Review of Vowel Harmony and Correspondence Theory
|Date: Mon, 19 Apr 2004 13:23:29 -0700
From: Eric Bakovic
Subject: Vowel Harmony and Correspondence Theory
AUTHOR: Krämer, Martin
TITLE: Vowel Harmony and Correspondence Theory
SERIES: Studies in Generative Grammar 66
PUBLISHER: Mouton de Gruyter
Eric Bakovic, University of California, San Diego
The topic of this book is the analysis of various aspects of
vowel harmony within Optimality Theory (OT; Prince &
Smolensky 1993/2002), with particular attention to the
Correspondence Theory of faithfulness (McCarthy & Prince
1995, 1999). The empirical and analytical issues addressed
in the book are all ones that have received substantial
attention in the vowel harmony literature regardless of
theoretical orientation, but the author's overall approach
differs in many interesting ways from previous work.
The book is divided into two parts. Part I, "The phenomenon
and the theoretical background", contains three chapters.
Chapter 1 introduces various aspects of vowel harmony,
highlighting those that are the focus of the book (see the
summary of Part II below). Chapter 2 introduces OT and the
particular types of constraints to be crucially employed in
the remainder of the book: positional faithfulness,
syntagmatic identity, and (three types of) "coordinated"
constraints. Chapter 3 is a preview of how specific examples
of some of the empirical phenomena introduced in Chapter 1
will be analyzed in Part II of the book with the constraint
types introduced in Chapter 2.
Part II, "Case Studies", also contains three chapters plus a
general concluding chapter. Each of the substantive chapters
addresses a coherent cluster of empirical phenomena, with
detailed analyses of examples from a variety of languages.
Chapter 4 addresses directionality and its (hypothesized)
morphological underpinnings. Chapter 5 addresses vowel
transparency. Chapter 6 addresses the problem of vowels that
are specified for a harmonic feature value that is the
opposite of the value that they impose on other vowels. (The
author dubs these vowels 'Trojan' vowels, a useful moniker
that I will adopt in this review.) Parasitic harmony -- a
situation in which harmonizing vowels must already share
some feature -- is also discussed in Chapter 6, as a part of
the extensive analysis of the well-known complex interaction
of vowel harmony and other processes in Yawelmani Yokuts.
The final, concluding chapter summarizes the main points of
the book and comments on four major theoretical themes:
serialism, underspecification, factorial typology, and
functional motivation. In addition to endnotes, a
bibliography and an index, there is a useful appendix of
proposed constraints and their definitions as well as an
appendix of languages analyzed in the book (with details
regarding properties relevant to vowel harmony).
Apart from a few relatively minor changes, additions and
omissions, this book is essentially identical to the
author's 2001 doctoral dissertation of the same title.
There are four main empirical phenomena that the author
chooses to focus on in this book, and which I will also
largely restrict my attention to in this review:
Vowel harmony originates in a particular vowel (e.g., a
vowel in the root) and proceeds outward from there.
2. Transparent vowels.
Vowel harmony may ignore certain vowels, skipping over
them and affecting vowels on the other side.
3. Trojan vowels.
Certain vowels may condition harmony for the opposite
value of the harmonic feature with which they surface.
4. Parasitic harmony.
Neighboring vowels may harmonize only if they already
agree in terms of another, attendant feature.
According to the abstract (pp. ix-x), the book's "central
goal is to give a unified account" of these four aspects of
vowel harmony (and some others) within OT. This goal is
explicitly contrasted with previous work on vowel harmony in
OT, in which the author claims that "a rich inventory of
theoretical devices has been applied and developed to
explain various aspects of vowel harmony". Briefly, the
analyses proposed for the four phenomena above are:
1. Root-to-affix directionality is due to a distinction
between root and affix faithfulness constraints (Beckman
1998), with pairs of such constraints ranked in such a way
that root feature values are preserved and extended at the
expense of affix feature values. Dominant-recessive harmony
is due to local conjunction (LC, violated only when both
conjuncts are violated; Smolensky 1993, 1995) of markedness
and faithfulness constraints, as proposed by Bakovic (2000).
