222 108 25MB
English Pages 402 [404] Year 1988
Advances in Nonlinear Phonology
Linguistic Models The publications in this series tackle crucial problems, both empirical and conceptual, within the context of progressive research programs. In particular, Linguistic Models will address the development of formal methods in the study of language with special reference to the interaction of grammatical components. Series Editors: Teun Hoekstra Harry van der Hülst Michael Moortgat
Other books in this series: 1 Michael Moortgat, Harry van der Hülst and Teun Hoekstra (eds.) The scope of lexical rules 2 Harry van der Hülst and Norval Smith (eds.) The structure of phonological representations. Part I. 3 Harry van der Hülst and Norval Smith (eds.) The structure of phonological representations. Part II. 4 Gerald Gazdar, Ewan Klein and Geoffrey K. Pullum (eds.) Order, Concord and Constituency 5 W. de Geest and Y. Putseys (eds.) Sentential Complementation 6 Teun Hoekstra Transitivity. Grammatical Relations in Government-Binding Theory
Advances in Nonlinear Phonology
Edited by
Harry van der Hülst INL, Dutch Lexicological Institute, Leyden
Norval Smith Institute for General Linguistics, University of Amsterdam
1985 FORIS PUBLICATIONS Dordrecht - Holland/Cinnaminson - U.S.A.
Published by: Foris Publications Holland P.O. Box 509 3300 AM Dordrecht, The Netherlands Sole distributor for the U.S.A. and Canada: Foris Publications U.S.A. P.O. Box C-50 Cinnaminson N.J. 08077 U.S.A. CIP-DATA KONINKLIJKE BIBUOTHEEK, DEN HAAG Advances Advances in nonlinear phonology / ed. by Harry van der Hülst, Norval Smith. - Dordrecht [etc.]: Foris. - (Linguistic Models ; 7) With ref. ISBN 90-6765-125-7 paper SISO 805.2 UDC 801.4 Subject heading: phonology.
ISBN 90 6765 125 7 (Paper) ® 1984 by the authors. No part of this book may be translated or reproduced in any form, by print, photoprint, or any other means, without written permission from the publisher. Printed in the Netherlands by ICG Printing, Dordrecht.
Table of Contents
Part I: Introduction Harry van der Hulst and Norval Smith The framework of Nonlinear Generative Phonology
3
Part II: Stress and accent Bruce andRule Stanistaw Puppel On theHayes Rhythm in Polish
59
Jack Hoeksema Formal Properties of Stress Representation
83
Jan G. Kooy and Mies van der Niet Stress, Stress Shift, and Morphology: the Case of Dutch -baar
101
Carlos Gussenhoven Intonation: a whole Autosegmental Language
117
Ron van Zonneveld Word Rhythm and the Janus Syllable
133
Part III: Prosodic Categories GeertE. Booij Coordination Reduction in Complex Words: a Case for Prosodic Phonology Grzegorz Dogil
143
Nonlinear Phonology in the Area of Speech and Grammar Pathology
161
Anneke Clitics inNeijt Arboreal Phonology
179
Marina Nespor The Phonological Word in Italian
193
VI
Contents
Domingos Prieto Prosodie Representation and Transformations in Galician-Portuguese
205
Irene Vogel On Constraining Prosodie Rules
217
Part IV: Autosegmental Phonology Ben Hermans The Relation between Aspiration and Preaspiration in Icelandic . . . .
237
Harry van derHulst Vowel Harmony in Hungarian: a Comparison of Segmental and Autosegmental Analyses John J. McCarthy Speech Disguise and Phonological Representation in Amharic
267 30S
Christine ter Mors Empty V-Nodes and their Role in the Klamath Vowel Alternations . .
313
Roland Noske and Syllable Changing Processes in Yawelmani Syllabification
335
Norval Smith Spreading, Reduplication and the Default Option in Miwok Nonconcatenative Morphology
363
Preface and Acknowledgements This book contains 18 original contributions dealing with problems in the area of nonlinear generative phonology. The subject matter of the articles (or in a few cases, a closely related topic) was presented during a workshop held in the Summer of 1983 in Amsterdam (August 8-12) but they were all written afterwards (thus dating from the end of 1983 to the end of 1984). The editors have added an introductory chapter presenting an overview of the framework of nonlinear phonology. As a whole this volume offers a good perspective on ongoing research in the field of nonlinear phonology, especially in the Netherlands. The workshop was organized by Harry van der Hulst, Anneke Neijt and Norval Smith and made possible through financial support of the Netherlands Organization for the Advancement of Pure Research. The Institute for General Linguistics of the University of Amsterdam allowed us to make use of their facilities. We would like to thank all who contributed to the success of the workshop.
Parti
Introduction
The Framework of Nonlinear Generative Phonology Harry van der Hülst and Norval Smith INL, Leyden and University of Amsterdam
1. INTRODUCTION
In this article we will provide an introduction to current views on the structure of phonological representations within the theory of generative grammar, thus providing the context within which the articles in this collection should be placed. Following a brief sketch of these views in section 2, section 3 and 4 will provide a discussion of the two main theories in this area, i.e. the theory of autosegmental phonology and the theory of metrical phonology. Section 5 offers a short summary of the articles in this volume*
2. A SKETCH OF GENERATIVE PHONOLOGY
Present-day generative phonology finds its starting point in the theory presented in Chomsky and Halle's The Sound Pattern of English (SPE; Chomsky and Halle 1968). Although there are many differences between the SPE approach and current approaches, a number of fundamental assumptions have been maintained without change, such as the distinction between underlying and surface representations, rule ordering and the desire that linguistic generalizations be reflected in the notational system: the more significant the generalization, the simpler the notation. The theory proposed in SPE has a derivational aspect and a representational aspect. The first aspect involves issues such as the formulation of phonological rules, rule application, rule ordering and interaction with morphological rules, whereas the second aspect involves the structure of phonological representations at each level of the derivation. The changes that we will discuss here mainly involve the representational aspect. This
* Parts of this introduction have been borrowed from Van der Hülst and Smith (1982a) and from introductory sections in Van der Hülst (1984). Some other collections of articles in the field of nonlinear generative phonology are: Aronoff and Oerhle (1984), Clements and Goldsmith (1984), Van der Hülst and Smith (1982), Safir (1979) and Sezer and Wetzels (1985).
4
Harry van der Hulst and Norval Smith
implies that we will not be primarily concerned with the theory of Lexical Phonology, which embodies a view on the derivational aspect different from that found in SPE. However, before proceeding to a discussion of the representational aspect, we will briefly characterize the main features of the lexical model. 2.1. The derivational aspect: lexical phonology The theory of lexical phonology is derived from a number of independent developments within the field of generative phonology and morphology. The theory of lexical phonology as advocated in Mohanan (1982) and Kiparsky (1982, 1983) incorporates a proposal found inPesetsky (1978). Pesetsky argued that rules which apply cyclically can be interpreted as rules applying in the lexical component after each application of a word formation rule. The advantage of this move is that cyclic application need not be stipulated, but follows from the organization of the grammar. In addition to this, the theory of lexical phonology revived a proposal advanced in Siegel (1974). Siegel proposed ordering word formation rules in blocks, and ordering certain phonological rules in between these blocks. This theory has become known as the ordered blocks (or levels) hypothesis. Combining these two proposals then gives rise to the following model: A stem enters the first level, and may be subject there to a number of morphological rules. After each application of a morphological rule the resulting complex structure passes through a list of lexical phonological rules relevant to that level. Every rule whose SD is met applies. The complex structure then enters a following level, where the same procedure is followed:
In this diagram one notices a class of post-cyclic rules. These are rules, that are more phonetically motivated then the cyclic rules, and that
Introduction
5
apply in more or less purely phonologically defined contexts. Post-cyclic rules may apply at the level of words (in which case we have post-cyclic lexical rules), and at the level of phrases and larger units (post-cyclic, post-lexical rules). Cyclic lexical rules are subject to a condition, called the Strict Cycle Condition. In essence this condition forbids a rule to change phonological information within a particular cycle unless the substring which meets the rule's structural description came into being on that cycle. For technicalities we refer to Mascaro (1976) and Halle (1978). In addition, all lexical rules (cyclic or post-cyclic) must apply in a "structure preserving" fashion. This does not mean that a lexical rule cannot change feature-values (it can, as long as it does not violate the Strict Cycle Condition), but it does mean that lexical rules do not create new segment types or new sequential or suprasegmental configurations, i.e. lexical rules do not violate constraints on lexical representations. If a language bars a segment type A from the inventory of underlying segments then a lexical rule is blocked in those cases where its application would create A. Similarly, if there is a sequential constraint barring the sequence AB then no lexical rule can apply to create AB. An extensive application of this model can be found in Rubach (1984), and Mohanan (1982), each of whom considers a large body of data from a single language, and in Kiparsky (1982, 1983) who shows that many classical problems can be solved in terms of the lexical model. The model utilized in these studies displays certain variations but it is clear that a consensus of opinion on the essential nature of the interaction between phonological rules and morphological rules has been firmly established.
2.2. The representational aspect: nonlinear phonology The changes that we have witnessed over the past few years with respect to the representational aspect have been caused primarily by a considerable extension of the empirical domain of the theory. The following quote from Chomsky (1955: 29) is still largely applicable to SPE: In this study, suprasegmental features (pitch, stress, juncture) have not been seriously considered. Ultimately of course, these phenomena must be incorporated into any full syntactic theory, and it may be that this extension still requires a more elaborate system of representation.
In SPE suprasegmental features are treated as if they were segmental features. The phonological representation is unilinear, i.e. it consists of a single sequence of segments and boundary symbols. Segments consist of an linearly unordered set of features, each of which has a binary value. The sequence of segments is associated with a hierarchical structure that is
6
Harry van der Hülst and Norval Smith
non-phonological, i.e. morphological and syntactic. Currently a phonological representation is considered to be a three-dimensional object, in which we find not just one sequence of segments, but several sequences. Hence the representation is termed multilinear or nonlinear. These sequences (called tiers) are linked to a central tier that consists of abstract units (called slots) to which the segments on the other tiers are associated. This pivotal tier (called the skeleton or skeletal tier) also constitutes the interface between the (morpho) syntactic hierachical structure and a phonological hierarchical structure. From today's point of view the quote given above sounds almost like an understatement. We will now sketch in abstract terms the two ways in which the SPE conception of phonological representations has been altered, before proceeding with a discussion of the type of data that have motivated these changes. Present-day nonlinear phonology incorporates the results of two independent theories which were proposed more or less at the same time by students at MIT. The first theory is called autosegmental phonology and proposes that representations are multilinear. The second is called metrical phonology and proposes that representations are hierarchical. 2.2.1. Autosegmental phonology The central claim of autosegmental phonology is that we must abandon the view of segments as unordered sets of specified features. It has been shown that the scope of a specified feature need not be a single segment. The scope of a feature may both be smaller and bigger than a single segment. Before giving some examples let us emphasize that the term segment must be redefined if features are to be allowed to have different scopes. We must first determine what it is that features may have within their scope. The current view is that features have scope over abstract units, called slots (also sometimes referred to as timing units). One view of slots is that they are completely unspecified units, represented by the symbol "X", another is that we have two types of slots, normally represented by the symbols "C" and "V". These different views will be discussed in section 4.2. For the time being we will use the symbol "X". The sequence of X-segments is referred to as the skeletal tier. There is of course no objection to calling the units that constitute this tier segments, as long as we realize that the interpretation of this term has been changed. The second point to be precise about is the status of features. The idea is that features are also segments; segments on their own, hence autosegments. If, in a particular language, there is good reason to represent the feature [F] as having scope over more than one slot then a sequence of such features is regarded as a sequence of [F\-segments. On the assumption that all other features are synchronized (i.e. have the same scope, in this example over one slots only), the following representation is the result:
Introduction
7
(2) X 1 +G -H
[-F] 1 X
X r
+N
1
-G -H
-i r
-N
1
-G +H +N
-i r
[+F] X
[F]-tier
X
skeletal tier
1
l-i r -G +G -H +H +N
-N
An understandable, but strictly speaking confusing, term to use for the third "cumulative" tier is segmental tier. It is possible to maintain that such a cumulative tier does not even exist and that it just happens to be the case here that the [G]-tier, the [H]-tier and all other tiers (except the [F]-tier) are subject to the same function that associates them to the skeletal tier. However, since wide scope is the special case rather than the norm it is defensible to assume that the third cumulative tier has a theoretical status, and that features are bundled in the traditional way, unless there is evidence to the contrary. On this one may consult Goldsmith (1979). But autosegmentalists have not only argued in favor of wide scope. Features must also be permitted to have narrow scope: (3)
[-F]
f+F] X
I
[...] The need to give features both wide and narrow scope (the reasons for which we will discuss in the next section) has led to the development of the theory ofautosegmentalphonology.
2.2.2. Metrical phonology The second change in our conception of phonological representations is logically independent of the first. As we said above, the only hierarchical structure that is imposed on the row of segments in SPE is of a morphosyntactic nature. This hierarchical structure tells us that substrings of segments constitute morphemes, words and finally phrases and sentences. Even in SPE it is pointed out that syntactic bracketing is not appropriate in all cases to characterize the phonological word, or domain over which intonational contours extend. To remedy this defect certain rebracketing operations are suggested. In subsequent work it has been pointed out that the mismatch between morpho-syntactic structure and other
8
Harry van der Hülst and Norval Smith
necessary kinds of structure is even more serious. The clearest examples involve rules that specify what sequences of segments are wellformed in a particular language. In SPE it was assumed that the domain of these rules was the morpheme, but many phonologists have pointed out that another unit would be more appropriate. This unit is the syllable. The logical conclusion of the introduction of phonologically motivated units comprising substrings of segments was that a complete phonological (or prosodic) hierarchical structure was assumed, distinct from the morphosyntactic structure, although not unrelated to it. This point of view did not develop as a result of the need for the units intonationalphrase, phonological word or syllable in themselves. New research contributed to the establishment of additional units in the phonological hierarchy. Developments in the treatment of stress, for instance, led to the recognition of the foot as a constituent intermediate between the syllable and the word. The new theory of stress was termed metrical phonology, but this term soon came to be used to refer to a theory of phonological constituent structure in general. The independence of morpho-syntactic and prosodic constituent structure is acknowledged by assuming two distinct tree structures imposed on the string of segments, which then forms the interface between these two structures. phWord Prosodic Hierarchy a
Morpho-syntactic Hierarchy Word "phWord" and "Word" stand for phonological and morpho-syntactic word, units which may but need not be isomorphic. A slightly more detailed discussion of the prosodic categories and their relation to morphosyntactic categories is presented below in section 4.
Introduction
9
2.2.3. Autosegmental and metrical phonology combined First of all it turns out now that the slots constituting the skeletal tier can be interpreted as the smallest units, the terminal symbols, of the phonological constituent structure. Putting the two theories together in this way we can no longer look upon a phonological representation as a two-dimensional object. In fact this would have been impossible anyway for a case in which two features were "autosegmentalized". This explains why the theory that is discussed here is called three-dimensional (as well as nonlinear). Since we now have an overall idea of the organization of a phonological representation, we may be more specific about the relation between the skeletal tier and the other phonetic tiers, using the latter term to refer to all other tiers. An important insight captured in autosegmental phonology is that the relationship between autosegments and slots need not always be stipulated. In certain cases this relationship is predicted by rule. In the simplest case the relationship can be brought about by associating autosegments to slots in a one-to-one fashion, going from left to right; this is called
mapping: (5)
[+F]
[-F]
X
X
[+F] ... X
.
The dotted lines indicate the structural change of the mapping rule that introduces the association lines. These are themselves represented by an unbroken line (cf. 7). If one-to-one association was the only possibility there would be little point in distinguishing between autosegments and slots in the first place. However, as we have already seen, autosegmental phonology was developed precisely because the association is not always one-to-one: autosegments may have wide or narrow scope, which implies that the relation may be many-to-many. The original claim of autosegmental phonology is that, where deviations from the one-to-one pattern arise, the number of autosegments is different from the number of slots. Two types of situation may arise. Either there are more slots than autosegments or there are more autosegments than slots: (6)
a.
[+F]
[-F]
X
X
b. X
[+F]
[-F]
X
X
[+F]
According to the theory of autosegmental phonology representations as in (6) may lead to cases in which features have either wide or narrow scope:
10
Harry van der Hülst and Norval Smith
(7)
a.
[+F]
[-F]
X
X
b.
[+F]
X
[-F]
[+F]
X
X
In (7a) we find spreading and in (7b) dumping; these are technical terms that we will continue to use here. Ignoring the issue of whether spreading and dumping are the norm or the exception, it will be clear that a mismatch between the number of slots and the number of segments always holds within a particular domain. The most general position is that autosegmental association can in principle be bound to any domain that the theory defines. This includes morpho-syntactic domains as well as phonological domains. It turns out to be the case then that we must be able to express the fact that a particular domain comprises not only a sequence of slots, but also one or more autosegments. The following figure illustrates this conception of phonological representations ("AS" stands for autosegment):
(8) word level AS's foot level AS's syllable level AS's segment level AS's X
X
X
X
X
X
X
morpheme level AS's word level AS's phrase level AS's
To provide an example, in most vowel harmony systems we find morpheme level autosegments. The study of autosegmental features in domains other than the morpheme or the word (usually meant as a morpho-syntactic domain) has so far not supplied us with many well-documented examples. We refer to Van der Hulst and Smith (1982b) and Hart (1981), where a slightly different point of view is presented. In forthcoming work Vago also applies and elaborates a version of autosegmental phonology, where autosegments occur on different levels. The above suffices to give the reader an idea of the theory of nonlinear phonology. In the following two sections we will discuss the various types of data that have played a crucial role in the emergence and further development of autosegmental and metrical phonology, which will also
Introduction
11
enable us to discuss the technicalities in greater detail. At the appropriate places we will show how the articles included in this volume fit in the framework at issue.
3. AUTOSEGMENTAL PHONOLOGY
3.1. The characterization of complex segments, tonal and non-tonal The standard theory is characterized by what Goldsmith (1976) has called the "absolute slicing hypothesis". An abstract representation of speech sound is split up into slices, called segments. Each slice is specified with exactly one value for each feature of the total set of features that is required to represent speech sounds. Hence segments have no linearly ordered subparts. In the majority of cases segments can be interpreted as functions from points in time to a particular state of the articulatory organs. So if a segment is specified as [+nasal] this means that while this segment is being articulated the velum is in lowered position. In real life it may be the case that the velum is lowered slightly earlier causing a preceding segment to be nasalized during part of its production. The absolute slicing hypothesis embodies the claim that such a half-nasalized segment is not to be represented as [-nasal, +nasal]. The definition of segments on which this hypothesis is based only allows specifications such as [@F], where @ ranges over + or -. In the SPE theory it is also possible to have an integer as the feature value, indicating a certain degree of, for example, nasalization, but integers are used for phonetic representations only. Since in this example the nasalization is dependent on the presence of a neighbouring nasal, the integer would indeed only be required as part of the phonetic representation, and not as part of the underlying representation, where the segment in question would be specified as [-nasal], thus abstracting away from the nasalization (which does not in this case distinguish between phonemes, in the classical sense). In some cases, however, we cannot abstract away from the fact that an articulatory state changes during the production of one slice, because in such cases the change is an intrinsic property of the segment, i.e. it is not caused by some neighbouring sound. Examples are numerous: affricates, pre- and postnasalized consonants, pre- and post-aspirated consonants, (short) diphthongs (see Ewen 1982 for a discussion of many of these segment types, usually referred to as complex segments). During the production of an affricate we have a change from a stop to a fricative type of articulation. Strict obedience to the absolute slicing hypothesis would require that we characterize such segments with a feature underlyingly specified with an integer (which would not reflect what was going on) or a separate feature directly refering
12
Harry van der Hülst and Norval Smith
to the change (which would reflect what was going on). In SPE the first possibility is blocked because all features are binary at the phonological level. The feature characterizing affricates is termed [delayed release]. (One can see here an attempt in the name of the feature to capture what is going on.) Similarly, we would need features like [prenasal], [postnasal], [preaspirated], [postaspirated], [diphthong] etc. With respect to tones too, there would be need for what we might call here contour features (as opposed to level features), e.g. features like [rise] and [fall]. There is no objection as such to adding contour features to the inventory of features. Formally they are not different from level features; the difference is merely terminological. The reason for questioning this strategy comes from the fact that we miss certain generalizations by using the contour features. The crucial argument has been advanced by Anderson (1976) with special reference to the contour features [prenasal] and [postnasal] and by proponents of autosegmental phonology (Goldsmith 1976) with reference to tonal contour features. This argument involves the presence of frequently recurring conjunctions in the formulations of phonological rules. Conjunctions of contour tone features and level tone features are a case in point. They show up again and again and may constitute the norm rather than the exception. When the context bar is on the right a low tone typically appears together with a falling tone, when it is on the left a low tone typically appears with a rising tone. Recurrent conjunctions require an explanation. There must be some factor that the conjoined environments have in common, and our formalism must be able to express this. An explanation is available if we decide to abandon (i.e. [fall]) contour features and replace them with sequences of level features. So F is replaced by a sequence consisting of a high tone feature and a low tone feature (HL). It will be clear that given such a move we no longer face the problem of recurrent conjunctions, since a structural description "{F, L } — " (i.e. a conjunction of F and L) can be replaced by "L — " . The mere decision to abandon contour features does not necessarily entail that we must also abandon the absolute slicing hypothesis embodied in SPE. In fact the decision to eliminate tonal contour features had already been taken by Woo (1969). The conclusion she drew from this was that it is impossible for short vowels to have contour tones. Assuming that long vowels can underlyingly be represented as a sequence of two short vowels Woo predicted that only long vowels can bear contour tones. There are indeed languages in which contour tones may not occur on short vowels. A case in point is Lithuanian where complex tones (i.e. rising or falling tones) can only occur on long vowels or sequences of short vowels and sonorant consonants. This in itself supports the position whereby these tones are represented with two level features. Languages of
Introduction
13
this type can be said to have a constraint prohibiting more than one tonal feature linked to a single segment. If contour features are used, such a constraint is hard to express, since, as pointed out, contour features and level features are not formally different. Unfortunately the prediction as such is false. There are many languages in which short vowels bear a contour tone, either underlyingly or at a later stage in the derivation. It seems then that our conception of segments must be altered in order to allow single segments to have two different specifications for one feature. This is the crucial motivation for abandoning the absolute slicing hypothesis: (9)
L
H X
[...] The same conclusion regarding "complex features" is found in Anderson (1976), who gives an argument that is completely parallel to the tonal example involving nasalization of vowels preceding either a nasal consonant or a prenasal consonant. In the cases he discusses there is no evidence for representing the complex consonants as two segments underlyingly (they are not contrastively "long"), hence the decision to represent a prenasal consonant as [+nasal] and [-nasal] directly implies that we must revise our conception of segments.
3.2. Supporting arguments for representing tonal features on an independent tier The study of tone provided the main impetus for the development of autosegmental phonology, and its application in this area has convinced many more phonologists than its application in other areas such as vowel harmony. This volume does not contain a paper on tone, but to understand better where autosegmental phonology has its roots (and, according to some, its most convincing application) we will discuss the tonal motivation in some detail. The first phenomenon that supports the autosegmentalization of certain features involves what has been called stability. In (10) a schematic example of a typical case of vowel deletion is given, leading to the emergence of a contour tone:
14
Harry van der Hülst and Norval Smith
(10)
H
L
H
L
I \ oIm Io
om o
\
Mississippi mud Rhythm rules are quite common: in recent literature one finds them proposed for French (Dell (forthcoming), Phinney (1980)), German (Kiparsky (1966)), Italian (Nespor and Vogel (1979)), Dari (Bing (1980)), Tiberian Hebrew (McCarthy (1979)), Passamaquoddy (Stowell (1979)), and Finnish (Hayes (1981)). Since not too many languages have been checked, it may turn out to be the norm, not the exception, for stress languages to have rhythm rules. Rhythm rules are typically optional. But at least for Polish and English, the variants generated by a rhythm rule are not free, strictly speaking. Rather, the propensity to apply the rule varies, depending on the changes in rhythmic structure that the rule induces. For example, in English it would probably be the norm to apply the rule to the phrase Mississippi législature, but in Mississippi legislation the rule would be quite unlikely to apply. Following Liberman and Prince (1977), we assume that this situation requires a twofold explanation: we distinguish the phonological operation of stress retraction from the system of rhythmic principles that determines when retraction is favored. For English, the Rhythm Rule itself can be expressed quite simply under the metrical theory of stress, as in (2):
60
Bruce Hayes and Stanistaw Puppel
(2)
English Rhythm Rule w s -*• s w \ / \ / cf. Mississippi mud s w s w s
where s is not the strongest syllable of its phrase. -*•
Mississippi mud s w s w s
To determine the propensity for stress to shift, Liberman and Prince invoke a separate representation, the metrical grid, which can be thought of as embodying the rhythmic structure of a text. A metrical grid is an abstract set of units arrayed in rows and columns. The height of the columns represents the stress prominence of syllables, while the rows may be viewed as series of rhythmic beats on different levels. The grids are automatically projected from metrical trees by a set of rules, the details of which may be found in Liberman and Prince (1977, 315-316, 322323). For our purposes the following rules will suffice:
(3)
Grid Construction a. Give every syllable a grid mark as a place marker. b. Referring to the metrical tree, add sufficient additional marks to the grid so that the strongest syllable of every strong constituent has a higher column than the strongest syllable of its weak sister.
Liberman and Prince further assume that certain configurations of the grid are selected as highly valued, or "eurhythmic". The variable propensity of the Rhythm Rule to apply is determined by the degree of rhythmic improvement it provides in the grid. The full story behind the application of the Rhythm Rule in (2) is that the grid of the output form (4b) is more eurhythmic than that of the input form (4a). (4)
a.
x x
x X
X
X
X X
X
b.
x X
X
x x
X
X X X
Mississippi mud s w s w s
X
x X
X
Mississippi mud s w s w s
Rhythm Rule in Polish
61
The grid obtained from applying the Rhythm Rule to Mississippi legislation, however, is said to be no more eurhythmic than the grid of the unretracted input form, so that application of the Rhythm Rule is inhibited.
(5)
a.
x x x x x x x x x x xxxx Mississippi legislation s w s w swsw
b.
x
X X
X
X X X X XXX X x x x x Mississippi legislation s ws w s w s w
V5 Vw Vw s/ \/w V
W Ww s w s \w/ Vs
Our intention is to determine explicitly what the principles are that specify when one grid is more eurhythmic than another. We will describe two proposals, that of Liberman and Prince (1977) and the proposal advanced in Hayes (1984); then use evidence from Polish to determine which theory is correct. The problem of rhythmic tendencies in speech has occupied linguists since long before the invention of metrical theory. There have been two general approaches to the problem, which we will call the "stress clash" theory and the "rhythmic interval" theory. (For the former, see van Draat (1912), Bolinger (1965); for the latter Pike (1945), Abercrombie (1964). Bolinger (1981) would represent a mixed approach). The stress clash theory holds that the paramount goal of rhythmic phonology is to avoid adjacent stresses. Liberman and Prince advance a version of this theory which defines stress clashes as adjacent marks on any level of the grid except the lowest. An application of the Rhythm Rule that is motivated by clash reduction is shown under (6):
(6) *x *x X X
X
X X X X
X
X X
Tennessee Ernie s w s s w
\//
\/
X
X X X
X
X
Tennessee Ernie s w w s w
62
Bruce Hayes and Stanislaw Puppel
An apparent defect of the stress clash theory is the applicability of the Rhythm Rule to forms with no adjacent stressed syllables, as in (4), Mississippi mud. Liberman and Prince provide an ingenious remedy for this: they generalize the notion of "adjacent" to mean "adjacent with respect to the immediately lower level". Specifically, two grid marks constitute a stress clash if they are adjacent on a level, with no mark occurring between them on the next lower level. Under this definition, the stress shift in (4), Mississippi mud, can be seen to be motivated by clash reduction as well. The relevant portion of the grid is as in (7): (7)
*x *x x x x Mississippi mud
x x x x x Mississippi mud
Compare example (5), Mississippi legislation, which resists the Rhythm Rule because there is no clash in the relevant sense: (8)
a.
x x X X X X XXX X x x x x Mississippi legislation x
b.
x x x x x x Mississippi legislation
An alternative to stress clashes is the "rhythmic interval" theory. This is based on the familiar notion of stress isochrony, the principle that stresses tend to occur at regular intervals, at least in the perceptual domain. 2 In Hayes (1984), it is proposed that stress isochrony is not just a perceptual effect, but forms the basic guiding principle for the Rhythm Rule and other rules of rhythmic phonology. The propensity of the rule to apply is dependent on the degree to which the stresses of a phrase are realigned in isochronous positions. Moreover, the spacing of the stresses is not freely chosen, but appears to strive for a target interval of around four syllables. We will refer to this principle as the Quadrisyllable Rule: (9)
Quardrisyllabic Rule A metrical grid is eurhythmic when it contains a row of marks spaced about four syllables apart.
Rhythm Rule in Polish
63
The greater the divergence of the actual grid intervals from the quadrisyllabic norm, the greater the dysrhythmy induced, particularly when the interval is smaller than required. Under this theory, "stress clashes" are merely the extreme on a continuum of dysrhythmy. The data noted above are explained under the rhythmic interval theory as follows: relabelling in Tennessee Ernie ((6)) and Mississippi mud ((4)) is strongly favored, as the interstress intervals are shifted from one to three and from two to four syllables, respectively. In Mississippi legislation ((5)), the interstress interval of the input is already the requisite four syllables, so that the Rhythm Rule is inhibited. The evidence from Polish we invoke here will help to decide between the stress clash and rhythmic interval theories. The outline of the paper is as follows. Section 1 motivates a metrical analysis of the Polish word stress rules and the Polish Rhythm Rule. Section 2 contains data and arguments in favor of the rhythmic interval approach. In the third section, we demonstrate that the Rhythm Rule data for Polish and English are strikingly parallel in several ways, and show that the parallels are explainable on the assumption that both languages adhere to the Quadrisyllable Rule. We conclude by considering the possibility that the Quadrisyllabic Rule is a universal of rhythmic phonology, presenting tentative data from other languages in support of this view.
1. STRESS IN POLISH
The variety of Polish described here is that of the second author, and is close to the standard of educated Poles. The basic facts of word stress in Polish are well known: in words of more than one syllable, main stress falls on the penult; and if at least two syllables precede the penult, then a secondary stress falls on the initial syllable; cf. Warszawa 'Warsaw', kinematografka 'little female cinematographer'. This simple situation admits of a few complications, which we describe below. A restricted set of words, mostly borrowed, displays antepenultimate stress in formal styles of speech. In colloquial style, the penultimate norm reasserts itself. In the words below, the "formal" stressing is given. (10)
fizyka Kórsyka òpera okólica
'physics' 'Corsica' 'opera' 'surroundings'
When a word with antepenultimate stress bears a zero or disyllabic case ending, penultimate stress occurs in all styles:
64
Bruce Hayes and Stanisiaw Puppel
(11)
a. b. c. d. e. f.
matematyk-a matematyk matematyk-ami matematyk matematyk-a matematyk-owi
'mathematics (nom.sg.)' (gen.pl.) (ins. pi.) 'male mathematician (nom.sg.)' (gen. sg) (dat. sg.)
The account we propose for these facts essentially follows that of Dogil (1979). The principal rules of stress assignment in Polish are the Main Stress Rule and the Secondary Stress Rule, stated as follows: (12)
Main Stress At the right edge of a word, form a binary foot, labelled s w>;cf. a. Stacha s w
V
'Stach (gen.)' (13)
b. Warszawa s w
V
'Warsaw'
c. kinematografka s w
V
'little female cinematographer'
Secondary Stress Form all remaining syllables into an unbounded left-branching structure, with sister nodes labelled sw, and adjoin it to the main stress foot. a.
b. Warszawa w s w
These rules derive the ordinary patterning of stress in most words. To account for cases of antepenultimate main stress, we follow Halle and Vergnaud (forthcoming) and Dogil (1979) in invoking a rule of extrametricality assignment. (For the theory of extrametricality rules and the evidence supporting it, see Hayes (1982).) The extrametricality rule for Polish must be something like (14): (14)
Extrametricality syl -*• [+ex] / X
in formal style, where X is one of a restricted set of stems.
65
Rhythm Rule in Polish
In the forms of (10) and (11a, e) rule (14) marks the inflectional ending -a as extrametrical. This syllable is thus ignored by the Main Stress Rule, so that the main stress foot is constructed over the antepenultimate and penultimate syllables. The final syllable is then adjoined to this foot by the universal convention of Stray Syllable Adjunction (cf. Liberman and Prince (1977, 294), Hayes (1982, 235)), and other rules give the correct result: (15)
[maternatyk](a)
Extrametricality
[maternatyk](a) sw V
matematyka - màtemâtyka sw w s ws ww
Main Stress
Stray Syllable other rules Adjunction
V
\/
\l
In ( l i b ) , matematyk, there is no post-stem syllable, so that Extrametricality cannot apply, and penultimate stress results. In (11c), matematy k-ami, Extrametricality can apply, but it runs afoul of the Peripherality Condition (Harris 1983, Hayes 1982), which requires that any extrametricality marking not at the edge of the stress assignment domain be erased. This leads to penultimate stress, as shown below: (16) [matematykjami
[matematyk](a)mi ->• matematykami -*• matematykami s w ww s w
Ml IV s I s \ls \lw
input form
Extrametricality
Peripherality Condition
stress rules
There is evidence that the Peripherality Condition is universal, so we exclude it from the set of rules specific to Polish. The extrametricality approach to these exceptional forms has two advantages. First, it predicts that the regular stress rule may not be violated arbitrarily, but only in cases where stress falls exactly one syllable to the left of its normal location. This pattern is a typologically common one, 3 so it is proper that it should fall out naturally from the devices of
66
Bruce Hayes and Stanisiaw Puppel
phonological theory. Second, the theory accounts for the appearance of penultimate stress in forms with disyllabic or zero endings - if we had instead posited a stem-governed rule of antepenultimate stress, this fact would have gone unexplained. We assume, then, the rules of Extrametricality, Main Stress, and Secondary Stress as the basic rules of word stress assignment in Polish. These rules are all word-level. They are followed by the Rhythm Rule, which applies phrasally. The Polish Rhythm Rule is much like the English one under (2), but there are two differences. First, unlike its English counterpart, the Polish rule may shift stress in either direction. A typical case of leftward shift is shown under (17): (17)
Bogusi-awa Bozek s w s w s w
2 3 1 Bogus4-awa Bozek s w s w s w
Rightward shift is much rarer, as the relevant input forms seldom arise. To our knowledge, rightward shift is found only in verb forms ending in the disyllabic clitics -bys'my and -bys'cie.4 The clitics are given penultimate stress at the word level, then are adjoined as weak sisters of the preceding verb, as in (18): (18)
1 2 poszlibysmy jedlibyscie
'we would have left' 'you (pi.) would have eaten'
When such verb-clitic units are preceded by a stronger stress (provided, for example, by contrastive emphasis), their stress shifts rightward: (19)
12 3 my poszlibysmy ws w
13 2 my poszlibysmy ws w
'we would have left'
My poszlibysmy thus receives the same stress contour as my opuszczamy 'we are leaving', where the verb is stressed opuszczamy in isolation. It is clear that the rightward and leftward versions of the Polish
Rhythm Rule in Polish
67
Rhythm Rule can be collapsed with some version of mirror image notation. One possible formulation is stated below: (20)
Polish Rhythm Rule s w // w s' where s'is the strongest stress of the phrase.
A second difference between the Polish and English Rhythm Rules concerns the tendency of native speakers to apply them. Although rhythmic stress shift is clearly possible in Polish (and the effects clearly audible), the propensity to apply the rule is generally less than it would be in English. For example, in a phrase like Bogusiawa Bozek, the statistical norm would be for the Rhythm Rule not to apply, whereas for a comparable English phrase like California dreamer the opposite holds true. In general, the Polish judgments for a given prosodie configuration are shifted relative to English in the direction of not applying the rule. The more obscure cases are still accessible to introspection, however, if they are enunciated in rhythmic speech styles, for example as accompaniment to tapping on a table. In this context, the judgments become more liberal, and approximate the tolerance of English.
2. ON THE REQUIREMENTS OF EURHYTHMY
Returning to our main issue, we wish to contrast rival views of the rhythmic goals that the Rhythm Rule strives to achieve. These are repeated below: (21)
a. Stress Oash Theory (Liberman and Prince (1977)) A metrical grid is eurhythmic when it contains no stress clashes. A stress clash consists of two grid marks X j , X2 that are adjacent on a representation consisting of their own level and the immediately lower level, b. Rhythmic Interval Theory (Hayes (1984)) A metrical grid is eurhythmic when it contains a row of marks spaced as close as possible to four syllables apart (the Quadrisyllable Rule), with greater divergence implying greater dysrhythmy.
We have located three areas in which the two theories can be tested against each other. Specifically, one may (a) vary the interstress intervals of the input, retaining stress clash throughout; (b) vary the interstress intervals of the output, again keeping the presence of clash constant; and (c) create stress clashes while keeping the interstress intervals constant. We treat these areas in turn below.
