Features, Segmental Structure and Harmony Processes: Part 2 [Reprint 2019 ed.] 9783110250497, 9783110132939

Citation preview

Features, Segmental Structure and Harmony Processes (Part II)

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

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 7 Harry van der Hülst and Norval Smith (eds.) Advances in Nonlinear Phonology 8 Harry van der Hülst Syllable Structure and Stress in Dutch 9 Hans Bennis Gaps and Dummies 10 Ian G. Roberts The Representation of Implicit and Dethematized Subjects 11 Harry van der Hülst and Norval Smith (eds.) Autosegmental Studies on Pitch Accent 13 D. Jaspers, W. Klooster, Y. Putseys and P. Seuren (eds.) Sentential Complementation and the Lexicon

Features, Segmental Structure and Harmony Processes (Part II) Edited by Harry van der Hülst University of Leiden Norval Smith University of Amsterdam

1988

FORIS PUBLICATIONS Dordrecht - Holland/Providence RI - U.S.A.

Published by: Foris Publications Holland P.O. Box 509 3300 AM Dordrecht, The Netherlands Distributor for the U.S.A. and Canada: Foris Publications USA, Inc. P.O. Box 5904 Providence RI 02903 U.S.A. Distributor for Japan: Toppan Company, Ltd. Shufunotomo Bldg. 1-6, Kanda Surugadai Chiyoda-ku Tokyo 101, Japan

ISBN 90 6765 430 2 (Paper) © 1988 Foris Publications - Dordrecht No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission from the copyright owner. Printed in The Netherlands by ICG Printing, Dordrecht.

Table of Contents

Preface and acknowledgements

vii

Contents of part I

viii

John Anderson & Jacques Durand Vowel harmony and non-specification in Nez Perce

1

Jennifer Cole & Loren Trigo Parasitic harmony

19

Hamida Demirdache Transparent vowels

39

Harry van der Hulst The geometry of vocalic features

77

R. Armin Mester Dependent tier ordering and the OCP

127

Catherine Ringen Underspecification theory and binary features

145

Willebrord Sluyters Vowel harmony, underspecification and rule mechanisms: the dialect of Francavilla-Fontana

161

Robert Vago Vowel harmony in Finnish word games

185

Preface and Acknowledgements

In June 1986 we organized a workshop on Features, Segmental Structure and Harmony Processes which was held at the Netherlands Institute for Advanced Studies (NIAS) in Wassenaar, and was sponsored by the Netherlands Organization for Scientific Research (NWO). It was decided then that the papers presented at the workshop should be published. Some of the papers included in this collection were in fact written by participants in the workshop, but several are by authors who we invited in the course of 1987 to contribute a paper dealing with the original theme. Part II contains 7 articles dealing with vowel structure and harmony processes. Part I (of which we specify the contents overleaf) contains articles on the geometrical organization of non-vocalic features. The article by Den Dikken & Van der Hulst in part I can serve as an introduction to both volumes, although it was not specifically written with the present collection in mind. First of all, we would like to thank Keith Snider for the care he bestowed on the preparation of the manuscripts. Secondly, we are grateful to Aglaia Cornelisse, Marcel den Dikken, John van Lit, René Mulder, Martina Noteboom, Rint Sybesma for their assistance with proofreading. Harry van der Hulst Norval Smith

Contents of Part I

Marcel den Dikken & Harry van der Hulst Segmental hierarchitecture Grzegorz Dogil Natural classes, sonority and syllabicity Haike Jacobs & Leo Wetzels Early French lenition: A formal account of an integrated sound change Glyne Piggott The parameters of nasalization Elizabeth Sagey Degree of closure in complex segments Norval Smith An integrated hierarchical system of place features Keith Snider the representation of tone: A three-dimensional approach Towards

Vowel harmony and non-specification in Nez Perce John Anderson and Jacques Durand University of Edinburgh and University of Salford