Other instances of directionality (morpheme-internal,
affix-to-root) are accounted for with other types of
positional faithfulness constraints, some of which are
understood as logical constraint conjunction (violated when
either conjunct is violated; Crowhurst & Hewitt 1997). (The
author attempts to demonstrate on pp. 85-86 that other
positional faithfulness constraints might be understood as
cases of a third type of constraint coordination; the
example discussed, however, is unrelated to vowel harmony
and the matter is not discussed anywhere else in the book.)
2. Vowel transparency is due to an LC of two constraints:
one demanding that a vowel agree with neighboring vowels in
terms of the harmonic feature, and another demanding that a
vowel disagree with neighboring vowels in terms of the
harmonic feature. The effect is supposed to be that a vowel
unable to agree with the harmony trigger (for markedness
reasons) will prefer to disagree with both neighbors (the
hallmark of transparency) than to initiate its own harmonic
domain (the hallmark of vowel opacity).
3. Trojan vowels are the result of an LC of two constraints:
one demanding output faithfulness to the input harmonic
feature value, and another demanding that a vowel disagree
with neighboring vowels in terms of the harmonic feature.
The effect is supposed to be that a vowel in a harmony-
triggering position that is unable to surface with its
underlying harmonic feature value (for markedness reasons)
will prefer to disagree with its neighbors, the apparent
result being that the vowel triggers harmony with its
underlying rather than surface harmonic feature value.
4. Parasitic harmony is the result of an LC of two
constraints: one demanding that a vowel agree with
neighboring vowels in terms of the harmonic feature and
another demanding that a vowel disagree with neighboring
vowels in terms of the attendant feature on which harmony is
parasitic. The effect is supposed to be that a vowel can
*either* disagree with its neighbors in terms of the
harmonic feature *or* agree with them in terms of the
attendant feature, but not both.
The unification that the author claims to have achieved in
this book appears to be the central role of constraint
coordination in each of these cases, primarily in the form
of LC. Close inspection of the analyses proposed in the book
reveals several serious complications, particularly with the
interaction of LCs. I address some of these complications in
what follows, arranged in order according to each of the
four main empirical phenomena already mentioned. None of
these complications is addressed in much depth here, due to
space limitations and the formal complexity of the issues.
I should add that as a person who also wrote a dissertation
on vowel harmony in OT and who also wrestled with the
extremely complex (but deceptively simple) ways in which LCs
interact in an OT grammar, I am simultaneously impressed
with the book's analytical coverage and sympathetic with the
author's not having confronted many of these complications.
Directionality is arguably the central concern of work on
vowel harmony and related phonological processes. (I will
refrain from citing any of the many relevant works here.)
One of the most difficult obstacles to a restrictive theory
of directionality is the ubiquitous tension with observed
facts. In the case of vowel harmony, any attempt to restrict
directionality in a principled way can be countered with a
pattern that cannot be adequately described.
One recent proposal for restricting directionality in vowel
harmony is my own cyclic account (Bakovic 2000, 2003). This
account downplays the significance of morpheme-internal
directionality and denies the existence of (strictly)
affix-to-root directionality. This proposal is directly
countered in the current book with the case of Futankoore
Pulaar (Fula; Paradis 1992), in which harmony appears to be
best described as proceeding leftward from the rightmost
affix (see Hyman 2002 for other examples of right-to-left
harmony that cannot be interpreted cyclically). For the
purposes of this review, I accept that the Fula case is a
direct counterexample to the claim that vowel harmony is
only either cyclic (stem-controlled) or dominant-recessive.
The account of Fula harmony proposed by the author requires
three distinct parts that individually contribute to the
overall right-to-left nature of the pattern:
(a) high ranking of a positional faithfulness constraint
favoring the rightmost vowel of a word (p. 140), to
account for the fact that this vowel is the trigger.
(b) low ranking (or absence) of any logical conjunctions
favoring the edgemost vowels of the root (p. 140), to
account for the fact that this vowel is a target.
(c) exceptional reversal of "the almost universal ranking"
between two positional faithfulness constraints (pp.
143-144), to account for the fact that a vowel between
the trigger and an opaque vowel agrees with the trigger.
