
s suffix
c non-final morpheme which can stand in isolation (i.e. clitic):
prefix, preposition, conjunction, interjection or particle
i inflectional ending
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 morphemes 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 pronominal
a adjectival
n substantival
v 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"--&)
r resultative process
(A@@"%.)(',-.$, #"0-.("'',-.$, !"986$.&.)(',-.$ 4!,7"--&)
i iterative process
()."!&.)(',-.$, 4,(.,!>"+,-.$, +',5,%!&.',-.$ 4!,7"--&)
c completed process
(4!":,#)+,-.$ 4!,7"--&)
Finally, the last character in the code may be a digit, irrespective of the
morphemes 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: 142-
143)
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 Morpheme Features Cross-reference to
Carstairs (1987)
in1
!!
nom., acc. sg. page 142, no. 21b
in2
!!
gen., acc. pl. page 142, no. 22a
in3
!%!
nom. sg. n/a
in4
!%!
nom., acc. pl. (neuter nouns) page 142, no. 20
in5
!%+',!
instr. pl. n/a
in6
!%+!
dat. pl. n/a
in7
!%C!
loc. pl. n/a
in8
!1!
dat., loc. sg. page 142, no. 23
in9
!17!
gen., acc. pl. page 142, no. 22a
in10
!,!
nom., acc. pl. page 142, no. 21a
in11
!67!
instr. sg. n/a
in12
!6"!
gen., acc. pl. page 142, no. 22a
in13
!7;!
instr. sg. n/a
in14
!+',!
instr. pl. n/a
in15
!%!
gen., acc. sg. page 142, no. 22b
in16
!%!
nom. pl. (masc. nouns)
27
n/a
in17
!%!
nom. sg. (neuter nouns)
28
n/a
in18
!6!
nom., acc. sg. page 142, no. 20
in19
!6+!
instr. sg. n/a
in20
!,!
dat., gen., loc. sg. page 143, no. 25
in21
!;!
acc. sg. n/a
in22
!;!
dat. sg. n/a
in23
!1!
nom. pl.
29
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!,@"--,!& !906
c2
+?'1#
rv
+#
sa
+60
sn
+%
in16
! professors.
28
For example, in )+> !,+'
rn
+%
in17
! name.
29
For example, in 5!&*#&'" !B0%5$
rv
+%4'
sa1
+
sn1
+1
in23
! citizens.
170
morphs are represented with the subscript outside the obliques, as follows: !#!
c
,
!+,#-!
rn
, !!
in1
, etc.
The translation of Oliveriuss 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 Oliveriuss 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 Oliveriuss 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"!"# !9'10'1$!
c
in front of
( !"!
c
in
4,%& !962%!
c
while
,: !6C!
c
oh!
These words, which are all prepositions, conjunctions and interjections, can be
subsumed under the general term particle. The first CFG rules which we need are
therefore ones to license such words. These rules will be as in (97):
(97) word particle
particle c
The terminal rules of the CFG have already made explicit which morphemes
are c morphemes. For example, we already have the rules in (98).
(98) c !9'10'1$!
c !"!
c !962%!
c !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%& !0;2
rn
+%
in3
! hand
3"+,#&' !&8'1+6$%4
rn
+
in1
! suitcase
4,6" !96-'
rn
+6
in18
! field
These words can all be accounted for by the context-free rule in (100).
(100) noun rn in
30
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 noun
word c
noun rn in
c !9'10'1$!
etc.
rn !0;2!
etc.
in !%!
etc.
We next turn to derived nouns with a single suffix. Some examples are shown
in Table 19.
Table 19. Further categories of morphological tokenization
Type of tokenization Example Number of words
of this type in
Oliverius (1976)
rv+sn+in (verb-root + noun-suffix + infl.) *)9'$
!5,
rv
+*4'
sn
+
in
!
life
100
ra+sn+in (adj-root + noun-suffix + infl.) !&#,-.$
!0%$
ra
+6#&'
sn
+
in
!
joy
46
Total 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 rv sn in
noun ra sn in
(104) noun noun-stem in
noun-stem rn
noun-stem verb-stem sn
noun-stem adj-stem sn
verb-stem rv
adj-stem ra
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. Oliveriuss (1976) tokenization of *"'1)'& /jcnjtj'ina/ woman
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)'& /jcnjtj'ina/ woman
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
jcn
rn
jtj'
sa
in
sn
a
in
noun-stem
jcn jtj' in a
word
noun
in
adj-stem
noun-stem
rn
sa
sn
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) !#9'1&#
rv
+,7
sn
+%-'
sa
+4
sa
+6#&'
sn
+
in
!
(b) !#9'1&#
rv
+,7
sn
+%-'
sa
+,7
sn
+%-'
sa
+4
sa
+6#&'
sn
+
in
!
(c) !#9'1&#
rv
+,7
sn
+%-'
sa
+,7
sn
+%-'
sa
+,7
sn
+%-'
sa
+4
sa
+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 rn
noun-stem verb-stem sn
noun-stem adj-stem sn
verb-stem rv
adj-stem ra
adj-stem noun-stem sa
(108) noun-stem rn
[semantics:concrete]
noun-stem verb-stem sn
[semantics:deverbal] [semantics:concrete]
noun-stem adj-stem sn
[semantics:action] [semantics:past_participle]
verb-stem rv
[semantics:concrete]
adj-stem ra
[semantics:concrete]
adj-stem noun-stem sa
[semantics:desubstantival] [semantics:concrete]
adj-stem noun-stem sa
[semantics:desubstantival] [semantics:deverbal]
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 far
32
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
! (example: +,3$ !+6&8'
rv
+
sa
+
ia
! to be able)
!&'
sa
+
ia
! (example: 3).&.$ !&8',&
rv
+%
svi
+&'
sa
+
ia
! to read)
!&',
sa
+
ia
! (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 iv
verb-stem verb-stem sv
[+derived] [derived]
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).&.$-> !&8',&
rv
+%
svi
+&'
iv
+#'%
c
! to be read):
(112) verb verb !#'(%)!
c
[+reflexive] [reflexive]
3.2.2.4 Prefixation
As pointed out in Townsend (1975: 149), prefixation plays a very minor role
in formation of nouns. This is confirmed in Oliveriuss (1976) analysis, where 837 of
the 967 prefixed forms include either a verb root or suffix. Most Russian prefixation
can thus be dealt with under the heading of verbal word-formation, and it is this which
I shall concentrate on in this section.
Prefixation in Russian verbal forms is very closely intertwined with verbal
aspect. As Townsend states:
With certain limited exceptions the addition of a prefix to a simplex imperfective
stem perfectivizes that stem. In traditional analysis and most grammar books a prefix
is either nonsemantic or semantic: if the former, it merely perfectivizes the stem
and is regarded as the perfective partner of the imperfective verb (Townsend 1975:
116).
181
Thus imperfective 3).&.$ !&8',&
rv
+%
svi
+&'
sa
! to read can be perfectivized by adding the
prefix !906!
c
(which in this case is non-semantic), forming !906
c
+&8',&
rv
+%
svi
+&'
sa
!. The
question is: what is the correct structural description of !906
c
+&8',&
rv
+%
svi
+&'
sa
!? The
candidates which one might consider are shown in Figure 15.
Figure 15. Three alternative representations of /pro
c
+tj'it
rv
+a
svi
+t'
sa
/
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 ones
intuitions about the semantic correlates of morphological structure. In most cases, in
906
c
&8',&
rv
infin
906
c
&8',&
rv
infin
906
c
&8',&
rv
infin
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 Ig, unprefixed, is glossed as burn, the
combination of podH- and Ig 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-g/podo-
gla, 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 podogla and podg
suggests the bracketing of (27a) below, the semantics seems to suggest the bracketing
of (27b):
(27) a. [ prefix [[root] inflection]]
b. [[prefix [root]] 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 type
33
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;#2
rv
+
sn
+
in
!
leave, related to ,.48-%&.$ !6&
c
+9;#2
rv
+%
svi
+&'
sa
! to let go. Instead of deriving

33
Booij and Rubachs 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;#2
rv
+
sn
+
in
! from !6&
c
+9;#2
rv
+%
svi
+&'
sa
! by some kind of subtractive operation
(which presents problems for phrase-structure analysis), both !6&
c
+9;#2
rv
+
sn
+
in
! and
!6&
c
+9;#2
rv
+%
svi
+&'
sa
! would be derived from the base !6&
c
+9;#2
rv
!. The corresponding
linear rule schemata would therefore have to be those in (113).
(113) (a) verb-stem (c) rv
[derived]
verb-stem verb-stem sv
[+derived] [derived]
(b) noun-stem verb-stem sn
(The rule in (113b) accounts for forms such as !6&
c
+9;#2
rv
+
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 (c) rn
[derived]
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
Havlks 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 [stole], prep sg; [glupij] stupid and [glupet] to become
stupid, just as [den] day and [dna], gen sg, or [bolen] sick, masc sg and
[bolna], 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) <,<!& !A6A0+%! beaver fur (gen. sg.)
<,<"! !A6A'10+! beaver fur (nom. sg.)
This approach has never been widely accepted since, as pointed out in a
number of works (Gussmann 1980, Szpyra 1992, Bethin 1998), there are instances
where epenthesis apparently does not occur in environments identical to those in
which it does occur, as shown by the examples in (116).
(116) <,<!& !A6A0+%! beaver (gen. sg.)
<,<! !A6A0+! beaver (nom. sg.)
Thus, as Szpyra (1992: 281) remarks, the logical conclusion is that the
presence vs. absence of yers is largely unpredictable and must therefore be marked in
the lexical representation of the relevant items. (It should be noted, however, that
Piotrowski 1992 manages to maintain an epenthesis account of the jers, accounting
for (115) versus (116) by stipulating that the !0! in (115) is lexically specified as
unsyllabifiable.)
The remainder of the works on vowel-zero alternations exploit the descriptive
devices of autosegmental phonology in representing jers. Spencer (1986) treats Polish
jers as skeletal vowel-slots with no feature melodies attached. Rubach (1986) (for
Polish) and Melvold (1989) (for Russian) take the opposite view, assuming that jers
are feature melodies with no skeletal slots. Szpyra (1992) assumes that jers are to be
represented as empty root nodes devoid of any melodic features such as
188
[consonantal] (p. 298), and that their manifestation as vowels (rather than zero) is
governed by syllabification principles reminiscent of epenthesis (specifically, they
vocalize when adjacent to unsyllabified consonants).
Here, however, I shall propose an analysis of vowel-zero alternations in
Russian which has the following properties:
- it accounts for regularities in the morphological inventory and accordingly reduces
the size of the inventory;
- jers are part of the lexical representation of morphemes;
- it does not postulate additional phonological entities beyond those already
mentioned in Table 7, abstract or otherwise;
- syllabification is not the main factor which determines whether jers vocalize or
delete, although the surface results of the vowel-zero alternations are of course
subject to syllabification;
- it does not require the postulation of an additional skeletal tier, as in
autosegmental phonology;
- it remains within the framework of context-free grammar.
The analysis of vowel-zero alternations proposed in this section rejects two claims
which have previously been made in accounting for the same facts; one of these is
specific to the Lexical Phonology framework, and the other is a general assumption
which has been made about the representation of the jers, the linguistic objects which
underlie the vowel-zero alternations in Slavic languages.
189
The first claim is that prefixes must be immediately dominated by the word
node for the correct predictions to be made regarding vocalization and deletion of the
jers (Pesetsky ms 1979, Booij and Rubach 1984). Pesetsky, Booij and Rubach are led
to make this claim because the framework in which they formulate their theory of jer-
vocalization and deletion relies on rule-ordering and cyclicity: the rules governing
jer-lowering (when jers surface as vowels as opposed to when they delete) apply to
the output of successive ordered word-formation rules (Pesetsky ms 1979: 14-15,
Rubach 1984: 31, Booij and Rubach 1984: 5-7, Rubach and Booij 1990: 128-129).
The current analysis does not accept rule ordering, and consequently is free to allow
morphologically more intuitive structures, yet is still able to account for the unusual
behaviour of prefixes with respect to jer-lowering.
The second claim is that jers must have their own phonological
representations. This assumption, exemplified in Szpyra (1992: 286) (Since yers
belong to the phonological inventory of Polish...), appears to be a non-sequitur from
the (correct) conclusion, on the basis of examples such as those in (115)-(116), that
jers must be present in lexical representations (as opposed to being derived by
epenthesis, for example), because their vocalization or deletion cannot be described
solely in terms of their environment.
Before proceeding, we need to consider the constraints under which we are
working. First, since jers may sometimes be deleted altogether, it must be the case
(assuming that jers can be represented at all in context-free grammar) that jers are
represented as part of the pre-terminal hierarchical structure, since rules of the form A
are permitted in CFG, but not rules of the form A. (Kiparsky 1982: 74
explicitly adopts just such a rule for jer vocalization, where zero rewrites as non-zero;
190
Kiparskys solution is unacceptable, since his phonology as a result must be more
powerful than context-free.) The second constraint is that jers must be part of the
lexical representation, as already mentioned. This means that jers must be given as
one of the constituents of the rules for individual entries in the morphological
inventory.
The only way in which these two constraints can be reconciled is if jers are
morphological, rather than phonological, objects. This means that they are not, strictly
speaking, vowels; it would be more accurate to say that they are morphological
objects which can surface as vowels. (However, it will be convenient in the following
discussion to refer to them as vowels, since they sometimes pattern together with the
[consonantal] segments in the phonological inventory. An example is in the
specification of the left edge of morphemes in Russian, which will shortly be
discussed.) Under the approach followed in this dissertation, the morphology and
syllable structure grammars intersect at their terminal elements: this ensures that the
language at their intersection is no more powerful than context-free. Thus, if jers are
represented pre-terminally and they are to be encoded in the morphological inventory,
then they must be pre-terminal in the morphology rather than in syllable structure.
What is needed, therefore, is a representation such as the following:
Figure 16. Representation of the morpheme /lask/~/lasok/ weasel
morpheme
# 2
jer
- %
191
This structure would represent the root morpheme in 6&-%& !-%#2%! weasel (nom. sg.)
(cf. 6&-,% !-%#62! weasels [gen. pl.]), distinguishing it from that in 6&-%& !-%#2%!
caress (nom. sg.), which does not alternate in the genitive plural (cf. 6&-% !-%#2!).
The constituent jer in Figure 16 differs from the constituents !-!, !%!, !#! and !2! in
that the latter are terminal symbols which are directly subject to the syllable structure
grammar, while jer is a pre-terminal symbol in the word-formation grammar. What
jer means is spelled out by further rules in the word-formation grammar, vocalizing
or deleting jer as appropriate. The consequence of all this is that jers are represented
lexically, but no additional phonemes need be postulated.
An important point about this method of representing jers is that it involves
centre-embedding context-free (not linear) rules. I use context-free rules because these
appear to capture the structure of morphemes including jers most aptly. Linear
grammar does not permit centre-embedding, as can be seen from Table 1 on page 20.
Indeed, as (Chomsky 1963: 394) shows, any language described by a context-free
grammar including centre-embedding rules and recursion cannot be described by a
linear grammar. Although I claimed on page 176 above that Russian morphology is
not recursive, it is still possible that there is no equivalent linear grammar which could
describe the structures necessary to account for vowel-zero alternations. It is for this
reason alone that I do not claim that the overall Russian word-formation grammar
proposed in this chapter is more restricted than context-free.
Now, this analysis is not yet explicit about when jers vocalize and when they
delete. The context-free rules governing this will clearly have to be able to tell
whether a jer is strong or weak, and this in turn will involve looking at the content
of following morphemes.
192
An important insight of Pesetsky (ms 1979) and subsequent followers of
Lexical Phonology was that the vocalization or deletion of jers depends on whether
they are located in a prefix, suffix, or inflectional ending. In the following discussion,
jers will be represented as E and O. (No attempt is made in this dissertation to account
for the alternation between front and back jers by rule.) Pesetsky suggested that the
following rules (which I shall refer to as the Lexical Phonology theory) govern the
vocalization and deletion of the jers:
(117) (a) Jers in i (inflectional) morphemes always delete, because they must
either be followed by a word-boundary (as in the masculine nominative
singular inflection !O!) or by a non-alternating vowel (as in the
feminine instrumental singular inflection !E7;!). (The evidence that
these jers exist is their effect on jers in preceding morphemes.)
(b) Jers in r (root) and s (suffix) morphemes always vocalize if
followed by a jer, and delete if followed by a (non-jer) vowel.
(c) Jers in c (prefix) morphemes vocalize if followed by a deleted jer,
and delete otherwise (i.e. if followed either by a non-alternating vowel
or by a vocalized jer).
Instead of modelling this theory using cyclic rules, as suggested by Pesetsky
and subsequent researchers in Lexical Phonology, what is proposed here is a context-
free encoding of the Lexical Phonology theory. This entails using four features,
[strong], [tense], [high], and [left:X], in the existing word-formation grammar.
The third and fourth of these features were used in the syllable structure module:
193
[high] was introduced on page 102, and [left:X] on page 104. The feature [strong]
follows the terminology traditionally used by Slavists (strong jers vocalize and
weak jers delete: cf. Isa;enko 1970), and should not be confused with the
strong/weak distinction in metrical phonology (cf. Liberman 1975). The feature
[tense] is introduced to differentiate between jers and non-jer vowels (in conjunction
with [high]), following Gussmann (1980), Rubach (1984), and others. However, this
is not to imply that additional segments are to be included in the phonological
inventory, as Gussmann and Rubach suggest.
First, we shall consider the distribution of [left:X] in the terminal rules. In this
case, X stands for the feature matrix [o tense, | high], which unifies with the
specifications of the first vowel (whether jer or non-jer) in the morpheme. (It is
assumed that these features spread across any intervening consonants.) Thus a
morpheme is [left:[+high, tense] if it contains a jer, and no vowels occur to the left
of the jer within that morpheme
34
. Accordingly, terminal rules of the type shown in
(118) are obtained.

34
There are two exceptions to this, which will be dealt with below.
194
(118) in ! O ! (the masc. nom. sg. ending)
[left: [+high,
tense]]
in ! % + ! (the dat. pl. ending)
[left: [high,
tense]]
rn ! - % # O 2 !
[left: [high,
tense]
Next, we turn to the distribution of [strong]. This is the feature which
encodes the actual phonological interpretation of jers (whether they vocalize or
delete): a jer in a [+strong] morpheme vocalizes, and one in a [strong] morpheme
deletes. Reflecting the fact that there are three ways in which jers behave (depending
on whether they occur in inflectional endings, roots or suffixes, or prefixes), there are
three basic distributional possibilities.
First, inflectional (i) morphemes must always be [strong] in accordance
with (117a). Thus in these morphemes the only feature which can vary is the feature
[left:X].
Secondly, we shall consider the distribution of [strong] in roots and suffixes.
The facts can be captured only if we allow [strong] to percolate up to pre-terminal
symbols in the tree. In other words, not just morphemes, but whole stems need to be
specified for [strong]. We shall stipulate that a stem is [o strong] if its rightmost
constituent is also [o strong]. (Because the word-formation rules are all left-
embedding, this constituent must be a terminal symbol.) The leftmost constituents
specification for [strong] is then determined by the specification of the suffix for
[left:X]: if the suffixs leftmost vowel is a jer, then the leftmost constituent is
195
[+strong]. Otherwise, it is [strong]. These facts are encoded in the rule schemata in
(119).
(119) stem stem suffix
[o strong] [+strong] [o strong,
left:[+high, tense]]
stem stem suffix
[o strong] [strong] [o strong,
left:[+tense]]
stem stem suffix
[o strong] [strong] [o strong,
left:[high]]
The distribution of features is exactly the same if the rightmost constituent is
an inflectional ending instead of a suffix, except, as already stated, the inflectional
ending must be [strong]:
(120) X stem inflection
[+strong] [left:[+high, tense],
strong]
X stem inflection
[strong] [left:[high],
strong]
X stem inflection
[strong] [left:[+tense],
strong]
We then have to consider the distribution of [strong] in the prefixation rules,
which, as we know from (117c), is different again. As Pesetsky (ms 1979) pointed
out, the relevant issue here is not whether there is a jer in the following r morpheme,
but whether there is a non-vocalized jer. The jer in the prefix !96$O! surfaces as a
196
vowel in !96$O+5EB+-+%! because there is a jer in !5EB!, and it has deleted. It fails to
surface in !96$O+5EB+-+O!, however, because the jer in !5EB! has vocalized.
I propose that these facts should be accounted for not in the word-formation
rules, but in the lexicon as a fact about root morphemes. This is the first of the two
exceptions referred to in footnote 34. For root morphemes whose first vowel is a jer,
the morphemes specification for [left:X] can be made dependent on its specification
for [strong]. Specifically, if the morpheme is [+strong], its jer vocalizes and it is
treated as a morpheme whose leftmost vowel is not a jer (with the result that any jer in
the prefix to its left will delete). If the morpheme is [strong], its jer deletes and it is
treated as a morpheme whose leftmost vowel is a jer (with the result that any jer in the
prefix to its left will vocalize). The following rules for !5EB! exemplify this.
(121) rn ! 5 E B !
[left: [high, [+strong]
tense]
+strong]
rn ! 5 E B !
[left: [+high, [strong]
tense]
strong]
Adopting this approach means that the prefixation rule schemata are similar to those
in (119) and (121). The actual rule schemata are as follows:
197
(122) stem prefix root
[o strong] [+strong] [o strong,
left:[+high, tense]]
stem prefix root
[o strong] [strong] [o strong,
left:[+tense]]
stem prefix root
[o strong] [strong] [o strong,
left:[high]]
Finally, we need to consider the other exception referred to in footnote 34. If a
morpheme contains neither jers nor non-alternating vowels (for example, the suffix !-!
in Pesetskys example !96$O+5EB+-+%!), we want the feature [left:X] to spread
leftwards from the following morpheme. For example, in !96$O+5EB+-+%!, the reason
the jer in !5EB! deletes is that the next vowel is !%!, so !-! should share the feature
[left:X] with !%!. This can be modelled if the lexical rule for !-! is as shown in (123).
(123) sv ! - !
[left: [o high,
tense],
o strong]
It needs to be underlined that the rules in (119), (120) and (122) are rule
schemata: they describe the distribution of [strong] and [left:X] in the existing CFG
morphology rules which match their descriptions, rather than being additional rules in
the morphology with an equal status to, say, those in (114).
35

35
In Appendix 2, the jer vocalization and deletion constraints are dealt with by defining a separate
predicate vowel_zero which is the same for all word-formation rules. This predicate is then defined
separately, using unification of feature structures. This is simply an abbreviatory device which makes it
possible to avoid stating the constraints in full for each word-formation rule.
198
Because the jers are themselves morphological constituents, the word-
formation grammar must contain rules to make explicit when they vocalize and when
they delete. These rules are now given in (124).
(124) O [consonantal, +back, high, low]
[+strong]
O !!
[strong]
E [consonantal, back, high, low]
[+strong]
E !!
[strong]
In accordance with these rules, the trees in Figure 17, Figure 18 and Figure 19
respectively show the structures of #"'23"% !$E4'+O&8'+O2+O! (surface form
!$14'+6&8'+62+!) day (double diminutive), 4,#,*56& !96$O+5EB+-+%! (surface
form !96$6+5B+-+%!) she set fire to and 4,#*,5 !96$O+5EB+-+O! (surface form
!96$+56B+!) he set fire to (without accounting for the alternation of !-! with zero in
the latter two forms).
199
Figure 17. Structure of #"'23"%
Notes to Figure 17:
- The jer in the inflectional ending deletes in accordance with (120).
- The jer in the second suffix vocalizes (is [+strong]), because it is followed by a
[left: [+high, tense]] morpheme.
- The jer in the first suffix vocalizes (is [+strong]), because it is also followed by a
[left: [+high, tense]] morpheme.
- The jer in the root vocalizes (is [+strong]), because it is also followed by a [left:
[+high, tense]] morpheme. It is [left:[high, tense]] in accordance with (121).
in
[strong,
left: [+high,
tense]]
sn
[+strong,
left:[+high,
tense]]
sn
[+strong, left:
[+high,
tense]]
rn
[+strong,
left:[high,
tense]]
noun
n-stem
[+strong]
n-stem
[+strong]
$'
E
[+strong]
4' &8' 2
O
[+strong]
O
[strong]
1 6 6
O
[+strong]
200
Figure 18. Structure of 4,#,*56&
Notes to Figure 18:
- The inflectional ending is [left:[high, tense]] because it consists only of !%!.
- The suffix shares the specification for [left:X] with the following morpheme,
because it contains neither jers nor vowels.
- The jer in the root deletes (is [strong]), because it is followed by a [left:[high,
tense]] morpheme. The root is [left: [+high, tense]] in accordance with (121).
- The jer in the prefix vocalizes because it is followed by a [left: [+high, tense]]
morpheme.
sv
[strong,
left: [high,
tense]]
verb
v-stem
[strong]
iv
[strong,
left: [high,
tense]]
v-root
[strong]
rv
[strong,
left: [+high,
tense]]
c
[+strong, left:
[high,
tense]]
O
[+strong]
6 9
E
[strong]
- B % 5 6 $
201
Figure 19. Structure of 4,#*,5
Notes to Figure 19:
- The jer in the inflectional ending deletes in accordance with (120).
- The suffix shares the specification for [left:X] with the following morpheme,
because it contains neither jers nor vowels.
- The jer in the root vocalizes (is [+strong]), because it is followed by a [left:
[+high, tense]] morpheme. The root is [left:[high, tense]] in accordance with
(121).
- The jer in the prefix deletes because it is followed by a [left:[high, tense]]
morpheme.
To conclude this section, I have shown that jers cannot have a representation in a
context-free syllable structure grammar: instead, they must be morphological objects
whose phonological representation is worked out by the word-formation grammar.
sv
[+strong,
left: [+high,
tense]]
verb
v-stem
[+strong]
iv
[strong,
left: [+high,
tense]]
v-root
[+strong]
rv
[+strong,
left: [high,
tense]]
c
[strong,
left: [high,
tense]]
E
[+strong]
O
[strong]
9 - B 5 6 $
O
[strong]
6
202
The different behaviour of jers in prefixes, roots/suffixes and inflectional endings
need not be accounted for by ordered cyclic rules, but may instead be encoded in
unordered context-free rules. These, like all the other context-free rules in the
grammar proposed in this dissertation, can be interpreted both as generative rules and
as rules which parse (assign structure to) phonological representations.
Although ordered rules and cyclicity are not a feature of the current
framework, it was shown that the traditional Lexical Phonology theory of jer-
vocalization and deletion can nevertheless be expressed in the current framework: my
grammar makes the same predictions as Pesetskys. Most important, the elimination
of rule-ordering and cyclicity removes the need for postulating a semantically
counter-intuitive structure for prefixed forms, and thereby solves the bracketing
paradox identified by Pesetsky. The correct structure for these forms is simply that
demanded by the semantics.
3.4 A heuristic for deciding between multiple morphological analyses
As more rules are added to the grammar, some strings come increasingly
frequently to have more than one possible parse. This is a common problem in
morphological parsing, noted by Church (1986: 156) and Baayen and Sproat (1996).
For example, Baayen and Sproat write:
Syncretism and related morphological ambiguities present a problem for many NL
[natural language PC] applications where lexical disambiguation is important;
cases where the orthographic form is identical but the pronunciations of the various
functions differ are particularly important for speech applications, such as text-to-
speech, since appropriate word pronunciations must be computed from orthographic
forms that underspecify the necessary information. (p. 156)
203
I shall take an example from the current CFG to illustrate this point. For this, we shall
return to the rules from (104) and (114) relating to nouns and noun-stems. We also
assume pre-terminal rules to account for the morphemes !+%+!
rn
as in +&+& !+%+%!
mummy, !$6!
c
as in #,":&.$ !$671C%&'! approach, and !O!
in
as in -.,6 !#&6-! table.
The relevant rules are listed in (125).
(125) noun noun-stem in
noun-stem prefix rn
noun-stem rn
rn !+%+!
c !$6!
in !O!
Now suppose we wish to parse the string #,+&+ !$6+%+! houses (dat. pl.) using
(125) together with the appropriate pre-terminal rules. Not just one, but two parses
will be found: these are shown in (126).
(126) (a) !$6+
rn
+%+
in
! houses-[dat. pl.]
(b) !$6
c
++%+
rn
+O
in
! to-mothers-[gen. pl.]
Clearly, multiple parses present a practical obstacle to checking the accuracy
of the grammar. Success in parsing a word does not equate with correctness in parsing
it, and one would want to avoid a situation where the grammar parses a string in
several ways, but only one parse is correct.
On the other hand, the fact that the grammar produces multiple parses need not
be seen as a weakness. For example, the English word lighter can be a deverbal
noun denoting an object used for lighting cigarettes, or a comparative adjective
204
meaning less dark, and in this case it would be a weakness of the grammar if it were
not able to parse this string in two different ways.
It might be argued that the case of !$6+%+! is unlike that of lighter: one may
object to the parse in (126b) on the grounds that it is spurious. This objection is not
unreasonable, but it needs to be supported by a clear statement of why (126b) is not a
possible parsing of !$6+%+!. One line of reasoning, for example, might run as
follows: the CFG of Russian morphology should include only rules which are
productive in modern Russian, since it seems unlikely that the human brain would
parse words on the basis of non-productive rules. While !96
c
+#;$
rn
! might have been
a valid morphological parsing according to word-formation grammar rules which
were productive at some previous stage in Russian, in contemporary Russian !96#;$!
has now become lexicalized as a single morph. Thus the second rule in (125) should
be dispensed with in a grammar of modern Russian morphology.
Adopting this theoretical stance would involve dispensing with the
commitment to maximum morphological segmentation which underpins Oliveriuss
approach; his morpheme inventory would have to be restructured to include, for
example, !96#;$! as a morph in its own right. This move would bring the morphology
more into line with the word-based view that the human brain prefers to store related
lexical items separately and in full, rather than storing one item, then deriving other
items from it (see for example Bybee 1988). However, one should be careful not to
equate morphological parsing with a statement of what occurs when the human brain
recognizes words:
One should not suppose that all utterances are constructed by speakers on the basis of
words and syntactic rules; even less should one suppose that speakers construct
utterances on the basis of morphemes and two levels of syntax (one governing the
formation of words from morphemes, and one the formation of sentences from
205
words). The speakers of a language can remember not only whole words (cf. the
established tradition of dictionaries), but also whole sentences (cf. the tradition of
phrase-books). [my translation PC] (Oliverius 1976: 13).
In any case, there is an alternative to striving to ensure that each string can be
parsed only in plausible ways. If morphological productivity is a question of degree,
then it should follow that the acceptability of different parses is also a question of
degree. Therefore, the best approach may be to accept that a word such as !$6+%+!
does indeed have more than one possible parse, but the different possible parses
should be ranked in terms of their acceptability or lexical prior probability (LPP)
(Baayen and Sproat 1996: 156), and in the absence of other constraints (such as
syntactic constraints) the most probable parse should be chosen as the correct one.
This is precisely the approach that Church (1986) takes in his morphological parsing
system for English:
Costs were placed on the arcs to alleviate overgeneration. Note that the grammar
produces quite a number of spurious analyses. For example, not only would formally
be analyzed as form-al-ly but it would also be analyzed as form-ally and for-mal-ly.
The cost mechanism blocks these spurious analyses by assigning compounding a
higher cost than suffixation and therefore favoring the desired analysis. (p. 156)
I also implemented a heuristic for assigning costs to competing analyses, but my
approach differs somewhat from Churchs: the former assigns costs to affixes, while
the latter assigns costs to particular rules.
3.4.1 Assigning costs to competing analyses
A reasonable assumption might be that a morphological analysis which relies
on low-frequency morphemes should be assigned a higher cost than a competing
analysis based on high-frequency forms. Oliverius (1976) provides a ready measure of
morpheme frequency (which I shall call Freq(M)): this is defined as the sum of the
206
word-frequencies of all the words in which that morpheme occurs. To be more
specific, using Freq(M), we would estimate the probability of a particular morph M
x
(i.e. the probability of selecting a particular morph M
x
from the complete set of
morphs M
1
, M
2
, , M
n
) as the expression in (127):
(127) Freq(M
x
)
P(M
x
) =
Freq(M
1
)+Freq(M
2
)++Freq(M
n
)
For a given corpus, the denominator of the fraction in (127) will be a constant: it will
simply be the number of tokens (morphs) in the word-list. For the word-list in
Oliverius (1976), this constant is 1,318,181. In order to manage very low values of
P(M
x
), the log of P(M
x
) can be used instead.
Continuing this reasoning further, we would define the cost of a given parse as
the product of P(M
x
) for each of the morphs represented in the parse, or alternatively
E(log P(M
x
)). Because P(M
x
) is a measure of probability, it is always between 0 and 1.
A number of facts follow from this. First, log(P(M
x
)) must be a negative number;
secondly, so must E(log P(M
x
)); and third, the greater the likelihood of the overall
parse, the closer the value of E(log P(M
x
)) associated with that parse will be to zero.
Of all the competing parses, the one with the lowest cost is chosen as the one most
likely to be correct. This approach is exactly the same as that used for deciding
between multiple syllabifications, as described in 2.4: the only difference is that in 2.4
it was necessary to calculate the relative frequencies of application of pre-terminal
rules in the syllable structure grammar from first principles, whereas with the pre-
terminal rules of the word-formation grammar this information is already given in
Oliverius (1976).
207
As an example, I have given the parse tree for -4"7)&6$',-.$
!#9'1&#+,7+%-'+4+6#&'+! speciality and included the appropriate values of log(P(M
x
))
(see Figure 20). As can be seen in Figure 20, just as each symbol within the parse tree
describes a partial grammatical structure, each partial grammatical structure can be
said to have its own value of log(P(M
x
)).
Returning to (126), we can now evaluate which of the two alternative parses is
more probable. This is done in (128), where we see that basing acceptability on
morpheme frequency in this case yields the result we had hoped for.
(128) Parse Log probability
!$6+
rn
+%+
in
! (J3232) + (J1412) = (J4645)
!$6
c
++%+
rn
+
in
! (J2959) + (J3665) + (J1365) = (J7989)
Using this heuristic to choose between parses, we have in principle resolved
the issue raised above, that of checking the grammar when multiple parses arise. Of
all the possible parses, that with the lowest cost is chosen. It is extremely unlikely that
two possible analyses would have the same likelihood when measured in this way, but
if this were the case, either parse could be chosen. The fact that the cost of a parse is a
function of the product of the probabilities of the individual morphs means both that
simpler (shorter) parses will be favoured, and that parses with more frequently-
occurring morphs will be favoured.
208
Figure 20. Parse tree for -4"7)&6$',-.$ (with log probabilities)
noun
-16810
noun-stem
-15445
infl. ending
-1365
adj-stem
-12109
sn
-3336
adj-stem
-10110
sa
-1999
noun-stem
-6575
sa
-3535
rv
-4009
sn
-2566
sp'cts
-4009
ij
-2566
al'
-3535
n
-1999
ost'
-3336 -1365
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 MLE
36
, 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. 156-
157)
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 high-
frequency ranges, thus making the overall MLE a less reliable predictor of the
properties of the low-frequency and unseen cases (p. 163).
Baayen and Sproats 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(M
x
) may not
be the best basis on which to compute the cost of a given parse. This is because P(M
x
)
is computed from the entire lexicon, not just low-frequency words. In other words,
Baayen and Sproats results might lead us to believe that a parser based on P(M
x
)
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
!1
c
+&
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) ~ J2462. 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 J2554. (!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
+1
svr
+&'
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(M
x
) to measure the LPPs of competing parses does not appear to
bear out Baayen and Sproats results: I found no clear examples of wrong parses of
low-frequency words which were attributable to the inclusion of morphs with high
P(M
x
) but low productivity. In fact, the opposite appeared to be the case: high-
frequency words were parsed incorrectly using productive affixation rules which
generally apply to low-frequency words. For example, +/ !+,! we is the 16
th
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 Sproats 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.)
(
T
h
e

