..... .

.. u. .x

.r,.a4.:
1.....41...

. . v.

_. . . .

..._. . .

..._. . .

..:

. I.
. are; q
.. . Ifr.
J... . . . tn...
a......._v. . .
1 7...

. 1..
: cam:

. . i»... 1......

iii. 1...}: 1....“

4.... 4 . 1.4;. .

.3...
if...

:1... ..
.caaf-zib .5. fr. 71.1.
XII-17,1. lJ: 33!"... 7‘
. .'3
2.2.5.4.}. I :i
. .13... 25... at 2
. . . (.5 '1 z
1 . XII-ullllr.
. 5:; 3.5.1....
3,. 1...:
.13..

v;
..\...
... .55

y (in

. . . 17.....(oi 02.1011

. . . 14.1.5.5; r!- .1111!
v .. o a y '11:) . .

3...) (r 55!...v’nyli

. L. :fiélti)...

’41:...) 3. :7:

1.
, .1. 3..
5.1;: .

,5. .

.
n!f€.ll.(if
.5 I115. .i..‘
13:11.}.
. .5

. . .. . :1. .
.r... .

Q. . .«viov.
: . . . . «y; .o...r..
.1. r r/
51 . .

.. 3.1:?
. 3.1!..r;
. . . .A
.7... . . .
), . .

.5! 1..» a
.. 31.1...
3,. :5}

7.1.1.1....
. ..}4 s... I
.;a:v,'.;..i.

.yylzarlll.l..rv

Z...) . . Zirt, LI... .pc...l .5.
. I. I .1! «ft! . .I.!I.rv

.
2.3.2:,

. . (r.
. 5f. V... 1.2.; (a f.
r, 3.1.... S (3:...

2.... 5L. 7. ; r .
; . l :1. b .I 3.

$1.3; .
. 7. 4..

.7. I...
. . , it)" (4
. iori)?!’
17...} . .

 

 

TH 8.81%

3 1293 00899 0255

This is to certify that the

thesis entitled

A CONSTRAINT-BASED APPROACH TO
RUSSIAN VOICING ASSIMILATION

presented by

Takayo Sugimoto

has been accepted towards fulfillment
of the requirements for

 

 

 

M.A. degree in Linguistics
. 7 \
4?? Q; 2..

Major rofessor . .
Professor en-Hwe1 L1n

[hue 12th October 1992

 

0-7639 MS U is an Affirmative Action/Equal Opportunity Institution

 

  
   
   

  

LIBRARY
Michigan State
‘ Halves-city

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.

DATE DUE DATE DUE DATE DUE

MAY p22 12005

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MSU Is An Affirmative Action/Equal Opportunity Institution
I c:\circ\datedue,pm3-p.1

A CONSTRAINT-BASED APPROACH TO
RUSSIAN VOICING ASSIMILATION

By

Takayo Sugimo to

A THESIS
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
MASTER OF ARTS

Department of Linguistics, Germanic,
Slavic, Asian, and African languages

1992

 

ABSTRACT

A CONSTRAINT BASED APPROACH TO
RUSSIAN VOICING ASSIMILATION

BY

Takayo Sugimoto

This thesis investigates Russian voicing assimilation within the
framework of non-linear phonology. Two topics dealt with are the nature
of assimilation rules and asymmetry effects.

In autosegmental phonology, a ’[Spread aF]’ analysis has been
considered a sole mechanism of assimilatory processes. This single
spreading analysis, which requires a binary feature system when both
values of a feature assimilate, is problematic if a feature is indeed privative.
Assuming Lombardi’s (1991) privative analysis of [voice], I propose a
constraint-med approach to assimilation, which allows a feature to be
expressed in any type of feature systems. Russian voicing assimilation is
expressed by two types of processes: spreading or delinking of [voice].
Persistent constraint-motivated rules of assimilation eliminate extrinsic
ruleordering.

Explanation of the asymmetrical segment v in Russian requires
Radial Underspecification. The segment must be underlyingly unspecified
for the contrastive feature [sonorant], which cannot be so under Contarstive

Underspecification.

 

ACKNOWLEDGMENTS

I was lucky to have had as my thesis committee three phonologists
from three different theoretical frameworks. I am indebted to my advisor,
Y-H Lin for her patient guidance and insightful comments, which led to
considerable improvements of this work. To other members of my thesis
cormnittee, Grover Hudson and David Lockwood, I also owe a major debt
for their helpful criticisms and suggestions on this thesis.

I also thank faculty and graduate students of Linguistics at MSU. In
particular, Mutsuko Endo Hudson, Kaz Fukushima, Paul Kershaw, Chege
Ithiora, Mieko K. Philps, and Yoshi Yamakawa for their suggestions and
help in many respects I am very grateful to Svetoslav Pavlov and Adam
Malacynzky for their help in Russian.

I am very grateful to friends from Santa Cruz Hisa Kitahara, Tohru
Noguchi, J unko Shimoyama and Akira Watanabe for their encouragement
and help in getting necessary material for this research Special thanks also
go to Jean-Remi B, Raul Tomas and Sanjay Yedur. Astrid Baviere and
Tatsuko Enami have been very supportive. I express my deepest thanks for
their understanding and friendship over years.

Finally, with love and gratitude, I would like to dedicate this thesis to
my family and Ban, without whom I could have never enjoyed

underspecification of my life.

iii

 

Chapter]:

Chapter 2:
2.1:

2.1:
2.1.2:

Table of Contents

Introduction

Theoretical Framework

The Theory of Autosegmental Phonology
Autosegmental Representations

Rules and Constraints

on autosegemental representations

2.1.3:

Feature-changing and

non-feature-changing assimilation

2.2:

2.2.1:
2.2.2:

Theories of Underspecification
Radical Underspecifieation Theory
Contrastive Underspecification Theory

2.2.3: Redundancy Rules under RUT
2.2.4: Redundancy Rules under CUT

2.2.5:

2.3:
2.3.1:
2.3.2:
2.3.3:
2.3.4:
2.3.5:

Chapter 3:
3.1:
3.2:

Summary: RUT and CUT compared
Theories of feature systems
Trubetzkoy’s feature system
Binary feature hypothesis
Privative feature theory
Privative vs. Binary features
Privative features and

Assimilation processes
Russian Voicing Assimilation
The Russian sound pattern and transcriptions
Language Data

«>0er

I4

33
34
34
36

Chapter 4: Historical Perspectives
4.1: Russian labio dental fricatives
4.1.1: Interpretation of labio-dental fricative v
4.1.2: Abstract /w/ and Underspecification
4.2: Approaches to two processes

4.3: Summary
Chapter 5: A Constraint-Based Approach
to Russian Voicing Assimilation

5.1: Underspecification
5.1.1: Obsu'uent cluster and
Obstruent-Head Licensing
5.1.2: Obstruent-Cluster Licensing
5.2: Privative features and Underspecification
5.21: Privative Voicing and
Radical Underspecification Theory
5.2.2: Radical Underspecification and
Hayes’s prosodic rule
5.2.3: Privative features and Principles of
Contrastive Underspecification

Chapter 6: Conclusion

List of References

41
41
41
46
47
55

56
S7

59

61

65

66

73

79

86

88

1. Introduction

Recent nonlinear phonological research has revealed the importance of
focusing primarily on representations rather than rules. Phonologists’
attention has shifted from the conditions on rules to the conditions on the
representations based on the assumption that we cannot have the right rules
without defining the right representations (McCarthy 1988: 84). The correct
representations may entail the correct rules, but not conversely.

As the importance of phonological representations in constructing
phonological rules and grammar has been realized, approaches appealing to
conditions or constraints on representations have begun to take a central role in
current nonlinear phonological theories.

It is also recognized in nonlinear phonology that a set of modular theories
is more restrictive and thus preferred to a homogeneous one. A formal
theory does not itself have to explain every phonological phenomenon
(Kiparsky 1985: 85, McCarthy 1988: 84). One subtheory explains certain
aspects of phonological phenomena such as asymmetry of a phonological
segment and another subtheory accounts for some other aspects like vowel
harmony.

In this thesis, I will reanalyze widely assumed non-linear approaches to

phonological phenomena, examining Russian voicing assimilation.

One issue considered here is the nature of assimilation rules. In
autosegmental phonology, which uses multi-tiered representations of segment-
internal structure, assimilation phenomena, just like harmony processes, have
been analyzed in terms of the spreading of features, usually being preceded by
the delinking of a formerly or inherently associated feature value. This is
pmticularly convenient for a type of assimilation which involves both values of
a feature since we can express the assimilation of both values by a single
spreading rule. This has been a strong claim for autosegmental spreading in a
binary feature system.

As a component of phonological representations, the n-arity of a feature
has become an important issue. Recently, Trubetzkoy's notion of a mixed
feature system has captured phonologists’ attention (Goldsmith 1985, Steriade
1987, Mester & Ito 1989, Lombardi 1991). T‘rubetzkoy (1969 (1939))
considered three types of feature oppositions: privative, binary and multi-
valued oppositions based on the behavior of individual features One extreme
position in non-linear phonology is that all features are privativei, single-
valued. While it is not clear whether phonological features are in fact all
privative, several desirable consequences of the use of privative featm'es (not
necessarily all features) have been brought into discussion. Goldsmith (1985)
uses the combination of binary and privative featm'es to account for vowel
harmony processes of several languages, based on the assumption that the
choice between privative and binary feature systems is determined by each

 

1In this thuis,the term ‘ptivuiw' it our! inurehugobly with tingle-valued orumry future: as it
it and in ncont works (Goldsmith 1985. Mutor ls Ito 1989, Lombardi 1991).

language. Others argue that some features are universally privative. Steriade
(1987) claims that some features like [round] are single-valued and that treating
such features as single-valued enables us to eliminate a number of redundancy
rules and issues of rule-ordering. Mester & Ito (1989) suggest the adoption of
privative voicing and point out that a privative feature system can restrict a
theory more than a binary system can. Lombardi (1991) provides evidence for
privative laryngeal featm'es.

There are some difficulties for privative features, one of which is
assimilations of both values. These must be analyzed by using binary features
since there is no way to express the spreading of the minus value of a feature
which is simply expressed by the absence of the feature (den Dikken & van der
Hulst 1988: 45).

However, I argue that it is not a privative feature system itself that is
problematic in accounting for assimilations of both values. What is
problematic is a rule which always requires any feature to be binary regardless
of the tram of each feature when both values of a feature assimilate when a
feature is indeed proved to be privative. This does not follow one of the basic
assumptions of non-linear phonology that right representations entail ' right
rules. Therefore, we have to know first what the correct representations or
more sepcifically feature systems of features are. Lombardi (1991: 9-11),
using the privative [voice], analyzes Polish voicing assimilation, which involves
both values, by two independently motivated mechanisms: delinking and
spreading. In her analysis, assimilaiton of both values is accounted for by two
separate processes: voicelessness assimilation is expressed by the delinking of

[voice]; voicedness assimilation by the spreading of [voice].

Another issue considered in this paper involves the treatment of
asymmetry of a phonological segment. It is not uncommon that a phonological
segment behaves differently from other segments of a group in a single
language system. With help from the autosegmental framework, theories of
underspecification have provided an account for this phenomenon. According
to one theory of underspecification, Radical Underspecification, a particular
segment shows asymmetrical behavior not because phonological rules treat
such an asymmetry of a segment as special. It is rather because of the
difference in segment-internal structure. More specifically, phonological
segments are expressed in terms of an organized set of features on multi-tiered
representations. Phonological rules also refer to the segment-internal
structure: they apply to a particular tier or feature value in a hierarchically
organized segment-internal structure Some segments simply lack or are
unspecified for certain features in their phonological representations. When a
phonological rule refers to a certain feature, segments which lack the feature
fail to undergo the rule.

Two types of underspecification available within non-linear phonology
are Radical Underspecification. Theory (Kiparsky 1982, Archangeli 1984,
Archangeli & Pulleyblank 1986; henceforth RUT) and Contrastive
Underspecification Theory (Steriade 1987, Mester & Ito 1989; henceforth
CUT). Both theories agree that certain predictable feature values are
unspecified underlyingly, while they differ in the degree of underspecification.
RUT requires underspecification of all predictable features, both redundant
and unmarked feature values. It also assumes that at most one value is present

underlyingly. CUT, on the other hand, distinguishes redundant feature values

from contrastive or distinctive feature values and is limited to
underspecification of the former type and requires distinctive features to be
underlyingly specified for both values

RUT’s strong claim is that it explains asynunetry effects better than CUT
(Archangeli 1988). Under RUT, which insists on maximal underspecification,
a phonological segment can be treated as featurelessZ. CUT, on the other hand,
does not allow completely featureless segments and has difficulty in accounting
for asymmetry effects under a binary feature system.

Comparing the phonological derivations under the two theories, RUT has
fewer specifications but needs more redundancy rules than CUT needs to
supply all unspecified feature values. According to proponents of CUT, RUT
involves complexity in derivations (Christdas 1988). I will show that this

problem can be resolved by adopting the privative features.

Russian voicing assimilation is, in these respects, particularly interesting.
It involves the assimilation of both values of the voicing feature and has an
asymmetrical segment. Briefly, voicing assimilation in Russian takes place
within an obstruent cluster. All members of an obstruent cluster must agree
with the final member of the cluster in voicing. It also takes place across
sonorant consonants, which are transparent to the process. Among obstruents,

the voiced labio-dental fricative v is asymmetrical: it shows dual-

 

3 By s featureless segment, I mean a segment without autosegmentslizsble phonological features;
thus, while a segment may be without this type of features, the presence of the segment is
expressed by the presence of an entity on the CV tier, which is as a core tier not underspeciflshle.

characteristics. It fails to trigger voicing assimilation like other sonorant
consonants while it undergoes the assimilation like other obstruents.

In traditional analyses, it is assumed that the Russian voiced labio-dental
fricative v is derived from a sonorant w (Lightner 1965, Coats & Harshenin
1971, Hayes 1984, Kiparsky 1985: 103-114). In order to account for this
asymmetric segment within the context of voicing assimilation, any segmental
iterative approach had to construct extrinsically ordered rules, one of which
changes the feature of the segment: [+sonorant] -> [sonorant] (w->v). Voicing
assimilation was considered to apply to all consonants in an obstruent cluster
iteratively; a devoiced intervening sonorant consonant becomes revoiced by a

separate rule (Hayes 1984).

In this paper I will provide a less costly solution to Russian voicing
assimilation, integrating subtheories of non-linear phonology: autosegmental
and underspecification theories and the privative feature system. In particular,
assuming Lombardi’s (1991) privative analysis of voicing, I will provide an
account for the phenomena of voicing assimilation by using a structural
constraint on obstruent-cluster voicing, privative [voice] and Radical
Underspecification. A constraint on structmal conditions of obstruent clusters
allows the use of privative [voice] in Russian. Since a structural constraint
allows any type of feature systems, we only have to investigate the n-arity of a
feature depending on each f eature’s behavior with respect to phonological
processes. I will also show that when privative voicing is adopted, not only can
we eliminate a number of redundancy rules, but also the assimilation processes

become non-feature-changing operations. Since cross-linguistically feature-

changing assimilation is considered very rare (Vago 1988), Russian voicing
assimilation, which is not a rare type of assimilation, should be expressed in
non-feature-changing rules. This constraint on obstruent-cluster voicing
motivates two language-specific processes, namely delinking and spreading,
referring to linking relations of the voicing feature to both the trigger and
targets. In other words, the constraint-motivated rules apply, depending upon
whether or not the trigger has [voice]: spreading of [voice] in the case of
voicedness assimilation; delinking of [voice] in the case of voicelessness
assimilation. These rules have persistent application: they apply whenever the
structural descriptions are met. As an effect of privative voicing, the devoicing
of word-final obstruent also becomes a single process: delinking of [voice].