The last of these three crucial parts of the analysis is
particularly problematic. If this exceptional reversal did
not hold of the grammar of Fula, then a vowel between the
trigger and an opaque vowel would agree with the opaque
vowel -- a pattern that is unattested in all cases of vowel
harmony, regardless of the relative positions of the harmony
trigger, the root, and the opaque vowel. The generalization
that appears to be necessary to capture is that harmony
always proceeds outward from the trigger; in other words,
that the directionality of harmony and the fate of a vowel
trapped between the trigger and an opaque vowel are two
sides of the same fact, to be captured with one analytical
mechanism. This is not so in the author's account of Fula.
I will not pretend here that my own account (Bakovic 2000,
2003) does a better job of capturing this generalization; I
think it succeeds rather well with cases of root-to-affix
directionality but admit that it encounters problems in
other cases (including Fula). But the author's account of
directionality in this book also fails to account for the
fate of a vowel trapped between two disagreeing opaque affix
vowels because it depends on the relative ranking of root
and affix faithfulness constraints. Root faithfulness is
irrelevant in this kind of case and there is a tie on affix
faithfulness, so the fate of the trapped vowel is predicted
to fall to other constraints. But as Anderson (1980) has
shown with a relevant example from Turkish, the fate of the
trapped vowel should be determined by whatever mechanism is
responsible for the directionality of harmony.
2. Transparent vowels.
The author shows that the LC account of transparent vowels
requires a third conjunct in addition to those noted above:
a markedness constraint defining the set of transparent
vowels. Without this, other disharmonic vowels (e.g., vowels
in disharmonic roots, or other vowels in the same language
that may be opaque rather than transparent) are incorrectly
predicted to behave transparently. The problem is that the
set of transparent vowels is already defined by a markedness
constraint against their harmonic counterparts, which are
independently absent from the vowel inventory of the
language (Kiparsky 1981). Because this is viewed as a
coincidence in the author's analysis, the prediction made is
that there could be a language in which only certain vowels
in disharmonic roots behave transparently (namely, those
targeted by the markedness component of the relevant LC),
while all other disharmonic vowels behave opaquely. Such a
pattern appears to be unattested.
Another consequence of the author's proposal is that it
distinguishes a single transparent vowel from strings of
more than one. Most accounts of transparency predict that a
string of transparent vowels behaves just like a single one;
a maximal string of transparent vowels either agrees or
disagrees with the vowels on either side. In the author's
proposal, each transparent vowel is evaluated independently
by the LC: each individual one must either agree or disagree
with both of its neighbors. Even in the simplest case of a
string of two transparent vowels, then, the only way to
satisfy the LC is (i) for the transparent vowels to become
harmony triggers, enforcing their harmonic feature value on
their neighbors on both sides or (ii) for the transparent
vowels to alter their values of other features so that they
can (vacuously) satisfy the LC. For a string of transparent
vowels to differ in its behavior from a single one in either
of these two ways appears to be unattested.
What is attested -- though subject to a significant amount
of variation -- is a kind of variant of (i): a string of
transparent vowels may behave opaquely, enforcing their
harmonic feature value on their neighbors *but on the side
opposite the harmony trigger only*. The author claims that
this "is perfectly accounted for in this approach", and that
the observed variation could be due to variability of the
domain assessed by the LC (pp. 166-167). The problem, of
course, is that the unattested cases above are also
predicted to exist, casting doubt on the overall account.
3. Trojan vowels.
The LC responsible for the behavior of Trojan vowels states:
'if unfaithful, then disharmonic'. This has the desired
effect, but only if the feature mentioned by both parts of
the Trojan LC is the harmonic feature; if it is some other
feature -- or if the language has no harmony at all! -- then
apparently random disharmony patterns are expected to be
caused by input vowels that must surface unfaithfully due to
constraints responsible for the vowel inventory. This
problem can perhaps be handled by imposing further
restrictions on the ranking of LCs with respect to other
constraints, but the solution is by no means obvious.
Given LCs demanding harmony or disharmony (responsible for
transparency) and LCs demanding faithfulness or disharmony
(responsible for Trojan vowels), what about LCs demanding
faithfulness or harmony? It isn't clear that such an LC
would have any effect if it mentions a feature that is
otherwise harmonic, but it would be easily discerned if it
mentions a different feature: apparently random harmony
(rather than, as above, disharmony) patterns are expected to
be caused by input vowels that must surface unfaithfully due
to constraints responsible for the vowel inventory.