x
-
a
x
i
s

o
f

t
h
e

g
r
a
p
h

s
h
o
w
s

o
n
l
y

e
v
e
r
y

1
9
t
h

l
e
x
e
m
e
.
)
F r e q u e n c y
214
To summarize, the reason that the CFG based on P(M
x
) 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 Churchs (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 grammars 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 grammars 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) The grammar was unable to assign any parse to the word
(b) The grammars morphological tokenization did not match Oliveriuss
morphological tokenization
(c) The grammars morphological tokenization was not significantly
different from Oliveriuss morphological tokenization
(d) The grammars morphological tokenization matched Oliveriuss
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;2
rn
+,
in10
! does not
significantly differ from !0;2
rn
+,
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 grammars 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 5,216 4620
Parse was not significantly different from Oliverius's 2,291 2029
Subtotal 7,507 6649
Parse did not match Oliverius's 1,978 1752
Parsed words: total 9,485 8401
Unparsed words 1,805 1599
TOTAL 11,290 10000
Figure 22. Analysis of coverage of morphology grammar
Parse was the same as
Oliverius's
46%
Parse did not match
Oliverius's
18%
Unparsed
16%
Parse was not
significantly different
from Oliverius's
20%
218
The performance of the parser could be improved by redefining the morpheme
inventory, reinterpreting complex stems as single morphemes, particularly for high-
frequency stems. A consequence of this move would be an enlargement of the
morpheme inventory. However, it is worth noting that even the existing word-
formation 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 grammars 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. Oliveriuss 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 grammars 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 1599% 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 (9095% 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 (905%), 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 905% of cases that represent
exceptions to Pesetskys 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 Lexical form Gloss
-,("-.$ *-("-.$ !#O+"'1#&'+! conscience
-,<)!&.$-> *-<)!&.$-> !#O+A',0+%+&'+#'%! to gather
-,(". *-(". !#O+"'1&'+! council
-,</.)" *-</.)" !#O+A,&'+,7+1! event
(,9!&-. *(9!&-. !"O*+0%#&'+! age
(,-%6)%'8.$ *(-%6)%'8.$ !"O*+2-',2+4;+&'! to exclaim
All of the starred forms in (130) with a deleted jer are historically attested in Russian
(Barxudarov et al. 1975), and in one case the predicted form still exists, but with a
221
different meaning (-(". society). In contemporary Russian, doublets of the
-,("./-(". type are quite common: in each case, the member of the pair which
exhibits exceptional jer-vocalization is historically a borrowing from Old Church
Slavonic (OCS), a South Slavic dialect, while the other member of the pair is a native
Russian (East Slavic) word. Usually, the two will also be distinguished semantically,
with the former more likely to relate to abstract concepts (e.g. -,</.)" occurrence)
and the latter more likely to belong to the vernacular vocabulary (e.g. -</.)" sale).
These doublets could be accounted for in the current framework by revising
the word-formation grammar so that the start symbol of the grammar is specified as
[OCS], a feature which was proposed in Lightner (1972). This feature would then
percolate down to the prefix, with the grammar rules making explicit the fact that jers
in [+OCS] prefixes vocalize no matter what follows.
A further 72 of the 237 exceptions seem to motivate a rule of epenthesis. In
these cases, the form predicted by the Lexical Phonology theory has a deleted jer, and
an unsyllabifiable consonant cluster. All of the cases which fall unambiguously into
this class involve the suffix !E#&"!. Some examples are given in (131)
(131) Actual form Predicted form Lexical form Gloss
,<1"-.(, *,<1-.(, !6A8&8'+E#&"+6! society
+',*"-.(, *+',*-.(, !+465+E#&"+6! multitude
-81"-.(, *-81-.(, !#;8&8'+E#&"+6! being
What these examples suggest is that it would be wrong to rule epenthesis out
entirely as playing a role in jer vocalization. Although the Lexical Phonology theory
222
of the jers accounts correctly for 9095% of cases, its coverage of the data would be
increased to at least 9370%
37
if a rule were included to vocalize jers where an
unsyllabifiable cluster would otherwise result. Clearly, though, these figures
demonstrate that epenthesis is not as significant a factor in explaining jer deletion and
vocalization as the Lexical Phonology theory.
3.6 Conclusion
In this chapter I have applied the methodology set out in 1.4 to develop a
morphology of Russian which accounts for the formation of particles, non-derived
nouns, derived nouns, verbs (non-reflexive and reflexive) and prefixation. I have not
discussed the formation of adjectives or adverbs, although rules governing these are
presented in the final form of the grammar in Appendix 2. The grammar was
equipped to deal with morphophonological alternations between vowels and zero by
incorporating the features [left:X], [strong], [tense] and [high] into the word-
formation grammar. The grammar, in its final form, meets the original objective of
employing a grammar formalism no more powerful than context-free. In section 3.4, I
introduced the problem of morphological ambiguity and suggested a heuristic for
choosing among competing parses based on the frequency with which morphs used in
competing analyses occur. Finally, in section 3.5 I presented the results of testing the
grammar in Appendix 2 for its ability to assign correct morphological structures to
11,290 Russian words. The grammar was also tested to determine whether it correctly
predicted the vocalization and deletion of jers, and it was found that the Lexical

37
The increase in coverage could be even greater, because some forms like -,9'&')" !#6*4%4,71!
conscience could be accounted for either by epenthesis or by the [OCS] modification to the
grammar proposed above.
223
Phonology theory was, by and large, accurate, although I suggested that the grammar
should also incorporate a separate mechanism to deal with borrowings from OCS and
a small number of cases of epenthesis which the Lexical Phonology account could not
accommodate.
224
Chapter 4: Stress assignment: three existing theories
4.1 Introduction
4.1.1 Two approaches to stress in Russian: the Slavist and the generative
approaches
Russian is traditionally described by textbooks as a language with free word-
stress (e.g. Avanesov 1956). The reason that stress in Russian has been described this
way is that stress may be attached to any syllable of the word, and may surface on
virtually any type of morpheme. However, it is not clear that Russian stress really is
free rather than fixed in the sense used by Tornow (1984: 46ff) and Hayes (1995:
31). Although, as Shapiro (1986) says frankly, Russian stress is a puzzle, it is not
completely arbitrary: Slavist and Western linguists have suspected for a long time that
stress system of Russian appears to be governed by morphological principles, which
places Russian clearly within Hayess (1995: 30-31) typology, but not as a truly free-
stress language.
The work done by Slavists on synchronic and diachronic Russian accentology
is a long-established strand of research in this field. It is impossible to give a
comprehensive list of citations, owing to the proliferation of work in the field;
however, some of the principal landmarks are Stang (1957), Illi;-Svity; (1963), Garde
(1976) and Dybo (1981), historical accounts; Zaliznjak (1967 and 1977), Redkin
(1971), Fedjanina (1976) and Afaneseva (1984), classifications of modern Russian
lexemes according to their stress pattern; and Zaliznjak (1985), a historical and
synchronic account. Concurrently, the attention of mainstream generative linguists
225
following in the tradition of Chomsky and Halle (1968) has also turned to this
perennial problem (Halle 1973, 1975 and 1997, Halle and Kiparsky 1979, Pesetsky
ms 1979, Melvold 1989 and Gladney 1995). It is clear from Halle and Kiparsky
(1981) that this line of research has been influenced in particular by Gardes work.
Finally, in the last few years, Nesset (1994) and Brown, Corbett, Fraser, Hippisley
and Timberlake (1996) have produced work which can be seen as a recasting of
Slavist ideas on Russian stress in a declarative framework. The basic approach of this
work is similar to that of most of the synchronic work by Slavists: lexemes, rather
than morphemes, are classified by their accentual features. It is unfortunate in
particular that the tradition of mainstream generative linguistics is somewhat apart
from the rest of the work mentioned, since each tradition has produced substantial
contributions which have been neglected by the other, to the detriment of the overall
state of research in the field.
The primary strength of the Slavists work rests in its commitment to a
thoroughgoing coverage of the data of Russian: generally, this tradition has produced
more substantial treatments of Russian stress than the generative tradition. Examples
of this are the comprehensive dictionary of Zaliznjak (1977) and his monograph
giving a detailed synchronic and diachronic account of stress in Russian (Zaliznjak
1985); Fedjaninas (1976) account, which classifies approximately 1,500 modern
Russian lexemes with mobile stress by accentual type; and Redkins (1971) account,
which has one index, like Fedjaninas, with a classification of approximately 3,000
non-derived words and another index listing approximately 250 suffixes with their
accentual properties (Redkin 1971: 216-221). Typically, this work has been aimed at
least partially at students of Russian, so an important goal has been to enable the
226
reader to discover facts about the stress of a particular word in which he or she is
interested. Zaliznjaks (1977) dictionary goes the furthest in achieving this, reliably
indicating the stress of approximately 100,000 Russian words plus their inflected
forms. The introduction to the dictionary contains detailed and extensive notes to
complement the information given for each lexical entry, enabling the user to
determine without ambiguity the precise form of any inflection of any
headword, together with its stress. Indeed, this dictionary is a standard source of
information about stress in Russian. In general, it is a reflection of Zaliznjaks
encyclopaedic knowledge of the subject that references to his work appear almost de
rigueur in the bibliography of any work on Russian stress, including the work of
generative linguists such as Halle and Melvold.
A strength of the generative work, on the other hand, is that it makes more
general linguistic claims: an important objective for generative linguists has been to
develop not just language-specific rules which apply to Russian (or any one other
language), but universal rules which apply to all (or as wide as possible a range of)
languages. Thus, in his (1997) account of Russian stress, Halle argues that the stress
system of modern Russian, along with that of Lithuanian and Sanskrit, survives more
or less unchanged from Indo-European. On an even broader level, both Halles and
Melvolds accounts are grounded in parametric metrical theory (Booij 1983, Hayes
1995, and see Dresher and Kaye 1990 for a discussion of a computational learning
algorithm based on parametric metrical theory), which conceives of stress assignment
as being governed by a universal set of parameters whose settings are selected
according to language.
227
While the coverage of Russian generally by the Slavist accounts has been
excellent, they have a number of weaknesses. First, synchronic Slavist accounts have
given little, if any, consideration to the ways in which the stress system of Russian
might be similar to that of other languages. As mentioned above, there is a tendency
in this work to list accentual properties as lexical features of whole lexemes (e.g.
Redkin 1971, Zaliznjak 1977, 1985): however, by definition this kind of approach
cannot be easily transferred to other languages, since many languages do not have
inflection at all, and even if they do, the categories for which lexemes inflect may be
radically different from those of Russian.
Secondly, Slavists have often glossed over the question of how different
modules of grammar integrate. In particular, where the focus has been on classifying
lexemes by accentual types, there has tended to be an impression that aspects of
Russian grammar such as morphological composition are either irrelevant or bear no
relation to stress assignment. To a lesser extent, this criticism can be levelled against
those who have considered the relationship between stress and affixation in Russian
(e.g. Redkin 1971, Zaliznjak 1985, Lagerberg 1992), since by definition such an
approach is morpheme-based and thus requires some consideration of how the
morphological structure of a word affects its stress. However, Zaliznjak (1985: 22)
states that the stress of non-derived words, in contradistinction to that of derived
words, is a lexical idiosyncrasy of each word and cannot be predicted formally
from any other characteristic of the given word. This is clearly opposed to Halles
(1997: 275) view that in modern Russian, as in Indo-European, stress is determined
by morphological structure, even in non-derived words. Zaliznjak therefore limits the
228
role for grammatical integration between morphology and stress assignment in a way
which Halle does not; in this chapter I shall show that Halle was right in this instance.
A third weakness of many Slavist accounts is that little thought is given to the
formal properties of the grammars proposed. For example, one obvious formal
disadvantage of Zaliznjaks (1977) listing stress on a word-by-word basis is that the
size of the lexicon required is much greater than it would have been if accentuation
had been listed on a morpheme-by-morpheme basis, since the number of morphemes
in Russian must be considerably less than the number of words in the language. A
second problem is that Slavists ignore the question of the power of grammar rules.
The potential pitfall here is that theories which are proposed may in fact be
unrestricted rewriting systems, and, as Chomsky (1963: 360) states, if all we can say
about a grammar of a natural language is that it is an unrestricted rewriting system, we
have said nothing of any interest.
A shortcoming of generative accounts, on the other hand, is that they often
rely on only a few illustrative examples to justify grammar rules: unlike, for example,
Zaliznjak (1977) they tend to be neither comprehensive nor detailed in their treatment
of Russian (or any other language). It hardly needs to be said that their treatment of
language-groups or, for that matter, language universals, is even less comprehensive
and detailed. Thus the coverage of generative theories, that is, the extent to which
they account for the data without exceptions, is often very much open to question. In
particular, there is no way of determining the coverage of Halles (1997) or Melvolds
(1989) theories of Russian stress without first knowing the accentuation of all entries
in the morpheme inventory; however, this information is not provided in either of the
two accounts. Clearly, the analyses which list the accentual properties of lexical items
229
on a word-by-word basis, such as Redkin (1971), Zaliznjak (1977), and Brown,
Corbett, Fraser, Hippisley and Timberlake (1996), do not have this weakness, since by
definition their coverage, in percentage terms, is the number of words listed divided
by the number of words in the entire lexicon, multiplied by 100. As far as Russian
stress is concerned, a consequence of the generative theories lack of coverage is that
they are of little practical use in determining the stress placement in a particular word,
since the likelihood is that the word in question will not be listed in the lexicon. Thus
the interest of generative theories remains on a theoretical level; as a practical tool for
stress assignment, they fall short of the standard set by Slavists.
To date, the goal of synthesizing the best features of the two approaches
coverage, constrainedness, and grammatical integration has been elusive. In my
view, only one analysis of Russian stress comes close to achieving it: this is Brown,
Corbett, Fraser, Hippisley and Timberlakes (1996) Network Morphology account. In
many ways, this research is similar to that of Redkin, Zaliznjak and other Slavists: it
achieves 100% correct coverage of 1,536 high-frequency nouns. Although it lists the
accentual properties of words on a lexeme-by-lexeme basis, it achieves economies of
description by identifying correlations between inflectional classes. Brown, Corbett,
Fraser, Hippisley and Timberlakes theory is implemented computationally in DATR
(Evans and Gazdar 1989a, 1989b and 1995). While the theory accounts only for stress
in inflection, Brown, Corbett, Fraser, Hippisley and Timberlake state that it could be
extended to account for stress in derived words also. However, the account relies on a
stem-by-stem (or lexeme-by-lexeme) instead of a morpheme-by-morpheme listing of
accentual properties. The theory thus makes firm predictions only about those
lexemes it lists, with a default mechanism predicting the most likely stress pattern of
230
previously unencountered lexemes. However, like Gladney (1995), I would argue that
an approach which classifies stems rather than morphemes in fact treats the
explicandum as the explicans. The advantage of morpheme-based accounts such as
those of Halle (1997) and Melvold (1989) is that any word analyzable in terms of the
entries in the morpheme inventory can be covered by the theory. Furthermore, with a
morpheme-based account the amount of information which needs to be listed lexically
is substantially reduced. For example, the 2,493 commonly-occurring Russian words
given by Oliverius (1976) can be morphologically parsed using an inventory of
approximately 1,600 constituent morphemes. These figures suggest that for a
realistically-sized lexicon, a stress assignment program which listed the lexical accent
for each morpheme would need at most 64% (1600/2500) of the memory space
required by another program which listed the lexical accent for each word. For these
reasons one would generally wish to base a theory on the smallest possible linguistic
units.
The reason that it has proved so difficult to formulate satisfactory generative
accounts of stress assignment in Russian is that stress assignment involves the
interaction of many different grammatical constraints. But this is precisely the kind of
area that computational linguistics is well-placed to deal with. As noted by Shieber
(1985: 190-193), Grishman (1986: 5) and Bird (1995: 14), computational linguistics
offers a solution to the problems both of lexical data management and of determining
whether complex linguistic theories actually work in the way they are intended to. A
list of morphemes can easily be stored, added to and updated as necessary, even if it
contains thousands of items. Each time the grammar or lexicon is updated, it is a
231
simple matter to re-test the theory. Without the aid of computer technology, however,
this would be impracticable.
The lack of comprehensive lexical information is perhaps the most significant
omission of generative theories to date: indeed, Gladney (1995) remarks upon the fact
that one cannot realistically list the accentual properties of a large number of
morphemes. However, such statements do not take into account the fact that
systematic compilation of this kind of information is greatly facilitated by
computational tools. In this chapter, computation allows us not only to establish the
coverage of the selected theories empirically, but also to provide exactly the kind of
comprehensive list of morphemes together with accentual properties which eluded
Gladney. Theoretically, this could be done for an unlimited number of morphemes: in
Appendix 3, I suggest accentual specifications for 1,094 morphemes and 1,509
allomorphs of Russian which, together with a stress assignment theory based on
Melvold (1989), can account correctly for the stress of 9507% of 9,512 nouns (see
section 5.3). A measure of the extent to which this list is a contribution to Slavic
linguistics is the fact that the longest existing list of this type of which I am aware is
Redkins (1971) index of suffixes: this list contains approximately 250 items. It is not
my intention to defend Melvold (1989) or the accentual specifications I suggest
as the definitive statement of Russian accentology, or to claim that this theory
explains every problem, but I observe that of the existing theories tested, Melvolds
appears preferable on the grounds that it provides the best coverage of the data with
the minimum theoretical complexity. I anticipate that the results which I have
obtained will at least clarify for accentologists which data are problematic, and set a
standard against which future theories can be measured.
232
In Chapter 5, I show that the essential features of Melvolds theory can be
recast within a declarative context-free grammar consisting of two modules operating
in parallel, a morphological module (as described in Chapter 3) and a syllable
structure module (as described in Chapter 2). The fact that Melvolds theory can be
readily implemented within the framework described leads to an important result for
general (rather than Slavic) linguistics: it is shown that there is no difficulty in
principle in integrating an account of Russian stress assignment into a general account
of syllable structure and word-formation.
4.1.2 Aims of this chapter
My aim in this chapter is to select three existing theories of Russian stress and
to conduct an independent and critical examination of them on equal bases to
determine which of them best fits the data of Russian. In all three cases, it is not clear
at the outset what the theorys coverage is. The theories selected are those of
Zaliznjak (1985), Halle (1997) and Melvold (1989). From the results of the
computational tests, I shall conclude that Melvolds theory is superior to the others in
three ways: (a) its stronger claims (Melvolds theory claims to account for a broader
range of words) are empirically justified, (b) it is a simpler theory, and (c) it fits the
data slightly better than Zaliznjaks theory, as the percentage of words for which it
accounts correctly is higher.
As will be shown in 4.3, there are numerous obstacles which must be
overcome before we can even attempt to answer the question of the theories
coverage. In short, a problem with all three theories is that they are insufficiently
explicit to be tested. This does not mean that the theories do not fit the data; it simply
means that unless we substantially add to the theories, they are of little practical use as
233
a means of assigning stress. In particular, the lack of explicitness is manifested in a
failure to specify the accentuation of every relevant morpheme in the morpheme
inventory. Thus an important subsidiary aim of this chapter is to establish what the
accentuations of a sizeable number of Russian morphemes are. Until this aim has been
achieved, we are not interested in identifying exceptions to any of the three theories
piecemeal: this approach could mislead us into modifying a particular theory in ways
which actually decrease its overall coverage. In order to ensure that we are in a
position to make fully-justified modifications to any of the derivational theories under
consideration, we need first to obtain a systematic and comprehensive list of words
which each theory fails to account for. This is where computational tools come into
their own.
Once these obstacles are overcome, however, the results of the tests carried
out suggest that the ideas in Halle (1997), as well as those of Melvold (1989), are
fundamentally sound. Melvolds theory, however, is broader in scope. The ideas from
Melvold (1989) are therefore incorporated into the overall context-free phonology set
out in this thesis: Chapter 5 describes in detail how this is done.
4.2 Three theories of stress assignment
4.2.1 Halle (1997)
Halles (1997) theory, which is a simple derivational, morpheme-based
account of Russian stress, is briefly summarized in (132):
234
(132) Accentual categories
Every morpheme in Russian is lexically specified to be either accented or
unaccented (Halle 1997: 278). The accent of an accented morpheme is
either borne by one of the morphemes constituent vowel segments, or the
accented morpheme is post-accented (in other words, post-accented is a
subcategory of accented).
Every stem in Russian is categorized as either marked or unmarked
with respect to stress assignment.
Phonological constituency of morphemes
Every morpheme in Russian consists phonologically of segments (phonemes)
which are drawn from a limited stock of vowels and consonants. Most of these
segments are uncontroversial and are thus not listed exhaustively by Halle.
The one (controversial) segment which is explicitly mentioned in Halle (1997)
is the jer. Halle views jers as a type of vowel, in accordance with standard
generative phonology (see section 3.3).
Stress assignment rules
Stress assignment in a given word is determined by (1) the accentual
specification and phonological constituency of the words morphemes, and (2)
the accentual specification of the words stem. The following procedure,
according to Halle, is sufficient to determine the stress of a word:
235
Rule 1: If a word contains no accented morphemes, it is stressed on the first
syllable and no further rules apply; otherwise, the leftmost accented
morpheme determines the word-stress
38
.
Rule 2: If the morpheme which determines the word-stress is accented but not
post-accented, then stress is on the constituent segment which bears the
accent; otherwise (if the morpheme is post-accented), stress is on the
first vowel to the right of the morpheme.
Rule 3: If, following the application of rule 2, stress is on a jer, then it is
retracted (moved left) to the first available vowel which is not a jer.
Rule 4: Following the application of rule 3, if the word is plural and its stem is
marked, then the stress is retracted (moved left) by one syllable.
Scope of theory
Halle (1997) does not specify whether the theory is limited in its scope (for
example, he does not state whether the theory applies only to non-derived
nouns).
There are some similarities between the analysis in Halle (1997) and that in
Halle (1973): the ideas that stress is morphological, surfaces on the leftmost lexically
accented vowel or otherwise on the leftmost syllable of a word, and is retracted to pre-
jer position when the stressed jer is deleted are not new as far as Halle is concerned.

38
This is the Basic Accentuation Principle from Kiparsky and Halle (1977: 209).
236
The principal differences (apart from the fact that Halle in the more recent
article is arguing from modern Russian back to a putative origin in Indo-European)
are as follows. First, in Halle (1973) nouns with stress on the ending in all inflected
forms are considered to be accentless. Their final stress is accounted for by the
Oxytone rule which assigns stress to the inflection in words with accentless stems.
Thus Halle (1973) categorizes as unaccented the same stems that are post-accented
in Halle (1997). For nouns where stress alternates between the inflection and the stem,
Halle suggests in the earlier article that the Oxytone rule applies in all cases where the
stress is on the inflection, and that it is blocked in all other case forms: he states (p.
317) that which case forms are marked [Oxy], if any, is an idiosyncratic property of
each noun. This analysis would require the listing of a sizeable amount of
information in the lexicon; perhaps it is because of this that Halle (1997) abandons the
rather arbitrary Blocking rule. Instead, these are the nouns which in Halle (1997) are
considered to have unaccented stems. Indeed, if one resorts to listing the case forms
where stress is or is not retracted as part of the lexical entries of stems, it seems that
there is little advantage in proposing a morpheme-based account of Russian stress at
all: after all, the account of Brown, Corbett, Fraser, Hippisley and Timberlake (1996)
achieves this objective in a much neater way.
Some example derivations are given in Table 21 to demonstrate the operation
of the rules in Halle (1997):
237
Table 21. Derivations of six Russian words in accordance with Halle (1997)
Word
5,: !,#8 5,!,#&: +) %,!,6?:
Meaning town (dat. sg.) towns (instr. pl.) king (dat. sg.)
Underlying forms
39
!B606$+;! !B606$+/%+',! !2606-'/+;!
Rule 1
40
!B/606$+;! !B606$+/%+',! !2606-'/+;!
Rule 2 n/a
!B606$+/%+',! !2606-'+/;!
Rule 3 n/a n/a n/a
Rule 4 n/a n/a n/a
Other rules n/a n/a n/a
Surface forms
!B/606$+;! !B606$+/%+',! !2606-'+/;!
Word
%,!,:6$ %,6<&-8: %,6<&: -&+)
Meaning king (nom. sg.) sausage (acc. sg.) sausage (instr. pl.)
Underlying forms
41
!2606-'/+O! !kolbas'+;! !kolbas'+/%+',!
Rule 1
!2606-'/+O! !kolbas'+;! !kolbas'+/%+',!
Rule 2
!2606-'+/O! !kolbas+/;! !kolbas+/%+',!
Rule 3
!260/6-'+O!
n/a n/a
Rule 4 n/a n/a
!26-A/%#+%+',!
Other rules
!260/6-'+!
n/a n/a
Surface forms
!260/6-'+! !26-A%#+/;! !26-A/%#+%+',!
4.2.2 Melvold (1989)
Melvolds (1989) theory about stress in Russian is more detailed than Halle
(1997). Part of her theory is summarized in (133), which is in the same format as used
in (132).

39
Accentuation is represented by the symbol / ; where this symbol immediately precedes a segment,
that segment is accent-bearing; where it is placed in morpheme-final position, the morpheme is post-
accented.
40
The morpheme which determines stress is underlined.
41
Jers are represented by the symbols O and E. Marked stems are shown in bold.
238
(133) Accentual categories
As far as the categorization of Russian morpheme is concerned, Melvolds
analysis is exactly the same as Halles, except that she introduces the
additional lexical distinction of dominant versus recessive which applies
only to suffixes. This distinction is independent of the accented versus
unaccented distinction.
As in Halle (1997), every stem in Russian is categorized as either
marked or unmarked with respect to stress assignment.
Phonological constituency of morphemes
Melvolds analysis is the same as that in Halle (1997). In particular, she also
acknowledges the existence of jers.
Stress assignment rules
Stress assignment in a given word is determined by the same kind of
considerations as in Halle (1997). In Melvolds theory, however, stress is
cyclic. This means that in derived words stress is first assigned to the
innermost stem; the output of this stress assignment is the input to the next
cycle, and so on. Thus in the following rules stem appears where word
would have appeared in Halle (1997):
Rule 1: If a stem contains no accented or dominant morphemes, it is initially
stressed and no further rules apply; otherwise, if it contains no
dominant morphemes, the leftmost accented morpheme determines the
239
stem-stress; otherwise, the rightmost dominant morpheme determines
the stem-stress.
Rule 2: If the morpheme which determines the stem-stress is unaccented, then
stress is on the initial syllable of the stem and no further rules apply;
otherwise, if the morpheme is accented but not post-accented, then
stress is on the constituent segment which bears the accent; otherwise
(if the morpheme is post-accented), stress is on the first vowel to the
right of the morpheme.
Rule 3: If, following the application of rule 2, stress is on a jer, then it is
retracted (moved left) to the first available vowel which is not a jer;
otherwise, the stress remains where it is.
Rule 4: Following the application of rule 3, if the stem is plural or belongs to
certain other grammatical categories (Melvold 1989: 26) and its stem is
marked, then the stress is retracted (moved left) by one syllable.
Scope of theory
Nouns (derived and non-derived, in direct and oblique cases).
Comparison of (132) and (133) makes clear that Melvold (1989), in a sense,
can be seen as an extension to Halle (1997): all the rules in (132) are represented in
(133), but not vice versa. In particular, the introduction of the dominant/recessive
distinction and associated rules, as well as the fact that Melvolds rules can apply
cyclically, means that Melvolds theory has means of explaining surface stress which
are not available in Halle (1997). Halle does not explicitly state the scope of the rules
240
given in his 1997 article, but if we are to believe Melvold (1989), these rules are
limited in their application to words which contain at most one derivational suffix;
this is because in words which contain two or more derivational suffixes, the
accentual properties of the root would always be obscured by the effects of cyclicity.
Furthermore, where words contain exactly one derivational suffix, Halles theory
could apply only if the suffix were recessive. If these proposals are correct and
Melvolds extra machinery is genuinely needed to account for words containing
derivational suffixes, then one would in general expect Halles theory to cope less
well with derived words than Melvolds theory. In section 4.4, I shall present
evidence which confirms this expectation.
Some further miscellaneous differences between Melvold (1989) and Halle
(1997) are as follows:
(a) Stress retraction can occur in grammatical categories other than the plural
(Melvold 1989: 26)
(b) Melvold explicitly allows for the possibility that stress retraction may occur
even when the stem is not post-accented (ibid., pp. 26-27). Halle (1997) does
not allow for this possibility (op. cit., p. 283).
(c) Melvold accounts for facultative variants of stress in plural short form
adjectives such as +)6/ kind, phonologically either !+',-/,! or !+'/,-,!, by
stipulating that there are two alternants of the plural inflection and that only
241
some lexically marked adjective stems take the accented alternant. Halle
(1997) makes no reference to facultative variants.
(d) Melvold accounts for short form adjectives where there are facultative variants
in both the plural and the neuter singular, such as <"6, white (neuter sg.),
<"6/ white (pl.), phonologically either !A'1-/6!, !A'1-/,! or !A'/1-6!, !A'/1-,!, by
stipulating that certain lexically marked post-accenting adjective stems have
optional stress retraction in the neuter singular and plural.
(e) All long-form adjectival inflections are assumed to be accented by Melvold.
(f) Patterns such as :,!,B,: !C6068+/6! good (neut short form), :,!,:B)0
!C60/68+,7! good (nom. masc. sg. long form) are accounted for by stipulating
that some lexically marked adjective stems undergo stress retraction in the
long form.
It follows from the observations made above that all the words in Table 21
would be derived in exactly the same way by Melvold as by Halle. Table 22,
however, gives examples of words whose derivation is possible only by virtue of
Melvolds rules:
242
Table 22.
42
Derivations of five Russian words in accordance with Melvold (1989)
Word
8')("!-).":.8 8')("!-).":.-&+) #8!&%8:
Meaning university (dat. sg.) universities (instr. pl.) fool (dat. sg.)
Underlying forms
!;4',"'/10#'+!"#$%"+;! !;4',"'/10#'+!"#$%"+/%+',! !$;0+%2/+;!
Rule 1
!;4',"'10#'+,&'/1&! !;4',"'10#'+,&'/1&! !$;0+%2/!
Rule 2
!;4',"'10#'+,&'/1&! !;4',"'10#'+,&'/1&! !$;0+%2/!
Rule 3 n/a n/a n/a
Cycle 1
Rule 4 n/a n/a n/a
Rule 1
!;4',"'10#',&'/1&+;! !;4',"'10#',&'/1&+/%+',! !$;0%2/+;!
Rule 2
!;4',"'10#',&'/1&+;! !;4',"'10#',&'/1&+%+',! !$;0%2+/;!
Rule 3 n/a n/a n/a
Cycle 2
Rule 4 n/a n/a n/a
Other rules n/a n/a n/a
Surface forms
!;4',"'10#',&'/1&+;! !;4',"'10#',&'/1&+%+',! !$;0%2+/;!
Word
+,:6,#,-.>+) ,:<,!,."'$
Meaning youth (instr. pl.) werewolf (nom. sg.)
Underlying forms
!+6-6$+6#&'+/%+',! !6A60/6&'+%&#+O!
Rule 1
!+/6-6$+6#&'! !6A60/6&'+14'!
Rule 2 n/a
!/6A606&'+14'!
Rule 3 n/a n/a
Cycle 1
Rule 4 n/a n/a
Rule 1
!+/6-6$6#&'+/%+',! !/6A606&'14'+O!
Rule 2
!+/6-6$6#&'+%+',! !/6A606&'14'+O!
Rule 3 n/a n/a
Cycle 2
Rule 4 n/a n/a
Other rules n/a
!/6A606&'14'+!
Surface forms
!+/6-6$6#&'+%+',! !/6A606&'14'+!
As Melvold (1989: 48) notes, the cyclicity of the rules is the device which
guarantees that the stress of derived words (words which include a derivational suffix)
will, in most cases, be on the same syllable for all inflected forms (see also Zaliznjak
1985: 30, Lagerberg 1992: 11), even if the stem contains only unaccented morphemes