68
Bruce Hayes and Stanisiaw
Puppel
2.1 Stretching the Input Interval Consider first the pair of examples under (22):
(22)
a. *x
x *x
X X
X
X X
X
X X
X
XX X X
XX
XX XX
XX
telewizor Kota sw sw s w
Vw V \ l
telewizor Kota s w sw s w
'Kot's television'
V \l \ l
w b. *x X X XX XX
x *x X X X X
telewizor Zagi-oby s w sw wsw
x x x x x x x x xx x xx telewizor Zag4oby 'Zagloba's television' s w sw w s w
By LP's definition, a clashing configuration is converted to a non-clashing one in both of these examples, so that relabelling should be equally likely in both. This is not so: it is considerably more natural to retract stress in (22a) than (22b). This difference follows directly from the rhythmic interval theory. In (22a), the best available "scansion" of the grid involves a row having marks spaced two syllables apart, while the stress-shifted version satisfies the requirement of (9) perfectly with a quadrisyllabically spaced row. By contrast, (22b) involves a shift from a trisyllabic to a pentasyllabic interval. This is no improvement, so no relabelling is expected. Examples (22a) and (22b) are representative, as ((23)) shows: (23)
a. prezydlnta zona 'the president's wife' b. zieloniutki ptaszek 'greenish little bird'
2
3
1
b. ?prezydenta samochôd 'the president's car' ?ziiloniutka ptaszyna 'greenish little bird'
Rhythm Rule in Polish
69
It makes no difference if the extra stressless syllable is added at the end of the first word rather than at the beginning o f the second: (24)
2
filozofia :Kota 'Kot's philosophy'
3
I
?matematyka Kota 'Kot's mathematics'
(N .B.: Filozofia is quadrisyllabic.) These facts form part o f a larger argument. The rhythmic interval theory predicts that eurhythmy should be gradient, with a whole range o f values depending on how close the interstress intervals are to the quadrisyllabic ideal. In contrast, the stress clash theory predicts that eurhythmy judgments should fall into two categories, depending on whether a clash is alleviated or not. In ( 2 5 - 2 6 ) , we present series o f examples that can test this hypothesis. (25)
a.
x X X
X
X X X
x
x X X
X X
prezydenta zona s w s w s w
VV V
X
2 1 b. prezydenta samochód 2 1 c. prezydenta fotografìa 2 1 d. prezydenta telewizorek (26)
X
X X X XXX XXX prezydenta zona s w s w s w
'the president's wife'
VV V
2 1 prezydenta samochód 2 1 prezydenta fotografia 2 1 prezydenta telewizorek
'car' 'photo' 'little television'
Bogusiawa Bozek
a. Bogusiawa Bozek 2 1 b. Boguslawa Zagtóba 2 1 c. Bogusfawa Paderewska
2 1 Bogusiawa Zagtòba 2 1 BogusJ-awa Paderewska
d. Bogusiawa Kowalikowna
Bogusfawa Kowalikowna
Native judgments of the applicability of the Rhythm Rule to ( 2 5 - 2 6 ) indeed follow a continuum, ranging from definite plausibility on one end to clear impossibility even in rhythmic speech on the other. This follows directly under the rhythmic interval theory. In (25-26a), the interval shift is from two to four syllables, a clear improvement in eurhythmy, while in ( 2 5 - 2 6 b ) , the shift from three to five moves the text no closer
70
Bruce Hayes and Stanisiaw Puppel
to the target. In (25-26c,d), the Rhythm Rule actually decreases the eurthythmy of the text, by an increasing degree. Thus, the rhythmic interval theory correctly predicts continuous judgments. 5 In contrast, the stress clash theory incorrectly predicts two discrete degrees of acceptability, encompassing (25-26a,b) (in which clash is resolved) and (2526c,d) (where the input form doesn't clash). The predicated but nonexistent boundary is illustrated below with the input grids for (25) b and c: (25)
b.
x *x
*x X
X
c. x
X
X
x x x x x x x prezydenta samochöd
X
X
x x X
x x x x x x x x prezydenta fotografia
The prediction of continuous judgments made by the rhythmic interval theory can be put to a further test: if an input form contains an interstress interval of just one syllable, this should be the most dysrhythmic of all, and the most amenable to relabelling. Polish abbreviations provide a convenient means of checking this prediction, as they regularly bear final stress-cf. PWN ([pevu ? en]) = Panstwowe Wydawnictwo Naukowe 'State Scientific Publishers' or the borrowing BMW ([be 9 emvu]). In (27), the examples under a. undergo the Rhythm Rule more readily than those under b., just as the rhythmic interval theory predicts. (27)
a. BMW Jana 2 i PWN zepsul si?
2
1
b. telewizor Jana
->- BMW Jana 'Jan's BMW' 2 I PWN zepsul si? 'The PWN got worse.' 2
1
-> telewizor Jana 'Jan's television set'
saturator zepsui si? -*• saturator zepsul si? 'The soda water machine broke down.' As before, the stress clash theory is unable to discriminate these examples. In both (27a) and (27b), the Rhythm Rule relieves a stress clash. The examples so far have been based on the expansion of the Rhythm Rule that shifts stress leftward. But the same argument is applicable to the rightward version as well. For example, the cases under (28) show the same increasing reluctance to relabel as the interstress interval is increased in length:
71
Rhythm Rule in Polish (28)
12 3 my poszlibysmy 1 2 3 jabika jedlibysmy 1 2 3 grafikç dalibysmy
my poszlibysmy 'we would have gone' 1 3 2 jabl-ka jedlibysmy 'we would have eaten the apples' 1 3 2 'grafikf dalibysmy 'we would have given the graphic works'
Just as before, the stress clash theory would be unable to discriminate among the data, as a clash is resolved in all cases. In summary, the Rhythmic Interval theory correctly predicts inhibition of relabelling to the extent that the interval of the input is already close to a comfortable size - about four syllables. This gradient prediction cannot be replicated with the coarser notion of stress clash. 2.2 Stretching the Output Interval The Rhythmic Interval theory predicts cases in which the Rhythm Rule is blocked, not because the input interval already has an appropriate size, but because the output interval is too large, being greater than the four syllable optimum. With sufficiently long words, this prediction can be tested: (29)
x
a. X
X
X X X X
X X X
X
X
biblioteka Jana s ws w s w
'Jan's library'
b. tllewizorek Jana
telewizorek Jana 'Jan's little television'
c. telewizoreczek Jana
telewizoreczek Jana 'Jan's little, little television'
As one goes from (29a) to (29c), the propensity to relabel decreases continuously. This is just what the Rhythmic Interval theory predicts: the output dictated by the tree structure is increased from four to six syllables, thus increasing in dysrhythmy and discouraging relabelling. The stress clash theory incorrectly predicts all of these examples to have equal status, as they all involve resolution of a clash.
72
Bruce Hayes and Stanisiaw
Puppel
2.3. Removing Clash In the cases described above, the rhythmic interval theory correctly distinguishes between examples that are incorrectly treated alike under the stress clash theory. We have located one case that goes the other way around, with the stress clash theory predicting a difference ignored by the rhythmic interval theory. The crucial examples are (30-31). In (30), the stress clash theory correctly predicts that the Rhythm Rule will not ordinarily apply, as the marks in boldface are not adjacent in the relevant sense (cf. the diminished grid under (30b)): (30)
a. x X
X
X
x x x
X X X X X X X X
prezydenta fotografia sws w s w s w
'the president's photograph'
\w/ Vs V \/s w \ / w ^ ^
b.
x X
X
X
\ / s x x
Example (31) is the same as (30) in the spacing of the principal stresses. However, it contains a stress clash, as the two main stresses are adjacent on the relevant representation: (31)
a. x
x x
X X X X X X
X
X
X
X X
matematyka samochod s w s w w w s w
V/
w b. X
X
X
'the mathematician's car'
Rhythm Rule in Polish
73
Since it contains a clash, (31) should readily undergo the Rhythm Rule by the stress clash theory. It does not; its status is essentially the same as (30). Once again, it is the rhythmic interval theory that correctly predicts the facts: both examples involve an interval shift from four to six syllables, which is no improvement in eurhythmy. 3. SUMMARY AND COMPARISON WITH ENGLISH
At this point we summarize the arguments that favor the rhythmic interval theory over the stress clash theory for Polish: (32)
a. Judgments of the applicability of the Rhythm Rule are gradient, rather than falling into two categories, and depend on interval size. This implies two things: (i) Propensity to apply the Rhythm Rule increases as the length of the input interval decreases below four syllables. (ii) Propensity to apply the Rhythm Rule decreases as the length of the output interval increases above four syllables. In all cases, presence or absence of clash appears to be irrelevant. b. If a clash is introduced by removing an "intervener" stress, while keeping the interstress interval constant, there is no increased pressure to apply the Rhythm Rule.
We take these facts to argue strongly in favor of the rhythmic interval account. Now what is remarkable about the phenomena we have discussed is that they are almost entirely reproducible in English. Example (33) shows that just as in Polish, input forms of English that clash equally but differ in interstress interval also differ in their propensity to undergo the Rhythm Rule: (33)
a.
x 4-x x x x-2-x x x x x x xx Alabama relatives SWS w s ww
b. acrobatic feats 2
3
!
c. Mississippi Mabel
more natural than
x 5 x x x-3 x xx x x x x x Alabama connections s ws w w s w
more natural than
acrobatic contortions
more natural than
Minneapolis Mike
2
3
1
74
Bruce Hayes and Stanisiaw
Puppel
As in Polish, these examples form only part of a continuum of propensity to relabel. This can be shown by examples such as those under (34). (A useful aid to native intuitions here is to read the examples of (34) consecutively, first applying the Rhythm Rule, than not.) (34)
2 l a. Mississippi abbreviations
2 1 b. Mississippi legislation 2 1 c. Mississippi connections ? i d. Mississippi relatives e. Tennessee relatives
? i Mississippi abbreviations (5 syllables) Mississippi legislation (4 syllables) Mississippi connections (3 syllables) Mississippi relatives (2 syllables) Tennessee relatives (1 syllable)
Observation (32a.ii) was that as the interstress interval of the output is increased beyond four syllables, propensity to relabel diminishes. It can be replicated in English with cases such as the following. (35)
2 1 a. Alabama relatives 2 1 b. Alamogordo relatives 2 1 c. Apalachicola relatives
(36)
2 1 a. telegraphic speech 2 1 b. palatographic measurements c. onomatopoeic verse
Alabama relatives (output 4 syllables) Alamogordo relatives (5 syllables) Apalachicola relatives (6 syllables) vs. telegraphic speech (4 syllables) vs. palatographic measurements (5 syllables) vs. onomatopoeic verse (6 syllables)
Under (32b) we observed that stress clash makes no difference when interstress spacing is held constant. This holds as well of English. For example, there is no greater propensity to apply the Rhythm Rule to (37a) than to (37b), even though (37a) contains a stress clash and (37b) does not. The examples under (38) make the same point.
Rhythm Rule in Polish (37)
75
a. x x x
x
X X X X X X
X
x X
X
Minneapolis connections s w s ww w s w
(38)
X
\l V \l \/ w s w s V \/
1
b. Potawotomi traditions 2
X X X X X X
Minnesota legislation s wswswsw
a. hypothetical connections 2
x x x x x
1
c. recreational facilities
vs.
sympathetic conversations
vs.
Narragansett occupations
vs.
Indonesian capabilities
2
3
2
3
1
1
There are further parallels between the rhythmic behavior of Polish and English. For example, the tendency observed in Note 5 for Polish to form longer sequences into multiple quadrisyllabic intervals is shared by English. Examples like (39) are similar to Polish examples such as (i) under Note 5, while examples like (40) parallel (ii): 2 (39)
^
1
2
Democratic Apalachicola automatic onomatopoeia
(40)
^
1
Democratic Apalachicola -*•
automatic onomatopoeia
2
2 1 twenty-seven Mississippi legislators -> 2 2 1 twenty-seven Mississippi legislators
Another possible parallel involves an additional principle of eurhythmy. Hayes (1984) argues that the eurhythmic target involves not just a single row, but is hierarchical: if the stresses of a phrase have been arranged into quadrisyllabic intervals, then there is further rhythmic pressure to divide the quadrisyllabic intervals evenly with weaker stress beats. This "Disyllabic Rule" can account for a number of phenomena. To give just one example, the rule can explain certain double applications of the Rhythm Rule, as in (41):
76
Bruce Hayes and Stanisiaw Puppel
(41)
a.
x x
x X X
X X
X
X
X
X
X
x x
X
X
a hundred thirteen men s w w s s
c.
b.
x X X
X X
X X
a hundred thirteen men s w w s s
x x x x x x x x x x x a hundred thirteen men s w s w s
In this example, the relabelling of the phrase hundred thirteen satisfies the Quadrisyllabic Rule, and the subsequent relabelling of thirteen satisfies the Disyllabic Rule. Other arguments for the Disyllabic Rule may be found in Hayes (1984). There is at least one aspect of Polish rhythmic phonology that appears to reflect the Disyllabic Rule. Polish words having more than one pre1 2 tonic syllable normally bear an initial secondary stress-cf. telewizorek 2
1
'little television set', Kowalikowny 'Kowalikowna (gen.)'. However, if a word having three pretonic syllables occurs immediately after a stressed monosyllable, then the initial weak stress may optionally appear one syllable later: X x b. (42) a. X X X x X X—X X X x X X X X X X X XX X x sen Kowalikowny sen Kowalikowny 'Kowalikowna's dream' It seems plausible that the function of this stress shift is to accommodate the Disyllabic Rule.6 The Disyllabic Rule may then be a further rhythmic principle shared by Polish and English.
Rhythm Rule in Polish
77
4. DISCUSSION
The similarities between Polish and English rhythmic phonology observed here are noteworthy in that the prosodic systems of the two languages are otherwise quite different. Polish is impressionistically very much a syllable-timed language, with essentially fixed stress, while English is a stress-timed language, having vowel reduction, great variation in syllable durations, and an irregular, semi-free pattern of stress. While it is obviously premature to propose linguistic universals on the basis of resemblances between just two languages, we believe that such resemblances are good candidates to be checked further for universal status, especially when, as in this case, explanations based on genetic relationship or areal diffusion can be ruled out. Accordingly, we offer a conjecture for the universal basis of the English-Polish resemblance. Recall that the Liberman/Prince program for explaining the applicability of the Rhythm Rule factors the phenomenon into two domains: (a) the phonological rule itself; (b) the rules of rhythmic evaluation, i.e. the rules that project grids from tree structure and the rules of eurhythmy. Although we have suggested a substantial revision to the latter rules, we feel that the basic division remains correct (for further argument, see Hayes (1984)). Now it is clear that Rhythm Rules may vary in their phonological form. For example, we have seen that the Polish Rhythm Rule is bidirectional, while English shifts stress only to the left, freely tolerating dysrhythmic configurations that could have been adjusted by rightward shift. The Polish Rhythm Rule also differs from the English one in a lower overall propensity to apply in any given context - although its rhythmic target is the same, it is less willing to distort basic stress contours to achieve that target. Furthermore, Rhythm Rules are not the only rules of rhythmic adjustment. Rules of destressing, length assignment, and "beat addition" also participate in achieving eurhythmic patterns (cf. Hayes (1984), Dell (forthcoming)). Thus, although the degree of interlanguage variation in the purely phonological portions of Liberman and Prince's account is surely not unlimited, the relevant rules must be stated explicitly in the grammars of individual languages, just as with other phonological rules. We believe that for the second class of rules, the situation may be markedly different. While languages differ widely in their basic stress patterns and their rules of rhythmic adjustment, the rhythmic principles that govern these rules - i.e. the rules that project grids from trees, and the Disyllabic and Quadrisyllable Rules - may be universal. The findings of Dauer (1983) bear on this issue in an interesting way. Dauer surveyed a rather diverse group of stress languages, testing whether they have a tendency to place stress at equal intervals. Her measurements revealed
78
Bruce Hayes and Staniskiw
Puppel
that all the languages under investigation showed a tendency towards stress isochrony, irrespective of whether they are "stress-timed" or "syllable-timed". The interstress intervals she found correspond roughly to the target interval we posit for English. Dauer suggests that all stress languages may be stress-timed, and that the traditional distinction between stress timing and syllable timing reflects only the degree to which a language's phonology insists on equal syllable durations. This conclusion is in close accord with our own results: under our view, the principles that determine syllable duration - e.g. syllable canons, vowel reduction, and certain timing rules - are purely phonological, and thus are expected to be language specific to some degree. But the pressure for isochrony of stress lies in the principles of rhythmic evaluation; i.e. in the rules that project grids from trees and in the principles of eurhythmy. As we conjecture that these principles are universal, that we would expect the pressure for isochrony - and more generally, the particular kinds of isochrony strived for - to be independent of the phonological pattern of a language. The comparison of English and Polish is particularly striking in this regard. The two languages lie on opposite sides of the stress-timed/ syllable-timed division, and differ substantially in the form of their Rhythm Rules. Nevertheless, the rhythmic principles that determine stress shift in the two languages appear to be the same. Chomsky (1980) has argued that linguistic competence should be viewed as a separate "mental organ" - that is, as a specialized domain of the mind, whose basic representations and principles of computation have evolved independently, and are not derivable from general psychological principles. We feel that the results of current work in linguistic theory largely support this view. The domain of rhythmic phonology, however, may constitute an isolated exception. As Liberman (1979) points out, the basic principles of rhythmic evaluation at work in phonology (repetition at even intervals, and division of intervals into equal subintervals) are clearly involved in other kinds of rhythmic behavior, such as music or dance. More concretely, Woodrow (1951) has found that experimental subjects listening to series of evenly-spaced, identical sounds often impose a rhythmic structure on them that obeys the Quadrisyllable and Disyllabic Rules. Popular traditions of versification (as opposed to art verse) often adopt binary meters with dipodic structure (cf. Attridge (1982), Burling (1966)). These are precisely the meters that obey the Quadrisyllabic and Disyllabic Rules. Facts of this nature suggest that rhythmic evaluation in phonology may be only a subcase of more general principles of rhythmic behavior. If this turns out to be the case, we will have automatically explained the bifurcation of rhythmic phonology that Liberman and Prince discovered: the rules of eurhythmy must be independent from the phonological rules, as they are not based
Rhythm Rule in Polish
79
on linguistic principles in the first place. In addition, the study of rhythmic phonology will be seen to be a tool for developing a more powerful theory of rhythmic structure in general.
NOTES 1. In all examples, we use numbers to designate rank order of stress prominence; as will be seen below, we deny them theoretical status. The reader unfamiliar with Polish orthography should know two things. First, the letter i before a vowel never forms a separate syllable, but spells either a [j] glide or palatalization. Second, acute accents in Polish indicate special segmental values, not stress. 2. For some experimental confirmation of perceptual isochrony, see Donovan and Darwin (1979), Lehiste (1977). 3. Cf. Spanish (Harris (1983)) and other Romance languages, Chamorro (Chung (1982)), Buhid (Barham (1958)), and English verbs (Hayes (1982)). 4. Dogil (1979) states that in his dialect of Polish certain compounds, such as koscidl protestancki 'Protestant church', are given falling stress, like English ¿levator dperator. In the dialect we are describing these forms don't exist, as compounds uniformly receive rising stress. However, Dogil reports (personal communication) that the results for verbal forms we describe below carry over to compounds in his speech. 5. An interesting pattern emerges when we take this process one step further, testing an interval shift from six to eight syllables. Here, the rhythmic interval theory raises the further possibility of parsing the phrase into two quadrisyllabic intervals instead of one: (i)
a.
x X
X
X
X X
X X X
X X X
X X X X X X X
prezydenta telewizoreczek
the president's little, little television'
b. Bogus+awa Zawalidrozanka And as the theory predicts, it is indeed somewhat easier to apply the Rhythm Rule to prezydenta telewizoreczek and Boguslawa Zawalidrozanka than it is to the slightly shorter forms of (25d) and (26d), prezydenta telewizorek and Boguslawa Kowalikowna. In the latter forms, the trisyllabic interval following the secondary stress in the second word is too short to serve as the isochronous counterpart to a preceding quadrisyllabic interval, so that relabelling is disfavored. Double quadrisyllable intervals can also be created by multiple application of the Rhythm Rule, as in (ii): (u)
2
2
1
Bogustawy zieloniutka torba -»• Bogusiawy zieloniutka torba
'Bogusiawa's greenish handbag'
6. It is not clear to us what specific phonological mechanism gives rise to the shift. The shift is probably related to another phenomenon, the optional appearance of multiple weak stresses in longer Polish words; for example, telewizoreczek 'little, little TV' may be pronounced either with one weak stress (tilewizoriczek), or with two (telewizoreczek). In words with three pretonic syllables, it is at least marginally
80
Bruce Hayes and Stanisiaw
Puppel
possible to place weak stresses on both the initial and the second syllables, as in Kdwalikowny. Our conjecture is that stressings of this sort represent an intermediate stage of the phonological derivation. In the framework of Hayes (1981), they could be arrived at simply by applying the Main Stress Rule iteratively, thus parsing the entire word into maximally binary feet: (i)
Kowalikowny I sws w
1 V
w
v
w
s
\ / s When the word appears in isolation, the medial foot would normally be deleted. This would largely satisfy the Quadrisyllabic Rule following the application of Stray Syllable Adjunction (cf. (iia)). b.
(ii)
X
x
X
X
X
X
X
X X X
X
Kowalikowny s ww s w
X
X X X
X
X X X
X
sen Kowalikowny w sw s w
V, 1/
\/ However, if there is an immediately preceeding stressed syllable, than both the Quadrisyllabic and the Disyllabic Rules may be satisfied by deleting the initial foot, as in (iib). This account seems appealing, but we can offer it only as conjecture, as there are many other facts of tertiary word stress in Polish that are sufficiently elusive that we have not yet arrived at a satisfactory analysis of them.
REFERENCES Abercrombie, D. (1964) "Syllable Quantity and Enclitics in English", in D. Abercrombie et al., eds., In Honour of Daniel Jones, Longmans, Green and Co., London. Attridge, D. (1982) The Rhythms of English Poetry, Longmans, London. Barham, M. (1958) 'The Phonemes of Buhid", Oceania Linguistic Monographs 3, University of Sydney, 4-9. Bing, J.M. (1980) "Linguistic Rhythm and Grammatical Structure in Afghan Persian", Linguistic Inquiry 11, 437-463. Bolinger, D. (1965) "Pitch Accent and Sentence Rhythm", in Forms of English: Accent, Morpheme, Order, Harvard University Press, Cambridge, Massachusetts. Bolinger, D. (1981) "Two Kinds of Vowels, Two Kinds of Rhythm", Indiana University Linguistics Club, Bloomington, Indiana.
Rhythm Rule in Polish
81
Burling, R. (1966) 'The Metrics of Children's Verse: a Cross-Linguistic Study", American Anthropologist 68, 1418-1441. Chomsky, N. (1980) Rules and Representations, Columbia University Press, New York. Chung, S. (1982) "Transderivational Relationships in Chamorro Phonology", Language 59, 35-66. Dauer, R. (1983) "Stress-timing and Syllable-timing Reanalyzed", Journal of Phonetics, 11, 51-62. Dell, F. (forthcoming) "L'accentuation des phrases en français", in F. Dell, D. Hirst, and J.-R. Vergnaud, eds., Les Représentations en Phonologie, Hermann, Paris. Dogil, G. (1979) Autosegmental Account of Phonological Emphasis, Linguistic Research, Inc., Carbondale, Illinois and Edmonton, Alberta. Donovan, A. and C. Darwin (1979) "The Perceived Rhythm of Speech", in E. FischerJ^rgensen, ed., Proceedings of the Ninth International Congress of Phonetic Sciences (Vol. 11), University of Copenhagen, 268-274. van Draat, P. (1912) "Rhythm in English Prose", Anglia 36, 1-58, 492-538. Halle, M. and J.-R. Vergnaud (forthcoming) Three-Dimensional Phonology. Harris, J. (1983) Syllable Structure and Stress in Spanish: A Nonlinear Analysis, MIT Press, Cambridge, Massachusetts. Hayes, B. (1981) A Metrical Theory of Stress Rules, 1980 Doctoral dissertation, MIT. Revised version distributed by Indiana University Linguistics Club. Hayes, B. (1982) "Extrametricality and English Stress", Linguistic Inquiry 13, 227-276. Hayes, B. (1984) 'The Phonology of Rhythm in English", Linguistic Inquiry 15, 33-74. Kiparsky, P. (1966) "Über den deutschen Akzent", Studia Grammatica 7,69-98. Lehiste, I. (1977) "Isochrony Reconsidered", Journal of Phonetics 5, 253-263. Liberman, M. (1979) The Intonational System of English, Garland Press, New York. Liberman, M. and A. Prince (1977) "On Stress and Linguistic Rhythm", Linguistic Inquiry 8, 249-336. McCarthy, J. (1979) Formal Problems in Semitic Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Massachusetts. Phinney, M. (1980) "Evidence for a Rhythm Rule in Quebec French", Proceedings of the Tenth Meeting of the North East Linguistic Society, Cahiers Linguistiques d'Ottawa 9, University of Ottawa. Pike, K. (1945) The Intonation of American English, University of Michigan Press, Ann Arbor. Stowell, T. (1979) "Stress Systems of the World, Unite!" MIT Working Papers in Linguistics 1, 51-76.
Formal Properties of Stress Representations* Jack Hoeksema State University of
Groningen/ZWO
1. BACKGROUND
In the 1960's several attempts were made at formalising phonological theory, resulting in systems as diverse as those of Peterson and Harary (1961), Batog (1967) and Chomsky and Halle (1968). Of these works, the most impressive is the last. SPE, as it is commonly referred to, contains a lengthy discussion of English word stress. Batog (1967), which is probably the most thorough in its development of the formal foundations, does not include a discussion of suprasegmental or prosodic phenomena, although the phonological theory he sets out to formalise, Harris' (1951) structuralistic account, is by no means silent about such issues. In this respect, there is a marked difference between SPE and Batog's little book. Batog himself remarks: "There are two things absent from our system: the theory of junctures and the theory of suprasegmental elements of utterance. The reason of our omitting them is that both of them are still immature and they require a more full elaboration by linguists themselves. In their present state they are not fit for logical analysis and formal treatment." (op. cit., 120)
Chomsky and Halle made an attempt in SPE to treat suprasegmental phonology as segmental phonology, that is, they extended the feature framework to treat suprasegmental phenomena like stress, and segmental phenonema in a unified manner. In so-called standard generative phonology, which is the phonological tradition originating in SPE, this has been one of the leading ideas. A rather sharp reaction to this segmental approach in the mid-1970's has given rise to metrical phonology, which employs several types of hierarchical representation to deal with suprasegmental phonology. The metrical theories that now dominate the field seem to agree that the SPE policy of representing stress as segmental * This research was partially supported by the Foundation for Linguistic Research which is funded by the Netherlands Organization for the advancement of pure research, ZWO.
84
Jack Hoeksema
features is basically incorrect and that an entirely different prosodic theory is needed for the proper treatment of stress and other suprasegmental phenomena, like vowel harmony or tone. In a sense, then, modern metrical theory appears to agree with Batog's assessment of the state of suprasegmental phonology in the 1960's. This state of affairs in contemporary phonology raises the question which this paper tries to answer: Just how different are the 'flat' segmental representations of standard generative phonology and the 'hierarchical' ones of metrical theory? My aim is to show that segmental representations using numerical features can be defined that are isomorphic with metrical trees and metrical grids (cf. Liberman and Prince 1977 on these notions). This implies that every rule which can be stated for metrical trees or metrical grids, can be stated for numerical representations in the style of SPE as well.
2. NUMERICAL REPRESENTATIONS AND METRICAL GRIDS
In the notation of SPE, stress patterns are indicated in the phonological representation by assigning natural numbers to the vowels. SPE uses the interpretation convention that when the stress goes up, the numbers go down. So a vowel with the stress feature [1 stress] is more heavily stressed than a vowel with the feature [2 stress]. To make the comparison with metrical grids a little easier, we will adopt the somewhat more natural convention that the numbers go up when the stress goes up in this section. (Of course, this convention does not make any difference for the representations at hand, since we are interested in their form, not in their interpretation.) Formally, we will define numerical stress representations as strings of ordered pairs. The left hand member of each ordered pair will indicate a syllable, the right hand member will indicate the stress level of that syllable. Here we deviate from SPE, but since we may assume some percolation convention which sends the stress number of a syllable to its vocalic nucleus, this does not really matter. The stress levels are taken from some interval I = [1 ,n] of N, the set of natural numbers. Definition 1 Fix a set S of syllables. Let I be some interval [1 ,n] of N. Then we define a numerical stress representation as a string over S x I. The notion of a numerical stress representation as defined above is of course a very crude notion. For phonological purposes, many such representations are uninteresting, so further restrictions on the set of possible
Formal Properties of Stress Representations
85
numerical stress representations are called for. Later on, we will encounter some of these futher restrictions. Before considering these, we will first consider the related notion of a metrical grid. What metrical grids are, is usually made clear by giving examples of diagrams of particular metrical grids. Here is an example from Prince (1983: 21):
(1)
Diagram of a metrical grid x x x x x x x Jim saw her in
x x x x x the park
In such diagrams, the height of the column of marks above each syllable indicates in a pleasant visual way its stress level. It is immediately clear that the nuclear stress of this sentence is on the syllable park, and that Jim is more heavily stressed than saw. In Liberman (1975: 280) a formal definition of the notion of a metrical grid is given, which is presented below in definition 2 in a somewhat simplified form:
Definition 2 A metrical grid is a structure (£,F), where £ is an ordered set of ordered sets (Lj)i < j < n and F is a function mapping each member of L m + i onto some member of L m in an order preserving way: if Cj and £j are members of L m + i , then F(Cj) is ordered before F(fij) just in case Cj is ordered before £j. Actually, Liberman states an additional requirement on well-formed metrical grids, which we will consider shortly. In Liberman's definition, F provides the columns of the grid: for two elements are in the same column on adjacent levels iff the function F maps one of them onto the other. The ordered sets Li, on the other hand, provide the horizontal rows. For the purposes of stress theory, Liberman's definition is too liberal, because it enables one to distinguish between grids that should be equivalent. For example, one could distinguish all of the following four metrical grids:
86 (2)
Jack Hoeksema
x New
X
New
x x York
y New
y y York x
y x York
y New
y York
Let us therefore adopt a more restrictive definition of metrical grids, which allows them less expressive power.
Definition 3 A proper metrical grid is a structure (X,*), where X C A x I, I as in def. 1, A a string of syllables, and * the ordering of X derived from the ordering on A in the following way: (a,m)*(b,n) iff a comes before b in the string A. We further require that (a,n) in X if (a,n+l) is, and that for every a in A: (a, 1) is in X. To see the motivation behind this definition, note that a diagram such as (1) above is completely specified by giving (1) the labels of the rows, (2) those of the columns and (3) the horizontal and the vertical specifications of the marks. It is easy to see that A provides the horizontal coordinates of the diagram, I the vertical coordinates (number the rows 1 to n from the bottom to the top) and X those pairs of coordinates on which there is a mark. By way of example, consider the following proper metrical grid for the expression New York: < ( (New,l), (York,l) (York,2) I , {(New.l) *(York,l), (New,l)*(York,2) | > . In the form of a diagram, this becomes:
(3)
Diagram of a proper metrical grid x New
x x York
2 1
It is important to note that the distinctions in (2) above cannot be expressed in the present formalisation, since the marks only indicate whether a particular pair of coordinates is a member of X. Therefore we have only two options available in the diagram for each pair of coordinates: either a mark or a blank. Proper metrical grids are rather redundant. Therefore we want to consider a less redundant structure.
Formal Properties of Stress Representations
87
Definition 4. A reduced proper metrical grid is a substructure R of a proper metrical grid G, such that (x,n) in XR iff no pair ( x , m ) , where m is less than n, is in XG, and *R is the restriction of *G to XR. A reduced proper metrical grid gives only the maximal members of each column in the grid's diagram. Those are the only entries we need t o know: all lower entries are in the non-reduced grid as well, and all higher entries are of course not in the non-reduced grid.
Fact 1 On reduced proper metrical grids, the ordering * is linear. To see this, note that ( x , m ) and (y,n) are unordered in a metrical grid only if x = y. In reduced metrical grids, no pairs ( x , m ) and ( y , n ) exist, such that x = y, and m + n. Next we will consider the relation between reduced proper metrical grids and numerical stress representations. This relation turns out to be extremely straightforward:
Fact 2 (X,*) is a reduced proper metrical grid if and only if it is a numerical stress representation. According to this fact, the notions of a reduced proper metrical grid and a numerical stress representation coincide. To see this, just note that any string A can be written as a structure (X,*), where X is the set of A's elements and * is the linear ordering of the string (i.e. a*b means that a comes before b in A). Since it is clear that we can construct a unique reduced proper metrical grid for every proper metrical grid and reconstruct the proper metrical grid from its reduced counterpart, we can safely state the following result:
Fact 3 The notions of a numerical stress representation, a proper metrical grid and a reduced proper metrical grid are notational variants. In connection with definition 2, it was mentioned that Liberman's definition of a metrical grid is in fact somewhat more complicated than stated there. Liberman requires that grids like (1) above are ruled out by definition. He states the requirement that if two marks are adjacent on some level, they should be separated by at least one and at most two marks on the next lower level, if such a level exists. In diagram ( I ) , the marks
88
Jack
Hoeksema
for Jim and saw are adjacent on the second level, yet the marks for these two syllables are adjacent on the first, the lowest, level as well. Similarly, saw and park have adjacent marks on level two, but in this case the number of intervening marks on the next lower level is too high: 3. In the formalism of proper metrical grids, Liberman's requirement can be formulated as follows: Definition 5. An alternating proper metrical grid is a proper metrical grid G, such that if (x,n+l) and (y,n+l) in X q and there is no (z,n+l) such that (x,n+l)*(z,n+l) and (z,n+l)*(y,n+l), then we have in X q at least one and at most two (z,n), such that (x,n)*(z,n)*(y,n). For numerical stress representations, a similar restriction can be formulated : Definition 6. An alternating numerical stress representation is a numerical stress representation such that if (x,m) and (y,n) in the string, 1 < m < n, and there is no (z,k) between (x,m) and (y,m) such that k > m, then we have at least one and at most two (z,k) such that (z,k) between (x,k) and (y,k) in the string and k = m - 1. By inspection of these definitions, we have: Fact 4. (X,*) is a reduced alternating proper metrical grid iff it is an alternating numerical stress representation. This fact is completely analogous to fact 2. As a corollary, we further have: Fact5. Alternating numerical stress representations, alternating proper metrical grids and reduced alternating proper metrical grids are notational variants. Just how useful alternating proper metrical grids are as a tool for describing stress patterns, is not clear. We saw that a reasonable grid description like (1) above is outside the scope of alternating grids. In fact, even such pleasant alternating grids as the one in (4) below are ruled out:
Formal Properties of Stress Representations (4)
x x x a
x x b
x c
x x d
x e
89 x x
x x f
x g
x h
x i
3 2 1
Grid (4) is ruled out because there are three elements between the marks for a and i on level 2, although the marks for a and / are adjacent on level 3. For the purposes of linguistic theory, it seems better to drop the strong alternation requirement on proper metrical grids (in Prince (1983), the best paper so far on the role of metrical grids in the description of stress phenomena, no such requirement is made). On the other hand, ordinary proper metrical grids allow stress descriptions that never occur in anyone's stress theory. For example, grids like the one in (5) below should be prohibited :
(5)
A stupid grid x x x a
x x x x x x b c
3 2 1
So it seems reasonable to state at least the following well-formedness condition on metrical grids:
Definition 7. A good grid is a proper metrical grid with the property that m n whenever (a,m) and (b,n) are adjacent and maximal members of X. Grid (5) is not a good grid, since (a,3) and (b,3) are adjacent and maximal in their column, yet their right hand members are equal. The same is true for (b,3) and (c,3). The corresponding notion for numerical stress representations is stated in the following definition:
Definition 8. A good numerical stress representation is a string of pairs (x,n), such that for any two adjacent pairs (y,m) and (z,k): m k. It is reasonable to require that every stress representation be good in the sense defined above. It is doubtful whether a stronger condition should be postulated, to wit, that any two syllables should have different
90
Jack
Hoeksema
stress levels. Call a stress representation ideal when it conforms to definitions 7 or 8, but with the adjacency condition dropped. We will see in the section about metrical trees that tree representations are ideal stress representations. A nice consequence of ideal representations is the following: Fact 6. Ideal stress representations have a unique syllable bearing maximal stress ('designated terminal element'). It seems desirable that every stress description contains a unique designated terminal element bearing the main stress. It is not clear, however, that the formalism of the representation should take care of this. Since we need stress rules as well, to specify the position of the main stress in particular languages, we have another mechanism at our disposal. On the other hand, we might state in our general definition of metrical grids that they should contain a unique designated terminal element, without having recourse to ideal metrical grids. Definition 9. A pointed metrical grid is a good grid with the property that it contains a pair (x,n) such that for every (y,m): m < n. Call this pair its point. It is obvious that the point of a metrical grid is its designated terminal element. Definition 10. A pointed numerical stress representation is a good numerical stress representation such that there is a pair (x,n) in the string such that for every (y,m): m < n. Summary. In this section, we have considered a great many definitions of metrical grids. Of these, good grids and pointed grids appear to be most useful for linguistic purposes. For every type of grid, there is an equivalent numerical stress representation. These numerical stress representations are identical with reduced grids. So we might conclude that numerical stress representations of the types considered above are just less redundant versions of metrical grids. As a corollary we have that every rule statable on metrical grids, is statable on the corresponding numerical stress representations as well.
Formal Properties of Stress Representations
91
3. METRICAL TREES AND NUMERICAL STRESS REPRESENTATIONS
Today, the most popular way of representing stress is drawing a so-called metrical tree. Now what is a metrical tree?
Definition 11. A metrical tree is a rooted binary branching tree, whose root is labeled by R, whose other nonterminal nodes are labeled by S or W, and whose terminal nodes are labeled by syllable symbols. Furthermore, if a node is labeled W (S), then its sister is labeled S (W). Here is an example of a metrical tree: (6)
R S
W
s
w
s
w
ten
nis
play
er
In these trees, S means 'Strong' and W means 'Weak'. From (6) we learn that tennis has a heavier stress than player, since the node dominating the string tennis is S, and the node dominating player is W. Likewise, in player, the first syllable has a heavier stress than the second. Naturally, the syllable that is dominated by S only is heavier than all other syllables. In (6), this syllable is ten. It is an immediate consequence of definition 11 that there is always a unique syllable dominated by S only.
Fact 7. The designated terminal element of a metrical tree is the terminal node that is dominated by S and R only.