A Trojan LC must also be stipulated to apply only to those
(input) vowels that are independently missing from the
surface vowel inventory. As with transparency, this is
accomplished with a third part of the LC: a markedness
constraint mirroring the one responsible for the inventory.
The arbitrariness of the coincidence in this case is even
more problematic than in the account of transparency, due to
the fact that another part of a Trojan LC is a faithfulness
constraint. The problem is that vowels in positions that are
targets of harmony (e.g., affixes) may violate faithfulness
when they undergo harmony, which makes them potentially
vulnerable to the Trojan LC. Again, random patterns of
harmony and/or disharmony are expected by vowels attempting
to avoid violation of the Trojan LC.
This last problem is particularly apparent in a comparison
between the analyses of Trojan vowels in Hungarian (p.
188ff) and those in Yoruba (p. 200ff). In Hungarian, Trojan
vowels are some instances of [i] and [e] (disregarding vowel
length), and the [+back] harmonic counterparts of these
vowels are independently missing from the surface vowel
inventory. In Yoruba, Trojan vowels are some instances of
[i] and [u] (disregarding nasality), and the [-ATR] harmonic
counterparts of these vowels are independently missing from
the surface vowel inventory. The Trojan behavior in each
case is accounted for in the manner proposed by the author:
(i) by assuming that the Trojan vowels are underlyingly
specified with the opposite value of the harmonic feature
than they surface with and (ii) with an LC that takes
advantage of the fact that these vowels will independently
surface unfaithfully, and requiring vowels adjacent to an
unfaithful Trojan vowel to disagree with the Trojan vowel.
In the case of Hungarian, there is little more that needs to
be said; because the same vowels [i] and [e] are transparent
in harmony target positions, they fortuitously satisfy the
Trojan LC whether or not they are underlyingly [-back] or
[+back]. In Yoruba, however, [i] and [u] are opaque in
harmony target positions and thus either satisfy the Trojan
LC (if they happen to be underlyingly [+ATR]) or violate it
(if they are underlyingly [-ATR]). This forces the author to
redefine the Trojan LC for Yoruba (pp. 206-207) such that
the faithfulness conjunct only refers to the harmony trigger
position. The prediction, of course, is that there could be
a pattern in which some of the relevant vowels are opaque
while others are transparent, depending on their underlying
source. This is unattested. Not only does the markedness
conjunct of a Trojan LC accidentally refer to gaps in the
surface vowel inventory, it must also duplicate an
independently necessary definition of the harmony trigger in
languages like Yoruba with opaque vowels.
4. Parasitic harmony.
The analysis of parasitic harmony depends on a condition
noted in passing on p. 109: that faithfulness to the
attendant feature must be ranked higher than faithfulness to
the harmonic feature. The opposite ranking of these
constraints would result in a case in which a parasitic LC
would be preferably satisfied by a change in the attendant
feature. (This consequence is noted by the author, again in
passing, on p. 220.) A case of this type -- which may be
characterized as dissimilation in terms of one feature only
when there is already disagreement in terms of another, or
'parasitic dissimilation' -- is not attested.
It is worth noting at this point that the common denominator
of all of the LCs discussed above is a constraint directly
demanding disagreement in terms of some feature, and this is
arguably the source of much of the problems noted with these
LCs. This is perhaps expected, since there is very little
evidence for constraints demanding disagreement in terms of
some feature in any case: to my knowledge, there exists no
language in which every vowel disagrees with its neighbors,
such that the result is an unbounded alternating sequence of
vowels. Actual cases of dissimilation do require an account,
but an account in terms of a dissimilation constraint of the
type assumed by the author in the proposed LCs is bound to
predict cases of the unattested type outlined here.
I began my critical evaluation by drawing attention to the
author's goal (stated in the book's abstract) of providing a
unified analysis of four aspects of vowel harmony, among
others: directionality, transparency, Trojan vowels, and
parasitic harmony. Throughout the book, the particular
analyses proposed are contrasted with others found in the
literature in terms of alignment, cyclicity, neutrality,
privativity/underspecification, targeted constraints,
floating features, sympathy theory, optimal domains theory,
etc. The extraordinary variety of alternatives discussed
makes it easier to accept the claim that the author's
account in terms of positional faithfulness and local
conjunction is in fact the most unified account to date.