42
Symbols are as in Table 21; additionally, dominant suffixes are italicized.
243
and the inflectional ending is inherently accented (cf +,:6,#,-.$ !+/6-6$+6#&'+!
youth [nom. sg.], +,: 6,#,-.>+) !+/6-6$+6#&'+%+',! youths [instr. pl.]).
Essentially, Melvolds use of cyclicity develops the framework of Halle and
Vergnaud (1987). One of the postulates of this framework is the Stress Erasure
Convention (SEC), which states that cyclic affixes obliterate stresses assigned on
earlier passes through the cyclic rules. Halle and Vergnaud (1987: 97-98) assume that
examples such as !;4',"'10#',&'/1&+%+',!, where !%+',! is assumed to be a cyclic affix,
indicate that Russian is exceptional to the SEC; if it were not, the accented
inflectional affix !%+',! would obliterate the stress assigned in Cycle 1 to
!;4',"'10#',&'/1&!). Halle and Vergnaud attribute the exceptionality of Russian to the
existence of a special stress copy rule in the language. However, as they admit, this
does not explain the behaviour of suffixes such as !14'! in !/6A606&'14'+!. The
difference between Melvolds account and Halle and Vergnauds is that according to
Melvold, dominant (rather than cyclic) suffixes erase stress assigned on previous
cycles. This allows Melvold to account for all the examples above, including the ones
which are problematic for Halle and Vergnaud.
It should be noted that Harris (1989) criticizes Halle and Vergnauds use of
the SEC on the basis of the way in which they apply it to explain cyclic stress deletion
in Spanish words:
The account proposed in H&V (1987) attributes [the stress deletion step] to the SEC,
whereas the one proposed here assigns it to the Main Stress Rule. The present
account thus makes it possible to sustain H&Vs claim that an independent rule,
whose sole function is to delete stress, is unnecessary in the grammar of Spanish. On
the other hand, this analysis renders the SEC totally otiose; stress deletion in Spanish
words is tied directly to the basic idea in H&V (1987) that stress contours are
manifestations of the organization of metrical entities into headed constituents:
reapplication of the Main Stress Rule results in the merger of two such constituents
into one, with the concomitant loss of one head that is, automatic erasure of one of
two stresses. (p. 360)
244
For now, we shall do no more than note Harriss objections to the SEC. In Chapter 5,
however, I shall return to the issue of stress erasure (which might appear to be
problematic for declarative grammar in any case, since it involves destructive
modification) and show how the basic idea of stress erasure by dominant affixes can
be implemented in context-free grammar non-destructively and in a way very
reminiscent of the organization of metrical entities into headed constituents.
4.2.3 Zaliznjak (1985)
Zaliznjaks (1985) theory about stress for derived nouns has many of the same
characteristics as that of Halles and Melvolds theories: not only are the rules
themselves similar, but his theory involves lexically determined accentual
specifications for morphemes. Furthermore, Zaliznjaks basic rules are at a similar
level of detail to those of Halle and Melvold. Zaliznjaks (1985: 37-38) basic rules
are set out in (134), again in the same format as (132) and (133).
(134) Accentual categories
Every morpheme falls into one of the following categories:
(recessive pre-accented)
+ (recessive accented)
(recessive post-accented)
D (dominant pre-accented)
+D (dominant accented)
D (dominant post-accented)
As in Halle (1997) and Melvold (1989), if a morpheme is accented (classified
+), then the accent is borne by a particular vowel in the morpheme.
245
Zaliznjak does not restrict dominance to suffixes, as Melvold (1989)
does:
A strong stem in which the root has acquired the property of dominance is itself
dominant The following are examples of dominant stems: #$%&''- !4%0/6$!
people, #$%&' '#-() !4%0/6$4+,7! national, #$%&''#*+,-./-&
!4%0/6$4,&8'1#&"+6! populism, &0#$%&''&/$-.1 !6A4%0/6$6"%+&'! promulgate,
0,2/%,''#-() !A'1*"0'/1$4+,7! harmless, &0,2/%,' '*-.1 !6A'1*"0'/1$',+&'! to
render harmless, 3#$/&'2*-.1 !;4%"/6*',+&'! to manure, #$4%&5$' 2*-.1
!4%9062/%*',+&'! to play pranks. [my translation PC] (Zaliznjak 1985: 53)
Finally, in Zaliznjak (1985), there is no accentual categorization of
stems as there is in Halle (1997) and Melvold (1989).
Phonological constituency of morphemes
Zaliznjak (1985) does not acknowledge the existence of jers. Otherwise, his
theory is similar in terms of the phonological constituency of morphemes to
that of Halle (1997) and Melvold (1989).
Stress assignment rules
Rule 1: If a word contains one or more dominant morphemes, stress is
determined by the rightmost dominant morpheme; otherwise, if the
word contains one or more morphemes classified as or +, stress is
determined by the leftmost of these morphemes; otherwise, stress is
determined by the rightmost morpheme (which must be post-accented).
Rule 2: If the morpheme which determines stress is classified , then stress is
attached to the following morpheme; if the morpheme which
determines stress is classified , then stress is attached to the
246
preceding morpheme. (Although Zaliznjak does not say so, we may
assume that if the morpheme which determines stress is classified +,
then stress is attached to that morpheme.)
Rule 3: If the morpheme to which stress is attached is classified +, then stress
falls on the vowel which bears the accent; if the morpheme to which
stress is attached is classified , then stress falls on the syllable
immediately to the left of the morpheme; if the morpheme to which
stress is attached is classified , the stress falls on the syllable
immediately to the right of the morpheme.
Scope of theory
Derived words in the nominative singular only (Zaliznjak 1985: 29).
The following differences between Zaliznjaks theory and Melvolds theory
thus emerge. First, Zaliznjak does not allow for the possibility that morphemes are
unaccented, while Melvold and Halle do. Secondly, the category pre-accented which
is included by Zaliznjak, is not recognized by either Halle or Melvold.
However, the difference between the theories which is the most significant is
the distinction between the morpheme which determines stress and the morpheme to
which stress is attached, a distinction which does not exist in either Halle (1997) or
Melvold (1989). The importance of this difference can be seen by considering the
following example given by Zaliznjak. He claims that the explanation for the stress in
3"!.2*)% !&8'10&'+/65+,2! small sketch is that !&8'10&'! is post-accented, !65! is
dominant post-accented, and !,2! is pre-accented, as in (135).
247
(135) &8'10&' 65 ,2
D
Because !65! is the rightmost dominant morpheme, it determines the stress placement.
But because it is also post-accented, stress is attached to the following morpheme.
According to Halle and Melvold, the accentuation of !,2! at this point would be
irrelevant: stress would simply surface on the first vowel to the right of !65!. But
because of Zaliznjaks final rule, stress surfaces on the syllable to the left of !,2!, since
!,2! is pre-accented. In other words, the stress returns to !65!.
As an aside, it is worth noting that the analysis of !&8'10&'+/65+,2! is
problematic for Melvolds theory. The basis on which Zaliznjak (1985) classifies !65!
as dominant post-accented is presumably the existence of forms such as 3"!.2*
!&8'10&+/65! sketch, where stress is on the initial syllable of the inflection for all
inflected forms, e.g. 3"!."*&: !&8'10&+65+/%! sketch (gen. sg.), 3"!."*8: !&8'10&+65+/;!
sketch (dat. sg.). The only possible analysis of !&8'10&+/65!, !&8'10&+65+/%!, and
!&8'10&+65+/;! within Melvolds theory would be to stipulate, as Zaliznjak does, that
!65! is dominant post-accented. However, the rules in (133) then cannot provide an
analysis of !&8'10&'+/65+,2!. This, however, is not the place to pursue this point further:
as stated above, we are not interested in identifying exceptions to any of the three
theories piecemeal.
The fact that Zaliznjaks theory is unnecessarily complex is the main objection
which can be raised against it. An additional point, as we have seen, is that Zaliznjak
states that his theory is applicable only to derived words and Melvolds theory claims
to account for both derived and non-derived words. Thus Melvolds theory makes the
248
stronger claim. Of course, this is a shortcoming of Zaliznjaks theory only if it does
not hold for both derived and non-derived words, whatever Zaliznjak claims. We shall
see in 4.4.3-4.4.5 that Melvolds claim is generally borne out by the data, but I do not
present any results which confirm or refute Zaliznjaks statement about the scope of
his theory. Thus the question of whether Melvolds theory does have a greater scope
remains open, and the principal proven advantage that Melvolds theory has over
Zaliznjaks is its greater simplicity.
4.3 Derivational theories and underdeterminacy
Derivational grammar rules are usually thought of as subjecting underlying
forms to various rewrite operations which result in generation of a surface form. Both
the rules and the underlying representations are integral elements in a grammar, since
the existence of any surface form relies on both a corresponding underlying
representation and the rules which link the two together; with both the rules and the
underlying representations it should be possible to test a derivational theory
empirically.
However, generative grammar often glosses over the question of what the
underlying forms are. Rules are often developed in considerable detail, but typically
only a few illustrative underlying representations are given as proof of the rules
validity. No doubt this is because the process of compiling such lists is more tedious
than the formulation of general rules, which can be done comparatively quickly.
Indeed, for a realistically-sized morpheme inventory, compiling comprehensive lists
of underlying representations by hand may well be impracticable. Yet if a list of
underlying representations does not accompany the rules, the grammar can never be
properly tested: the grammar is underdetermined.
249
Although Zaliznjaks account gives the most detail, listing over a hundred
morphemes in an appendix, all three theories presented in 4.2 have this shortcoming:
stress assignment rules are discussed in some detail, but the underlying forms of
comparatively few morphemes are specified. Thus these theories cannot be tested on
a large corpus as they stand: the question of how well they fit the data on a large scale
therefore remains open.
Since we are interested in obtaining an answer to the question of the respective
coverage of different theories of Russian stress assignment, we need somehow to
obtain, for each theory and dataset to be tested, the underlying form for each
morpheme represented in the given dataset. In other words, we need to build up a
morpheme inventory with fully-specified underlying forms which best fits both the
theory and the data. It goes without saying that the specification of each underlying
form must be consistent with the given theory. In sections 4.3.1-4.3.3 I consider
whether there is any principled way of obtaining this information.
The first method, brute force, involves specification of the morpheme
inventory in every possible way to determine which specification is optimal. This
method, however, has to be rejected: the solution space is so large that solution of the
problem in this way is not computationally feasible. However, the discussion in 4.3.1
serves at least to demonstrate the scale of the problem.
The second, computational learning (or inference) of underlying
accentuations, is feasible and well-founded in logic. It is also useful in that it can
provide conclusive evidence that a given theory is problematic (i.e., inconsistent).
However, as soon as a theory is shown to be problematic, one can no longer obtain
reliable information about underlying forms for that theory by logical inference. In
250
any case, even if the theories under consideration were unproblematic, it is highly
unlikely that any of them would be sufficiently constrained such that logical inference
would allow the accentuations of all the entries in the morpheme inventory to be
uniquely determined. In summary, for the theories under review, all that logical
inference can reliably tell us is when their coverage is less than 100%.
The third method involves the stipulation of defaults to achieve full
specification of the morpheme inventory. When the inventory is specified in this way,
one at least has a basis on which to test a given derivational theory and thus measure
its coverage. The stipulation of defaults effectively involves adopting an additional
hypothesis, and is thus more than just a method of obtaining underlying forms.
Indeed, it might well be argued that the (unspoken) hypothesis of Halle (1997) and
Melvold (1989), and possibly also of Zaliznjak (1985), is that certain default
accentuations are stipulated. However, to say this runs the risk of attributing to these
theories ideas which Halle, Melvold or Zaliznjak might wish to disown. However, it
turns out (as I shall show in 4.3.3) that the adoption of this hypothesis as a firm
principle means that the fit of the theory to the data is suboptimal: a better fit can be
obtained if one effectively abandons the hypothesis and determines the optimal
accentual specification of the morpheme inventory for a given theory by trial and
error.
In a sense it is immaterial how we obtain the underlying forms which fit a
given initial set of rules. However, the different methods of establishing the accentual
properties of the morpheme inventory are of interest not just because their output is a
more fully-determined theory. They are also of interest because they shed light on
other aspects of the theories: first, the scale of their underdetermination; secondly,
251
cases which are problematic for all three theories; and thirdly, that while the
hypothesis of default underlying forms can be adopted to solve the problem of
underdeterminacy, it will not solve this problem optimally.
4.3.1 Computing underlying accentuations by brute force
In this section, I use the term morpheme inventory to mean a set of
morphemes with underlying forms which are accentually fully-specified. The length
of the morpheme inventory is defined as being the number of entries in the morpheme
inventory, and the lexical solution space (LSS) is defined as the number of possible
morpheme inventories, given a list of accentually unspecified morphemes. Thus the
lists in (136) and (137) are both of length 2, and, moreover, contain the same
morphemes, but are distinct morpheme inventories, because the accentual
specifications differ in at least one morpheme: !$"60/!
rn
in (136) versus !$"60!
rn
in
(137).
(136) !$"60/!
rn
!$/6&8'!
rn
(137) !$"60!
rn
!$/6&8'!
rn
Assuming any morpheme can be either accented, unaccented or postaccented, the LSS
of the list consisting of the (accentually unspecified) morphemes !$"60!
rn
and<!$6&8'!
rn
would be 9 (3 3, or 3
n
where n = 2 [the length of the list]). The morpheme inventory
in Oliverius (1976) has a length of 1,759, so for Halles theory and this dataset the
LSS could be estimated as 3
1,759
( ~ 1804 10
839
). This figure is greater than the
number of atoms in the universe (let alone the number of bytes of memory space
252
available in the universe), and for this reason it is computationally impossible to
compare the results of using every possible morpheme inventory when large datasets
are involved.
4.3.2 Backwards phonology and the Accent Learning Algorithm
Instead of attempting to compute the underlying representations for the
morphological inventory by brute force, one could, in principle, use a given theorys
rules to reconstruct or learn all the possible underlying accentuations from the
surface forms, a process described by Bear (1990) as backwards phonology (cf. also
Leben and Robinson 1976):
I shall start by making explicit what it means to apply a phonological rule in the
backwards direction. The basic idea is straightforward and will be, I think,
uncontroversial.
a b / % _ | (1)
A rule like the one in (1) transforms the string !%a|! into the string !%b|!. Here % and
| are strings of characters over some alphabet, e.g., the phonemes of a language. I
take it that such a rule can also be interpreted as mapping the string !%b|! into the
string !%a|!, when it is applied backwards.
To take a more linguistically realistic rule, let us consider the simple rule in (2).
4 N / _ B (2)
The rule in (2) says nothing about whether it is optional or obligatory in the
backwards direction. Optionality in the backwards direction is entirely independent
of optionality in the forward direction. (p. 2)
In other words, we know from (2) in the quotation above that any underlying !4B!
must surface as (NB), but, without further information, we cannot assume that any
surface (NB) results from an underlying !4B!: all we know is that a surface (NB) could
be the reflex of an underlying !4B!. Indeed, it would be unusual for a generative
253
grammar to state explicitly whether its rules were obligatory or optional in the
backwards direction.
Halle (1997), Melvold (1989) and Zaliznjak (1985) are no exceptions: under
any given theory, any given underlying form can map to one and only one surface
form, but we do not know whether the rules of the theories are optional or obligatory
in the backwards direction. Thus we need to consider all possible ways in which
underlying forms can map to surface forms. Table 23 illustrates, for a simple example,
the multiplicity of ways in which a surface form may map to underlying forms,
assuming the rules of Halles theory hold. In this example, O represents an underlying
jer and O a zero-morpheme which does not contain an underlying jer. Throughout this
chapter, it is assumed that we have no a priori knowledge of where jers occur in the
lexical representations of morphemes.
254
Table 23. Possible solutions for -.,6 /st'ol/ table (nom. sg.)
Surface form
43
Underlying forms Rules applied
!#&/6-! (a) #&6-<+<O For (a-h):
(b) #&6-<+<O/ !#&6-! is unaccented and the
(c) #&6-<+<O inflection is unaccented or
(d) #&6-<+<O/ postaccented, so stress is on
(e) O#&6-<+<O the initial (and only) syllable.
(f) O#&6-<+<O/ (NB O/ = postaccented O;
(g) O#&6-<+<O O/ = postaccented O.)
(h) O#&6-<+<O/
(i) #&6-/ + O For (i-l):
(j) #&6-/ + O/ !#&6-! is post-accented, but no
(k) O#&6-/ + O vowel or jer follows so stress
(l) O#&6-/ + O/ on initial (and only) syllable
(m) #&/6-<+<O For (m-v):
(n) #&/6-<+<O/ The first full vowel is lexically
(o) #&/6-<+<O accented.
(p) #&/6-<+</O
(q) #&/6-<+<O/
(r) O#&/6-<+<O
(s) O#&/6-<+<O/
(t) O#&/6-<+<O
(u) O#&/6-<+ /O
(v) O#&/6-<+<O/
(w) #&6-/<+ O For (w-ab):
(x) #&6-/<+ /O !#&6-! is post-accented, but stress
(y) #&6-/<+ O/ is retracted as a jer follows
(z) O#&6-/<+ O
(aa) O#&6-/<+ /O
(ab) O#&6-/<+ O/
(ac) #&6-<+ /O For (ac-ad):
(ad) O#&6-<+ /O A lexically accented jer causes
stress to be retracted

43
NB !stol! has a nominative singular inflectional morpheme, which is a zero-morpheme.
255
Thus, for this word which consists of only two morphemes, there are a total of 30
underlying forms, all of which can be mapped to the surface form in a way which is
consistent with Halles rules. For longer words, the number of possible underlying
forms will be much greater. This will apply particularly in cases where the stress
surfaces in an early morpheme, since the rule which states that the leftmost accented
morpheme surfaces as stressed means that without further information one
cannot know whether morphemes to the right of the morpheme which surfaces as
stressed are unaccented, accented or post-accented. The number of possible
underlying forms for a word in which n morphemes follow the morpheme stressed in
the surface form is likely to be at least 3
n
, since any combination of these n
morphemes can be unaccented, accented or post-accented. For some words, therefore,
there may be hundreds of possible underlying forms.
However, what has been said is not the whole truth about Halles theory: the
number of underlying forms for each word will in fact be less than suggested above,
since one word may provide evidence about the underlying form of another. It is in
fact a feature of all the theories under consideration that underlying representations
for one word may have to be ruled out if they are disproved by other words
containing the same morphemes. To take an example, the morpheme !#&6-! appears in
both !#&/6-! (nominative singular) and !#&6-/%! (genitive singular). Assuming Halles
theory holds, the possible underlying forms for !#&6/-%! are listed in Table 24:
256
Table 24. Possible solutions for -.,6& /stol + 'a/ table (gen. sg.)
Surface form Underlying forms Rules applied
!#&6- + /%! (a) #&6-<+</% For (a-b): !#&6-! is unaccented and the
(b) O#&6-<+</% inflectional ending is accented
(c) #&6-/<+ % For (c-h): !#&6-! is post-accented,
(d) #&6-/ + %/ so stress is on the following vowel
(e) #&6-/ + /%
(f) O#&6-/ + %
(g) O#&6-/ + %/
(h) O#&6-/ + /%
None of the possibilities for !#&6-/%! includes accented !#&/6-!: the morpheme !#&6-!
must be either unaccented or post-accented, or any inflected form of !#&6-! would
always be stressed on the stem. Since the underlying accentuation of such morphemes
must be the same no matter which words they occur in, there will be many such cases
where one word provides additional constraints on the admissibility of accentuations
for another. In the general case, if some word W contains some morpheme M for
which M
u
is a candidate accentuation, but if M
u
is not included at least once in the
candidate underlying solutions for W, we can say that M
u
must be eliminated as a
possible underlying representation of M. Accordingly, on the basis of the possible
accentuations listed for !#&6-/%!, one can go back to !#&6-! and eliminate possibilities
(m-v) in Table 23 (all those which include !#&/6-!, an underlying form which is now
known to be untenable).
Following this logic further, %,!,6>: !2606-'/%! king (gen. sg.) and %,!,:6$
!260/6-'! (nom. sg.) both exert constraints on the morpheme !2606-'!: the only
accentuation of this morpheme allowed by Halles theory which would be consistent
with these forms is !2606-'/! (post-accented), which in turn implies that the nominative
257
singular inflectional morpheme contains a jer. This now enables further solutions for
!#&6-! in Table 23 (a-b, e-f, i-l) to be eliminated.
4.3.2.1 A concise encoding of solutions
It was shown in Table 23 and the discussion following it that even with only
five rules in the grammar, the number of backwards solutions for some words can be
very great (in some cases, hundreds or even thousands for a single word). Listing
solutions as we did in Table 23 and Table 24 thus tends to be very cumbersome.
However, there is a more concise way of listing the solutions. Once again, I use
Halles theory to illustrate this. Suppose we have a word W with four morphemes M
1
,
M
2
, M
3
and M
4
, which have the respective structures: CVC, V, CV, and CVCV. Suppose
further that W is stressed on its second syllable, i.e. on the only vowel of M
2
. The
following is an exhaustive list of the valid possibilities within the constraints of
Halles theory:
(138) (a) M
1
is inherently unaccented and M
2
is inherently accented: there are no
constraints on the underlying forms of M
3
or M
4
.
(b) M
1
is inherently post-accented, M
2
does not contain an initial jer, and
there are no constraints on the underlying forms of M
3
or M
4
.
(c) M
1
is inherently unaccented, M
2
is inherently post-accented, M
3
contains an initial jer, and there are no constraints on the underlying
form of M
4
.
258
From (138) we know that W cannot provide any information which will enable any
underlying representations for M
3
or M
4
to be eliminated
44
: the lack of constraints on
M
3
and M
4
means that all the possible underlying representations for these
morphemes will be listed among the solutions for W. W provides no more information
about M
3
and M
4
than a word which contains neither M
3
nor M
4
: this word implies no
constraints on these morphemes, and listing all their possible underlying forms is thus
redundant.
In the list of solutions, we can therefore omit morphemes on which there are
no constraints. Omission thus encodes the fact that all candidate forms for the
morpheme are possible. Using this scheme, the solutions for W can be represented as
follows:
(139) Accentuations Jers
(a) M
1
: CVC, M
2
: /V no constraints
(b) M
1
: CVC/ M
2
: V (no initial jer)
(c) M
1
: CVC, M2: V/ M
3
: JCV (initial jer)
To give a real-life example (again assuming Halles rules hold, for the
purpose of illustration), Table 25 repeats the information of Table 23 (which
represented all the solutions for !#&/6-! ) in abbreviated form:

44
However, if it can be shown by some other word(s) that (a) and (b) of (138) are not valid
possibilities, then the constraint that M
3
must contain an initial jer will come into operation.
259
Table 25. List of solutions for /st'ol/ (revised)
Grammar rule Possible
accentuations
Possible jers
Initial syllable stress
M
1
:<#&6-<or<#&6-/
M
2
: no jer
Inherent accent< M
1
:<#&/6- <
Inherent post-accent with following jer< M
1
:<#&6-/
M
2
: O
Inherently accented jer< M
1
:<#&6- M
2
: /O
In this case, then, the omission of M
2
in the second solution encodes the fact that if
!#&/6-! is the underlying representation of M
1
, then we know nothing about either the
accentuation of M
2
, or whether it contains a jer.
4.3.2.2 Formalization of the Accent Learning Algorithm
At this point, we can formalize a method by which underlying accentuations
can be learnt from the surface forms, given a particular theory about stress, as
follows:
(140) The Accent Learning Algorithm (ALA)
(a) For each morpheme in the dataset, compute all the candidate underlying
representations allowed by the given theory;
(b) For each word in the dataset:
(i) compute the list of solutions possible in the given theory using
backwards phonology;
(ii) eliminate from the list computed in (b) (i) any solutions which
contain an underlying representation not currently included in (a)
(NB: the list in (a) is continuously updated; see (b) (iii) below);
(iii) from the output of (b) (ii) make all possible inferences about
which representations from (a) should be eliminated (i.e. check that
each underlying representation from the relevant morphemes in (a)
is either included at least once in the output of (b) (ii), or omitted
altogether in at least one of the solutions), and update (a)
accordingly.
(c) Repeat (b) until no further eliminations are possible.
260
The ALA goes through the corpus of words, making all possible inferences from each
word in turn. After considering each word, it updates the database of possible
underlying morpheme representations.
Now, suppose that there are two words in the sample, W
1
and W
2
, and W
1
is
processed by the ALA before W
2
. Suppose further that both W
1
and W
2
contain some
morpheme M on which W
2
, but not W
1
, introduces some constraints. While the
underlying representations for M will be updated by the ALA when W
2
is considered,
the possible solutions for W
1
will not. However, it is important that the relevant
solutions for W
1
should be ruled out, since this may allow further constraints relating
to other morphemes in W
1
to come into effect.
A simple method of ensuring that all constraints are eventually applied is to
stipulate that once all words have been considered, the ALA returns to the beginning
of the word-list and re-evaluates all the words. Once the second iteration is complete,
it starts a third iteration, and so on. On each successive iteration, the number of
candidate accentuations per morpheme will shrink. As soon as an entire iteration
passes with no further changes being made, we know that no further inferences can be
made, at which point the ALA can be terminated.
The ALA is no more than a procedure by which all possible deductions can be
made, given a particular theory and dataset. The procedure itself is theory-neutral, and
it is not necessarily the only possible algorithm which could be used to yield the same
results. As the ALA proceeds, there will be convergence of two (related) types: first,
convergence of the number of solutions possible for each word, and secondly,
convergence of the number of possible underlying forms for each morpheme. Thus,
even though Halle, Melvold and Zaliznjak are not explicit in specifying the
261
accentuation of items in the morpheme inventory, we find that for each theory we can
apply logic and use the theory itself to deduce the accentual specification of items in
the morpheme inventory. Another way of looking at this is to say that morphemes
start as being accentually underspecified, but can become more fully specified
depending on the surface stress of words in which they appear: each new word can
imply new constraints on the underlying forms of its constituent morphemes. In view
of this, it is reasonable to wonder whether the theories themselves are sufficient to
imply one underlying form per morpheme, given a large enough corpus of words. If
they are, then the criticism that they are under-determined would not hold.
This is the key question which we would like to answer by implementing the
ALA and using it to process a large sample of words. However, we cannot reliably
obtain an answer to this question where there are cases which are problematic for a
theory, as will become clear in section 4.3.2.4. Moreover, even for Zaliznjaks theory
(which, as will be shown in section 4.3.2.6, need not encounter problematic cases),
logical inference is not sufficient as a means of obtaining a full accentual specification
of the morpheme inventory. The objection that the theories omit an explicit
specification of the accentual properties of the morphemes of Russian is therefore
sustained in all three cases.
4.3.2.3 A small-scale demonstration of the ALA on a non-problem combination
In this section, I demonstrate the use of the ALA on a dataset of 9 words,
which are listed in (141). For the purposes of this demonstration, we again assume
that Halles rules hold. One aim of this section is to show that the ALA makes the
correct inferences in accordance with Halles theory: on a dataset of this size it can
readily be seen that the ALA is producing the desired results.
262
(141)
45
#("!": 0 !$"'10'
rn
+/17
in9
! door (genitive plural)
#(":!) !$"'/10'
rn
+,
in20
! door (genitive singular)
5,!,#,:( !B606$
rv
+
sn1
+/6"
in12
! town (genitive plural)
5,:!,#" !B/606$
rv
+
sn1
+1
in8
! town (locative singular)
%,!,:6$ !260/6-'
rn
+
in1
! king (nominative singular)
%,!,6?: !2606-'
rn
+/;
in22
! king (dative singular)
)5!/: !,B0
rn
+/,
in20
! game (genitive singular)
)5!8: !,B0
rn
+/;
in21
! game (accusative singular)
):5!/ !/,B0
rn
+,
in10
! 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
words 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 Jers Stem
!$"'10'!
rn
, !/17!
in9
unmarked
!$"'10'/!
rn
!17!
in9
: no jer unmarked

45
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 #(": !) !$"'/10',! door (gen. sg.). As the
continuation of the output log shows in (144), two solutions are listed: both of these
rely upon the operation of the rule which stresses the words initial syllable when no
other rules can operate. If no other information were available about how !$"'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) !$"'/10'
rn
+,
in20
! 2
Accentuations Jers Stem
!$"'10'!
rn
, !,!
in20
!$"'10'!
rn
, !,/!
in20
!$"'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
! 3
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !/6"!
in12
unmarked
!B606$!
rv
, !/!
sn1
!6"!
in12
: no jer unmarked
!B606$/!
rv
!!
sn1
: no jer; !6"!
in12
: no jer unmarked
!B/606$!
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
+1
in8
! 2
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !1!
in8
!B606$!
rv
, !!
sn1
, !1/!
in8
!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 Jers Stem
!260/6-'!
rn
!2606-'/!
rn
!O!
in1
46
!2/606-'
rn
! is not a possible accentuation.
!2606-'
rn
! is not a possible accentuation.
!2606-'
rn
+/;
in22
! 1
Accentuations Jers Stem
!2606-'/!
rn
!;!
in22
: no jer
!260/6-'!
rn
is not a possible accentuation.
!,B0
rv
+
sn1
+/,
in20
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in20
: no jer
!/,B0!
rv
is not a possible accentuation.
!,B0!
rv
is not a possible accentuation.
!!
sn1
does not contain a jer.
!,B0
rv
+
sn1
+/;
in21
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !;!
in21
: no jer
!/,B0
rv
+
sn1
+,
in10
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in10
: no jer marked
!,B0
rv
+
sn1
! is a marked stem

46
Jers are represented by the symbol O.
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
+/17
in9
! 1
Accentuations Jers Stem
!$"'10'!
rn
, !/17
in9
! unmarked
!17!
in9
is not a possible accentuation.
!17/!
in9
is not a possible accentuation.
!$"'/10'
rn
+,
in20
! 2
Accentuations Jers Stem
!$"'10'!
rn
, !,!
in20
!$"'10'!
rn
, !,/!
in20
!B606$
rv
+
sn1
+/6"
in12
! 1
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !/6"!
in12
unmarked
!6"!
in12
is not a possible accentuation.
!6"/!
in12
is not a possible accentuation.
!B/606$
rv
+
sn1
+1
in8
! 2
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !1!
in8
!B606$!
rv
, !!
sn1
, !1/!
in8
!260/6-'
rn
+
in1
! 1
Accentuations Jers Stem
!2606-'/!
rn
!O!
in1
!O!
in1
contains a jer.
267
!2606-'
rn
+/;
in22
! 1
Accentuations Jers Stem
!2606-'/!
rn
!;
in22
!: no jer
!,B0
rv
+
sn1
+/,
in20
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in20
: no jer
!,B0
rv
+
sn1
+/;
in21
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !;!
in21
: no jer
!/,B0
rv
+
sn1
+,
in10
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in10
: no jer 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) !$"'10'
rn
+/17
in9
! 1
Accentuations Jers Stem
!$"'10'!
rn
, !/17!
in9
unmarked
!$"'/10'
rn
+,
in20
! 2
Accentuations Jers Stem
!$"'10'!
rn
, !,!
in20
!$"'10'!
rn
, !,/!
in20
!B606$
rv
+
sn1
+/6"
in12
! 1
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !/6"!
in12
unmarked
!B/606$
rv
+
sn1
+1
in8
! 2
Accentuations Jers Stem
!B606$!
rv
, !!
sn1
, !1!
in8
!B606$!
rv
, !!
sn1
, !1/!
in8
!260/6-'
rn
+
in1
! 1
Accentuations Jers Stem
!2606-'/!
rn
!O!
in1
268
!2606-'
rn
+/;
in22
! 1
Accentuations Jers Stem
!2606-'/!
rn
!;!
in22
: no jer
!,B0
rv
+
sn1
+/,
in20
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in20
: no jer
!,B0
rv
+
sn1
+/;
in21
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !;!
in21
: no jer
!/,B0
rv
+
sn1
+,
in10
! 1
Accentuations Jers Stem
!,B0/!
rv
!!
sn1
: no jer; !,!
in10
: no jer 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'/!
!$"'10'!
!17!
in9
!17!
!/17!
!17/!
!/17!
!,!
in20
!,!
!/,!
!,/!
!,!
!,/!
!B606$!
rv
!B606$!
!B/606$!
!B60/6$!
!B606$/!
!B606$!
!!
sn1
!!
!/!
!!
!6"!
in12
!6"!
!/6"!
!6"/!
!/6"!
!1!
in8
!1!
!/1!
!1/!
!1!
!1/!
!2606-'!
rn
!2606-'!
!2/606-'!
!260/6-'!
!2606-'/! !2606-'/!
!!
in1
!!
!/!
!!
!/!
!;!
in22
!;!
!/;!
!;/!
!;!
!/;!
!;/!
!,B0!
rv
!,B0!
!/,B0!
!,B0/! !,B0/!
!;!
in21
!;!
!/;!
!;/!
!;!
!/;!
!;/!
!,!
in10
!,!
!/,!
!,/!
!,!
!/,!
!,/!
270
This simple example suffices to show that even with logical inference, Halles 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 3
x
(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 3
13
, or 159 10
6
. More precisely, for a morpheme inventory where the
number of morphemes with x underlying forms is n
x
, the exact size S of the overall
LSS will be given by (148):
(148)
Clearly S will be zero when n
0
> 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 Halles 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 159 10
6
, is given in (149):
(149) 0
0
1
0
2
2
3
9
4
2
~ 126 10
6
x
n n n
x S = ... 1 0
1 0
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) 0
0
1
7
2
3
3
3
= 216 10
2
The fact that logical inference reduces the LSS size from 1,259,712 to 216
(which is a reduction by a power of approximately 26) 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 Halles (1997) theory, with the following words:
(151) -,7)&:6$'/0 !#6&# + ,7 + /%-' + 4 + ,7! social
(152) -,7)&6)-.): 3"-%)0 !#6&# + ,7 + %-' + ,#&' + /,&8' + 1#2 + ,7! socialist
For both of these words the three initial morphemes are the same:
272
(153) ! #6&# + ,7 + %-'<!
47
On the basis of (151), a backwards application of Halles theory would lead one to
conclude that at least one of the following propositions must be true:
(154) (a) !%-'! is accented
(b) !%-'! is post-accented, and there is a following jer
(c) !,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 nave 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 999%, 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 01% 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 G
1
that
contains a general rule along with certain special rules governing exceptions and a
grammar G
2
that gives up the general rule and lists everything as an exception, then
we will prefer G
1
. 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 Halles (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)
Halle (1997) Melvold (1989) Zaliznjak (1985)
Accented morphemes
are possible
yes yes yes
Post-accented
morphemes are possible
yes yes yes
Unaccented morphemes
are possible
yes yes no
Pre-accented
morphemes are possible
no no yes
Dominant morphemes
are possible
no (all morphemes are
effectively recessive)
yes (suffixes only) yes
Jers can affect stress yes yes no
Features of stem can
affect stress
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 Halles
theory is smaller than the number of ways it can be explained by Melvolds 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,
Halles theory can be said to be more constrained than Melvolds theory (although, as
we shall see in 4.4, the price which is paid for this is that Halles theory is not as well
equipped to deal with derived words).
For instance, with Halles 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 Melvolds 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 Melvolds 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
morphemes
48
), 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 Melvolds
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 Zaliznjaks theory are somewhat different from those of the other
theories: in order to assign stress in Zaliznjaks 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 Zaliznjaks theory has the consequence that
surface stress in a given word will, on average, have many more possible explanations
in Zaliznjaks theory than in Melvolds 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
+/65
sn
+,2
sn
+
in1
!, it turns out that in
Zaliznjaks 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 Melvolds theory implies
49
. Thus two related facts
about Zaliznjaks theory become clear: first, that Zaliznjaks 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 Halles theory for !#&6-! would be 30 (see Table 23).
278
reduce the extent to which logical inferences can be made, and secondly, that
Zaliznjaks theory is considerably less constrained than Melvolds.
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 Halles theory
was used to illustrate this, we clearly also have the specific result from 4.3.2.4 that
Halles 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 Halles theory must be less than 100%.
When I applied the ALA to Melvolds theory using a dataset of 9,512 nouns as
the basis for drawing logical inferences
50
, I obtained the result that this theory is also
problematic. The following example will suffice to demonstrate this.
If we assume Melvolds 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 Melvolds theory is problematic
Word Morphological
tokenization, following
Oliverius (1976)
Inferences
'&3&: 6, beginning !4%
c
+&8/%
rv
+-
sn1
+6
in18
!
!-!
sn1
is not dominant
51
4!&:()6, rule !90/%"'
rv
+,
svr
+-
sn1
+6
in18
!
If !,!
svr
is dominant, it must also
be unaccented
!,#):."6$ parent !06$'
rv
+/,
svr
+&'1-'
sn
+
in1
!
!,!
svr
cannot be dominant
unaccented, so !06$'!
rv
must be
post-accented
!,:#)'& homeland !0/6$'
rv
+,4
sn1
+%
in3
!
!in!
sn1
must be dominant
unaccented
568<)'&: depth !B-;A'
ra
+,4
sn1
+/%
in3
!
568<)'&: is a problem word, as
there is no possible analysis
given the previous inferences.
Thus '&3&: 6,, 4!&:()6,, !,#):."6$, !,:#)'& and 568<)'&: collectively represent
a combination which is problematic for Melvolds theory, and we have therefore
proved that the coverage of Melvolds theory must be less than 100%.
In the case of Zaliznjaks theory, the result is not so clear. Again, the ALA
was applied, this time using a dataset of 524 nominative singular derived nouns
(nouns containing at least one suffix) as the basis for drawing inferences. The reason
for restricting the dataset to nominative singular derived nouns is that this is the stated
scope of Zaliznjaks theory (see 4.2.3). In this case, more explanation is given for
each inference, as the theory is more complex. Assuming Zaliznjaks theory holds,
then the following chain of inferences would be made:

51
If !-!
rn1
were dominant post-accented, the stress would be '&3&6,:; if it were dominant unaccented the
stress would be '&:3&6,; and it cannot be dominant accented as it does not contain a vowel.
T
a
b
l
e

3
0
.