The use of privative features brings us some desirable effects on
underspecification as well. A privative feature system eliminates some
redundancy rules (complement rules in RUT) and their ordering. This is a
desirable consequence especially for RUT, whose derivation has been claimed
to be more complex than that of CUT (Christdas 1988: 85). The elimination of
exstrinsic rule ordering contributes to the simplification of the grammar.

Russian data provide a support for Radical Underspecification Theory.
RUT provides a natural account without requiring any extrinsic rule-ordering,
obeying a constraint on the interaction of redundancy rules with phonological
rules (the ROC, see 2.2.3). Radical Underspecification of w replaces the
feature-changing rule of w->v by a redundancy rule. CUT is at pain in
accounting for the asymmetrical segment v, which must be specified for

[sonorant] under the binary system.

This thesis is divided into six chapters. Chapter two introduces the
theories which focus primarily on phonological representations and constraints
on representations, namely Autosegrnental Phonology and two theories of
Underspecification. Two kinds of feature systems, privative and binary
features, will also be introduced. Chapter tluee examines phenomena involved
in Russian voicing assimilation. There, we will see the assimilation of both
values of the voicing feature and the dual-characteristics of the voiced labio-
dental fricative v. Chapter four presents some historical perspectives on
voicing assimilation and the special behavior of the labia-dental fricative v. In
particular, we attend to Hayes’s (1984) ordered set of rules and Halle &
Vergnaud’s (1981) foot incorporation. Chapter five provides a constraint-
based approach to the problems, using Radical Underspecification and
privative voicing. I will then discuss the preferability of my solution by
pointing out problems associated with a spreading analysis and the principles of
Contrastive Underspecification.

 

 

2. Theoretical Framework

This section presents an overview of the theories which are assumed in
the later chapters: Autosegmental Phonology and two theories of
Underspecification, Radical Underspecification and Contrastive
Underspecification. Then, we will see two types of feature systems, privative
and binary features and their effects on autosegmental representations and
underspecification approaches.

2.1. The Theory of Autosegmental Phonology

2.1 1 Autosegmental Representations

The autosegmental framework characterizes phonological properties or
features by separate autonomous tiers: the CV (skeletal) tier and melodic tiers
(Clements & Keyser 1983). The CV tier is the core tier in autosegmental
representations, playing the role of anchor for other elements on other tiers.
The CV tier consists of C and V slots (positions), to which consonantal and
vocalic elements, respectively, are to be linked. For instance, given a Swahili
word mbu ’a fly’, we can assign each segment to the CV -tier as follows:

(1)

C V-tier C C V

In (1) above, the segments [m,b,u] themselves are not autosegments but
just phonological segments. Phonological segments are decomposed into or

9

 

10

expressed by a set of features, which are represented on melodic tiers. Each
melodic tier consists of autosegments, a feature-valued set such as [+nasal] and
[nasal] (in binary terms) for the nasal tier. Autosegmental tiers are associated
with or linked to one another by association lines. Association lines indicate
temporal associations or linkages between an autosegment P and a P-bearing
unit.

We can associate autosegments to the CV tier. Since each phonological
segment consists of a set of features, we can decompose each phonological
segment with respect to phonological features: m consists of [+son,+nasal,
+labial...], b-[-son,-continuant,+labia1,-nasa1,+voiced...], and

u-[+sonorant,+high, ~1abial,..].

There is a good reason to suspect that phonological segments are not
simply an unorganized set of features. That is, a certain related class of
features such as [round] and [labial] function togetheras a group in undergoing
phonological processes (McCarthy 1988, Lombardi 1991). This fact has led to
the innovation of feature geometry (Clements 1985). In feature-geometric
approaches, features are hierarchically organized to represent a segment-

internal structure.

In this work, I assume the slightly modified version of the feature
geometry proposed by McCarthy (1988), in which the feature [sonorant]
occupies the root node and the voicing feature is not directly but indirectly
linked to the root node through the laryngeal node. The CV -tier is added to his

geometry for the reason that Russian voicing assimilation requires a consonant-

ll

vowel distinction to identify obstruent clusters, to which the process applies.

The geometry is given below:

(2) (modified version of McCarthy 1988:105)
C

[sonorant] Root node

  
  
 

Laryngeal node

 

[ continuant] [ nasal]
[voice]

pl node

[labial] [coronal] [dorsal] [pharyngeal]

[round] [distributed] [anterior] [lateral]

It is important to note that, in McCarthy’s model, the major class feature
[sonorant] is notfunctioning as anautosegment: it can neither be spread nor be
delinked by rules (1988: 97). This becomes crucial later when we come to

formulate rules for voicing assimilation in Russian.

So each phonological segment of the word mbu ’a fly” can be partially
autosegmentalized with respect to the features [nasal] and [labial] and indirectly
linked to the CV (skeletal) tier through nodes by association lines.

12

(3)
m b u phonological segments
C V-tier C C V core tier
l I l associationline
Root node [+ on] [-son] [-son]
associationlines
nasal tier [+nasal] [nasal] [-nasal] autosegmental tier
place node "
[ associationlines
labial tier [+labial] [+Iabial] [-la ial] autosegmental tier

For the purpose of this thesis, we will be concerned only with the relevant
features of the Russian voicing assimilation process such as [sonorant], [voice],
and the CV tier(skeleton). Since there is no justification for the direct
dependence between sonorant and voicing features, they are separately linked
to the CV tier (through the laryngeal node, which is understood). For

simplicity of representation, they are expressed as below:

(4) sonorant tier [-son] [+son]
l |
CV tier C V
l l

voicing tier [+voice] [+ voice]

We see that association lines indicate temporal association or linkage of
an autosegment P to a P-bearing unit such as C in the CV tier. All association

lines are subject to a constraint prohibiting line-crossing as in (5).

(5) Line-crossing prohibition (Goldsmith 1976):
Association lines between the same two tiers may not cross.

13

This constraint motivates Tier-Conflation, which is introduced later in this
section. In addition, a P-bearing unit cannot be associated to two opposing
feature values such as [+voice] and [-voi_ce] at the same time.

In this paper, we will deal with two basic mechanisms of autosegmental
phonology (Clements 1985; Hayes 1986; McCarthy 1988):

(6) a. delinking of association lines
b. spreading a feature or an autosegment P, which is already
associated to a P-bearing unit, to another P-bearing unit

Delinking is a deletion of association lines and spreading is an insertion of
them. These mechanisms can be used for a feature-changing process like (7)
where the voicing feature of C1 spreads onto C2 and the voicing feature of C2
gets delinked.

(7) C1 C2
L-w’"?

[-voice] [+voice]

Autosegmental representations place phonological features on different
tiers so that assimilation processes such as vowel harmony and voicing
assimilation can be expressed autosegmentally with rules referring to
associations of only relevant features. Given that X is a target and Y a trigger
of voicing assimilation as in (8), we can express assimilation by inserting the
assimilation feature of Y [-voice] into the target X, delinking the association
line between X and [+voice]. The example in (8) is the case of leftward spread,

which is observed in Russian voicing assimilation:

I4

(8) X Y X Y
£‘~~-l ->
[+ voice] [- voice] [- voice]

Strictly speaking, two types of processes are available in autosegmental
phonology to explain assimilation phenomena (Mester 1986: 74): fusional and
spreading harmony. In fusional harmony, there must be a feature shared by
triggers and targets. Fusion cannot apply to a sequence of different values of a
feature. In spreading harmony, the voicing feature of the trigger spreads onto
all targets, delinking any feature stranded by the process. Fusional and
spreading processes are illustrated below as in (9) a and b, respectively:

(9)8. b.
C C -> C C C C -> C
I I \/ rur
[or lab] [or lab] [or lab] [or lab] [-or lab] {-0 lab]

I assume that voicing assimilation is explained by f usional harmony when
the targets share the same voicing feature as the trigger; otherwise, it is
explained by spreading harmony when the targets and the trigger do not agree

in voicing.

2.1.2 Rules and constraints on autosegmental representations
Apart from autosegmental representations, the nature of rule applications
in autosegmental phonology can be characterized as persistent or non-
persistent: rules and constraints can be either persistent or nonpersistent (Halle
& Vernaud 1987, Myers 1991). i

15

N onpersistent rules (language-specific ordered rules) and universal
constraints, on the one hand, block any application of rules which create a
violation. Thus, we can predict that there is no occurrence of representations
which are to be ruled out by the latter type of rules.

Persistent rules, on the other hand, apply whenever structural conditions
are met, changing the ill-formed representations to well-formed ones.
Syllabification is claimed to be persistent; so is Foot Formation (I-Ialle &
Vergnaud 1987). Rules of this type never block occurrence of ill-formed
representations or block application of a rule which creates a violation.

Persistent rules have reasons to exist. First, since they fix ill-formed
representations by changing them to well-formed ones, underlying
representations of a language do not always have to be well-formed (Myers
1991: 341). This aspect explains the fact that when a foreign word is
introduced into a language, the language changes the sound pattern of the word
to the permissible sound pattern of the language. Second, persistent application
of rules simplifies the grammar: persistent rules reduce burdens associated
with rule-ordering. This concludes the illustration of non-persistent rules and

universal constraints.

Tier Conflation:

Finally, the principle of Tier Conflation (McCarthy 1986, McCarthy &
Prince 1986; henceforth TC) has to be defined. According to McCarthy
(1986), it is a process which reshapes multi-tiered representations into single

linearized representations. When TC linearizes representations, some multiply

l6

linked associations may have to be broken in order to satisfy the line-crossing
constraint (McCarthy 1986: 7).

For example, suppose that a language has a representation like (10a), in
which a link of the voicing feature to the obstruents is created by spreading
across any sonorant consonants or vowels. When a rule supplies the
intervening sonorant the redundant feature [+voice], the multiple-linking of the
voicing feature of the obstruents must be broken since association lines carmot
cross (10b). We thus have to break the multiple-association of the voicing
feature to the obstruents so that the redundant feature [+voice] is inserted to be
linked to the sonorant (10c):

(10) a. [son] [+son] [-son] b. ’ [-son] [+son] [-son]
I I I I I I
C C C C C C
N
[-voice] [+voice] [-voice]

c. [~son] [+son] [son]

I | I
C C C

[-voice] [+voice] [-voice]

2.1.3 F eature-changing and non-feature changing assimilation

It has been assumed in the autosegmental framework that spreading is the
sole mechanism of assimilation processes (Halle & Vernaud 1980; McCarthy
1984, 1986; Hayes 1986; Goldsmith 1990). That is, any type of assimilation,

l7

regardless of the value of a feature, can be expressed simply by the spreading
of the assimilation feature value: spread [01F]. In this type of approach, which
presupposes the use of a binary feature system, voicing assimilation in both
values can be expressed by a single spreading rule.

Assimilation processes can be motivated either by non-featm'e-changing
or feature-changing rules (McCarthy 1984). A non-feature-changing rule fills
in a feature to an unspecified p-bearing unit by means of spreading. Thus, any
non-feature-changing rule is explained by a single spreading process without a
prior delinking operation. A feature-changing rule delinks the association of a
feature value to a p-bearing unit, and then spreads another feature value to the
dissociated p-bearing unit. Any feature-changing operation has two processes,
the delinking of a feature value and the spreading (or insertion) of another.

I will argue that the use of the privative feature [voice], which amounts to
inherent underspecification", enables us to reduce the feature-changing

processes of Russian voicing assimilation to non-feature-changing ones.

22 Theories of Underspecification

The notion of underspecification is not new to non-linear phonology. In
fact, Halle (1959) considered leaving some redundant features unspecified.
This attempt was objected to by Lightner (1963) and Stanley (1967) for the

 

3 The term inherent underspecification is due to Archangeli (1988), where she talks about several
aspects of the underspcificstion effects of single-valued features. See also Steriade (1987) end
Christdes (1988) for their role in underspecification.

l8

reason that underspecification in a binary system creates the third value, i.e., a
zero value besides plus and minus values. Chomsky & Halle’s (1968) solution
to this problem was the full specification of features prior to application of any
phonological rule; underspecification was permitted only in the lexicon, in
which unmarked features such as voicing for sonorants are unspecified
(Chomsky & Halle 1968, Archangeli 1984: 74).

Since SPE it has been assumed that only idiosyncratic information is
stored in the lexicon and all redundant information is left unspecified. Such
unspecified inf orrnation is supplied by rules. Recent developments in
nonlinear generative phonology have led us to be concerned with the role of the
lexicon more than before. Underspecification approaches provide a way to
investigate what kind of information is in fact present or absent from the
lexicon, more specifically, underlying representations. Underspecification
theories also provide us with a way to supply such unspecified features by rules
(Archangeli 1988: 183).

There are two types of underspecification theories which share the
assumptions mentioned above: Radical Underspecification Theory (RUT) and
Contrastive Underspecification Theory (CUT). These two types share the
position that all redundant feature values must be unspecified underlyingly and
are derived by redundancy rules. But they differ in underlying specification of
distinctive feature values: both values of distinctive features in a language are
underlyingly specified in CUT; at most one value of each such feature is
specified in RUT.

19

2.2.1 Radical Underspecification Theory (Kiparsky 1982; Pulleyblank 1983;
Archangeli 1984, 1988; Abaglo & Archangeli I989; henceforth CUT)

Kiparsky (1982) was the first to propose to underspecify all predictable
features in lexical entities to prevent the occurrence of ternary contrast in
underspecification (see 2.2). This idea was further developed by Pulleyblank
(1983) with respect to tonal underspecification and by Archangeli (1984) for
the underspecification of phonological features in general.

The theory of Radical Underspecification claims that a grammar is highly
preferred when it differentiates all phonemes of a language by the minimum
number of distinctive features (Archangeli 1984). It approaches this goal by
simplifying underlying phonological representations and grammar, that is,
eliminating of all predictable features from underlying specifications.

Therefore, this theory may be characterized as maximum minimization
of underlying representations: in addition to the underspecification of all
redundant feature values, the theory requires that for a given feature only
either one of the two values + or - can be specified underlyingly. Accordingly,
there is no feature which has both values present underlyingly; there is no
feature which is assigned to all phonemes of a language. All unspecified
feature values are supplied by redundancy rules (discussed in section 2.2.4).

RUT also claims that underspecification must be determined at least in
part on a language specific basis (Archangeli 1984, Abaglo & Archangeli 1989;
Archangeli & Pulleyblank 1989). While RUT takes the notion of markedness
into account, markedness does not always hold cross-linguistically.

Underlyingly specified feature values are unpredictable features in a language

20

and do not have to be universally marked ones (Archangeli & Pulleyblank
1989: 197). For a given feature the marked (or unmarked) feature value may
be determined on a language-specific basis. Since if the marked feature value is
specified, the unmarked counterpart is predictable, the unmarked value of
distinctive features must also be unspecified underlyingly.

Radical Underspecification Theory gives the following underlying

specification of distinctive features for [m,p,b,f,w,y] in Russian:

(11) Distinctive Features for consonants /m,p,b,f ,w,y/

a. Full specification b. Radical underspecification
m p b f w y m p b f w y

sonorant + - - + + sonorant - - -

labial ++ ++ +- labial ++ ++ +

continuant - - + + + continuant + + +

voiced + + + + + + voiced +

nasal + - - - - - nasal +

(The blanks indicate the unspecified feature values.)