Fact 8. Every metrical tree has a unique designated terminal element. These two facts turn out to follow directly from a more general proposition about metrical trees:
Fact 9. Metrical trees are ideal stress representations.
92
Jack Hoeksema
Ideal stress representations were defined in the previous section as stress representations in which no two syllables have the same stress values. We will say that two syllables in a metrical tree have identical stress values iff their t o p - t o - b o t t o m paths are labeled identically. For example, the t o p - t o - b o t t o m paths in ( 6 ) are labeled RSS, RSW, RWS and R W W . So in ( 6 ) no two t o p - t o - b o t t o m paths are labeled identically. We will now derive fact 9 by proving the following claim:
Fact 10. Let X and Y be two different t o p - t o - b o t t o m paths in a metrical tree T . Then X and Y are labeled differently. Proof: Let X be ( x j , . . , x n ) , where x j , . . , x n are (non-terminal) nodes in T and Y be ( y p - . y , ^ ) - For any node x, let v ( x ) denote its label. Since T is rooted, X j = y j . Let k be the maximal number such that ( x j , . . ^ ) = ( y j v M y ^ ) - Since x j = y j , k is at least 1, and since X ¥= Y , k is less than n. The string ( x j , . . ^ ) is the common part o f X and Y . N o w suppose that X and Y are labeled identically. Then for every i, v(x^) = v ( y j ) . Therefore
v
(xk+i)
=
v ( y j c + | ) . We know that x ^ + j and y ^ + j are sister
nodes, since they have a common immediate dominating node: x^ = y^. But then one of them should be labeled W and the other S, by definition 11. This contradicts our assumption that v i x ^ j ) = v i y ^ j ) . Since the latter assumption follows from our initial assumption that X and Y were labeled identically, this initial assumption must be false. Q.E.D. Another
elementary property
o f metrical trees has been stated in an
earlier paper o f mine (Hoeksema 1982):
Fact 11. Any metrical tree T is completely characterised by the l e f t - t o right ordering o f its t o p - t o - b o t t o m paths. For instance, if T is the metrical tree in ( 6 ) , then the ordered set o f its top-to-bottom
paths can be written as ((ten, RSS), (nis, RSW), (play,
RWS), (er, R W W ) ) . N o t every tree is completely characterised by the set o f its t o p - t o bottom paths in the linear order o f its terminal elements. Take for example the following tree:
(7)
A B
B
b
b
Formal Properties of Stress Representations
93
This tree has the following string of vertical paths: ((b, AB), (b, AB)). This string, however, is not unique for (7), since the tree in (8) below has exactly the same string of vertical paths: (8)
A B b
b
These examples do not contradict fact 11, however, since both (7) and (8) are ruled out as metrical trees. Let us now prove fact 11. We will do this by proving a somewhat stronger claim. We say that a phrase structure grammar G has the PathTree Property iff every one of its derivation trees is uniquely characterised by the linear string of its vertical paths. Here is the claim we will prove:
Fact 12. A context-free phrase structure grammar G has the Path-Tree Property if every production rule has the form A -*• x j ...x n , where xj # x i + 1 for every i such that 1 < i < n. Proof: Let T be any tree, then its path-representation p(T) is the string ( ( a j , X j ) , ... , ( a n , X n )), where a j . . a n is the terminal string of T and Xj is (v(xj),..., v(x n )), where Xj..x n is the string of nodes dominating aj, such that Xj immediately dominates X j + j . We have to show that the mapping p is one-to-one on the domain of trees generated by a grammar of the form specified in fact 12 above. In other words, if p ( T j ) = p C ^ ) , then T j = T2. We will prove this by showing how T can be recovered from p(T) in a unique fashion. First note that for every Xj in p(T), its first element is identical to that of any other Xj, to wit, the root of T. We will write this first element as f j ( X j ) . In general, f„(X) is the n - t h element in the string X. So f j ( X j ) = f j ( X 2 ) ... = f , ( X n ) = the root of T. Now we try to find the daughters of this root. For every Xj, going from 1 to n, draw a new node labeled by f2(Xj) iff f2(Xj) ^ ty i + 1 ^ ' t o t ' i e right of the previously drawn daughters of the root. However, if f2(Xj) = f 2 ( X | + j ) , then these symbols must correspond with the same node in T, because two adjacent nodes labeled identically are ruled out by the condition on G in fact 12. For every new node established in this fashion, we can find its daughters etc., until the whole tree T is recovered. Since the instructions to draw new nodes are unambiguous, fact 12 follows. Q.E.D.
94
Jack Hoeksema
In order to show that fact 12 implies fact 11, we have to show that any metrical tree is generated by a context-free grammar of the type specified in fact 12. From the definition of metrical trees, it is immediately clear that any such tree is generated by a context-free grammar employing the following production rules: (9)
PS-rules for metrical trees
x->swiws
X -»• x (where X = R,S or W and x is a syllable symbol) Inspection of these rules suffices to see that they are of the required type. Let us now turn to the main question of this section: Can metrical trees be written as numerical stress representations? In other words, is it possible to define a class of numerical stress representations that is equivalent to the class of metrical trees? The answer to this question turns out to be rather simple, once we use the concept of a path-representation. In this respect, a path-representation plays the same role here as reduced grids did in the section about metrical grids. First I will sketch a rather attractive mapping from metrical trees to numerical stress representations which I have presented in my above mentioned paper, Hoeksema (1982). 1 Definition 12. Let f be a mapping from |S,W,R| * into { 1,0^ | *, where G indicates the empty string, such that f(S) = 0, f(W) = 1, and f(R)_= £. For strings(xj.,x n ) we define f(xj..x n ) = (f(xj)...f(x n )). Let X, where X is any string(xj..x n ), indicate its reverse ( x n . . x | ) . If T is a metrical tree, then its associated numerical stress representation n(T) = ((a,, f(Xj)),.., (a n , f(X n )), where ((a,, X , ) ( a n , X n ) ) = p(T). What we get by this mapping is a stress representation in binary numbers. To get ordinary numerical stress representations, we only have to translate the binary numbers into decimal numbers. By way of example, we will derive the numerical representation of the tree in (6). First, we form its path-representation: ((ten, RSS), (nis, RSW), (play, RWS), (er, RWW)). Next, we turn the paths around and substitute 1 for W, 0 for S and delete R: ((ten, 00), (nis, 10), (play, 01), (er, 11)). Finally, we translate the binary numbers into the decimal system: ((ten, 0), (nis, 2), (play, 1), (er, 3)). Notice that this numerical representation is different from the ones considered earlier, in that the numbers go down when the stress goes up. If we keep that in mind, we will see that the
Formal Properties of Stress Representations
95
mapping is quite appropriate, since it correctly assigns the main stress to ten, and it assigns heavier stress to play than to nis. It is useful to compare the mapping of definition 12 with the algorithm given in Liberman and Prince (1977: 259) for the conversion of trees into number representations. Before doing so, we will first state the following proposition:
Fact 13. The mapping from metrical trees into numerical stress representations described in definition 12 is one-to-one. This property is very important, for it guarantees that we always can recover a tree from its numerical representation. If that were not possible, then the two representations would not be equivalent. Since fact 13 immediately follows from definition 12, a proof is not necessary. Let us now consider the mapping given by Liberman and Prince. Here is the slightly more elegant formulation of this mapping by Prince (1983:22):
(10)
Stress numbering For any terminal node a, determine the first W that dominates a. Count the number of nodes that dominate this W. Add 1. This is the stress number of a.
This mapping is certainly not one-to-one. To see this, consider the following two metrical trees: (11)
a.
R
b.
R
The stress numbering system of Liberman and Prince treats these trees as equivalent. Each has the following numbering: ((a,2), (b,4), (c,3), (d.l)). In contrast, our numbering system would yield different numerical representations for these trees: the associated numerical stress representation of tree (1 la) is ((a, 1), (b, 5), (c, 3), (d, 0)), and that of tree (1 lb) is: ((a,l), (b,7), (c,3), (d,0)). Notice that our stress representations,
96
Jack Hoeksema
though different from those of Liberman and Prince, agree with theirs in the relative values of the stress numbers. More precisely, we claim that: Fact 14. For any two terminal nodes x,y in a metrical tree: if x has a higher stress number than y under the Liberman and Prince mapping, then x has a higher stress number than y under our mapping. Recall that under the Liberman and Prince mapping x has a higher number than y iff the path from the lowest W dominating x upward to the root is longer than the corresponding path for y. But the length of these paths is equal to the length of the stress numbers in binary in our mapping, as we may skip the O's coming before the first 1. So if x has a longer binary stress number than y, then evidently x has a higher decimal number than y. Let us now compare the metrical tree representations with the structures studied earlier, metrical grids and numerical stress representations. From fact 13 we know that we can define a class of numerical stress representations that are equivalent with metrical trees. That does not imply, however, that trees and numerical representations in general are notational variants. One important difference between grids and numerical representations on the one hand and metrical stress representations is that there is only a finite set of metrical trees for some string X, but an infinite number of grids and numerical representations. In fact, the number of metrical trees for a given string X is a function of the length of X. In order to make a fair comparison we will assume that there is some upper bound to the number of grids or numerical representations that one may associate with a given string. The following requirement seems natural: Definition 14. Call a metrical grid well-behaved iff it has at least as many columns as rows. According to this definition, a metrical grid is well-behaved if the number of stress levels it can distinguish is not higher than the number of syllables in the string. Likewise, a well-behaved numerical stress representation on a string X can be defined as a string over XxN, such that for any (x,n) in the string n < |X|. Even if we restrict our attention to well-behaved good grids, there are still many more such grids than metrical trees. For instance, on a string of four elements, there are 40 possible metrical trees, but 108 wellbehaved good grids. Only in the case of well-behaved ideal grids, is the num-
Formal Properties of Stress
Representations
97
ber less than the number of trees: 4! = 24. But if we require that each tree has uniform [S W] or [W S] labeling or obeys the condition that a right node is strong iff it branches, or the condition that a left node is strong iff it branches, we are left with only 18 different trees (cf. Halle and Vergnaud 1978 for discussion of these further restrictions on the labeling of metrical trees). Now we must face something of a dilemma. Should we prefer metrical trees to grids or numerical representations just because there are less of them (if we ignore the 'ideal' representations for the moment)? On the other hand, we might need more distinctions than the tree theory provides us with. For example, if we want to make a distinction between 7 stress patterns and tree theory allows for just 4 patterns, then we have to make arbitrary assignments for at least 3 of the patterns in question. The issue is further complicated by the fact that tree theory may be supplemented by more non-terminal vocabulary (cf. the introduction of feet-symbols in Selkirk 1980). A drawback of tree theory is that constituency is often rather arbitrary. As this point has been stressed already by several people (e.g. Prince (1983), Van Zonneveld (1982)), I will not dwell on it here. In this respect both numerical representations and metrical grids are superior, because they are not committed to the existence of the constituents in question. Another point in favor of grids and numerical representations is that they need not be ideal stress representations in our technical sense. For the unprejudiced observer, it appears to be possible that two non-adjacent (or even adjacent) syllables in a string have the same stress level. Trees deny this possibility, as do all other ideal representations. An example that comes to mind is the stress pattern in a Dutch word like genade /xana.da/ 'mercy'. Here the two shwa's appear to be equally weak. Finally, it should be noted that relative stress is often a very opaque notion in tree theory. For example, it is impossible to tell whether b is stronger than c in the following tree or vice versa: R
(11) S
a
W
b
c
d
In this tree, it is quite obvious that a is stronger than b, that c is stronger than d, that ab is stronger than cd, for that is what the labeling tells us.
98
Jack Hoeksema
What reason is there to suspect, however, that the weak daughter of a strong node, like b, is stronger than the strong daughter of a weak node, like c? One has to invoke some numbering convention, such as the one of Liberman and Prince considered above. However, as Prince (1983: 22) puts it, there is nothing inherent in the tree system that would lead to the particular rank ordering of terminals entailed by this numbering convention. So this convention involves an auxiliary hypothesis of considerable complexity and indeed - from the metrical point of view - arbitrariness. To phrase it differently, if trees were really good representations of stress, then we would not need any auxiliary representations such as numerical representations or grids, unless, perhaps, these representations would follow directly from the tree notation. Taken together, these arguments against the use of metrical trees in the description of stress are quite compelling to my mind. What exactly will turn out to be a good representation of stress, I do not profess to know. Here I agree with Bruce Hayes,2 who remarked that there are theories of stress employing grids, and ones employing trees, as well as theories employing both, but the correct theory will probably employ neither.
NOTES 1. There is an obvious reformulation of the mapping for those tree theories which allow labels other than R, S and W (cf. Selkirk 1980). If « is the number of labels in the theory under consideration, then we will translate the paths into w-ary numbers. Again, the mapping is o n e - t o - o n e . 2. In his talk at the ZWO-Workshop on Nonlinear Phonology, Amsterdam, August 8 1983.
REFERENCES Batóg, T. (1967), The axiomatic method in phonology, London, Routledge, Kegan Paul Ltd. Chomsky, N. and M. Halle (1968), The Sound Pattern of English, New York etc, Harper and Row. Halle, M. and J.-R. Vergnaud (1978), 'Metrical structures in phonology (a fragment of a draft)', MIT. Harris, Z.S. (1951), Methods in Structural Linguistics, Chicago, University of Chicago Press. Hoeksema, J. (1982), 'Enkele Stellingen omtrent metrische representaties'. in TABU 12-2,86-94. Liberman, M. (1975), The intonational system of English, diss. MIT, published by Garland Publishing Inc., New York and London, 1979. Liberman, M. and A. Prince (1977), 'On Stress and Linguistic Rhythm', Linguistic Inquiry 8-2, 249-336.
Formal Properties of Stress
Representations
99
Peterson, G.E. and F. Harary (1961), 'Foundations of phonemic theory", in: R. Jakobson (ed.), The Structure of Language and its Mathematical Aspects, Providence, Rhode Island, Am. Math. Society, pp. 139-165. Prince, A. (1983), 'Relating to the Grid', Linguistic Inquiry 14-1, 19-100. Selkirk, E.O. (1980), T h e role of Prosodic Categories in English Word Stress', Linguistic Inquiry 11-3, 563-605. Zonneveld, R. van (1982), 'Met de ritmischc hangmat in the metrische boom', TABU 12-2, 68-85.
Stress, Stress Shift, and Morphology: the case of Dutch -baar* J a n G. Kooij and Mies van der Niet Department of General Linguistics of State University of Leiden
0. SUMMARY
A test case for the evaluation of non-linear explanations in phonology is the interaction between prosodie structure and morphological structure, in particular, the (re)organization of stress patterns in derived words. It appears from the extensive literature on the subject that a number of questions in this area are still unanswered, and that issues originally raised in the context of linear phonology keep coming up. We therefore wish to discuss in some depth the nature of stress shift in Dutch adjectival derivations, in particular, complex verbal stems suffixed with the productive suffix -baar, '-able'. We will show that relocation of main stress in these adjectives is primarily a prosodie phenomenon where the morphological distinction between Root suffixation and Word suffixation plays no role.
1. From SPE onwards it has been claimed that derivational affixes can be subcategorized according to their behavior under stress assignment. In the area of suffixation, to which we will restrict ourselves, this has led to the well-known distinction between neutral and non-neutral suffixes. A suffix is non-neutral if word-level stress rules apply to a derived form essentially as they apply to a non-derived form, with stress shift as a possible consequence. A suffix is neutral if the application of stress rules to a derived form disregards the presence of the suffix as far as the major stress pattern is concerned. Within the SPE framework, the difference is represented by distinct boundary symbols between stem and suffix: + for non-neutral suffixes, and # for neutral suffixes. Typical English examples are:
* Research for this paper was, in part, carried out for the Project 'Prosodic structures' of the Faculty of Letters at the University of Leiden. We are grateful to participants at the Symposium on Non-Linear Phonology for their comments, and to Anneke Neijt also for showing us the unpublished paper 'Attractive -baar'.
102
Jan Kooij and Mies van der Niet
(1)
parent
parent+al parent#hood
Typical Dutch examples would be: (2)
vijand^
'enemy'
vijind+igA vijand#schap^
'hostile' 'enemy-hood'
The distinction between non-neutral, or, Class I suffixes and neutral, or, Class II suffixes, is supposed to carry over to other aspects of morphological structure. A central claim in work by Siegel (1974) and Allen (1978) is that Class I suffixes do not follow Class II suffixes in derived forms, and this generalization is still an important guideline in recent studies of morphological structure, e.g. Kiparsky (1982) and Selkirk (1982). Nevertheless, it can be demonstrated that much of the original argumentation in favor of the distinction is inadequate, both phonologically and morphologically. Detailed criticism can be found in Strauss (1979); here, we will only summarize three major difficulties that the distinction would cause in a description of Dutch morphology. (i) As long as location of main stress in a derived form is the decisive criterion, some suffixes are unavoidably members of both classes at the same time. In Dutch, this would be the case with the suffix -boar, the equivalent of English -able, which derives adjectives from verbs and is the subject matter of this paper. 1 In the derivations of the type (3a), main stress shifts towards the end of the stem, but in derivations of the type (3b), it does not: (3)
a. b. klemtoonjyj 'stress'
uitvo£ry 'perform' beklemtodny 'stress'
uitvoer-baar^ 'perform-able' beklemtoon-baar^ 'stress-able'
The example (3a) is a verb-particle compound with a separable particle; in independent clauses, the inflected form of the verb would be 'voer ... uit'. The example (3b) is based on a nominal compound and formed with the inseparable transitiving prefix be-. The different locations of main stress in the derived adjectives may be either due to the difference in morphological structure of the verbal stem, or they may be explained by a difference in prosodic structure, as we will show later on. Since both types of adjectives are productive and transparent formations, it does not seem to be very attractive to conclude that -baar is a Class I suffix in (3a) and a Class II suffix in (3b).
103
Stress Shift and Morphology
(ii) If Class II suffixes are defined by their neutrality under stress assignment, the class is heterogeneous by definition. It would, for instance, contain a number of [+native] suffixes that behave as elements of compounds in various respects: they receive secondary stress, they can undergo rules of Gapping (see Booij, this volume, and Kiparsky (1975) for the German cognates), and the one suffix in this group that begins with a vowel, the suffix -achtig, '-like', does not syllabify with the stem, as vowel-initial suffixes usually do (see Kooij (1980), for more discussion). Examples of compound-like derivations without stress shift are given under (4): (4)
uitzicht
'prospect'
uitzicht-lods
vijand
'enemy'
vijand-schap
potlood
'pencil'
potlood-achtig
'prospect-less desperate' 'enemy-hood, hostility' 'pencil-like'
On the other hand, Class II contains suffixes that cannot be stressed and that prosodically integrate with the preceding stem. Examples are the nominalizing suffix -ing and the agentive suffix -er: (5)
uitvoer afzend
'perform' 'send, dispatch'
uitvoering afzender
'performance' 'sender'
The often made claim (repeated in Selkirk's recent monograph) that Class II suffixes typically do not syllabify with the preceding stem is clearly untenable for Dutch. Phonological generalizations of this nature would lead to a classification of 'weak' and 'strong' suffixes that cuts right across the distinction between non-neutral and neutral suffixes. The weak, adjective-forming suffix -ig, for instance, causes stress shift (see (2)), but the equally weak nominalizing suffix -ing does not (see (5)). (iii) Generally, the stress criterion unites suffixes that are otherwise quite different and splits up categories of suffixes that can be identified on independent phonological and morphological grounds. A case in point are the strong suffixes of Dutch that receive secondary stress in derivations. There is no shift of main stress in the examples under (4), above, but stress shift does occur in a number of formations with the suffix -heid, '-ness' as well as in a number of formations with the suffix -baar, '-able'. These observations all point to the conclusion that stress shift, and stress assignment in general, is a property of derivations, and not a defining property of classes of suffixes. The picture that is presented by Dutch derivational processes confirms this:
104
Jan Kooij and Mies van der Niet
(a)
[+native] derivational as well as flectional suffixes are predominantly stress-neutral (or, extrametrical in the sense of Hayes (1982)); shift of main stress towards the last stressable syllable of the word preceding the suffix is more typical for derived adjectives than it is for other types of derivation (Schultink (1979)); reapplication of stress rules where the original stress pattern is destroyed or disregarded2 and where, furthermore, main stress often goes on the suffix itself, is typical for the Dutch Romance vocabulary:
(b)
(c)
(6)
a. úitvoer V inlèid^ b. úitvoer^ inlèidy
c. Amérika N relígie N
'perform' 'introduce' 'perform' 'introduce' 'America' 'religion'
uitvòeringj^ inlèidendp art ùitvoérbaar^ inleídendy Àmerikàan N rèligieûs^
'performance' 'introducing' 'perform-able' 'introductory' 'American' 'religious'
We do not deny that it is important for morphology to distinguish between more central and more peripheral derivational processes, and that the distinction between Class I suffixation and Class II suffixation can be used to express this difference. Also, the observation that stress shift often characterizes derivations that are more lexical or more opaque is correct; compare, for instance, the words inleidend, 'introducing', and inléidend, 'introductory' under (6), above. But the conclusion that these and other instances of relocation of main stress can all be explained by one and the same morphological distinction is unwarranted. To fully appreciate what happens at the interface of morphological and phonological structure, we will need a more sophisticated theory of morphology and we will also have to investigate prosodie structure in more detail.
2. Recently, several authors have tried to represent morphological structure in a different, non-linear fashion without giving up the difference between Class I suffixation and Class II suffixation or the Affix Ordering Generalization formulated by Siegel and referred to above. One such attempt is Selkirk's study on the syntax of words (1982). Essentially, Selkirk claims that there are two types of word formation rules for English suffixation: a rule of the type Word -*• Word Affix, and a rule of the type Root -»• Root Affix (where 'Word' replaces the notion 'Stem' used in an earlier version of her study). The distinction between boundaries has been replaced by a difference in morphological constituency. Class I affixes
105
Stress Shift and Morphology
attach to stems -using 'stem' in a traditional sense -that are Roots, and the resulting derived form undergoes the lexical stress rules of the language. Class II suffixes attach to stems that are Words, and these derivations are not subject to the lexical stress rules. The difference between the formations parental and parenthood, mentioned above, is represented as follows: (7)
a.
Word
b.
Root Root parent
Word Word
Affix
ll
I I parent
Affix
Root
hood
The rule Word -* Root connects the two levels of representation. Selkirk also allows for re-analysis of an original Word-formation as a Rootformation in those cases where a Class I suffix follows a Class II suffix in violation of Siegel's Affix Ordering Generalization. One example is the English derivation govern - government (Class II) - governmental (Class I). For (native) compounding, Selkirk proposes a rule Word -*• Word Word. The rule predicts that only those suffixes that are introduced by the rule Word -> Word Affix can be attached to compounds. It is this type of derivation that raises a principled issue, discussed at some length in Williams (1981) and also by Selkirk. Since quite a number of Dutch derivations with -baar are based on compound verbs (compare (3), above) we have a good reason to go somewhat more deeply into this aspect of Selkirk's proposal. Quite obviously, the prediction that only Word-forming suffixes can attach to compound stems is incorrect as it stands. Consider the Dutch noun Zuidamerikaan, 'South American', derived from Zuid-Amerika, 'South America'. The prefix Zuid- in Zuid-Amerika receives secondary stress and the noun, furthermore, is subject to Gapping, cf. phrases like Noord- en Zuid-Amerika, 'North America and South America' are grammatical in Dutch. This noun is either a Pref Word formation or it is a Word Word compound. For the present discussion, the difference is not crucial, we only must make sure that Zuid-Amerika is a Word formation. Accepting the position that it is a Pref Word derivation explains that the noun is stressed w - s and not s - w as would be predicted by the regular Compound Stress Rule. Semantically, Zuidamerikaan is a compositional function of the meanings of the complex noun Zuid-Amerika and of the suffix -aan which derives geographical names. Zuidamerikaan means 'person from South America' and not 'southern American'. In an informal fashion, we can represent this by the following bracketing:
106
Jan Kooij and Mies van der Niet
(8)
N(N(Zuid
Amerika) N aan) N
But it is also clear that the semantic structure is at odds here with morphological structure. For all purposes, including stress assignment, Dutch -aan is a Root suffix and, in addition, it receives main stress (compare (6c), above). The suffix attaches to Amerika and the resulting stress pattern is Zuidamerikaan .The bracketing of the derived structure is rather as in (9): (9)
j^[Zuid[Amerikaan]]^
Selkirk discusses similar examples in English (1982:106 ff) and accepts the position of Williams (1981) that the derived structure of these formations is right-branching. This implies that she also accepts Williams' definition of 'lexical relatedness' which states that in a configuration (10a) where Z is the Head of the derivation, X-Y is lexically related to X - Y in the configuration (10b) although X-Y is not formally a constituent in (10a). (10)
a.
\
b.
We will not discuss here the notion of relatedness, but we believe that Williams' claim about the structure of derivations from compounds or compound-like Word formations is fundamentally correct. It seems to be the only reasonable interpretation of the position that derivation 'precedes' compounding, and, as we shall see, it also makes the correct prediction about the prosodic structure of these derivations. Notice, however, that under the assumptions of Selkirk's analysis, the derivation Zuidamerikaan must be a Root at the lowest level of its structure since derivations with -aan change the original stress pattern. At the higher level, the derivation must remain a Word formation. Contrary to what orthography may suggest, the prefix Zuid- in Zuidamerikaan is still an independent element and the derivation can still be gapped the phrase Noord- en Zuidamerikanen, -North Americans and South Americans', is also perfectly grammatical. We can now represent the derived morphological structure as (1 la) and the derived prosodic structure as ( l i b ) .
Stress Shift and Morphology (11)
a.
107
W Aff
b. W R
R Zuid Amerika
w Aff aan
Zuid
Ameri
s kaan
In ( l i b ) , w and s designate 'weak' and 'strong' and 2 stands for 'foot',as usual. At the lowest level, dominated by R in the structure (1 la), the rules for Romance derivations will assign the correct stress pattern to Amerikaan. At the higher level, dominated by W in (11a), the stress rules correctly predict that the major stress pattern in this type of formation is w - s. Details aside, the derived stress pattern is Zuidamerikaan. It can be seen from this example that hierarchical representation of derived morphological structure has clear advantages over linear representation. The separate application of Root level and Word level stress rules ensures the correct assignment of stress, and boundary symbols can be entirely dispensed with. The additional assumption that derivations from compound stems as a rule produce right branching structures makes it possible to translate derived morphological structures adequately into derived prosodic structures. At the same time, however, it can be demonstrated that shift of main stress to word final position that typically occurs in Dutch derived adjectives cannot be uniformly explained by the assumption that these derivations are Root formations, or, that it can be so explained only at the expense of unnecessary and unmotivated complications. In order to see why, consider again the derivations uitvoering, 'performance' and uitvoerbaar, 'perform-able', referred to earlier under (3) and under (5), respectively. Both derivations are based on a complex verb with a separable particle, uitvder, 'perform'. The verb must be minimally characterized as a Word formation; since the stress pattern is s - w as it is in regular Dutch compounds, we will assume that uitvder is generated by the rule Word -»• Word Word. Complex verbs with prefixes that are not separable from the stem in inflected forms and, furthermore, have a stress pattern w - s will be generated by the rule Word -*• Pref W; an example would be the verb overwin, 'conquer'. Again, nothing too important is involved in making the distinction, as long as both verbs are generated as Word formations, but we assume that it is an adequate way of dealing with the different behavior of the prefixes and with the stress difference. We will have to allow for a third type of complex verb: verbs that are prefixed with weak and dependent affixes like be- in bespreek, 'discuss' are most probably Root formations in Selkirk's analysis.
108
Jan Kooij and Mies van der Niet
The nominalizing suffix -ing in uitvoering is stress neutral, as observed, but it is also a weak, unstressed suffix that prosodically integrates with the preceding stem. Extending Williams' proposal for adequacy, the derived structure uitvoering must be right branching though -ing would be a Class II suffix in the given framework. We can represent the derivation as in (12a), and the derived prosodic structure as in (12b). At the lowest level, voering is a foot, branching s - w by convention; at the higher level, the major stress pattern remains s - w because there is no shift of main stress in -ing nominalizations.
The derived prosodic structure uitvoerbaar would pattern along the same lines: a right branching structure where -voer and -baar are separate feet and where -baar, being a suffix, cannot be strong. Since, however, main stress is on -voer, -voer must be s in an s domain, which means that, at the higher level, the stress pattern s - w must be reversed to w - s. But the question now is how to ensure this reversal within a framework that distinguishes between Word formations and Root formations. Recategorization of voerfbaar) as a Root, comparable to the recategorization of Amerika(aan) in (11a) would not give the correct result; as long as uitvoerbaar is a compound at the higher level, the stress pattern would, wrongly, be s - w: úit-voerbaar. Minimally, we will have to assume that uitvoerbaar is restructured as a Pref W formation with a w - s stress pattern as in overwin, 'conquer' -overwinbaar, 'conquerable'. There is one fundamental objection against this analysis, namely, that it lacks independent motivation; the morphological restructuring of uitvoerbaar is forced upon the analysis by the facts of stress assignment. The same objection applies, and even more strongly, if we were to assume that the entire formation uitvoerbaar is reanalyzed as a Root formation. Under standard assumptions about the role of Root suffixes in derived forms, this analysis, too, would correctly predict final stress (provided that -baar itself, being a heavy syllable, is extrametrical). But the analysis would leave the question why uitvoerbaar is a Root formation wide open. As long as absence or presence of stress shift is the primary motivation for the choice of a particular morphological structure, these problems seem to be inescapable. In this particular case, choice of category at the
Stress Shift and Morphology
109
lower level determines choice of category at the higher level, but similar problems arise if we look at the question the other way around. Consider the nominalization bespreking, 'discussion', derived from the verb bespreek, 'discuss'. A comparison of the various types of complex verbs in Dutch favors an analysis of bespreek as a Root formation, as noted above. Consequently, the affix -ing in bespreking attaches to a Root, though considerations of stress would classify -ing as a Word-level affix because it is stress-neutral. The second objection against an analysis where suffixation with -baar leads to morphological recategorization is more obvious: quite a large class of compound verbs are suffixed with -baar without relocation of main stress. While the existence of these counterexamples cannot decide against the Root analysis, it certainly can be taken as an indication that something else is going on. All in all, the difficulties connected with the Root analysis of -baar formations are enough reason to propose a different analysis. This analysis is based on the following assumptions: (i) (ii)
suffixes that are phonologically strong are never Root suffixes; in [+native] derivations, stress shift is explained by rule.
For the derivations under discussion, we propose the following rule: (13)
In [[X Y]y Z ] ^ , X Y a compound,Z -baar,main stress is relocated on the last stressable syllable in Y.
The effect of the rule can be represented as the change from (14a) to (14b):
In the next section, we will discuss rule (13) in more detail, and show that the change in stress pattern that it predicts also occurs in other derivations with strong suffixes.
3. As we briefly observed in section 1, shift of main stress towards word final position is a typical feature of derived adjectives in Dutch. It may
110
Jan Kooij and Mies van der Niet
well be that the tentative rule (13) that we propose is actually a subcase of a more general rule: 3 (15)
In [[Y]x Z]^, main stress is relocated on the last stressable syllable in Y.
This would account for the location of main stress in a variety of derived adjectives, compare: (16)
alcohol vijand onzin wetenschap
'id' 'enemy' 'nonsense' 'science'
alcohólisch vijándig onzínnig wetenscháppelijk
'alcoholic' 'hostile' 'nonsensical' 'scientific'
Rule (15) would not be without exceptions and classes of exceptions, and it would take us too far afield to discuss the desirability of a uniform account of stress shift in derived adjectives. Rather, we will concentrate on the derivations with -baar and, for expository purposes mainly, accept the existence of the specific rule (13). Notice that -baar has to be mentioned in the structural description because there is no stress shift in adjectives derived with the suffix -bos, '-less'. The -baar derivations are an interesting case precisely because -baar is a strong suffix. If we accept Williams' theory that suffixes attach to compounds by creating right branching structures, strong suffixes of the type exemplified under (4), above, seem to be the exception. Compare, for instance, the derived prosodic structure of the formation uitzichtloos, 'prospect-less, desperate' with the derived prosodic structure of uitvoerbaar.
Though intuitions about these structures are not always clear, it is hardly to be doubted that the realization of uitzichtloos is [ [s w] s w] and not [s[s w] w ]. The contrast between (17a) and (17b) is thus twofold: relocation of main stress on voer entails a reversal of the morphologically determined stress pattern s - w, and the resulting structure is right branching rather than left branching.4
Stress Shift and Morphology
111
Consider next the relevant examples of -baar adjectives derived from complex verbal stems: (18)
a. uitvoer uitschakel óverdraag b. uitprobeer ónderverhuur c. uitéenplaats d. terugvorder
'perform' 'switch off 'transfer' 'try out' 'sublease' 'put apart' 'reclaim'
e. overwin f. beklémtoon
'conquer' 'stress'
uitvóerbaar uitschâkelbaar overdrâagbaar uitprobeerbaar ónderverhuurbaar uitéenplaatsbaar terugvorderbaar, terugvórderbaar overwinbaar beklémtoonbaar
Stress shift occurs only (and to our knowledge, exceptionlessly) in the derivations of the type (18a), where the compound stem consists of two elements that are single feet, branching or non-branching. The constraint on the application of rule (13) that is exemplified by (18b) as well as by (18c) can be summarized in the following condition (Van der Niet (1982); Van Zonneveld (1983)): (13)'
... Unless X or Y or both branch w - s.
In the case of uitprobeerbaar, the explanation is straightforward. The verb probeer has a lexically designated final stress: probeer, try'. Whether we represent the derived structure as left branching ((19a)) or as right branching ((19b)), the foot preceding -baar is s, whereas in uitvoerbaar, prior to the application of rule (13), it is w (compare (14a), above). (19)
b.
a.
S s w beer baar
2w 2s beer baar
It is more difficult to explain the examples (18c). Both in a left branching structure and in a (maximally) right branching structure, plaats would be w.
112
Jan Kooij and Mies van der Niet
Location of main stress on plaats, however, would cause a stress clash, w s s w, where s is main stress, and would have to be resolved by a reversal on uiteen, yielding swsw,uiteenplaatsbaar. Though this pattern is acceptable in itself, the stress clash apparently blocks the application of (13) to these forms. The example (18d) confirms the prosodic analysis in a subtle way. If the particle, terug, is realized w - s, rule (13) fails to apply, and terugvorderbaar patterns like uiteenplaatsbaar. If the particle is realized as a monosyllable after contraction, trug, the rule can apply since the derivation now patterns like uitvoerbaar. The conclusion is that application of the proposed rule (13) is constrained by features of the prosodic structure of the derivation. We may actually consider the possibility that instead of (13) we have a rule like (21): (21)
In [[X Y]y Z]^, X Y a compound and Z -baar, main stress is relocated on the last stressable syllable in Y iff X and Y are single feet.
This automatically includes derivations of the type overwinbaar in (18e), where the rule would apply vacuously. It would exclude derivations of the type beklemtoonbaar in (18e) under the assumption that beklemtoon is not itself a compound. Notice that a formulation of (21) that would not exclude the example (18e) on morphological grounds, could still exclude it on the grounds that the leftmost element in the structural description must not branch w - s. Whatever the correct generalization, it is sufficiently clear that shift of main stress in complex verbs suffixed with -baar is constrained (or possibly, triggered) by features of prosodic structure. The possibility that stress shift in these derivations is the exception rather than the rule, and that its motivation is entirely prosodic rather than morphological should not be excluded.5 It is interesting, in this connection, to consider the derivations in (22) that all involve sequences of phonologically strong suffixes:
Stress Shift and Morphology (22)
a. werkloos vindingrijk uitzichtloos b. systeemloos kritiekloos
'unemployed' 'resourceful' 'desperate' 'systemless' 'critiqueless'
113 werkloosheid 'unemployment' vindingrijkheid 'resourcefulness' uitzichtloosheid 'desperateness' systeemloosheid 'systemless-ness' * systeemloosheid kritiekloosheid 'critiqueless-ness' *kritiekloosheid
As often observed by Dutch grammarians, main stress shifts in complex formations suffixed with the strong suffix -heid, '-ness'. The pattern in the first two examples in (22a) is exactly the same as in uitvoerbaar: main stress is relocated on the last foot preceding the suffix -heid, and the major stress pattern is reversed, [[s w]w] becomes [w[s w]]. The same reversal occurs in the third example in (22a): [[[s w]w]w] becomes [s w] [s w]s- In the examples under (22b), a lexically strong foot precedes the syllable that is the focus of the rule and no shift occurs; these derivations pattern like the example uiteenplaatsbaar discussed above. Stress shift occurs optionally in derivations like nakomeling, 'offspring' - nakdmeling, and in compounds like kinderbijslag, 'children allowance' - kinderbifslag. The obvious similarity between the derivations in (22) and the derivations in (18) indicates that Dutch simply has a few phono logically strong suffixes, among them -boar and -heid, which are lexically marked for the fact that they trigger a relocation of main stress in particular prosodic contexts, some, it appears, obligatorily, and some optionally.
4. The main purpose of this paper has been to show that stress shift in Dutch derivations cannot be uniformly explained by connecting it with the difference between Word formations and Root formations and the concomitant distinction between suffixes that are stress-neutral and suffixes that are not stress-neutral. Actually, we believe that this type of explanation is adequate for the Romance vocabulary only. 6 In the [+native] vocabulary, stress shift is governed by rule and, to an extent, determined by morphological class; in that respect, the observed relocation of main stress in Dutch derived adjectives is comparable to the difference between English permit^ and permit y. We have also shown that derivations with phonologically strong suffixes form a separate class and that, in these derivations, relocation of main stress is largely determined by features of prosodic structure and rhythmical organization. This raises interesting and relatively unexplored questions about the autonomy of prosodic structure in its interaction with morphological structure.