D
e
m
o
n
s
t
r
a
t
i
o
n

t
h
a
t

Z
a
l
i
z
n
j
a
k

s

t
h
e
o
r
y

i
s

p
r
o
b
l
e
m
a
t
i
c
281
The reason that this chain of inferences does not conclusively show that
Zaliznjaks theory is problematic is that it hinges on inferences about the suffixes
!,4!
sn1
and !6&!
sn1
, whose accentual properties, as Zaliznjak (1985: 67, 73) points out,
vary depending on whether they function as deadjectival suffixes (as in %!&-,.&:
!20%#+6&+%! beauty) or deverbal suffixes (as in !&<,:.& !0%A+6&+%! work). In effect,
this is tantamount to claiming that there are two !6&!
sn1
suffixes, which are
distinguished from one another both by meaning and accentuation; this is a claim
which is in opposition to the analysis of Oliverius (1976). Indeed, it turns out that if
we distinguish the two suffixes for example, by relabelling !6&!
sn1
as !6&!
sn2
in all
the cases where !6&!
sn1
has a deadjectival meaning and re-run the ALA on
Zaliznjaks theory, no problem words are found for this dataset. Thus the existence of
problem words, in the case of Zaliznjaks theory, could be attributed to an invalid
morphological tokenization in Oliverius (1976).
For the sake of argument, let us assume this is the case, and move on to
consider the extent to which logical inference helps draw conclusions about the
underlying forms of morphemes which would be consistent with Zaliznjaks theory.
Table 31 shows how many underlying forms were ruled out as a result of using the
ALA:
282
Table 31. Number of candidate accentuations against
number of morphemes before and after operation of ALA
Number of candidate
accentuations x
Number of morphemes
before operation of ALA
Number of morphemes
after operation of ALA
0 0 0
1 0 15
2 0 62
3 0 44
4 72 119
5 0 90
6 463 237
7 0 0
8 54 24
9 0 0
10 3 1
Total number of
morphemes
592 592
The LSS before and after application of logical inference, therefore, is given in
(155) and (156) respectively.
(155) 4
72
6
463
8
54
10
3
~ 2507 10
455
(156) 1
15
2
62
3
44
4
119
5
90
6
237
8
24
10
1
~ 2021 10
390
The conclusion to be drawn from these figures is that even if the ALA does
not disprove Zaliznjaks theory, the scale of underdeterminacy which remains after all
logical inferences have been drawn is so great that the theory is effectively
unverifiable. Note, also, that in this case logical inference has reduced the LSS size by
a power of approximately 12. The corresponding figure derived for Halles theory
from Table 26 was 26 (see page 271). Although the dataset in (141) is different from
the 524-word dataset from which inferences about Zaliznjaks theory were drawn, and
283
one should therefore be cautious about directly comparing Halles and Zaliznjaks
theories based on figures from Table 26 and Table 31, it is nonetheless interesting that
this comparison appears to provide yet more evidence confirming the hypothesis that
Zaliznjaks theory is significantly less constrained than that of Halle.
To summarize the principal conclusions which have been obtained about
Halles, Melvolds and Zaliznjaks theories by using the ALA, then, we have the
following two results:
- The coverage of both Halles and Melvolds theories must be less than 100%.
- The coverage of Zaliznjaks theory could be 100%, but there is no way of proving
this as the theory is too underdetermined.
4.3.3 Unique specification of the morpheme inventory by defaults
In previous sections, I have argued that the theories under consideration are
under-determined. Up to this point, we have implicitly assumed that in each case
where there are many candidate underlying forms which map to a single surface form,
the candidates are all equally likely. In a sense, this is a neutral assumption: it allows
the ALA to make inferences which are as generous as possible, given a particular
theory. However, we now need to consider the possible counter-argument that a
stronger and thus more constrained assumption would be more in the spirit of Halle,
Melvold or Zaliznjak.
It is a generally accepted principle in generative phonology that an underlying
form should be the same as the corresponding surface form unless there is evidence to
suggest otherwise. This general constraint on phonological analysis is embodied in
Kiparskys (1973) alternation condition (hereafter referred to as the AC).
284
Summarizing the discussion in Kiparsky (1973: 18), Cole (1995: 74) explains the
force of the AC as follows:
The Alternation Condition constrains underlying representation in several ways. It
requires the lexical representation of a nonalternating form to be identical to its
surface form (with low-level, automatic phonetic processes factored out); it requires a
single underlying representation for distinct morphemes which are always identical in
surface form; and it requires morphemes which are always distinct in surface form to
have distinct underlying phonological representations.
While neither Halle (1997), Melvold (1989) nor Zaliznjak (1985) makes any
reference to the AC, we need to consider the possibility that their theories assume
these constraints, albeit implicitly.
First, the AC cannot be built in to the ALA, since the whole point of the
ALA is to define an initial LSS large enough that it must include the optimal
morpheme inventory for a given theory (and then, of course, to reduce this solution
space by considering the evidence furnished by words in the language). In the case of
the ALA, the constraints are no more than those which are already integral to Halles,
Melvolds and Zaliznjaks theories. The constraints of the AC, on the other hand, are
additional to those stated in Halle (1997), Melvold (1989) and Zaliznjak (1985). The
ALA and the AC should therefore be seen as separate: they address similar issues of
abstractness and unconstrainedness in underlying representations, but the ALA, as we
know, (a) highlights when a theory is problematic and (b) (when a theory is
unproblematic) provides a measure of the extent to which a theory is under-
determined, while the AC constrains the abstractness of phonological analysis, as
Cole (1995: 73) puts it.
It should also be noted that the AC is clearly inapplicable in practice where the
size of the LSS is zero, as is the case for theories which encounter problem words. As
285
we have seen, the fact that the size of the LSS is zero does not in itself mean that a
theory should be abandoned, as the theory could still have a very high coverage.
Herein is a paradox: if a theory has any coverage at all, there must be underlying
forms to complement the theory, yet if we say that a given theory has an LSS of size
zero, we are saying that there are no underlying forms which complement the theory.
To reconcile these alternatives, we need to acknowledge that for a theory with
problem words logic (as in the ALA) may not be the most appropriate means of
testing the theory; yet logic is needed to ascertain whether the theory has problem
words in the first place. For a problematic theory, we are led to the conclusion that
the AC should be adopted as an alternative to the ALA as the best heuristic for
determining underlying forms.
Where the final LSS is not zero, some form of the AC might be used in
addition to the ALA. For example, it might not be possible to infer anything about a
particular morpheme which is stressed in all instances (in other words, it is non-
alternating) using the ALA, but in this case the AC would presumably lead one to
conclude that the morpheme is underlyingly accented. To take a real-life example,
consider again the information in Table 26. The right-hand column of this table shows
that after the application of logical inference, there are still two candidate underlying
forms for !,!
in20
, !1!
in8
, and !!
in1
. In all three cases, the morphemes are either unaccented
or post-accented. In these cases, it might be argued that the underlying specification
unaccented is more faithful to the AC, since the other option, post-accented,
involves the addition of accentual content to the morpheme. In the same way, it is
reasonable to suppose that the AC would discriminate against allocating jers to
morphemes, and against marked stems. The AC could therefore be used as a basis for
286
eliminating all but one candidate underlying form in every case, reducing the size of
the LSS for any dataset to 1.
However, even assuming the possibility of ranking different candidate
underlying forms in terms of their divergence from the surface form
52
, there are
practical difficulties in deciding how to implement the AC with respect to theories
such as Halles, Melvolds and Zaliznjaks and within the current framework, as I
shall now show.
First, suppose we adopt the following as an implementation of the AC. We
consider the first morpheme in the inventory, M
1
; all words W containing M
1
are
listed. We check whether the underlying form U
1
most closely related to the surface
form would give the correct stress assignment in all W. If U
1
fits, we will then in
general be led to make inferences about the underlying forms of many other
morphemes in W, and there is no guarantee that these underlying forms will be the
least divergent candidates for their respective surface forms. Effectively, the list of
morphemes is like a network, and a decision about the underlying form of one
morpheme will have effects in other parts of the network. If we had considered a
morpheme other than M
1
first, the results for the network might have been different.
Since the number of ways in which we can order a list of morphemes of length L is
L!, we are once again faced with a huge number of different possible results for a

52
In the case of Russian accentology, there are clearly factors other than the AC which constrain
underlying forms. For example, the argument that unaccented is the default underlying form for every
morpheme appears to be refuted by the fact that most (916%, according to Halle 1997: 281) nouns in
Russian have fixed stem stress. If root morphemes were unaccented by default, one would, for
example, expect the majority of feminine non-derived nouns to have mobile stress, as in the inflected
forms of !8%&: . These facts lead to the conclusion that there is an overriding lexical principle in
Russian, independent of the AC, which states that roots are accented.
287
large morpheme inventory: in summary, the result of applying this procedure will
differ depending on the order in which we consider the entries in the morpheme
inventory. This is surely not what Kiparsky, Halle, Melvold or Zaliznjak would have
envisaged: by a process of reductio ad absurdum we are therefore led to the
conclusion that this is not a viable implementation of the AC.
It is also possible to contemplate applying a similar procedure on a word-by-
word basis: using this method, we might suppose that where there is a choice of rules
in a theory which could have explained a given words surface form, one rule is more
likely than another to be used. For example, returning to the example in Table 25, we
could postulate that the surface stress of !#&/6-! is most likely to be explained by the
second rule, the inherent accent rule. But again this decision will force us to make
inferences about the underlying form of the morpheme !#&6-!, which may in turn
constrain the underlying forms of other words. Again, this may happen in such a way
that the least marked rule which could have applied to the other words is ruled out.
This brings us back to exactly the same problem that we faced above: the results of
applying the procedure will depend on the order in which we consider words in the
dataset.
Consider now a third possibility, where we disregard the logical consequences
of each decision about underlying forms. For each morpheme which has multiple
candidate underlying forms after the application of the ALA, we select the underlying
form most faithful to the surface form (and/or lexical principles such as that which
states that roots are accented); this is done without reference to the surface forms of
words which include the morpheme, in order to avoid the network effect referred to
above. However, the effect of disregarding the logical consequences of each decision
288
is that the morpheme inventory obtained will not necessarily be the optimal inventory
for a given set of stress assignment rules (the inventory which gives the rules the
greatest coverage). I shall now provide evidence of this by an illustration. In section
4.3.2.6, a 524-word dataset was the basis on which logical inference was applied to
Zaliznjaks theory. As we know from Table 31, a morpheme inventory of length 592
is required to account for this dataset; as shown above, Zaliznjaks theory is not
problematic when applied to this dataset. Having subjected Zaliznjaks theory to the
process of logical inference, we can now apply the set of preferences in Table 32 to
the outcome.
53
Table 32. Ranking of underlying morpheme forms
Accentuation* Dominance* Morpheme
type 1
st
preference
2
nd
preference
3
rd
preference
4
th
preference
1
st
preference
2
nd
preference
prefix + on final
syllable
+ not D D
root + on final
syllable
+ not D D
suffix + on final
syllable
+ not D D
inflection + on final
syllable
+ not D D
*It is assumed that dominance preferences take precedence over accentuation
preferences, so for example it is deemed more likely that a prefix is than D.
When these defaults are applied to the morpheme inventory, it turns out that
the theorys coverage of the 524-word dataset is 6489%: 340 of the words are

53
I assume this particular set of defaults as it appears to accord with Zaliznjaks stated preferences (e.g.
regarding the accentuation of prefixes see Zaliznjak (1985: 58) and with the observations I have
already made about the proclivity of nouns to fixed stem stress. Unfortunately, the number of
permutations of default ranking in this system is so great that it is impossible to test all potential
rankings: the number of permutations is (8!)
4
~ 264 10
18
.
289
correctly stressed (i.e. 184 were incorrectly stressed). Yet it will be shown in 4.4 that
a different specification of the morpheme inventory results in a much greater coverage
of 9198% (482 words). We must therefore conclude that if the procedure just
outlined is a faithful implementation of the AC, then one of the following four
propositions must be true: (a) the AC must be a suboptimal principle, or (b) we need
to consider a different set of defaults, or (c) the coverage of Zaliznjaks theory is
really only 6489%, or (d) the AC must be incompatible with Zaliznjaks theory.
The fourth and final possible implementation of the AC which we shall
consider is as follows. Suppose we apply defaults to the morpheme inventory as
outlined above to obtain a situation where the LSS is 1. This can be done after using
the ALA, to eliminate residual indeterminacy. This then enables us to test the given
theory, including the accompanying morpheme inventory, to determine its coverage
(which may well be suboptimal). So far, the implementation is exactly the same as
that outlined in the previous paragraph. The difference is that we now inspect the list
of exceptions to the theory the 184 words which were incorrectly stressed, in the
case of Zaliznjaks theory to determine whether there are morphemes which occur
repeatedly in incorrectly stressed words. If we suspect that the incorrect stress of these
words is attributable to an incorrect accentual specification of the morpheme common
to them, we can override the common morphemes accentual specification. Each such
modification of the morpheme inventory may mean that some words in the corpus
which were previously incorrectly stressed now become correctly stressed, although it
is equally possible that some other words which were previously correctly stressed
may now be incorrectly stressed. However, providing the number of words in the
former category exceeds the number in the latter, the modification will result in an
290
increase in coverage. A point will eventually be reached where no further increases in
coverage are possible: at this point, the coverage of the theory will approximate to the
optimal coverage possible for the given theory.
This is in fact the method which I used to specify the morpheme inventory for
the tests described in 4.4; it seems that for the theories under consideration and within
the current framework of backwards phonology, there is in practice no other way of
obtaining the morpheme inventory which best fits a given theory and dataset. It
should be noted, however, that for purely mathematical rather than linguistic
reasons the method is almost guaranteed to succeed, since it is based on a
reanalysis of previous failures. As such, the method is of no linguistic interest in
itself. In a sense this does not matter (cf. Halles 1959: 12 statement that the methods
by which a scientific description is discovered are not of essential concern to the
science in question). However, given that the procedure outlined is a procedure
driven by mathematics rather than linguistics, we need to be careful to acknowledge
that the procedure is not a straightforward implementation of the AC (although it may
embody aspects of the AC): after all, if we were to claim otherwise, we would be
saying that the AC is also of no linguistic interest. Note also that the underlying forms
derived in this way will not necessarily correspond to those one might expect on the
basis of (for example) historical data from Russian.
I conclude this section by stating that none of the possible implementations of
the AC within the current framework appears to yield a morpheme inventory which is
empirically correct (or even close to correct): the only method which was found to
achieve this result relies on an extragrammatical statistical procedure. It seems,
therefore, that in this particular case the AC is not sufficient to provide an optimal
291
solution to the problem of under-determinacy, and that the optimal solution is in fact
non-linguistic. While the AC might well be an appropriate criterion by which to build
a theory rules and morpheme inventory from scratch, the aim of the current
study is slightly different. In order to establish the morpheme inventory in conjunction
with which Halles, Melvolds and Zaliznjaks theories are tested, we are forced to
take their rules as read. The only changes which we are permitted to make are to the
morpheme inventory. It is thus perhaps unsurprising that the AC turns out not to be
appropriate within the framework of backwards phonology.
4.4 Tests to ascertain the coverage of the three theories
All the tests documented in sections 4.4.1-4.4.3 involved the following
procedures. An appropriate subset of the complete dataset was chosen as the sample
for the test. (For details of the composition of the complete dataset, see 1.5.) What
was appropriate was determined by reference to the scope of the theory or theories
being tested: thus, for the comparison of Zaliznjaks theory and Melvolds theory,
only nominative singular derived nouns in Sample A were used as the dataset: this is
because Zaliznjak states that his theory applies only to derived non-inflected words,
and the relevant part of Melvolds theory applies to nouns but not other parts of
speech. In order to ensure that each entry in the morpheme inventory had one and
only one underlying form, under-determinacy was eliminated using the extra-
linguistic procedure described in 4.3.3.
The Stress Checking Algorithm was then run. This essentially involves the
straightforward subprocedures in (157):
292
(157) The Stress Checking Algorithm (SCA)
(a) For each morphologically-tokenized stressed word, remove the stress and
look up the underlying forms of the constituent morphemes in the
morpheme inventory.
(b) Assign stress to the word (how this is done clearly depends on which
theory is being tested).
(c) Compare the stress computed in (b) with the actual stress from (a). Keep
count of the number of words correctly stressed and the number of words
incorrectly stressed.
Once the SCA has processed all words in the dataset, we will have the total
number of words incorrectly stressed and the total number correctly stressed.
Determining how many words fall into each category clearly gives a good measure of
the theorys accuracy.
4.4.1 Test of Halles theory on non-derived nouns
This section sets out the results of testing Halles theory on 3,234 non-derived
words.
The results of running this test, given in Table 33, show that the coverage of
Halles theory is extremely high (almost 98%): it seems that this conclusively
disproves two of Zaliznjaks (1985) claims:
As far as nonderived words are concerned, the only way of approaching the task [of
predicting the stress of word-forms of a given lexeme] is to list [the stress-pattern of
every lexeme]. In other words, the stress-pattern and the placement of the stress
within the stem are effectively lexical idiosyncrasies of each word: this information
cannot be predicted formally from any other characteristics of the word. [my
translation PC] (p. 22)
the stress-pattern and placement of stress within the stem for derived words differs
from that of nonderived words in that it can be determined from a words
morphological composition [my translation PC] (p. 29).
293
Table 33. Results of testing Halles theory on non-derived words
No of words % of total
Correctly stressed 3,163 9780%
Incorrectly stressed 71 220%
TOTAL 3,234 10000%
The results above show that the stress of non-derived words can generally be
determined by looking at their morphological composition; and this approach involves
listing far less information lexically, since what is listed is morphemes rather than
words. This is perhaps unsurprising; one reason that Halles theory achieves such high
coverage is that the vast majority of non-derived nouns have either fixed stem stress
(accounted for in Halles theory by stipulating that the root is accented) or fixed
ending stress (accounted for by stipulating that the root is post-accented). The real
question is whether the theory still holds as well when applied to word-categories
other than non-derived nouns; will the same accentual specification of the morpheme
inventory and the same theory fit these words too, when many of the non-inflectional
morphemes in non-derived nouns will be common to other word-categories? The test
in the next section provides an answer to this question.
4.4.2 Test of Halles theory on non-derived and derived nouns
To determine whether Halles theory is able to explain stress assignment for
derived nouns as successfully as for non-derived nouns, the test in 4.4.1 was repeated
on a dataset consisting of both derived and non-derived nouns. This dataset included a
total of 9,512 words: the 3,234 nouns from Table 33 plus 6,278 derived nouns. The
results of this test are shown in Table 34.
294
Table 34. Results of testing Halles theory on non-derived and derived nouns
No of words % of total
Correctly stressed 8,416 8848%
Incorrectly stressed 1,096 1152%
TOTAL 9,512 10000%
It is clear from these results that Halles theory is less adequate at explaining stress
assignment for derived nouns than for non-derived nouns. The coverage is still quite
high, but as we shall see in the next section, Melvolds theory succeeds significantly
better at accounting for the same dataset.
4.4.3 Test of Melvolds theory on non-derived and derived nouns
This section sets out the results of testing Melvolds theory on derived and
non-derived nouns (both headwords and inflected forms), the same dataset as used in
4.4.2.
The test in 4.4.1 was again repeated, this time with the appropriate
modifications to the SCA to reflect Melvolds rules of stress assignment. The results
of this test were that Melvolds theory accounted correctly for over 95% of the dataset
tested, as Table 35 shows.
Table 35. Results of testing Melvolds theory on non-derived and derived nouns
No of words % of total
Correctly stressed 9,069 9534%
Incorrectly stressed 443 466%
TOTAL 9,512 10000%
295
4.4.4 Test of Melvolds theory on nouns, non-reflexive verbs and adjectives
A re-test of Melvolds theory on a dataset including not just nouns, but also
non-reflexive verbs and adjectives, yielded the results in Table 36:
Table 36. Results of testing Melvolds theory
on nouns, adjectives and non-reflexive verbs
No of words % of total
Correctly stressed 9,914 9278%
Incorrectly stressed 771 722%
TOTAL 10,685 10000%
Comparing Table 35 and Table 36, we see that the error rate increases when we
attempt to apply Melvolds theory to all parts of speech, rather than just nouns. I
believe that this can be attributed largely to the fact that Melvolds theory, as it stands,
does not account adequately for stress in the infinitives of verbs, since many of the
771 exceptions are verbal infinitives. (The corpus used does not include anything
other than infinitives for verbs; thus, we cannot draw any conclusions about stress in
verbal conjugation from the tests carried out.) The difficulty in accounting for stress
in verbs can be explained as follows.
The suffixes !,!
svr
and !%!
svi
are frequently found in nouns, and they are almost
always stressed
54
. This suggests that these suffixes are dominant and accented. Two
examples are given below:

54
There is a single class of counter-examples to this consisting of nouns derived from participial stems,
e.g. ,<>:9&'',-.$ !6A'
c
+/%*
rv
+%
rvc
+4
ra1
+4
ra1
+6#&'
rn
+
in1
! duty. These forms, including the participles
from which they are derived, tend to be stressed one syllable earlier than the suffix !%!
rvc
, and this
implies that they can be explained by a retraction rule. Thus these exceptions are not a serious
296
(158) 4!"#-"#&: ."6$ !90'1$
c
+#'1$
rv
+/%
svi
+&'1-'
sn
+
in1
! president
83):6)1" !;&8
rv
+/,
svr
+-'
sn1
+,8&8
sn
+6
in18
! college
However, numerous infinitives containing these suffixes are stressed on a
syllable before the suffix, a fact which is inconsistent with the hypothesis that !,!
svr
and !%!
svi
are dominant and accented. A whole class of such infinitives is represented
by derived imperfective verbs, which end in !,"
sa
+%
svi
+&'
sa2
! (/(&.$ / )(&.$). In
these verbs, the stress falls on the syllable preceding !,"!
sa
. Two other examples
outside this class are given below:
(159) 46&:%&.$ !9-/%2
rv
+%
svi
+&'
sa2
! cry
8#&:!).$ !;
c
+$/%0'
rv
+,
svr
+&'
sa2
! hit
If Melvolds theory could be modified to account for cases such as these, the
overall error rate for nouns, adjectives and verbs would be much more satisfactory. I
leave this as a goal for future research.
4.4.5 Test of Zaliznjaks theory on nominative singular derived nouns
This section sets out the results of testing Zaliznjaks theory on 524
nominative singular derived words.
Once again, the test in 4.4.1 was repeated with appropriate modifications to
the SCA to reflect the fact that Zaliznjaks stress assignment rules were being tested.