Under RUT, only the marked value [+voice] for obstruents is specified
underlyingly. Since voicing for sonorants is predictable, it is unspecified and

is later supplied by a redundancy rule.

RUT account of asymmetrical segments:

Asymmetrical segments are those which behave asymmetrically to any
other segment in a group of sounds. For instance, in Russian the voiced labio-

dental fricative v shows such a characteristics in voicing assimilation. It

21

functions as an obstruent in undergoing voicing assimilation while acting like a
sonorant in failing to trigger the process. Radical Underspecification Theory
is claimed to be successful in accounting for asymmetry effects by fully
underspecifying a segment (Archangeli 1984, Pulleyblank 1988). Since under
this theory, at most one value, either plus or minus, may be underlyingly
specified, it is possible for a segment to be without features in the underlying
representation though an entity C or V of the CV tier must present for a
segment to exist.

One of principles of RUT that underlying representations are determined
at least in part language-specifically predicts the fact that asymmetrical

segments may differ among languages.

2.2.2 Contrastive Underspecification Theory (Steriade 1987, Christdas 1988,
Mester & Ito 1989; henceforth CUT)

This theory distinguishes redundant feature values from distinctive ones
(Steriade 1987,Christdas 1988). It restricts underspecification to all and only
no nco ntrastive (i.e., no n-distinctive) redundant features and requires the
specification of distinctive features for both values (Steriade 1987, Christdas
1988, Mester & Ito 1989). For example, in English the feature [voice] is
redundant for all sonorants, and so it is unspecified. On the other hand, voicing
is distinctive for obstruents: N and /d/ contrast only in voicing, so the feature
[voice] is specified for both values underlyingly.

While RUT treats features as phonological primitives, CUT takes

combinations of features as primitives. CUT assumes that any theory of

22

feature specification must express natural classes of segments at all levels
(Christdas 1988:92). Thus, it insists on full specification of features which are
necessary to group segments into natural classes. It regards distinctiveness as
idiosyncratic information of a language and takes it into consideration in

determining the underlying specification of phonological segments.

As for the distinction between redundant and distinctive feature values,
Christdas (1988) classifies phonological features into two types: primary and
secondary features. Primary features like [sonorant] and [consonantal] are
those which are necessary to express natural classes of phonological segments
in all languages. They are distinctive and are fully specified underlyingly.
Secondary features are those which may not have to be present in underlying
representations. Examples are laryngeal features like [voice] and [aspiration],
and primary vowel features like [high], [back], [low]. Unlike primary-features
which are always distinctive, secondary features may function redundantly as
well as distinctively depending on each language. When secondary features
function in distinctive capacity in a language, they must be specified for both
values underlyingly. For instance, if voicing is distinctive for sonorants in a
given language, the feature must be specified for both values underlyingly.
According to Christdas, secondary features (but not primary features) are

candidates for one-valued or unary features.

Like RUT, CUT underspecifies all redundant features; unlike RUT, it
specifies contrastive features for both values underlyingly.

 

 

 

23

(12) Distinctive Features for consonants /m,p,b,f ,w,y/

Contrastive underspecification cf. Radical underspecification (13b)
mpbfwy mpbfwy

sonorant - - - + + sonorant -

labial + + + + + - labia] + + + + +

continuant - - + + + continuant + + +

voiced - + - voiced +

nasal + nasal +

Only noncontrastive and redundant features of the language are unspecified
under CUT. For example, m, like any other nasal consonant, is unspecified for
the feature sonorant since it is redundant. With respect to voicing, this feature
is specified for all obstruents for both values since it is contrastive for

obstruents, while all sonorants are unspecified for voicing.

For both theories, nasals are unspecified for the feature sonorant since it
is predictable in the language. Both theories can have a redundancy rule such
as [ ] -> [+son] / [_ ,+ nasal]. As for any other sonorants such as w and y,
they are specified for [+son] under CUT since the theory requires both values
of contrastive features to be specified underlyingly. For the same reason,
obstruents are specified underlyingly as [-son]. On the other hand, with respect
to the voicing feature, the two theories make different underlying
specifications: [-voice] is unspecified and [+voice] is specified for obstruents
under RUT, while both values are specified under CUT since voicing for
obstruents is contrastive. It differs from RUT in that RUT unspecifies [-voice]
for obstruents while CUT requires [+voice] as well as [~voice] to be

underlyingly specified for obstruents since the voicing feature is contrastive

24

for obstruents. In addition, the feature sonorant must also be specified for both

values under CUT

2.2.3 Redundancy Rules under RUT (Archangeli 1984)

Redundancy Rules are rules which assign segments the unspecified
values. Hence, they are feature-filling rules as opposed to feature-changing
rules. In the underspecification frameworks, both RUT and CUT, redundancy
rules were assumed to apply as late as possible unless specific constraints
motivate them to apply earlier in the derivation. Under RUT, redundancy
rules are of two types: complement and default rules. This classification is not
really crucial here, but for the purpose of clarity I will briefly mention the
difference. Default rules are universal rules while complement rules are rules
which require knowledge of the underlying representations of a given
language. To put it simply, complement rules assign the value which is
opposite to the underlyingly specified value. Suppose, in a given language,
[-son] is underlyingly specified for consonants and the voicing feature for
sonorants is unspecified. Then we have redundancy rules as follows:

(13) a I l ->[+sonl

b. [ ] -> [+voice] / [+son]

Rule (13b), which assigns [+voice] to segments specified as [+son], is a
default rule since it applies to any segment specified as [+son] and needs no
language-specific information. Rule (13a), which assigns [+son] to a consonant
is a complement rule since the rule applies after the underlyingly specified
value [-son] is identified.

25

We have seen that Kiparsky (1982) proposes what we now call Radical
Underspecification to prevent such an undesirable consequence as ternary
oppositions. Even in an approach using RUT with binary features, however,
we cannot avoid creation of ternary oppositions (Pulleyblank 1983). This
problem was solved by Archangeli (1984), who introduces a constraint on the
ordering of redundancy rules and phonological rules:

(14) Redundancy-Rule Ordering Constraint (Archangeli 1984: 52; henceforce
RROC)
A redundancy rule assigning or to F, where or is ”+” or ”-”, is

automatically ordered prior to the fust rule referring to [or F] in the structural
description.

This constraint states that if a rule 1’ refers to a feature value [01F] in its
structural description, then the redundancy rules assigning the feature [0F]
must automatically apply prior to the application of rule P. For example, if we
have a rule which refers to [-son], any redundancy rule which assigns [-son] to
unspecified segments must automatically apply before the rule. This constraint
hold for RUT, not for CUT. Once we adopt this system of automatic
application of redundancy rules, we no longer need rule-ordering among
redundancy rules and phonological rules“.

Moreover, as Archangeli points out, the RROC avoids the occurrence of
ternary oppositions of a feature (Archangeli 1984: 54-55). Since [02F] can

 

4 Porthe argument against the RROC, see Mester 8: Ito (1989: 265-266).

26

refer to both [+F] and [-F], they must be present prior to application of the rule
which refers to [GP]; there cannot be any segment unspecified for the feature.

This also means that when a rule specifically refers to only one value of a
given feature like [sonorant], the ROC motivates only a redundancy rule
supplying the value [-son] but not the one supplying [+sonorant]

2.2.4 Redundancy Rules under CUT

As underlying specification of segments differs between the two theories of
underspecification, redundancy rules of the two theories differ as well. CUT
defines redundancy rules as rules which fill in redundant feature values. In this
sense, those rules correspond to default rules of RUT. Unlike RUT, it has no
need to supply unmarked distinctive feature values, which are already present
in underlying representations. Because the ROC does not hold in this theory,
redundancy rules apply before or after a phonological rule.

2.25 Summary: RUT and CUT compared

RUT CUT
a. mmecifiedfeaures redmiantand unmarked onlyredurxiantfeaures
valueof distinctivefmtrres
b. underlyingly specified atmostonevalue bothvaluesfor mstinctivefmtrrcs

fmurevalues

c. redrmdancyrules langmg—mivcmlani language-miverml
lam-W1C

d rule-ordaingofRRs bytheRROC beforeordter pl'nmlogicalrules

27

2.3 Theories of feature systems

Binary and privative oppositions are not new to generative phonology.
In fact, a pre-generative phonologist Trubetzkoy (1969 (1939)) considered
such oppositions. What has been needed is the formal theory of feature
systems. In early generative phonology (Chomsky & Halle 1968), the binary
feature system was adopted as a formal device and claimed to be universal.
Recently, Trubetzkian feature oppositions have captured attention, which has
been deserved for decades.

In this section, I will introduce the two feature systems, privative (or

unary) and binary feature systems.

2.3.1. Trubetzkoy's feature system

Trubetzkoy (1969) classifies phonological features into three types of
oppositions: privative, gradual, and equipolent (binary). Some features such
as nasality, rounding and voicing are privative, i.e., single-valued features
which are expressed, for example, as [nasal] and [voice]. They are
characterized only as presence or absence of the feature, not in terms of binary
or multi-valued features. The property of being nasal is simply present or
absent.

Some other features are by nature to be expressed by virtue of the
difference in degree (gradation) of the same property. Vowel height and tonal
features in a tone language belong to this feature type, according to
Trubetzkoy.

28

Equipollent features are a type of features that are also known as binary
features and those whose opposition has an equal status: two opposing
properties are logically equivalent (den Dikken & van der Hulst 1988). This
type of opposition cannot be described either as privative or gradual
oppositions: it is the most frequent type (T rubetzkoy 1969: 75 ).

2 3.2 Binary Feature Hypothesis (Chomsky & Halle 1968)

The binary feature hypothesis claims that all features have binary values,
two opposing values such as + or -. In The Sound Pattern of English’ (1968;
henceforth SPE), Chomsky & Halle suggest that all features be fully specified
in binary feature-values prior to any application of phonological rules,
responding to the problem observed by Lightner (1963) and Stanley (1967)
that underspecification of a feature may allow ternary oppositions during the
derivations. So, since SPE, phonologists followed underlying specification of
features prior to the application of any phonological rules until
underspecification came back in Kiparsky (1982), Pulleyblank (1983) and
Archangeli’s (1984) work solve the problem of the three-way distinction in
underspecification.

Underspecification theories of phonological segments within generative
phonology were originally developed in the early 1980’s with the use of binary
features (Kiparsky 1982, 1985, Archangeli 1984) while recent versions of the
underspecification framework claim that certain features are by their nature
privative. For example, Steriade (1987) treats the feature [round] as privative

29

and [voice] and [lateral] as underlyingly privative in some languages.
Goldsmith (1985) claims that the determination of the privativity or the binary
nature of a given feature is language-specific, accepting the combination of

privative and binary features in a language system.

2. 3. 3 Privative Feature Theory

Privative or single-valued features play a role in phonological representation
with respect to the presence or absence of a feature. So, the feature [+voice] in
the binary system is expressed by the presence of [voice] and [-voice] by the
absence of the feature in the privative feature system.

In recent research, Trubetzkoy's feature system has been revived, as in
Goldsmith (1985), Mester & Ito (1989), Lombardi (1991). In this thesis, I
adopt Mester & Ito and Lombardi’s claim that voicing is privative. I do not
assume that all features are privative unless there is so me evidence. Thus, I

keep a binary system for other features such as [sonorant].

2 3.4. Privative vs. Binary Features

Some have pointed out advantages of the use of a privative feature system
for certain features (Goldsmith 1985, Steriade 1987, Christdas 1988, Mester &
Ito 1989, Lombardi 1991). For instance, the adoption of single-valued features
eliminates the issue of language-particular ordering among redundancy rules
(Steriade 1987). The privative feature [voice] reduces a number of processes

involved with voicing assimilation and simplifies the process (Mester & Ito

30

1989). Privative features will also result in simplification of the
underspecification approach (Mester & Ito 1989).

On the other hand, there is a case against privative features. den Dikken
& van der Hulst (1988: 45), though arguing against binary features, see the case
of the spreading of both values as problematic for the privative feature system.
Kiparsky (1991) also argues that assimilation or harmony where both values
spread must be analyzed in terms of binary features. If this is the case, then
Russian voicing assimilation argues for binary [voice]. However, this position
is based on the assumption that the spreading of both values is symmetrical: the
negative value [-F] spreads in the same way as the positive value [+F]. For
example, if the harmonizing feature is expressed in binary terms, then the
feature only needs to be spread by a single rule, regardless of the value, since
the spread of both positive and negative values are in symmetry in such a case.

The claim that privative features cannot work if a rule refers to both
values is reasonable since given a single spreading rule while the rule can
spread the plus value, there is no way to spread the opposite value.

However, this type of argument may seem questionable. Suppose that a
language hasassimilation of both values of a feature [F]. As far as the feature
[F] is binary, we have no problem with expressing the assimilation of both
values in terms of a single spreading rule ’spread [orF]’. However, when the
feature is indeed proved to be privative, then a single spreading rule, which
requires any feature to be binary, becomes problematic.

Recall that the basic assumptions of non-linear phonology that right
phonological representations entail correct rules. In the case above, a single

31

spreading rule cannot be right since the rule requires the feature [F] to be
binary and the feature is not properly expressed It is rather correct
phonological representations that yield correct rules Given this assumption,
we must f‘ust ask whether a feature is privative, binary or even multi-valued.
If a feature is indeed privative, then a rule which expresses assimilation of both
values of the feature cannot be a single spreading rule. With respect to Russian
voicing assimilation, which involves both values of the voicing feature, an
assimilation rule with privative voicing has to be something other than a single
spreading rule. Below we will see some evidence for the privativity of the

laryngeal features.

Lombardi (1991) convincingly argues for the privativity of laryngeal
features, based on the behavior of laryngeal features in laryngeal
neutralization. She observes that in laryngeal neutralization, which is
characterized as a loss of laryngeal distinctions, laryngeal features such as
voicing, aspiration, and glottalization often behave as a group; neutralized
obstruents are always plain voiceless. For example, Sanskrit (Whitney 1885,
1889 (cited in Lombardi 1991: 5)) has voiceless, voiced, voiceless aspirated
and voiced aspirated stops. In this language, only plain voiceless stops are
found word-finally. r

(15) [0I - an aspirated C]
a. lagnimath/ ->[agnimat] ’being near the fire’

b. /viirudh/ -> [viirut]

c. /suhrd/ -> [suhrt] 'friend’

32

(Translations for 17a and b are due to Cho 1990:147-148, for b, it is not given
in Lombardi 1991: 5)

This fact can be explained only in the feature-geometric analysis by
means of a delinking of the Laryngeal node of a segment, which dominates the
laryngeal features. The delinking of the Laryngeal node results in the
delinking of laryngeal features altogether. Moreover, the hypothesis that
laryngeal features are privative explains the fact that a neutralized obstruent is
always plain voiceless. If they are binary, then, we have to insert the minus
value of each feature involved. If this is the case, then it must be possible that a
neutralized obstruent can retain some of the laryngeal distinctions such as
aspiration while losing the voicing distinction. This, according to Lombardi, is
not the case. Thus we cannot explain why laryngeal neutralization always
results in loss of laryngeal distinctions altogether unless we take her privative
analysis of laryngeal features.

She also considers consequences of privative laryngeal features. She
regards voicing assimilation of both values not as a traditionally assumed single
rule but as two independently motivated mechanisms: one is the delinking of a
feature and the other the spreading of the feature.

Sanskrit voicing assimilation appears to support the privative analysis of
the laryngeal features as well. One example given in Cho (1990: 147) shows
that devoiced consonants result in plain voiceless.