114
Jan Kooij and Mies van der Niet
NOTES 1. Actually, -able is an ambiguous suffix too, as pointed out by Selkirk (1982) and, earlier, by Aronoff (1976). 2. There are strong arguments for the position that stress assignment in the Romance vocabulary of Dutch is not cyclic; for a more detailed discussion, see Kooij, forthcoming. 3. The idea that stress shift in Dutch adjectives derived from complex verbs is primarily a matter of prosody is not new. It was also proposed in Schultink (1979), following suggestions by the Danish linguist Hammerich. We do not agree with Schultink's description of stress shift in uitvoerbaar as a 'reaction to stress clash', but we dö accept his viewpoint that the prosodic restructuring observed in these derivations is a much more general phenomenon of Dutch complex words. 4. Neijt and Zonneveld (1981a, 1981b) explain the reversal of the main stress pattern by the assumption that -baar forms a foot with the preceding syllable. This seems to be an unnecessarily strong assumption; it derives directly from their hypothesis that -baar is a Root suffix, which we reject. In Zonneveld (1983), the idea that -baar is a Root suffix is abandoned. 5. Notice, incidentally, that in the exact German couterparts of these formations, with the suffix -bar attached to compound verbs with separable particles, stress shift does not apply. For the large majority of productive derivations of the type ausführbar, German pronunciation dictionaries give an s - w - w stress pattern; the same applies to German derivations like Wissenschaft - wissenschaftlich, 'science - scientific', where Dutch has wetenschap - wetenschdppelijk. The argument is, admittedly, indirect, but it is still true that the difference between the two languages is easier and more elegantly explained by the assumption that German lacks a rule of the type (13) than by the assumption that -bar is a Word suffix in German but a Root suffix in Dutch. 6. A number of adjectives derived from, especially, nominal compounds do not exhibit shift of main stress, contrary to expectation. Compare: düizend-pöot, 'centipede (lit. 'thousand-leg')' - düizend-pdtig, 'centipede-like'. The picture is futher obscured by the fact that in some formations with the suffix -ig and with its near synonym -erig, stress shift is optional. AU in all, it seems quite surprising for Dutch grammarians to have accepted so readily the idea that stress and re-allocation of stress is a reliable diagnostic for morphological structure, in spite of the fact that irregularities abound and have been pointed out quite frequently. For more data, compare Schultink (1979), and earlier publications mentioned there.
REFERENCES Allen, M., 1978. Morphological investigations. Doctoral diss., University of Connecticut. Aronoff, Mark, 1976. Word Formation in Generative Grammar. Cambridge Mass.: The MIT Press. Hayes, Bruce, 1982. 'Extrametricality and English Stress'. LI 13, 227-276. Kiparsky, P., 1975. 'Stress, Syntax, and Meter'. Lg 51, 576-616. Kiparsky, P., 1982. 'From cyclic phonology to lexical phonology'. In: H. van der Hülst and N. Smith (eds), The structure of phonological representations I, p. 131-175. Dordrecht: Foris. Kooij, Jan G., 1980. 'Morpheme Boundaries and Syllable Boundaries: A Case for
Stress Shift and Morphology
115
Natural Phonology'. In: W. Zonneveld, F. van Coetsem, and O.W. Robinson (eds), Studies in Dutch phonology, p. 61-78. The Hague: Martinus Nijhoff. Kooij, Jan G., forthcoming. 'Stress, Rhythm, and Morphology'. Paper read at the Symposium on Rhythm in Language, Literature, and Music, Bielefeld, December 1983. Neijt, Anneke and W. Zonneveld, 1981a 'Attractive -baar\ Unpublished paper Dept. of General Linguistics, University of Utrecht. Neijt, Anneke and W. Zonneveld, 1981b. 'De aantrekkingskracht van -baar'. Glot 4,215-228. Schultink, H„ 1979. 'Reacties op 'Stress Clash". Spektator 8,195-208. Selkirk, Elizabeth, 1982. The Syntax of Words. Cambridge Mass.: The MIT Press. Siegel, Dorothy, 1974. Topics in English Morphology. Doct. diss., Cambridge Mass. Strauss, S.L. (1979) 'Against boundary distinctions in morphology'. Linguistic Analysis 5, 387-419. Van der Niet, Mies, 1982. 'Accentaantrekking in afleidingen'. Unpublished thesis, Dept. of Dutch, University of Leiden. Van Zonneveld, Ron, 1983. Affix-Grammatica. Dordrecht: Foris. Williams, Edwin, 1981. 'On the notions 'Lexically Related' and 'Head of a Word". LI 12,245-274. Zonneveld, W., 1983. 'Lexical and Phonological Properties of Dutch Voicing Assimilation'. In: M. van den Broeckc, V. van Heuven, and W. Zonneveld (eds), Sound structures;studies for Antonie Cohen, p. 297-312. Dordrecht: Foris.
Intonation: a Whole Autosegmental Language Carlos Gussenhoven State University of Nijmegen
1. INTRODUCTION
In the account of the structure of English intonation presented below the assumption is made that the prosodie component of intonation languages and the component of syntax and lexis are parallel subsystems. These linguistic subsystems are assumed to be organised in similar ways, as illustrated in Figure 1. LINGUISTIC EXPRESSED
OPTIONS BY
NON-SEGMENTAL MEANS
SEGMENTAL MEANS
INTONATION
SYNTAX/LEXIS
Figure 1. Organisation of the linguistic system as two subsystems
The model assumes that an intonation language like English employs two phonetic resources to encode its morphemes. The left-hand subsystem is encoded in Fo variations against time, the right-hand subsystem
118
Carlos Gussenhoven
in spectral composition variations against time (non-segmental and segmental encoding, respectively). The subsystems are otherwise on a par, with outputs of the two subsystems being mapped onto each other at the end of the respective derivations. The theoretical advantage of this parallel view is that questions like 'What is the function of intonation?' (i.e. is it grammatical, attitudinal, does it mark information weight, etc.) resolve as pre-theoretical issues to which no answer is possible other than by giving a statement of the lexis, syntax, semantics and phonology of the prosodic system. Indeed, one would not seriously try to answer such broad questions as 'What is the function of the subsystem of syntax and lexis?'. Rather, such questions are narrowed down to specific rules of syntax or specific (sets of) lexical items, or specific phonological (variable) rules. Another advantage of a representation like that in Figure 1 is that other aspects of sound structure that have been treated in an autosegmental framework, but which are not of the same order as the phenomena we are dealing with here, can be relegated to their proper, separate position in the model. Tone in SE Asian languages, consonantal roots in Arabic, vowel harmony in languages like Turkish and Finnish and nasal prosodies in Guarani form part of the segmental structures they are mapped onto in a way that intonation does not. In some of the above cases, the autosegmental tier captures purely phonological features of a 'segmental' morpheme. For example, a Chinese morpheme like ma ('mother') cannot be decomposed into a segmental element [ma] and a tonal element [level tone] without breaking up the minimal set of phonological features that define the morpheme, i.e. without underspecifying it. The tonal feature, in other words, functions in the phonology of the 'segmental' half of the linguistic system. In an intonation language like English, by contrast, the tonal characteristics of an utterance 'Mother!' and its segmental features constitute separate morphemes, as is readily evident from the fact that the morpheme mother can be pronounced with a number of different pitch contours without underspecifying it, as well as from the circumstance that these different pitch contours are independently invested with meanings (Gussenhoven 1983). In other cases, the autosegmental tier is used to represent morphemes in their own right, as in the case of Bantu tone languages, many of which employ Fo variations for this purpose. However, tonal morphemes in tone languages always figure as bound morphemes in segmental morpheme paradigms, like 'aspect' in the verb phrase. Rather than constituting a fully-fledged subsystem, these morphemes are integrated in the 'segmental' subsystem. By contrast, tonal morphemes in intonation languages cannot be associated with any morpheme paradigm that is also segmentally encoded, and therefore cannot be said to be integrated in the 'segmental' subsystem in the way Bantu tones are.
Intonation
119
Within either phonological component there may be justification for representing certain aspects of sound structure autosegmentally. Arguably, in the non-segmental phonological component features like downdrift (the gradual downward slope of the - imaginary - baseline of intonation contours), 'speaking up' (the realisation of a contour in a higher area of the available pitch span, Liberman & Pierrehumbert 1984), high onset (the high beginning of unaccented syllables, such as high-pitched a m A book?) and range (the peak height of a tone) might be interpreted as being autosegmental with respect to the tone elements making up the tones of English discussed below. One consequence of this mapping of the outputs of two separate phonological components would appear to be that it is no longer very clear just what the P-bearers are onto which P-elements are mapped. For the purposes of mapping tones in the 'segmental' phonological component, it seems fairly clear that the tone segments are the P-elements and that syllables are the P-bearers. When intonation contours are mapped onto texts, however, it would not be realistic to say that a tone element is associated with a syllable. Although (1) would be a convenient formulation of the modification DELAY (see below), it is not really true that the modification is governed by syllable structure, as suggested by the representation in (1). (1)
&
a
Rather than forcing it onto the next syllable, application of DELAY causes the association target to be moved to the right by a measure that would seem to correlate with the intensity of the meaning of DELAY, and it could in fact end up both to the right and to the left of the syllable peak of the syllable after the accented one. This gradual nature of the modification DELAY has in fact been incorporated in the phonetically fairly detailed descriptions of British English intonation (e.g. Kingdon 1958: 135). For ease of exposition, however, we will represent mappings as being between tone elements and syllables, as in Figure 2. Observe that this circumstance in fact provides a further indication that the mapping of tones in tone languages and the mapping of intonation in intonation languages are different processes. Not only do we find that association targets can be meaningfully shifted through the text without tolerating constraints from the phonological structure of the text, there is also evidence of meaningless fluctuation, or instability, of association targets. An example is provided by the final tone element of
120
Carlos Gussenhoven
fiLH, as discussed in section 4. As far as I know, such instability does not occur in tone languages. It may for this reason be preferable to abandon the term 'association' altogether in the case of intonation, and with Ladd 1983 - to speak of 'alignment'. This article concentrates on the structure of the lexicon of tones in English. This lexicon is organised in terms of free and bound morphemes, or tones and modifications, respectively. The discussion attempts to leave semantic aspects out of account. However, in order to appreciate some of the arguments used below, we need to point out that the modifications DELAY, HALF-COMPLETION and STYLISATION are viewed as representing points on a semantic continuum of 'specialness': DELAY signals 'specialness' or 'non-routineness', STYLISATION the opposite, while HALF-COMPLETION occupies a position in between, lying closer to STYLISATION than to DELAY. The article also discusses two rules for linking tones. It will be argued that the description makes it possible to distinguish between intonational differences that are low-level phonetic and those that are the realisation of meaningful linguistic options.
2. TONES
The English tone system comprises three tones, ftL, ftLH and L h . In (2) their association with the word unicorn is illustrated. (2)
unicorn HL
TiWo/n HLH
¡y
5nj6orn
LH
The starred element is the first in all three tones. 1 This implies that the choice of a tone does not determine the Fo specification of the text to its left. This seems the desired result. The pitch of It's a in It's a unicorn can vary from high to low in all three cases, as illustrated in (3). (3)
It's a unicorn
It's a unicorn
It's a unicorn
The pre-accentual Fo specification would appear to constitute a paradigm which is independent of the tone paradigm. A neutral value for English here is 'mid' (Willems 1981), but this fact does not justify the inclusion
Intonation
121
of a M tone element in the tonal words, as in Goldsmith (1976) and Leben (1976). Association conventions are straightforward. Mappable structures assembled in the two phonological components are to remain intact, and mapping is therefore exhaustive. The segmental domains onto which the tones are mapped stretch from a 'starred' syllable to the next 'star' or to the end of the tone group. We need to state that only the second tone element of a tone can spread. This will cause the starred element to be linked to the accented syllable, and the final H in HLH to the final syllable. If the text is only one syllable long, all tone elements will be associated with that syllable. This, of course, captures what is surely a commonplace in intonation description. In the case of £H, a special provision is necessary to the effect that the H is incremental: in a longer stretch of unaccented syllables, each tends to be higher than the one before, as in (4a). Alternatively, the increment is restricted to the last syllable (cf. Pierrehumbert's boundary tone H%, 1980), as in (4b). Note that our notation makes explicit that the difference between (4a) and (4b) is non-systemic, i.e. of a low-level nature.
Wh$ did you say telephoned? (4b)
WhS did you say telephoned?
3. MODIFICATIONS
One of the important features of the description of the English tonal paradigm presented here is that the three tones are exhaustively crossclassified with Modifications. That is, no statement of a relationship between two tones should be possible without at the same time stating what the modification is that they have in common, or without specifying what the third tone from the set of three looks like under the same modification. Modifications are bound morphemes, and as such cannot occur without a tone. Modifications are phonetically implemented by means of phonological operations on the tone elements of the three tones. These operations are not restricted to deletion or alteration of tone elements,
122
Carlos Gussenhoven
but may also concern the time domain as well as affect the association conventions given above. 3.1 Delay The modification DELAY exclusively operates in the time domain: the tone elements themselves as well as the associations remain unaltered. The only statement that is necessary is that the starred element is associated at a point to the right of the starred syllable (usually the next syllable), which leaves the starred syllable itself unspecified. A neutral value here would appear to be 'low', but, as in the undelayed case^that specification is independent of tone choice proper. The delayed HL is the rise-fall of other descriptions, the delayed HLH is the rise-fall-rise, while delayed LH would not appear to have been noted in the literature as a tone (variant) separate from the unmodified l u . In Figure 2 associations are illustrated with It's a unicorn as the text.
UNMODIFIED
w
It's a unicorn
It's a
*
HL DELAYED
It's a unicorn
//
It's a
*
STYLISED
CRESCENDOED
It's a
unicorn
III
unicorn *
HLH
HL HALF-COMPLETED
unicorn
V
LH
It's a unicorn
II
HLH
LH
It's a unicorn
It's a
HL - HM
HLH - HMH
LH -
It's a unicorn
It's a
unicorn
It's a unicorn
HL — MtM4-
HLH -
MtLH
LH - M
It's a unicorn
It's a
unicorn
It's a unicorn
17
V
w *
HL
unicorn
Vh
w
HLH
V
LM
P7
*
w
LH
Figure 2. Cross-classification of three tones and five modificatory states with It's a unicorn as the text
Intonation 3.2
123
Half-completion
HALF-COMPLETION exclusively operates in the domain of pitch span. The common phonetic denominator is the confinement of the Fo trajectory after the starred tone element below or above the midline. In effect, this causes the second tone element in a tone to be rewritten as M. Association and timing are as in the unmodified case. Half-completed falls have been noted in the literature as 'suspended falls' (Crystal 1969) and, for Dutch, as (one type of) E-falls ('t Hart & Collier 1975). A halfcompleted fall-rise is described in Crystal (1969) as a 'high, narrow fallrise', while, again, no separate mention is made of half-completed t u . This tone is readily recognisable in a word like cowardliness, in which cow- would be associated with £ and -ardliness with M, forming a mid plateau. In the unmodified case, -ardliness would have a rising contour, as noted above. This tune-text combination could occur in the following example: (5)
It is not difficult to mention a few undesirable qualities: laziness, intemperance, intolerance, cowardliness... *.M * LH->
3.3 Stylisation STYLISATION was established as a modification by Ladd (1978), who also identified the meaning 'routine' for this morpheme. Ladd set up the 'call contour' or 'vocathrechant' (Liberman 1975) and the level tone as stylised variants of the fall (HL) and rise (Î-H) respectively.2 Stylised AL consists of two level stretches somewhat above and below the midline respectively, stylised I n of a single mid-level stretch only. The modification is clearly phonetically more complex than the preceding two. A process of fusion must be postulated in order to account for the level stretches:
L *
*
• .
and H fuse completely, to M, while H and L fuse partially, to MtMI. Variable lengthening of the syllables of the text onto which the tones are mapped is a frequent accompanying feature. In addition, a special association statement is required: the starred tone element is subject to spreading. This necessarily goes for M, the only tone element in its word. The starred tone element of MtMI spreads to any [-stress] vowels, except the last. Note that this gives the desired results in all cases (cf. the characterisation of this tone in Liberman 1975, 20-21): in cSwardliness the starred Mt gets associated with the first three syllables, leaving the last for Ml. In Abercrombie, Mt is associated with the first two syllables,
124
Carlos Gussenhoven
since spreading is blocked by the [+stress] [o]: the last two syllables are left for Ml. In a case like jShn, both tone elements must be associated with the one syllable, causing it to be 'broken up' into two sections to accommodate the two elements. The fact that potentially more than one tone element in a tone can spread first brings up the question of direction of association. On the basis of at least some informant responses, it would appear that English conforms to the supplement to Goldsmith's Well-formedness Condition: mapping proceeds from left to right. These infomants associated the Ml of stylised fiL with table book rather than with book when asked to pronounce cSffee table book with stylised HL, which suggests that spreading of Mt proceeds from the left until the first [+stress] vowel is encountered, leaving the remainder of the text for Ml. Our framework suggests that there must be a third stylised tone, the stylised HLH. Basing our requirements on the characteristics of stylised tones given above, stylised HLH should typically cause lengthening of all syllables,^have fused Mt, and have spreading apply as in the case of the stylised HL. The missing tone can best be called up by imagining a child opening a conversation with its governess, as follows: (6)
Na:: y:: nny:
When this tone is mapped onto texts like cowardliness and Abercrombie, spreading of Mt takes place as in the case of the stylised HL. Observe how the mid-level plateau created by stylisation of £ H produces a tone which, in the case of a text with many unaccented syllables after the accented one, resembles half-completed . The only difference is that in the latter case the accented syllable gets associated with a low tone element and in the former with a mid tone element. This clear difference in linguistic structure should of course correlate with a difference in meaning, if our description has any validity. As was observed above, STYLISATION enhances the effect of routineness relative to HALFCOMPLETION. This is precisely what seems to be the difference between the tone given in (6) and the one given in (7). (7)
We were supposed to be able to mention all the undesirable qualities listed in his book: laziness, intemperance, intolerance, cowardliness .... I forget the fifth.
£h->1V1
Intonation
125
Secondly, observe how the two calling contours given by Leben (1976) fall out of our analysis precisely where they should. Leben notes that train conductors in Canada may be heard to call out (8)
Next stop Ottawa! H MM
to announce the next stop, rather than HHM, his notation for the 'vocative chant'. Example (8) is a half-completed fall in our analysis, which so happens to be shouted. A stylised fall (Leben's HHM) would be inappropriate in this context, as it would infelicitously emphasise the meaning 'routine'. Note, too, how lengthening would be much more natural with 'HHM' than with 'HMM', another prediction made by our description. 3.4 Crescendoing A fourth - and extremely rare - modification is CRESCENDOING. CRESCENDOING is realised by a gradual increase of the Fo change that characterises the tone: the movement (falling or rising pattern) begins very slowly and tends to accelerate its rate of change towards the end of the text. A crescendoed fall is given in (9). •
(9) HZ
*
Unmodified HL
Crescendoed HL
320-
240
160
80
0
16
¿8
64
'
80
96
'
112
128
1U
'
160
CSEC
The auditory effect of crescendoed tones is enhanced by a lengthening of the final syllable, which carries the brunt of the Fo change. It will be clear that the manner in which the starred tone element enlarges its scope on the segmental tier differs from the straight spreading observed for stylised tones above. The tone element is rather 'dragged along' through the segmental tier. The crisscross representation of the associations in Figure 2 is an attempt to represent this feature.
126
Carlos Gussenhoven
Another difference with stylised tones is the manner in which lengthening is applied: STYLISATION allows free lengthening of all syllables, CRESCENDOING only of the last. In (10) a crescendoed fall-rise and a crescendoed rise are compared with their unmodified counterparts. (10) HZ
320
2(0 160
80
0
16
32
48
6«
Unmodified HZ
80
96
112
128
1U
160
CSEC
128
1«
160
CSEC
Crescendoed Î.H
320
0
16
32
¿8
64
80
96
112
As noted above, DELAY, HALF-COMPLETION and STYLISATION can be seen as morphemes that are spaced along a single semantic dimension (from 'special' to 'routine'). The meaning of CRESCENDOING is unrelated to those of the other three modifications, however. CRESCENDOING has to do with hearer-ratification. A speaker who uses a crescendoed tone thereby explicitly signals that the only hearer ratified to receive the communication is the addressee. Probably, crescendoed tones only occur with vocatives. If a speaker says (11)
jtf^rey!! (i.e. Don't say that when other people can hear you) *
Intonation
ill
(s)he thereby indicates that any other people within earshot are not supposed to pay any attention to the communication and are expected to 'respect' it as intended for the addressee only. The same implication attaches to crescendoed HLH and f-H. Imagine a couple, entertaining a friend. The friend is facetiously teasing the woman about her total inability to perform some simple task. She might then, in mock-distress, appeal to Jier, possibly temporarily abstracted, partner, and use a crescendoed HLH (or conceivably f-H) on a vocative (e.g. Jeffrey) as if to indicate: 'I pretend to be utterly helpless in the face of this harassment, please come to my rescue'. A pre-theoretical impressionistic description of that utterance can do little more than characterise it as a nonassertive, drawled whine. Application of our framework, by contrast, allows immediate classification of the tone, and can explain its effect as being due to (mock) hearer-ratification: she explicitly leaves the friend out of the exchange, who is supposed to pretend not to pay any attention to her call for assistance and allow the altercation to be brought to an end by the intercession of the third party. It also explains how a tone so rare can nevertheless have so clear a meaning. I am fairly sure that I have never heard a crescendoed fall-rise, certainly not in the situation described. Yet I would have no trouble decoding it. It would appear that the tone/modification framework presented here crucially lays bare the linguistic parameters that define the English tonal paradigm.
4. RESIDUE
The three-tone paradigm does not account for all Fo variation reported in tonal contours. If our analysis is correct, this residue should be relegated to low-level, phonetic variation, and hence not be invested with linguistic meaning. One example of residue has been given in (3). Clearly, the two contours are communicatively equivalent. It may even be the case that speakers would be hard to put to it to hear the distinction. Another example of residue is given in (12). (12) a. That's what I call cowardliness
ThSt's what I call cowardliness Pattern (a) is generally given by British authors describing the realisation
128
Carlos Gussenhoven
of the 'divided fall-rise' or the 'compound fall-plus-rise' in British English. The rise element (the final H) is described as beginning at the last (secondarily) accented word (Gimson 1980: 268). Pattern (b) is given by Pierrehumbert (1980) for American English: in her description the equivalent of the final H of HLH is associated with the last syllable. The difference resolves as one of unstable timing of the association of the final tone element in HLH, which cannot be incorporated as a separate modification. Again, this would appear to be precisely what we want: the difference is dialectal at best, and in fact not noticeable for most speakers. Cf. the striking difference in effect between unmodified t u and delayed ,H on a word like unicorn.)
Î
5. LINKING
In this section, which must be seen as tentative, two rules are discussed that operate on sequences of tones. Their effect is to simplify the contours spanning the tones relative to the unlinked situation. These rules are the TONE LINKING RULE (TLR) and the DOWNSTEP RULE (DR). TLR can be stated as follows: (13)
T
^
f f 0
That is, of a sequence of two tones the rightmost tone element of the lefthand tone is moved to a position adjacent to the left-hand tone element of the right-hand tone, or is deleted. The intervening stretch is bridged by a contour linking the two innermost tone elements, as illustrated in (14). (14)
a. A l b. A c. A
Al
A
A l
A
——A
lAl /
\
A contour like (a) could be referred to as an unlinked contour, one like (b) as a partially linked one, and one like (c) as a completely linked one. For unlinked tones a low baseline is assumed. Figure 3 gives the complete set of contours for all combinations of the three tones. Note that complete linking of HLH and I n must be blocked in order to prevent the result from being homophonous with HL + £ h , and that in a number of cases the distinction between partial and complete linking is neutralised. DOWNSTEP creates a series of descending terraces (cf Pierrehumbert 1980, Ladd 1983), as in (15).
Intonation
129
PARTALLY
UNLINKED
*
*
HL HL
*
*
HLH HL
_ A
A
-
W
A
.
LINKED
LINKED
- J -
» *
V _
— - V
LH H L
* * H L LH
J
/
* * H L H LH
* * LH LH
COMPLETELY
— y
—
'
(im text)
^ L /
/
Figure 3. Unlinked, partially linked and completely linked contours
The rule could be formulated as in (16), where the square brackets are assumed to contain an iterative element. (¡3, a 'downstepped high', is taken to form a plateau with the preceding L.) (16)
D R : f t L ftL [ A l ] ^ ft ftL [!&.]
In terms of our framework the important question is of course whether TLR and DR are rules of substance or of form. Is there a meaningful option 'TLR' in the way that there is a meaningful option 'passive', or should TLR be seen as a purely phonological rule? The examples in (17) seem to suggest that TLR is in fact a meaningful option. Note that as far as the contribution from syntax/lexis is concerned (17) is structurally ambiguous, as indicated on the right. (17)
John and Mary and Bill
[John] and [Mary] and [Bill] [John and Mary] and [Bill] [John] and [Mary and Bill]
130
Carlos Gussenhoven
If we assume that all three proper names are assigned an accent, each of which is realised by HL (say in the case of an answer to the question 'And who do YOU consider examples of the "right sort of people" then?'), no application of TLR will impose structure (a), linking of John and Mary will impose structure (b), while structure (c) is appropriately realised by linking Mary and Bill. The application of TLR is apparently determined by whether constituents are structurally more closely integrated than other constituents. DR, by contrast, imposes no such structure. A downstepped contour on (17) leaves the ambiguities unresolved. (Disambiguation could still be achieved by durational means, but this leaves the fact that DR does not resolve the ambiguities unaffected). A similar point can be made with respect to old men and women. Other examples of how downstep can operate across boundaries which cannot be bridged by linking are given in (18) to (21), where we cannot link the last two accents. (18)
You will stSrt at the bSttom rung of the ladder
(19)
T\ i>0 X = N, A, V, P, Q
In principle, we find reduction in all structures generated by (8). The following examples serve to illustrate this:
Coordination Reduction in Complex Words (9)
147
X = N, coordination of N's [ l a n d ^ M ] N en [tuinbouw] ^ 'agriculture and horticulture' coordination of NP's: [de l a n d W M ] j^p en [de tuinbouw] j^p t h e agriculture and the horticulture' X = A, coordination of A's een [ e l [ t w a a l f j a r i g e ] A j o n g e n 5 'an eleven year old, twelve year old boy' X = P, coordination of P's [voor/tf] p of [achterin] p de boot 'in the front or at the back of the boat' X = V, coordination of V's: Jan kon [ g l i m f j i ^ t f ] y of [schaterlachen] y 'John could smile or roar with laughter' coordination of VP's (V') .. dat Jan [eerst a p p e l ^ y p en [daarna druivesap 'that John first apple juice drank and then grape juice dronk] y p drank' coordination of S's (V"): .. dat [Jan a p p e l ^ tff^i^Jg en [Piet druivesap dronk] ^ 'that John apple juice drank and Peter grape juice drank' coordination of S's (V'"): Jan vroeg [wie er a p p e l ^ H44M\ g en 'John asked who there apple juice drank and [wie er druivesap dronk] g who there grape juice drank' X = Q, coordination of Q's Hij verkocht [ z e s t i g M i t M i ] Q, [zeventigduizend] Q 'he sold sixty thousand , seventy thousand exemplaren copies'.
Following Neijt (1979: 5) I assume that initial coordination is only possible for maximal projections of lexical categories. The relevant rule schema is (10): (10)
X1 -»• (conj X 4 ) n , i maximal, n > 2 .
An example of reduction in a structure with initial coordination is:
148 (11)
GeertBooij hetzij h o o f d ^ l M f , hetzij nevenaccent 'either main stress or secondary stress'
In the examples (9) and (11) deletion has taken place on the left side of the conjunction. However, as we already saw in (1), it is also possible to delete on the right side, as shown in (12): (12)
Dutch: [ [regel] N [ordening] N ] N en [ [ M i l ] N [toepassing] N ] N 'rule ordering and rule application' German: [ [Herren] N [mantel] N ] N und [ {iMHit] N [schuhe] N ] N 'men's coats and men's shoes'
The data presented so far suggest that the gap must be adjacent to the conjunction. This is confirmed by the data in (13): (13)
i. *in the l a n d ^ M van Nederland en de tuinbouw van Belgie 'in the agriculture of Holland and the horticulture of Belgium' ii. *.. dat Jan a p p e l ^ dronk en Piet druivesap dronk 'that John apple juice drank and Peter grape juice drank' iii... dat Jan appel^ft en Piet druivesap dronk 'that John apple juice drank and Peter grape juice drank'
In (13i) and (13ii) the gap is separated from the conjunction, and hence they are ungrammatical. In (13iii) the verb dronk has also been deleted (by V-deletion) and thus the gap of sap is adjacent to the conjunction. The ungrammaticality of the example of right reduction in (14) also confirms this condition: (14)
""de regelordening en de /¿j^ftoepassing 'the rule ordering and the rule application'
The next question to be asked is: what is the nature of the deleted constituent? Should it be defined phonologically, morphologically or syntactically? Phonological identity is not sufficient: witness, for example, the ungrammatical phrase en vlegels 'birds and impertinent persons'. The same holds for morphological identity, as shown by the ungrammaticality of the reductions in (15):
Coordination Reduction in Complex Words (15)
149
Dutch'. *[[bIauw]A«]Aen[[rod]Aig]A 'bluish and reddish' * [ [absurd] A W i ] N en [ [banal] A iteit] 'absurdity and banality'
N
German: * [ [ S a l z ] N « ] A u n d [[Mehl]Nig]A 'salty and mealy' * [ [Beambt] N oder [ [Arbeiter] N i n ] N 'civil servant, fem. or worker, fem.' * [ [Bestraf] y t f ^ ] N oder [ [Beforder] yung] 'punishment or promotion'
N
It should also be remembered that deletion can take place although the identical consitituents do not share the same syntactic status (cf. (2)). My answer to the problem of the nature of the deletable constituents is that they are phonological words (or projections thereof). By means of the notion 'phonological word' we can express the fact that there is not always a one-to-one-correspondence between syntactic words and their phonological correlates. In some languages, articles, clitics and the like are not independent phonological words, but fuse phonologically with a preceding or following word. Classic examples are the Latin conjunctions -que 'and' and -ve 'or', which fuse with the preceding word. On the other hand, a phonological word may also be smaller than a syntactic word. For instance, the constituents of Dutch compounds (and also certain affixes cf. section 4) have to be considered independent phonological words. This is clear from their syllabification patterns: the internal morphological boundaries of compounds always coincide with a syllable boundary, even when this would violate the Maximal Onset Principle. The following minimal pairs illustrate this:
syllabification :
(16) i
[balk] [anker] [bal] [kanker] ii [wet] sjtaal] [wet] [staal]
'beam brace' 'testicle cancer' 'legal language' 'knife sharpener'
(balk)ff(an)a(ker)a (bal)a(kan)a(ker)a (wets) a (taal) a (wet)a(staal)a
The fact that the phonological word is the domain of syllabification follows from Selkirk's prosodic theory (Selkirk 1978, 1980a, b) which assumes the prosodic hierarchy of syllable (a), foot (F), phonological word (co), phonological phrase (0 (ii) Y is adjacent to a conjunction (iii) There is a remnant that, like its counterpart, can function as focus constituent.
I assume that the requirement that Y must have an identical counterpart follows from the principle of recoverability of deletions. 6
4. REDUCTION IN COMPLEX WORDS WITH AFFIXES
Rule (21) predicts that those affixes which form independent phonological words can in principle be deleted. This prediction is correct, both for Dutch and German, as the examples in (22) illustrate:
(22)
Dutch: [ [zwanger] '] n en [ [moeder] N schap] N 'pregnancy and motherhood' [ [eenzijdig] A M 4 ] n en [ [ partijdig] A heid] N 'one-sideness and partiality' [ [zicht] v M H \ A en [ [tast] ybaar] A 'visible and tangible' [[storm] ftUWi] a e n [ [regen] N achtig] A 'stormy and rainy' [[eer] A en [ [deugd] N zaam] A 'respectable and virtuous' [ [oever] NW B / [ D . X _ Y . . . ] D .
An obvious question to raise in an attempt to place restrictions on prosodic rules is whether all six prosodic categories can serve freely as the domains Dj and Dj in all three types of rules. That is, can the prosodic category foot, for example, appear in the position Dj in domain span, domain juncture and domain limit rules? If we consider this question first in abstract terms, we find that, in fact, Dj and Dj cannot always be filled freely by the six prosodic categories. A summary of the possibilities is presented in Figure 1, where '+' indicates that a given combination of
219
Constraining Prosodie Rules
prosodic category and rule type is possible, or at least, cannot be ruled out a priori, and indicates that a given combination is not possible. RULE TYPE:
PROSODIC CATEGORY (Dj)
a
£ u>
I U
SPAN
JUNCTURE a £ cj
+ + + + +
+ + + +
+
+
(domain Dj) 0 I U
LIMIT
+ -
+
-
+
-
+
+
+ +
+ +
-
-
+
-
-
+ +
+ + + +
Figure 1. Prosodic categories and rule types.
What Figure 1 shows is that all six of the prosodic categories may, in principle, serve as the domain (Dj) in span rules, as indicated by the plus signs in the first column. The same is true for limit rules, as indicated by the plus signs in the last column. The situation with juncture rules, however, is somewhat more complicated since these rules make reference to two categories, the category of the adjacent units involved in the juncture (D:), and the unit within which such a juncture must occur in order for tne rule to apply (Dj). Since the prosodic categories are organized hierarchically, it follows that the juncture of two units of a particular type can only occur within a larger unit that comprises the two units in question. We must thus exclude a priori any domain juncture rule in which Dj, the domain of juncture, is not smaller than Dj, the domain within which the juncture occurs. Such cases are indicated by the minus signs in the columns under the juncture rules. The other combinations, which are logically possible, are indicated with plus signs. Thus, of the 48 ways of combining prosodic categories and rule types, only 27 are logically possible. In considering the logically possible rule types thus far, we have, in fact, made the tacit assumption that all prosodic domain rules are precisely of the forms given in (1) - ( 3 ) , and that the variables represented by '...' do not contain any reference to other prosodic categories. In other words, we have been assuming that in the cases of domain span and domain limit rules, only one prosodic category appears in the formulation of the rule, and in the case of domain juncture rules, two categories appear. There is, however, no a priori reason that this must be the case. Since prosodic categories do not correspond to boundary symbols in a terminal string of segments, but rather represent the phonological structure of a given sequence, it is not possible to manipulate them in the way boundary symbols were typically manipulated in traditional
220
Irene Vogel
SPE-type generative phonology. That is, while in traditional generative phonology it made sense to talk about erasing boundaries as a phonological derivation proceded from one cycle to the next, in prosodic phonology it does not make sense to eliminate (prosodic) categories once certain rules have applied. In other words, the information about the prosodic structure or levels of representation of a given string is constantly present throughout the application of all the phonological rules of the language from P-rulei through P-rule n , in much the same way as the syntactic structure of a given string is present throughout the application of all the transformational operations, from T t through T n . In fact, domain juncture rules, by their very nature, make use of this aspect of the representation of prosodic categories, in that they must simultaneously see not only the domain of juncture (Dp, but also the domain within which this juncture must occur (Dj). Given that all the categories of the prosodic hierarchy are present throughout the application of the set of phonological rules of a language, the question now is whether phonological rules may, in fact, make reference to any number of these categories. It is not difficult to imagine the formulation of a rule that applies, for example, at the left limit of a foot that is at the right limit of a phonological word, that is, a domain limit rule that makes reference to two categories, or another rule that applies, for example, within the domain of the phonological phrase between two phonological words when the first word ends in a strong foot, that is a domain juncture rule that makes reference to three categories. If we do not limit the number of categories a rule may refer to, however, the number of possible rules increases dramatically. In the present attempt to place restrictions on prosodic rules, I propose therefore to explicitly limit the number of categories referred to by domain span and domain limit rules to one, and by domain juncture rules to two, and thus retain only the 27 logically possible rule types arrived at above. This is, of course, ultimately an empirical question, and if it turns out that there are large numbers of phonological phenomena that cannot be handled without rules that refer to more categories, the restriction will have to be reconsidered.