challenge to the hypothesis that !%!
rvc
is dominant accented. Two other counter-examples which do not
fall into this category and are thus perhaps true exceptions are .!": <,(&')"
!&0'/1A
rv
+6"
ra
+%
rvc
+4'
ra1
+,7
rn1
+6
in18
! demand and 9":!%&6, !*'/10
rv
+2
ra
+%
rvc
+-
rn1
+6
in18
! mirror.
297
The results of this test were that the coverage of Zaliznjaks theory is approximately
92% of the dataset tested, as Table 37 shows.
Table 37. Results of testing Zaliznjaks theory
on nominative singular derived nouns
No of words % of total
Correctly stressed 482 9198%
Incorrectly stressed 42 802%
TOTAL 524 10000%
Of course, these figures are not directly comparable with those in 4.4.1-4.4.4, because
the dataset used for the test is different. For the purpose of direct comparison,
therefore, we need to know how Melvolds theory deals with the same dataset. This
test was performed and the results are set out in the next section.
4.4.6 Test of Melvolds theory on nominative singular derived nouns
The test in 4.4.3 was repeated (using the same morpheme inventory), but this
time using exactly the same dataset which was used in 4.4.5. The results are presented
in Table 38.
Table 38. Results of testing Melvolds theory
on nominative singular derived nouns
No of words % of total
Correctly stressed 498 9504%
Incorrectly stressed 26 496%
TOTAL 524 10000%
298
Comparison of Table 37 and Table 38 thus reveals that the coverage of the two
theories was not radically different, although one is led to conclude on the basis of
this sample of relatively high-frequency words that the coverage of Zaliznjaks theory
is not as high. The difference between the two theories coverage is fairly significant
(with p < 0045). Furthermore, it appears that as far as Melvolds theory is concerned,
nominative singular derived nouns are a representative subset of the dataset used in
4.4.5, since the coverage of 9504% is comparable to the equivalent figure from Table
35.
4.4.7 Analysis of errors in Melvolds and Zaliznjaks theories
The two lists of words not correctly accounted for by Melvolds and
Zaliznjaks theories are given in Table 39 and Table 40 respectively.
299
Key to Table 39 and Table 40
Word common to Table 39 and Table 40
D
denotes that a morpheme is dominant
O and E denote underlying jers
! denotes accentuation/post-accentuation (depending on position in morpheme) in Table 39
> denotes accentuation/post-accentuation (depending on position in morpheme) in Table 40
< denotes pre-accentuation
Table 39. Analysis of words incorrectly stressed by Melvolds theory
Actual stress Computed
stress
Gloss Underlying form
1
!"#$%&'()*"" *!"" #$%&'()*"
majority
#$!%&'
ra
+
D
(
sa
+!)*
ra2
+E+,-!
rn
+%
in18
#
2
*"+'," **""+',
war
#-%.
rv
+E*
ra1
+
rn2
+/
in3
#
3
-&./" 0)". *-&"./0)".
director
#0')1'23
rv
+,
ra2
+%1
rn
+O
in1
#
4
1,*""- *1,"*"-
factory
#4/
c
+-%0
rv
+
rn1
+O
in1
#
5
1,"',*/( *1,',*/" (
curtain
#4/
c
+*/
c
+-'!2+
rv1
+
rn1
+O
in1
#
6
1,"2,3 *1,2,"3
smell
#4/
c
+5!/6
rv
+
rn1
+O
in1
#
7
1,2&"(0, *1,2&(0,"
note
#4/
c
+5')+
rv
+O3!
rn1
+/
in3
#
8
1/".0,#" *1/.0," #"
mirror
#4'21
rv
+3
ra
+
D
!/
rvc
+&
rn1
+%
in18
#
9
0.,(")," *0.,("" ),
beauty
#31/+!
ra
+%,!
rn1
+/
in3
#
10
',.""- *',"."-
people
#*/
c
+1%0
rv
+
rn1
+O
in1
#
11
'/"4/5 *'/4/" 5
German
#*'27'
ra
+!2,+
sn
+O
in1
#
12
"!".6"-"*,'&/ *"!".6-"*," '&/
equipment
#%$
c1
+%1!80
rn
+%-
ra
+
D
!/
rvc
+E*'
ra1
+!).
rn1
+%
in18
#
13
")*/" )()*/''"()$ *")*/)()*/" ''"()$
responsibility
#%,
c
+-'!2,
rv
+E+,-'
rn
+
D
!2*
ra
+E*
ra1
+%+,'
rn
+O
in1
#
14
2"-.6" 7, *2"" -.67,
girlfriend
#5%
c
+0189
ra
+
rn2
+/
in3
#
15
2""/1- *2"/" 1-
train
#5%
c
+.240!
rv
+
rn1
+O
in1
#
16
2""4"8$ *2"4""8$
help
#5%
c
+7%(,('!
rv
+
rn2
+O
in1
#
17
2.,"*&#" *2.,*&"#"
rule
#51/-'
rv
+
D
!)
rvr
+&
rn1
+%
in18
#
18
2.&.""-, *2.&"."-,
nature
#51')
c
+1%0
rv
+
rn1
+/
in3
#
19
2."9/"((&: *2."9/((&":
profession
#51%
c2
+:'2+
rv
++'
ra
+!).
rn2
+/
in3
#
20
2."9/"((". *2."" 9/((".
professor
#51%
c2
+:'2+
rv
++
ra
+%1
rn
+O
in1
#
21
(*/"-/'&/ *(*/-/"'&/
information
#+
c
+-'!20'
rv1
+
D
!2
rvs
+E*'
ra1
+!).
rn1
+%
in18
#
22
(#6" ;,+ *(#6;,"+
case
#+
c
+&!8,('
rv
+
D
!/.
sn
+O
in1
#
23
(2""("! *(2"("" !
means
#+
c
+5%
c
++!%$
rp
+
rn2
+O
in1
#
24
(),)$:" *()," )$:
article
#+,/
rv1
+E,'
ra1
+.
rn1
+/
in3
#
25
)./"!"*,'&/ *)./!"*,"'&/
demand
#,1'!2$
rv
+%-
ra
+
D
!/
rvc
+E*'
ra1
+!).
rn1
+%
in18
#
26
6;/'&" 0 *6;/"'&0
pupil
#!8,('
rv
+
rvr
+
D
!2*'
ra
+)3!
sn
+O
in1
#
300
Table 40. Analysis of words incorrectly stressed by Zaliznjaks theory
Actual stress Computed stress Gloss Underlying form
1
!"#$'&"5, *!"" #$'&5,
hospital
#$>%&'
rv
+<*'
ra1
+>),+
rn1
+>/
in3
#
2
*""1-63 **"1-6" 3
air
#-%4>
c
+0>86
rv
+>
rn1
+<
in1
#
3
*""1.,() **"1.,"()
age
#-%4>
c
+1>/+,
rv
+>
rn1
+<
in1
#
4
7#6!&'," *7#6" !&',
depth
#9&>8$'
ra
+)*>
rn1
+>/
in3
#
5
7"(2"-&"' *7"(2""-&'
gentleman
#9%+5>%0'
rn
+)*>
rn1
+<
in1
#
6
1,!""), *1,!"),"
worry
#4>/$
rv
+
D
%,>
rn1
+>/
in3
#
7
1,"',*/( *1,',*/" (
curtain
#4/>
c
+*/>
c
+-'>2+
rv1
+>
rn1
+<
in1
#
8
1,"2,3 *1,2,"3
smell
#4/>
c
+<5/6
rv
+>
rn1
+<
in1
#
9
1/".0,#" *1/.0," #"
mirror
#4'>21
rv
+3>
ra
+
D
>/
rvc
+
D
<&
rn1
+<%
in18
#
10
0,.)&"', *0," .)&',
picture
#3>/1,'
rn
+)*>
rn1
+>/
in3
#
11
4&'&()/".()*" *4&'&"()/.()*"
ministry
#7')*'>)+
ra
+,'>21
rn
++,->
rn
+<%
in18
#
12
46<;&" ', *46"<;&',
man
#7>8;
rn
+,('>
ra3
+)*>
rn1
+>/
in3
#
13
'/-"()," )"0 *'/-"(),)""0
lack
#*'2>
c
+0%>
c
++,/>
rv1
+,>
ra1
+>%3
rn1
+<
in1
#
14
""!#,0" *"!#,"0"
cloud
#%$>
c1
+&>/3
rv
+>
rn1
+<%
in18
#
15
""!#,()$ *"!#,"()$
region
#%$>
c1
+&>/+
rv
+,'>
rn1
+<
in1
#
16
"!".6"-"*,'&/ *"!".6-"*," '&/
equipment
#%$>
c1
+%1>80
rn
+%->
ra
+
D
>/
rvc
+<*'
ra1
+<).
rn1
+<2
in18
#
17
""!.,1 *"!.,"1
shape
#%$>
c
+1>/4
rv
+>
rn1
+<
in1
#
18
"!:"1,''"()$ *"!:1,"''"()$
duty
#%$'>
c
+>/4
rv
+
D
>/
rvc
+*>
ra1
+*>
ra1
+>%+,'
rn
+<
in1
#
19
""2=) *"2=")
experience
#%>
c
+5),>
rv
+>
rn1
+<
in1
#
20
"")-=3 *")-=" 3
rest
#%,>
c
+0)6>
rv
+>
rn1
+<
in1
#
21
"")26(0 *")26"(0
leave
#%,>
c
+5>8+3
rv
+>
rn1
+<
in1
#
22
"";/./-$ *";/./"-$
queue
#%>
c1
+,('21'>20'
ra
+>
rn
+<
in1
#
23
2""*"- *2"*"" -
cause
#5%>
c
+->%0
rv
+>
rn1
+<
in1
#
24
2""-*&7 *2"-*&"7
exploit
#5%>
c
+<0-')9
rv
+>
rn1
+<
in1
#
25
2""/1- *2"/" 1-
train
#5%>
c
+.>240
rv
+>
rn1
+<
in1
#
26
2""&(0 *2"&"(0
search
#5%>
c
+<.)+3
rv
+>
rn1
+<
in1
#
27
2"#"*&" ', *2"" #"*&',
half
#5>%&
ra
+>%-'
ra
+)*>
rn1
+>/
in3
#
28
2""#$1, *2"#$1,"
use
#5%>
c
+<&'4
ra
+>
rn3
+>/
in3
#
29
2""4"8$ *2"4""8$
help
#5%>
c
+7>%(,('
rv
+>
rn2
+<
in1
#
30
2"":( *2":"(
belt
#5%>
c
+<./+
rv
+>
rn1
+<
in1
#
31
2.,"*&#" *2.,*&" #"
rule
#51>/-'
rv
+
D
>)
rvr
+
D
<&
rn1
+<%
in18
#
32
.,!""), *.,!"),"
work
#1>/$
rv
+
D
%,>
rn1
+>/
in3
#
33 ./!>'"0
*./!/'"" 0
child
#1'>2$'
rn
+
D
>%*
rn1
+
D
%3>
sn
+<
in1
#
34
(*/"-/'&/ *(*/-/"'&/
information
#+>
c
+-'>20'
rv1
+
D
>2
rvs
+<*'
ra1
+<).
rn1
+<2
in18
#
35
(/./-&"', *(/./"-&',
middle
#+'21'>20'
rn
+)*>
rn1
+>/
in3
#
36
(""*/()$ *("*/"()$
conscience
#+%>
c
+-'>2+
rv3
+,'>
rn1
+<
in1
#
37
(2""("! *(2"("" !
means
#+>
c
+5%>
c
++>%$
rp
+>
rn2
+<
in1
#
38
)&%&'," *)&"%&',
silence
#,'>)(
ra
+)*>
rn1
+>/
in3
#
39
)./"!"*,'&/ *)./!"*," '&/
demand
#,1'>2$
rv
+%->
ra
+
D
>/
rvc
+<*'
ra1
+<).
rn1
+<2
in18
#
40
6" <,( *6<," (
terror
#8>
c
+;>/+
rv
+>
rn1
+<
in1
#
41
6" ."*/'$ *6.""*/'$
level
#8>
c
+1>%-'
ra
+>2*'
rn3
+<
in1
#
42
6;/'&" 0 *6;/"'&0
pupil
#>8,('
rv
+>
rvr
+
D
>2*'
ra
+>)3
sn
+<
in1
#
301
It should be stated at the outset that it is impossible to give explanations for all the
items in these lists, since in each case an incorrect stress could either be attributed to
an incorrect accentual specification of some morpheme(s) within a word, or it could
be attributed to a deficiency in the overall stress assignment rules proposed. We are
more interested in the latter, since identification of deficiencies in the rules is likely to
facilitate the positing of better rules. Yet there is never any way of being sure that an
incorrect stress is caused by a deficiency in the rules rather than some incorrect
accentual specification, since, as was shown in 4.3, the coverage of a particular set of
rules depends crucially on the accentual specification of the accompanying morpheme
inventory, and there are so many ways in which a realistically-sized inventory can be
specified that it is impossible to deduce the optimal way by trying every possibility.
However, two general observations emerge from the data. First, over 40% of
the words (11 out of the total 26) incorrectly stressed by Melvolds theory are also
incorrectly stressed by Zaliznjaks theory, and, furthermore, they are without
exception assigned the same incorrect stress. The fact that neither theory accounted
correctly for these words strongly implies that their stress assignment would be
exceptional, no matter which theory we adopted. The 11 words are given in Table 41:
302
Table 41. Exceptions common to Zaliznjaks and Melvolds theories
Actual stress Incorrect stress Gloss Morphological string
1 9&:'&("- *9&'&(":-
curtain
!*%
c
+4%
c
+"'1#
rv1
+
sn1
+
in1
!
2 9&:4&: *9&4&::
smell
!*%
c
+9%C
rv
+
sn1
+
in1
!
3 9":!%&6, *9"!%&: 6,
mirror
!*'10
rv
+2
sa
+%
svi
+-
sn1
+6
in18
!
4 ,<,!8:#,(&')" *,<,!8#,(&: ')"
equipment
!6A
c1
+60;$
rn
+6"
sa
+%
svi
+4'
sa1
+,7
sn1
+6
in18
!
5 4,:"9# *4,":9#
train
!96
c
+71*$
rv
+
sn1
+
in1
!
6 4,:+,1$ *4,+,:1$
help
!96
c
++68&8'
rv
+
sn2
+
in1
!
7 4!&:()6, *4!&(): 6,
rule
!90%"'
rv
+,
svr
+-
sn1
+6
in18
!
8 -(":#"')" *-("#": ')"
information
!#
c
+"1$'
rv
+1
svd
+4'
sa1
+,7
sn1
+6
in18
!
9 -4,:-,< *-4,-,:<
means
!#
c
+96
c
+#6A
rp
+
sn2
+
in1
!
10 .!":<,(&')" *.!"<,(&: ')"
demand
!&0'1A
rv
+6"
sa
+%
svi
+4'
sa1
+,7
sn1
+6
In18
!
11 83"'):% *83":')%
pupil
!;&8'
rv
+
svr
+14'
sa
+,2
sn
+
in1
!
The reason that stress was incorrectly assigned in 3, 4, 7, 8, 10 and 11 is that the
suffixes !%!
svi
, !,!
svr
,<!1!
svd
and !14'!
sa
were all specified as dominant and accented in
both theories, on the compelling basis of the many words ending in &: ')"
!%
svi
+4'
sa1
+,7
sn1
+6
in18
!, ":')" !1
svd
+4'
sa1
+,7
sn1
+6
in18
! or !14'
sa
+,7
sn1
+6
in18
!, and ):."6$
!,
svr
+&'1-'
sn
+
in1
!. For this reason, it seems reasonable to concur that these words are best
treated as exceptional in their stress assignment. (Russian also has -("#": ')" with the
more technical meanings reduction and squaring of accounts, but presumably with
the same morphological structure as 8.)
The remaining cases 1, 2, 5, 6 and 9 have the common feature that all
consist of prefix(es) and a root, and in all cases the two theories both assigned stress
to the root where it should have been assigned to the initial syllable. This brings us to
the second general observation, which is that words of this type accounted for more
than half of the words incorrectly stressed by Zaliznjaks theory (22 out of 42): these
are listed in Table 42.
303
Table 42. Prefixed nouns stressed incorrectly by Zaliznjak
Actual stress Incorrect stress Gloss Morphological string
1 (,:9#8: *(,9#8: :
air !"6*>
c
+$>;C
rv
+>
sn1
+<
in1
!
2 (,:9!&-. *(,9!&:-.
age !"6*>
c
+0>%#&
rv
+>
sn1
+<
in1
!
3 9&:'&("- *9&'&(":-
curtain !*%>
c
+4%>
c
+"'>1#
rv1
+>
sn1
+<
in1
!
4 9&:4&: *9&4&: :
smell !*%>
c
+<9%C
rv
+>
sn1
+<
in1
!
5 ,:<6&%, *,<6&:%,
cloud !6A>
c1
+->%2
rv
+>
sn1
+<6
in18
!
6 ,:<6&-.$ *,<6&:-.$
region !6A>
c1
+->%#
rv
+&'>
sn1
+<
in1
!
7 ,:<!&9 *,<!&:9
shape !6A>
c
+0>%*
rv
+>
sn1
+<
in1
!
8 ,:4/. *,4/:.
experience !6>
c
+9,&>
rv
+>
sn1
+<
in1
!
9 ,:.#/: *,.#/::
rest !6&>
c
+$,C>
rv
+>
sn1
+<
in1
!
10 ,:.48-% *,.48:-%
leave !6&>
c
+9>;#2
rv
+>
sn1
+<
in1
!
11 ,:3"!"#$ *,3"!":#$
queue !6>
c1
+&8'10'>1$'
ra
+>
sn
+<
in1
!
12 4,:(,# *4,(,:#
cause !96>
c
+">6$
rv
+>
sn1
+<
in1
!
13 4,:#()5 *4,#():5
exploit !96>
c
+<$"',B
rv
+>
sn1
+<
in1
!
14 4,:"9# *4,":9#
train !96>
c
+7>1*$
rv
+>
sn1
+<
in1
!
15 4,: )-% *4,): -%
search !96>
c
+<7,#2
rv
+>
sn1
+<
in1
!
16 4,:6$9& *4,6$9&:
use !96>
c
+<-'*
ra
+>
sn3
+>%
in1
!
17 4,:+,1$ *4,+,:1$
help !96>
c
++>68&8'
rv
+>
sn2
+<
in1
!
18 4,:>- *4,>:-
belt !96>
c
+<7%#
rv
+>
sn1
+<
in1
!
19 -,:("-.$ *-,(":-.$
conscience !#6>
c
+"'>1#
rv3
+&'>
sn1
+<
in1
!
20 -4,:-,< *-4,-,:<
means !#>
c
+<96>
c
+#>6A
rp
+>
sn2
+<
in1
!
21 8:*&- *8*&: -
horror !;>
c
+5>%#
rv
+>
sn1
+<
in1
!
22 8:!,("'$ *8!,: ("'$
level !;>
c
+0>6"'
ra
+>14'
sn3
+<
in1
!
There is a relatively large class of words of this kind in Russian: many of these words
appear to be deverbal nouns with zero suffixes of the type discussed on pages 183-
184, as can be seen from Table 42. The fact that they are all stressed on the initial
syllable is key: Melvolds theory, of course, is able to account correctly for most of
them because in the morpheme inventory complementing the theory, prefixes are
unaccented and the roots in question are also usually unaccented for the words in
304
question
55
, and the default rule (stress the leftmost syllable) thus applies. The basic
rules of Zaliznjaks theory can never account for words of this type since he stipulates
(p. 58) that prefixes (apart from !",!,<!9%!<and<!4'1!) are always post-accented (he does
not acknowledge the existence of unaccented morphemes). Thus, even if the
following morpheme is pre-accented and stress attaches to the prefix (see Table 28),
it must surface on the syllable immediately following the prefix. Since such a large
proportion of the exceptions to Zaliznjaks theory are examples of this type, we can
thus safely state that the poorer coverage of Zaliznjaks theory is largely attributable
to its weaker treatment, in the formal rules, of prefixed deverbal nouns.
It is also interesting that where the same stems appear to be contained within
the constituent structure of morphologically more complex words, the stress is usually
no longer on the initial syllable. Examples can be found for almost all of the words in
Table 42: a number are given in Table 43.

55
In the sample tested, incorrect assignment of stress by Melvolds theory in these cases was in fact
always attributable to the fact that the root was not unaccented.
305
Table 43. Words derived from prefixed stems
Noun from Table 42 Word(s) with same stem and different stress
(,:9#8: air (,9#8:B'/0 !"6*$;84,7! air (adj.)
(,:9!&-. age (,9!&-.&: .$ !"6*0%#&%&'! to grow up
,:<6&-.$ region ,<6&-.',: 0 !6A-%#&467! pertaining to a region
,:<!&9 shape <"9,<!&:9)" !A'1*6A0%*',71! disorder
,:.#/: rest ,.#/:&:.$ !6&$,C%&'! to rest
,:.48-% leave ,.48-%',: 0 !6&9;#2467! pertaining to leave
,:3"!"#$ queue ,3"!"#',:0 !6&8'10'1$467! next
4,,3"!2#'/0 !966&8'10'6$4,7! next
4,:(,# cause 4,(,#): .$ !96"6$',&'! to lead around
4,:#()5 exploit 4,#(): 5&.$ !96$"',B%&'! to move around
4,:"9# train 4,":9#).$ !9671*$',&'! to drive around
4,:+,1$ help 4,+,:1')% !96+68&8'4',2! helper
4,:>- belt 4,>-',: 0 !967%#467! pertaining to a belt
-4,:-,< means -4,-,:<-.(,(&.$ !#96#6A#&"6"%&'! to facilitate
8:*&- horror 8*&:-'/0 !;5%#4,7! terrible
8:!,("'$ level 8!,('>:.$ !;06"4'%&'! to level
However, note also some counterexamples (which are rarer) in Table 44:
Table 44. Further words derived from prefixed stems
Noun from Table 42 Word with same stem and same stress
,:4/. experience ,:4/.'/0 !69,&4,7! experienced
4,:+,1$ help <"-4,:+,1'/0 !A'1*96+68&8'4,7! helpless
-,:("-.$ conscience -,:("-.).$ !#6"'1#&',&'! to be ashamed
One might suppose on the basis of the examples above that the rules of stress
assignment in nouns differ from those in other parts of speech. This analysis is
implied by Zaliznjak (1985: 39-43) (one of the examples he adduces is in fact ,: .48-%
from Table 43 above), who states that for verbs stress to the right of the root is more
characteristic whereas in suffixless deverbal nouns stress is usually on the root, or
in a significant minority of cases a different stress placement is found, usually on the
prefix.
306
In order to account for the examples in Table 43 within Melvolds theory, one
way forward would be to reclassify all the noun-forming suffixes !!
sn
, !!
sn1
, !!
sn2
, !&'!
sn1
and !14'!
sn3
as dominant and unaccented: their effect would therefore be to obviate any
accentual effects of the preceding morphemes and assert the default initial-syllable
accentuation. This would be entirely consistent with Melvold (1989) and Zaliznjak
(1985), who both adduce the word ,:<,!,."'$ !/6A606&'+14'! werewolf (cf. ,<,!,: .
!6A60/6&! turn) as evidence that !14'! should be classified as dominant and
unaccented (D Init, in Zaliznjaks terminology), and Zaliznjak (1985: 68) also lists
!&'! and various other morphemes which are claimed to be of this type.
This move entails certain problems, however. The first, as Zaliznjak states,
and I confirmed by analysis of the 524-word dataset, is that the words of this type
where stress is on the initial syllable are actually in a minority: a larger number have
stress on the root. In particular, there are other words in Oliverius (1976) which
contain !!
sn1
and !!
sn2
where the actual stress would not be consistent with a dominant
unaccented specification for these morphemes, e.g. 9&(,: # !*%
c
+"/6$
rv
+
sn1
+
in1
!
factory, 4,#!8:5& !96
c
+$0/;B
ra
+
sn2
+%
in3
! girlfriend. Interestingly, however, all the
words in Oliverius (1976) containing !!
sn
, !&'!
sn1
and !14'!
sn3
do have initial stress. One
might therefore conclude that words such as 9&(,:# and 4,#!8:5&, which Oliverius
claims include empty noun-forming suffixes as part of their constituent structure, in
fact do not include these suffixes. An alternative, equally valid, conclusion would be
that they include other zero noun-forming suffixes which are not dominant
unaccented.
We now explain the difference in stress assignment between the words in the
first and second columns of Table 43 by stipulating that the words in the second
307
column do not have dominant unaccented noun-forming suffixes as part of their
morphological structure. For those words with zero noun-forming suffixes, this
stipulation seems entirely justified on semantic grounds, since noun-forming suffixes
would not be needed to form the non-nouns in the right column. In the case of
8!,('>:.$, 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 Zaliznjaks basic rules appear to be
inferior to Melvolds theory in three ways: first, their scope is more limited than that
of Melvolds 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 Zaliznjaks theory shows that it can objectively be said to be less
elegant than Melvolds, since it contains more rules. These two reasons are the
principal reasons for rating Zaliznjaks theory less highly. The third reason is that the
coverage of Melvolds 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 Melvolds theory, is sufficient
to account for 9534% 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 Melvolds, 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 Melvolds 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, Melvolds provides the most promising starting-
point 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 Melvolds theory and show how they can
be implemented within this phonology of Russian. This re-implementation
demonstrates that although Melvolds 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 phrase-
structure grammar (even one more powerful than context-free) can be viewed non-
derivationally. The grammar in this chapter formalizes the salient aspects of
Melvolds 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
Melvolds theory. In certain respects, I depart from, or add to, the theory. For one
thing, Melvolds 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 Melvolds 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 (Melvolds)
theory about Russian stress assignment. It is important to appreciate, however, that
Melvolds 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 Melvolds 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
Melvolds 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 post-
accentuation as the projection of a left bracket (Idsardi 1992, Halle 1997). In other
words, Melvolds 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 Melvolds
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 words syllable structure?
- What is a words 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 Melvolds 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 words 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 Melvolds 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 Melvolds 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 words [+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 Melvolds 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 words 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-stem
[+stress]
adj-stem
[+stress]
sn
[stress]
adj-stem
[+stress]
sa
[stress]
noun
[+stress]
infl. ending
[stress]
noun-stem
[stress]
sa
[+stress]
rv
[stress]
sn
[stress]
sp'cts
[stress]
ij
[stress]
al'
[+stress]
n
[stress]
ost'
[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) X stem inflection
(b) stem stem suffix
(c) stem root
(d) stem prefix 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 Melvolds theory, namely:
(161) a) A dominant suffix always wins the stress from a stem to its left.
b) An accented left-hand constituent always wins the stress from its
[dominant] sister.
c) A [+accented] right-hand constituent always wins the stress from its
[accented] sister.
d) 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, [stress] [+dominant,
o stress] o stress]
Thus, for example, (160b) becomes (163):
(163) stem stem suffix
[X] [Y] [Z]
{X [+accented, o stress],
Y [stress],
Z [+dominant, o stress]}
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 Melvolds 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, o 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 prefix root
[X] [Y] [Z]
{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, [+accented, [dominant]
o stress] o stress]
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] right-
hand constituent inherits the mother nodes specification for [stress] if its sister is
[accented]. The necessary constraints are as follows:
(166) [+accented, [accented] [+accented,
o stress] o 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#&'+/%+',!<youths [instr. pl.]; cf. Table 22
on page 242). In conjunction with the other constraints, its effect will be to assign the
feature [+stress] to none of the morphemes in the word. The syllable structure
grammar will then do the work of assigning stress to the initial syllable of a word
consisting entirely of [stress] segments; how this is done will be discussed in more
detail in section 5.2.6 (see page 334).
This section, then, has provided an outline of how the constraints in Melvolds
theory on which morpheme determines the word-stress can be incorporated into the
context-free rules: effectively, the values for [stress] are blanks in the lexical
entries for each morpheme which are filled in by the relevant CFG rules.
5.2.2 Polysyllabic morphemes
In some cases, we need to mark the association of stress with a particular
segment to avoid ambiguity. Without this information, for instance, there is no way of
knowing which vowel in a polysyllabic morpheme (such as !2606"!
rn
) is [+stress] if
the morpheme itself is [+stress]. (In the case of !2606"!
rn
, the stress always falls on the
second !6! if !2606"!
rn
is stressed.) This information therefore needs to be lexically
specified in the morpheme inventory. Perhaps the most obvious way of doing this is
319
to allow the feature [stress] to apply to individual segments. Thus the terminal rule
which lists !2606"!
rn
in the morpheme inventory would be as in (168):
(168) rn !<<<<2 <<<<<6 <<<<<0 <<<<<6 <<<<<"<<<<!
[o stress] [o stress]
This rule states that the second vowel is the one which has the special status as far as
stress is concerned.
5.2.3 Post-accentuation
None of the three theories of stress surveyed in 4.2 states explicitly what the
role of the syllable is in stress assignment: in particular, Halle (1997) and Melvold
(1989) associate stress with individual phonemes or syllabic nuclei (cf Halle 1997:
276), rather than with whole syllables. The treatment of stress assignment as if it were
independent of syllabification is in fact typical of much work in mainstream
generative grammar (as mentioned in 2.1, Chomsky and Halle 1968 is a prime
example of this). Yet in this, Halle and Melvold go against the grain of metrical stress
theory:
The general relevance of syllables to rules of stress assignment has long been clear
(see for instance Kahn 1976). A theory in which stress rules refer only to the
structural properties of syllables is far more constrained than a theory (such as that of
SPE) in which stress rules refer directly to segments Following earlier work (e.g.
Jakobson 1931), I adopt the view that in stress languages, the stress-bearing unit is
the syllable Halle and Vergnauds view on stress-bearing units is quite different:
they argue that stress rules may include statements of the form the stress-bearing
elements are X, where X can be for example vowels I suggest that this proposal
states as part of language-particular grammars what arguably should be part of
universal grammar. It seems unlikely that we would ever find a language in which
the stress-bearing units are consonants, or nasals, or coronals, and so on. The
phenomena that lead HV to suppose that segments can be stress-bearing can be better
accounted for by adopting an explicit theory of syllable structure and syllable weight.
(Hayes 1995: 48-49)
320
Up to this point, syllabification has played no role in the context-free account
of stress assignment either. However, I shall show in this section that referring to
syllabification explicitly in stress assignment not only brings it into line with metrical
stress theory and reconciles two important modules of generative grammar which
have often been thought of separately, it also simplifies the way in which we can think
of post-accentuation and stress retraction. The current analysis therefore is fully in
support of Hayess proposition that the syllable cannot be ignored in stress
assignment. Yet, I also propose, contrary to Hayes (1995), that in Russian, stress is
also borne by individual segments of morphemes, which may even in some cases be
consonants.
56
One of the important contributions of my proposal, therefore, is to show
how Hayess viewpoint can be reconciled with that of Halle and Vergnaud.
In all the post-accented morphemes identified by Halle and Melvold, the final
segment is a consonant. This, in fact, is not surprising, since only roots and
derivational suffixes may be post-accented
57
, and Slavic roots and derivational
suffixes generally have to end in consonants (Isa;enko 1970: 88) (although some
verbal roots may end in vowels). In most cases, the final consonant of a post-accented
morpheme will syllabify with the initial segments of the following morpheme, as in
!26-!A%!"#+%!. In view of this, there are good grounds to see post-accentuation as being
in some way determined by the final consonant of the [+stress] morpheme: if the first
vowel in the following morpheme is stressed, then this is simply because it and the

56
This analysis finds a number of parallels with Piotrowski, Roca and Spencers (1992) treatment of
lexical syllabicity in Polish, which aims to explain the reflexes of jers in Polish, a phenomenon closely
related to those under discussion here.
57
Prefixes are usually unaccented (or occasionally accented, e.g. !",!
c
, !9%!
c
), and it does not make
sense to speak of post-accented inflectional endings, since by definition no morpheme may ever follow
an inflectional ending.
321
consonant in question reside together within the same syllable. To represent this fact
in the grammar, two adaptations need to be made. First, in each post-accented
morpheme, we need to mark the final consonant as having the special status for stress
assignment, just as we marked the second vowel of !2606"!
rn
in (168). Thus the final
consonant of post-accenting morphemes can be lexically marked [o stress]. For
example, the terminal rule for !26-A%#!
rn
would be as in (169).
(169) rn !<<<<2 <<<<<6 <<<<<<- <<<<<<A <<<<<<<% <<<<<<<<<#<<<<!
[o stress] [o stress]
If we then stipulate in the syllabification grammar, by rules such as those in
(170), that [+stress] spreads to all segments within the same syllable, the correct
syllable will always be stressed.
(170) syllable onset rime
[X] [Y] [Z]
{ X [o stress],
Y [o stress],
Z [o stress] }
rime nucleus
[X] [Y]
{ X [o stress],
Y [o stress] }
etc
Using rules of this type, #8!&%&: !$;0
ra
+%2
sn
+%
in3
! fool (gen. sg.), in which the
suffix !%2
sn
! is dominant post-accented, will have a dual morphological/phonological
structure as in Figure 24.
322
Figure 24. Morphological/phonological structure of #8!&%& /dur
ra
+ak
sn
+a
in3
/
noun
[+stress]
noun-stem
[+stress]
infl. ending
[stress]
adj-stem
[stress]
sn
[+stress]
$ ; 0 % 2
[+stress]
%
[+stress]
onset
nucl.
rime onset rime
nucl.
onset
[+stress]
nucl.
[+stress]
rime
[+stress]
syllable syllable syllable
[+stress]
weakfoot strongfoot
phonological word
323
In certain identifiable respects, the syllabification interpretation of post-
accentuation differs in its predictions from Melvolds theory. First, it is conceivable in
Melvolds 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 Melvolds theory that there might be a case of the type
!V
1
C
1
+C
2
V
2
!, where the morpheme ending in C
1
is post-accented, but C
1
C
2
is not
a possible onset. According to Melvolds theory, V
2
would be stressed, but according
to the syllabification view V
1
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 Melvolds theory: perhaps the syllabification interpretation fails to
cover certain words in Russian which are covered by Melvolds 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
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 G
1
and G
2
, which define
languages L
1
and L
2
respectively, then L must be at the intersection of L
1
and L
2
: the
strings of L must comply with the constraints of both grammars in order to be well-
formed. 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.
utterance
syntax
phonology semantics
morphology
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 Melvolds 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 nucleus coda
[X] [Y] [Z]
{ X [o stress],
Y [o stress],
Z [o stress] }
To see how this works, consider the constituent structure of #8!&:% !$;0
ra
+%2
sn
+
in1
!
fool (nom. sg.), shown in Figure 26.
327
Figure 26. Morphological/phonological structure of #8!&% /dur
ra
+ak
sn
+
in1
/
noun
[+stress]
noun-stem
[+stress]
infl. ending
[stress]
adj-stem
[stress]
sn
[+stress]
onset
nucl.
rime onset
[+stress]
rime
[+stress]
nucl.
[+stress]
syllable syllable
[+stress]
weakfoot strongfoot
phonological word
$ ; 0
[+stress]
%
[+stress]
2
[+stress]
coda
[+stress]
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
Melvolds 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 Melvolds 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 !V
1
C
1
+JC
2
V
2
!, where V
1
C
1
is a post-accented morpheme,
C
1
C
2
is a possible onset, J is a jer and V
1
, V
2
are vowels. In this case, Melvolds
theory would predict that stress surfaces on V
1
, whereas the syllabification
interpretation predicts that stress will be on V
2
.
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 non-
derived 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 Siegels (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 -)'- !,4
sn1
! and -,.- !6&
sn1
! (e.g.
568<)'&: !B-;A'
ra
+,4
sn1
+%:
in3
! depth, (/-,.&: !",#
ra
+6&
sn1
+%:
in3
! height) all exhibit
stress retraction in the plural. While this refutes the proposition that all derived words
have fixed stress, it is not a problem as far as the context-free implementation is
concerned: we simply stipulate that !,4!
sn1
and !6&!
sn1
, the stem-final morphemes in
these cases, are also examples of [+plural_retraction] morphemes. All word-formation
rules of the type in (172) would then be subject to the plural_retraction constraints in
(173).
(172) noun noun_stem in
[Y] [X] [Y]
(173) plural_retraction(X,Y):-
X [+plural_retraction],
Y [o plural, o plural_retraction].
plural_retraction(X,Y):-
X [plural_retraction],
Y [plural_retraction].
(where X represents the features of the stem, and Y represents the features of
the inflectional ending and noun as a whole)
Now, in order to encapsulate the idea of stress retraction within the
phonological module of the grammar, the only module which incorporates the concept
of the syllable, we need to stipulate a new feature, which might be called
[actual_stress]. (Again, the use of an extra feature is necessitated by the restriction
that the grammar must be no more powerful than context-free.) For forms which
331
exhibit stress retraction, the syllable which is [+actual_stress] must immediately
precede the [+stress] syllable, as in (174):
(174) strongfoot syllable syllable (weakfoot)
[+pl_retr] [+actual_stress] [+stress]
In other words, stress retraction could be considered as a misalignment of
[+actual_stress] and [+stress] syllables. The [+actual_stress] syllable is always at the
beginning of a strongfoot; the [+stress] syllable is the second constituent of a
strongfoot in [+plural_retraction] forms. (In all other cases mentioned so far,
[+actual_stress] will of course align with [+stress].) Plural stress retraction is therefore
more justifiably a case of retraction than the rule of jer stress retraction as
formulated in (171).
To illustrate the application of these rules, Figure 27 gives the structure of the
word (/-,:./ !",#
ra
+/6&
sn1
+,
in
! heights which exhibits plural stress retraction.
332
Figure 27. Morphological/phonological structure of (/-,./ /vis
ra
+'ot
sn1
+i
in
/
noun
[+stress,
+pl_retr,+plural]
noun-stem
[+stress, +pl_retr]
infl. ending
[stress, +plural]
adj-stem
[stress]
sn
[+stress, +pl_retr]
" , #
[+actual_
stress]
6
[+actual_
stress]
&
[+stress]
,
[+stress]
onset
nucl.
rime onset
[+actual_
stress]
rime
[+actual_
stress]
nucl.
[+actual_stress]
onset
[+stress]
rime
[+stress]
syllable syllable
[+actual_stress]
weakfoot strongfoot
[+plural_retraction]
phonological word
[+plural_retraction]
nucl.
[+stress]
syllable
[+stress]
333
Again, how does this interpretation differ in its predictions from Melvolds
theory, if at all? The only difference is that in Melvolds theory, it is theoretically
possible that there could be two stems with, say, different roots but ending in the same
morpheme, with one stem marked and the other unmarked. In the theory I propose,
this is impossible, because the markedness of a stem depends only on the markedness
of the stem-final morpheme, and this is a fixed lexical property of the morpheme.
Again, we shall see in section 5.3 whether my analysis or Melvolds accounts better
for the data.
5.2.6 Dominant unaccented morphemes
The final problem is to encode in context-free grammar the idea that where a
dominant unaccented morpheme is [+stress], the word is stressed on the initial
syllable (which may be far removed from the [+stress] morpheme). This problem can
be solved as follows. If we use terminal rules like the one in (175) for dominant
unaccented morphemes, then it will be possible for a morphological word and a
dominant unaccented morpheme to be [+stress], but for all its constituent segments
to be [stress], a special case which up to now has been excluded.
(175) sn !<<<<1 <<< <4' !
[o stress] [stress] [stress]
[+dominant]
The phonological grammar as it currently stands cannot syllabify a string
consisting entirely of [stress] segments, so we need to add the extra rule in (176),
which can be thought of as a default rule to cover this special case. The effect of this
rule is that any word which contains no [+stress] segments consists of an initial
334
stressed syllable and then a weakfoot (consisting entirely of unstressed syllables).
(This rule will also be used for cases like 5,: !,#8 !B606$
rn
+;
in22
! town [dat. sg.] and
+,:6,#,-.>+) !+6-6$+6#&'+/%+',!<youths [instr. pl.], discussed on page 318, where
stress is assigned to the initial syllable by default).
(176) prosodic_word syllable weakfoot
[+actual_stress]
[stress]
As far as the behaviour of dominant unaccented morphemes is concerned, the
predictions of the context-free grammar are exactly the same as the predictions of
Melvolds grammar.
An example of an instance where rules (175)-(176) are used is illustrated in
Figure 28.
335
Figure 28. Morphological/phonological structure of 8!,("'$ /'u
c
+rov'
ra
+cn'
sn
+
in
/
noun
[+stress]
noun-stem
[+stress]
infl. ending
[stress]
adj-stem
[stress]
sn
[+stress]
;
[stress]
0
[stress]
6
[stress]
"'
[stress]
1
[stress]
4'
[stress]
nucl.
[+actual_stress]
rime
[+actual_
stress]
onset
[actual_
stress]
rime
[actual_
stress]
nucl.
[actual_stress]
onset
[actual
stress]
rime
[actual_
stress]
syllable
[+actual_stress]
syllable
[actual_stress]
weakfoot
phonological word
nucl.
[actual_stress]
syllable
[actual_stress]
c
[stress]
ra
[stress]
coda
[actual_stress]
336
5.2.7 Concluding comments about the context-free phonology
We saw in 5.2.3-5.2.6 that the context-free analysis which I suggest accounts
for Russian stress, although modelled closely on Melvold (1989), differs from it in
only a limited number of respects in the predictions it makes. As mentioned in 4.2.2,
Melvolds theory explains Russian stress assignment in cyclic terms, following Halle
and Vergnaud (1987). Her theory includes a form of the Stress Erasure Convention, in
that dominant suffixes delete stress assigned on previous cycles. Although it might
seem that structure-changing of this type would be problematic for context-free
grammar, the present analysis offers an interpretation of cyclicity which avoids
structure-changing by stating the distribution of the feature [stress] in the word-
formation grammar solely as a function of the placement of the features [dominant]
and [accented].
However, some comparatively minor (but real) modifications were made to
the theory: some of the time, the context-free grammar can potentially make different
predictions from Melvolds theory, and therefore the two theories cannot be said to be
entirely equivalent. These modifications were introduced for two reasons: first, they
simplify the overall theory, and secondly, they bring it more into line with mainstream
generative phonology, emphasizing the role of syllabification in stress assignment. It
is important to emphasize that the modifications were not made because Melvolds
theory cannot be implemented unmodified as a context-free grammar.
The modifications to Melvolds theory have three important implications for
the lexicon: in all cases, the effect of the modifications is to reduce the amount of
information which needs to be stored therein. These implications are as follows:
337
- Jers are not regarded as significant in themselves with respect to stress
assignment.
- Stems (as opposed to morphemes) are not regarded as significant in
determining whether stress is retracted in the plural. Instead, plural stress
retraction is ultimately a function of the lexical properties of individual
morphemes.
- Accentuation and post-accentuation are not distinct features of morphemes in
the context-free lexicon, although one might surmise the opposite from Halle
(1997) and Melvold (1989). Instead, whether a morpheme is post-accented
or accented (and, one might add, which vowel of a polysyllabic morpheme is
accented) is encoded by the association of the single feature [stress] with a
particular segment in the morpheme.
The consequences of these simplifications are that certain types of word can,
at least in theory, be explained by Melvolds theory but not the context-free theory.
(Whether there are in fact any such words will be discussed below.) Specifically, the
words of this type fall into the following categories:
(177) (a) A word containing the sequence of morphs M
1
xM
2
, where M
1
is a post-
accented morpheme which determines the stress, the final segment of M
1
is
C
1
, x represents a sequence of zero or more morphemes composed entirely
of consonantal phonemes C
2
C
n
, the initial segments of M
2
are C
n+1
V
1
,
and C
1
C
2
C
n
C
n+1
is not a possible onset.
338
(b) A word containing the sequence of morphs M
1
xM
2
, where M
1
is a post-
accented morpheme which determines the stress, the final segment of M
1
is
C
1
, x represents a sequence of zero or more morphemes composed entirely
of consonantal phonemes C
2
C
n
, the initial segments of M
2
are JC
n+1
V
1