(16) foodh + syate/ -> [biotsyate] ”wake up’
1 2

33

In example (16), Cho (1990) regards this devoicing process as a coda
devoicing. Observe that in (18) the aspiration is lost as consonant 1 becomes
devoiced. Although more data are needed, this phenomenon may also support
the privativity of laryngeal features since if the process is simply the delinking
of [+voice] and the insertion of [voice], the aspiration feature does not have to
be lost.

If we adopt privative features, voicing assimilation of the Russian type
must be treated as asymmetrian That is, the positive value [voice] can spread
whereas there is no way to express the spreading of the negative value, which is
simply absent in privative feature terms. The use of privative features,
therefore, requires us to have at least two rules to account for the data: one for
voicedness assimilation and the other for voicelessness assimilation.
Voicedness assimilation may be expressed by a single spreading rule such as
’spread [voice]’ while voicelessness assimilation needs to be expressed by
’delink [voice]’.

2 3.5 Privative features and assimilation processes

Section 2.1.3 defined the notion of feature-changing and non-feature-
changing processes in autosegmental phonology. When an assimilation rule
expresses an assimilation feature by unary-features in place of binary ones, the
rule becomes non-feature-changing. Similarly, Russian voicing assimilation,
which was once analyzed in terms of a feature-changing process (Kiparsky
1985), is treated as non-feature-changing ones under a privative feature
system.

3. Russian Voicing Assimilation

This chapter addresses the facts and problems of Russian voicing
assimilation. In particular, I will call the reader’s attention to assimilation of
both values within an obstruent cluster, the dual-characteristics of the voiced
labio dental fricative v, and the transparency of sonorants (nonvocalic
consonants) to voicing assimilation.

3.1 The Russian sound pattern and transcriptions

Russianis a Slavic language with thirty six consonants and five vowels as
phonemes. Palatalization is phonemic in this language. The phonemic
inventory for Russian (Halle 1959, Jones & Ward 1969) is provided in (17):
(17) Russian Phonemic Inventoryd (C-pdatalized C)

bilabial labio- alveolar palatal velar

dental
plosive(stop)
voiceless p p’ t t’ k k’
voiced b b’ d d’ g
fricative
voiceless f f’ s s’ E :e x
voiced z z’ i "
affricatc
voiceless t' 5
nasal m m’ n n’
flap r r’
lateral l l’
glide w w’ j

 

5 The phonemes wand w' never appear on surface; however, for the purpose of the discussions,
they are assumed to he the underlying forms of v and v; respectively.

34

 

35

Vowels
front back
i u high
e o
a low

As for the transcriptions of Russian data, basically I use broad
transcriptions throughout (Ladefoged 1982: 34-38, Lockwood, personal
communication). The underlying representation is indicated in / /. With
respect to a surface representation of Russian sounds, only sounds relevant to
discussions such as obstruents and intervening sonorants are phonetically
transcribed within brackets [ ]; unbracketed sound, which are not directly
relevant to discussions, remain unchanged (except w and w’)‘. For instance,
given /jug/-> ju[k], jug in / / is the underlying representation and k in [ ] is
phonetically transcribed and other sounds remain unchanged in terms of
transcriptions.

For palatalized consonants, C’ stands for a palatalized C. While I use
most of the data from published works, the sound pattern is due to Moscow

dialect. s’ and z’ are found in the Moscow version of standard Russian7.

 

6 Because lassume that [v] is derived from lw/ and both underlying and surface representations of
the segment is at issue in this thesis, lwl is transcribed as [v] as the surface representation even
when the segment is not in an obstruent cluster.
7 For instance, these sounds are not found in St. Petershurg version of standard Russian
(Lockwood, personal communication) as the comparison ofthe pronunciation of two words below
show:
SP: lji‘do/ 'stiII‘ Moscow: liil'ol

/ dald‘al 'of rain‘ Idai'a/

36

3.2 Language Data

It has been argued that Russian has two phonological processes involving
obstruents: neutralization and voicing assimilation (Kenstowicz & Kisseberth
1979; henceforth K&K). The neutralization rule is the devoicing of a word-
final obstruent. This rule is argued to feed the rule of voicing assimilation
(K&K 1979), a regressive assimilation that requires the members of an
obstruent cluster to agree with the cluster-final obstruent in voicing, either
voiced or voiceless. In addition, both processes are independent of place of
articulation. Russian voicing assimilation is not determined by syllabification.
Rather, Russian voicing assimilation takes place word-internally (across a

morpheme-boundary) as well as across a word boundary.

(18) Word-Final Obstruent Devoicing (Neutralization )

a /zapad/-> zapa[t] ‘west’

b. Aug /-> ju[k] ’south’

cf. na yu[g’]e ’in the south’

c. /mu§/ -> mu[§] ’a husband’ cf. muéja ’husbands’
d /sos’ed/- > sos’e[t] ’a neighbor’ cf. sos’edi ’neighbors’
(l9) Voicing Assimilation

Word-internal ( + : morpheme boundary)8

a /pros’+ba/ -> pro [z’b]a ’request’

b. Is + daéa/ -> [zd]aca ’submission’

c. /sos’ed+ka/ -> sos’eltk]a ’a female neighbor’

 

8 I provide a morpheme boundary and a boundary between aproclitic and sword where itis relevant
to the discussions.

37

Across the word-boundary ( - : boundary between a proclitic and a word)
e. /k - zapadu / -> [gz]apadu ’to the west’
f. Atak - d’ela/ -> ka[gd']ela ’How are things ?’
g. /ot -doma/ -> o[dd]oma ”from the house’
cf. o[t k]uda ’from where ?’
h. lb’ez-kukli/ -> be[sk]ukli ’ without a doll ’

(20) Interaction of Final Devoicing and Voicing Assimilation
a. /vizg/ -> vizk -> vi[sk] ’scream, nomsg.’

cf. vi[zg]a ’scream, gensg.’

While all obstruents do undergo voicing assimilation, not all obstruents
can trigger it. For instance, the voiced labio-dental fricative /v/ fails to trigger

the assimilation:
(21) a. [tvloj ’your. masc.sg.’
b. [sv]o j ’ one’s own, mascsg ’

c. zdra[stv]ujt’e ’ hello’

CL mo [skv]a ’Moscow’

e. ka[k vlipa’éivaet’e ’ how are you?’
cf. /kak d’ela / -> ka[g d’]ela ’ how are things ?’
f. [s vlam’i ’with you’

[1'] can be derived either from 117 or /v/. In addition, [1'] derived from /v/
and the one not derived from /v/ behave differently in triggering the voicing

assimilation: derived f, just like v, fails to trigger the voicing assimilation

 

 

38

(Reformackij 1971 9, Jakobson 1978, Hayes 1984, Kiparsky 1985) whereas
underived f does trigger the assimilation:
(22) [I] from /v/
V a. ltr’ezv/ —> tr’e[zf] (" tr’e[sf] ) ’sober, predicative adj.’
(Hayes 1984, Kiparsky 1985)

cf. [tr’ezvij] ’sober, attributive adj.’
b. /r’ezv/ -> r’e[zf] (‘r’e[sf]) ’frisky’ (Jakobson 1978)

cf. [r’ezvijt’s’a] 'frisk’
c. /xorugv’/ ~> xoru[gf’] (‘xoru[kf']) ’gonfalon’ (Jakobson 1978)

This fact strongly argues for the asymmetry of the segment v: it never
triggers voicing assimilation while it undergoes it. Although there is a
controversy over the pronunciation of words like the above. Some argue that
word-final v not only undergoes the word-final devoicing but also triggers
voicing assimilation. For instance, /tr’ezv/ is pronounced rather as [tr’esf],
which is a result of devoicing and voicing assimilation. Some other argue that
as a fast speech effect after final devoicing tr’e[zf] becomes tr’e[sf] (Hayes
1984). Yet, I take the pronunciation given in (22), as it has been analyzed in
recent works (Jakobson 1978, Halle &Vergnaud 1981, Hayes 1984, Kiparsky
1985).

Underived /f/, on the other hand, can both undergo and trigger the
voicing assimilation.

(23) If/asatarget of the assimilation A7 ->[v]
a. /prof-b’uro/ -> pro [vb’] uro ’trade union bureau’ (Grebnev 1959)

 

9 [am grateful to Prof. D. Lockwood for introducing me this important article.

39

In normal speech, many Russian proclitics, e.g. prepositions, are
pronounced as if they belong to the following word: so the last segment of a

preposition undergoes voicing assimilation.

(24) A7 as a trigger of the assimilation: a voiced consonant becomes voiceless
when it is followed by underived f :
a. fo’ez- fonet’iki/ -> b’e[sf]onet’iki ’without phonetics’
cf. lb’ - gori/ -> b’e[zg]ori ‘without a mountain’
b. A2 - frant'ii / -> i[sf]rant'ii ’ from France’

From these facts, we can state that all obstruents except v both trigger and
undergo voicing assimilation; v undergoes voicing assimilation but fails to

trigger voicing assimilation.

We still need to examine cases in which voicing assimilation affects an
obstruent cluster which contains v or sonorant consonants like nasals.
(25) [v] as a target of voicing assimilation
a. /ot -vzriva/ -> o[d vzlriva ‘from an explosion’ (J)

b. /ot- vzdoxa/ -> o[d vzdjoxa ’fromasigh’ (J)

Observe that intervening sonorant consonants10 in obstruent clusters are
transparent to voicing assimilation in Russian (Jakobson 1978: 108, Hayes
1984: 318-320):

 

1° hrapid speech, Russian sonorant comments may devoice before a voiceless obstruent, e.g.,
[gt )a 'mouth, gensg.‘ (Jones 8: Word 1965 (cited in Hayes 1984: 321)).

(26) Voicing assimilation across sonorants:

a. fiz-mt'enska/ ->i[smt']enska ’froth'enska’ (Hayes 1984)
b. lot-mzdi/ -> o[dmzd]i ’from the bribe’ (H)
c. Inad- rtut’yu/ -> na[trt]ut’yu ’ over mercury’ (J)

(I. /k- lzivim slovam/ -> [glz]ivim slovam ’to the mendacious words’(J)
e. /s- rd’ejust’im rum’ant‘em/ -> [zrd’]ejust'im rum’ant'em

‘with a glowing color’ (.1 )

In short, we have seen the phenomena of Russian voicing assimilation.
The assimilation is of both values. It takes place within an obstruent cluster, in
which sonorant consonants may intervene. Word-final obstruent devoicing
seems to interact with voicing assimilation. As for the asymmetrical behavior
of the labio-dental v, the segment, just like a sonorant, never triggers voicing
assimilation while it undergoes both final devoicing and voicing assimilation as
any other obstruent does. Intervening sonorant consonants in an obstruent
cluster are transparent to the process. One aim of this thesis is to provide a
natural account for the special behavior of v in comparison with underived If/

and with other obstruents within the context of voicing assimilation.

 

4. Historical Perspective

Russian voicing assimilation and final devoicing has been studied for
decades with respect to the definition of labia—dental fricatives and
classification of the sounds which involve the process. Russian voicing
assimilation was quite challenging for segmental iterative approaches since it is
unbounded propagating assimilation which affects the voiced labio-dental
fricative v in an obstruent cluster. Although several attempts have been made
within an autosegmental framework, none has provided a uniform account for
both the asymmetrical behavior of the voiced labio dental fricative v and
voicing assimilation of both values. In this section, I will provide a summary
of published analyses of labio-dental fricatives in section 4.1 and those of

voicing assimilation in Russian in section 4.2.

4.1 Russian labio dental fricatives
4.11 Interpretation of Russian labio-dental fricatives

The peculiar behavior of Russian [v] mentioned in the previous chapter
has been studied for decades in the literature. Many researchers have suggested
that [v] is underlyingly /w/, a sonorant and becomes an obstruent v at some
point of derivation (Jakobson 1948, Halle 1959, 1973, Lightner 1965, Coats
& Harshenin 1971, Hayes 1984, Kiparsky 1985). This analysis seems plausible
since v behaves like an obstruent on the one hand, and like a sonorant on the
other. On the other hand, Barkai &. Horvath (1978: 82) have objected that
deriving a segment from another which never surfaces is not realistic and such

analysis requires extrinsic rule-ordering. From the standpoint of second

41

42

language acquisition, as observed by Shapiro (1966: 193), native speakers of
Russian learning English tend to substitute v for w.
(27) a. wine -> [v]ine b. woman ~> [vjoman

c. Taiwan -> Tai[v]an d. lower -> lo[v]er

e. warsaw -> [v]arsa[v]a

Below we will see how such a peculiar segment v has been conceived of in
relation to voicing assimilation. Although various authors, both phonologists
and phoneticians, disagree on some data like tr’e[zf] vs. tr’e[sf], we take the
analysis that v fails to trigger voicing assimilation like other sonorants but does
undergo the process like other obstruents. Among /w/ analyses, I present that
of Lightner (1965) and Hayes (1984). As for the one against /w/ analyses,
Barkai & Horvath’s analysis (1978) is introduced.

Lightner (1965 (cited in Barkai & Horvath 1978: 79)) takes the abstract
w analysis. He distinguishes two types of /w/: w at the word-final position and
before an obstruent; w in all other environments While w in any
circumstances becomes obstruent v, only the former type undergoes voicing
assimilation. He constructed extrinsically ordered rules to prevent the latter
type of w from triggering the assimilation.
(28) (Lightner 1965 (Cited in B & H 1978 :79-80))
a. w -‘> v/_{ [sonorant] , #}
b. Voicing Assimilation

{-son] ->[a voice] /_ [-son, a voice]

c. w ->v (elsewhere)

(condition: rule-ordering a-b-c)

43

(29) a. / b’ez Wpuska/ /tvoj/
by 28a 2 v p n/a
by 28b 2 f p man
by 28b sf p n/a
by 28c n/a tv
b’e[s fp]uska [tv]oj

As derivation (29a) clearly shows, rule (28b) entails that v which is derived by
rule (28a) does trigger the voicing assimilation. His rules thus fail to capture
the generalization that v never triggers voicing assimilation. This is a
problematic aspect for any segmental iterative approach, according to which a
segment can affect only an adjacent one at a given time. We will come back to

this issue in the following section.

Barkai & Horvath (1978) argue against the abstract /w/ analysis of the
Russian labio-dental f ricative v. According to them, an approach of deriving
[v] from /w/ is unrealistic and undesirable since the abstract /w/ never surfaces
in Russian and it also requires extrinsic rule ordering (B&H 1978: 82).

Their close examination of the peculiar behavior of v found in three
unrelated languages Hebrew, Hungarian and Russian has led them to introduce
a sonority-based approach, in which the labio-dental fricative v is ranked
between fricatives and nasals (B & H 1978: 83).

(30)
stops fricatives v nasals y r l
l 2 3 4 5 6 7

(the higher number, the higher degree in sonority)

Based on the sonority scale provided above, Barkai & Horvath propose a
voicing assimilation rule for Hungarian and Hebrew(B & H 1978: 83):

(31) m. sonorant -> [or voice] /_ [ n. sonorant,or voice]
where m;-: 3 and n 5 2

The rule above states that a segment with sonority equal to or less than
degree three assimilates to the following segment in voicing if its sonority is
equal to or less than degree two. All stops and fricatives except v can be the
potential trigger of voicing assimilation and v can only undergo the process.