3. ATTESTED TYPES OF PROSODIC RULES
While we have just seen, in abstract terms, that there are potentially 27 well-formed types of prosodic rules, this does not necessarily mean that all of these "logical possibilities" are actually real possibilities. An examination of the types of prosodic rules that have thus far been attested, in fact, reveals some interesting gaps in the system. In this section, I will
Constraining Prosodie Rules
221
attempt to distinguish the accidental gaps from the systematic ones, and on the basis of what appear to be systematic gaps, I will propose further restrictions on the set of possible prosodic rules. Let us first consider the following inventory of prosodic rules. For each of the 27 logically possible rule types, I have listed all of the prosodic rules I have come across in the literature, unless otherwise indicated.2 For each attested type, however, only one example is given. (4)
Domain span a. a: b. 2: c. u>: i) rules: IT s-voicing, GR nasal deletion, nasal assimilation, stop voicing, SKT (classical) Ruki, Nati ii) example: IT s-voicing: s -*• [+voice] / [^...V—V...]^
d. 0:
e. I:
(casa -*• ca[z]a'house', ficcanasi ficcana[z]i "busybodies', chiesetta -»• chie[z]etta 'little church') i) rules: AE rhythm rule, SKT (vedic) Ruki, Nati 3 ii) example: SKT Ruki: s -* s / L...luL[+seg] 9 9 ([+seg]a#r/r) iff (vi - syati -*• vis-yati) i) rules: IT spirantization, gorgia toscana, consonant lengthening, GR s-voicing, SP spirantization ii) example: IT spirantization: {t J" j -*• [+cont] / [j-..V_ V...]j 3
(amici -*• ami[/]i'friends', la cena^-la [/]ena 'the dinner', vende giornali -*• vende [ 3]iornali 'he sells newspapers') f. U: i) rules: AE flapping, BE r-insertion, SKT glide formation, vowel contraction, m-assimilation, SP s-voicing ii) example: AE flapping: t -*•£/ [y...V_V...]y (water -»• wa[X]er, right angle righ[X] angle, Pat asked -> Pa[X] asked, Please wait. I'll be right back. -*• ...wai[i] I'll...) In the series of span rules in (4), we see that examples are missing for two prosodic categories, the syllable and the foot. 4 Since there are abundant examples in all the other categories, it might seem that the absence of these two types of rules is accidental and that such examples would turn up with further research. On the other hand, it might not be accidental that span rules are missing only in the two smallest categories. Since in some sense the function of span rules is to provide phonological unity throughout a given string, it may be that the strings defined by the syllable
222
Irene Vogel
and the foot are so small that they present, in any case, a certain unity. Furthermore, it is possible that these two units are too short in general to provide the repeated environments required in order to say that a rule operates throughout the strings they define. If this is the case, and the absence of syllable and foot span rules is not accidental, but rather is systematic, we would be able to restrict the possible prosodic rules to those in which Dj is greater than 2 in the case of domain span rules. At this point, however, such a conclusion is somewhat speculative, and awaits additional data from more languages for further confirmation. (5)
Domain Limit a. a: i) rules: AE glottalization, FR nasalization, DU schwa insertion, SP glide to obstruent, /--strengthening, liquid gliding, lateral depalatalization, nasal depalatalization, aspiration, velarizations ii) example: AE glottalization: t ->t ? / [ a ...[-cons]_] a (set ->• se [t ? ], wait -*• wai[t ? ], atlas -»• a[t ? ]las) b. 2 : i) rules: AE aspiration, FR e, 3 -»• e ii) example: AE aspiration:
[+asp] / [ s _ . . . ] s
(tell -»• [t^Jell, repeat -*• re [p^]eat, connect [k^Jonnect) c. co: i) rules: AE vowel tensing, IT vowel lengthening, SKT final cluster reduction, depalatalization, deaspiration/ devoicing ii) example: AE vowel tensin:
[ + tense] / [ ¿ j . . ^ ] ^
(city -»• cit[i], veto -»• vet[o], potato -»• potat[o]) d. «P: e. I: (pause,prepausallengthening) 6 f. U: i) rules: SKT Visarga at pause, 7 SP glide devoicing ii) example: SKT Visarga at pause: i ?! h/ (devas -*• devah, punar
punah)
The domain limit rules in (5) provide a more straightforward case than the domain span rules. The only category for which a limit rule has not been found is the phonological phrase. This appears to be a fairly clear case of an accidental gap since there are limit rules operating on domains both larger and smaller than the phonological phrase, and since there is, furthermore, no particular reason that rules might apply at the ends of all the prosodic categories other than 0. It seems reasonable, therefore, to assume that Dj may, in fact, be any of the prosodic categories in domain limit rules.
223
Constraining Prosodie Rules (6)
Domain juncture a. a o n Z : b. a on co: c. a on : d. a on I : e. a o n U : f. g. h. i.
E 2 2 £
on on on on
co: 0 : I : U:
j. co on 0 : i) rule: IT raddoppiamento sintattico ii) example: C - [+long] /
V]2]w[a_
[ - S s a ] " V - V where a (avrà letto -»• arvà [l:]etto 'will have read', città vecchia -*• città [v:]ecchia 'old city') k. co on I : 1. co on U : i) rules: SKT final voicing, stop to nasal* ii) example: stop to nasal: [ X n t ] -[+naSl1 l u - L " - J t J + ^ l -
L-lu
(tat nanias tad namas OPT tan namas, dik nagah -»• dig-nagah -»• OPT din-nagah) m. on I : n. 0 o n U : *
o.I onU: Finally, the pattern exhibited by the domain juncture rules is the most difficult to interpret. Of all the possible combinations of categories in juncture rules, so far only two types have been found. In both of these, the domain of juncture (Dp is the phonological word; in one case the larger domain (Dj) is the phonological phrase, and in the other, the phonological utterance. The question now is whether the absence of the other types of juncture rules is accidental or systematic. If we look first at the co juncture rules, we see that there are no cases of such juncture rules operating within the intonational phrase, a gap that is most likely accidental, since there are examples of rules operating in both 0 and U. As far as the other missing rule types are concerned, two obvious hypotheses are a) that it is "accidental" that no examples have been found of rules applying at
224
Irene Vogel
syllable, foot, phonological phrase and intonational phrase junctures, and b) that it is not "accidental" that a, E, (¡> and I juncture rules are missing. In favor of the first hypothesis is the fact that the phonological word, the juncture category for which the only juncture rules have been found, is in the middle of the prosodic hieararchy. It could thus be argued that it seems strange that both the smaller and the larger categories are not involved in juncture rules, and the fact that only the middle one is, in the present data, is therefore most likely accidental. In relation to the second hypothesis, on the other hand, it could be argued that the fact that only co juncture rules have been found is evidence that the phonological word is in some sense the privileged category with respect to this type of rule. At this point there is no overwhelming evidence in favor of either position, although the first hypothesis seems somewhat less likely than the second one. That is, if we assume that all the possible types of juncture rules actually exist, it seems quite unlikely that only phonological word juncture rules have "accidentally" turned up so far, especially given that there is a good distribution across categories in the other two types of rules. Furthermore, there may, in fact, be a reason that phonological rules apply at the junctures of co but not other units. It may be that the junctures between phonological words, because of their similarity, if not identity, to those between morphological words, are in some sense more salient for speakers than the other junctures, and thus tend to be involved in phonological processes. If this were the case, we would be able to further limit the number of well-formed prosodic rules by excluding all juncture rules that did not have the phonological word in the position of D;.
4. TEST CASES
In the introduction to this paper, it was mentioned that one thing that was needed in prosodic phonology was a way to determine whether a prosodic rule was well-formed or not, and in this light, several restrictions have been proposed in sections 2 and 3. In particular, it has been proposed that the set of well-formed prosodic domain rules can be restricted in several ways by placing constraints on the number of prosodic categories a particular type of rule may refer to, and in some cases, also the nature of the category or categories a rule may refer to. If we now examine the formulation of the prosodic domain rules that have been proposed thus far, we find, in fact, that two of them deviate from the notions of well-formedness just proposed.9 This means that if the proposed restrictions are correct, we must conclude that the rules in question are ill-formed, and another way must be found to express the phono-
225
Constraining Prosodie Rules
logical phenomena they were originally intended to account for. On the other hand, if it can be demonstrated that a number of phenomena escape treatment with any of the accepted rule types, it will clearly be necessary to re-evaluate the proposed restrictions. In this section, the two prosodic rules that do not conform to the notions of well-formedness will be examined, and it will be shown that these rules can, in fact, be reformulated in accordance with the restrictions. The two rules that deviate from the proposed forms are found in Italian: Vowel Lengthening (VL) and Raddoppiamento Sintattico (RS) (cf. Nespor, in press).10 In both of these cases, the deviation lies in the fact that the rules makes reference to more than the suggested number of prosodic categories. As can be seen in (7), VL is a limit rule that applies at the right limit of a prosodic word, but instead of referring only to one prosodic category, as was proposed for limit rules, the formulation in question makes reference to three categories. (7)
Vowel Lengthening (Nespor's rule (20)) V-[+long]/[aj...[2[aC0_]aa...]2]aj
This rule states that a vowel is lengthened when it is at the left limit of the rightmost foot of a prosodic word, as illustrated by the examples of long and short vowels in (8). (g)
a
" UzIaPaUo^UsL-^
b
- Ids[ffPaUonetU2l2[atoUaneU2]w 'type of cake'
c
- [ c o [ 2 [ a r a ] a t a P i ] a [ a d o U 2 L ^ l r a : P i d o ] 'raPid'
d e
- iJ s [ < 7 t a n U a t o U 2 i c o ca
h
- [wIs[o Wa Ua 'calamity'
mi
1 1 6
]
'bread'
ttant°l 'so ta
much
W2isia lff]z]w
[panetto:ne]
' [kalamita]
The second rule, RS, is a juncture rule, and as has been seen already in (6j), this rule (repeated in (9)) makes reference to three prosodic categories rather than two, as proposed above for juncture rules. (9)
Raddoppiamento Sinatattico (Nespor's rule (22)) +son~| wnere x
OJ
[
-nasj - b - h
According to this rule a consonant is lengthened within the phonological phrase domain at the juncture of a phonological word with another pro-
226
Irene Vogel
sodic category that is either a phonological word or smaller unit (e.g. foot), when at the right limit of the first co there is a foot consisting solely of C 0 V, as illustrated below. (10)
a. [ 0 [ t i ; [ 2 t r e n t a ] 2 [ 2 t r e ] 2 ] Ù J [ ( O [ 2 p o r t e j 2 ] a ; ] ^ te] '33 doors' b
- U U s ^ U J z ^ l z U * ra
- [tékidldo] 'hot tea'
c
- [0[Js ]2[2 Ì2L[a;[2: 'he will invite Marco'
d
- U ta;[strental2Ìcj[io[2Porte]2]wV '30 doors'
f
invite
- \- [ c j [ z P o r t a i s ] c o l 2 l o i s l 0
[trentatrép:ór-
Marco
]s]a;l0
[inviteram:arko]
[tre'ntaP°rtel
[ P ó r t a l o l ' b r i n 8 it'
A closer examination of the two rules of Italian reveals that they both apply in relation to the final foot of a phonological word. More specifically, VL applies when the vowel in question is in the first, but not only, syllable of the final foot, and RS applies after a vowel that is in the first, and only, syllable of the final foot. By referring to the first vowel of the final foot of a word, both rules are actually identifying the position of primary stress, or of the "designated terminal element". The effect of the proposed formulations of VL and RS, however, is that of obscuring the nature of the context by making it appear to consist of a particular combination of segments and prosodie categories. If we make direct reference to stress or the DTE in formulating the rules, however, as in (11) and (12), we are not only able to reduce the number of prosodie categories referred to in accordance with the restrictions proposed above, but we express more directly the nature of the contexts of the individual rules, and in this way, also what they have in common. (11)
Vowel Lengthening (revised) +son v-^r+iongi l c 1° rL-nasJ l. 1j cj i b j //L[ ,c, j. . . r L+DTE-"
(12)
Raddoppiamento Sintattico (revised)
where x < co We are thus able to maintain the restrictions proposed above by reformulating the domain limit rule, VL, with one prosodie category, and the do-
Constraining Prosodie Rules
227
main juncture rule, RS, with two prosodie categories. Furthermore, the resulting reformulations have the additional advantage of revealing more clearly the nature of the contexts of the two rules. We can therefore conclude that the proposed restrictions serve not only as a way of limiting the possible forms of phonological rules, but they also yield formulations that are to be preferred on independent grounds.
5. RULE ORDERING
As soon as a language has more than one prosodic rule, it is necessary to have some way in which to determine the order of application of these rules. Thus far, the only explicit remark to this effect is a tentative suggestion of Selkirk's (1980a: 121-122) that rules apply according to a principle similar to that of the cycle, that is, rules with smaller domains apply before rules with larger domains. An alternative principle might be that all the rules of one type apply before those of another type. We can thus formulate the following two general ordering hypotheses: (13)
Hypothesis 1: Prosodic rules apply according to the size of the domain (e.g. smaller domains before larger domains).
(14)
Hypothesis 2: Prosodic rules apply according to the type of rule (e.g. span then limit then juncture rules).
It is clear, however, even without looking at any prosodic rules, that neither of these hypotheses is adequate. That is, they fail to make any predictions about the cases in which there is more than one rule operating in the same domain or more than one rule of the same type, respectively. One possibility is, instead, that the two aspects of prosodic rules mentioned in the first two hypotheses interact along the lines of the following two hypotheses. (15)
Hypothesis 3: Prosodic rules apply according to the type of rule in each domain according to its size (e.g. span then limit then juncture rules in smaller then larger domains).
(16)
Hypothesis 4: Prosodic rules apply according to the size of the domain for each type of rule in order (e.g. span rules in small to large domains then limit rules in small to large domains, etc.).
228
Irene Vogel
Finally, it is possible that a language has more than one rule of a given type operating in the same domain, and in this case, too, it is necessary to specify the order in which the rules apply when they can affect the same segments. In this case, we can make another type of hypothesis, as in (17). (17)
Hypothesis 5: Prosodic rules apply according to the type of relation among the rules themselves (e.g. feeding order before bleeding order).
As a first test of these hypotheses, let us now consider two sets of prosodic rules in which the order of application of the individual rules is crucial, thus requiring an explicit statement of ordering. The first is the set of rules that accounts for the well-known phenomena of aspiration, flapping and glottalization of t in American English.12 (18)
Aspiration ( 2 limit) t-*[+asP]/[2_...]2 (e.g. [t^Jea, [t^Joday, re[t^]urn, re[t^]ail, see [t^Jheresa)
(19)
Flapping (U span) t^X/[u...[-cons]_V...]u (e.g. wa[X ]er, doub[ £]ed, bright[£] orange, Pa[£] answered)
(20)
Glottalization (a limit) t + t ? /[ a ...[-cons]_] f f (e.g. wai[t ? ], wi[t?]ness, doub[t?]less, brigh[f ] red)
Rule (18) must be ordered before rule (19) since the latter takes as input only those f s that are not aspirated. In this order, we correctly generate sue Ted with [t^] and suit Ed with [-C]; in the opposite order, we would incorrectly generate [X] in sue Ted. Rule (20) also takes only unaspirated [t] as input, but since there is no overlap between the contexts of (18) and (20), the order of these two rules is not crucial. Instead, the order of (19) and (20) is crucial. The order (19) before (20) correctly generates write often with [X] and write back with [t ? ]; the opposite order would incorrectly generate [t ? ] in write often. Thus, rule (18) must apply before rule (19), and (19) must apply before (20). Such an order is not predicted by any of the hypotheses that take into consideration prosodic domain and rule type (Hypoth. 1 - 4). The first rule applies in a relatively small domain (£), the second in the largest domain (U), and the third in the smallest domain (a), demonstrating that the ordering does not depend
Constraining Prosodie Rules
229
on the size of the domains. It is also clear that the ordering does not depend on the type of rule, since (18) is a limit rule, (19) a span rule and (20) another limit rule. The only factor that seems to be constant in the relations between (18) and (19) and between (19) and (20) is that in both cases the first rule of the pair bleeds the second. Thus, at least for this set of rules, it seems that Hypothesis 5 is confirmed in that the ordering depends on the relations among the rules themselves. The second set of rules to be considered here consists of the following four rules of Tuscan Italian.13 (21)
Raddoppiamento Sintattico (co juncture on 0 domain) C - [+long] / V U x _ [ S ] ...] x ...] 0 where x < co |+DlfcJ (e.g. città [v:]ecchia 'old city', blu [k:]iaro 'light blue', ha [f:]atto 'he has done')
(22)
Consonant Lengthening (I span) C x - [+long] / []...V_V...], (C x = J.A.ji, ts, dz) (e.g. li[ J:]o 'smooth', fi[X:]o 'son', la [ J:]ena 'the scene')
(23)
Spirantization (I span)
(e.g. pa[J]e 'peace', la [3]ente 'the people, hanno [ /Jercato 'they have sought') (24)
Gorgia Toscana (I span) -cont "l
[
-voice [+cont] / [j...[-cons]_[-cons]...]j -del .rei. (e.g. to[0]o 'mouse', la [6Javola t h e table', vede [h]arlo 'he sees Among Carlo') these rules, (21) must apply before (23) and (24); its order in relation to (22) is not crucial. The order (21) before (23) correctly generates avrà cercato 'he will have sought' with [t: J ] and hanno cercato they have sought' with [ J ], while the opposite order would incorrectly generate avrà cercato with [ J : ] . Similarly, the order (21) before (23) correctly generates avrà cantato 'he will have sung' with [k:] and hanno cantato they have sung' with [h], while the opposite order would incorrectly generate avrà cantato with [h:]. In addition, rule (22) must apply before rule (23) to generate hanno scelto they have chosen' with [ J : ] and hanno cercato with [ J ] ; the opposite order would incorrectly yield
J
230
Irene Vogel
hanno cercato with [ J :]. The order of (22) and 23) with relation to (24) is not crucial. As far as the ordering hypotheses are concerned, it could be said that the Italian rules confirm any of the first four hypotheses since the
0 where V_n = neutral vowel (29)
The readjustment rule marks all roots ending in e (and also some ending in e) as [-28]. These roots will then be subject to the optional rule in (29), but a form like Jözsef is also marked as [-29]. Furthermore roots ending in a sequence . . . V n C 0 e ( ' ) C 0 (cf. 14) are marked [-29]. To see how the rules apply consider the following derivation: (30)
a. rAdir + tÖl +B - B
b. radir + töl -WFC-*
+B
-B
-(28)-
Hungarian Vowel Harmony c. radir + t 6 l
I
+B q>
277 d. radir tol
-WFC-+
+B
Finally the fronting rule in (24) turns i into i. Roots ending in e or more than one front unrounded vowel either only undergo the WFC (in which case e acts as harmonic) or the same four steps given for radir, with the difference that step c is due to rule (29) in those cases (in which case it acts as neutral). The analysis that Clements proposes has a consequence with respect to the representation of the vowel e. Vago (1980a) points out that in suffixes this vowel must be [+low] since it alternates with the low vowel a. In roots, however, the e must be specified as [-low], at least if it occurs in the final syllable. To see this consider the vacillating mixed root Agnes. This root may take either front or back suffixes. Back suffixes arise if the optional rule (27) applies, deleting the autosegment that is associated to the vowel in the final syllable. When this happens, the autosegment [+B] associated to the first vowel spreads: (31)
agnes
II
¿gnEs
I
¿gnAs
V
+B -B (29) +B 0 (WFC) +B
agnes (24)
II
+B -B
If the second vowel is specified as [+low] spreading of [+B] will turn it into a. The fronting rule cannot apply to the resulting form since in general a's cannot all be turned into e's; the vowel a belongs to the set of surface vowels. Hence the final vowel must be [-low] giving, after spreading, an abstract vowel / A/ that is fronted by the neutralization rule. Since only one e appears on the surface Clements' analysis presupposes a second neutralization rule, which collapses the mid e and the low e. The final aspect of Clements analysis involves roots with front unrounded vowels. As noted in the quote above, some of them are marked as an exception to the morpheme structure condition which requires front unrounded vowels to be associated to [ - B ] . The form in (32a) is regular and subject to the MSC and the form in (32b) is marked as exceptional. Clements represents these exceptional forms with an autosegment [+B], thus accounting for the fact that these roots take back suffixes: (32)
a. kek I -B
b. hid +B
It is not clear whether it follows from anything that roots exempted from
278
Harry van der Hülst
the morpheme structure condition have a [+B] autosegment, rather than no autosegment at all, unless Clements assumes yet another morpheme structure condition which requires that all roots have (at most) one floating autosegment. 2.2. An alternative autosegmental analysis In this section I will propose an autosegmental analysis that differs from the one I have just summarized. I too will posit two sources for the surface root vowel e, one harmonic source and one neutral source. To account for the behaviour of rteutral vowels, I propose that they will be associated to an autosegment [-B] at the end of the derivation, rather than at the beginning, on the assumption that predictable information is in general supplied as late as possible. Before I proceed let me point out that the treatment of neutral vowels that was offered in Clements (1976b) is altered in Clements & Sezer (1982). They assume that an analysis of harmony involves assigning values to five parameters: (33)
a. The class of P-segments (melody units) which constitute the autosegmentally-represented harmony features; b. The class of P-bearing units (melody-bearing units) defined as the class of units to which P-segments are associated under the universal Well-formedness Conditions; c. The (possible null) class of opaque segments, defined as those which are underlyingly associated with a P-segment; d. The (possibly null) class of transparent segments which must be formally excluded from the class of P-bearing units; e. The domain within which the Well-formedness Conditions initially apply
("P" stand for "harmony" or "tone" etc.) Clements & Sezer are not explicit about the way in which transparent (= neutral) segments are to be represented formally, either underlyingly or at the surface. They merely say (p. 218): Transparent segments are a subset of the segments characterized under [33b] which are "neutral" to the system in the sense that they do not associate with P-segments under the Well-formedness Conditions, but receive their feature values by independent specification.
The proposal that I adopt here can be seen as the formal implementation of the suggestion that neutral vowels are "formally excluded" from the class of P-bearing units.
Hungarian Vowel Harmony
279
As the first step toward an alternative analysis let me be explicit about the set o f underlying vowels. I assume five underlying vowel segments and one harmonic segment, being the P-bearing segments and the P-segment respectively: (34)
[+high] [-high, - l o w ] [+low]
vowels segments
harmonic segment
(P-bearing segments) [-round] [+round] I U E 0 A
(P-segment) [back]
In addition the grammar contains a set o f structure conditions, concerning segmental, autosegmental and sequential structure, capturing predictable value-assignments which are left unspecified in the lexical representations. Among segmental and autosegmental structure conditions are two conditions which are relevant for the phenomenon o f VH: (35)
t
-, -round I
r
-low
[-B]
J
(36)
[-B]
[
[-low]
-round j -high
/ X
X'
J [+low]
Condition ( 3 5 ) accounts for the fact that Hungarian does not have back unrounded, nonlow vowels. Condition ( 3 6 ) says that an unrounded nonhigh vowel is mid if long (-+ e) but low otherwise (-»• e, a, a). I will now assume that a condition like ( 3 5 ) blocks the universal association conventions from linking [ + B ] to vowels that are [-round, - l o w ] , blocking following from the universal principle o f proper inclusion precedence. ( 3 6 ) does not block association, since its structural change is not in conflict with that o f the association conventions. However, ( 3 6 ) adjusts the result o f association. In particular it turns ( 3 7 a ) into ( 3 7 b ) thus capturing the fact that a alternates with e (a low vowel) but a with e (a mid vowel):
280
Harry van der Hülst
(37)
a.
[-B]
[
-high "I
b.
[-B] I ["-high "I
-low I +low -roundl I-roundl The idea that /E/ predictable information is added at the latest possible stage /AI (found for example in the theory of natural generative phonology, e.g. Hooper 1976) has received a new stimulus from the work of Kiparsky (1982, 1983) and Pulleyblank (1982) who combined this with the idea that values which are supplied at the end of the lexical derivation are considered to result from default rules. The other suggestion, i.e. to block association in the case of neutral vowels, derives in essential respects from ideas that have been expressed by Paul Kiparsky.8 Finally, to let structure conditions adjust the output at each stage of the derivation is nothing but resuscitating the device of linking, proposed by Chomsky and Halle (1968). After this sketch of what I think is the most promising variant of a "standard" autosegmental theory of VH, we can turn to the question how roots and suffixes are now represented. Harmonic roots and non-alternating suffixes will be represented as in the previous autosegmental analysis: (38)
Roots a. +B b. -B VArOs OrOm c. +B d. -B hAz f6z Suffixes (non-alternating) e. +B kOr
Alternating suffixes will be represented as "defective", i.e. as lacking a segment on the backness tier: (39)
nAk
The representation of alternating suffixes as defective is in conflict with the claim advanced in Vago (1976, 1980c) that at least for some suffixes it can be shown that they must be underlyingly specified for the feature [back] (Vago 1976: 245-6):
Hungarian Vowel
281
Harmony
The directional case morphemes normally function as alternating suffixes attached to nominal stems [...]. However, following a personal pronoun, the case morphemes function as stems to which a possessive suffix is attached that agrees with the personal pronoun in person and number; the personal pronoun may then be deleted, in (en) ndlam
(*6nnit) 'at me", (en) tolem (*intol) 'from me', en is the first person singular pronoun, nal- and tol- are the adessive and ablative case morphemes that function as stems, and -am/-em is the first person singular suffix. The fact that some stem forms, like ndl-, have back vowels, but others, like tol-, have front vowels is explained quite naturally if we assume that the stem forms reflect the harmonic shapes of the underlying suffix forms; otherwise, ad hoc statements are required to specify the harmonic shapes of the stem forms. In this analysis, the underlying adessive and ablative suffixes are /nal/ and /tol/. It can further be shown that the conditional - n a / - n e / - n a / - n e has an underlying front vowel: /ne/. While this suffix alternates regularly (e.g. hozndnak 'they would bring', nizninek 'they would see'), in the first person singular indefinite forms only the alternate -ni appears: e.g. hoznek 'I would bring', niznek 'I would see'. These facts suggest that the underlying form of the conditional suffix is /ne/, which is exempted from undergoing VH before the first person singular indefinite suffix /k/.
We might adopt two strategies to escape from Vago's position. Although Clements (1976b) does not address the issue explicitly, he mentions a "universal convention" which stipulates that alternating suffixes always lack an autosegment (p. 52). If we wanted to maintain that there is one underlying form for the case suffixes and the independent stems, the autosegment that forms part of that morpheme would be stripped by the universal convention if the morpheme appears as a suffix. 9 The second strategy would be to question the assumption that suffixes and stems must be derived from one underlying representation. It is a fairly common phenomenon that some suffixes bear a resemblance (both in form and meaning) to stems. The explanation for this is that many suffixes derive historically from stems. Without going into the matter in depth I claim that we are not forced to assign a single underlying representation to such historically related morphemes. The second argument involving the conditional suffix can be countered as follows. Vago assumes a rule that assigns an exception feature to this suffix when it precedes the indefinite suffix /k/. The facts on which the second argument is based can be handled in the autosegmental model by assuming an allomorphy rule (in the sense of Aronoff 1976) which renders the suffix /nA/ opaque in the context of the suffix /k/: (40)
nA
^
ne I -B
/
+k
282
Harry van der Hülst
I conclude that it is possible to maintain that all alternating suffixes are defective underlyingly. Disharmonic roots will be represented as follows: (41)
+B-B
I I
soffQr So far the analysis has been the same as the one presented in the previous section. Let us now turn our attention to mixed roots. The representation of mixed A (neutral) roots will be: (42)
+B rAdlr
It was argued at the beginning of this section that association of [+B] to the second vowel will be blocked by the presence of a language-specific structure condition. This implies that f will be skipped when the autosegment is associated to the vowel(s) of alternating suffix(es): (43)
+B radlr
+B -B +B nak
by (35)
I I I
radirnak
Let us now turn to other mixed roots. There are two remaining categories, typically containing an e in the final syllable. In the present analysis the surface vowel e has two possible sources: (44)
a. -B
I
A
b. E by (35)
-B
-B
I
I
E by (36) A
In the case of (44a) we are dealing with the harmonic vowel /A/ that is associated to the autosegment [-B] and in the case of (44b) we are dealing with the neutral vowel /E/ that is assigned an autosegment [-B] by rule (35), the result of which is adjusted by rule (36). The dual source of surface e is the key to the analysis of the categories mixed B (vacillating) and mixed C (disharmonic). The point is that a root with an e in the final syllable cannot be mapped onto a unique underlying representation. With this in mind it comes as no surprise that roots of this type typically vacillate:
Hungarian Vowel Harmony (45)
283
mixed B (mixed vacillating) i
I
âgnes —
+B-B âgnes [+low]
1
+B
ÂgnEs [-low]
In the case of Jozsef native speakers have apparently decided to favor the disharmonie interpretation.
(46)
mixed C (mixed disharmonie) +B - B jôzsef [+low]
There are, as we have seen in section 1, no roots ending in e that take back suffixes only. 10 Zonneveld (1980: 34) points out that this gap cannot be explained in an analysis that posits two sources for e. If there are words that vacillate between the two sources for e (e.g. like Agnes), and there are words having only the low-e (e.g. like Jozsef), why then are there no words that have been "fixed" on the non-low-e? Zonneveld furthermore claims that this gap is explained if we assume only low-e. The reason is this. The low-e cannot be included in the set of neutral vowels because it alternates with a. Hence roots containing this vowel (and a back vowel as well) must be marked as exceptions to vowel harmony: they are disharmonic. In Zonneveld's analysis such roots are furthermore either marked with [+D] (in the case of Agnes) or not so marked (in the case of Jozsef) and this diacritic feature is referred to by an optional disharmony rule. (This analysis will be discussed in more detail in the next section). It follows then that roots ending in e that take back suffixes cannot exist, just as there are no roots of the type soffor which take back suffixes. Zonneveld's point is well-taken. Suppose we conclude from this that the autosegmental analysis must be modified. The way to do this would be as follows. We assume only one source for e, i.e. /A/ ([+low]), which belongs to the set of harmonic vowels. To account for the vacillation we then postulate an optional rule that inserts a floating autosegment [+B], after the autosegment that is associated to the /A/. Forms like Jozsef are marked as exceptions to this rule:
284
Harry van der Hülst
(47)
0
[+B] / [-B] I /A/
The one-source-for-e analysis starts out then from the assumption that (48) is excluded (cf. fn. 7): (48)
*X /E /
i.e. the melodic segment /E/ must be linked to two slots. A form like Agnes is now represented with two /A/'s: 1 1 (49)
+B - B
+B - B +B
I I AgnAs
"(4V)(opt.)
N
| | AgnAs
K
XX
XX
I am not convinced that we must accept this complication of the autosegmental analysis, however. Vago (1976) claims that roots taking only front suffixes develop from roots that vacillate. Apparently then there is a tendency to regard the e as harmonic only. Given this tendency and the fact that vacillation is the expected initial stage (given two sources for e) we also explain the gap. But there is a more compelling reason to reject the approach just outlined. As we know, e too may occur in the final syllable of a vacillating root (cf. 11). But unlike e, e may also be neutral (as in tanyer\ cf. 8). We now face the problem of how to represent these two types of roots. It seems to me that a proponent of the one-e analysis must be in favour of a one-e analysis: (50)
a. Vacillating +B -B AffAr
N
xx
b. Not vacillating +B - B tAnyÀr
A A
XX
XX
affer undergoes the optional rule in (47), just like Agnes, but observe now that we must say that tanyer undergoes this same rule obligatorily. Apart from technicalities involving the optional/obligatorily distinc-
Hungarian Vowel Harmony
285
tion, it is rather strange to see that e triggers the insertion of a [+B] autosegment, if and only if a back vowel precedes. If a root contains only e (one of more) we find as a rule [-B] affixes (ignoring the two exceptional forms in 16a). It seems to me then that instead of (50b) we must prefer (51): (51)
+B tAnyEr
K K
XX
XX
If this reasoning is convincing I fail to see why anyone would be in favour of a double source for e, but against a double source for e. So I conclude that the analysis that I proposed in the first place can be maintained. Let us proceed with a discussion of the treatment of the other root types. Vacillating roots that involve a final sequence of two front unrounded vowels (cf. 12 and 14) call for a separate rule: (52)
-B
Obligatory if the final vowel is e or e Optional in other cases Roots like federativ that do not end in e or e, nor in a sequence of two front unrounded vowels, and yet are vacillating according to Vago (1978: 120) (cf. the examples in 15) can only be accounted for by assuming that in sporadic cases neutral vowels are optionally associated to [-B]: (53)
-B +B(-B)
AI
federativ Finally, consider the monosyllabic roots that contain a neutral vowel. Crucial to an account of those roots is the fact that harmony in Hungarian is root-controlled. We can express this fact by assuming a morpheme structure condition saying that each root must contain an autosegment, be it [+B] or [-B]. As it turns out we find both possibilities: (54)
a. -B
b. +B
viz
hid
286
Harry van der Hülst
The autosegment in (54a) will be associated to the root, but this is not possible in example (54b). In both cases, however, the autosegment will be associated to a following alternating suffix. b.
+B
I
hid nak
/ - B +B
II
\ hidnak,
To complete the analysis I must still provide an account of the truncated forms in pet names (cf. 18). Recall that Clements (1976b) provides two arguments in favour of representing neutral vowels with [-B] underlyingly. The first argument involves the diminutive of Tibor (Tibinek/*Tibinak) and the second regular neutral vowel roots. In both cases we find front suffixes, and the question is how we can explain this. As for neutral vowel roots, we have just seen how the analysis works here: each root, including neutral vowel roots, must have an autosegment. As for the diminutive forms, let us simply see what will happen if we add nothing to the analysis: (56)
Tib + I + nAk
-(35)-
Tib + i + nAk -?-»• ...
I
-B
I
-B
It seems to me that the most likely thing to happen after assignment of default values is application of the WFC, leading to front suffixes. The present analysis makes crucial use of the dual source of e. Let us now return to the three properties of e that we mentioned above. It appears that we can interpret these properties as follows (cf. 17). The /E/ source does not occur in affixes, nor in monosyllabic words that take back suffixes. In all other morphemes, however, we may find a free variation between /A/ and /E/ as the source for e. The fact that the [-low]-e does not occur in suffixes or monosyllabic roots that take back suffixes remains an accidental fact, but, as was pointed out, other neutral vowels too are somewhat defective in their distribution. / does not occur in neutral suffixes and there are only two monosyllabic roots with e that take back suffixes. 2.3. A different approach to neutral vowels Booij (1984) offers an analysis of Hungarian vowel harmony that is different from the two preceding analyses with respect to the treatment of neutral vowels, ignoring for the moment other differences that are of minor importance. Booij recapitulates an argument provided by Anderson
Hungarian Vowel Harmony
287
(1980) against Clements' (1976b) treatment of neutral vowels, pointing out that the same argument can be used against the proposal of Clements & Sezer (1982). Booij claims that both approaches are based on the assumption that the presence of a neutral segment A, being either [+F] or [-F], implies the absence of a segment that differs from A in the specification for [F] only. Recall the first proposal, i.e. that found in Clements (1976b). Neutral ¡1/ is turned into an abstract vowel * which is turned into a front vowel i by an absolute neutralization rule. Such an approach would be impossible if Hungarian has surface t's. In the second proposal (as found in Clements & Sezer 1982) the set of neutral vowels must be identified (parameter 33d). Booij says (p. 4) that if the backness value of neutral vowels is specified after the application of the backness rule: ... parameter d cannot 'see' the difference between a front vowel and its back counterpart. This is no problem in the case of Hungarian, because of the absence of the (+back) counterpart of the three neutral vowels /i, i, e/. However, it would be a problem for a language in which a front vowel is neutral with respect to backness harmony, but its [+back| counterpart is not. An analogical reasoning holds for other kinds of harmony.
Anderson (1980) gives two examples where the implication on which both approaches to neutral vowels are based, is not satisfied. Firstly, in Khalkha Mongolian /i/ is neutral with respect to rounding harmony. Yet the language has a "rounded /i/", i.e. ii. Secondly, in Finnish we find backness harmony, as in Hungarian. Apart from i and e that are neutral (and have no back counterpart) it appears to be the case that occasionally u and even o are neutral. Yet the vowels ii and o occur as well. Booij then continues (p. 4): Although Anderson's insight that neutral vowels must be skipped is correct, this does not necessarily lead to a rejection of the autosegmental analysis of vowel harmony. The autosegmental framework as developed in [Clements & Sezer 1982 J solves the problem of neutral vowels, provided that the neutral vowels are specified for the relevant feature at the underlying level.
The specification of neutral vowels at the underlying level is interpreted as being at the segmental level}2 I will first consider the two arguments that Anderson puts forward, because I think that they do not, contrary to what Booij believes, carry over to the Clements-Sezer approach (nor to the approach that I have adopted). If I am correct then it follows that Booij's suggestion is unnecessary. Secondly, I will argue that Booij's suggestion, though seemingly possible, is at odds with the spirit of autosegmental phonology.
288
Harry van der Hülst
In both the Clements-Sezer approach toward neutral vowels and the more explicit (and compatible) approach that I have adopted some care is taken to prevent the WFC from associating an autosegment to some Pbearing vowels, i.e. the neutral ones. Yet this does not exclude the possibility of associating P-bearing units having the same featural make-up lexically to this autosegment. The point is, however, that such P-bearing units will be opaque. In other words a vowel / may be neutral with respect to backness harmony, although the language in question hast. The prediction is that in that case t will always function as an opaque vowel. We predict then that rounded front vowels are opaque in Khalka Mongolian. The analysis offered in Chinchor (1979) shows that rounded high vowels are indeed special in the sense that they do not participate in the harmonic process as "normal" harmonic vowels do. The high rounded vowels block the spreading of an autosegment [+R], so in that sense they behave as opaque. It seems to me this is sufficient reason to cast doubt on the conclusion drawn by Booij, although I admit that the process of rounding harmony in Mongolian needs further study. With respect to the Finnish case, Campbell (1980) states that "back and front harmony are both possible in most cases with neutral y" and "back harmony is considered more prestigious, more learned, while front harmony is more colloquial" (p. 250-251). I claim that in colloquial styles we find the things that are phonologically natural. The prestigious style may just as well contain a non-phonological rule inserting the autosegment [+B] in the case of loans that typically belong to the learned vocabulary. Let me now point out why Booij's proposal is at odds with the fundamental claim of autosegmental phonology. I will first quote his "summary" of the autosegmental theory (p.2): In an autosegmental analysis of this (i.e. Hungarian; HvdH) harmony the roots can be lexically specified with an autosegment [-B(ack)] or [+B(ack)J, which is associated with the vowels of the root by means of the Association Convention. This implies that the vowels in the segmental core are archisegments, i.e. they are underspecified since they do not contain a specification for the feature (back). The same holds for the alternating suffixes like nek/nak.