(where J is a jer), and C

1
C
2
C
n
C
n+1
is a possible onset.
(c) A noun whose behaviour in the singular suggests that its stem-final
morpheme M is post-accented, and whose behaviour in the plural suggests
that it has a marked stem, where M also occurs as the stem-final
morpheme in another noun whose behaviour in the singular and plural
suggests that the other noun cannot have a marked stem.
In the next section, we move on to consider whether there is an empirical price
to be paid for the simplifications to Melvolds theory. There are two possibilities:
either Melvolds theory over-generates, in which case the simplifications are
desirable, or the simplifications damage Melvolds theory, in which case a reduction
in coverage will result. I shall argue that the former is the case, since, as I shall show,
the simplifications result in no significant loss of coverage.
5.3 A test of the entire grammar
In Chapters 2-3, two grammar modules were presented and discussed, a
syllable structure grammar and a word-formation grammar. These two modules are
both enriched as discussed in 2.5, 3.3, and 5.2.1-5.2.6 in order to account for stress
assignment and various other phonological processes. The two grammars are
summarized in Appendix 1 and Appendix 2 respectively, and my claim is that Russian
words are described by the intersection of these two grammars. Of course, much of
339
the inspiration for this overall grammar comes from established, more or less
mainstream, theories. We saw that the syllable structure grammar is based as much as
possible on the SSG; the word-formation grammar is indebted in many ways to
Oliverius (1976); the phonological process enrichments model the observations of
Avanesov (1956), Halle (1959), and various other Russian phonologists; and the
stress assignment enrichments are based on Melvold (1989).
In this section, I describe a test of the entire grammar in its final form. The test
has two aims: first of all, to show that the grammar is coherent, and secondly, to show
that my theory about Russian stress assignment is preferable to Melvolds. This is not
because my theory has a greater coverage than Melvolds; instead, it is because my
theorys coverage is the same as Melvolds, but is a simpler theory.
The subject of the test is stress assignment. Because my claim is that both
word-formation and syllable structure are involved in Russian stress assignment,
stress assignment represents an excellent testing-ground for the overall grammar,
since if either of the modules fails to function, or makes wrong predictions, then the
ability of the grammar to assign stress correctly may be compromised. By testing the
grammars ability to assign stress, we test both its coherence and its coverage in
comparison to Melvold (1989), meeting both of the objectives described in the
previous paragraph.
The actual results of testing the overall grammar for its ability to assign stress
to nouns yielded the following results:
340
Table 45. Results of testing the overall phonology for its ability to assign stress
Context-free phonology Comparative
figures for
Melvolds theory
No of
noun-
forms
% of
subtotal
% of
total
No of
noun-
forms
% of
total
Correct stress assigned 7,188 8794% 7557% 9,069 9537%
Incorrect stress assigned 986 1206% 1037% 443 466%
Subtotal 8,174 10000% 8593% 9,512 10000%
No syllable-structure parse 184 193%
No morphological parse 1,154 1213%
Total 9,512 10000% 9,512 10000%
It is important to appreciate that this test is a much more stringent test than those
described in 4.4. Effectively, Halles, Melvolds and Zaliznjaks theories were fed
with correct morphological parses. After all, none of these theories has any
mechanism for assigning morphological (or syllable) structure. My grammar, on the
other hand, assigns stress on the basis of syllable and morphological structure which it
computes itself, and if either of these are wrong, the stress assignment may be wrong.
To illustrate this, we consider again the alternative structures in (126), which are
repeated here for the readers convenience:
(178) (a) !$6+
rn
+%+
in
! houses-[dat. pl.]
(b) !$6
c
++%+
rn
+O
in
! to-mothers-[gen. pl.]
If it were the case that !$6+!
rn
, !$6!
c
, and !+%+!
rn
were unaccented, and !%+!
in
were
accented, then stress would be assigned in (178a) to the second syllable and in (178b)
to the first syllable. In fact, there is an extremely high degree of correlation between
341
the CFGs ability to assign morphological structure correctly (in accordance with
Oliverius 1976), and its ability to assign stress correctly
59
(p < 552 10
-68
).
For true comparability, therefore, we need to test Melvolds theory on the
same 4,416 nouns to which my grammar assigned correct morphological structure.
The results of this test are shown in Table 46.
Table 46. Results of testing Melvolds theory on 4,416 nouns
Melvolds theory The CFG
No of noun-
forms
% of total No of noun-
forms
% of total
Correctly stressed 4,218 9537% 4,205 9507%
Incorrectly stressed 205 463% 218 493%
TOTAL 4,423 10000% 4,423 10000%
It would sidetrack the discussion to list the words incorrectly stressed by the CFG but
correctly stressed by Melvolds theory, then explain the discrepancies painstakingly.
The reason for this, as pointed out in section 4.2.3, is that evaluating stress assignment
theories of this type cannot be done piecemeal. Each theory is tested together with
the morphological lexicon which best fits that theory (since we have no a priori
knowledge of the lexicons accentual specification). As a consequence, there is not
necessarily a simple mapping between the lexicon used in conjunction with Melvolds
theory, and that used in conjunction with the CFG, and therefore the factors which

59
Of the 7,188 words correctly stressed, 4,139 were morphologically parsed in accordance with
Oliverius (1976), and of the 986 words incorrectly stressed, 709 were incorrectly morphologically
parsed.
342
cause a word to be incorrectly stressed under one theory but not under another are
complex. However, it is possible to extract the generalization that the slight loss of
coverage caused by adopting the CFG appeared to be entirely attributable to the
existence of cases of the type mentioned in (177a). For example, Melvolds theory
explains the stress of !$"60
rn
+&#
sn
+,
in10
! by assigning !$"60!
rn
the lexical specification
[+post-accented], while !&#!
sn
is [accented]. The equivalent in the CFG would be to
stipulate that !$"60!
rn
is [+accented], with [stress] associating with !0!, and !&#!
sn
is
[accented]. However, !$"60
rn
+&#
sn
+,
in10
! would then be stressed on the first syllable:
!0&#! is not a possible onset, so !0! must syllabify as the coda of !$"60!.
In my opinion, it is more revealing to focus on the fact that the coverage of my
grammar is only very slightly less than that of Melvolds theory. I would argue that
the benefits of adopting the CFG outweigh the small cost in three major ways. First,
the CFG is simpler: by using syllable structure, I dispense both with the feature
[post-accented] and with the idea that jers have unusual accentual properties. The
CFG is also simpler in that it interprets plural stress retraction as resulting from a
lexical property of morphemes rather than stems, which means there is no need to list
stems separately in the lexicon: morphemes, of course, have to be listed anyway.
The second major advantage of the CFG is that it is arguably more in line with
general linguistic theory, because as we saw in section 2.2, the phonological
significance of syllable structure is independently motivated by a host of cross-
linguistic evidence, whereas it is less obvious what the cross-linguistic basis is for
assuming the existence of post-accentuation and jer stress retraction.
The third advantage of the CFG is that it fulfils functions above and beyond
those fulfilled by Melvolds theory: it not only assigns stress, but also assigns syllable
343
structure, morphological structure, enforces phonological constraints such as voicing
assimilation, and vocalizes and deletes jers, all within a single grammar.
5.4 Conclusions
In Chapter 1, I emphasized the fact that computer technology makes it
possible to test linguistic theories in a way which could not have been done before.
The merits of adopting a computational approach will by now be clear. In this
dissertation, I have tested a number of theories, compiled an extensive accentual
morphological lexicon, and built up coherent two-module grammar. None of this
would have been possible without computational tools. I hope that other (non-
computational) linguists may see what can be achieved, and be encouraged to use the
computational facilities which exist to subject their own theories to closer scrutiny in
the same way.
But this is a linguistics dissertation rather than an advertisement for
computational linguistics, and I hope that the results which I have achieved will be of
linguistic interest whether or not the reader is interested in computation. I have
examined a number of different approaches to a variety of areas of Russian phonology
(syllable structure and phonotactics; consonant-vowel interdependencies; vowel
reduction; word-final devoicing; voicing assimilation; word-formation; vowel-zero
alternations; and stress assignment), and drawn these together into a single context-
free grammar. Because all these areas are integrated into a single phonology, the
grammar is able to deal with situations where one area interacts with another. In this, I
would argue, the theory I propose is of more interest than the majority of theories
which deal with only a single aspect of Russian phonology. The grammar follows in
the spirit of generative linguistics, being formally explicit and aiming to capture
344
generalizations by rule where possible. However, this does not compromise the
grammars declarativeness: there is no rule-ordering, no feature- or structure-
changing, and the grammar is clearly reversible (that is, it models both the creative
and parsing aspects of language). This, of course, all follows from the fact that the
grammar is context-free.
In a number of areas, my computational phonology achieves high levels of
coverage in areas in which theories have not previously been tested:
- The CFG assigns syllable structure in 9851% of cases. To the best of my
knowledge, this is the first time that a level of coverage has been measured in
this way for Russian syllabification.
- The CFG assigns correct allophonic representations (within the limitations of
the grammars scope, and on the basis of comparison with phonetic
representations reported by Halle 1959) to all the cases tested.
- The CFG assigns morphological structure to 8401% of cases tested. It assigns
correct or nearly-correct morphological structure (on the basis of comparison
with morphological representations given in Oliverius 1976) to 6649% of the
same sample.
- The CFG assigns correct stress to 9507% of Russian word-forms to which a
correct morphological analysis had been assigned. There is a negligible
difference between this and the coverage of 9534% measured for my
implementation of Melvolds theory.
345
The key results of this dissertation are in the area of stress assignment. The
existing theories of stress assignment examined in Chapter 4 clearly make interesting
and non-trivial predictions about stress assignment in Russian, but the problem with
these theories is that they do not give all the information which is needed in order to
be able to evaluate them properly. I hope that it will no longer be regarded as
acceptable to assume a morphological lexicon and word-formation grammar, now that
I have produced a grammar which makes this information explicit.
I would argue that the other weakness of existing theories of stress assignment
in Russian is that they ignore the role of syllable structure. The lack of a formal means
of accounting for syllable structure has been a weakness of many generative analyses,
a fact which was first demonstrated thirty years ago in critiques of Chomsky and
Halle (1968). But it seems both that this weakness of generative linguistics still
persists, and that there are major areas of linguistics where the importance of syllable
structure has not yet been fully exploited (or in some cases even considered), Russian
stress assignment being a case in point. Contrary to the impression given by most
existing accounts of the subject, I have shown that syllabification plays an integral
role in Russian stress assignment, since the latter is best seen as a function of the
intersection of morphology and syllable structure. Not only is this analysis simpler
than that of the existing accounts, but it is shown to be just as general.
346
Appendix 1: Russian syllable structure grammar
Notes
Phrase-structure rules are in bold type. Feature matrices attached to a node are shown
below the node, in non-bold, non-italic type and enclosed in square brackets.
Auxiliary constraints are shown in italics and enclosed in braces. The notation is
based on Prolog DCG notation, but has been modified to be easily read by a linguist
not familiar with Prolog.
Cross-ref.
to text
above
Rule
(32),
(34)-(35)
prosodic_word weakfoot strongfoot
[right:R] [left:L]
{R [peripheral:+],
voicing_assim(L,R)}
(32),
(34)-(35)
prosodic_word strongfoot
(32),
(34)-(35),
(73)
weakfoot syllable weakfoot
[left:LL,right:RR] [left:LL, right:LR] left:RL,right:RR]
{LR [actual_stress:, stress:, peripheral:],
voicing_assim(RL,LR)}
(32),
(34)-(35),
(73)
weakfoot syllable
[left:L,right:R] [left:L, right:R]
{R [actual_stress:,stress:]}
(32), (34),
(72)
strongfoot syllable
[left:L,right:R] [left:L, right:R]
{R [actual_stress:+,stress:+,pl_retr:,vfv:]}
(32), (34),
(35), (72),
(84)-(85)
strongfoot syllable weakfoot
[left:LL, right:RR] [left:LL, right:LR] [left:RL,right:RR]
{LR [actual_stress:+,stress:+],
RR [pl_retr:,peripheral:,vfv:],
voicing_assim(RL,LR)}
(174) strongfoot syllable syllable
[left:LL,right:RR] [left:LL, right:LR] [left:RL, right:RR]
{LR [stress:,actual_stress:+],
RR [stress:+,actual_stress:,pl_retr:+,vfv:],
voicing_assim(RL,LR)}
(174) strongfoot syllable syllable weakfoot
[left:LLL, [left:LLL, [left:LRL, [left:RRL,
right:RRR] right:LLR] right:LRR] right:RRR]
{LLR [stress:,actual_stress:+],
LRR [stress:+,actual_stress:],
RRR [peripheral:,pl_retr:+,vfv:],
voicing_assim(RRL,LRR),
voicing_assim(LRL,LLR)}
347
(27), (29),
(34),
(55)-(56),
(73)
syllable onset rime
[left:LL, [left:LL, [left:RL,
right:RR] right:LR] right:RR]
{LR [stress:H2,peripheral:H3],
RL [stress:H2,peripheral:H3],
cv_interdependency(RL,LR)}
(27), (31),
(35)
syllable rime
[left:L, left:L,
right:R] right:R]
{L [onset:[]]}
(47) onset segment segment segment segment
[left:P1, [P1] [P2] [P3] [P4]
right:P4]
{P6 [sonor:,
cont:+,
coron:,
anter:+],
P2 [cont:+,
sonor:,
coron:+,
anter:+],
P3 [cont:,
del_rel:],
P4 [cons:+,
voc:+],
voicing_assim([sonor:+],P4),
voicing_assim(P4,P3),
voicing_assim(P3,P2),
voicing_assim(P2,P6),
cv_interdependency(P4,P3),
cv_interdependency(P3,P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H],
P3 [stress:H],
P4 [stress:H]}
348
(42) onset segment segment segment
[left:P1, [P1] [P2] [P3]
right:P3]
{P6 [cont:+,
sonor:,
coron:,
anter:+],
P2 [cont:+,
sonor:,
anter:+],
P3 [cons:+,
sonor:+],
voicing_assim([sonor:+],P3),
voicing_assim(P3,P2),
voicing_assim(P2,P6),
cv_interdependency(P3,P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H] ,
P3 [stress:H]}
(43) onset segment segment segment
[left:P1, [P1] [P2] [P3]
right:P3]
{P6 [cont:+,
sonor:,
coron:,
anter:+],
P2 [cont:+,
sonor:,
anter:+,
coron:+],
P3 [cont:,
del_rel:],
voicing_assim([sonor:+],P3),
voicing_assim(P3,P2),
voicing_assim(P2,P6),
cv_interdependency(P3,P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H],
P3 [stress:H]}
349
(45),
Table 13
onset segment segment segment
[left:P1, [P1], [P2] [P3]
right:P3]
{P1 [cont:+,
sonor:,
anter:+],
P2 [cont:,
del_rel:],
P3 [cons:+,
sonor:+,
nasal:],
voicing_assim([sonor:+],P3),
voicing_assim(P3,P2),
voicing_assim(P2,P6),
cv_interdependency(P3,P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H],
P3 [stress:H]}
(36),
Table 13
onset segment segment
[left:P1, [P1] [P2]
right:P2]
{P1 [cont:+,
sonor:,
coron:+,
anter:+],
P2 [cont:],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(37),
Table 13
onset segment segment
[left:P1, [P1] [P2]
right:P2],
{P1 [cont:+,
sonor:,
coron:+,
anter:],
P2 [cont:,
del_rel:],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
350
(38),
Table 13
onset segment segment
[left:P1, [P1] [P2]
right:P2]
{P1 [cont:+,
sonor:,
anter:+],
P2 [cons:+],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(33),
Table 13
onset segment segment
[left:P1, [P1] [P2]
right:P2]
{P1 [cons:+,
voc:,
del_rel:],
P2 [cons:+,
voc:+],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(40a) onset segment segment
[left:P1, [P1] [P2]
right:P2]
{P1 [sonor:,
del_rel:],
P2 [cons:+,
voc:,
sonor:+,
nasal:+],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(40b),
Table 13
onset segment segment
[left:P1, [P1] [P2]
right:P2]
{P1 [sonor:],
P2 [cons:+,
sonor:+,
nasal:],
voicing_assim([sonor:+],P2),
voicing_assim(P2,P6),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(28),
Table 13
onset segment
[left:P, [P]
right:P]
{P [cons:+],
voicing_assim([sonor:+],P)}
351
(27), (29) rime nucleus
[left:L, right:L] [left:L]
(27), (30),
(35), (73)
rime nucleus coda
[left:L, right:R] [left:L] [right:R]
{R [stress:H6,stress:H2,peripheral:H3],
L [stress:H6,stress:H2,peripheral:H3]}
(28),
Table 13
coda segment
[right:P] [P]
{P [cons:+],
voicing_assim(P,[sonor:+],
cv_interdependency([cons:+],P)}
(48),
Table 13
coda segment segment
[right:P2] [P1] [P2]
{P1 [cons:+,
voc:+],
P2 [cons:+,
voc:],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(49),
Table 13
coda segment segment
[right:P2] [P1] [P2]
{P1 [voc:,
del_rel:],
P2 [cont:,
coron:+,
anter:+,
del_rel:],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
352
(50) coda segment segment
[right:P2] [P1] [P2]
{P1 [cons:+,
voc:,
sonor:+,
nasal:+],
P2 [cont:+,
sonor:,
coron:+],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(50) coda segment segment
[right:P2] [P1] [P2]
{P6 [cons:+,
voc:,
sonor:+,
nasal:+],
P2 [cont:],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
Table 10 coda segment segment
[right:P2] [P1] [P2]
{P1 [cont:,
coron:],
P2 [cont:+,
sonor:,
coron:+,
anter:+],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
353
Table 10 coda segment segment
[right:P2] [P1] [P2]
{P1 [sonor:,
cont:+,
coron:+,
anter:+],
P2 [cont:,
coron:,
anter:],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
Table 10 coda segment segment
[right:P2] [P1] [P2]
{P1 [sonor:,
cont:+,
coron:+,
anter:],
P2 [cont:,
coron:+,
anter:],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
Table 10 coda segment segment
[right:P2] [P1] [P2]
{P1 [sonor:,
cont:+,
coron:+,
anter:+],
P2 [cons:+,
voc:,
anter:+],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
354
Table 10 coda segment segment
[right:P2] [P1] [P2]
{P1 [cont:,
del_rel:|,
P2 [cons:+,
voc:+,
coron:+],
voicing_assim(P2,P6),
voicing_assim(P6,[sonor:+],
cv_interdependency([cons:+],P2),
cv_interdependency(P2,P6),
P6 [stress:H],
P2 [stress:H]}
(28),
Table 13
nucleus segment
[left:P] [P]
{P [cons:, voc:+]}
segment phoneme
[F] [F1]
{F F6)}
(81)
/*Definition of voicing assimilation constraints*/
voicing_assim(X,Y):-
X [sonor:+],
Y [sonor:+,vfv:+].
voicing_assim(X,Y):-
X [sonor:+],
Y [sonor:,voice:V,vfv:V].
voicing_assim(X,Y):-
X [sonor:,vfv:V],
Y [vfv:V].
(56)
/*Definition of consonant-vowel interdependency constraints*/
cv_interdependency(X,Y):-
X [cons:,
back:Vowel_back,
high:Vowel_high,
onset: [back:Cons_back,
high:Cons_high]],
Y [cons:+,
high:Cons_high,
back:Cons_back,
rime: [high:Vowel_high,
back:Vowel_back]].
/*The following rule makes the simplifying assumption that a consonant is always
[rime:[+back]] when followed by another consonant.*/
cv_interdependency(X,Y):-
X [cons:+],
Y [cons:+,
rime: [back:+]].
355
Appendix 2: Russian word-formation grammar
The notation is as in Appendix 1. It is assumed that each node contains semantic
features such that the grammar is non-recursive, but the identity of these features is
unimportant in this dissertation.
Cross-ref.
to text
above
Rule
(by analogy
with (110))
word noun
[M] [M]
(110) word verb
(by analogy
with (110))
word adj
(by analogy
with (110))
word adv
(97) word part
(104) noun noun_stem in
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3),
pl_retr(F6,F2,F3)}
(111) verb verb_stem iv
[F1] [F2] [F3]
{F6 [stress:+],
F2 [reflexive:],
vowel_zero(F6,F2,F3)}
(113) verb verb /s'a/
c
[F1] [F2] [F3]
{F6 [reflexive:+],
F2 [reflexive:],
F3 [reflexive:+]}
(by analogy
with (104))
adj adj_stem ia
[F1] [F2] [F3]
{F6 [stress:+],
vowel_zero(F6,F2,F3)}
(by analogy
with (104))
adv adj_stem iau
[F1] [F2] [F3]
{F6 [stress:+],
vowel_zero(F6,F2,F3)}
(97) part c
(104) noun_stem rn
[F] [F]
(114) noun_stem c rn
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
356
(114) noun_stem noun_stem sn
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(104) noun_stem adj_stem sn
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(104) noun_stem verb_stem sn
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(104) adj_stem ra
[F] [F]
(by analogy
with (114))
adj_stem c ra
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(by analogy
with (104))
adj_stem adj_stem sa
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(105) adj_stem noun_stem sa
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(by analogy
with (104))
adj_stem verb_stem sa
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(104),
(113a)
verb_stem rv
[F] [F]
(113a) verb_stem c rv
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(111),
(113a)
verb_stem verb_stem sv
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(by analogy
with (104))
verb_stem adj_stem sv
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
357
(by analogy
with (105))
verb_stem noun_stem sv
[F1] [F2] [F3]
{assign_stress(F6,F2,F3),
vowel_zero(F6,F2,F3)}
(124) /E/ /c/
[strong:+]
(124) /E/ //
[strong:]
(124) /O/ /o/
[strong:+]
(124) /O/ //
[strong:]
(162)
assign_stress(X,Y,Z):-
X [acc:+,stress:S],
Y [stress:],
Z [dom:+,stress:S].
(165)
assign_stress(X,Y,Z):-
X [acc:+,stress:S],
Y [acc:+,stress:S],
Z [dom:].
(166)
assign_stress(X,Y,Z):-
X [acc:+,stress:S],
Y [acc:],
Z [acc:+,stress:S].
(167)
assign_stress(_,Y,Z):-
Y [acc:],
Z [acc:, dom:].
(119),
(120),
(122)
vowel_zero(X,Y,Z):
X [strong:P]
Y [strong:Q],
Z [strong:P, left:[high:Q,tense:]].
(119),
(120),
(122)
vowel_zero(X,Y,Z):-
X [strong:P],
Y [strong:],
Z [strong:P, left:[tense:+]].
(173)
pl_retr(X,Y,Z):-
X [pl_retr:+],
Y Z [plural:P, pl_retr:P].
(173)
pl_retr(X,Y,Z):-
X [pl_retr:],
Y Z [pl_retr:].
358
Appendix 3: Morphological inventory
Key to symbols:
The symbol / means that the next segment is [+stress] if the morpheme is [+stress].
The symbols O and E denote jers.
Morpheme Morpheme
frequency
Allomorph Allomorph
frequency
Accented Dominant Plural stress
retraction
!!
ip
8,738
!!
8,738
!!
ipu
12,467
!!
12,467
!!
ia1
443
!%!
443
!!
ia2
47,481
!!
47,481
!!
in1
56,855
!!
56,855
!!
sp
37,973
!!
37,973

!6A'!
14 +
!EA!
sn
447
!E/A!
433 + + +
!5$!
44 !E$!
sn
236
!E$!
44 +
!E4'!
sn3
102
!E4'!
102 +
!,4!
55
!E4!
12,621
!E4!
sa1
15,097
!E4'!
2,421
!8&8'!
163 +
!&8'!
46 +
!E#2!
sa
2,676
!E#2!
2,467
!E#&/"!
1,792 + !E#&"!
sn
2,519
!E#&"'!
727
!E/&#!
962 + !E&#!
sn
1,894
!E&8'!
932
!!
10,316
!O&8'!
463
!O2!
sn1
12,860
!O/2!
2,081 +
!-'!
275 !O-!
sa
1,008
!O-!
733
!"!
49 !O"'!
sn
134
!O/"'!
85 + +
!%!
ip
7,483
!%!
7,483
!%!
ipu
732
!/%!
732 +
!%!
in15
39,763
!%!
39,763
!%!
c1
4,020
!%!
4,020
!%!
svd
2,753
!%!
2,753 + +
!!
843
!/%!
16,539 + +
!/%7!
73 + +
!/%"!
28 + +
!%!
svi
17,907
!,"!
424

359
!%2!
22 !%$!
c
48
!%9!
26
!/%$!
67 + + !%$!
sn3
122
!%/$'!
55 +
!%$0'1#!
rn
37
!%$0'1#!
37
!%B%!
c
25
!%B%!
25
!%B6B!
ra
22
!%B/6B!
22 +
!%B0!
rn
36
!%B0!
36
!%7!
sn
212
!/%7!
212 +
!%78!
sa
47
!/178!
47 + +
!%2!
rv
96
!/%2!
96 +
!%2!
sn
120
!%/2!
120 + +
!%2!
6,626 !%2!
sp
6,742
!/%&8'!
116 +
!%2%$'1+'!
rn
23
!%2%$'/1+'!
23 +
!/%-!
230 + + !%-!
sa
615
!/%-'!
385 + +
!%+'10',2!
rn
85
!%+'10'/,2!
85 +
!%+',!
In5
50,920
!/%+',!
50,920 +
!%+!
in6
50,989
!/%+!
50,989 +
!%+!
sp
471
!%+!
471
!%4!
sa1
125
!%4!
96
!/%4!
231 + + !%4!
sa2
268
!7/%4'!
37 + +
!%0!
rn
99
!/%0!
99 +
!%0!
sa
130
!%/0'!
130 + +
!/%0!
124 + + !%0!
sn
212
!/%0'!
88 + +
!%0+'!
rn
97
!/%0+'!
97 +
!%0&'!
rn
45
!/%0&'!
45 +
!/%8!
27 + +
!/,8!
42 + +
!%8!
sn
165
!/68!
96 +
!18!
26 !%8!
sa2
99
!%8!
73
!/%8&8'!
303 + + !%8&8'!
sa
541
!;8&8'!
238
!%&!
sa
108
!/%&!
108 + +
!/%&!
847 + +
!/%&8'!
17 + +
!%&!
sn
966
!/64!
102 + +

!%/&8'!
100 + + !%&8'!
sa
121
!/;&8'!
21 + +
!%"!
sn
27
!%/"!
27 + +
!%"B;#&!
rn
32
!/%"B;#&!
32 +
!%"&!
ra1
93
!%"&!
93
!%"&!
ra2
57
!/%"&!
57 +
360
!%C!
in7
50,989
!%C!
50,989 +
!%C!
c
56
!/%C!
56
!%5!
sn
31
!/%5!
31 +
!%5$!
sp1
587
!/%5$!
587 + +
!%5$!
sp2
84
!/%5$!
84 + +
!A'1$!
rn
63
!A'1/$!
63 + +
!A'/1$!
70 + !A'1$!
rv
123
!A'1$'!
53
!A'/15!
247 + !A'1B!
rv
295
!A'/1B!
48 +
!A'1-!
ra
193
!A'/1-!
193 +
!A'10'1B!
rn
141
!A'10'1B!
141
!A'10'1&8'!
rv
19
!A'10'/1&8'!
19 +
!A'1#'1$!
rn
91
!A'1#'/1$!
91 +
!A'1*O!
c
412
!A'1*O!
412
!A'/,!
164 +
!A'/,"!
19 +
!A',!
rv
332
!A/67!
149 +
!A',A-',7!
rn
35
!A'/,A-',7!
35 +
!A',-'!
rn
45
!A'/,-'!
45 +
!A',-'!
sa
28
!A'/,-'!
28 +
!A'/,4!
68 + !A',4!
ra
128
!A/%74!
60 +
!A';0!
rn
43
!A';/0!
43 +
!A'/,0!
210 + !AO0!
rv
1,023
!AO0!
813
!A%A!
rn
90
!A/%A!
90 +
!A%-&'!
rn
31
!A/%-&'!
31 +
!A%"'!
rv
47
!A/%"'!
47 +
!A%*!
rn
39
!A/%*!
39 +
!A,!
c
1,820
!A,!
1,820
!A/,#&0!
220 + !A,#&0!
ra
265
!A/,#&0'!
45 +
!A-'1$!
ra
23
!A-'/1$!
23 +
!A-'1#&'!
rv
49
!A-'/1#&'!
49 +
!A-'/,5!
67 + !A-',*!
ra
156
!A-'/,*!
89 +
!A-';$!
rv
31
!A-'/;$!
31 +
!A6!
76
!A65!
22
!A6B!
rn
211
!A6B!
113

!A67!
rv
131
!A67!
131
!A62!
rn
37
!A62!
37
!A6-'!
ra
1,362
!A/6-'!
1,362 +
!A6-'!
rv
214
!A6-'!
214
!A6-!
rn
30
!A6-!
30
!A6-6&!
rn
46
!A6-/6&!
46 +
361
!A60'!
113 !A60'!
rv
156
!A60/6!
43 +
!A606$!
rn
15
!A60/6$!
15 +
!A60&!
rn
45
!A60&!
45
!A6&'!
rn
21
!A6&'!
21
!A0',B!
rv
260
!A0'/,B!
260 +
!A0/%&!
64 + !A0%&!
rn
129
!A0/%&7!
65 +
!A06$'!
rv
28
!A0/6$'!
28 +
!A0/6#!
42 + !A06#!
rv
170
!A0/6#'!
128 +
!A06"'!
rn
30
!A06"'!
30
!A/,!
4,451 +
!A/,"!
252 +
!A/;$!
304 +
!A;$!
rv
5,224
!A/;$'!
217 +
!A;?'1&!
rn
16
!A;?'/1&!
16 +
!A;2"!
rn
41
!A/;2"!
41 +
!A;+/%5!
51 + !A;+%B!
rn
153
!A;+/%B!
102 +
!A;0!
rv
15
!A/;0!
15 +
!A;05!
rn
29
!A/;05!
29 +
!A;#!
rn
35
!A;#!
35 +
!A;&,-!
rn
31
!A;&,-!
31 +
!$'!
sp
1,222
!$'!
1,222
!$'E4'!
rn
1,197
!$'E4'!
1,197
!$'%$'!
rn
100
!$'/%$'!
100 +
!$'14'EB!
rn
96
!$'14'EB!
96
!$'/1!
91 +
!$'1/7!
289 +
!$'1!
rv
394
!$'/1"!
14 +
!$'1!
c
115
!$'1!
115
!$'1$!
rn
24
!$'/1$!
24 +
!$'1-!
1,598 !$'1-!
rv1
1,687
!$'/1-'!
89 +
!$'/1-!
75 + !$'1-!
rv2
205
!$'/1-'!
130 +
!$'1+!
rn
29
!$'1+!
29
!$'10'1/"!
194 +
!$'10'1/"'!
46 +
!$'10'/1"7!
288 +

!$'10'1"!
rn
288
!$06"!
48
!$'105!
rv
189
!$'/105!
189 +
!$'1#'/%&!
443 +
!$'1#'/%&'!
153 +
!$&#/%&!
31 +
!$'1#'%&!
rn
1,330
!$&#/%&'!
703 +
362
!$'/1&!
128 + !$'1&!
rn
318
!$'1&'!
190
!$'1"'/%!
28 +
!$'1"'/%&!
66 +
!$'1"'%&!
rn
358
!$'1"'/%&'!
264 +
!$'1"!
rn
585
!$'/1"!
585 +
!$'15/;0!
49 + !$'15;0!
ra
73
!$'15/;0'!
24 +
!$',0'12!
rv
109
!$',0'12!
109
!$',#&#,9-',4!
rn
19
!$',#&#,9-'/,4!
19 +
!$'/,"'!
106 + !$',"'!
rv
185
!$'/,"-'!
79 +
!$',"%4!
rn
26
!$',"/%4!
26 +
!$/%!
824 + !$%!
rv
1,279
!$/%"!
455 +
!$%!
c
1,328
!$%!
1,328
!$%-'!
ra
572
!$%-'!
428
!$/%0!
101 + !$%0!
rn
150
!$/%0'!
49 +
!$/%0!
45 + !$%0!
rv
120
!$/%0'!
75 +
!$%"!
ra
228
!$/%"!
228 +
!$,+!
rn
25
!$,+!
25
!$-'%!
c
967
!$-'%!
967
!$4!
11 !$4!
rn
21
!$/64'7!
10 +
!$6!
c
1,450
!$6!
1,450
!$6A!
rv
153
!$/6A!
153 +
!$62!
rv
115
!$62!
115
!$6-'!
ra2
19
!$/6-'!
19 +
!$6-'!
rv
21
!$6-'!
21
!$/6-5!
148 + !$6-B!
ra
331
!$/6-B!
183 +
!$/6-5!
329 + !$6-B!
rn
355
!$6-B!
26
!$6+!
772 !$6+!
rn
820
!$/6+'!
48 +
!$606B!
ra
117
!$60/6B!
117 +
!$60/65!
48 + !$606B!
rn
282
!$60/6B!
234 +
!$6#O2!
rn
89
!$6#O/2!
89 +