It cannot explain the case that an obstruent cluster contains [v] cluster-
medially. Their rule is a segmental iterative one, under which [v] blocks
voicing assimilation because their rule cannot allow [v] to trigger voicing
assimilation. For instance,

(32) a. b’ez vpuska -> " b’e[z fp]uska ’without permission’
cf. b’e[s fp]uska

b. kak vdova —> " ka[k vd]ova ’like a widow’

cf. ka[g vd]ova

Hayes (1984), supporting the view of treating v as a sonorant /w/
underlyingly, reanalyzes Coats & Harshenin’s (1971) approach. He proposes a
set of extrinsically ordered rules. Among them, the devoicing of a word-final
obstruent and voicing assimilation which apply to any consonant including
sonorants. For now. we concentrate on the treatment of v and we will discuss

his autosegmental rule for voicing assimilation in the next section.

45

(a simplified version of Hayes 1984: 319-320)
(33) Final Devoicing
C -:> [-voiced]/__#
(34) Voicing Assimilation
Assirnilate all consonants in a cluster to the voicing of its rightmost
obstruent.
(35) W strengthening
[C, -cons,+labial] -> [-son]
(36) SonorantRevoicing
[+son] -> [+voiced]

(condition: rules are ordered as 34-35-36-37)
(37) Derivations:
UR (a)/o’ez-wpuska/ (b) Iiz-mt'enska/ (c) /tr’ezv/

by (33) 2y
(34) S ‘yp 5 mt' n/a
(35) s fp n/a zf
( 36) n/a 5 mt' n/a
SR [b’es f puska] [is mt‘enka] [tr’ezf]

In derivation (37a), rule (33) devoices z and w then w becomes f by w
strengthening. In (37b), voicelessness assimilation also takes place; the
devoiced sonorant m gets revoiced by a sonorant revoicing rule. In (37c), w
devoices as a word-final consonant and becomes f by the w strengthening rule,
which bleeds the sonorant revoicing rule.

Thus, sonorants become devoiced by undergoing neutralization (i.e.,
word-final devoicing) and voicing assimilation Hayes, then, has to construct a
rule which later revoices the devoiced sonorant, assuming that /v/ is not a
phoneme in Russian while If/ is. During the derivation, M/ is treated as a

sonorant and gets devoiced by the former rule. But he constructs the rule /w/

strengthening. /WI ->[v] and lgv/ -> [f]. This rule is ordered to apply before the
sonorant revoicing rule. As in (37b), all sonorants except /w/ have to go
through two processes, devoicing and revoicing in order to explain the
asymmetry of [v]. Thus, such operations should be replaced by a simpler set of
rules.

The difficulty that Hayes’s analysis has seems to stem from the nature of
his assimilation rule since the rule affects all consonants in an obstruent cluster,
both obstruents and sonorants, even though at the surface only obstruents show

agreement in voicing.

4. 1.2 Abstract /w/ and underspecification

We have seen the abstract /w/ analysis and the argument against the M/
analysis. Barkai & Horvath (1978) are right in pointing out that /w/ never
surfaces in the language. Yet we have no evidence that it is underlyingly not
sonorant. Underspecification, specifically, Radical Underspecification,
enables us to treat it as a neutral segment. while we assume that it is
underlyingly /w/ not /v/. According to RUT, at most one value of a feature can
be present in underlying representations, we can specify [sonorant] and so
unspecify [+sonorant]. Consequently, v, which is assumed to be /w/, is
underlyingly underspecified for the feature sonorant. So underlyingly it is
neither sonorant nor obstruent. It becomes [sonorant] by a default rule [ ] ->
[sonorant] / [_, +continuant, +labial]. We can explain without any feature-
changing rule why it fails to trigger the assimilation while it undergoes it. We
no longer need to decide when the labia-dental fricative v underlyingly belongs

to one of the two groups.

47

4.2 Approaches to Voicing Assimilation and Final Devoicing

In the last three decades or so, several attempts have been made to
provide a uniform account for the phenomena involved with the two
phonological processes final devoicing and voicing assimilation in Russian. In
early generative phonology, the phenomena were analyzed and expressed by
means of segmental or directional iterative rules. For instance, Kenstowicz
and Kisseberth (1979) formulate segmental iterative rules within the

framework of generative phonology:

(38) Final Devoicing: (K&K, P409)
[-sonorant] -> [voiceless] /_ #.

(39) Voicing Assimilation

[-sonorant] -> [a voice] /_ [-sonorant ,or voice]

Rule ordering: 38-39

‘ Rule 38 feeds rule 39.

(40) Derivations

a. /vizg/ b. /pros’+ ba/
by38 zk n/a
by 39 sk z’b

vi[sk] pro [z’b]a

48

Their rules are not meant to cover every phenomenon of Russian voicing
assimilation: they deal with simple cases like /vizg/ ->vi[sk] but they fail to
explain cases like ltr‘ezv/ -> tr’e[zf] or lb’ez wpuska/ -> b’e[s fp]uska.

Aside from this, there is a case in which segmental iterative approaches in
general are shown to be inadequate to explain phenomena involved with
Russian assimilation

For example, Hayes (1984) provides examples to argue against the
segmental iterative approach to voicing assimilation. He claims that Russian v
allows unbounded propagating assimilation, which, according to him, can be
better accounted for by a prosodic approach than by a segmental iterative one,
supporting the analysis of Halle & Vergnaud(1980). ‘

The unbounded propagating assimilation in the Russian case means that,
in a given obstruent cluster, the voicing assimilation trigger, i.e, the last
member, affects all targets, thus its adjacent segment as well as nonadjacent
ones. In other word, the voicing feature of the trigger spreads onto all
consonants until the end of the cluster. Consider the examples below:

(41)

a. lb’ez z’ertv bj/ -> b’ez z’er[dv b]j ’without sacrifices (Shapiro 1966)
b. Ikakvdova/ -> ka[g vd]ova ’like the widow’ (S)

c. /vot v d’etstv’e/ -> vo[dv d’]etstv’e ’but inchildhood’ (S)

d. /ctob vzjat/-> cto[b vz]jat ’ in order to take’ (8)

e. lb’ez-wpuska/ -> b’e[s fp]uska ’without permission’

49

(42) Assimilation across nasals:

a. 152 mt'enska/ -> i[s mt']enska ’ from Mcensk’ (Hayes 1984 )
b. /ot mzdj / -> o[d mzd] j ’ from the bribe’ (H)

(43)

a. [tvloj ’your, masc.sg.’ b. [s v]ami ’with you’

c. lu’ezw/ -> tr’e[zf] ’sober’

First, nasals are transparent to assimilation (42). Second, the voiced
labio-dental fricative v is not transparent (41). As (43) shows, v cannot trigger
the voicing assimilation. Therefore, it follows that assimilation is not a process
of transmitting the voicing f eature of the final member of each obstruent
cluster in a segmental iterative way such that a trigger affects only the adjacent
segment, which in turn affects the segment to its left. It is rather a propagation
process, by which the final member of the cluster affects all targets.

Under a segmental iterative rule, any segment that undergoes the rule is
at the same time a potential trigger (Halle and Vergnaud 1981; Hayes 1984;
Kiparsky 1985). Accordingly, segmental iterative analysis is problematic since
it require the cluster-medial labia-dental fricative v to affect the adjacent one in
propagating assimilation. To illustrate the point,

(44) b’ez w p uska -> b’ez f puska -> b’e[s fp]uska
C1 C2 C; C1 C2 C3

Under a segmental iterative approach, the last member C3 triggers C2,
which in turn triggers C1. On the surface it seems that voicing assimilation has
taken place iteratively. The facts that v fails to trigger voicing assimilation as
in words like tvofyour’ and svof’one’s own’ and also that sonorant consonants

are transparent to voicing assimilation tell us that it is not the case.

50

Any segmental iterative analysis such as that of Lightner (1965), Coats &
Harshenin (1971) and Kenstowia & Kisseberth (1979) fails to explain a case
like (44) in which the C3 triggers C2 and C1 iteratively without making C; a
trigger. As we have seen, Lightner (1965) had to classify v into two types: v
which both triggers and undergoes voicing assimilation and v which does not.
Thus, under a segmental iterative analysis, phenomena like Russian voicing

assimilation are never interpreted as unbounded propagation.

An autosegmental analysis, on the other hand, accounts for unbounded
propagation phenomena by affecting segments at a long distance by means of

spreading.

(45) a. lb’ez w p uska/ ->b’ez f puska -> b’e[s f pluska
C1 C2 C3 C1 C3 C3 C1. C2 C3
| \‘j ‘ .
[-voice] [-voice] [~voice]
b. /ot m2 dj/ ->o[d m z d]j ’ from the bribe’

C1 C2 C3 C1 C2 C3

x
\
§

[+voice] [+voice]

Hayes formulates what he calls a prosodic rule for Russian voicing

assimilation as follows:

(46) (Hayes 1984: 325)
In the configuration [son], reassociate [a voice]
I
[or voice]
with the maximal string of consonants to the left, deleting any autosegments
stranded in the process.

51

e.g.
{-son]
I
b’ez w puska -> b’es w puska
l I l
[+voice] [+voice] [-voice] [-voice]

As I mentioned above (p46-47), Hayes’s assimilation rule, even when
prosodically formulated, applies to any type of consonant including sonorants:
any sonorant may be devoiced if the process is triggered by a voiceless
obstruent. Thus, under his approach, we need a rule which revoices sonorants
devoiced by the earlier application of rules. In addition, his rules are so
ordered that the /w/ strengthening rule bleeds the sonorant revoicing rule in
that v, which is already an obstruent, cannot undergo so norant revo icing.
Hayes justifies his sonorant revoicing rule by saying that it is not a phonological
rule but phonetic implementations of the language.

Hayes’s approach consists of an ordered set of rules. Unlike a segmental
iterative rule, in his prosodic rule the final member of an obstruent cluster is
the only trigger affecting all the consonants in the cluster. Although this is a
great improvement, the existence of the two rules, w strengthening and
sonorant revoicing, assures that the assimilation process takes place iteratively,
without skipping any consonant which never shows agreement in voicing on the
surface.

In other words, Hayes’s autosegmental rule can prevent a sonorant from
triggering the voicing assimilation, but it cannot prevent the trigger from
skipping or affecting sonorants since the rule does not refer to the feature
sonorant for the targets. My claim, then, is that if we can have a rule which

52

refers only to relevant features of targets, i.e., the feature obstruent, we will no

longer need such rules as w strengthening and sonorant revoicing.

Halle & Vergnaud’s (1981) also make use of autosegmental rules and are
quite successful in affecting only obstruents in voicing assimilation Their
identifying obstruent clusters prior to any phonological rules is the key to such
asuccess.

They consider Russian voicing assimilation as directional harmony: the
process takes place in one direction, being triggered by the final member of an
obstruent cluster. Their approach is simpler than Hayes’s in the sense that their
voicing assimilation rule affects all and only obstruents by constructing a tree
that consists of only obstruents. This is done by what they call foot
incorporation Foot incorporation determines an obstruent cluster and takes
place before any phonological rules. It reshapes a tree, which is already
formed by tree construction, by terminating a branch dominating a sonorant so
that a tree comes to consist of only obstruents. Only elements constituting a
tree participate in voicing assimilation Voicing assimilation is expressed in
terms of copying the voicing feature of the last member of the tree onto all
other members. Since sonorants are disregarded by foot incorporation,
voicing assimilation takes place skipping intervening sonorants.

Halle & Vergnaud provide a set of rules as below:
(47) (slightly modified version of Halle & Vergnaud 1981: 17)
a Tree Construction

b. Foot incorporation

53

c. w ->v _
d. word-final obstruent devoicing

e. copy feature on root

Let us see how these rules generate voicing assimilation Only two types
of obstruent clusters are given here; The tree construction is limited to
segments relevant to voicing assimilation for the purpose of discussion:

(48)

males
liz- mt'enska/ -> i[s mulenska ’from Mcensk’
lot-mzdi/ -> o[d mzd]i ’ from the bribe’
Deductions.
Rules
3. iz-_mt'enska oL-rnzdi
b. iz-_mt'enska ot;mzdi
c. n/a n/a
d. n/a n/a
/[.\V] [K
e. Menska od-mzdi

(Tree construction of only relevant segments are given)

First, by. initial tree construction (a), all segments are clustered together
to form trees with the condition that vowels are opaque, i.e., triggering the
harmony. Then, by foot incorporation (b), sonorants are eliminated from the
trees and adjacent obstruents are incorporated into a single tree. At this point,
obstruent clusters are recognized. The trees are right-headed; that is, the final
member of a tree governs the obstruent cluster voicing as in copying rule (e).
By (e), the voicing feature of the final member of the tree is copied onto other

members of the tree.

Unlike Hayes’s assimilation rule which affect all consonants in an
obstruent cluster, Halle & Vergnaud’s analysis allows the assimilation trigger
to affect only obstruents by the copying rule. Thus, they do not need two
processes, devoicing intervening sonorants and revoicing them.

Although Halle & Vergnaud (1981) are successful in disregarding
sonorants, which never participate in Russian voicing assimilation, their rules
have not explained the asymmetrical behavior of v. They have to add an ad hoc
rule for the voicing of v (Hayes 1984:320) as below:

(49) {-sonorant, +labial, +continuant] -> [or voice] /_ [ or voice, -sonorant]

(H &V 1981: 17)

This problem must be worked out; yet I take their analysis as significant
and as an important step toward the most appropriate explanation for
assimilation phenomena of the Russian type. As we will see in the following
chapter, their notion of foot incorporation is somehow modified and adopted in

underspecification approaches to the phenomena.

55

4.3. Summary

As we have seen, many researchers have given a very insightful
explanation for some of the Russian voicing assimilation phenomena over the
decades, however, none have provided a uniform account for every single
phenomenon

The traditional analysis of deriving [v] from /w/ seems plausible, given
the segment’s dual-characteristics. When the /w/ analysis is assumed, a feature-
changing w strengthening rule is required to explain the asymmetry of the
voiced labio-dental fricative v in the approaches which fully specify /w/.

While Hayes’s assimilation nrle expresses propagation phenomena by
spreading the voicing feature of the trigger onto targets, his rules are
extrinsically ordered. They devoice sonorants and then revoice them, which is
irrelevant to the surface realization of obstruent-cluster voicing.

Halle & Vergnaud’s autosegmental analysis has shown that an
autosegmental rule of voicing assimilation can affect all and only obstruents.
Their approach identifies obstruent clusters early in the derivation before any
phonological rules apply.

In the following chapter, I will propose an approach to Russian voicing
assimilation, solving the problems that previous analyses had.

 

 

 

5 . A Constraint-based Approach
to Russian Voicing Assimilation

In this chapter, I will present an approach which uses a structural
constraint on obstruent-cluster voicing, Radical Underspecification and
privative voicing. This approach solves the problems that previous studies had
and gives a uniform account for the phenomena of Russian voicing
assimilation.

In my approach, a structural constraint on obstruent-cluster voicing
motivates persistent application of two processes: spreading and delinking of
[voice]. By Radically underspecifying the voiced labio-dental fricative v, we
can eliminate the need for the f eature-changing rule [+sonorant] to {-sonorant]
for the segment. By adopting privative [voice], we can reduce feature-
changing assimilation to a non-featm-e-changing type. In this approach,
assimilation triggers are identified by Obstruent-Head Licensing (OHL) prior
to any application of rules, both redundancy and phonological ones, and thus
before the unspecified w becomes v. Voicing assimilation affects all and only
obstruents by referring to the contextual feature [sonorant]. By OHL, all
obstruents except v can trigger voicing assimilation Application of OHL is not
subject to extrinsic rule ordering because it is not like any phonological or
redundancy rules. It rather provides a domain, in which voicing assimilation
takes place. .