The crucial term here is archisegment. In recent literature this term is used to refer to P-bearing units, but we must realize that in the theory as conceived by Clements (1976a) P-bearing units are not strictly speaking archisegments. Segments in the "segmental core" are not "underspecified" in the sense that they contain a blank for the feature [Back]. The essence of the autosegmental approach toward backness harmony is that the feature [Back] is not present in the segmental core, but is present on a different tier. On a more intuitive level, representing one and the same
Hungarian Vowel Harmony
289
feature on separate tiers strikes me as artificial in the light of the phonetic interpretation of the model, but I will not pursue that point here, since little is known in this area.13 It seems to me that the modification of the autosegmental framework that Booij proposes is a very fundamental one. Implicitly Booij casts serious doubt on Clements' original claim that the same model that is used to represent tonal phenomena can be extended to the phenomenon of vowel harmony. Now it may eventually turn out that tone and harmony call for different approaches. It is pointed out in Van der Hulst (1984: chapter 1), that there are a number of differences between tone and harmony that appear somewhat mysterious if both are represented in the same fashion. I claim, however, that the treatment of neutral vowels does not force one into a revision of the framework proposed by Clements in various articles. Moreover, Booij's approach to neutral vowels faces the problem that there are two pieces of evidence showing that neutral vowels are associated to autosegments. This evidence was mentioned in the summary of Clements' analysis. Neutral vowel roots as well as the diminutive of a proper name like Tibor take front suffixes. To explain this Booij must assume a default rule that introduces the value [-Back] for suffixes that remain unassociated to an autosegment. Observe that in Booij's analysis it remains an accidental fact that the default rule introduces [-Back] rather than [+Back]. The analysis that I have proposed in the preceding section makes the appearence of front suffixes after neutral vowel roots non-accidental: suffixes are front because neutral vowels are front. 2.4. Conclusion I claim that the analysis offered in section 2.2. qualifies as "the best autosegmental analysis" and that there is no need to give up a basic claim of the autosegmental model about the structure of phonological representations.14 In the next section I will discuss segmental analyses of the same Hungarian data.
3. SEGMENTAL ANALYSES
3.1. The alternatives During the past few years a number of segmental analyses have been proposed. Vago developed the first extensive analysis within a generative model (Vago 1975, 1976, 1980c). His 1976 article triggered three reactions (Jensen 1978, Phelps 1978 and Ringen 1978). A large portion
290
Harry van der Hülst
of Vago's reply (Vago 1978) is devoted to the issue of dealing with exceptions to vowel harmony. Zonneveld (1980) takes up this central issue, presenting an analysis that differs from the accounts that had already been proposed. In reaction to Zonneveld, Ringen (1982) defends her analysis and Battistella (1982) contributes to the discussion by offering yet another analysis. In this section I will not spell out in detail all the proposed analyses. In particular I will not consider here the analyses proposed by Phelps (1978) and Battistella (1982), since these authors make use of powerful exception devices that are non-standard and, it seems, unnecessary within a segmental approach.15 Since these authors offer no compensation by eliminating other devices or by covering more data their analyses can not qualify as "better" than analyses that make use of a more restrictive theory of exception devices, without covering less data. Furthermore I will concentrate on points where the segmental analyses seem to differ in essential respects. I will discuss the differences by showing how the various types of roots have been handled, discussing them in the same order as in section 1: (57)
a. harmonic roots b. disharmonic roots c. mixed roots mixed A (mixed neutral) mixed B (mixed vacillating) mixed C (mixed harmonic/vacillating) mixed D (vacillating disharmonic) d. neutral vowel roots
(a) harmonic roots The controversy surrounding this class is the following. The type of agreement that exists between vowels inside a root and across a suffix boundary is undoubtedly the same. Yet it has been proposed to deal with VH inside roots in terms of a separate statement. Within this view root vowels are fully specified and VH is accounted for in terms of a morpheme structure condition (MSC). Suffix vowels may be fully specified or archisegmental (depending on the evidence for underlying specification) and a phonological rule either changes or fills in their value in accordance with the value of the final root vowel. Several arguments have been advanced in favor of this type of approach (cf. Ringen 1980).16 Firstly, Kiparsky (1968) argues that the two VHrules (i.e. the MSC and the P-rule) have different classes of exceptions:
Hungarian Vowel Harmony
291
[...] it is impossible to derive both root harmony and affix harmony by a single rule, because they have different classes of exceptions, and are therefore different processes. Finnish (as well as Turkish and Hungarian) contains harmonically incompatible vowels, e.g. afaari 'affair', olympiasaiset 'Olympic games'. Obviously these words are exceptions to the stem harmony morpheme structure condition. Yet that all strictly observe the vowel harmony rule; their affixes undergo exactly the same rule as do those of native words [... ] (p. 29) However, Ringen (1980) correctly points out that we have a non-sequitur here. Disharmonic roots may be marked as [-vowel harmony rule], implying that each segment of such roots is thus marked. Hence no segment of such roots will undergo VH. According to SPE conventions it is not the case that segments that are marked as [ - V H ] cannot function as the environment of the rule. Hence there is no reason why suffix vowels could not agree with the final vowel of a disharmonic root. Apparently Kiparsky had a different theory of exception features in mind, as will also become apparent below (cf. footnote 21). A second argument in favour of using a morpheme structure condition relates to the use of "blanks" in phonological representations. An obvious way of dealing with VH inside roots and across suffix boundaries using only a P-rule involves setting up roots in which all vowels except one are unspecified for the harmonic feature. 1 7 The specified vowel is usually considered to be the vowel in the first syllable. Stanley (1968) claimed that it is possible t o (mis)use blanks as a third value and that lexical representations must therefore contain fully specified feature bundles, thus ruling out the one-rule approach to VH just mentioned. Regularities holding at the morpheme level cannot be expressed in terms of P-rules that fill in or change underlying valued. All regularities at this level must be expressed in terms of a separate category of morpheme structure conditions. Again Ringen points out that the argument is not compelling (cf. also Ringen 1977). It is possible to modify other aspects of the theory in order to prevent the misuse of blanks, while allowing P-rules to apply morpheme-internally to underspecified segments. The change that Ringen proposes also affects the third argument in favour of the MSC-approach, which I will now discuss. This argument is based on proposals advanced in Kiparsky (1973), where he argues in favour of a principle which requires that non-automatic neutralization rules must be prevented from applying in non-derived environments. In this way Kiparsky blocks the possibility of setting up abstract underlying representations that violate his earlier alternation condition (cf. Kiparsky (1968, 1973)). Since VH rules (at least in Hungarian and all other cases I know o f ) neutralize underlying contrasts and
292
Harry van der Hülst
are non-automatic in the sense that they have exceptions, we may infer that VH inside roots cannot be effected by a rule. Ringen (1977, 1980) posits a single rule of VH. She is able to capture the spirit of Kiparsky's principle by reformulating it so that it forbids a rule from changing feature values when applied in a non-derived environment. The requirement, thus formulated, prevents both the misuse of blanks that Stanley rightly condenms and the type of abstractness that Kiparsky judges so harmful for the generative theory. By this ingenious move Ringen has managed to eliminate two arguments in favour of the two-rule approach. 18 Only in Vago's analysis of Hungarian VH do we find an explicit defence of the position that VH inside roots is accounted for by a morpheme structure condition. Vago (1976, 1980), however, does not rely on the three arguments that have just been discussed. His line of argumentation is the following: (58)
i. Neutral vowel roots (regularly) take front suffixes; ii. Since neutral vowels are front we explain i. if we assume that neutral vowels trigger VH; iii. Mixed vowel roots like Tibor have an underlying front vowel in the first syllable; cf. Tibinek; iv. By extension the same holds true for bika, although there is no diminutive evidence v. bika does not become bike; I— VH does not apply root internally
Vago argues that roots need not be marked as [-VH], if Kiparsky's principle is accepted. In all other segmental approaches the use of a separate MSC is explicitly or implicitly rejected. Ringen (1980) uses, as said, one harmony rule. In roots the first non-neutral vowel is fully specified and all other vowels are unspecified for the feature [back]. Suffixes (at least those for which there is evidence) are fully specified. The harmony rule only fills in values when it applies root internally and changes values when applied to suffixes. Neutral vowels are not affected by the rule: 19 (59)
[~+syll 1 f+syll "| +low + [aback] / |+syll |(C 0 -low ) Q CQ — l+roundj I aback I I -roundl
A neutral vowel may trigger the rule (i.e. is not skipped) if it is the only vowel that can function as the triggering environment after the VH rule has
Hungarian Vowel Harmony
293
applied the value for [back] in neutral vowels is filled in by a rule ( c f . Ringen's footnote 13, p. 152), that I formulate here as follows: 2 0 (60)
|~+syll
"I
-low
[-back]
L-round J In the analysis given by Jensen ( 1 9 7 8 ) and Zonneveld ( 1 9 8 0 ) there is also one rule: (61)
[+syll]
[aback] / [+syll
| C 0 — ( l e f t - t o - r i g h t iterative)
I .¿back I Neither author is explicit with respect to the representation of roots and affixes; i.e. it is not clear whether in roots only one vowel is specified and whether affixes are unspecified or not. In any case Jensen and Zonneveld make no use o f Kiparsky's principle, since they specify disharmonic roots (like soffor)
as [ - V H ] . They also specify mixed roots like beka as
[ - V H ] , Hence they eliminate premiss five in Vago's argument, which means that the conclusion no longer follows.
(b) disharmonic roots & (c) mixed roots A n analysis like that of Jensen, using only one rule and an exception feature [ - V H ] , faces a serious problem, as Vago ( 1 9 7 8 ) points out. Let us see how such an analysis would work. As noted above, disharmonic roots are marked as [ - V H ] , as well as mixed roots like bika and Jdzsef. Mixed roots like radir, however, are not so marked and therefore subject to the rule. Consider the following derivation: (62)
radir+nek
radir+nek
radir+nak
The back unrounded vowel 4 is subsequently fronted by a rule o f absolute neutralization: 21 (63)
r+syll
"I
I -roundl |_-low
->
[-back]
J
The vacillating behaviour o f roots like Agnes is accounted for (by Jensen) by supplying them optionally
with the rule feature [ - V H ] ,
294 (64)
Harry van der Hülst a
b e Agnes + nek AgnAs + nak Agnesnak -> Agnesnak (the second e of Agnes is [-low, -back], the e of the suffix is [+low, -back]) c Agnes + nek ->• Agnes + nek -*• Agnesnek [-VH] [-low] [+low] where; a = VH, b = absolute neutralization rule, c = a rule converting /e/ to [ e]
Vago (1978) correctly points out that Jensen's use of the feature [-VH] is problematical. Consider the trisyllabic roots given in (15). These roots are disharmonic and must in Jensen's account be marked as obligatorily [-VH], but they are also vacillating and therefore at the same time optionally [-VH]. The problem under discussion does not exist if we claim, as was first done in Ringen's work, that the vacillating character of roots like Agnes is due to a separate Disharmony rule, which is triggered by a diacritic feature [D]: (65)
P+syll 1 +low
[~+syll "1 [-aback] / a b a c k l ( C c
|_+roundJ
L
J
D
[
+syll
I
Condition: obligatory when root contains vowels, " l ™only )neutral oCo r o u n d j optional otherwise This disharmony rule can be collapsed with the harmony rule given in (59) above: (66)
|"+syll
"I
+low L+roundJ
[aback] /
+syll \ [aback] (
H
ack LVL- D
])
[
+syll "I "low )0C0. -roundl
Zonneveld (1980) also proposes a disharmony rule: (67)
[+syll]
[aback] / T+syll I aback C Q — L+D J
The disharmony rule and the harmony rule given in (61) above can also be collapsed into one rule by means of braces:
Hungarian Vowel Harmony
(68) [+syll] -»• [aback] /
295 +syll' [aback])
nri
—
Despite the similarities in the rules there are a number of important differences between Ringen's and Zonneveld's analysis. A first difference is that disharmonic roots are marked with an exception feature in Zonneveld's analysis, while Ringen does not need this device. In her approach to disharmonic roots the vowels are specified for the feature back. This alone takes care of the fact that the VH rule is not applicable here: it cannot change values in non-derived environments. A second difference involves the use of the diacritic feature which triggers the disharmony rule. A minor point is that Ringen assumes that diacritics are simplex in order to prevent reference to the complement class of the exceptional class (Ringen 1982, 195). The more important difference between the two analyses become apparent when we consider neutral vowel roots. (d) Neutral vowel roots We have seen that in the regular case such roots take front suffixes and that only about 60 roots select a back suffix. The regular case is handled by both analyses without problems, since both rule (65) and rule (67) derive front suffixes in this case. With respect to the irregular class two approaches are possible. One, advocated by Vago and supported by Jensen (1978), Phelps (1978) and Zonneveld (1980), involves the postulation of underlying abstract segments. It is claimed that hid has in its lexical representation a back unrounded vowel 4-. As such this root causes suffixes to surface with a [+back] vowel. The rule of absolute neutralisation, ordered after VH, turns * into / to generate the surface form. This treatment of "neutral vowels" is not chosen by Ringen, who adheres to a concrete theory of phonology. In this theory it is impossible to posit segments in the underlying representation of a morpheme that never reach the surface. The possibility of positing such abstract segments represents an undesirable richness of the theory, as was first argued by Kiparsky (1968). The impossibility of using such segments follows in Ringen's theory from the fact that the rule which must turn the abstract segment into some surface vowel (or zero) would be applying in a feature changing manner in a non-derived environment. In her theory, then, Ringen accounts for backing after forms like hid in terms of the disharmony part of the proposed rule. Hence a form like hid is stored with its surface vowel, but marked as [D].
296
Harry van der Hülst
Ringen's approach faces the problem that the disharmony rule, which applies optionally in the case of polysyllabic roots that are marked [D], must apply obligatorily to monosyllabic roots like hid. Zonneveld's analysis circumvents this complication. With respect to neutral vowels in their other occurences (where they do not behave exceptionally as in hid) Ringen and Zonneveld also have different views. They agree on the point that roots containing the low vowel e as well as a back vowel are to be treated as disharmonic. So Zonneveld marks them as [-VH], whereas Ringen uses full specification. However, in Zonneveld's analysis the other unrounded front vowels are subject to the rule of VH, i.e. i, i and e when occurring in roots like radir are converted into back unrounded vowels by the rule of VH. The rule of absolute neutralization given above in (59) converts these vowels into front unrounded vowels after the rule of VH has applied. One might say that in Zonneveld's analysis (as in that of Jensen) "neutrality" is a surface phenomenon only. Apparently Ringen assumes that abstract segments are not only impossible underlyingly, but also at intermediate stages of the derivation. Hence her rule only applies to vowels that are either [+low] or [+round] and skips front unrounded (i.e. neutral) vowels if possible. In this way Ringen avoids the appearance of a vowel like 4 in forms such as /radir/. Let us now observe, however, that if Ringen adopted a rule that does not skip neutral vowels, she could allow the neutralization rule (which she has to adopt in any case, cf. fn. 21) to change / to i, 4 tot i and X to e. The crucial point is that such an application of the neutralization rule would not be prevented by the constraint on rule application, since the abstract vowel in radir constitutes a derived environment. So it is not at all clear why Ringen should adopt a VH rule that is different from the rule that Jensen and Zonneveld use. Before evaluating the foregoing discussion, let us consider the segmental analysis proposed by Vago (cf. footnote 1). Vago argues that two rules must be postulated in addition to an MSR. His argument rests on the presence of mixed roots, and in particular on the presence of mixed harmonic roots like Joszef and mixed vacillating roots like Agnes. Vago starts out by assuming a simple harmony rule, essentially the one given in (61) above repeated here for convience: (69)
[+syU]
[aback] / T+syll l C 0 labaclu
Given this rule and given a root like radir, which selects back suffixes, one might argue that the lexical representation of this root is /rad+r/, with a back unrounded vowel. But what about Agnes, which may select both [-back] and [+back] suffixes?
Hungarian Vowel Harmony
297
Vago plainly rejects the possibility of having two underlying representations. Hence the lexical representation must be /dgnes/. Vago extends then this conclusion t o radir, whose lexical representation is assumed to be /radir/. However, radir selects a back suffix and it must therefore follow that there is a second harmony rule (called marked vowel harmony) that skips unrounded front vowels. Forms like Agnes are optionally exempted from conditioning this second harmony rule: (70)
f+syll "I [+syll] -
[ + b a c k ] / f + s y l l | C 0 ( -back C ^ I +backj I-roundl
Forms like Jozsef are obligatorily [-marked VH]. Rule (29) and rule (70) cannot be collapsed according to Vago since forms may be (obligatorily or optionally) exempted from undergoing rule (70), while they must undergo rule (69). As Vago says (1976: 255): The validity of the above argument rests on the assumption that lexical features may refer to an abbreviated rule schema, but not to its subrules. All things being equal, a theory that accepts this premise is more restrictive, hence stronger, than a theory that rejects it. Granting this, let us see what the essential difference is between the t w o rule analysis proposed by Vago and the two-rule analyses proposed by Ringen and Zonneveld. At first sight it appears that the crucial difference is that the latter two analyses invoke a rule of disharmony, whereas Vago has two harmony rules. The difference is only apparent, however. This is observed by Zonneveld (1980: 36): [.. .1, notice that the label "marked" VH is little more than a mnemonic disguise for what appears to be called, with equal reasonableness, a rule of vowel disharmony. Because of the subscript 1 in [(66)] there will obligatorily appear a front vowel immediately to the left of the back-vowel focus, a situation which is not avoided by suggestive nomenclature. I think Zonneveld's objection to the marked harmony rule is convincing. Having discussed the major segmental "variants", let us now ask the question which segmental analysis qualifies as the optimal one.
3.2. Abstract or concrete The first conclusion that we may draw is that a segmental account necessarily involves two P-rules: a harmony rule and a disharmony rule. T w o -
298
Harry van der Hülst
rule analyses of this type are proposed by Ringen and Zonneveld (and implicitly as we just saw by Vago as well). The second conclusion is that it is not necessary to assume that there is a separate morpheme structure condition to account for VH inside roots. This leaves us with the analysis proposed by Ringen and Zonneveld. Let us therefore focus again on the major differences between these two analyses. Ringen attributes the behaviour of both vacillating and neutral vowel roots that take back suffixes to the presence of diacritic [D], thus claiming that both types of roots belong to the same category. Zonneveld, on the other hand, attributes both types of roots to different categories. Vacillating roots trigger the optional disharmony rule with a diacritic feature, whereas the neutral vowel roots are represented underlyingly with an abstract vowel, thus triggering the obligatory rule of VH. A point in favour of Zonneveld's position is that the two types of roots indeed behave differently. The neutral vowel roots at issue do not vacillate. This follows from Zonneveld's analysis, but it must be stipulated by Ringen (cf. the condition on the rule cited under (65)). One might argue that Zonneveld pays a price for his abstract vowel analysis, viz. the rule of absolute neutralization. This rule, however, expresses a true generalization about the segment inventory of Hungarian, which implies that any analysis of Hungarian must contain such a rule. We observed above that Ringen assumes the same rule in her analysis (cf. (60) and footnote 21). The two analyses also differ, as we have seen, in their treatment of disharmonic roots: Zonneveld uses an exception feature and for Ringen it is sufficient to say that the vowels of such roots are fully specified. There can be little doubt that Ringen's approach is superior here. This difference is not independent from the abstractness issue, since the principle that allows Ringen to use full specification as a means to encode exceptionality also forbids her to posit abstract vowels. With respect to the treatment of neutral vowels, 1 have shown that by skipping them Ringen makes her rule (and theory) more complicated than necessary and for this reason I will prefer Zonneveld's approach: neutral vowels participate in the harmony process and a neutralization rule takes care of the fact that only the front couterparts appear at the surface. In terms of overall economy I think the best segmental analysis should adopt from Ringen the use of full specification as a means to encode exceptionality (we don't need the exception feature), and from Zonneveld the treatment of neutral vowels (the rule becomes simpler, there is no "danger" of misusing the derived abstract segments and forbidding them requires extra principles). I do not find it possible, however, to argue in favour of Zonneveld's abstract segment in forms like hid, although as we saw above Ringen's
Hungarian Vowel Harmony
299
approach toward these roots is not satisfactory either. The question then as to whether or not abstract segments may be present in lexical entries, raised 16 years ago by Kiparsky, cannot receive a definite answer. I would like to claim that the issue is undecidable in principle, at least within the context of a segmental theory. Either one adopts a theory in which abstract segments are banned from the lexicon, as Ringen does, or one adopts a different theory in which such segments are legitimate entities. The former theory allows a rather harmless use of abstract segments (e.g. as here), but also the kind of use that Kiparsky objected to in the first place. The second theory disallows any kind of abstractness and leads in general to more complex analyses. It is not clear to me how the issue is ever going to be settled.
4. THE COMPARISON
An obvious difference between the autosegmental analysis and the (two) "best" segmental analyses is the fact that in the former there is no harmony rule to begin with. There is only a universal wellformedness condition. A segmentalist might adopt the point of view, however, that the harmony rule used for Hungarian could also be looked upon as a universal rule. It is compatible with the analysis of both Ringen and Zonneveld to say that in harmonic roots only the first vowel is specified for the feature [back], and also that all suffixes have underspecified vowels. Underlying representations with underspecified vowels, our segmentalist might argue, entail the presence of a harmony rule by universal convention. In this sense an autosegmental and a segmental analysis would be equally simple. What about the rule of disharmony, however? Recall that both Ringen and Zonneveld make use of such a rule to account for the behaviour of vacillating roots. Our autosegmental analysis accounted for these roots by postulating different underlying representations. This difference appears to be independent of the framework we choose. The view that vacillating roots have two possible underlying representations is also possible in a segmental approach (cf. the analysis of Jensen 1978). We also saw that an autosegmental analysis with one source for e is in principle possible. In that case we have to postulate an autosegmental counterpart of the segmental disharmony rule (cf. (47)). Let us turn therefore to the "heart of the matter", i.e. to the neutral vowel roots that take back suffixes. Here Ringen uses the diacritic [D], while Zonneveld uses abstract vowels. I would now like to claim that the autosegmental analysis strikes a perfect balance between these two approaches: 22
300
Harry van der Hülst
(71)
a. / h i d / [D]
b. /hid/ [+B]
c. / h i d /
[+B] By postulating a floating feature we have n o t c o m m i t t e d ourselves t o t h e claim that abstract vowels are possible. The autosegmental representation is of course abstract in that it allows a floating feature that need not appear on the surface, but this is only possible, I would like t o argue, because we make use of an autosegmental tier, for which the language offers a b u n d a n t evidence. This makes the autosegmental t y p e of abstract ness more constrained than the type of abstractness that is allowed by Zonneveld, Jensen and m a n y other generative phonologists. The autosegmental representation has the advantage over Ringen's diacritic and shares with Zonneveld's abstract vowels that it does not trigger a separate rule. In Ringen's analysis it is essentially a coincidence that the diacritic [D] triggers a rule that refers to the feature [back], rather than to a rule that refers t o [nasal]. Given that a separate disharm o n y rule is formulated, such a state of affairs would nearly be as easy to describe. This is not t h e case in the autosegmental analysis if t h e constraint on the use of floating features just suggested is a d o p t e d , because there is n o independent motivation for placing the feature [nasal] o n a separate tier in Hungarian. I claim that the advantage of adopting the autosegmental analysis lies in the fact that this f r a m e w o r k o f f e r s a principled answer t o the "abstractness c o n t r o v e r s y " as it is manifested in the analyses of vowel h a r m o n y . It allows us to be concrete and disallows m a n y of the abstract analyses that are possible in an S P E - t y p e of approach. Yet we are not forced into a diacritic approach in cases of the type we have discussed here. This is a welcome result, since it does not seem likely that "ghost segm e n t s " will ever stop haunting the p r o p o n e n t of a purely segmental framework.
NOTES 1. In very recent work Vago has offered an autosegmental analysis of VH in Hungarian. The autosegmental theory that Vago proposes deviates in several ways from the standard theory. Vago's autosegmental analysis reached me too late to be included in this study. 2. Miklos Kontra and Catherine Ringen have been working on the treatment of loanwords with respect to Hungarian. Their work, known to me from an oral presentation, reached me too late to be discussed here. 3. This was pointed out to me by C. Ringen and A. Kornai. 4. C. Ringen (p.c.).
Hungarian Vowel
301
Harmony
5. I refer to Clements' article for some discussion on the technical details associated with the application of this rule. 6. A harmless ambiguity is that kosztiim could also be represented with one of the association lines missing. In Vago (forthc.) it is argued that the first autosegment must be unassociated, but nothing that I say here hinges on a decision, so I will ignore this point. 7. The difference between short and long vowels is due to their being associated to one and two "slots". In full the phonetic interpretation of vowels refers to more complex structures: +B
I
X
I
A
8. During the 2nd Salzburg Summer Institute (1982) Paul Kiparsky showed me some unpublished work of his dealing with vowel harmony. The present proposal concerning the treatment of neutral vowels is inspired by that work. 9. This convention forces us to represent suffixes like /kor/ (cf. 38) with a bound autosegment and should refer to floating autosegments as being impossible in suffixes. 10. Recall our discussion of the form maszek in the section 1. 11. In examples (49), (50) and (51) I deviate, for the sake of clarity, from the convention of using small letters for P-bearing units that are bound to an autosegment. I also indicate length in terms of slots instead of with a diacritic. 12. In Van der Hulst and Smith (1982b) it is also argued that neutral segments can be handled in terms of segmental specification. 13. Ringen (1984) makes the same point when she discusses Halle and Vergnaud's proposal to specify all vowels at the segmental level with the "unmarked value" (cf. Halle and Vergnaud (1980). This proposal can easily be replaced by another proposal in which the unmarked value, where necessary, is assigned by a "default" rule that inserts an autosegment. This is the proposal suggested in Pulleyblank (1982), taken over in the analysis that was presented in section 2.2. 14. In some "underground" variants of the autosegmental model it is proposed to use single-valued features as autosegments, i.e. in this case [B] or nothing. In considering this possibility during an oral presentation of his paper, Booij pointed out that in this alternative it follows that neutral segments must be specified nonautosegmentally. It seems to me that this is a very interesting approach, but it faces the problem of dealing with opaqueness. For example, in Akan the low vowel is opaque, yet it is [-ATR), whereas the spreading autosegment is [+ATR]; Clements (1981). 15. These exception devices involve segment-sized exception features, exception features that differentiate between "environment and "focus", and others. I refer to Zonneveld (1980) and Ringen (1982) for some discussion of the proposals of Phelps and Battistella. 16. Kiparsky (1968) gives some other arguments based on markedness theory and morpheme structure conditions. In the light of subsequent developments these arguments are no longer relevant. 17. I refer to Kiparsky (1968) for some discussion of some of the early proposals that involve this approach.
302
Harry van der Hülst
18. In the theory of "lexical phonology" (cf. Kiparsky 1982, 1983 and Mohanan 1981) Ringen's proposals have been incorporated. 19. In giving this rule and others I have made some harmless changes such as the replacement of "V" by [+syll]. 20. A problem in Ringen's approach is that neutral vowel roots have underspecified vowels only, i.e. Ringen (1980, 147) gives /hi:d/ for hid. This is probably an oversight, since if only neutral vowels are present one of them must be fully specified in order to function as the determinant in the harmony rule. 21. In her footnote 14 (p. 152) Ringen (1980) claims that her analysis does not contain the absolute neutralization rule, but this is not really true, since in footnote 13 she introduces a rule (given in (63) in this paper), that is identical to the rule of absolute neutralization that Jensen proposes. In fact no analysis of Hungarian can do without such a rule, since it expresses a true generalization about the segment inventory of Hungarian. 22. Kiparsky (1968) proposes a third way of dealing with the irregular abstract vowel roots. He marks them as [-VH1. This exception feature is of the type [-environment VH) as opposed to the feature (-context VH) that Kiparsky (1968) will also have to use for disharmonic roots. Hence the analysis proposed by Kiparsky (1968) is subject to the same criticism that we levelled against Phelps (1978) and Batistella (1982) in that it presupposes a theory of exceptions that is extremely powerful. REFERENCES Anderson, S. (1980), Problems and perspectives in the description of vowel harmony. In: R. Vago (1980b), 1-48. Anderson, S. (1982), Differences in rule type and their structural basis. In: H. van der Hülst and N. Smith (1982), Part II, 1-25. Aronoff, M. (1976), Word formation in generative grammar. Cambridge/Mass/, MIT Press. Battisteila, E. (1982), More on Hungarian vowel harmony. Linguistic Analysis 9, 95119. Booij, G. (1984), Neutral vowels and the autosegmental analysis of Hungarian vowel harmony. Vrije Universiteit Working Papers in Linguistics. Amsterdam. Campbell, L. (1980), The psychological and sociological reality of Finnish vowel harmony. In: R. Vago (1980b), 245-270. Chomsky, N. and M. Halle (1968), The sound pattern of English. New York. Chinchor, N. (1979), On the treatment of Mongolian vowel harmony. Cunyform Papers in Linguistics 8,171-187. Clements, G.N. (1976a), Vowel harmony in non-linear generative phonology, IULC 1980. Clements, G.N. (1976b), 'Neutral vowels in Hungarian vowel harmony: an autosegmental interpretation', in J. Kegl, D. Nash and A. Zaenen (eds.), Proceedings of the 7th annual meeting of the North Eastern Linguistic Society, MIT department of Linguistics, Cambridge, Mass. Clements, G.N. (1981), Akan vowel harmony: a nonlinear analysis. In Harvard studies in phonology 2 IULC, 108-177. Clements, G.N. (1984), Vowel harmony in Akan: a consideration of Stewart's word structure conditions. Ms. Clements, G.N. and E. Sezer (1982), 'Vowel and consonant disharmony in Turkish', H. van der Hülst and N. Smith (1982), Part II, 213-239.
Hungarian Vowel
Harmony
303
Goldsmith, J. (1976), Autosegmental phonology. IULC. Halle, M. & J.R. Vergnaud (1980), Harmony processes. In: W. Klein & W. Levelt (eds.), Crossing the boundaries in linguistics. Dordrecht, Reidel, 1-22. Hooper, J. (1976), An introduction to natural generative phonology. New York. Hülst, H.G. van der (1984), Syllable structure and stress in Dutch. Dordrecht, Foris. Hülst, H.G. van der & N. Smith (1982) (eds.), The structure of phonological representations. Part I & II. Dordrecht, Foris. Hülst, H.G. van der & N. Smith (1982), Prosodie domains and opaque segments in autosegmental theory. In H. Van der Hülst & N. Smith (1982). Part II, 311336. Jensen, J.T. (1978), A reply to 'Theoretical implications of Hungarian vowel harmony", Linguistic Inquiry 9, 89-97. Kiparsky, P. (1968), How abstract is phonology. IULC. [Reprinted in P. Kiparsky, Explanation in Phonology. Dordrecht, Foris, 1983J. Kiparsky, P. (1973), Abstractness, opacity and global rules. IULC. Kiparsky, P. (1982), From cyclic to lexical phonology. In H. van der Hülst and N. Smith (1982), Part 1,131-175. Kiparsky, P. (1983), 'Some consequences of lexical phonology', Ms. Mohanan, K.P. (1981), Lexical phonology, MIT dissertation. Phelps, E. (1978), Exceptions and vowel harmony in Hungarian, Linguistic Inquiry 9, 98-105. Pulleyblank, D. (1982), Tone in lexical phonology. MIT, diss. Ringen, C. (1977), Vowel harmony: Implications for the alternation condition. In: W. Dressier & O. Pfeiffer (eds.),Phonobgica 1976. Innsbruck, 127-132. Ringen, C. (1978), Another view of the theoretical consequence of Hungarian vowel harmony, Linguistic Inquiry 9, 105-115. Ringen, C. (1980), A concrete analysis of Hungarian vowel harmony. In R. Vago (1980b), 135-155. Ringen, C. (1982), Abstractness and the theory of exceptions. Linguistic Analysis 10, 191-202. Ringen, C. (1984), Vowel harmony: linear or non-linear? Paper presented in Eisenstadt at the 5th intern, phonology meeting. Stewart, J. (1983), Akan vowel harmony: the word structure conditions and the floating vowels. Studies in African linguistics 14. Vago, R.M. (1975), Hungarian generative phonology, IULC. Vago, R.M. (1976), Theoretical implications of Hungarian vowel harmony. Linguistic Inquiry 7, 243-263. Vago, R.M. (1978), Some controversial questions concerning the description of vowel harmony. Linguistic Inquiry 9,116-125. Vago, R.M. (1980a), A critique of suprasegmental theories of vowel harmony. In R. Vago (1980b), 155-183. Vago, R.M. (1980b) (Eds.), Issues in vowel harmony. Amsterdam, Benjamins. Vago, R.M. (1980c), The sound pattern of Hungarian. Washington, Georgetown University Press. Vago, R.M. (1980d), Introduction. In R. Vago (1980b). XI-XX. Vago, R.M. (forthc.), Morpheme-level harmony in a multi-levelled autosegmental framework. Ms. Zonneveld, W. (1980), Hungarian vowel harmony and the theory of exceptions in generative phonology. Linguistic Analysis 8, 21-39.
Speech Disguise and Phonological Representation in Amharic* John McCarthy University of Texas, Austin
Recent work on speech disguise systems or language games, as they are often known, has demonstrated that there is a rich interaction between the processes forming play languages and the theory of phonological representation (Clements 1983; Haraguchi 1983¡McCarthy 1982a, 1982b, forthcoming; Wright 1983; Yin 1983; Yip 1982). Specifically, the facts of a wide variety of speech disguise systems support the conception of phonological and morphological structure developed in McCarthy (1979, 1981) and subsequent works, where a CV-skeleton representing the segmental timing of utterances is posited to be independent of its associated phonemic melody (or melodies, in some cases). Quite generally, even in languages with exclusively concatenative morphological systems, language games may manipulate the CV-skeleton and the phonemic melody entirely separately from one another. In one Arabic language game (McCarthy 1982a), for example, the root consonants are transposed without affecting either the CV-skeleton or the other, vocalic phonemic melody. This is perhaps to be expected, given the overwhelming evidence internal to the language for the division between roots and other elements of the phonological representation. Yet strikingly, in Austronesian languages like Tagalog (McCarthy 1982b) or Bantu languages like Luganda (Clements 1983), we find precisely the same phenomenon in a transposition language game. In these languages, the phonemic melody albeit without the consonant/vowel dichotomy that is characteristic of Semitic - is transposed, leaving behind the representation of segmental quantity by the CV-skeleton. In this note, I will propose an analysis of a speech disguise system, based on the Ethiopian Semitic language Amharic, with a property that has not yet been observed: manipulation of the CV-skeleton while the root phonemic melody remains unaltered. In this respect it contrasts with the Arabic, Tagalog, and Luganda language games mentioned above, •Although this paper was not presented at the workshop, its obvious connections with the work of the other authors represented in this volume seemed to justify its inclusion here. This work was supported by the National Science Foundation under grant BNS8121002.