!$/6&8'!
29 + !$6&8'!
rn
58
!$6&8'10'!
29
!$65$'!
rn
70
!$65/$'!
70 +
!$0'1"!
ra
19
!$0'/1"!
19 +
!$06A!
rv
18
!$0/6A!
18 +
!$06B!
rv
17
!$0/6B!
17 +
363
!$0/;5!
220 +
!$0;B!
554
!$0;*'!
554
!$0;B!
ra
1,881
!$0;*'7!
553
!$0;5!
rv
25
!$0/;5!
25 +
!$;!
rv
18
!$/;!
18 +
!$;2!
rv
79
!$/;2!
79 +
!$/;+!
717 + !$;+!
rv
736
!$/;+'!
19 +
!$;0!
ra
20
!$/;0!
20 +
!$;8!
rn
185
!$;/8!
185 + +
!$/,8!
29 +
!$,C!
87
!$/6C!
71 +
!$;/8!
185 +
!$;C!
rv
512
!$;C!
140
!$"'10'!
rn
320
!$"'10'!
320
!$"',5!
112 !$"',B!
rv
168
!$"',B!
56
!$"!
1,054 !$"!
rp
1,197
!$"'!
143
!$"60!
145 !$"60!
rn
190
!$"6/0'!
45 +
!1!
ipu
2,438
!/1!
2,438 +
!1!
in8
65,320
!1!
65,320
!1!
c
4,719
!1!
4,719
!1!
svd
3,685
!/1!
3,685 + +
!17!
in9
8,255
!/17!
8,255 +
!17!
c
16
!17!
16
!17!
429
!8!
1,980 +
!17!
sa
2,791
!&8'!
57
!12!
rn
24
!12!
24
!12#!
c
25
!12#!
25
!12*%+'14!
rn
45
!12*/%+'14!
45 +
!1-'/12&0!
35 + !1-'12&0!
rn
64
!1-'/12&0'!
29 +
!1-!
rp
36
!1-!
36
!14'!
sa
1,528
!14'!
1,528 + +
!14'10B!
ra
21
!14'/10B!
21 +
!19'10'!
sp
545
!19'/10'!
545 +

!/10!
619 + +
!10'!
98
!10!
sn
1,003
!/,0!
286 + +
!1#!
rp
230
!1#!
230
!1&',7!
sp
207
!1&',7!
207
!1&!
sn
149
!/1&!
149 + +
364
!1&%5!
rn
58
!1&%/5!
58 +
!1C!
c
53
!1C!
53
!?'10+!
rn
55
!?'/10+!
55 +
!?'1#!
rv
99
!?'1#!
99
!?'1"0%-'!
rn
29
!?'1"0%/-'!
29 +
!?',B!
rv
29
!?'/,B!
29 +
!?',&#!
rv
46
!?'/,&#!
46 +
!?',*'!
rn
45
!?'/,*'!
45 +
!?%A0'!
rn
111
!?/%A0'!
111 +
!?%2!
rv
45
!?/%2!
45 +
!?%2;-'!
ra
32
!?%2;-'!
32
!?%+',-'!
rn
48
!?%+'/,-'!
48 +
!?%8!
rn
47
!?/%8!
47 +
!?-%B!
rn
22
!?-/%B!
22 +
!?-6&!
rn
26
!?-/6&!
26 +
!?64!
rn
68
!?/64!
68 +
!?64%0'!
rn
33
!?64%/0'!
33 +
!?60+!
rn
89
!?/60+!
89 +
!?6&!
rn
41
!?6&!
41
!?0%4&#!
rn
31
!?0/%4&#!
31 +
!?0%*!
rn
26
!?0/%*!
26 +
!?064&!
rn
75
!?064/&!
75 +
!?;&!
rn
30
!?;&!
30
!B%$!
rv
27
!B/%$!
27 +
!B%-#&;2!
rn
38
!B/%-#&;2!
38 +
!B%*'1&!
rn
186
!B%*'/1&!
186 +
!B1!
rn
16
!B1!
16
!B12&!
rn
99
!B12&!
99
!B14'10!
rn
50
!B14'/10!
50 +
!B1067!
rn
109
!B/1067!
109 +
!B,A!
rv
47
!B/,A!
47 +
!B,B%4&!
rn
19
!B,B/%4&!
19 +
!B-'/%!
151 +
!B-'/%$!
155 +
!B-'%$!
rv
476
!B-'/%$'!
170 +
!B-%*!
rn
458
!B-%*!
458
!B-;A!
66 !B-;A!
ra
90
!B-;A'!
24
!B-;9!
ra
23
!B-/;9!
23 +
!B-;C!
ra
32
!B-;C!
32
!B6$!
rn
840
!B6$!
840

!B/6$!
45 + !B6$!
rv
65
!B/6$'!
20 +
!B6-!
ra
26
!B/6-!
26 +
365
!B-/%8!
30 +
!B-/%#!
27 +
!B-/%#'!
98 +
!B6-6#!
rn
381
!B6-6#!
226
!B-%/"!
220 + +
!B6-6"!
rn
513
!B6-6/"!
293 +
!B6-;A!
45 !B6-;A!
rn
104
!B6-;A'!
59
!5/%0!
32 + !B60'!
rv
206
!B/60'!
174 +
!B60',*64&!
rn
20
!B60',*/64&!
20 +
!B60!
rn
97
!B6/0!
97 +
!B60%*$!
ra
37
!B60/%*$!
37 +
!B/60$!
31 + !B60$!
ra
50
!B/60$'!
19 +
!B60-!
rn
20
!B/60-!
20 +
!B606$!
591
!B0/%5$!
55 +
!B606$!
rv
670
!B0/%$!
24 +
!B6#96/$!
29 + !B6#96$'!
rn
58
!B6#96/$'!
29 +
!B6#&'!
rn
174
!B6#&'!
174
!B6#;$!
rn
124
!B6#/;$!
124 +
!B6&/6"!
125 + !B6&6"!
ra
288
!B6&/6"'!
163 +
!B6"/%0'!
80 +
!B6"/60!
146 +
!B6"60!
rv
1,505
!B6"/60'!
1,279 +
!B0'/%*!
27 + !B0'%*!
rv
52
!B0'/%*'!
25 +
!B0%?'!
rv
57
!B0/%?'!
57 +
!B0%+!
rv
121
!B0/%+!
121 +
!B0%4'!
rn
22
!B0%4'!
22
!B06+!
ra
132
!B0/6+!
132 +
!B06+!
rv
99
!B0/6+!
99 +
!B0;A!
ra
20
!B0/;A!
20 +
!B0;$'!
rn
60
!B0;$'!
60
!B0;99!
rn
140
!B0/;99!
140 +
!B0;#&!
ra
30
!B0/;#&!
30 +
!B0;*!
ra
18
!B0/;*!
18 +
!B;A!
rn
54
!B;/A!
54 +

!B;-'!
rv
19
!B/;-'!
19 +
!B;#&!
ra
43
!B;#&!
43
!,!
ip
10,534
!,!
10,534
!,!
ipu
352
!,!
352
!,!
in10
69,487
!,!
69,487
!,!
c
14,596
!,!
14,596
366
!,!
svr
16,365
!/,!
11,967 + +
!,A6!
c
26
!,A6!
26
!,$'17!
rn
35
!,$'/17!
35 +
!/,$!
633 + !,$!
rv
3,048
!7!
2,415
!,B0!
rn
262
!,B/0!
262 + +
!,B0!
rv
262
!,B/0!
262 +
!,7!
ip
22,332
!/,7!
22,332 +
!/17!
32 + + !,7!
sn2
1,829
!/,7!
1,797 +
!,764!
sn
110
!,764!
110
!,7;-'!
rn
64
!,7/;-'!
64 +
!,7;4'!
rn
42
!,7/;4'!
42 +
!,/2!
1,379 + !,2!
sn
1,408
!,&8'!
29
!,2!
sn
355
!,2!
355 +
!,2!
sa1
127
!,2!
127
!/,2!
645 + !,2!
sa2
1,168
!/,&8'!
523 + +
!,-',!
c
493
!,-',!
493
!/%!
378 +
!/,+'!
208 +
!7/%!
186 +
!7/6+!
54 +
!4'/%!
957 +
!,+'!
rv
2,587
!4'/,+!
804 +
!/,+'!
176 + !,+'64!
rn
286
!,+'64!
110
!/,+!
167 + +
!/6+!
149 + +
!,+!
sa
393
!/6+'!
77 + +
!,4'!
sn4
22
!/,4'!
22 +
!,4!
rp
286
!,4!
286
!,4!
151 !,4!
c
183
!,4'!
32
!,4!
sn1
1,022
!/,4!
1,022 +
!,4!
sn2
74
!,/4!
74 + +
!,4&'10'1#!
rn
321
!,4&'10'/1#!
321 +
!,4514'!
rn
107
!,45/14'!
107 +
!/,#2!
75 + !,#2!
rv
95
!7,#2!
20

!,#&!
sa
16
!/,#&!
16 + +
!,#&60'!
rn
154
!,#&/60'!
154 +
!,8!
c
16
!,8!
16
!,8&8'!
sn
67
!/,8&8'!
67 +
!,&'!
sn2
32
!,&'!
32
367
!,&'/1&!
81 + + !,&'1&!
sn
113
!&'/1&!
32 + +
!,&!
sa
57
!,/&!
57 + +
!,&#!
rv
32
!/,&#!
32 +
!,&#!
sn1
524
!/,&#!
524 +
!,&8'!
sn
18
!/,&8'!
18 + +
!,"!
sa1
244
!/,"!
244 + +
!,"!
sa2
287
!/,"!
287 + +
!,*!
sv
860
!/,*!
860 + +
!,*O!
c
3,090
!,*O!
3,090
!17!
36
!%8!
1,483
!7!
sa
4,265
!67!
2,746
!/,7!
3,584 + !7!
sn
3,823
!/7!
239 +
!7%!
rp
5,242
!7/%!
5,242 +
!7%A-!
rn
43
!7/%A-!
43 +
!7%0!
ra
85
!7/%0!
85 +
!7%#!
ra
176
!7/%#!
176 +
!7%#!
rv
18
!7%#!
18
!7%8&8'!
ra
77
!7/%8&8'!
77 +
!7/%"'!
210 + !7%"'!
rv
370
!7/%"-'!
160 +
!7%*,2!
rn
125
!7%*,/2!
125 +
!71$'/,4!
69 +
!6$'/,4!
1,386 +
!71$',4!
rp
1,667
!/6$4!
212 +
!71$"%!
c
57
!71$"/%!
57 +
!712!
rv
27
!7/12!
27 +
!71-'1!
c
29
!71-'1!
29
!7/1#!
1,103 + !71#!
rv1
1,103
!7/1#!
1,103 +
!/1$!
78 + !71#!
rv2
199
!7/1#!
121 +
!718&8'6!
c
1,117
!718&8'/6!
1,117 +
!7/1C!
445 +
!7/1*5!
54 +
!71*/$!
180 +
!71*$!
rv
717
!7/1*$'!
38 +
!7151!
c
19
!7151!
19
!7;!
in13
3,797
!7;!
3,797

!7/;5!
29 +
!7;B!
rn
47
!7/;B!
18 +
!7;4!
ra
278
!7/;4!
278 +
!7;*!
rv
181
!7/;*!
181 +
368
!B!
615
!2!
7,660
!2&!
990
!&8'!
750
!2!
rp
16,327
!&8'&!
6,312
!2!
c
2,370
!2!
2,370
!2!
sa
354
!2!
354
!/,&8'!
38 + + !2!
sp
1,729
!2!
1,691
!/,&#!
136 + !2!
sn2
199
!2!
63
!2%!
rp
104
!2%!
104
!2%!
c
41
!2%!
41
!2%A'/,4!
24 + !2%A',4!
rn
97
!2%A'/,4'!
73 +
!2%$0!
rn
22
!2/%$0!
22 +
!2/%+'14!
29 +
!2/%+'14'!
56 +
!2%+'14!
rn
85
!2%+4'!
29
!2%4$',$!
ra
25
!2%4$'/,$!
25 +
!2%9'E-'!
rn
18
!2/%9'E-'!
18 +
!2%9',&!
rn
114
!2%9'/,&!
114 +
!2%0%4$%8!
rn
38
!2%0%4$%/8!
38 +
!2%0+%4!
rn
76
!2%0+/%4!
76 +
!2/%0&!
53 + !2%0&'!
rn
203
!2%0/&'!
150 +
!2%#!
rv
32
!2%#!
32
!2%&8'!
rv
22
!2/%&8'!
22 +
!2%*!
rv
2,912
!2/%*!
2,912 +
!2,!
rv
19
!2/,!
19 +
!2,-!
rn
129
!2,-!
129
!2,4!
rn
50
!2,/4!
50 +
!2,09',&8'!
rn
22
!2,09',/&8'!
22 +
!2,"!
rv
39
!2/,"!
39 +
!2-'1&!
rn
42
!2-'/1&!
42 +
!2-',2!
rv
27
!2-'/,2!
27 +
!2-';&8'!
rv
59
!2-';/&8'!
59 +
!2-/%$!
154 + !2-%$!
rv
183
!2-/%#!
29 +
!2-%##!
rn
410
!2-/%##!
410 +
!2-;A!
rn
87
!2-/;A!
87 +

!24'/,5!
66 + !24',B!
rn
295
!24'/,B!
229 +
!26-'/14!
31 + !26-'14!
rn
61
!26-'/14'!
30 +
!26-!
109 !26-'6#!
rn
149
!26-'6/#!
40 +
369
!26-!
rp
518
!26-!
518
!26+!
rp
650
!26+!
650
!26+!
rn
47
!2/6+!
47 +
!26!
406
!2/6-!
123 +
!26+!
128
!26+!
c
680
!2/60!
23 +
!26++;4'!
ra
246
!26++/;4'!
246 +
!26+4%&!
rn
309
!2/6+4%&!
309 +
!264'!
rn
34
!264'!
34 +
!2/%4!
15 +
!264!
230
!264!
ra
831
!264'!
586
!264&60!
rn
23
!264&/60!
23 +
!264&06-'!
rn
25
!264&0/6-'!
25 +
!264&#10&!
rn
56
!264&#/10&!
56 +
!260'!
rn
18
!2/60'!
18 +
!260',$!
rv
57
!260',/$!
57 +
!260%A-'!
rn
54
!260%A/-'!
54 +
!2/60+!
26 + !260+!
rn
43
!2/60+'!
17 +
!2606&!
ra
75
!2606&!
75
!2606"!
rn
56
!260/6"!
56 +
!2609;#!
rn
40
!2/609;#!
40 +
!26#&'O0!
rn
62
!26#&'O/0!
62 +
!26#&';+!
rn
42
!26#&'/;+!
42 +
!265!
rn
16
!2/65!
16 +
!20'1!
rv
155
!20'/1!
155 +
!20'19!
ra
108
!20'/19!
108 +
!20'1#'!
rv
17
!20'/1#'!
17 +
!20'1#-!
rn
46
!20'/1#-!
46 +
!20'1#&'!
rn
37
!20'1#&'!
37
!20'/,2!
82 + !20',2!
rv
262
!20'/,&8'!
180 +
!20%7!
rv
89
!20%7!
89
!20%/#!
347 + !20%#!
ra
456
!20/%#'!
109 +
!20%&'!
ra
29
!20/%&'!
29 +
!20/,!
435 + !20,!
rv
493
!20/,"!
58 +
!20,/-!
38 +

!20,-'!
15
!20,-!
rn
67
!20/,-7!
14 +
!206+'!
rn
88
!20/6+'!
88 +
!206+!
ra
17
!20/6+!
17 +
!206"'!
rn
33
!206"'!
33
!206"%&'!
rn
56
!206"/%&'!
56 +
370
!20;5!
124 !20;B!
rv
329
!20;B!
205
!20;9!
rn
87
!20/;9!
87 +
!2;2;0;*!
rn
32
!2;2;0/;*!
32 +
!2;-'!
rv
141
!2/;-'!
141 +
!2;-%2!
rn
35
!2;-%/2!
35 +
!2/;9!
22 + !2;9!
rv
77
!2/;9'!
55 +
!2;0!
rv1
84
!2/;0!
84 +
!2;0'!
rv
37
!2/;0'!
37 +
!2;0!
rv2
22
!2/;0!
22 +
!2;0&!
ra
18
!2/;0&!
18 +
!2;#!
rv1
289
!2/;#!
289 +
!2;#!
rv2
65
!2;#!
65
!2;#&!
rn
36
!2;#/&!
36 +
!2;&8'!
rv
39
!2/;&8'!
39 +
!2;C!
rv
66
!2/;C!
66 +
!2"%$0!
ra
20
!2"/%$0!
20 +
!2"%0&'!
ra
120
!2"/%0&'!
120 +
!-'E$!
rn
76
!-'E$!
76
!-'/15!
46 +
!-'6/B!
196 +
!-'*!
98
!-'1B!
ra
499
!-'*'!
159
!-'17&'14!
rv
30
!-'17&'/14!
30 +
!-'12!
rv
156
!-'/12!
156 +
!-'14'!
rn
76
!-'/14'!
76 +
!-'14&!
rn
19
!-'/14&!
19 +
!-'1#!
rn
234
!-'1#!
234
!-'/1&'!
94 + !-'1&'!
rv
199
!-'/6&!
105 +
!-'1&!
rn
579
!-'1&!
579
!-'1"!
ra
46
!-'/1"!
46 +
!-'/1#!
42 + !-'1*!
rv
118
!-'/1*!
76 +
!-',!
c
1,074
!-',!
1,074
!-',A6!
c
26
!-',A6!
26
!-',7!
rv
23
!-'/,7!
23 +
!-',2!
21
!-'/,&8'!
222 +
!-',2!
rn
505
!-',/&#!
262 + +
!-',4!
ra
56
!-'/,4'!
56 +
!-',#&!
47 !-',#&!
rn
93
!-'/,#&7!
46 +
!-',8!
ra
91
!-'/,8!
91 +
!-',8!
c
118
!-',8!
118
!-',&'10!
rn
50
!-',&'/10!
50 +
371
!-';A!
178 !-';A!
rv
497
!-'/;A'!
319 +
!-';$'!
rn
558
!-';$'!
558
!-!
381 !-!
sn1
788
!-'!
407
!-OA!
rn
53
!-OA!
53
!-%A60!
rn
20
!-%A/60!
20 +
!-/%$!
66 + !-%$!
rv
86
!-/%$'!
20 +
!-%$64'!
rn
41
!-%$/64'!
41 +
!-'/15!
178 +
!-'/1&8'!
49 +
!-%B!
174
!-%B!
rv1
838
!-/65!
437 +
!-%B!
rv2
98
!-/%B!
98 +
!-%2!
rv
31
!-%2!
31
!-%+9!
rn
49
!-/%+9!
49 +
!-%9!
rn
26
!-/%9!
26 +
!-%#2!
rv
24
!-/%#2!
24 +
!-,A!
rv
217
!-/,A!
217 +
!-6$!
rn
52
!-/6$!
52 +
!-62O&'!
rn
31
!-62O&'!
31
!-69%&!
rn
25
!-69/%&!
25 +
!-68%$'!
rn
69
!-68%$'!
69
!-/6"!
19 + !-6"!
rv
45
!-/6"'!
26 +
!-;4!
rn
58
!-;/4!
58 + +
!-;&8'!
ra
332
!-/;&8'!
332 +
!-;&8'!
rn
41
!-;/&8'!
41 +
!-;&8'!
rv
677
!-/;&8'!
677 +
!+'%B!
ra
28
!+'/%B!
28 +
!+'%#!
rn
39
!+'/%#!
39 +
!+'%&8'!
rn
44
!+'/%&8'!
44 +
!+'1$-'!
ra
89
!+'/1$-'!
89 +
!+'1-'2!
rv
15
!+'/1-'2!
15 +
!+'1-!
rv
70
!+'/1-!
70 +
!+'14'!
ra
57
!+'/14'!
57 +
!+'/14!
114 + !+'14!
rv
192
!+'/14'!
78 +
!+'/14&!
22 + + !+'14&!
sn
57
!;+'/14&!
35 + +

!+'/10!
39 +
!+'10'/1!
37 +
!+'10'1!
rv
94
!+'/60!
18 +
!+'10!
rv
246
!+'/10!
246 +
!+'1#'%&#!
rn
145
!+'/1#'%&#!
145 +
372
!+'/18&8'!
38 +
!+'1#&!
435
!+'1#&!
rn
752
!+'/1#&'!
279 +
!+'18!
rn
35
!+'18!
35
!+'18!
rv
93
!+'/18!
93 +
!+'/1&!
24 + !+'1&!
rv1
49
!+'/1&!
25 +
!+'/1&!
19 +
!+'/1&'!
164 +
!+'1&!
rv2
304
!+'/1&8'!
121 +
!+'1&/%--!
57 + !+'1&%--!
rn
92
!+'1&/%--'!
35 +
!+'1&6$!
rn
23
!+'/1&6$!
23 +
!+'1&0!
rn
152
!+'/1&0!
152 +
!+'1&8'&!
rn
121
!+'/1&8'/&!
121 +
!+'1C%4'!
rn
114
!+'1C/%4'!
114 +
!+'15$!
rn
215
!+'/15$!
215 +
!+',!
in14
69
!+'/,!
69 +
!+',B!
rv
16
!+'/,B!
16 +
!+',20!
ra
17
!+',20!
17
!+',-!
ra
54
!+'/,-!
54 +
!+',--'!
rn
67
!+',--'!
67
!+',+6!
c
66
!+',+6!
66
!+',4',#!
ra
50
!+',4',/#!
50 +
!+',4!
ra
54
!+',4!
54
!+'/%!
77 +
!+',4!
105
!+',4!
rv
537
!+4'!
355
!+',4;&!
rn
219
!+',4/;&!
219 +
!+',0!
rn
255
!+',0!
255
!+',#!
rv
25
!+'/,#!
25 +
!+',&'!
rv
150
!+'/,&'!
150 +
!+!
rp1
608
!+!
608
!+!
rp2
2,616
!+!
2,616
!+!
sn1
251
!/+!
251 + + +
!+!
sn2
225
!+!
225
!+%B%*',4!
rn
41
!+%B%*'/,4!
41 +
!+%7!
rn
35
!+/%7!
35 +
!+/%-!
256 + !+%-!
ra
849
!+/%-'!
593 +
!+%+!
rn
285
!+/%+!
285 +

!+'/14!
67 + !+%4!
rn
85
!+/%4!
18 +
!+/%4$!
152 + !+%4$!
rv
231
!+/%4$'!
79 +
!+%0&!
rn
30
!+/%0&!
30 +
!+%08!
rv
15
!+/%08!
15 +
373
!+%#!
ra
214
!+%#!
214
!+%#!
rv
31
!+/%#!
31 +
!+%##!
rn
30
!+/%##!
30 +
!+%8,4!
rn
317
!+%8/,4!
317 +
!+%&'!
rn
247
!+/%&'!
247 +
!+%&'1+%&'!
rn
16
!+%&'1+/%&'!
16 +
!+%&'10'!
rn
96
!+%&'10'!
96
!+%&!
ra
30
!+/%&!
30 +
!+%&06#!
rn
74
!+%&0/6#!
74 +
!+%C!
rv
21
!+/%C!
21 +
!+,!
rv
19
!+/,!
19 +
!+/,8-'!
42 +
!+/,#-!
45 +
!+,#-!
rn
237
!+/,#-'!
150 +
!+4/65!
59 + !+46B!
ra
805
!+4/6B!
746 +
!+6!
rv
28
!+6!
28
!+6/8&8'!
217 +
!+/65!
548 +
!+/6B!
138 +
!+6B!
rv
2,221
!+/6&8'!
1,318 +
!+62!
rv
29
!+/62!
29 +
!+/6-!
156 + !+6-!
rn
650
!+/6-'!
494 +
!+6-!
c
22
!+6-!
22
!+-/%$!
39 +
!+6-6$!
397
!+6-6$!
ra
858
!+6-6$'!
422
!+6-62!
rn
103
!+6-6/2!
103 +
!+6-6&!
rn
17
!+6-6&!
17
!+6-&8'!
rv
252
!+/6-&8'!
252 +
!+6+'14&!
rn
42
!+6+'/14&!
42 +
!+64&!
rv
17
!+/64&!
17 +
!+60!
58 !+60!
rn
230
!+60'!
172
!+606*!
rv
35
!+60/6*!
35 +
!+6#26"!
rn
59
!+6#2/6"!
59 +
!+6#&!
rn
38
!+6#/&!
38 +
!+6&!
ra
40
!+6/&!
40 +
!+&8'!
rv
17
!+&8'!
17
!+;8&8'!
rv
14
!+/;8&8'!
14 +

!+/;*!
59 + !+;*!
rn
91
!+/;*'!
32 +
!+;5!
61 !+;5!
rn
122
!+;57!
61
!4'1!
c
7,735
!4'1!
7,735
374
!4'/1A!
45 + !4'1A!
rn
91
!4'1A'1/#!
46 +
!4'/1+!
53 + !4'1+'!
ra
106
!4'1+'!
53
!4'1#!
rv
299
!4'/1#!
299 +
!4'1&!
c
763
!4'1&!
763
!4'15!
ra
21
!4'/15!
21 +
!4',!
c
1,702
!4',!
1,702
!4',2!
rv
17
!4'/,2!
17 +
!4',*!
ra
107
!4'/,*!
107 +
!4!
rp
1,157
!4!
1,157
!4!
sa2
1,471
!4'!
1,471
!4%!
c
11,313
!4%!
11,313
!4%$!
ra
576
!4/%$!
576 +
!4%$!
c
341
!4%$!
341
!4%,!
c
32
!4%,!
32
!4%&#!
rv
43
!4/%&#!
43 +
!4/,4!
57 + !4,4!
ra
83
!4/,4'!
26 +
!46!
c
1,966
!46!
1,966
!46/5!
22 + !46B!
rn
221
!46/B!
199 +
!467%A0'!
rn
19
!467%A/0'!
19 +
!46+'10!
rn
109
!46+'10!
109
!4/6+!
36 + !46+!
ra
60
!4/6+'!
24 +
!468!
91 !46#'!
rv
227
!4/6#'!
136 +
!46#!
rn
97
!46#!
97
!46#&!
rn
28
!4/6#&!
28 +
!46&8'!
rn
249
!46&8'!
249
!46"!
948 !46"!
ra
1,027
!4/6"'!
79 +
!465!
rv
24
!46/5!
24 +
!40%"'!
rv
138
!40/%"'!
138 +
!4&!
193 !4&!
sa
306
!4&#!
113 +
!4;!
c
711
!4;!
711
!4!
115
!4%"-'!
24
!4/6"!
40 + +

!46"'!
236
!46"-'!
25
!4;!
svc
1,585
!4/;!
1,145 +
!4;&0'!
rn
28
!4/;&0'!
28 +
!4/;5!
402 + !4;5!
rv
420
!4;5/$!
18 +
375
!4;5$!
rn
18
!4;5/$!
18 + +
!6!
ip
990
!/6!
990 +
!6!
ipu
10,232
!/6!
10,232 +
!6!
in18
19,598
!6!
19,598
!6!
4,002 !6!
c1
5,385
!6A'!
380
!6A!
rp
96
!/6A!
96 +
!6A8&8'!
ra
442
!/6A8&8'!
442 +
!6B!
sn
20
!/6B!
20 +
!6BO4'!
rn
92
!6BO/4'!
92 +
!6B$!
sp
1,702
!6B$!
1,702
!67!
ip
6,547
!/67!
6,547 +
!67!
in11
15,808
!67!
15,808
!67!
c
42
!67!
42
!62/4!
107 + +
!62/6!
21 +
!626!
rn
235
!6264!
107
!62&'%A0'!
rn
32
!62&'%A/0'!
32 +
!6-'!
sp
1,729
!/6-'!
1,729 +
!6+!
in19
33,898
!6+!
33,898
!64'!
sa
325
!/64'!
325 +
!/64!
9,199 + !64!
rp
12,252
!/64'!
3,053 +
!69'%&'!
c
174
!69'/%&'!
174 +
!69'10!
rv
29
!69'/10!
29 +
!60!
sa
16
!/60!
16 + +
!60!
509 !60!
sn
603
!/60'!
94 + +
!60B%4'!
rn
255
!60B/%4'!
255 +
!6021#&0!
rn
22
!602/1#&0!
22 +
!60;$!
rn
30
!60/;$!
30 +
!/6#'14!
24 + !6#'14!
rn
72
!6#'/14'!
48 +
!6#46"!
rn
71
!6#4/6"!
71 +
!6#&!
53 !6#&!
sn
661
!6#&'!
608
!6#&0!
ra
37
!/6#&0!
37 +
!6#&06"!
rn
34
!6#&06"!
34
!6&'/1&8'!
21 + !6&'E&#!
rn
231
!6&'E/&#!
210 +
!6&!
ip
6,559
!6&!
6,559

!6&!
1,485 +
!6&'!
20 +
!6&!
sn
1,881
!6&8'!
376
!6&O!
c
3,182
!6&O!
3,182
!6&60!
sp
1,348
!6&/60!
1,348 +
!6&8'!
rn
51
!6&8'!
51
376
!6&8'14'!
c
742
!6&8'14'!
742
!6"!
in12
23,238
!/6"!
23,238 +
!,4!
55
!6"!
1,923
!6/"'!
56 + +
!6"!
sa
2,328
!;7!
75
!6C!
c
24
!6C!
24
!6*'60!
rn
37
!/6*'60!
37 +
!6*!
sa
111
!/6*!
111 + +
!65!
sn
696
!/65!
696 + +
!9'%&'E4!
rn
32
!9'%&'E/4!
32 + +
!9'/%&!
191 + !9'%&!
rn
689
!9'/%&'!
498 +
!9'1!
rv
312
!9'/1!
312 +
!9'1$!
rn
22
!9'1$!
22
!9'10'1!
374 !9'10'1!
c
570
!90'1!
196
!9'10',6$!
rn
17
!9'10',/6$!
17 +
!9'10'1$!
336
!9'10'1$'!
66
!9'10'/6$!
9 +
!90'15!
51
!90'15$'!
99
!90'1$!
640
!9'10'6$!
c
1,302
!90'1$'!
19
!9'10!
rv
30
!9'1/0!
30 +
!9'10"!
rp
934
!9'/10"!
934 +
!9'1#!
ra
43
!9'1#!
43
!9'18!
ra
17
!9'18!
17
!9'1&8'!
rv
117
!9'/1&8'!
117 +
!9'1&8'%&!
rn
69
!9'1&8'/%&!
69 +
!9',!
rv
75
!9'/,!
75 +
!9',64'!
ra
466
!9',/64'!
466 +
!9',#!
520 !9',#!
rv
808
!9'/,#'!
288 +
!9',&!
rv
85
!9',&!
85
!9'71#!
rn
26
!9'7/1#!
26 +
!9!
rv
19
!9!
19
!9/%$!
87 + !9%$!
rv
250
!9/%#!
163 +
!9%-'&!
rn
59
!9%-'/&!
59 +

!9/%-!
21 + !9%-!
ra
88
!9/%-'!
67 +
!9%-%&!
rn
28
!9%-/%&!
28 +
!9%9!
rn
131
!9/%9!
131 +
!9%0'E4'!
rn
109
!9%0'E4'!
109
!9%0!
rn
32
!9/%0!
32 +
377
!9%0!
rv1
52
!9%0!
52
!9%0!
rv2
26
!9%0!
26
!9%02!
rn
30
!9/%02!
30 +
!9/%0&!
61 + !9%0&!
ra
290
!9/%0&'!
229 +
!9%#!
50 !9%#!
rv
126
!9%#'!
76
!9%##!
rv
31
!9/%##!
31 +
!9%;*!
rn
85
!9/%;*!
85 +
!9%C!
rv
55
!9%C!
55
!9,-'!
rn
25
!9/,-'!
25 +
!9,&!
rv
247
!9,&!
247
!9,C!
rv
21
!9/,C!
21 +
!9-'14!
ra
69
!9-'/14!
69 +
!9-'1&8'!
rn
126
!9-'1/&8'!
126 + +
!9-',&!
rn
18
!9-',/&!
18 +
!9-%2!
rv
49
!9-/%2!
49 +
!9-%2%&!
rn
52
!9-%2/%&!
52 +
!9-/%+'!
10 + !9-%+'!
rn
19
!9-%+'64!
9
!9-%4!
rn
172
!9-/%4!
172 +
!9-%&'!
rv
24
!9-/%&'!
24 +
!9-%&!
63 !9-%&!
rn
107
!9-/%&'!
44 +
!9-%"!
rv
26
!9-/%"!
26 +
!9-,!
rv
23
!9-/,!
23 +
!9-6$!
rn
18
!9-6/$!
18 +
!9-6&!
rn
22
!9-/6&!
22 +
!9-68&8'!
ra
105
!9-/68&8'!
105 +
!9-6C!
ra
153
!9-6C!
153
!9/%!
77 +
!96!
c
10,589
!96!
10,512
!96$O!
c
1,297
!96$O!
1,297
!961!
rv
32
!96/1!
32 +
!962%!
c
195
!962/%!
195 +
!96267!
rn
187
!962/67!
187 +
!96-'!
rn1
194
!96-'!
194
!96-'!
rn2
32
!9/6-'!
32 +
!96-!
ra
105
!9/6-!
105 +
!96-!
rn
95
!96-!
95
!9/6-4!
162 +
!96-4!
ra
390
!9/6-4'!
228 +
!96-6#!
rn
39
!96-6/#!
39 +
!96-6&4!
rn
21
!96-6&/4!
21 + +
!9/60!
227 + !960!
rv
264
!9/60'!
37 +
!9606B!
rn
37
!960/6B!
37 +
378
!960&?'1-'!
rn
24
!960&?'/1-'!
24 +
!960&0'1&!
rn
58
!960&0'/1&!
58 +
!96#!
rv
25
!96/#!
25 +
!96#-'1!
c
331
!96#-'1!
331
!96&8'&',!
c
201
!96&8'&'/,!
201 +
!96&8'"!
rn
20
!9/6&8'"!
20 +
!9/6*5!
24 + !96*$!
ra
82
!9/6*$!
58 +
!90'%+!
ra
161
!90'/%+!
161 +
!90'%&!
rv
25
!90'/%&!
25 +
!90'%5!
rv
16
!90'/%5!
16 +
!90'1+'!
ra
19
!90'/1+'!
19 +
!90',!
c
2,658
!90',!
2,658
!90',4&#,9!
rn
20
!90'/,4&#,9!
20 +
!90%2!
rv
27
!90/%2!
27 +
!90%"!
377
!90%"'!
372
!90%"!
rv
865
!90/%"-'!
116 +
!90%*$!
ra
141
!90/%*$!
141 +
!90/,5!
21 + !90,B!
rv
42
!90/,B!
21 +
!906!
c1
1,828
!906!
1,828
!906!
c2
324
!906!
324
!906A!
rv
71
!90/6A!
71 +
!906A-'1+!
rn
21
!906A-'/1+!
21 +
!90/%8!
154 +
!90/6#!
253 +
!906#!
rv
1,016
!90/6#'!
609 +
!9068&8'!
16
!906#&!
290
!906#&!
ra
346
!906#&'!
40
!906&',"!
c
135
!906&',"!
135
!906&8'!
ra
98
!90/6&8'!
98 +
!9&'!
rn
84
!9&'!
84
!9;A-'!
ra
225
!9/;A-'!
225 +
!9;B!
rv
46
!9/;B!
46 +
!9;42&!
rn
21
!9/;42&!
21 +
!9;#2!
102
!9;#&!
59
!9;#&!
rv
515
!9;#&'!
354
!9/;&!
49 +
!9;&!
rn
294
!9;/&'!
245 +
!9;&!
rv
40
!9/;&!
40 +
!0'/%5!
15 + !0'%$!
rv
568
!0'/%$!
553 +
!0'1!
196 !0'1!
c
253
!0'1$!
57
379
!0'1A'!
rn
913
!0'/1A'!
913 +
!0'1$!
ra
35
!0'/1$!
35 +
!0'12!
rn
76
!0'1/2!
76 +
!0'/1!
16 + !0'1#!
rn
241
!0'/1#!
225 +
!0'18!
rv
478
!0'/18!
478 +
!0'1"!
ra
126
!0'/1"!
126
!0'1*',4!
rn
21
!0'1*'/,4!
21 +
!0',#!
rv
57
!0'/,#!
57 +
!0!
sa
128
!0!
128
!0O&!
rn
38
!0O/&!
38 +
!0%A!
rv
1,629
!0%/A!
1,629 +
!0%$',!
c
24
!0%$',!
24
!0/%$'!
32 + !0%$',7!
rn
75
!0/%$',7!
43 +
!0%$!
ra
230
!0/%$!
230 +
!0%7!
rp
47
!0%7!
47
!0%764!
rn
180
!0%7/64!
180 +
!0%2!
rn
31
!0/%2!
31 +
!0%4'!
rn
23
!0/%4'!
23 +
!0/%4!
63 + !0%4!
ra
211
!0/%4'!
148 +
!0/%#!
153 +
!0%#&!
31
!0/%#&'!
28 +
!0/6#!
56 +
!0%#!
rv
321
!06#&!
53
!0/%"!
117 +
!0/6"!
53 +
!0%"!
ra
219
!06"'!
49
!0'/1*!
66 +
!0/%5!
69 +
!0%*!
1,079
!0%*!
rv
1,239
!0/%*'!
25 +
!0%*O!
c
1,406
!0%*O!
1,406
!0%*"'1!
c
126
!0%*"'1!
126
!0,A!
rn
55
!0/,A!
55 +
!0/,"!
19 + !0,"!
rv
37
!0/"!
18 +
!0,5!
ra
27
!0/,5!
27 +
!06A!
rv
24
!0/6A!
24 +