Two considerations need to be kept in mind, however.
First, very little of the previous work on vowel harmony has
as its analytical goal a grand unified theory of vowel
harmony. Most of the work that the author directly
challenges addresses some particular aspect or aspects of
vowel harmony, often within a single language or very small
group of languages, and with the more modest goal of
accommodating the analysis of the facts into a larger
framework of assumptions. It is thus appropriate for the
author to contrast his approach to a particular phenomenon
with previous approaches to the same phenomenon; it is
inappropriate, however, to conclude from a set of such
comparisons that previous approaches have failed to provide
unified analyses of the set of phenomena that the author has
uniquely chosen to unify.
Second, the unified analysis championed by the author is in
many respects at odds with fundamental assumptions of the
framework in which it is couched. One of the fundamental
assumptions of OT is that different constraint rankings are
the primary source of crosslinguistic variation (and the
only such source for some). A consequence of any OT analysis
is that each of the possible rankings of the constraints in
the analysis should, at least in schematic form, correspond
to a (different) pattern in some other language.
This assumption and consequence is challenged in various
places in the book, most notably at the end of Chapter 4
(pp. 154-155) and in the second-to-last section of the
general conclusion (pp. 256-258). This challenge -- that,
like constraints themselves, constraint interactions must
serve some 'higher' functional purpose -- is embodied here:
"The Bad Ranking Hypothesis: A constraint ranking is
counter-productive if it neither facilitates articulation
nor interpretation. Such grammars are avoided." (p. 155)
The problem with this challenge is not that it is wrong but
rather that it is unformalized. The author mentions such
concepts as "general strategies of information structuring",
"facilitation of information retrieval" and "maximisation of
interpretability" without defining them nor demonstrating
how any of the undesirable constraint interactions serve
none of these purposes. Since functional motivation appears
to be a major underpinning of the proposal, the author has
missed an opportunity to propose an explicit theory of it.
Anderson, Stephen R. (1980) Problems and Perspectives in the
Description of Vowel Harmony. Issues in Vowel Harmony, ed.
by R. Vago, pp. 1-48. John Benjamins.
Bakovic, Eric (2000) Harmony, Dominance and Control. PhD
thesis, Rutgers University. [ROA-360.]
Bakovic, Eric (2003) Vowel Harmony and Stem Identity. San
Diego Linguistic Papers 1/2. Linguistics Department, UCSD,
Beckman, Jill (1998) Positional Faithfulness. PhD thesis,
UMass Amherst. Published 1999, Garland. [ROA-234.]
Crowhurst, Megan and Mark Hewitt (1997) Boolean operations
and constraint interaction in optimality theory. ROA-229.
Hyman, Larry (2002) Is there a right-to-left bias in vowel
harmony? Presented at 9th International Phonology Meeting.
Kiparsky, Paul (1981) Vowel Harmony. Ms., MIT.
McCarthy, John and Alan Prince (1995) Faithfulness and
Reduplicative Identity. Papers in Optimality Theory, ed. by
J. Beckman, S. Urbanczyk and L. Walsh Dickey, pp. 249-384.
UMass Occasional Papers in Linguistics 18. [ROA-60.]
McCarthy, John and Alan Prince (1999) Faithfulness and
Identity in Prosodic Morphology. The Prosody-Morphology
Interface, ed. by R. Kager, H. van der Hulst, and W.
Zonnefeld, pp. 218-309. CUP. [ROA-216.]
Paradis, Carole (1992) Lexical Phonology and Morphology: The
Nominal Classes in Fula. Garland.
Prince, Alan and Paul Smolensky (1993/2002) Optimality
Theory: Constraint Interaction in Generative Grammar.
ROA-537. [In press, Blackwell.]
Smolensky, Paul (1993) Harmony, markedness, and phonological
Smolensky, Paul (1995) On the structure of the constraint
component Con of UG. ROA-86.
| ABOUT THE REVIEWER:
ABOUT THE REVIEWER
Eric Bakovic is an assistant professor in the Linguistics
Department at the University of California, San Diego. His
research interests include phonological theory, vowel
harmony, and Spanish phonology and morphology.