306
John J. McCarthy
where the CV-skeleton is ignored by the process of speech disguise. All data on the Amharic play language come from the valuable treatment by Teshome Demisse and Bender (1983). The language is used chiefly by what they call "freelance prostitutes" in Addis Ababa, although it is said to have originated with a group of schoolboys' football team. A disclaimer. Like the other Ethiopic speech disguise systems described in Leslau (1964), this Amharic one goes beyond the familiar norm of play languages like pig Latin by augmenting the phonological transformation with a number of lexical and grammatical peculiarities. I shall not be concerned with those here except as they bear on the purely phonological aspect of how the regular processes of speech disguise operate in this case. I shall also give rather short shrift to certain aspects of Amharic phonology, although there is no doubt that closer study of them might illuminate some aspects of the language game. Some regular and phonologically relatively transparent data from the Amharic language game appear in (1): Amharic form a. g w aro gin matt'a ktfu t'affa birr barr t'att'a balla gabba watt'a satt'a hed yet bet dagg wiSat waddada badda son man s'afa k'allala b. kabad wark'
Disguised form g w ayrar gayn an may t ' a t ' kayfaf t'ayfaf bayrar bayrar t'ayt'a t' baylal gayb ab wayt'at' sayt'at' haydad yaytat baytat daygag waystat waydad baydad saym am maynan s'ayfa f k'aylal kaybdad wayrk'a k'
Gloss 'backyard' 'but' 'come' 1 'cruel' 'disappear' 'dollar' 'door' 'drink' 'eat' 'enter' 'exit' 'give' 'go' 'where' Tiouse' Tcind' 'lie' love' 'make love' 'name' 'who' 'write' 'wishy-washy person' 'heavy, difficult' 'gold'
307
Speech Disguise sakkara wassada zaffana t amara saddaba gabbaza k'ald rajjim sigara t'ayyak'a c. tawallagad birc'ik'o
saykra r' waysdad zayfnan taymrar saydbab gaybzaz k'ayldad rayjmam saygrar t'ayk'ak' waylgdad bayc'irkak
'drunkard' 'take' 'sing' 'learn' 'insult' 'invite to' 'joke' 'tall' 'cigarette' 'ask' 'stagger' 2 'glass, drinking
The quite systematic alternations in (1) between Amharic and the play language based on it constitute the vast majority of the forms cited in Teshome Demisse and Bender (1983). It is fairly clear that one or at most two processes are involved in forming the disguised words. Pretheoretically, we can observe that the Amharic form transmits no information to the language game other than its consonants. The canonical form and the vocalism of the base are irrelevant to determing the disguised form. The language game supplies the vocalism - the diphthong ay in the first syllable and a in the second - and it also supplies the canonical pattern assumed by the derived form. This canonical pattern is CVCVC in (la), CVCCVC in ( l b ) , and CVCCCVC in (lc). We can also observe that choice of canonical pattern for the disguised form among these three possibilities is determined by the number of different consonants in the Amharic word. That is, the forms in (la) have just one consonant (like t'att'a) or two different consonants, although they may have two, three, or four actual consonants when geminates and nonadjacent identical consonants are counted. On the other hand, all forms in ( l b ) have three different consonants and the forms in (lc) have four, when the prefix t9 is discounted. The theory of morphophonological representation applied to Semitic languages in McCarthy (1979, 1981) provides a straightforward unitary account of these disparate observations. The language game extracts from the Amharic base form only the consonantal root, the fundamental unit of lexical organization, which is represented on a separate autosegmental tier from either the vocalism or the CV-skeleton. That is, the language game is sensitive to the contents of the root tier alone. Roots in Arabic and Hebrew, and in Amharic as well, are subject to a constraint that is a version of Leben's Obligatory Contour Principle: adjacent identical elements are prohibited. It follows, then, that any reference to the number of different consonants in a surface form is equivalent to referring to the
308
John J. McCarthy
number of consonants in the root itself. Thus, the forms in (la) all have uni- or biconsonantal roots, those in (lb) have triconsonantal roots, and (lc) have quadriconsonantal roots. Finally, the language game itself supplies a CV-skeleton (or rather, a family of them), to which the ay- 3 vocalic melody is preassociated. This skeleton appears in (2):
(2)
CV(CXC)CVC
A
I a
ay
Let us first look at some representative Amharic forms and their disguised counterparts, and then turn to some details of the autosegmental association. In (3) I display words from each of the three types in (1): (3)
Amharic form a. t \
Disguised form ay a
CVC = gm
V I CVCVC = gay nan
1 / gn
gn
3 a
ay 3 M I
CVCCV = t'stt'a
CVCVC = t'ayt'at'
V t' b. 3
CVCCVC V = sskksrd skr
t' ay
3
NI
I
CVCCVC = saykrsr
i a
skr ay
\ \ \ CVCVCV = sigara
\J I CVCCVC = saygrsr
3
\ll/ sgr
I / / sgr c. t3
ay CVCVCCVCCVC = wlgd
t3W3llagged
3
N i CVCCCVC = waylgd3d I//A" wlgd
Speech Disguise
309
The first issue we must deal with is the pattern of association found in the language game forms. Although the original Amharic forms show a wide variety of idiosyncratic, lexically or morphologically determined associations, the disguised forms are all associated according to the unmarked or automatic rules. That is, they have one to one association of all nonfinal consonants and auto segmental spreading of the final consonant. In this respect, they conform precisely to what happens with geminate or ninth binyan forms in Arabic (McCarthy 1979, 1981). It follows, then, that no special rules of association are needed beyond the universal apparatus already justified. A second issue is the selection of the correct skeleton from among the choices embodied by the parentheses in (2). We have already observed that the size of the template - the expansion of the parentheses - is determined by the number of consonants in the root. This sort of behavior is extensively precedented in Semitic; some Arabic skeleta are prohibited with roots of four consonants (generally because one consonant would inevitably be lost), whereas others may not occur with roots of two or three consonants. Although the language game must apparently stipulate the connection between skeleton size and root length, such a stipulation is also countenanced by universal grammar on independently motivated grounds. An alternative account of these two problems can also be considered. We might suppose that the autosegmental spreading in the language game is not a result of universal conventions but is rather stipulated in the representation of the skeleton. In other words, the grammar would explicitly require that the last two C-slots of the skeleton in (2) be associated with the same segment. With this in hand, we can then appeal to universal considerations for the connection between skeleton size and root length. In McCarthy (1984) it is proposed, based on data of a quite different sort from Modern Hebrew and Cupeno, that optional skeletal slots are expanded only when some phonemic material would otherwise remain unassociated; in effect, the necessity of expressing all lexical material phonetically forces association with optional slots. The interaction between the stipulation about the manner of association and this universal principle also generates the correct constellation of facts. Since I know of no data that select between these two quite different accounts, and since they are roughly equivalent in explanatory power, I will leave this issue undecided. In many cases the source of the disguised form is not Amharic proper but English; generally we find that these loan words behave according to the analysis developed above:
310
John J. McCarthy
(4)
English form a. cash cassette date (appointment) gin (drink) shit! whiskey b. artificial Mirinda isprayt (Sprite)
Disguised form kaysas kaystat daytat jaynan say tat wayskak ayrtatfayslal mayrndayda d ayspraytat
The forms in (4a) require no further comment, given that some approximation to English phonetics is the input to the speech disguise system. The words in (4b) are all, in their undisguised forms, in flagrant violation of the canons of Amharic word structure. As Teshome Demisse and Bender (1983) suggest, they are split into two quasi-words for the purposes of the language game, a phenomenon that is quite well attested in many other languages (McCarthy forthcoming). This is particularly clear in the case of artificial, where the foot-sized units arti and ficial are separately input to the language game. The facts discussed to this point are representative of the most straightforward and general aspects of this Amharic speech disguise system. With some phonological and morphological background, most of the remaining data in Teshome Demisse and Bender (1983) are readily explicable as well. First, affixes of various sorts are ignored in the construction of the disguised form. This follows from the insensitivity of the language game to any aspect of the base word other than its root, since affixes are obviously not part of the root. We have already noted this fact in connection with (la), and it is also attested in maynkak from Amharic mankiya 'spoon'. Second, Amharic does not contrast labialized and nonlabialized consonants before round vowels. It is apparently the case that consonants in this environment are underlying (redundantly) labialized, because the labialization appears overtly before nonround vowels in the disguised forms. This fact is responsible for a large number of surface alternations attested in the following forms: (5)
Amharic form bunna k'onjo t'armus sukka k'ollaf
Disguised form b w aynan k' w aynjaj t'ayrm w aysas s w aykak k' w aylf a f
Gloss 'coffee' 'beautiful' 'bottle 'fork' 'lock'
Speech Disguise muzik'a k'ut'/r suk' c'oha k'ucc
311 m w ayzk' ak' k' w ayt'rar s w ayk'ak' c' w ayhah k' w ayc'ac' fWaydad v w aymt at
'music' 'number' 'shop' 'shout' 'sit down' 'food' 'vomit'
Apart from the emergence of underlying labialization, these language game forms are unremarkable. A final set of phonological effects in the speech disguise system is observed with the vowel-initial words in (6). In these forms, the initial C slot of the derived CV-skeleton apparently remains empty, contrary to what we would expect from the association procedure. 4 (6)
Amharic form ijijia me anta anci addara
Disguised form ayjiajia ayn ane ayntat ayncac aydrar
Gloss 'we' 'I' 'you (m. sg.)' 'you (f. sg.)' 'stay over night'
The explanation for this apparent anomaly is quite straightforward: the roots of all of these words have initial h, associated with the first C slot, which is deleted in all contexts (Bender and Hailu Fulass 1978:47). We may prefer to eschew this relatively abstract treatment in a complete phonological treatment of Amharic, but in any case the question is totally independent of the language game discussed here. This exhausts the data except for a few patently irregular forms that are noted as such by Teshome Demisse and Bender (1983). I conclude, then, that the CV-skeleton in (2), augmented by some stipulations about association, is a sufficient account of the formation of this speech disguise system in Amharic. What is of particular interest here is the similarity between this language game and the ordinary morphological resources of Semitic languages. Not only is the language game parasitic from the root-pattern dichotomy that underlies Amharic morphophonological representations, but it also supplies a CV—skeleton just as the Semitic processes of verbal derivation characteristically do. This observation supports not only the conception of morphophonological structure developed here but also the notion, developed in McCarthy (1982b, forthcoming), that language games are a kind of functionally empty morphology, sharing significant formal properties with the nonconcatenative morphological systems of the world.
312
John J.
McCarthy
NOTES 1. The apostrophe indicates glottalization. Geminate glottalized consonants are written with a single apostrophe. 2. The t3 sequence in the Amharic form is a prefix, which is systematically ignored in the language game, as I show below. 3. The i in the disguised form is a vowel cpenthesized into the medial triconsonantal cluster by a phonological rule (Bender and Hailu Fulass 1978:38). The disguised form is irregular in its loss of glottalization. 4. The first person pronouns in (6) are exceptional in retaining final vowels in the disguised form. REFERENCES Bender, M. Lionel and Hailu Fulass (1978) Amharic Verb Morphology, African Studies Center, Michigan State University, East Lansing, Michigan. Clements, G.N. (1983) "Compensatory Lengthening and Consonant Gemination in Luganda," unpublished paper, Cornell University, Ithaca, New York. Haraguchi, Sh. (1983) "An Autosegmental Theory: Its Expansion and Extension." Journal of Linguistic Research 2: 59-80. Leslau, W. (1964) Ethiopian Argots, Mouton, The Hague. McCarthy, J. (1979) Formal Problems in Semitic Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Mass. [(1981) Indiana University Linguistics Club, Bloomington, Indiana.] McCarthy, J. (1981) "A Prosodic Theory of Nonconcatenative Morphology," Linguistic Inquiry 12:373-418. McCarthy, J. (1982a) "Prosodic Templates, Morphemic Templates, and Morphemic Tiers," in H. van der Hülst and N. Smith, eds. The Structure of Phonological Representations 1, Dordrecht: Foris. McCarthy, J. (1982b) "Language Games in Nonconcatenative Morphology," paper presented at University of Chicago, University of Illinois, Harvard University, University of Massachusetts, University of Toronto, New York University and Pennsylvania State University. McCarthy, J. (1984) "Prosodic Structure in Morphology," in M. Aronoff and R. Oehrle, eds., Language Sound Structure, MIT Press, Cambridge, Mass. McCarthy, J. (forthcoming) Formal Morphology, MIT Press, Cambridge, Mass. Teshome Demisse and M. Lionel Bender (1983) "An Argot of Addis Ababa Unattached Girls, " Language in Society 12:339-347. Wright, M. (1983) A Metrical Approach to Tone Sandhi in Chinese Dialects, Doctoral dissertation, University of Massachusetts, Amherst. Yin, Y.-M. (1984) "Reduplication and Fanqie languages in Chinese," Texas Linguistic Forum 23. Yip, M. (1982) Reduplication and CV-skeleta in Chinese Secret Languages.Linguistic Inquiry 13:637-662.
Empty V-Nodes and their role in the Klamath Vowel Alternations* Christine ter Mors State University of
Groningen/ZWO
1. INTRODUCTION
What follows is an analysis of some important vowel alternations in Klamath within the framework of autosegmental phonology. All the data can be found in Barker (1963; 1964). The analysis draws heavily on that of Clements & Keyser (1980, henceforth C&K), but avoids the problems raised by their solution. The crucial developments in the present paper, which permit a more insightful account of the phenomena, result from the interaction of two assumptions: 1) 2)
a process of syllabification that has as one of its effects the identification of empty V-nodes, and a general rule of a -insertion that fills such empty V's providing certain syllabic conditions are met.
The problems posed by Klamath affixation are the following: 1) Prefixation results in an alternation between full vowels, a, and 0 i n the prefixed morphemes. The verbal form conwa 'he vomits', for instance, may combine with the causative prefix h Vs (where V stands for a variable vowel) and then surfaces as hoscanwa 'he makes vomit', while the distributive of paga 'he barks', namely /CV+pag+a/ (where the reduplicative prefix copies both the initial consonant and the vowel of the following morpheme) assumes the surface shape of papga 'distr. bark'. 2) There is also an alternation between glides and vowels: [w,W,w], and [y,Y,y] (i.e. the plain, voiceless, and glottalized variants of [w] and [j]) respectively alternate with [i(:)] and [o(:)]. Underlying /somalw/ forms the basis of both somalwo. ts 'writing instruments' and somalo:ye:ga 'starts to write'. Likewise /taby/ underlies both tapyap 'man's younger brother' and tabi.ni 'the youngest'. 3) Klamath also displays alternations between 0 a n d a within * I hereby want to especially thank Frank Heny and Deirdre Wheeler for their help in the development of the present analysis. This research was partially supported by the Foundation for Linguistic Research, which is funded by the Netherlands Organisation for the advancement of pure research, ZWO.
314
Christine ter Mors
certain consonant clusters. Forms like/tgalm/ 'west' and/gmocdk/ 'old, old person' respectively turn up as tgahm and gmocstk. It will be shown that a unified account can be given of all these alternations involving glides, full vowels, a's and 0, and that the present theory based on the idea of empty V-nodes provides solutions where former theories (Kisseberth 1973;Kean 1974; Thomas 1974; C&K) have failed.
2. THE FRAMEWORK.
The autosegmental framework of the present analysis follows C&K in representing the syllable as a hierarchical unit consisting of three levels: the a-tier (a being the symbol for syllable), the CV-tier (consisting of strings of the elements C and V) and the segmental tier, see (1). (1)
a-level CV-level segm. level
a
a
AS A IMIII s mo q ya
CCVCCV
It has not been necessary to postulate any additional levels in accounting for the processes under consideration. The analysis follows C&K also in employing syllable templates that constitute the a - and CV-tiers, and actively structure the segmental material coming from the lexicon. The template in the present theory is a sort of syllable 'model', a configuration of C's and V's to which all Klamath syllables must conform. These C's and V's stand for positions that can be encountered in the syllable; they are purely structural, phonetically uninterpretable units. One of the hypotheses of our analysis is that we can build such a template for any language by analyzing its surface forms and deducing from these the generalizations and restrictions on the form of syllables in that language. In this sense the template can be seen as a universal notion, as a definition of the syllable, with each language having to define for itself which segments can fill which positions in a syllable of that particular language. The Klamath template we will be working with is given in (3). It reflects the restrictions on syllable structure that can be drawn from the Klamath surface forms, a representative list of which is given in (2). (2)
c/yemis 'hunger': initial obstruent + resonant tqopo: 'thumb': initial obstruent + obstruent waicamc 'big old coyote': final resonant + obstruent
Empty V-nodes in Klamath cawiga/A: 'crazy person':
315 final obstruent + obstruent
'neighbour':
yswkkm
medial res.+obs.+obs.+res.
swafoAiis 'looking like inside': medial obs.+obs.+obs.+obs. neivfg-i 'runs d o w n ' :
medial res.+obs.+obs.
klocwa 'fishes by torchlight':
medial obs.+res.
cla^vamna 'carries a massive object around': medial res.+res.
(3) (C)
C
V
[ - son]
( C ( C ) ) [ - son]
This template encodes the following: a)
A syllable minimally consists o f one C f o l l o w e d by one V .
b)
Consonant clusters usually 1 contain no more than t w o consonants; word-medially
a word may have as many as four consonants,
analysable as a sequence o f a syllable-final and a syllable-initial cluster. c)
As indicated by the [ - s o n ] specification o f the syllable-marginal C's o f the template, resonants (liquids, nasals, and glides) must always be adjacent to vowels. 2 This implies that there are no w o r d initial or w o r d - f i n a l resonant clusters; these may only occur w o r d medially where they exhaustively constitute the cluster and belong to separate syllables.
A crucial assumption o f the present analysis is that in Klamath only true vowels can fill V-positions, and that all other segments function as C's. In cases where there is no underlying vowel t o fill the V - p o s i t i o n , this position is left open until, at a later stage o f derivation, it is filled or deleted. Such an empty V will be represented as t , indicating that - a s an obligatory part o f the syllable-the V is present though empty.
3. THK ANALYSIS.
3.1. Empty V-Nodes and the (f) / a Alternations. A striking result o f the above described framework, hingeing on template ( 3 ) , is that it allows a straightforward account o f the 0 / a alternations as in /gmoSdk/: gmocatk, given below.
and /tgalm/: tgahm.
Some more examples are
316
Christine ter Mors
(4)
/koml/ /b+e:+ys+m/' /sosanrigs/ /delwgs/ /sgoyw+alcwi/ /s?aywgs/
komal be:ysam sosan?aqs delwaks sgoywalcwi s?aywaks
'pelican' 'daughter's' 'wrestling' 'attack' 'send someone right up to' 'knowing'
Syllabification takes place - we assume - by mapping the template onto the segmental tier. We represent this in terms of an association of the segments to the C's and V's of the template. To derive the correct surface forms it appears necessary to assume a directional element in the association process. Consider the syllabifications of /gmocdk/ in (5). We see that a random mapping of the template onto the segmental tier is not enough. Just given the restrictions of the template /gmocdk/ could be syllabified in several ways. (5)
a. a
a
A A
b.
a
a
/l\ A
ccvcvcc
ccvccvc
gmoc0dk
gmocd0k
Mill
or
a
a
IV !\\ mini gmocdk0
ccvcccv
All forms in (5) contain acceptable syllabifications. However, (5a) is the only one that can possibly yield the correct surface form gmocsdk. Therefore, to get syllabification (5a), we must assume that the Rule of Association that links up the phonemes to the CV-tier, works from right to left (RL), and also maximizes C-clusters. Only then will we get syllabification (5a) with dk hooked up as the maximal syllable-final cluster. We accordingly formulate the Rule of Association as in (6). (6)
Rule of Association '. Map in a RL fashion the syllable template onto the segmental tier; maximize clusters.
If we now syllabify all the forms of (4) according to this rule (6), we get the following structures:
(7)
a
a
CCVCVC
A A
C VC VC
A A
CVCCVC
A A
t g a 10 m
k o m0 1
b e: y s 0 m
I I II I I
317
Empty V-nodes in Klamath a
a
a
a
a
K A\ /IV
/t\ /IV 1111 I I I
c c v c v c c
c v c v c c v c c
g m o c 0d k
s o s a n n 0g s
a
A
a
a
a
a
/IV A \ /K
/ I V c v c c v c c
c c v c c v c c c v
d e 1 w 0g s
sgoyw01cwi
1111 I I I ] 11
a
a
/TV / V
c c v c c v c c
s ? a y w 0g s
All we need add now, to derive the correct a -containing surface forms, is a rule that fills these empty V-nodes with a . This rule is formalized below. (8)
a -Insertion:
0
a/Y
With the help of this rule the examples in (7) attain their surface forms of tgalam, komal, be.ysam, gmocstk, sosan?aqs, delwaks, sgoywslcwi, and s?ayw3ks. An interesting point of the here assumed syllabification according to rule (6) is that it also excludes the o ^/"generation of s's as in hypothetical forms like gamocadaka or detewagasa (cf. 7). The maximizing of clusters restricts the occurrence of empty V-nodes to positions where this is expressly required by the template condition that each syllable have a vowel. Without this maximizing principle ^'s - and therefore a's - might be inserted in every possible position permitted by the template, i.e. after each C that is not followed by a vowel, thereby forming a long line of minimal syllables as in the incorrect forms above. We may therefore see this cluster maximizing as an expression of what 1 call the 'Minimal Loss of Identity Principle' (MLIP): the general tendency of languages to allow only strictly necessary changes in their underlying forms. This is, of course, related to all theories of markedness, in the sense that the less a rule or process changes the underlying structure, the less marked it is. 3.1.1. Advantages over Kean (1974) and C&K. Support for the present directional syllabification with a -insertion in empty V-nodes is found in a comparison with the analysis of Kean (1974). Kean concluded that the 3-insertion process in Klamath was triggered by certain clearly definable resonants ( /w/ and f+soiJ)> which she therefore mentioned as essential subparts of her two rules: 'Idiosyncratic /w/'
318
Christine ter Mors
( 9 ) and 'Sonorant Cluster' ( 1 0 ) . Both rules are followed here by some examples they account for.
(9)
(10)
Idiosyncratic I w/: 0 -*• > 3/ / [#xvrC
L+son
-jw-C(CZ) #](186)
/delwg+s/
~(9)-*
delwaks
/s?aywg+s/
—(9)->-
s?ayw3ks
/sgoywléwi/
-(9)-»-
sgoywalSwi
Sonorant Ouster. 0 ^ 3 / [#XC - ^ ^ J ( C Z ) #] (189) /tgalm/
-(10)^-
tgabm
/b+e:+ys+m/
—(10)->-
be:ys3m
/koml /
-(10)>
komal
Being purely linear, however, her analysis could not exploit the insights gained in a syllabic framework. For - with our present system o f syllabification identifying ^'s - it is possible to predict where Kean's rules will go wrong, namely in those cases where the template gives us the empty V - n o d e before instead o f after the /w/ o f rule (9), and also where it gives the after instead o f before the resonant o f rule ( 1 0 ) . In ( 1 1 ) below are given some examples o f this, set next to the correct surface forms as derived through the present analysis.
Empty
V-nodes in Klamath
319
(11)
/wqeplgs/'summer house'
a
° *
Kean: wqepalgs
a
/I / f c s I I I M II
cvccvcc (10)
present. analysis
*wqepalqs (other rules) o
w q e p 1 0 g s (6) w (8) o q e p 1 a g s wq e p 1 a q s (other rules)
/sgoyw+cn+a/ 'sends along someone' Kean: *sgoywacna
(9)
present. analysis
a
a
a
A/K/K 111111 M I s g o y 0 w c n a (6)
ccvcvcccv
sgoyswcna
(8)
s g o y o : c n a (14) (Vocalization discussed in 3.2.). /sgoyw+cn+bli/ 'sends someone back' Kean: sgoywacribli
sgoywacanbli
(9)
present . analysis
(10)
*sgoyw3C3mbli (other rules)
a
a
a
a
A A A A IIIIHiljlil sgoy0wc0nbli
ccvcvccvcccv (6)
s g o y s w c s n bl i (8) s g o y o : c 3 n b l i (14) (Vocalization) s g o y o : csmbl i (other rules)
Another point missed by Kean's analysis, but crucially predicted by the present system, is that 3 -insertion not only occurs next to certain resonants, but also in non-resonant environments where a vowel is lacking. As the forms gmocatk and sosan?sqs of (7) illustrate, none of these a's can be predicted by Kean's rules, since both these rules have resonants triggering the insertion process. The advantage of the present analysis of 3 -insertion in empty V-nodes
320
Christine ter Mors
over the rules given by Kean, is not only that we have reduced a -insertion to one single general principle, but also that it explains why we get a - i n sertion, namely because the nucleus of the syllable concerned is empty. Apart from this it should be pointed out that the third row of forms in (7) presents overwhelming counterevidence against the analysis as given by C&K. C&K also worked with template association, but their template was formulated as: a (C)
c
v
(C
(C))
I I I [-son] [-cons] [-son] A crucial difference from the present template (3) is that they allow for glides under the V. What they proposed was that in cases where the underlying structure lacks a vowel to fill the V (e.g. a )-and ccY^ s q oWwk would therefore be ill-formed in surface structure-, but contains a glide in a position where it can function as syllable peak, this glide is linked up to the V, and there vocalizes to the corresponding vowel (e.g. a a -*• ccv 1 ' 1X I I s qoWwk sqoWo:k). Though this analysis could account for a lot of glide/vowel alternations like those in (13, 15) below, it offers no explanation why, in forms like (7): delwsks, sgoywslcwi, and s?ayw9ks, the glide may not fill the V-position, and why we find instead a-insertion in these forms. Their template association would derive incorrect surface forms as in (12).
:
(12)
C&K
a
a
A / A \ c v c vcc 111111 Id e 1 w g s/ -»• *delo:ks a
a
/K /t\
a
A\
c c v c v c c c v
/sgoywlcwi/-* a
A\
*sgoyo:lcwi
a
/ I V c c v c v c c /s ? a y w g s/ -*• *s?ayo:ks
Empty V-nodes in Klamath
321
3.2. The Glide/ Vowel Alternations: Empty V-Nodes Again. In the present system with template syllabification identifying ^ ' s and a-insertion rule (8), the glide/vowel alternation forms mentioned in section 1 (e.g. tapyap versus tabi.ni) can be directly related to the a / 0 alternation forms accounted for above. For what the present framework shows us (see (13)), is that in all those cases where these alternating glides surface as vowels, the underlying structure is such that those glides fill a syllable final C-position, in front of which our template syllabification (rule (6)) identifies an empty V-node. (13)
/tabyni/ 'the youngest'
a
a
N f\N
/somalwye:ga/ 'starts to write'
a
a
a
a
A / T \ c c v c v c c
(6)
/delwga/ 'attacks'
a
1111111 s q oW0 w k
a
1 1 1 I I I I I I I I somal0wye:ga
(6)
/sqoWwk/ 'because of being spring'
a
a
A AA A A c v c v c v c c v c v
c v c v c c v 1 1 1 I I 1 1 t a b0y n i
a
a
a
A A A c v c v c c v 1 1 1 1 I I I d e 10 wg a
(6)
(6)
The empty V's trigger the application of our a-insertion rule (8), and the full vowel surface forms can be derived through an additional rule of vocalization (14, cf Kean (185)), which turns all tautosyllabic a +glide sequences into vowels homorganic with the original glides as in (15) below. For a discussion of the length of these vowels see Ter Mors (1983).
(14)
Vocalization: r - c o n s -j"| r+voc -| a - voc. II -> a round LaroundJI L+high J J
(15)
a
a
o
a
A A A c v c v c c v 1111111 t a b0 y n i t a b ay n i t a b i : n i
A A A A A c v c v c v c c v c v (6) (8) (14)
11 { 1 1 1 1 . s o m a 1 0 w ye g a (6) s o m a 1 a w ye g a (8) s o m a 1 o : ye g a (14)
322
Christine ter Mors a
a
a
A\ /TV ccvevcc s q oW0 wk s q oWa wk s q oWo : k
a
A A
a
K
cveVCCV (6) (8) (14)
d e 1 0wg a d e 1 a wg a d e 1 o : g a
(6) (8) (14)
Under the present analysis the only distinction between these glide/vowel alternations of (15) and the a -insertion cases of (7) is that in the forms of (15) the 3 is inserted before the glide, where the resulting a+glide sequence feeds Vocalization, while in the forms of (7) the a is inserted after the glide, and is therefore retained in surface structure. In this way the present system reduces the two alternations, i.e. the one between 0 and a (/tgalm/: tgatam), and that between glides and full vowels (tapyap versus tabi'ni), to "one single phenomenon: the identification of empty V-nodes in the syllabification process. This is a generalization that could not possibly be expressed in any of the previous analyses. Additional support for the present analysis with inserted a's in the cases of (15) is provided by Barker, who lists alternative surface forms such as e.g. niqygwwi for niqyo.wi, from underlying /niq+aywwi/. At no stage of C&K's derivation is there a 3 in this form, while under the present analysis ( a a a -(8)> n i q y s w w i - ( 1 4 ) + niqyo.wi)
c'V e'evt cv
! ' ' A ' I ! n i q y0 wwi it is only necessary to assume that Vocalization for some reason has not applied, or does not apply obligatorily in this form.
3.3. Vowel Deletion: New Perspectives. The third alternation problem of Klamath is that encountered in forms like conwa versus /hVs+conw+a/: hosc3nwa, and paga versus /CV+pag+a/: papga. The relevant data all show full vowels alternating with a ' s or 0 , wherever the syllable containing such vowel underlyingly is prefixed by a reduplicative morpheme copying its vowel and sometimes also its initial consonant (cluster). Among these reduplicative prefixes are the causatives snVand hVs, the reflexive sV, the prefix IV(meaning: act on a round object), and the distributive CV. For more examples of the copying process, consider the items in (16).
Empty V-nodes in Klamath (16)
a. po:ka /lV+po:lc+a/ ewe: lea /snV+cwe:k+a b. conwa /hVs+conw+a/ qlin / snV+qlin/ c. paga /CV+pag+a/ po:ka /sV+lV+po:lca/ /coq+peq+a/
323
lopo:ka -*• snecweilca hoscanwa -»• sniqtan -* papga -*• solpo:lca -> coqpqa
'puts on warpaint' 'puts warpaint on someone' 'is tough' 'makes tough' Vomits' 'makes vomit' 'choke' 'causes someone to choke' 'barks' 'bark (distributive)' 'puts on warpaint' 'puts warpaint on oneself 'puts the buttocks in someone's face'.
What we see is that the reduplicative prefix always contains a short copy of the first vowel following. If this vowel itself is long, it remains as in (16a). If short, it reduces to a(16b) or deletes (16c). In a syllabic framework, it is fairly easy to predict whether we will find a a or 0 . A a appears where the original vowel occupied a closed syllable (cf. template (3) and syllabification rule (6)), and 0 surfaces in case the original vowel was syllable final. A widely accepted way of dealing with these facts is therefore (cf. Kean 1974; Thomas 1974: C&K) to assume two rules: Vowel Reduction (VR) and Vowel Deletion (VD), with VR reducing all prefixed short vowels to a , after which VD deletes all a ' s that occupy open syllables. Below is given C&K's version of these rules together with some exemplary derivations. (17)
a. Vowel Reduction3
: [+syl, -long] -»• / [+syl] [-syl] 0 +[-syl] Q -
b. Vowel Deletion
: 3 -»• 0 / -]Q
c. hVs + conw + a hos + conw + a
Reduplication
a
a
hoscanwa
a
CV + pag + a pa + pag + a
a VR VD output
a
a
pa + p g + a papga
The present template analysis allows us to see these processes of VR and VD in a wholly new perspective. To begin with the directional element in the syllabification process makes it possible to rephrase the above rule of VD as a simple restriction on the association of the segmental tier.
324
Christine ter Mors
According to our Rule of Association (6), segmental strings are syllabified by means of a RL association of the segmental tier to the template. The surfacing of the VR a's in open syllables can be prevented by assuming a restriction (18) on the association of a's as syllable final elements.
(18)
Open Syllable Condition :* a r - x
v 3I
(where x indicates the absence of any element in that position) This condition prevents the association of 3's in syllable final positions, which - under the general convention that non-associated elements do not surface-accounts for the fact that we find no Klamath forms with VR s's in open syllables. (19) illustrates the effects of (18) on the VD forms of (16). (19)
/CV+pag+a/ pa+pag+a papsga
a
(17a)
a
/ I V l\ CVCVCV M i i i pa p 3g a p a p ga /coq+peq+a/ coq+paq+a
(6 + 18)
(17a)
/ I V . N CVCCVCV I I I * c o q p 3q a c o q p q a
(6+18) (output)
Reformulated as a condition on the association of the tiers, VD is now an integral part of the syllabification process. Also, the formulation of (18) quite rightly is that of a general restriction on 3's in open syllables, for it is true that the 3's of Klamath can never occur in that position. (There is a limited set of apparent counter examples to this condition
Empty V-nodes in Klamath
325
(18) containing forms like e.g. /de+dmesg+a/ that surfaces as dedmasga. RL syllabification would presumably link up o ^ as the final syllable, CCV i ii
sga
leaving the a in syllable-final position, where - b y condition (18)- it would delete. If this appears to be more than an exceptional case, we must invoke the power of the MLIP discussed in section 3.1. MLIP must then be assumed to try and protect a a from deleting by drawing a C into the syllable if possible; the syllable is then closed and the a saved
a
o
a
CVCCVCCV
11111111
a
MLIP
a
a
CVCCVCCV
111
(dedmasga ->• d e d m a s g a). It is to be expected that this can be done without any problems in all cases where the a is followed by two C's since according to the template the following syllable needs only one initial C and can easily "lend" one to its left neighbour.) However, we seem to be missing a generalization, if on the one hand we have a's created through a process of VR, while on the other hand we regard a's as the fillers of empty V-nodes (cf. rule (8)). The obvious question to be asked is: can we relate these VR a's too, to empty V-nodes and the a-insertion rule (8)? And in the present framework the answer is that this is certainly possible. Let us reconsider the forms of (16). We have assumed till now that morphemes came from the lexicon, and went through the affixation process, where VR (17a) reduced the appropriate vowels to a . The words then entered the phonological component where syllabification (6) took place and the subsequent rules of a-Insertion (8) and Vocalization (14) applied. If the VR a is now also to be accounted for through a-insertion in an empty V-node, we must not have a VR rule that in the course of affixation process has as its output a a (/hos+conw+a/ (l_7a) hoscanwa), but instead we need a rule that replaces the orginal vowel by ^ (/host+ Sonw+a/-»- hos2 Y nwa). To get a ^ in this position, though, the segements must -already underlyingly- be associated to V's and C's; only then is it possible to formulate a deletion rule that leaves empty V's as in CVC CVCC V CVCCVCCV 1 i t J. i i < • , il Li ii . . , , /h o s+c o nw+a/ h o s cy)nwa. Assuming that we have such an early association of the segments to C's and V's a single rule of VD, as in (20), can replace the earlier rules of VR (17a) and VD (17b or 18). (20)
Vowel Deletion4 (revised): [-long] -»• 0 / VC Q + C Q V.
If such an empty V-node created through rule (20) occupies a closed
326
Christine ter Mors
syllable, it is automatically filled by our rule of a -Insertion (8), as the examples of (21) show; we no longer need a separate rule of VR to account for such a's. (An accordingly revised formulation of the association process follows below). (21)
a CVC
CVCC
V
CVC
CVCC
V
/K
a A
a A
cvccvccv
I I 1111 I 111 1111 • gagirka
As we have seen, open syllbales do not have a vowel surfacing in such CVCVCV V-positions (/p a p a g a I
P a P8 a )- W e
first
accounted for this through
a restriction (18) on the association of 3's in open syllables (18) ^
(/papaga/
(17a)
papaga
(6,18) a a ^ / f ^ A ^P a P8 a )CVCVCV i I i * i i p a p a g a
But now that we assume C's and V's already at an underlying level, and with a rule of VD as (20), we must revise the open syllable condition so that it prevents the association not of a's, but of Y's as the rightmost element of a syllable. This revised version is given in (22), with some examples in (23) (22)
Open Syllable Condition:
* a r - x
v i
(where x indicates the absence of any segments in that position.)
Empty V-nodes in Klamath
327
a
/K A
cv c v c v (23)
a
cvcycv
|| I I I | (20,6,22) | | | * | | /CV + pag + a/ -»-p a + p a g + a -* p a p 0 g a ^ papga
a
a
/^k A 20 622 III ill I ( - ' )c o111 [ I! I /coq + p e q + a / -»• q p 0 q a -*• coqpqa CVC CVC V
cvccvcv
The here assumed lexical association o f segments to C's and V's o f course requires a revision o f the association process. Instead o f having syllabification take place by associating the segments with the C's and V's o f the template (rule ( 6 ) : /gagayka/ -*• a a^ a^ ), the process must be rec v c v c c y p a g a y k a presented as an association o f the C's and V's (with the segments linked up to them) to the a's on the a - t i e r conforming to the syllable template ( g a g ). The accordingly revised Rule o f AsCVCVCCV CV CVC CV i i i : i ,i i i i i j. ' :> i g a g0 y k a g a g 0 y ka sociation is given in ( 2 4 ) (24)
Rule of Association (revised): Syllabify all elements on the CV tier in a R L fashion according to the template. (Implying that, as before, Y is used to represent the lack o f a vowel where the template requires its presence.)
This revision does not come out o f the blue. It should be recalled that we stipulated the template as a universal representation o f the " m o d e l " syllable o f a language, with each language having to define somewhere which segments in that language can fill C - and V - positions. Assuming that this is done by labelling segments C or V - with C meaning C-filling element and V V-filling element - , the lexicon is just the place where we expect this labelling to happen. Therefore in the present framework it is only to be expected that at the input to the phonological component all segments are already associated to C's and V's. Additional arguments for this early association o f segments are provided by Marantz ( 1 9 8 2 ) . He showed that the presence o f C's and V's at the affixation stage allows a good insight into the processes o f reduplication; that reduplication can be described in such a system in a relatively simple manner, i.e. as a filling in o f the reduplication C's and V's with the
328
Christine ter Mors
melodic segments associated to the reduplicated C's and V's. In this line of thought the reduplicative process can be represented as in (25). (25)
CVC
CVCC
V
I M III I
h CV
CV
s+conw+a
l l l | VI
CVCC
CVC
CVCC
V
III l III I
-*• h o s + c o n w + a
II III] IV
CV C V C C
+ gayk+a
-»• g a + g a y k + a
CVC V + pag +a
CV C V C V pa +pag +a
The revision proposed here in the association process is certainly essential. I will not, however, go back and revise all the examples cited, but simply assume a reformulation of all preceding forms and derivations in accordance with these revisions. We may conclude that also the 'VR' a's can be retraced to empty Vnodes. This analysis allows an equal treatment of these 'VR' 3's - which are under the present analysis inserted in ^'s created by VD - and the other insertions a's, which were argued to derive from ^'s identified by the template syllabification process. We have now also provided the link between the VD cases and the Vocalization cases discussed in 3.2. above. As illustrated by e.g. the second case under (21), a's that fill VD ^'s, just likethea's of (15), do not turn up in surface structure when they precede a glide. Instead, we find full vowel forms in surface structure, which is exactly as predicted by our Vocalization rule (14). This justifies the proposal of that rule. Some more examples are given below.
(26) cv cyco V I
+ g a y lc+a/ 'distr. are silly'
CV CVCC V
II UK I
g a+g a y lc+a
Reduplication
ii mi i
(20)
CV CVCC V g a+g0yk+a
Empty V-nodes in Klamath a
a
329
a
AAA
cvcvccv
III11II
gagpylca g a g a y lea g a g i : lea
(24) (8) (14)
CCV CCVC CC V
Mil I I I
I
+ p n i w + b c + a / 'distr. blow out'
III III! Ij I
CCV CCVC CC V pni+pniw+be'+a
Reduplication
III Mil II I
CCV CCVC CC V pni+pn0w+bd+a
(20)
IKa / /a K Ka CCVCCVCCCV 5
II1III fill
pnipn0wbca pnipnawbda pnipno:bca pnipno:pca
(24) (8) (14) other rules
Those that are acquainted with C&K's analysis, will remember that they assumed a vocalization of glides in V-position. Vocalization cases like those under ( 1 5 ) therefore contained no a underlyingly, but were an-
a r\ a a
/N
a
M i l l '
I I
a a /-s, a a
/N
a
a
/N.
a Is q o W w k /
b)
a
CCVCV^C , or i i i 1 1 i1 sqoW0wk
created b y t h e morphological process o f V o w e l D e l e t i o n ( 2 0 ) , as in C V CVCC V
CVCVCCV
,11 III' I - 11 MM I •
Ig a + g a y k + a
gag(/)yka
C o m b i n e d w i t h a restriction o n e m p t y V - n o d e s in o p e n syllables ( 2 2 ) , an insertion rule that fills all remaining ^ ' s w i t h a ( 8 ) , and V o c a l i z a t i o n rule ( 1 4 ) , the analysis accounts for the K l a m a t h glide / V / a / 0 alternations as being all reflections o f the Klamath R L t e m p l a t e syllabification process.
NOTES 1. What is not discussed in this paper is that word-final clusters may contain not just two, but sometimes three or even four consonants, the last one or two of which then must be coronal. There is not much to be said about these; they are one of the irregularities of the language, and can be taken care of by allowing word-final clusters one of two extra coronal C's. This can be taken care of by a rule like the one given by Lapointe and Feinstein (1981), 31):
In I C
Out [+coronal|
The C is the final C of our syllable template; the consonants linked up to this C must obey the template conditions. The optional extra C's at the end of the word must be coronal as specified by , . P u t ... Some of these clusters are found at the end of , ., . . 1+ coronal] the followings words: golks lo:Lapks pakst gatbanwapkst
'pintail duck' 'believer' 'barking' 'arriving in the future'.