Then I will defend my solution by objecting to two other analyses, which
one might easily come up with. One is based on the modified version of
Hayes’s spreading rule using binary features. The other is privative voicing

S7

with Contrastive Underspecification. I will show why an approach using a
structural constraint is desirable, pointing out problems with Hayes’s spreading
rule. I also favor Radical Underspecification over Contrastive
Underspecification in the Russian case.

When examining how the use of privative voicing is more advantageous
than binary alternative with regard to representations and rule applications for
voicing assimilation in Russian, we keep binary values for any other feature,
since voicing is the only assimilation feature here and since we have no
evidence that other relevant features like sonorant are privative.

Before analyzing the data within the underspecification framework, let us

look at rules, constraints, and assumptions for underspecification.

5.1 Underspecification
The Russian data will be examined in terms of the underspecification
theories, which can differentiate v from any other obstruent and sonorant in
order to explain the Special behavior of labio-dental fricative v and the two
phonological processes in Russian more simply than Hayes’s pr0posal. There
are several assumptions to be made here within the framework of
underspecification
First, I take basically the traditional analysis of Russian labio-dental
fricatives: [v] is derived from /w/; [f] can be derived either from /f/ or /w/. [1']
is derived from /w/ if it has undergone a devoicing process. Underived f is
underlyingly an obstruent If/ and triggers voicing assimilation v is a
consonant with the features [+labial] and [+continuant] underlyingly and will

become an obstruent at some point of derivation by a redundancy rule.

58

Although it may be highly misleading, I assume that underlyingly lw/ just
stands for a consonant with the specification [+continuant, +labial] and is
unspecified for [sonorant] and [voice]. This consonant is expressed as /w/. So,
underlying /w/ is specified neither as a sonorant nor as an obstruent under RUT
(but not CUT). Thus, as far as features relevant to Russian voicing assimilation
are concerned, v is underlyingly a segment without features [sonorant] and
[voice] under RUT. Under CUT, on the other hand, w cannot be featureless (in
a sense that w without [+sonorant]) since the theory insists on underlying
presence of both values of the contrastive feature [sonorant]. As we will see in
the later section, the underlying presence of [+50 norant] brings CUT a

problem: it cannot explain the asymmetry of v as simply as RUT does.

It is widely accepted that an approach with extrinsic rule-ordering
involves complexity and is thus highly disfavored over one without (Hudson
1974, Barkai & Horvath 1978). In the previous chapter, we saw approaches
using a set of extrinsically ordered rules such as application of w strengthening
before sonorant revoicing. Those rules need rule-ordering because the
asymmetrical segment v is always specified for either value of the feature
so norant: it is underlyingly specified as sonorant and later becomes an
obstruent. In other words, in spite of its dual-characteristic in voicing
assimilation, in the previous approaches, the labio-dental fricative v always
had to belong to either of the two groups of consonants, sonorants or
obstruents.

Recall that the dual-characteristic of v is that it undergoes voicing

assimilation while it fails to trigger the assimilation Now that as a consequence

59

Recall that the dual-characteristic of v is that it undergoes voicing
assimilation while it fails to trigger the assimilation Now that as a consequence
of underspecification, we have v as a segment without the features [sonorant]
and [voice], we can easily differentiate the asymmetrical segment from any

other obstruent and sonorant.

5.1.1 Obstruent clusters and Obstruent-Head Licensing

We have seen in chapter four that various authors provide different rules
for the phenomena of Russian voicing assimilation Though those rules differ
in fundamental assumptions as well as in formulation, all seem to agree that it is
the final member of an obstruent cluster that triggers the assimilation

I take Halle & Vergnaud’s analysis, in which obstruent clusters are
identified prior to any application of phonological rules as significant But I do

not maintain such a process as tree incorporation because autosegmental ,

phonology does make it possible for a rule to refer to a particular group of
sounds like obstruents and sonorants. H &. V’s foot incorporation is a process
which excludes all sonorants from a tree so that a tree consisting only of
obstruents results. This process is not necessary because when an
autosegmental rule explicitly specifies {-son] in its structural description, then
segments with [+sonorant] are simply ignored in the application of the rule.
Instead, I propose below a rule which identifies the trigger of voicing
assimilation First, 1 term the final member of an obstruent cluster the
obstruent head. Second, I define an obstruent cluster as a sequence of
obstruents, within which sonorant consonants may intervene. My proposal is a

language-specific process called Obstruent-Head Licensing, which scans the

60

CV and sonorant tiers and identifies the head, the potential trigger of voicing

assimilation within an obstruent cluster:

(50) Obstr'uent-Head Licensing(OHL):
License the final member of an obstruent cluster identified in the

underlying representation as the trigger of voicing assimilation.
{-son] {-son]

C (C) C, (C1= Obstruent head)

The CV -tier is needed for this process since whatever is to be identified as an
obstruent cluster must be a sequence of consonants specified as {-son]. The
cluster may contain intervening consonants unspecified for {-son] while it
cannot allow any intervening segment specified as V, that is, a vowel.
Obstruent-Head Licensing applies only once after underlying representations
and prior to any application of rules, licensing the final member of an
obstruent cluster as the trigger of voicing assimilation OHL provides the
environment for the voicing assimilation rule. This OHL applies only to
underlying representations. It means that v, which is not specified as {-son]
underlyingly while becoming {-son] later in the derivation, is not subject to the
rule. Thus, v never becomes the trigger of voicing assimilation even in such a
configuration as [ _zv ] or [ tv_]: v never triggers voicing assimilation,
which is in accordance with the facts. Consequently, we no longer need
extrinsic rule-ordering, since as the effect of O-head Licensing, there is no

circumstance in which v becomes a trigger of assimilation

61

5.1.2 Obstruent-Cluster Constraint

In the previous chapter, we saw that Hayes’s assimilation rule affects all
consonants in an obstruent cluster and requires a sonorant revoicing rule while
at the surface only obstruents show agreement in voicing. We can economize
such seemingly unnecessary processes by formulating a rule which refers to the
contextual feature for the assimilation {-sonorant].

I propose a structural constraint on the voicing feature of obstruent-
clusters, which motivates persistent application of a rule changing ill-formed
representations to well-formed ones. I assume that voicing assimilation is
persistent to eliminate the need for rule-ordering. To see how such a structural
constraint motivates a persistent rule and how a persistent rule interacts with
redundancy rules, we first look at a constraint expressed with a binary feature.
Then, we will see a constraint using privative voicing.

(51) Obstruent-Cluster Constraint (OCC)
0
{-son] {-son]

I I
c c1

[or voice] [a voice]

This says that no members of an obstruent cluster can have an opposing value to
that of the head: all members of o-cluster must agree in the voicing feature
with the head. Observe that the rule simply refers to {-son] and voicing
features; any intervening sonorant consonants, which are unspecified for the
features sonorant and voice, are disregarded. This constraint motivates a

62

leftward spread of the voicing feature of the head and then motivates a
delinking of the voicing feature assigned to non-head members of the cluster as

illustrated in (52).

(52)
{-son] {-son]
I l
C C,
a J

[or voice] [or voice]

To be clear, it is constraint (51) that motivates the persistent application
of rule (52). I assume that the OCC-motivated rule applies persistently, that is,
it applies whenever structural conditions are met (for persistent rules, see
2.1.2). So this constraint-motivated rule interacts with redundancy rules,
changing ill-formed representations to well-forrned representations. Unlike
universal constraints, which prohibit the occurrence of ill-formed
representations like the existence of adjacent identical features, the OCC does
not block the application of a rule which creates ill-formed representations. In
(53), a redundancy rule (RR) supplies the unspecified feature value [a voice];
then by the OCC effect, the voicing feature of the head [b voice] becomes
multiply linked to all members of the obstruent cluster. The OCC does not
disallow the application of RR while the RR creates the representation which is
subject to the OCC. It is also true that the OCC, in this case, functions to

prevent ternary oppositions.

 

63

(53)

after OHL by R by OCC
['SOHII-SOHI {-5011} {-5011} {-5011} {-8011}
I I | I I I
C C1 -> C C1 -> C C1

[(1 voice] [or voice] [or voice] [or voice]
If we use privative features, which have only one value present, we need

two types of structural constraints on obstruent-cluster voicing:
(54) the OCC for privative voicing

a ’ b. ‘
{-son] {-son] {-son] {-son]
I I I |
C C1 C Cl
I l
[ voice] [voice]

Constraints (54) a and b will motivate processes (54) c and d, respectively:

( 54) C- d.
{-son] {-son] [—son] {-son]
I | I I
C\ C1 C Cr
‘ - ~ ~ J r:
[ voice] IVORY]

In (54c), the voicing feature of the o-hcad spreads its voicing feature onto the
obstruent(s) in the cluster whereas in (54b), a delinking of all non-head
occurrences of the voicing feature takes place.

Next, the devoicing of word-final obstruents is formulated below. (55a)
is expressed in binary terms and (55b) in privative terms:

(55) Word-Final Obstruent Devoicing

a b.
{-son] {-5011}
I I
C # C
| f
[+voice] -> [-voice] [voice]

We derive all predictable distinctive features by the interaction of
redundancy rules and rules motivated by a language-specific constraint on the
voicing of the representations of obstruent clusters. As mentioned before
(2.2.3). redundancy rules play a crucial role in underspecification theories.
Following the main tenets of RUT and CUT, a set of redundancy rules under

each theory is to be constructed.

I will deal with seven types of voicing assimilation shown below in (56).
We will consider which combination of underspecification theories and feature
systems explains the voicing assimilation phenomena best. First, we consider a
constraint-based approach using privative [voice] and Radical
Underspecification. Using privative features, we will consider their effects on
representations and rule application under RUT, since it requires the simplest
derivations and fewer rules; then comparisons among other combinations of

theories and feature systems will be made.

(56) a /wakzal/ -\ va[gz]al ’station’
b. /b’ez fonet’iki / -> b’e[s f]on’etiki ’without phonetics’
c. /wizg/ -> vi[sk] ’scream’ (gen.sg.)
d. /tr’ezw/ —> tr’e[zf] ’sober’

65

/ot- wzryva/ -> o[d vz]ryva ’ from an explosion’
fo’a wPuska/ -> b’e[s fp]uska ’without permission’
/iz- mt'enska/ -> i[s- mt']enska from Mcensk’
AWOj/ -> [tv]oj ’your,sg.’

Frown

5.2 Privative features and Underspecification

In this section, we will see how a constraint-based approach allows the
privative fcatme [voice] to play a significant role in accounting for Russian
voicing assimilation

The use of the privative feature [voice] requires us to regard assimilation
of both values as two independent operations taking place within the same
domain: devoicing and assimilation The voicelessness assimilation is
expressed by the delinking of [voice] and the voicedness assimilation by the
spreading of [voice]. Besides Lombardi’s (1991) observation, the adoption of
privative [voice] in place of its binary counterpm is strongly supported for the
following reasons.

One consequence is that the me of privative [voice] reduces feature-
changing assimilation processes to non-feature-changing ones. As we have
seen in the previous section, under privative feature [voice], voicing
assimilation of both values is explained by two types of single processes:
voicedness. assimilation by spreading and voicelessness assimilation by
delinking of [voice]. The processes are simpler than when the binary feature
[voice] is used. The privative voicing eliminates the need for a complement
rule supplying [-voice]. Another desirable consequence of the adoption of
privative [voice] is that it makes final devoicing a single delinking operation as

well. Final devoicing, which is found in many unrelated languages like
Sanskrit and Polish, is now expressed by a single delinking process.

The following section presents an approach using privative voicing in
Radical Underspecification since it gives simpler derivation and has no f eature-
changing rules. Because Russian voicing assimilation involves an asymmetrical
segment, RUT turns out to be preferred to CUT. As we go on to look at
privative features and CUT, we will consider problems associated with CUT in
dealing with voicing assimilation of the Russian type and implications for

feature theories.

5.2.1 Privative Voicing and Radical Underspecification Theory

In this section, I present what I believe to be the best solution to Russian
voicing assimilation, using Radical Underspecification and privative voicing
within the authentic autosegmental rules.

My approach is strongly supported for the following reasons. First, by
Radical underspecification of /w/, we can eliminate the need for Hayes’s w
strengthening rule, which involves the delinking of the major class feature
[sonorant]. Second, the use of privative [voice] brings RUT one advantage. It
resolves a problematic aspect of RUT that derivations are complex due to
maximal underlying simplification of segments. Since privative voicing means
that the unmarked value [-voice] is simply absent from the phonology, there is
no need to formulate or apply a complement rule inserting the value; thus, a

simplification of the grammar results.

67

(57) Obstruent-Headlicensing:
License the final member of an obsturent cluster identified in the underlying

representation to be the trigger of voicing assimilation
{—son] {-son]
I I
C (C) C1 (C1 = Obstruent head)

We have to make up a redundancy rule which assigns a segment two
features at once in order to prevent the rule from assigning [voice] to the
underived f. If we have two rules, one assigning {-son], the other assigning
[voice], the later rule derives not only v from /w/ but also v from If/ since the
difference between f and w is only the underlying presence or absence of [-

son]. Undesirable redundancy rules and derivations are provided below:

(58) a[ ]-> [-son]/[_,+cont, +labial]

b. [ ] -> [voice] / [ _ , -son, +cont, +labial]

(59) a. /W/ -> [v]
UR SR

byRRa RRb
{-son] {-son]
I |
C -> C -> C
|

[voice]

68

b. If/ -> [v]

[voice]

Consequently, we need redundancy rules as follows:
(60) Redundancy Rules
a. [ ] -> [-son,voice]/[_,+labial, +cont]
b. [ ] -> [+son]
c. [ ] -> [voice]/[_,+son]

The adoption of privative features leads to the elimination of the
complement rules. That is, once we adopt privative [voice], we no longer need
a complement rule for the voicing feature, a redundancy rule assigning the

value opposite to the underlyingly specified value, [voice] at point.

(61) Word-Final Obstruent Devoicing
{-50 n]
I
C #-

1:

[voice]

69

(62) Effects of Obstruent-Cluster Constraint (OCC)

a {-son] {-son] b. {-son] {-son]
I I
C: C1 C C1
‘ \ \ \I
[ voice] [ voice]

(Condition: persistent application; propagating within a given o-cluster)

(63) Derivations: ( [s] = sonorant; [v] =voice)
a. /wak_zal/ —> va[gz]a1 ’station’
UR SR

by OHL by OCC a

{-51 {-51 {-51 I-SI {-51 {-51
I I I I I I
C C —> C C, -> C C,

l |
[VI IV] [V]

After OHL, the voicing feature of the trigger spreads. Unlike binary voicing, it
is a spreading process, delinking of the target’s feature is not needed.

b. /b’ez fonet’iki/-> b’e[s f]on’etiki ’without phonetics’
UR
byOHL byOCC b

{-5} I-Sl I-Sl {-5} {-51 {-5}

I I | | | I

C C -> C C1 --\" C C1

I I

[V] [V]

This is the delinking of the [v], motivated by the structural constraint OCC b.
Assoon asOHL licenses the final member as the trigger, the OCCb-motivated
rule fixes the ill-formed representation.

70

c. /wizg/ -> vi[sk] ’scream’(gensg.)