!065!
30
!0/65$'!
37 +
!06/$!
632 +
!06$!
rv
878
!0/6$'!
179 +
380
!0'1&8'!
69
!0/62!
193 +
!062!
rv
284
!0/6&8'!
22 +
!06-'!
rn
49
!06-'!
49
!06*!
rn
22
!0/6*!
22 +
!0;/A!
100 + !0;A!
rv
121
!0;/A'!
21 +
!0;/2!
905 + !0;2!
rn
964
!0/;&8'!
59 +
!0;#!
rn
124
!0/;#!
124 +
!0;5!
rn
19
!0;5!
19 +
!#'!
1,188 !#'!
rp
1,765
!*!
577
!#'!
361 !#'%!
c
7,557
!#'%!
7,196
!#'1!
c
155
!#'1!
155
!#'1$!
ra
23
!#'1$!
23
!#'12;4$!
rv
59
!#'12/;4$!
59 +
!#'1-!
rn
54
!#'1/-!
54 + +
!#'1+'E7!
rn
99
!#'1+'E/7!
99 + +
!#'1$'+!
96
!#'1+!
21
!#'1+!
rn
417
!#'1+'!
300
!#'14!
rn
22
!#'14!
22
!#'14&'%A0'!
rn
26
!#'14&'/%A0'!
26 +
!#'10'1/$'!
18 +
!#0'/1$!
163 +
!#'10'1$'!
rn
328
!#0'/1$'!
147 +
!#'10'7!
ra
111
!#'/10'7!
111 +
!#'10!
ra
51
!#'/10!
51 +
!#'10$!
125 !#'10$!
rn
214
!#'10$'!
89
!#'/6#&'0!
24 + !#'1#&'O0!
rn
49
!#'1#&'O/0!
25 +
!#'1"'10!
rn
29
!#'/1"'10!
29 +
!#',B4!
rn
41
!#'/,B4!
41 +
!#'/,-!
245 + !#',-!
rn
421
!#'/,-'!
176 +
!#',4'!
ra
54
!#'/,4'!
54 +
!#',#&'1+!
rn
42
!#',#&'/1+!
42 +
!#'/1-'!
121 +
!#'6-!
rv
205
!#'6/-!
84 + +
!*!
67 !#O!
c
14,124
!#O!
14,057
!#/&'!
83 + !#O&!
rn
705
!#O/&!
622 +
381
!#'/1$!
157 +
!#'/1$'!
157 +
!#'/1#!
167 +
!#'/,$'!
298 +
!#%$!
132
!#%$!
rv
1,043
!#/%$'!
132 +
!#%+!
ra
1,381
!#/%+!
1,381 +
!#%96B!
rn
40
!#%96/B!
40 +
!#%0%7!
rn
28
!#%0/%7!
28 +
!#/,-!
22 + !#,-!
rv
96
!#-!
74
!#/,4!
44 +
!#,46"'!
43
!#,4!
rn
130
!#,46"'7!
43
!#2%+'17!
rn
28
!#2%+'/17!
28 +
!#260!
172 !#260!
ra
314
!#2/60'!
142 +
!#26&8'!
rv
74
!#2/6&8'!
74 +
!#2"6*'!
c
46
!#2"6*'!
46
!#-'1$!
463 !#-'1$!
rv
520
!#-'/1$'!
57 +
!#-'1#!
rv
52
!#-'/1#!
52 +
!#-'6*!
rn
49
!#-'6/*!
49 +
!#-%A!
ra
29
!#-/%A!
29 +
!#-67!
rn
20
!#-67!
20
!#-/%"!
61 +
!#-6"!
601
!#-6"!
rn
716
!#-/6"'!
54 +
!#-/,8!
317 +
!#-/;8!
274 +
!#-;8!
rv
615
!#-/;C!
24 +
!#-;5!
rv
72
!#-/;5!
72 +
!#+'1!
99 !#+'1!
rv1
247
!#+'17!
148
!#+'1!
rv2
95
!#+'/1!
95 +
!#+/%&0'!
26 + !#+6&0'!
rv
696
!#+/6&0'!
670 +
!#4'/15!
26 + !#4'1B!
rn
152
!#4'1B!
126
!#6!
rn
650
!#6!
650
!#'/1A'!
781 +

!#/6A!
222 +
!#6A!
rp
1,198
!#/6A'!
195 +
!#6A%2!
rn
52
!#6A/%2!
52 +
!#6-$%&!
rn
53
!#6-$/%&!
53 +
!#/6-4!
116 + !#6-4!
rn
150
!#/6-4'!
34 +
382
!#6062!
rn
172
!#/6062!
172 +
!#6#4!
rn
25
!#6#/4!
25 + +
!#6&#!
ra
84
!#/6&#!
84 +
!#6"!
rp
74
!#6"!
74
!#9'/1!
270 + !#9'1!
rv
289
!#9'/1"!
19 +
!#9'1&#!
rv
129
!#9'/1&#!
129 +
!#9',4!
rn
75
!#9',/4!
75 + +
!#!
40
!#4!
15
!#/64!
14 +
!#9!
rv
205
!#9!
136
!#964$'!
rv
23
!#9/64$'!
23 +
!#9/60&!
126 + !#960&!
rn
219
!#9/60&'!
93 +
!#&'12/-!
31 + !#&'12O-!
rn
62
!#&'/62O-!
31 +
!#&'1-'!
rv
29
!#&'/1-'!
29 +
!#&'14!
rn
153
!#&'1/4!
153 + +
!#&'19'!
rn
47
!#&'19'!
47
!#&'19'14!
rn
24
!#&'19'/14!
24 +
!#&',&!
rv
129
!#&'/,&!
129 +
!#&',C!
rn
63
!#&',/C!
63 +
!#&!
414
!#&%!
1,865
!#&/%"!
243 +
!#&/%"'!
403 +
!#&/%"-'!
167 +
!#&%!
rv1
4,021
!#&/67!
929 +
!#&%!
c
97
!#&%!
97
!#&%$',7!
rn
36
!#&/%$',7!
36 +
!#&%2%4!
rn
22
!#&%2/%4!
22 +
!#&%0!
411 !#&%0!
ra
501
!#&%0'!
90
!#&%0!
rv
125
!#&%0!
125
!#&,$!
ra
30
!#&/,$!
30 +
!#&6/-!
343 + !#&6-!
rn
405
!#&6-'!
62
!#&6-A!
rn
16
!#&6-/A!
16 +
!#&60!
rv
40
!#&/60!
40 +
!#&606/4!
274 +

!#&0%/4!
431 + +
!#&6064!
rn
760
!#&0%4'!
55
!#&60/65!
35 + !#&6065!
rv
103
!#&0/6B!
68 +
!#&0'1-'!
rv
19
!#&0'/1-'!
19 +
383
!#&0'/1+'!
44 + !#&0'1+!
rv
69
!#&0'/1+-'!
25 +
!#&0/%8!
96 + !#&0%C!
rn
124
!#&0/%C!
28 +
!#&0/%7!
32 + !#&067!
rv
769
!#&0/67!
737 +
!#&06/2!
19 + !#&062!
rv
44
!#&06/&8'!
25 +
!#&0;!
rv
22
!#&0/;!
22 +
!#&#14!
rn
51
!#&#/14!
51 +
!#&;$'!
rv
140
!#&/;$'!
140 +
!#&/;2!
31 + !#&;2!
rv
62
!#&/;&8'!
31 +
!#&/;-!
23 + !#&;-!
rn
45
!#&/;-7!
22 +
!#;!
rv
20
!#/;!
20 +
!#;$!
rn
87
!#/;$!
87 +
!#/;5$!
37 +
!#;/$!
53 +
!#;$!
rv
172
!#;$'!
82
!#;+!
rn
24
!#/;+!
24 +
!#;0!
ra
21
!#;/0!
21 +
!#;8&8'!
rv
55
!#;8&8'!
55
!#;&O2!
rn
31
!#/;&O2!
31 +
!#;C!
ra
30
!#;C!
30
!#"'%8&8'!
ra
42
!#"'/%8&8'!
42 +
!#"'102!
rv
15
!#"'/102!
15 +
!#"'1&!
rv
212
!#"'/1&!
212 +
!#"'15!
ra
44
!#"'/15!
44 +
!#"!
rp
1,776
!#"!
1,776
!#"6A/6$!
91 + !#"6A6$!
rn
111
!#"6A/6$'!
20 +
!8!
21
!/,8!
47 + +
!8!
sn
444
!/;8!
376 +
!8%B!
rn
124
!8%B!
124
!8%9!
rn
41
!8/%9!
41 +
!8%0!
rn
42
!8%0!
42
!817!
rn
44
!8/17!
44 +
!81#&!
102 !81#&!
rn
282
!81#&'!
180

!8/,A!
37 + !8,A!
rv
52
!8/,A'!
15 +
!8,0!
ra
160
!8,0!
160
!82%?!
rn
25
!82%?!
25
!82/6-!
702 + !826-!
rn
908
!82/6-'!
206 +
384
!86?'!
rv
79
!86/?'!
79 +
!8696&!
rv
17
!8/696&!
17 +
!86##!
rn
16
!86#/#!
16 +
!8&%A!
rn
54
!8&%A!
54
!8&;2!
rn
25
!8&/;2!
25 +
!8&;0!
rn
18
!8&/;0!
18 +
!8&8'/,8&8'!
16 + !8&8',&!
rn
38
!8&8'/,&!
22 +
!8&8'62!
rn
48
!8&8'6/2!
48 +
!8/;+!
83 + !8;+!
rv
111
!8/;+'!
28 +
!8/;&!
23 + !8;&!
rv
50
!8/;&'!
27 +
!&'!
42,690 !&'!
sa2
48,060
!&'/,!
3,959 + +
!&'E0!
sn
264
!&'E0!
264
!&'/%!
98 + !&'%5!
rv
254
!&'%5!
156
!&'1%&0!
rn
76
!&'1/%&0!
76 +
!&'12!
rn
35
!&'/12!
35 +
!&'1-'!
ra
79
!&'1-'!
79
!&'1-'!
sn
1,588
!&'/1-'!
1,588 +
!&'1-!
rn
37
!&'1-!
37
!&'1+!
45 !&'1+!
ra
121
!&'/6+!
76 +
!&'1+!
rn
66
!&'/1+!
66 +
!&'14'!
rn
29
!&'14'!
29
!&'10'!
rv
75
!&'/10'!
75 +
!&'109'!
ra
18
!&'/109'!
18 +
!&'100'!
rn
23
!&'/100'!
23 +
!&'1#!
rv
15
!&'/1#!
15 +
!&'1&0/%$!
21 + !&'1&0%$!
rn
67
!&'1&0/%$'!
46 +
!&'/1&8'!
17 +
!&'/1&8'!
16 +
!&'1&8'!
rv
57
!&/62!
24 +
!&'1C4'!
rn
209
!&'/1C4'!
209 +
!&',9!
rn
17
!&'/,9!
17 +
!&',8!
57 !&',C!
ra
259
!&'/,C!
202 +
!&'69!
rv
92
!&'69!
92

!&'6&'!
rn
61
!&'/6&'!
61 +
!&!
13,052 !&!
rp1
13,705
!&'!
653
!&!
2,166 !&!
rp2
2,349
!&"!
183
385
!&!
677
!&'!
492
!&8'!
169
!&!
sa1
1,356
!&"!
18
!#!
128
!#'!
119
!&!
1,372
!&'!
221
!&!
sa2
2,325
!&#!
485
!&!
229
!&'!
327
!&!
sn1
653
!&8'!
97
!&!
276 !&!
sn2
494
!&'!
218
!&%7!
rv
34
!&/%7!
34 +
!&%-'!
rn
75
!&/%-'!
75 +
!&%4'E&#!
rn
73
!&/%4'E&#!
73 +
!&%0'1-!
rn
25
!&%0'/1-!
25 +
!&/%8&8'!
52 + !&%#2!
rv
68
!&/%#2!
16 +
!&,-!
rn
45
!&,-!
45
!&,#'%&8'!
rn
406
!&/,#'%&8'!
406 +
!&6!
c
1,240
!&6!
1,240
!&62!
rv
36
!&/62!
36 +
!&6-!
rn
29
!&6-!
29
!&6-9!
rn
36
!&6-/9!
36 + +
!&6-#&!
ra
37
!&/6-#&!
37 +
!&64!
ra
52
!&/64!
52 +
!&64!
rn
42
!&64!
42
!&644!
rn
65
!&/644!
65 +
!&60/5!
84 + !&60B!
rv
99
!&/60B!
15 +
!&60/69'!
32 + !&6069'!
rv
52
!&60/69-'!
20 +
!&60&8'!
ra
17
!&/60&8'!
17 +
!&8&8'!
23 !&6#2!
rv
41
!&6#/2!
18 +
!&6&8'!
rv
149
!&/6&8'!
149 +
!&/6"!
33 + !&6"%0',8&8'!
rn
529
!&6"/%0',8&8'!
496 +
!&0'1A!
rv
125
!&0'/1A!
125 +

!&0'1"/65!
34 + !&0'1"6B!
rn
76
!&0'1"/6B!
42 +
!&0!
36 !&0!
rp
793
!&0'!
757
!&0%$',!
rv
28
!&0%$'/,!
28 +
!&0%2!
rv
108
!&0%2!
108
386
!&0%+"%7!
rn
32
!&0%+"/%7!
32 +
!&0%"!
rn
57
!&0%/"!
57 + +
!&06!
rv
17
!&0/6!
17 +
!&0;A!
rn
107
!&0;/A!
107 + +
!&0;/$!
632 + !&0;$!
ra
752
!&0;/$'!
120 +
!&;+%4!
rn
40
!&;+/%4!
40 +
!&;&8'!
rn
19
!&/;&8'!
19 +
!&"'/105$!
18 +
!&"'/10$'!
23 +
!&"'10$'!
ra
93
!&"'/60$!
52 +
!&"60!
rv
79
!&"/60!
79 +
!&#1!
rp
63
!&#1!
63
!&#1-'!
rn
66
!&#/1-'!
66 +
!&#/1-!
170 + !&#1-!
ra
199
!&#/1-'!
29 +
!&#14!
rn
20
!&#/14!
20 +
!&#14&!
rn
39
!&#/14&!
39 +
!&#14&0!
rn
117
!&#/14&0!
117 +
!&#19'!
rv
31
!&#19'!
31
!&#1#!
rv
39
!&#/1#!
39 +
!&#1C!
rn
176
!&#/1C!
176 +
!&#,?0!
rn
63
!&#/,?0!
63 +
!&#"'1&!
rv
173
!&#"'1&!
173
!&8'%7!
rn
57
!&8'%7!
57
!&8'%#!
rn
959
!&8'%#!
959
!&8'/%8&8'!
38 + !&8'%#&!
ra
140
!&8'/%#&!
102 +
!&8'/%#&!
277 + !&8'%#&!
rv
546
!&8'/%#&'!
269 +
!&8'1-6"'/12!
636 + !&8'1-6"'12!
rn
708
!&8'1-6"'/1&8'!
72 +
!&8'1+6$%4!
rn
43
!&8'1+6$/%4!
43 +
!&8'10'1$!
29 !&8'10'1$!
ra
96
!&8'10'1$'!
67
!&8'10'1*!
c
334
!&8'10'1*!
334
!&8'102!
rn
22
!&8'102!
22
!&8'/604!
128 + !&8'104'!
ra
148
!&8'/604'!
20 +
!&8'10&!
rn
41
!&8'10/&!
41 +
!&8'1&,0!
24

!&8'1&/,0'!
231 +
!&8'1&"'10!
36
!&8'1&"'10!
rp
399
!&8'1&"'/60!
108 +
!&8'1*!
rv
29
!&8'/1*!
29 +
!&8'/%!
763 + !&8',4!
rv1
940
!&8'/,4!
177 +
387
!&8',4!
rv2
41
!&8'/,4!
41 +
!&8',#&!
ra
108
!&8'/,#&!
108 +
!&8'/1#!
169 +
!&8',/#!
35 +
!&8',/#'!
35 +
!&8',&!
427
!&8',&!
rv
812
!&8'/6&!
146 +
!&8'-'14!
rn
122
!&8'-'/14!
122 +
!&8'10&'!
24 !&8'60&!
rn
48
!&8'/60&!
24 +
!&8';/$!
24 + +
!&8';$'1#!
rn
60
!&8';$'/1#!
36 +
!&8';&'!
c
132
!&8';&'!
132
!&8';"!
rv
248
!&8'/;"!
248 +
!&8';5!
ra
60
!&8';5!
60
!;!
in22
33,704
!;!
33,704
!;!
c
5,525
!;!
5,525
!;%*!
sa
29
!;/%*!
29 + +
!;$!
sp
633
!;$!
633
!;BO-'!
rn
21
!/;BO-'!
21 +
!;BO-!
rn
143
!;BO/-!
143 +
!;B0';+!
ra
15
!;B0'/;+!
15 +
!/,2!
43 +
!/,&8'!
121 +
!/;2!
66 +
!/;&8'!
1,179 +
!"/,2!
42 +
!;2!
rv
1,475
!"/,&8'!
24 +
!;-'!
rn
232
!/;-'!
232 +
!;/+!
66 + !;+!
rv
256
!/;+'!
190 +
!;+!
sn
106
!/;+!
106 +
!;4',"'10#'!
ra
64
!;4',"'/10#'!
64 +
!;4!
sn
57
!/;4!
57 + +
!/;9!
156 +
!;9'!
199
!;9!
sn
392
!/;9-'!
37 +
!;0!
sn2
220
!/;0!
220 + +
!;&!
sp
407
!;&!
407
!/;&0!
168 + !;&0!
rn
330
!"&0!
162

!;8!
25 !;C!
rn
51
!/;C!
26 +
!;C!
sn
28
!/;C!
28 + +
!;*!
ra
51
!/;*!
51 +
!;*!
sn
31
!/;*!
31 + +
388
!;5!
236 !;51!
c
1,249
!;5/1!
1,013 +
!;5,4!
rn
20
!/;5,4!
20 +
!"'1*!
30 !"'E#'!
rp
4,251
!"'E#'!
4,221
!%*!
157
!"'/%*!
30 +
!"'%*!
rv
257
!"'/%*'!
70 +
!"'/1$'!
144 + !"'1$'!
rv2
486
!"'/1#!
342 +
!"'1$'!
c
481
!"'1$'!
481
!/18!
63 +
!"'/18!
34 +
!"'/1$!
27 +
!"'/1$'!
18 +
!"'1$!
rv
307
!"'/1#!
165 +
!"'12!
rn
54
!"'12!
54
!"'1-'!
ra
127
!"'1-'!
127
!"'1-'!
rv
27
!"'/1-'!
27 +
!"'10'6"!
rn
26
!"'10'/6"!
26 +
!"'/10!
311 + !"'10!
rv
552
!"'/10'!
241 +
!"'/108!
98 + !"'10C!
rn
252
!"'/10C!
154 +
!"'1#'E/4!
61 +
!"'1#'E4!
c
122
!"'/6#'E4!
61 +
!"'1#'1-!
66 !"'1#'1-!
ra
158
!"'1#/'6-!
92 +
!"'1#'+%!
c
21
!"'1#'+/%!
21 +
!"'/1#!
21 +
!"'/1#'!
16 +
!"'1#!
rv
125
!"'/,#'!
88 +
!"'18&8'!
rn
88
!"'18&8'!
88
!"'1&'E0!
rn
117
!"'1&'E0!
117
!"'1&!
rn
54
!"'1/&!
54 +
!"'/1&!
756 +
!"'/1&'!
254 +
!"'1&!
rv
1,165
!"'/1&8'!
155 +
!"'1&8'10!
293 !"'1&8'10!
rn
354
!"&8'10!
61
!"'1*!
rv
64
!"'/1*!
64 +

!"',!
rv
77
!"'/,!
77 +
!/,$!
35 +
!/,$'!
23 +
!"'/,$!
450 +
!"',$!
rv
1,238
!"'/,$'!
730 +
389
!"'/,4!
35 + !"',4!
rn
56
!"'/,4'!
21 +
!"!
rp
2,290
!"!
2,290
!"O!
c
14,434
!"O!
14,434
!"%B64!
rn
74
!"%B/64!
74 +
!"%5!
rn
202
!"/%5!
202 +
!"1!
rp
36
!"1!
36
!",!
c
1,554
!"/,!
1,554 +
!"/,8!
55 + !",#!
ra
319
!",#!
264
!-%#!
59 !"-%#!
rv
114
!"-%#!
55
!"6$!
rn
281
!"6/$!
281 + +
!"/65!
24 +
!"6$!
682
!"6$!
rv
939
!"/6$'!
233 +
!"67!
rv
382
!"67!
382
!"62*%-!
rn
27
!"62*/%-!
27 +
!"6-'!
rv
265
!"/6-'!
265 +
!"6-';!
rv
41
!"6-'/;!
41 +
!"/6-4!
135 + !"6-4!
rn
166
!"/6-4'!
31 +
!"6-6#!
rn
75
!"6-6#!
75
!"64!
c
59
!"/64!
59 +
!"606&!
rn
73
!"60/6&!
73 +
!0/%8&8'!
50 +
!0/%&!
74 +
!0/%&'!
71 +
!"'10!
221
!"60/6&!
47 +
!"606&!
rv
522
!"0/%8&8'!
59 +
!"60&8'!
rv
14
!"/60&8'!
14 +
!"6&!
c
1,283
!"/6&!
1,283 +
!"6*'!
rv
21
!"/6*'!
21 +
!"6*-'1!
c
76
!"6*-'1!
76
!"0'/1+'!
647 + !"0'1+'64!
rn
701
!"0'1+'64!
54
!"0%B!
ra
49
!"0%/B!
49 +
!"0%&8'!
rn
42
!"0%/&8'!
42 +
!"#&0'/1&'!
140 + !"#&0'1&'!
rv
343
!"#&0'/1&8'!
203 +

!"!
55 !"8!
sa
132
!"8!
77
!&60!
28
!"&60!
335
!"&60!
rp
426
!"&60'!
63
390
!8!
171 + !C!
sn
330
!C!
159
!C%0%2&'10!
rn
64
!C%0/%2&'10!
64 +
!C,+'!
rv
27
!C/,+'!
27 +
!C-'1A!
rn
113
!C-'/1A!
113 +
!C+;0!
ra
19
!C+/;0!
19 +
!8!
131
!8/1!
26 +
!C/%5!
21 +
!C/65!
94 +
!C/6$!
261 +
!C6$!
rv
1,925
!C/6$'!
1,392 +
!C6-6$!
ra
108
!C6-6$!
108
!C60!
rn
45
!C/60!
45 +
!C6068!
ra
750
!C60/68!
750 +
!C/6&!
47 + !C6&!
rv
1,102
!C/6&'!
1,055 +
!C/6*!
592 + !C6*'%7!
rn
885
!C6*'/%7!
293 +
!C/;5!
38 + !C;$!
ra
321
!C/;$!
283 +
!C"/%&!
38 + !C"%&!
rv
169
!C"/%&'!
131 +
!C"6#&!
rn
71
!C"6#/&!
71 +
!*0'!
81 !*'E0!
rv
117
!*'E0!
36
!*'1-'!
ra
91
!*'/1-'!
91 +
!*'1+!
21 !*'1+!
rn
311
!*'1+'E/-'!
290 + +
!*'10"!
rv
34
!*'/10"!
34 +
!*',+!
rn
108
!*',/+!
108 + +
!*'604!
rn
50
!*'60/4!
50 +
!*/,"!
191 + !*O"!
rv
597
!*O"!
406
!*%!
c
5,296
!*%!
5,296
!*%A!
rv
60
!*%/A!
60 +
!*/%$!
139 + !*%$!
rn
174
!*/%$'!
35 +
!*%264!
rn
57
!*%2/64!
57 +
!*%-!
rn
93
!*/%-!
93 +
!*$!
rv
252
!*$!
252

!*$60/6"!
87 +
!*$60/6"'!
25 +
!*$0/%"!
92 +
!*$606"!
ra
221
!*$0/%"-'!
17 +
!*-!
ra
22
!*-!
22
391
!#'/14!
88 +
!#4'!
89 +
!*4'!
sn
831
!*4'!
654
!*4/%!
1,797 + !*4%!
rv
1,830
!*4/%"!
33 +
!*4/%+'!
42 +
!*4%+'/14'!
13 +
!*4%+'64'!
rn
68
!*4/%+'64!
13 +
!*6-6&!
64 !*6-6&!
rn
80
!*6-6&'!
16
!*;A!
rn
60
!*;A!
60
!*;-'&!
rv
81
!*/;-'&!
81 +
!*"'1/4!
11 + !*"'14!
rn
22
!*"'/14'7!
11 +
!*"'10'!
rn
40
!*"'10'!
40
!*"'6*$!
rn
68
!*"'6*/$!
68 +
!*"64!
59 !*"64!
rn
111
!*"/64'!
52 +
!*"/;2!
50 + !*";2!
rn
100
!*"/;&8'!
50 +
!5%!
rv
23
!5/%!
23 +
!51-!
96
!5/%-!
237 +
!5%-!
rv
450
!5/%-'!
117 +
!5%#!
rv
20
!5%#!
20
!51!
c
2,363
!51!
2,363
!51-'1*!
rn
88
!51-'/1*!
88 +
!51#&!
rn
23
!5/1#&!
23 +
!51&8'!
rv
25
!5/1&8'!
25 +
!5/,!
1,111 + !5,!
rv
1,269
!5,"!
158
!5$!
277 !5,$!
rv
412
!5/,$!
135 +
!56-&!
ra
51
!5/6-&!
51 +
!564!
rn
260
!56/4!
260 + +
!5;04!
ra
85
!5/;04!
85 +
392
Appendix 4: The computational phonology as a Prolog Definite Clause Grammar
/* A computational phonology of Russian.
For use in parsing and generating Russian word-forms.
The variety of Prolog used here is Poplog Prolog.*/
:- op(21,xfy,':').
/* Top-level predicate: this finds all possible structures for a
given string of phonemes (the variable Word), matched with their
respective likelihoods (Prob). Prob is computed as the log likelihood
of the syllable structure plus the log likelihood of the
morphological structure. The number of possible parses (Number),
whether the word-form is marked or unmarked for plural stress
retraction (Mark), a morphological structure (Mstructure), syllable
structure (Pstructure), and morphological tokenization (Tokenization)
are also returned. The predicate mpword stipulates that a well-formed
word-form (an mpword) of Russian must have both a morphological and
syllable structure (it must be both an mwd and a pwd), in
accordance with the grammar modules below.*/
mpword(Tokenization,Number,Prob,Mark,Mstructure,Pstructure,Word):-
fast_setof (
(A1,T,M,MS,PS),
( mwd(T1,M,MS,A2,Word,""),
pwd(A3,M,PS,Word,""),
A1 is -(A2 + A3)
),
[(Prob,T,Mark,Mstructure,Pstructure)|NT]
),
!,
length(NT,N),
Number is N + 1.
/* Russian syllabification grammar */
pwd(Z,R,pwd(W,S)) -->
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, [left:L,right:R],weakfoot(W1)) -->
syllable( Z,
[left:L,
right:R], W1),
{unif(R,[stress_det:0,stress:0|_])}.
strongfoot( Z, [left:L,right:R],strongfoot(S)) -->
syllable( Z,
[left:L,
right:R], S),
{unif(R,[stress_det:1,stress:1,pl_retr:0,vfv:0|_])}.
strongfoot( Z, [left:L,right:R],strongfoot(S,W)) -->
syllable( 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, [left:L,right:R],strongfoot(SA,SB)) -->
syllable( 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, [left:L,right:R],strongfoot(SA,SB,W)) -->
syllable( 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( Z,
[left:LM,
right:RM], syllable(O,R)) -->
onset( Z1,
[left:LM,
right:LRM], O),
rime( 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|_]|_], X),
unif( [cons:1,
high:F3,
back:F4,
rime:[high:F2,back:F1|_]|_], 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
su7estvitel'noe. Vorone: 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: Izdatelstvo Moskovskogo Universiteta.
Avanesov, Ruben Ivanovi; and S. I. Oegov. 1959. Russkoe literaturnoe proiznoenie
i udarenie: slovar-spravo7nik. Moscow: Gosudarstvennoe izdatelstvo inostrannyx i
nacionalnyx 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, mile. 1935. Origines de la formation des noms en Indo-Europen
(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. Crochetire, J.-C. Boulanger, and C. Ouellon (eds.). Proceedings of
the Fifteenth International Conference of Linguists. Quebec: Presses de lUniversit
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 Vasilevna. 1969. The syllable structure of speech and distinctive
features of phonemes. Phonetica 20, 1-40.
Bondarko, Lija Vasilevna. 1977. Zvukovoj stroj sovremennogo russkogo jazyka.
Moscow: Prosve;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. Cross-
serial 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, 53-
107.
Bulanin, Lev Lvovi;. 1970. Fonetika sovremennogo russkogo jazyka. Moscow:
Visaja kola.
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, 135-
158.
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), 81-
96.
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 dtudes Slaves.
Garde, Paul. 1980. Grammaire russe: I. Phonologie. Morphologie. Paris: Institut
dtudes 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. Trait de phontique. 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: Izdatelstvo 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, 150-
181.
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: Izdatelstvo Akademii
Nauk SSSR.
Isa;enko, A. V. 1970. East Slavic morphophonemics and the treatment of the jers in
Russian: a revision of Havliks Law. International Journal of Slavic Linguistics and
Poetics 13, 73-124.
It, 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 lvolution phonologique du russe compare
celle des autres langues slaves. Prague: Jednota Keskoslovenskch 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
Geruschlaute im Russischen. In Margarete Woltner and Herbert Bruer (eds.).
Festschrift fr 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 63
th
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 Indo-
European 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 Demsa
(6668-6669). Leningrad: Izdatelstvo 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:
Prosve;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 Bybees 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: Sttn pedagogick
nakladatelstv.
Oliverius, ZdenMk F. 1976. Morfemy russkogo jazyka: 7astotnyj slovar. Prague:
Univerzita Karlova.
Panov, Mixail Viktorovi;. (ed.). 1968. Russkij jazyk i sovetskoe ob7estvo: sociologo-
lingvisti7eskoe issledovanie. Moscow: Nauka.
Panov, Mixail Viktorovi;. 1979. Sovremennyj russkij jazyk: Fonetika. Moscow:
Visaja kola.
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. Luschtzky, 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.
Redkin, Vitalij Andrejevi;. 1971. Akcentologija sovremennogo russkogo
literaturnogo jazyka: posobie dlja u7itelej. Moscow: Prosve;enie.
Reformatskij, Aleksandr Aleksandrovi;. (ed.). 1970. Iz istorii ote7estvennoj fonologii.
O7erk. Xrestomatija. Moscow: Nauka.
Reformatskij, Aleksandr Aleksandrovi;. 1975. Fonologi7eskie tjudy. 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 gnrale. Paris: Librairie Payot.
Scalise, Sergio. 1984. Generative morphology. Dordrecht: Foris.
;erba, Lev Vladimirovi;. 1912. Russkie glasnye v ka7estvennom i koli7estvennom
otnoenii. 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. 337-
383.
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, 183-
204.
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. Grundzge der Phonetik. Leipzig: Breitkopf und Hrtel.
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.
vedova, Natalia Julevna 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 Fliers 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. Slovoobrazovatelnyj 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 hufigsten 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 Anatolevi;. 1967. Russkoe imennoe slovoizmenenie. Moscow:
Nauka.
Zaliznjak, Andrej Anatolevi;. 1977. Grammati7eskij slovar russkogo jazyka:
slovoizmenenie. Moscow: Russkij jazyk.
Zaliznjak, Andrej Anatolevi;. 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.