2. Résonants may sometimes be found in other environments, where Barker describes them as syllabic and not affecting the stress pattern. I assume that these segments are lexically marked for being able to exhaustively fill a syllable. Examples are: [m]sas - prairydog, [n Jcets - bark, |1 ]mac - metate, [w u ]so - chest. 3. The boundary symbol and the lack of reference to syllable structure make the rule seem a non-syllabic one. In the following sections it is shown how the present
Empty V-nodes in Klamath
333
analysis provides the means to account also for VR in a purely syllabic way, retracing 3's to empty V-nodes. 4. Since this rule mentions the (morphological) boundary symbol as an essential subpart of its Structural Description, it is assumed that VD is a morphological rule and applies before the segments come into the phonological component. 5. be is not a permitted a-initial C-cluster, therefore b is syllabified with the a to its left. REFERENCES Barker, M.A.R. (1963) Klamath Dictionary, University of California Publications in Linguistics, Volume 31, University of California Press, Berkely and Los Angeles. Barker, M.A.R. (1964) Klamath Grammar, University of California Publications in Linguistics, Volume 32, University of California Press, Berkely and Los Angeles. Clements, G.N. & S.J. Keyser (1980) "The Hierarchical Nature of the Klamath Syllable", Unpublished manuscript, Harvard University and M.I.T. Kean, Mary-Louise (1974) "The Strict Cycle in Phonology", Linguistic Inquiry 5.2, pp. 179-204. Kisseberth, C.W. (1973) "On the Alternation of Vowel Length in Klamath, A Gobal Rule", in M.J. Kenstowicz and C.W. Kisseberth, eds., Issues in Phonological Theory, Mouton, Den Haag. Lapointe, S.G. & M.H. Feinstein (1981) "The Role of Vowel Deletion and Epenthesis in the Assignment of Syllable Structure", manuscript, Johns Hopkins University and Hampshire College. Revised version is published in H. v.d. Hulst and N. Smith, The Structure of Phonological Representations, Part II, Foris, Dordrecht 1982. Marantz, A. (1982) "Re Reduplication", Linguistic Inquiry 13.3, pp. 435-483. Mors, C.H. ter (1983) "The Problem of Vowel Length in Klamath: a Solution", in Linguistic Flowers from the Botanic Gardens, ATW., Groningen. Revised version will appear in S. de Mey and E. Reuland, eds., Explorations in the Theory of Grammar, Foris, Dordrecht. Thomas, L. (1974) Klamath Vowel Alternations and the Segmental Cycle, Doctoral Dissertation, University of Massachusetts, Amherst.
Syllabification and Syllable Changing Processes in Yawelmani* Roland Noske University of Amsterdam/ZWO
0. INTRODUCTION.
It has been observed by Kisseberth (1970:293) that in Yawelmani, 'there are a number of processes, which, it might be said, "conspire" to yield phonetic representations which contain no word final clusters and no trilateral clusters.' The processes mentioned by Kisseberth are: 1 # (1)
0-*i/C_C
(2)
C -> 0 I C C + _
C
Rule (2) is ordered before (1) and applies only in forms containing a suffix which selects the 'zero' stem of the verb bases (in Newman's (1944) (the main source on Yawelmani) terminology) (I will discuss this selection process later in detail). Looking back at this today we can see that these rules work to avoid impermissible syllables in Yawelmani. If we look into the corpus provided by Newman (1944), we see that with very few exceptions, the possible syllable types are restricted to CV, CVC and (or CV:). Once we have established these three types, it can be seen that there are even more rules taking part in the conspiracy to restrict syllables to these three types than the ones mentioned by Kisseberth. Consider the following two rules proposed by Kuroda (1967) and Kisseberth (1969) 2 : C V J V J
(3)
Shortening V -> [- long] /
(4)
Elision V ->• 0 /
V
* This paper differs in important respects from the one presented during the n o n linear phonology workshop in August 1983. I would like to thank Norval Smith, Irene Vogel, Diana Archangeli, Stanley Newmanf and Harry van der Hülst for their comments. This work was supported by the Stichting Taalwetemchap, which is funded by the Netherlands Organisation for Pure Scientific Research (ZWO).
336
Roland Noske
Examples of these processes are given in (S) and (6): (5)
/laga:+t/ [lagat] 'spend the night' (passive aorist)
(6)
/laga:+in+hin/ [laginhin] (mediopassive+aorist)
Because the syllable was not recognised as a relevant phonological unit when Kisseberth wrote his article, he was not aware of the complete extent of the conspiracy. Kisseberth asks linguists to look at the relationship he felt between the rules. Since the publication of his article, conspiracies have been noted in a great many languages, but no attempt has been made, to my knowledge, to express the relationship formally. I will here express the relationship in a formal way, by making use of the concepts of syllabification and the CV-tier. But first of all, I will treat a proposal made by Archangeli (1983a,b).
1. ARCHANGELI'S PROPOSAL.
In Yawelmani, the regular verb, which Newman refers to as the basic verb, includes two or three consonants. There are three possible consonantvowel configurations, which are given in (7). (7)
Al CVC(C)
A2 CWC(C)
B CVCVV(C)
Al, A2 and B correspond to Newman's classification.3 When combined with the majority of affixes, verbal stems take one of the forms listed above. Examples are given in (8). (8)
stem selected by the base IA1 dub IA2 hiix IB lagaa IIA1 lulcl IIA2 wuu?y IIB biniit
aorist (-hin) 4 dubhun hexhin lagaahin lulculhun woo9uyhun binethin
passive aorist (-t) dubut heexit lagat lulclut wo9yut bineetit
gloss 'lead by hand' 'be fat' 'spend the night' 'bury' 'sleep' 'ask'
In these forms the working of several phonological processes in Yawelmani can be seen. A harmony process has taken place in the forms of IA1, IIA1 and IIA2. This process works directionally from left to right and
Syllabification in Yawelmani
337
rounds unrounded vowels if preceded by rounded vowels with the same specification for the feature [high]. This accounts for the u in dubhun. The process has been formulated by Archangeli as: 5 (9)
Vowel Harmony (VH) [+round] V [ahigh]
""V [ahigh]
Also, we can see the working of the Epenthesis process formulated by Kuroda and Kisseberth as in (1), in the passive aorist forms of the verbs given under IA1, IA2, IIA1, IIA2 and IIB. In the cases of IA1, IIA1, and IIA2, the inserted z's have turned into u by the working o f VH. An i has also epenthesised in the aorist form of IIA2, and has also subsequently been turned into u by VH. Another process to be noted is Lowering, given as formulated by Archangeli (1983a:361) in (10), which lowers long high vowels. (10)
Long Vowel Lowering (LVL) VV->VV
V
[-high] The working o f this process can be seen in the aorist and passive aorist forms of the verbs given as examples of types IA2 and IIA2. In the aorist form of the verb given as example o f type IA2, the underlying ii has been changed to ee by LVL. In the aorist form ee has subsequently been turned to e due to the Shortening process given in (3). In the verb given as example of type IIA2, the long uu has given rise to the harmony process, rounding the high vowels in the following syllables, has then been lowered to oo, and is then shortened to o in the passive aorist case. There are, however, a number of affixes which select a different consonant-vowel configuration for the verb stem that the one which, so to speak, 'belongs' to the verb in question. Examples are given in (11).
(11)
stem selected by the base
desiderativeaorist -(h) atn+ hin
IA1 dub IA2
hiix
reflexive/reciprocal adjunctive -wsiil
selects'. CVC(C)
selects : CVCVV(C)
dubhatinhin hixhatinhin
dubowsol hixewsel
338
Roland Noske IB HAI IIA2 IIB
lagaa lulcl wuu 9 y biniit
laghatinhin lulclatinhìn wu9yatinhin bintatinhin
lagawsel lulcoowsel wu9ooysel bineetiwsel
Archangeli gives the following account of the alternations of the verb stems in Yawelmani. She posits the three rules given in (12): (12)
a. insert CVCC b. insert CVVCC c. insert CVCVVC
The affixes which select a verb stem type of their own carry a diacritic which triggers one of the rules in (12). The other affixes, like the aorist in (8) do not carry such a diacritic. Furthermore, according to Archangeli, the verbs themselves carry a diacritic triggering one of the rules in (12). This diacritic comes into action only if there is no affix with its own diacritic, otherwise the diacritic of the affix takes precedence by means of a new version of the Elsewhere Condition. One can thus speak of a 'default template' supplied by the verb. Consider now the aorist forms in (11). The h is sometimes present, namely in the cases of the biconsonantal verb roots (the IA1, IA2 and IB cases), but the segment is absent in the cases involving a triconsonantal verb root (the IIA1, IIA2 and IIB cases). Archangeli accounts for this in the following manner: she assumes the representation as in (13). (13)
(h)
t n VCC a
apply (12a)
consonantal melodic tier skeleton (CV-tier) vocalic melodic tier
In (13), h is parenthesised to indicate that this element is extra-prosodic, i.e. 'it does not participate in the normal prosodic activity that elements are subject to, here association' (Archangeli (1983a:377)). When a verb melody is inserted the extramelodic h is no longer peripheral but is open for association with an element on the consonantal melodic tier of the affix. See (14a,b) which represent the IA1 and IIA1 desiderative aorist forms in (11). (14)
a. d •
b !
u
h
t n h : • •
a
n !
339
Syllabification in Yawelmani b. 1 ¿ 1 h t n1 h 1n i i i • ! i ii i i i | c vI c c + Iv c c c >v c : i i u a i i
¡ I
1
1
1
In (14b), ft remains unassociated because there is no empty slot for it to be linked to. Archangeli uses the alternations of the type displayed in -(h)atnas one of three arguments for the assumption of an independent CV-tier in Yawelmani, on which the templates supplied by the affix are located. Although I share part of her general conclusion and think that the assumption of the skeleton can account for morphologically conditioned alternations in Yawelmani which hitherto could not really be explained, I will challenge here her argument that alternations such as the h/0 alternation in -hatn- provide motivation for the assumption of the skeleton. First, it should be noted that the h/0 alternation (for which Kisseberth has formulated rule (2)) is analysed by Archangeli as being the result of purely an association process, while as we have seen it takes part in a conspiracy to avoid unpermissible syllable structure. The relationship with syllable structure of other processes taking part in the conspiracy, e.g., Shortening, is recognised by Archangeli (1983a: 361). 6 Second, it should be noted that the analysis given by Archangeli is counter to an essential feature of the theory of lexical phonology developed by Mohanan (1982), Kiparsky (1982, 1983), and Pulleyblank (1983), which she claims to assume (1983a:350). It is the assumption that the verb melody is inserted only after insertion of the affix, and that association has already taken place in the affix before it takes place in the verb stem. According to every proposal in lexical phonology, however, any affix cycle should follow a stem cycle. Therefore, there can be no such structure as the one posited by Archangeli in (13), where there is no verb melody and on which her analysis of the A/0 alternation in -(h)atncrucially depends. Any association takes place from left to right, therefore the first morpheme to be associated is the root morpheme. Bearing this in mind, it can easily be shown that without Archangeli's assumption which goes against the essence of lexical phonology, her system will not work. For this we have to consider two logical possibilities: the first is that insertion of the template by the statement apply (12a) in (13) takes place before association, the second is that it takes place after. In both situations we get the wrong result. Looking at the first possibility, we get a form like in (15):
340
Roland Noske
(15)
1 le 1
h
t n h n
CVCC + VCCCC
In (15), h is no longer peripheral and is therefore open for association. Association will produce (16), which is wrong. (16)
tt ni «h n >1\ 11 1\ i \ CVCC + v ' e f c e e 1
t
(16) will eventually create, after epenthesis, the phonetic form *[lulclahitnih] and the final n remains floating. The result would thus be an absent/ present alternation of the final segment of the affix, rather than of the first. If we assume, on the other hand, that association takes place before application of (12a), we get the the result in (17), creating the absurd outcome *[ulukluh]. (17) v
'
1 ¿1 h t n h n "Ox* >
v
>
S
N
N
\
\
VCCCC
The only way in which Archangeli's analysis could be saved, while maintaining the essence of lexical phonology, would be to assume that the nonalternating segments of the affix are underlyingly linked, i.e. linked in the skeleton, which would render her account extremely inelegant, since it cannot be assumed that the nonalternating segments of the stem are also underlyingly linked to the skeleton, as this skeleton is inserted by rule. Therefore, the key to the explanation of the h/0 alternation in -hatn- cannot be the alleged extraprosodicity of the alternating segment. This becomes even more clear in my third objection against Archangeli's proposal, which is that there are other affixes in which there is also a seemingly morphologically conditioned alternation process between the presence and the absence of a segment which cannot be accounted for in the same way as Archangeli accounts for the -hatn-j-atn- alternation, because the segments in question are not morpheme-initial. These affixes
Syllabification in Yawelmani
341
are: the consequent-gerandial -m(i), precative -x(a), imperative and consequent adjunctive -?(h)iy~. The affix -?(h)iy- does not appear in this form in Yawelmani, but in three other dialects of Yokuts, viz. Wikchamni, Gashowu and Choynimni (Newman, p. 163). It shows exactly the same behaviour as -(h)atn-, in that the h disappears if it finds two consonants right in front of it. The affix selects the 'strong' stem, as Newman terms it (which in Archangeli's system amounts to a CVCVVC template) and surfaces as - ?hiy- when combined with a triconsonantal verb root but as -?iy- when combined with a triconsonantal one. Examples (from Gashowu) are given in (18) 7 (Newman, p. 163-4). (18)
a. cese[ • ]- 9 hiy-a nim 9 o|su 9 '(he) stole my knive', -a objective, lakhy-
!
a Class III:
mussa-"be ashamed" tem. 6):
Class Ilia: pijj-
"pinch"
template 6):
m s > i 1 i [CVCCV] -*• mus?aI I I I I I u a p j i
i
! I [CVCCV] f I
i What does happen is that default rules associate the unmarked vowel lyl with any unlinked skeletal slot V, and the unmarked consonant /?/ with any unlinked skeletal slot C, provided other rules do not apply to associate unlinked slots. One problem that we require to face concerns the question of double segments. If spreading does not take place at all how are these segments derived. With vowels this problem is easier to solve. Only singleton unlinked V-slots, or double unlinked V-slots are associated with /y/ by the default rule. It would seem reasonable then to have a rule operating in the case of unliked V-slots immediately following a linked V-slot. (9)
x K V
V
This rule encapsulates a general principle of morpheme structure in Sierra Miwok that diphthongs are not allowed. In the absence of a category of specific morpheme structure rules (Kiparsky, 1983) this rule must be a lexical rule. During association care must be taken not to associate two successive vocalic segments with two immediately adjacent V-slots. One
368
Norval Smith
manner to achieve this would be to associate one vowel segment in each syllable. Possible syllables in Sierra Miwok are CV, CVV and CVC. This allows us to express the restriction on diphthongs in terms of syllable structure:
(10)
*
a V
V
x
y
More difficult is how we account for the double consonants in derived template 5) in particular. In terms of table 1) we would expect, all other things being equal: •polta? *kelti? *halhy? *tap?u? *kat?y?
instead instead instead instead instead etc.
of of of of of
pollat kellit hallyh tappu? katty?
There is also no sign of the operation of a default consonant rule in the basic templates of classes III and Ilia where the second consonant appears doubled. Numerous nominal stems also display doubled consonants. Two factors might appear to play a role here. Firstly, no derived template, either those illustrated in 1) or any other, displays both doubled consonants and doubled vowels. Secondly, default consonants (or vowels) only appear in derived structures. Now, clearly, spreading rules, or at least those operating to fill up previously unlinked slots, must operate prior to all default rules. This is not in question. How do we account for the facts here? Derived template pattern 5) must involve a special rule of some kind. The rule employed in McCarthy (1979) to resolve a similar problem in Arabic will not provide a solution here as biconsonantal stems would form an insoluble problem. (12)
k
t
b
I IA Va
cvccvc McCarthy has a rule - morphologically restricted in its operation - that
Nonconcatenative Morphology
369
dissociates the /b/ from the third C-slot. By the (then) automatic spreading convention /t/ was spread to this slot giving /kattab/. If we adopted this procedure it would not be possible then to prevent spreading of the second consonant to the final C-slot, i.e./*kattyt instead of /katty?/. The fact that no template displays doubled consonants and vowels is probably just a consequence of the possible combinations of - usually two - syllables in a template. Several of the possible combinations of double consonants and vowels are ruled out because they would not give syllabifiable combinations, e.g # C C W . . , ..WCC... We conclude that it is purely chance combined with the fact that the number of possibilities is very limited anyway that is responsible for the fact that there is no template displaying double consonants and double vowels. How are we to explain double consonants in the various types of stem? Nonderived nominal stems display such a variety of structures that we have in fact little reason to think double consonants are derived by rule in these cases. On the other hand, in verbal stems of classes III and Ilia, which differ only in their vowel structure, it is the second and last consonant that appears doubled. We suggest for the moment that this occurs by rule (13):
C V C C (V) [stem 1] Derived stems display, at least at first sight, a number of problems. A case of spreading that does not only affect adjacent consonants concerns the formation of intransitives. This category is marked by the lack of any suffix, and a template not hitherto discussed - CVCCVCC. (14)
cyymkotaah-
"ride" "move over
cymmymmkottahh-
'ride around' 'move over"
Note that this case could also be analysed in terms of a template CVCCVC and a suffix consisting of an unassociated C. We have here a case that has clearly certain parallels with the case of derived template type 5). In the case of triconsonantal intransitive stems we have repetition of the second consonant which also occurs with all stem types in the case of derived template type 5).
(15)
Basic template
Derived template 5
naathasuul-
natty? hassul-
"reach" "ask"
370
Norval Smith
With derived template 5) we have where necessary the operation of the default rules if the lexical entry for a particular verb does not contain more than one vowel, or more than two consonants. The last two segments in the first examples are the result of this. In this aspect then these forms differ from the intransitives. In other words the two forms both have second-consonant doubling, but only one type - the intransitives - has general spreading. How can we reach this result? Clearly whichever means we employ to achieve second consonant doubling, this must precede general spreading in the case of the intransitives. There are two possibilities. The most obvious would be by a rule of spreading that was restricted to sequences of C's. (16)
x
y
I K c v c c v x template 5 » intransitive ?
S
But this will not be a possible rule, as the normal left-to-right operation of the association convention will previously have associated a third segment z with the second of the two C-slots in the cluster, as in the Arabic case adduced by McCarthy (1979). However it has been demonstrated for tone languages that we need a language particular type of rule operating prior to the universal association conventions. These are the so-called Initial Tone Association Rules (ITAR's) of Clements and Ford (1979). We suggest that what we need in this case is the segmental melody analogue to the ITAR, which would then just be a special case of a more general category of Initial Association Rules (IAR's). The IAR that is necessary in this case would be the following: (17)
Cj
C2 x N c v c c x (template 5 ) (intransitive >
Note that this same rule could be extended to account for stem l's of verb classes III and Ilia. In this case a simple addition to the morphological categories involved in this rule would be sufficient. Now we will turn to an interesting pair of suffixes involving apparent suffixing reduplication in the template [CVCCVC]. The simpler case involves the suffix /-Ce-/ "qualitative".
Nonconcatenative (18)
Morphology
a. kyw-rje cuj-rjeb. ciile kojjoc. hulaw-rje?okiihd. hitphityp- qe-
371
"get cold" "get sweet" "red pepper" "salt" "forget" "beg for food' "get cold" > "get cold" >
kywkyw-wecujcuj-jecilcil-lekojkoj-je hulwaw-we?okhih-he-
"be cold" "be sweet' "taste peppery hot" "taste salty" "be late" "be pitiful"
hitpyp-pe
"be cold"
There appears not to be any completely unified description possible of these forms. We assume the following templates: (19)
biconsonantal stems: triconsonantal stems:
[CVC+CVC] [CVCCV+C]
Let us illustrate this in terms of the various cases: (20)
a.
y w
k
CVC + CVC
I
k b.
w
o 1
CVC + CVC
J
1
c.
o
1 k \ h
?
I
IA
C V C C V ? i
+ C
k o
d.
I
CVC + CVC
h i
I
372
Norval Smith
Even assuming different templates - in terms of how much of the reduplicated suffix they allow to appear - we still have a problem. From the cases illustrated in the first column, we can see that c) and d) require the suffix to be associated from right to left, while a) and b) are indifferent between right to left, and left to right association. On the other hand it might appear as if the example under b) in the second column required to be associated from left to right. This is however as such only a pseudoproblem, since in fact the elements in the consonantal or vocalic melody tiers are not a priori ordered with respect to the elements in the other segmental melody. The real problem is then to establish which of the segmental melody tiers has primacy of association. This is necessary as association is "phoneme"-driven. Let us say that the consonantal melody tier has primacy of association. The attentive reader will protest at this stage that it doesn't make the slightest difference which of the tiers is associated with the CV-skeleton first as they are associated with different types of slot, and we have just stated that there is a priori no ordering relationship between the two tiers. However there does appear to be one such ordering restriction. The two tiers are only indirectly linked via the CV-skeleton so we have to express the ordering restriction at this level. (21)
|2 * V C c
2
In other words the second element of a vocalic melody is not allowed to be associated with a V on the skeletal tier that precedes a C which is associated with the second element of a consonantal melody. In support of this we may quote another template, which unlike that in question involves spreading - [CVCVCCVCC] - which takes no derivational suffix, and has an iterative meaning. (22)
a. ?eep-
"spread out (fire)"
b. putta-
"bubble up"
cujjajuttac. hywaathyleet-
"pour" "shake" "run" "fly, be in the air"
?epeppepp- "spread out (liquid)' pututtutt- "bubble up fast' cujujjujj"fall in a stream' "shiver" jututtutthywattatt- "run around" hylettett"flop about (fish)"
Nonconcatenative
Morphology
petja-
"drop several things"
373 petajjajj"drop several things along the way
Let us illustrate the various cases: (23)
a. ?
p
cvcvccvcc e b. p
t
P
' C V C V C C V C C
c. p
t
j
I i
cvcvccvcc u
a
u
t
a
r^-.
C V C V C C V C C e
a
There are no problems with the derivation of a) or c). With b) however things are different. The result we would expect - in this case /*putattatt-/ does not occur. Observe however that in this case too ¡Cjl would be associated with slots in the skeleton following Iv^• It is again necessary therefore to assume primacy of association of the consonantal tier over the vocalic tier. If elements from the vocalic tier were associated first the result of the ordering restriction (21) would be /*pupappapp-/. It is now possible to explain the formation of /cilcil-le-/ from /ciile-/. In (20b) the reduplicative suffix is associated from right to left, as in cases c) and d). The consonantal tier is associated first - first the /1 / is associated to the last C-slot, then the /c/ is associated to the other available C-slot. Then starting with the rightmost vocalic element, we look for a slot to associate this to. However the only V-slot available is preceding the C-slot to which the second consonantal element is associated. This is disallowed in terms of ordering restriction (21), so we go on to the next vowel which is /i/, and associate this. Note that in the case of /kojjo-/ we have only one vocalic element in the melody, as the obligatory contour principle appears to operate in Sierra Miwok, as we will shortly demonstrate. Finally, we might wonder why in (20d) illustrating /hitpyp-/ we do not get association of the /i/ of the reduplicative suffix to the unfilled V-slot of the stem portion of the template.
374
Norval Smith
(24) h
t
p
C V C C V + c h
t i
As we will see later it is possible for spreading to take place across morpheme boundaries, but not association. Let us however examine first the other template making use of reduplicative suffixation, as this has slightly different effects, which tend to confirm our analysis of the previous case. This involves the suffix / - n y - / associated with the same template [CVCCVC] "multiple location iterative". Most examples of this have the semantic effect "perform some activity here and there". (25)
a. i. jooh?ynn?allii. cyym?uuk?allb.
?ojjajarj - ge-
c.
wy?iithukaa-jwynny-cc-
lili"
joh?u?-ny-
"kill h t " (ht = here and there) "come" ?yn?y?-ny"come often" "hear" ?al?y?-ny"listen around" "climb" cymcym-ny- "climb around h t " "enter" ?uk?uk-ny"go in and o u t " "hear" ?al?al-ny"listen around, go to hear something" "name" ?oj?a?-ny"call someone several different names" "lie someone down" "lie down restlessly jaq?e?-nyht" 'peek at" wy?tit-ny"peek in several different places" "smell" hukjaj-ny"sniff around" "walk around" "walk around wyncyc-nysightseeing"
Here we find two additional types. Under a) we find in addition to examples with apparent full reduplication, examples with default elements filling up the rest of the template. Under b) we find examples involving two elements on the consonantal tier associated to the first two C-slots. The other two C-slots are filled up by default consonants - /?/. The two elements on the vocalic tier are associated to the two V-slots.
375
Nonconcatenative Morphology
Both these types can be analysed in terms of [CVCCVC] templates, without however any sign of a reduplicative suffix. (26)
a.
a ?
1
l I
C V C C V b. o a
c
C V C C V
c
A reduplicative suffix in the first case would result in /*?al?yl/ in the case of a [CVCCV+C] template, or /?al?al/ (cf. (25)a.ii.) in the case of a [CVC+CVC] template. With example b) we would have respectively /*?oj?aj/ and /*?oj?oj/ (cf. /cilcil/ above). A case of a biconsonantal univocalic stem where we do have definite proof of a [CVCCV+C] template is the following case: (27)
cyt-rje-
"like"
cyt?yt-ny-
"like, taste good"
This is to be analysed as follows: (28)
y c
t I CVCCV
I
+ C
I
c
t y
This form although unparallelled is extremely useful in suggesting that what we have here is not simple consonantal spreading in a [CVCCVC] template, as forms like (25c) /wy?tit-ny-/ might have suggested. What generalizations can be drawn here? We can in fact assume that in each of the three types of template found we have a reduplicative suffix. However since in the [CVCCVC] variant there is no scope for the association of this suffix the effect is null. Let us sum up the distribution of the various types: (29)
a. -Cebiconsonantal stems: triconsonantal stems:
[CVC+CVC] [CVCCV+C]
376
Norval Smith b. - n y biconsonantal stems: triconsonantal stems:
[CVC+CVC] [CVCCVC] ([CVCCV+C]) [CVCCV+C]
There is in fact one example of another type, involving a slightly different template. However, because the second stem consonant is /?/, the precise internal structure of the template cannot be identified. (30)
cee?- "hit with thrown object" cee?y?-ny- "hit someone ht with thrown object"
It is not apparent how the relationship between the number of consonants in a verb stem and the extent of the reduplicative suffix should be formally accounted for, although there is a clear functional explanation. There are numerous indications of the operation of the Obligatory Contour Principle with Sierra Miwok, and we will restrict ourselves to discussing one example from (25) that would otherwise be left begging an explanation. The form in question is /jaq-rje-/ "he someone down", whose "multiple location interative" is not /*jaijgeg-ny-/ as might be expected, but /jar)?e?-ny-/. The reason is simply that the two separate cases of /rj/, of different origin, are adjacent on the consonantal melody tier, and are therefore regarded as a single case of /q/. That they are of different origins is abundantly clear, as /jaq-qe-/ is formed from a combination of /jaar)-/ "lie someone down" and /-rje—/ "medio-passive suffix". (31)
j,
0
I A
C V C - C V a e We will now turn to a a morpheme boundary not. (32) a. ?otii-ko?ojissamahhokana?aacab. ?etaalcyym?etal-nukku-
(/-ge-/ selects in this case a template CVC.)
case that demonstrates that spreading can cross in Sierra Miwok, but that simple association can"two" ?otki-liip"double; twins" ?ojsi-liip"four" "quadruplets" mahko-loop- 'quintuplets" "five" na?ca-laap"ten" "ten at a time" ?etla-lamh"return' "be ready to return' cym?y-lymh- "be ready to climb" "climb" take someone home" ?etal-nukku-lumh "be ready to take someone home"
Nonconcatenative
377
Morphology
Observe the cases in (32a). Three of the four examples have three vowels. However, the template selected by the suffix /-lVVp-/-[CVCCV] has only two slots. All other things being equal we might expect to find the first two vocalic elements associated to the first two vowel-slots, with the third vocalic element appearing in the unspecified V-slot of the suffix. (33)
?
t k
I I I
1
p
|
I c
*C V C C V - C V V
I
Ii
0
V o
This does not happen, however. The vocalic element associated to the final V-slot of the stem template is spread to the unfilled V-slots of the suffix. (34)
?
t
k
I I I
1
p
I
I
C V C C V - C V V C
1
i
o
o
In (32b) there is a potentially very interesting case - /cym?y-lymh-/ which appears to illustrate the spreading of a default vowel. Of course as the vowel in the suffix displays the same quality as the default second vowel of the stem, it could itself also represent such a default vowel. This raises the question of just where such default values are filled in. In terms of Kiparsky (1983) they are not structure changing, and they assign contrastive values of phonological features. Therefore they fall into the category of lexical rules. This means that they apply in each cycle after all sequential morpheme structure rules, by the Elsewhere Condition. What is the relationship between the rule of spreading and the default vowel rule? A form like /?etla-lamh-/ suggests that spreading applies before the default vowel rule. (35)
a.
?
t 1
1
m h
I I I I I I cvccv-cvcc e
a
b.
?
t 1
1
m h
I I I M l cvccv-cvcc e
a
378
Norval Smith SPR.
?
t 1
1
DEF. ?
mh
DEF.
a
1
I I I
I II I II c v c c v - c v c c I e
t 1
mh I I I
CVCCV-CVCC
^
?etla-lamh-
SPR.
e
a
y
*?etla-lymh-
What are the consequences of this ordering for forms involving default vowels such as /cym?y-lymh-/? Spreading can never take place in such forms, the suffix vowel being therefore also the result of the operation of the default rule. Let us examine the derivation of this form. (36)
Lexical Representation c m C V V C y Cycle 1 IAR: not applicable Assoc. Conv. c m I I C V V c I y
Rule (9) c
m
I I C V V c 1/ y
Cycle 2 c
m l
m h
I I I I I c v v c - c v c c 1 / y Template Selection c m
|
m h
C V C C V - C V C C y
379
Nonconcatenative Morphology Assoc. Conv. c
1
m
m h
C V C C V - C V C C y Default Rules c m ?
m h
1
C V C C V - C V C C y c
y m ?
by OCP =»
y 1
m h
C V C C V - C V C C y /cym?y-lymh-/ What we may ask has this discussion of Miwok nonconcatenative morphology to contribute to the theory of autosegmental phonology in particular? The main contribution we feel has been to provide analogies for a number of factors relating to tonal phenomena that have emerged in recent years, from the field of nonconcatenative morphology. The most significant of these concerns spreading. The study of tonal phenomena led Pulleyblank (1983) to conclude that spreading was not the result of a universal association principle, but was governed by language-particular rules. We have seen in Sierra Miwok that there are certain morphological processes that induce spreading, while others - under virtually identical phonological circumstances - do not have this effect. In other words, spreading in terms of segmental melody is also rule-governed. We have also demonstrated that the phenomenon of apparently unexpected consonant-doubling can best be explained in terms of an Initial Association Rule, parallel to the Initial Tone Association Rules of Clements and Ford (1979). We have also illustrated with Sierra Miwok data a number of restrictions on the operation of automatic association. These involved a restriction to one association from the vocalic melody tier per syllable, even if two successive V-slots were available, and a restriction against association across a morpheme boundary. Noske (this volume) suggests on the basis of data from Yawelmani - a related language - that the assumed universality of automatic association does not apply in that language. We do
380
Norval
Smith
not feel that the restrictions we have just stated for Siena Miwok would justify such a claim for this language. What does also appear, however, is an asymmetry of association between the consonantal melody tier and vocalic melody tier, whereby the consonantal tier takes precedence. This article represents by no stretch of the imagination a definitive account of the autosegmentally interesting processes in Sierra Miwok. We do not even touch upon the interesting processes of vowel harmony, or vowel-epenthesis, which play a significant role in this language. We hope however to return to these at a later date.
REFERENCES Broadbent, S.M., 1964. The Southern Sierra Miwok Language (= University of California Publications in Linguistics 38). Berkely & Los Angeles, University of California Press. Clements, G.N. and K.C. Ford, 1979. Kikuyu tone shift and its synchronic consequences. LI 10,197-210. Freeland, L.S., 1951. Language of the Siena Miwok (= Memoir 6 of the International Journal of American Linguistics) Indiana University Publications in Anthropology & Linguistics. Kiparsky, P., 1983. Some consequences of lexical phonology, ms. MIT. McCarthy, J.J., 1979. Formal problems in Semitic phonology and morphology. Doctoral dissertation, Cambridge, Mass., MIT. McCarthy, J.J., 1982. Prosodic Templates, Morphemic Templates, and Morphemic Tiers. In: H. van der Hulst & N.S.H. Smith (eds.), The structure of phonological representations. Dordrecht, Foris Publications. Noske, R.G., 1984. Syllabification and syllable changing processes in Yawelmani. In this volume. Pulleyblank, D.G., 1983. Tone in lexical phonology. Doctoral dissertation, Cambridge, Mass., MIT. Smith, N.S.H. and B. Hermans, 1982. Nonconcatenatieve woordvorming in het Sierra Miwok. GLOT, Jaargang 5, 263-284.
Umlaut in a Dutch Dialect* Camiel Hamans
1. INTRODUCTION
In a recent article, Halle and Vergnaud (1981) discuss harmony processes. They distinguish two types: dominant and directional harmony. Dominant harmony spreads a feature in both directions whereas directional harmony propagates a feature in one direction only, from left to right or from right to left. For the description of dominant harmony they make use of an autosegmental framework, in which an autosegment can be associated with segmental slots in a more or less classical way. For directional processes they propose a metrical theory, in which the feature specification of the designated terminal node is copied by a rule on to the root of the tree and percolates downwards from there to all terminal nodes of the tree. Halle and Vergnaud do admit that these harmony processes are similar in many respects, but in the interests of descriptional simplicity they propose different systems to account for the two types. Intuitively this idea is not very attractive. In spite of the differences between various kinds of harmony and assimilation that occur in various languages, they are all basically of the same nature. In fact all harmony processes can be characterized in the same way: the occurrence of a particular feature determines the value of the same or corresponding features in a restricted domain, for instance a word or a foot. For this reason I would suggest that all kinds of harmony processes be described within one and the same framework.
2. COUNTERARGUMENTS
2.1. Turkish There are not only intuitive counterarguments against Halle's and Vergnaud's proposal. There are factual counterarguments as well. For instance their description of Turkish vowel harmony is incomplete. They do not only not account for the exceptions to the normal rule of progressive
382
Camiel Hamans
assimilation (cf. Clements and Sezer (1982)), but also overlook the fact that there are instances of regressive vowel assimilation. This type of harmony is very interesting, working as it does mainly in recent loanwords. The rule, which is a very productive one, applies to loanwords that become assimilated. Lewis (1978 3 : 16-17) gives the following examples: (la)
imparator (emperor) menecer (manager of a football team) madalya (medal) iitiiv (sterilizer) apolet (epoulet)
< Serbo-Croat imperator < English manager < Italian medaglia < etiiv < French etuve < French epaulette
Regressive vowel harmony also applies to compounds: (lb)
bu gun (this day) o bir (the other)
> bugiin > biigiin (biigiin) > obir ( > obur)
These examples show that vowel harmony in Turkish can work from right to left too, although Halle and Vergnaud consider Turkish vowel harmony as a prototype of directional harmony. From the quoted loanwords we may conclude that it is not only progressive vowel harmony in Turkish that is productive, but regressive harmony too. There are more examples in Turkish which show that Halle's and Vergnaud's claim that Turkish is typical for directional harmony, e.g. left to right assimilation, is not true. Turkish is a CVCV-language. This means that loanwords often become adapted to this structure. In 'impossible' consonant clusters a vowel has been inserted, which subsequently undergoes vowel harmony (cf. Lewis 1978: 9-10). (2a)
Arabic ism > isVm > isim (name) gism > kisVm > kisim (part) satr > satVr > sat ir (line)
Backwards harmony of this type is possible also: (2b)
German Schlepp Groschen English train French classeur
> sVlep > > kVrus > >tVren > > kVlasor >
§ilep (cargo-boat) kurus (piastre) tiren (train) kilasor (file)
Vowel harmony in Turkish is bidirectional, as the examples prove. Halle and Vergnaud have only considered vowel harmony in morphological
Umlaut in a Dutch dialect
383
processes, and because of the fact that prefixation is rare in Turkish, they assumed that Turkish vowel harmony operates only in one direction, from left to right. The few instances of prefixation Lewis gives (1978: 55-56), do not contradict my conclusion: (2c)
apik (open) belli (evident) biitiin (whole) pabuk (quick) pevre (circumference) duz (flat)
apapik (totally open) besbelli (completely evident) biisbiitun (altogether, entirely) parpabiik (very quick) pepepevre/peppevre (all around) diimdiiz (absolutely flat)
2.2. Umlaut Another wellknown example of directional and local vowel harmony is umlaut. Umlaut could be considered therefore as another prototype of Halle's and Vergnaud's directional harmony. Traditionally umlaut has been described as a kind of regressive vowel assimilation operating locally on adjacent syllables. Both elements of this description appear to be incorrect. Boer (1924 2 : 58 & 169) has given examples of umlaut in trisyllabic forms ohg. zahari - mhg. zeheren). In my description of umlaut in Roermond (see below) we will find many, also recent, examples of umlaut operating in non-adjacent syllables, for instance: computer - computerize, hamburger - hamburgerke (-ke is the diminutive suffix). Umlaut is not only typical of Germanic languages. McCormick (1981: 127) shows that Sinhalese possessed a rule of progressive umlaut at one time: (3)
Sinh.
sevel (moss) poho (quarter of a lunar month) dolos (twelve)
cf. Pali sevala cf. Pali uposatha *dolasa
Schrijnen (1924) 2 gives many examples from Greek and Latin, which he interprets as progressive and regressive umlaut. (4)
Greek progressive regressive
2