UR SR
by OHL by FD by OCCa

[SI [SI [SI [-SI {-51 [SI [SI {-5}

I I I I | | I I

C C# -> C C1# ->C C1# -> C C1#

I | | I |

[V] [V] [VI [V] [VI

Here, unlike what is needed under binary voicing, the final devoicing is simply
a delinking process. The application of FD creates an ill-formed
representation By OCCa, the voicing feature of the target is delinked

d /tr’ezw/ -> tr’e[zf] ’sober’
UR SR
OHL n/a

byRR byFD

I-SI I-SII-SI I-SII-Sl
I | | I |
C C -> C C -> C C
I | | I
[VI [V] IV] [V]

OHL does not apply because /zw/ does not constitute an obstruent cluster.
Following the ROC, redundancy rule a supplies features and provides an
environment for the final devoicing.

e./ot-wzryva/ -> o[d vz]ryva ’fromanexplosion’
UR SR
byOHL byRRa byOCCa
I-Sl I-Sl [-SI {-5} {-51 [~51 {-51 {-5} {-5} {-5}
I I | I I I | | I I
CCC->CCC1->CCC1->CCC,
I l | l
M [V] [V] [V] [VI

71

Under privative voicing, there is no need to insert [-voice] to underlyingly
unspecified segments by Rs. RRa becomes activated prior to the application
of OCC, which make reference to features [-son] and [voice]. OCCa motivates
the spreading of [v] onto the targets.

f. /b’ez wpuska/ -> b’e[s fp]uska ’withoutpermission’
UR SR
by OHL by R a by OCCb

{-SI I-s] I-SI r-sr [SI {-51 [SI {-51 {-5} {-s1
| I I I l I I | l I
CCC1->CCC1->CCC1->CCC1
I I | I
IV] [V] [v] [v]

After OHL, RRa supplies /w/ features necessary to be an obstruent to undergo
voicing assimilation. Since the ill-formed representation is due to the presence
of [voice] in the targets, OCCb applies. In the case of voicelessness
assimilation, we have to simply delink the voicing feature of all the targets.

g. /iz- mt'enska/ -> i[s mt']enska ’from Mcensk’

UR SR
by OHL by OCCa by RRb &c

{-51 {-5} {-51 {-51 {-51 {-51 [51 [+51 {-51

I | I I I I I l I

CCC->CCC1-> CCC1->CCC1

| I I

IV] IV] [V]

This is the voicelessness assimilation across a nasal. The delinking of [v] takes
place ignoring the nasal, which is transparent to the process. On the surface.
the nasal must be realized as [+sonorant, voice], which has no effect on the
voicing of the o-cluster.

72

h. ftwoj/ ->[tv]oj ’your,sg.’

UR SR
OHLn/a; by RRa
{-51 [SI {-5}
I | |
C C -> C C

|

[V]

This case does not involve voicing assimilation since /w/ underlying form of v
is not subject to the OHL and thus these segments do not create an o-cluster.

To summarize, the approach presented above has several desirable
consequences. First, the use of privative voicing yields the simplest way to
explain the phenomena It results in the simplification of assimilation processes
as well as that of the derivations under RUT as a whole. In this analysis, a
complement rule which supplies [-Voice] to obstruents is eliminated due to the
use of privative [voice]. So, RUT with privative voicing requires fewer
redundancy rules than binary voicing while keeping maximally unspecified
underlying representations. As an effect of Radical underspecification of /W/,
w strengthening is unnecessary.

Second, no feature-changing rule is required: each rule involves one
process. Voicing assimilation processes are expressed by either one of the two
non-feature-changing operations, delinking and spreading.

Third and more importantly, these processes are motivated by a
language-specific constraint on obstruent-cluster voicing. The constraint
permits the privative voicing in assimilation of both values. The desirable

consequences mentioned above are therefore due to this type of approach

73

which allows privative features to play a role in expressing assimilation of both
values.

Finally, since the two mechanisms independently apply to different
contexts in a given obstruent cluster, it is fair to say that Russian voicing
assimilation is of two types: devoicing and spreading. The devoicing explains
voicelessness assimilation and the word-final obstruent devoicing. The

spreading takes care of voicedness assimilation of obstruents.

5.2.2 Radical Underspecification and Hayes’s prosodic rule

The preceding section presented an approach using constraint, Radical
underspecification, and privative voicing. It has shown that a rule which
appeals to structural conditions of obstruent-cluster voicing allows privative
[voice] to play a role in assimilation of both values. In this section, I favor this

constraint-based approach over a spreading analysis.

At the end of the previous chapter, we discussed that Hayes’s prosodic
rule (p54-55) can be simplified by constraining the rule to refer to the value of
the feature [sonorant] of targets. This allows us to eliminate the sonorant
revoicing rule since if the rule refers to {-son] of targets, it can skip [+son]
consonants, affecting only obstruents. We may ref orrnulate his rule below:
(64) Voicing Assimilation

In the configuration {-son], reassociate [or voice]

[or voice]

 

74

with the maximal string of obstruents to the left, deleting any autosegments
stranded in the process.

(Condition: Persistent application)

It is stated above that by underspecifying the labio-dental fricative v for
[sonorant] and assigning {-son] by a default rule, we can eliminate the ordered
rule of w strengthening as well.

Some necessary rules and constraints for underspecification are given

below:

(65) Obstruent-Head Licensing (OHL):
License the final member of an obstruent cluster identified in the underlying

representation to be the trigger of voicing assimilation
[-son] [-son]

C (C) C1 (C1= Obstruenthead)
(66) Redundancy Rules
a. [ ] -> [-voice]/[_,-son]
b. [ ] -> [-son]/[ , +labial, +cont]
c. [ ] -> [+voice]/[_,-son.+labial,+cont]

d I I -> [+son]

e. [ ] ->[+voice]/[ _,+son]

(The order is irrelevant to rule application; rules apply following RROC
(22.3))

As for the application of rules a and c, there is no possibility that rule a applies

to the context to which rule c also applies since more Specific rule c takes

75

precedence wherever both rules can apply (Kiparsky 1982, 1985. Hudson,
personal communication).

(67) Word-Final Obstruent Devoicing
a. {-son]

I
C#

[+voice] -> [-voice]

The final devoicing under the binary feature system must be a feature-changing
process, the delinking of [+voice] and the insertion of [-voice] since the final
devoicing triggers voicing assimilation by the spreading of [voice] as we have
seen in chapter three.

(68) derivation:

a /wakzal/ -> va[gz]al ’station’
UR SR
by OHL by RRa by VA
{-51 [IS] {-51 Il-SI I-Sl {-5} [=7] l-sll
I I | I

CC->CC1->CC,->CC1

I I I I \I

[+Vl [+Vl I-Vl I+Vl [+Vl

First, the OHL identifies /2/ as a trigger and licenses it; then the RRa supplies
the unmarked feature {-voice]; finally the voicing assimilation takes place,
deleting the association of [-v] to the target and spreading the feature of the
trigger [+v] onto the target. Both values of voicing feature must be present in
order for OCC to apply.

76

b. /wizg/ -> vi[sk] ’scream' (gensg.)
UR
by OHL by FD by VA
{-5} {-8} {-8} {-8} {-5} {-5} {-5} {-5}
I I | I I I I I
C C# -> C C1# -> C C1# -> C C1
I I I I I |

[+Vl I+VI I+Vl I+Vl [W] H] H]

After the OHL. [+v] of the trigger is delinked and instead [-v] is inserted by the
final devoicing; by VA the feature of the target is delinked and that of the
trigger spreads onto the target.

c. /b’ez-wpuska/ -> b’e[s fp]uska 'without permission‘
UR SR
by OHL by RRa,b,c by VA
{-5} I-Sl {-8} {-5} I-SlI-SII-Sl I-Sll-SII-Sl
| I I I | I I I | I
CCC->CCC1 ->CCC1->CCC1
I I I I I
[+Vl I+Vl I-V] [+Vl H] W]

This is acase of spreading of the negative value. RRs motivated by the ROC
apply supplying all the feature values relevant to VA. Here. the asymmetrical
segment becomes v. By VA, all the targets lose the voicing feature by
delinking and receive [-v] from the trigger by spreading.

d fiz-mt’enska/ -> i[s mtflenska ‘from Mcensk’
UR SR
by OHL byRRa &d by VA by R &TC
{-51 {-5} {-5} {-51 {-5} [+5] I-Sl I-Sl [+5] [-S] {-5} [+5] {-8}
I I I I I I I I I I I I I
CCC->CCC1->CCC1->CCCt->CCC1
I I I I

[+Vl I+Vl I+Vl H] W] H] I+Vl I-Vl

77

This is voicelessness assimilation across a sonorant m. The voicing assimilation
takes place across the sonorant m, which is transparent to the process, by
delinking of the voicing feature of the target and spreading that of the target.
Since the intervening sonorant must be realized as [+voice] on surface, the
multiple-linking of the [-v] to the obstruents is split by tier-conflation.

My approach presented in the previous section is preferred to this
spreading analysis with RUT for the following reasons.

First, under a spreading analysis like l-layes’s rule, the assimilation
feature must be binary to be expressed in a single spreading rule: the spreading
of an assimilation feature value of the trigger onto the target(s), regardless of
the value. Thus, rules entail representations. Privative features cannot work
here since the spreading of voicelessness, which is expressed by the absence of
the feature [voice], cannot be expressed It is necessarily a feature-changing
rule even though we adopt RUT, which does not specify [-voice] for obstruents
underlyingly, we have feature-changing processes as shown in (68) since the
RROC requires both values to be present prior to the application of voicing
assimilation. Any rule of voicing assimilation under binary features is
necessarily a feature-changing operation, changing [or voice] to {-01 voice]. A
rule which expresses assimilation by changing a feature is highly disfavored
(Hayes 1986).

Second, a single spreading analysis requires the devoicing of word-final
obstruents to be a two-step process: a delinking of [+voice] and insertion of
[-voice] (Mester & Ito 1989). If the devoicing of word-final obstruent is the

delinking of [+voice] and the insertion of [-voice], then, as Lombardi argues

78

(see 2.3.4). laryngeal neutralization phenomena found in many unrelated
languages remain unexplained

On the other hand. as we have seen in the preceding section. those
problems do not arise under an approach with structural constraints since it is

compatible with the privative feature system.

in the following section, providing derivations using Contrastive
Underspecification and privative features. I will show why Radical

Underspecification is favored over Contrastive in the Russian case.

5.2.3 Privative Features and the principles of Contrastive Underspecification

The last two sections have shown that privative voicing gives simpler
derivations than its binary counterpart does. It is also shown that Radical
Underspecification provides a natural account for the asymmetrical behavior
of the voiced labio-dental fricative v. In order for us to be really convinced of
this, we may have to see the reason why a privative voicing analysis with
Contrastive Underspecification is less favored Contrastive Underspecification
does not work nicely in explaining the asymmetry of v because its principle
does not allow underspecification of a distinctive feature, which must be
present and because the contextual feature for Russian voicing assimilation is
the major class feature [sonorant], which cannot be delinked

In this section. we consider a problem associated with CUT in using
binary features. According to CUT, contrastive features must be specified for
both values underlyingly. So the features to be specified for both values are the

features voicing and sonorant. Such underlying specification of the feature

79

sonorant for both values brings us a problem in accounting for the
asymmetrical segment v. If the feature sonorant is specified for both values,
then the underlying form of the voiced labio-dental fricative v must also be
specifiedas [+sonorant], which is to be changed to {-sonorant] in order for it to
undergo voicing assimilation. This is problematic and forces us to instead

adopt [obstruent] as a privative feature.

As far as binary features are concerned, under RUT. only the marked
feature [+voice] can be underlyingly specified for obstruents while CUT
requires both values to be specified underlyingly. Thus, it necessarily involves
feature-changing processes, delinking one value of the feature and inserting
another.

Privative features play a significant role especially in the case of
voicelessness assimilation. Once we adopt the privative feature [voice], there is
no way to indicate [-voice] other than simply to have it left absent. Thus, RUT
and CUT have the same underlying specification for the voicing feature and
they have it throughout the derivation. The major difference is the underlying
specification of [+son]: RUT does not specify it while CUT specifies it.

It is the underlying specification of [+son] that brings us a problem. That
is. when both values for the feature [sonorant] are specified underlyingly. v as a
glide must also be specified as [+son]. Consequently. we need a rule changing
w to v as [+son] -> {-son] before voicing assimilation applies. We need a more
complicated rule than under the principles of RUT. The derivation involving

the voiced labio-dental v is given below:

80

(69) Redundancy Rule:

a [+son] -> {-son, voice] / [ _, +cont, +lab]
(70) lb’ez wpuska/ -> b’e[s fp]uska ’without permission’
UR SR
by OHL by RRa by OCCb

{-5} [rs] I-sI I-s][+sl I-sI I-sl {-slI-sl I-sI I-sI I-sI
I I I I I I I I I I I I
ccc .>ccc1 -r» ccc, .>~cccl

I I I I
[VI [V] [V] [VI

As shown above, it is necessary to change the specification of the labio-
dental fricative from [+son] to [soul so that voicing assimilation affects the
segment. It is a feature-changing rule, the delinking of [+son] and the insertion
of {-son] by stipulation. Since redundancy rules are by definition all fill-in
rules. we have to retain the rule as a phonological one, which is no different
from w strengthening. Even if we regard it as a phonological rule, we carmot
have it because the major class feature [sonorant] can never undergo delinking
or spreading (see p13). To avoid this, we have two alternatives available. One
way is to specify the voiced labio-dental fricative v as obstruent {-sonorant]
underlyingly. This cannot be accepted since it incorrectly motivates v to
trigger voicing assimilation. Another way is to adopt privative [obstruent]
instead of the binary-valued [sonorant]. The privative feature [obstruent] is
used below while we have at present no evidence to assume that [obstruent] is

privative. The feature [+sonorant] in binary terms is then expressed by the

 

81

absence of [obstruent]. I will show the effects of the privative features [voice]

and [obstruent] on CUT below.

(71) Obstruent-Head Licensing (OHL)
License the final member of an obstruent cluster identified in the underlying

representation to be the trigger of voicing assimilation.

[obstr] [obstr]

I |
C (C) C, ([obstr]=obstruent; C1= Obstruent head)

(72) Redundancy Rules:
a. [ ] -T> [obstr,voice]/[_,+labial,+ cont]
b- I I ->'IV0iC€I/[.__I
(The RROC does not hold in this theory ( see p23))

(73) Word-Final Obstruent Devoicing (FD)
[obstr]

I
C #

2]:

[voice]

(74) Effects of ObstruentCluster Constraint (OCC)

a. [obstr] [obstr] b. [obstr] [obstr]
I I I
C: ‘ C] C Cl
‘ ‘ ~ \ I it
[ voice] [ voice]

(Condition: Persistent application; propagating within a given o-cluster)

82
(75) Derivations:

a. /ot-wzriva/ -> o[d vz]riva ’fromanexplosion’
UR SR

by OHL byRRa byOCCa

[oI Io] Io} Io] [O] I0] [0] I0] I0] I01
I I I l I l I I I I
CCC C CC1 —>CCCl —>CCCl
I I I I W

[V] IV] [V] [v] [v

L—J

After OHL, redundancy applies; then OCCa spreads [v] onto the targets.

b. lb’ez wpuska/ -> b’e[s fp]uska ‘without permission’
UR SR

byOHL byRRa byOCCb
[0] [OI [0] [0] [GI [OI [OI [0] IOIIOI
I I I | I | I I I I
CCC -> C CC] ->CCC1 ->CCC1
I I | I
[V] [VI [VIIVl

After OHL, RRa , then OCCb fixes the ill-formed representation by deleting all
the targets’ voicing features.

c. /iz- mt'enska/ -> i[s rnt']enska ’from Mcensk’
UR SR
by OHL byOCC b by RRb
[OI [0] I0] [0] [0] [0] [0] I0]

I I I I I I I I
CCC—>ccc1 ->ccc, -> CCC
I I I
IV] [V] [v]

83

Assimilation is taking place across the sonorant m. Ignoring the sonorant,
OCCb deletes the voicing feature of the target. RRb supplies the sonorant the
voicing feature.

d /twoj/->- [tv]oj

UR SR
OHLn/a;by RRa by RRb

[OI [OI [O] [O] [O]

I I | | |

C C C C -?> C C

While the problematic aspect associated with the principles of CUT might
be resolved by the use of the privative feature [obstruent] in addition to [voice],
again we have to know if [obstruent] is really a privative feature. If [obstruent]
is not a privative feature, then we might have to ask whether [sonorant] is
privative. There are still some difficulties for CUT. Under a version of CUT,
only redundant feature values can be candidates for a single-valued features.
As the theory implicitly treats the feature sonorant as a distinctive feature
(Christdas I988), which cannot be single-valued (p26) and must be specified

for both values under a binary feature system.

I have shown that the asymmetry of Russian v can be best explained by
means of Radical underspecification with privative voicing provided that the
contextual feature sonorant is binary. The constraint-based approach with
privative voicing gives the simplest and natural account for assimilation of both

values in Russian.

Contrastive Underspecification cannot work nicely here since it does not
allow the segment to be unspecified for the contextual feature sonorant, which
is distinctive. Even though the privative [obstruent] appears to resolve the
problem, the privativity of the feature obstruent is not yet justified.

The approach presented here, which appeals to structural conditions of
the trigger and targets of assimilation, allows privative voicing even in the case
of Russian. They are conditions on representations of obstruent-cluster
voicing: a wellformed representation in an o-cluster whose head is voiced, and
the other well-formed condition in an o-cluster whose head is voiceless. These
two constraints on representations motivate either one of the two non-feature-
changing operations: spreading or delinking.

The adoption of the privative feature [voice] reduces the assimilation
processes and redundancy rules. Assimilation of both values is expressed by
either one of two independent mechanisms, delinking and spreading. This also
means that assimilation under a privative feature system (not under a binary
one) is really two different processes: voicedness assimilation is an
assimilation process (spreading) and voicelessness assimilation is actually a
devoicing process (delinking). Assimilation is interpreted as the spreading of
the assimilation feature and devoicing as the delinking of it. The voicing
assimilation of both values becomes a pair of non-feature-changing processes.
The final devoicing also becomes a single operation, delinking of [voice]. We
no longer need the insertion of [voice] word-finally. The two universal
mechanisms account for the assimilation of both values and the word-final
obstruentdevoicing. Voicelessness assimilation and word-final devoicing now

become the same rule type. The delinking analysis of Voicelessness assimilation

85

and final devoicing explains the Sanskrit examples where all devoiced
obstruents either by laryngeal neutralization or by voicelessness assimilation
result in plain voiceless (p33-35).

O-Head Licensing relies on the assumption that v is underlyingly neither
obstruent nor sonorant but becomes the obstruent v by a redundancy rule
during the derivation. This is possible only under the framework presented
here. Only Radical Underspecification treats v as featureless with respect to

the relevant features for voicing assimilation.

6. Conclusion

We have seen that three modular theories, i.e., underspecification,
autosegmental representation, and the privative feature system, function
together to give a uniform account for the phenomena of Russian voicing
assimilation. The autosegmental framework, which allows an assimilation rule
to affect only segments specified for particular features like {-son], eliminates
the need for the rule of sonorant revoicing assumed in a previous analysis
(Hayes 1984). Underspecification together with redundancy rules eliminates a
need for the w strengthening rule and extrinsic rule-ordering. Finally, the use
of privative [voice] has several advantages in explaining the assimilation
processes. When privative voicing is adopted, there are only non-feature -
changing processes: delinking and spreading of [voice]. This is probably true
of any feature if it is privative.

Persistent application of OCC-motivated rules interacts with redundancy
rules without extrinsic rule-ordering. Since, by the effect of the ROC,
redundancy rules interact with phonological rules, we may say that the
approach presented accounts for Russian voicing assimilation without extrinsic
rule-ordering.

As for O-head Licensing, it is assumed to apply prior to any application
of rules, both redundancy and phonological, providing an environment for
voicing assimilation. OHL is neither a phonological nor a redundancy rule. It
does not change or affect linkings of features or supply missing values. It is
more like a syllabification process. Yet, OHL differs from syllabification in
that it does not apply persistently or change structural descriptions. With

respect to non-persistent application of OHL, it refers to the CV -tier and the
contextual feature value [sonorant]. Perhaps, the fact that OHL licenses an
entity in the CV -tier, which is by principle not subject to any phonological

processes like delinking or spreading, explains its special property.

A rule which expresses an assimilation by spreading any feature value
whatsoever is a rule merely using features. In non-linear phonology, which is
a theory of representation rather than one of phonological rules, the right rule
must be assumed to follow from the right represenation of segments or more
specifically, that of features, i.e, the n-arity of a given feature, instead of the
case that the n-arity of a feature follows from a rule. We have seen that a single
spreading analysis is not in accordance with the evidence for the privative
laryngeal features provided by Lombardi (1991). Given her evidence for
privative voicing and the Russian data, a constraint-based approach is favored
over a spreading one.

The objection that privative features cannot explain assimilation of both
values is eliminated by a constraint-based approach. The structural constraint
on obstruent-cluster voicing (OCC) allows the use of the privative [Voice] even
in the case of Russian voicing assimilation. The constraint presented here
makes reference to features and their associations to both the trigger and
targets in its structural description. This differs from any spreading analysis,
which makes reference to the assimilation feature and its association only to the
trigger.

Accordingly, the present approach predicts the relationship between the

n-arity of assimilation features and assimilation rules: if a feature is privative,

88

then assimilation of both values of the feature is asymmetrical and needs tWo
rules. If it is binary, then assimilation is expressed symmetrically by a single
spreading rule.

I have provided an analysis of Russian voicing assimilation, supporting
the position of [to & Mester (1989) and Lombardi (1991) that the voicing
feature is privative. Both values of a feature like voicing can be adequately
expressed by the opposition of presence or absence of the feature when a rule is
expressed to refer to structural conditions. Logically, we do not need binary
features to express binary oppositions. Of course this by no means supports the
position of the Strong Privativity (den Dikken & van der Hulst 1988) that all
features can be reduced to privative features. Some features by nature may not
be candidates for privative status. In fact, if the feature [sonorant] is used as
privative, any theory fails to explain the asymmetry of the Russian labio-dental
fricative v without requiring a phonological rule changing the values of the
major class feature [sonorant]. The question of whether a feature is privative
or binary depends on the nature of each feature. We might have to ask whether
the nature of features is language universal or language specific. Whatever the
n-arity of a feature, a constraint-based approach is shown to be compatible with
single valued features as well as binary ones; thus a selection of a feature system

follows not from a rule but from the nature of a feature.

 

LIST OF REFERENCES

List of References

Abaglo, Poovi & Diana Archangeli. 1989. Language Particular
Underspecification: Gengbe /e/ and Yoruba /i/. Lingo/3112' Inquiry
20 457-480.

Andersen, Henning. 1969. The phonological status of the Russian ’labial
fricatives’. Joumal of Lbrgwfsa'cs 1965. 121-127

Archangeli, Diana. 1984. Yawelmam‘Pfiono/ogy andMozp/Iologx Ph.D.
Dissertation, MIT , published in New York: Garland

. 1988. Aspects of Underspecification Theory. Archangeli
ed. Pizarro/cg}! .f 183-207.

& Doug Pulleyblank. 1987. Maximal and Minimal Rules:
Effects of Tier Scansion. MEZS 17, 16-35.

& . 1989. Yoruba Vowel Harmony.
[.z‘rrgwistzt‘ Inga/23’2”. 173-217.

Barkai, Malachi & Julia Horvath. 1978. Voicing Assimilation and the
Sonority Hierarchy: Evidence from Russian, Hebrew, and Hungarian.
bigwsa'cs. 212. 77-88.

Chomsky, Noam and Morris Halle. 1968. Hie SoundPattem of Eng/1's): New
York: Harper & Row.

89

90

Cho, Young-mee Yu. 1990. A Typology of Voicing Assimilation. WCCFZ 9.
141-155.

Christdas, Prathima. 1988. Phonology and morphology of Tamil. Ph.D.
Dissertation, Cornell University.

Clements, George N. 1985. The Geometry of phonological features.
Poona/oat Yearbook 2

. & Samuel J. Keyser. 1983. CVP/Iono/ogtc A Generative
7%on oft/1e Std/able Cambridge, MA: MIT Press.

 

den Dikken, Marcel & Harry van der Hulst. 1988. Segmental
Hierarchitechture. in van der Hulst & Smith, eds. Farm/es, Stag/Irena?!
Structure and Harmony Processes Vol. 1. Foris.1-78.

Ewen, CJ. & H.G. van der Hulst. 1985. Single-valued features and the
nonlinear analysis of vowel harmony. In H.Bennis & F .H.Beukema, eds.
Linguistics in tfieNetfier/ands. 1985. Dordrecht: Foris. 39-48.

Goldsmith, John A. 1976. Autosegmenfi/ Phonology. Ph.D. Dissertation,
MIT. Distributed by Indiana University Linguistic Club.

. I985. Vowel harmony in Khalkha Mongolian, Yaka,
Finnish and Hungarian. P11011010“. ’ Yearbook 2 251-74.

 

. 1990. A utosegmenaa/ & Metfica/ P/Iono/ogjz' Basil
Blackwell, Oxford & Cambridge: MA.

Gorecka, Alicja. 1989. P/Ionolog , ofArticzz/atron Ph.D. Dissertation, MIT.

Grebnev. 1959. Foneticeskoe i morfologiciskoe- oformlenie abbreviatur v
russkom jazyke. Bar/lea?) t/yzsoke ska/y raster/Io Jazfl'a a lite/army 3
5-22.

Gussmann, Edmund 1992. Resyllabification and Delinking: The Case of
Polish Voicing. Lingo/Isa? Inga/13V 23. 29-56.

9]

Halle, Morris. 1959. 771e Sound Pattern ofRumZan The Hague: Mounton.
. 1973. The accentuation of Russian words, Lg. 49. 312-348.

Halle, Morris & .I-R Vergnaud. 1980. Three dimensional phonology. Journal
of Lzrgwsabkesearcfi 1. 83-105.

.1981. Harmony Processes. In W.Klein &
W.Levelt (eds) 0055111g I/Ie boundanes 1'11 111100111306 Dorencht: Reidel.
1-22.

 

. 1987. Ajzfssat'on Stress Cambridge, MA:

 

MIT Press.

Hayes, Bruce. 1984. The phonetics and phonology of Russian voicing
assimilation. In Amoff & Oehrle (eds) Language Sou/1d Palfems 1984.
318-328.

. 1986. Assimilationas Spreading in Toba Batak. Lrizgwsffc
111911113” 12 467-499.

. 1988. Dipthongization and Coindexing. ms. Los Angeles:
UCLA.

Hudson, Grover. 1974. The Role of SPCs in Natural Generative Phonology.
Papers from Parasession on Natural Phonology. Chicago. Chicago
Linguistic Society 10. 171 -183.

lakobson, Roman. 1971. K voprosu o gluxosti i zvonksti russkix scelinnyix
gubnyx. Selected Wh‘tz‘ngs: The Hague-Paris, 728-733.

. 1978. Mutual assimilation of Russian voiced and voiceless
consonants. Stud/29 L112g1113I1'c’27 32. 107-1 10.

Jones, Daniel & Dennis Ward 1969. 771eP/1011et1'cs ofkussfan Cambridge
University Press: Cambridge.

92

Kenstowicz, Michael & Charles Kisseberth. 1979. Introduction to Generative
Phonology. Academic Press: Chicago.

Kiparsky, Paul. 1982. Lexical Phonology and Morphology. Linguistics 111 the
Morning Calm.edbyl.-S. Yang, 3-91. Hanshin,SeouI:Linguistic
Society of Korea.

. 1985. Some Consequences of Lexical Phonology. Phorzo/ogv
fearhooh 2. 82-138.

 

. 1991. In Defense of the number two. paper presented at the
Summer Institute of Santa Cruz.

 

Ladefoged, Peter. 1982. A Course in Phonetics. Harcourt Brace .Iovanovich,
Publishers. 34-38.

Lightner, T. M. 1963. A Note on the Formulation of Phonological Rules.
Quarterb-t Boyer: Report of the Research Laboratory offleezrom'cs 6'3
Cambridge, MA: MIT.

. 1965. Seer/11mm! Phonology of modem standard £15512”.
PILD. Dissertation, MITT.

 

Lombardi, Linda. 1991. Laryngeal Features and Privativity. ms. LSA
Summer Institute of Santa Cruz. '

McCarthy, John. 1984. Theoretical consequences of Montanes vowel harmony.
LJhgwsh'c 111411123" 1 5. 291 318.

. 1986. OCP Effects: Gemination and Antigemination.
Lihgwsak‘ hlquhy 17. 207-263.

 

.1988. Feature Geometry and Dependency: A Review.
Phonetica 45. 2-4. 84-108.

& Alan Prince. 1986. Hvsoah'c Momho/ogtr ms.

 

 

 

93

Mester, R. Armin. 1986. Studies in Tier: Sat/cane Ph.D. Dissertation,
UMASS

. 1988. Dependent Tier Ordering and the OCP. in van der
Hulst & Smith eds. Featznes, Segmean Senetnres and Harmon V
Hocesses Vol. 2. Foris. 127-144.

 

& Junko Ito. 1989. Feature Predictability and
Underspecification: Palatal Prosody in Japanese Minetics. Language 65
258-293.

 

Myers, Scott. 1991. Persistent Rules. Linguistic lnqniry 22. 315-344.

Pulleyblank, Douglas. 1983. Tone in Lenka] Phonology: PhD.
Dissertation, MIT. Published, 1986. Dordrecht: Reidel.

.1986. Underspecification and Low Vowel Harmony in
Okpe. Studies in African Linguistics 17.2. 1 19- 153.

 

.1988. Vocalic Underspecification in Yoruba. Linguistic:
lnqniry 19. 233-270.

 

Reformackij, AA. 1971. [zf] iz <zv>. Fonologidieshre Etjnay' Moscow:
Nauka. 129-133.

Shapiro, Michael. 1966.. On non-distinctive voicing in Russian. Journal of
Linguistics ll 189-194.

Stanley, R. 1967. RedundancyRules in Phonology. Language 41. 294-302.

Steriade, Do nca. 1984. Geel'pr'osoo’ies and the manner ofst’llebriiarnon
Ph.D. Dissertation, MIT.

. 1987. Redundant Values. In A Bosch, B. Need & E. Schiller
feds.) Papers 1129171 the parasession on autosegmental andrnem'ca/
phonology Chicago: Chicago Linguistic Society. 339-362.

 

94

Sugimoto, Takayo. 1992 Gemination and qoenthefie vo we]: in Japanese
ms. Michigan State University.

Trubetzkoy, N. 1939. Grundzuge derphono/ogze TCLP 7. English version,
(translation) C.Baltaxe, 1969. H7nczb/es ofphono/ogu Berkeley, Univ.
of California Press.

Vago, Robert M. 1988. Underspecification in the height harmony system of
Pasiego. Phonologyj 342-362.

.q,__,