1. INTRODUCTION

We are concerned here with some consequences of the theory of notation proposed within dependency phonology (DP), specifically the prediction that particular segments may be totally unspecified or unspecified with respect to an individual gesture, where a gesture is a subset of interactive features. Such possibilities follow from the unary feature (component) notation basic to DP representations. If segments are differentiated with respect to a particular gesture by the presence or absence of, say, the three hypothetical unary features A, B and C, giving, in the absence of more complex internal structure to the gesture, the potential segment types {A}, {B}, {C}, {A,B}, {A,C}, {B,C}, {A,B,C} (where "{ }" enclose gestural representations and " , " indicates simple combination), then an eighth possibility is { }, where all three features are absent. 1 Such a segment is unspecified with respect to that gesture; and a segment characterised uniquely by such gestural representations is totally unspecified. It may be that there are segments whose phonetic characterisation warrants a representation including an empty (or absent) gesture (cf. e.g. Lass 1976: ch.6; and, within DP, Durand 1987). We focus here, however, on the role of non-specification in contrastive representations; the nonspecifications we are concerned with arise as a result of minimising redundancy. The notation also permits a distinction to be made between an empty gesture and an absent gesture (cf. Anderson to appear; Anderson and Durand in press, to appear); again, we focus here on the notion of empty gesture, and ignore this further potential differentiation. The question arises as to what extent the occurrence of unspecified gestures is given by general principle. In the first place: do all languages show such segments? more particularly, do all languages show an unspecified vowel, for example? are there typological restrictions (parametrisation)? Further, to what extent is the selection of the unspecified segmenttype based on non-language-particular considerations (rather than simply descriptive advantages)? Anderson and Ewen (1981) propose unspecified consonant-types for English on the basis of considerations of markedness,

2

John Anderson and Jacques Durand

and this selection is shown to accord with descriptive advantages. Anderson and Durand (in press, to appear), in their discussion of various vowel systems, suggest, on the other hand, that in the cases they consider the vowel whose non-specification has the most striking descriptive advantages is also the one predicted to be unspecified by principles based on the geometry of the system. The systems discussed by Anderson and Durand are all asymmetrical; their principles are therefore not applicable, as they stand, to all vowel systems (given that there are symmetrical systems). But to the extent that their proposals are appropriate, clearly markedness too is not universally relevant to the selection of unspecified segments, in that the "asymmetrical" vowel is not always the least marked. It is yet to be determined how these two sets of principles, if they are appropriate, interact, and to what extent they are sufficient: what other principles need to be invoked? how extensive is residual language-particular selection (exceptions to any principle)? We do not venture into these larger issues here. Rather, we are concerned to develop the principles proposed by Anderson and Durand, summarised in 2, in applying them to a further vowel system, one which contrastively might appear to be symmetrical, that of Nez Perce.

2. SYSTEM GEOMETRY AND DEFAULT

The starting-point of our investigation of Nez Perce lies in the observation that couching phonological systems in the DP notation leads to interesting predictions as to the nature of unspecified segments. Archangeli (1984) suggests that, while the fully specified Yawelmani vowel-system (ignoring length) is as in (1): (1)

Fully specified Yawelmani system l

high low round back

+ -

-

a

+

o

-

+ + +

u

+ -

+ +

a more transparent account of Yawelmani is possible, if it is assumed that only one value is specified for each distinctive feature underlyingly and that some segments and features can be left unspecified as in (2):

Vowel harmony in Nez Perce (2)

3

Underspecified Yawelmani system (short vowels) i high round

a -

o +

u +

This representation captures the asymmetrical nature of / i / , which is the epenthetic vowel of Yawelmani. As a result of (2), rules which are featurechanging and require 'clean-ups' in classical treatments can now be reformulated in terms of feature-value additions. Over and above the postulation of underspecified systems modelled on (2), Archangeli puts forward a complex set of complement and default rules which turn onevalued systems into standard binary systems. In DP, starting from the three primitive articulatory components i (acuteness/palatality), a (compactness/lowness), u (gravity/rounding), the full componential make-up of the Yawelmani system would be as in (3) (with respect to the categorial gesture, all the segments are characterised as {V}, V alone, i.e. vowels; consonants involve combinations of the vocalic feature with the consonantal C): (3)

Fully specified DP representation of Yawelmani {i} / i /

{u} / u / {a,u} / o / {a} / a /

where the comma indicates combination (simple association) of components. It should be clear that the three components i, a, u are not all required to specify the Yawelmani vowels contrastively. In particular, given the four segments /i,a,o,u/, the features a and u will each be involved in the specification of two members of the /a,o,u/ series. The segment / i / is therefore the obvious candidate for the status of unspecified vowel as indicated in (4): (4)

Constrastively specified Yawelmani system {} / i /

{u} / u / {a,u} / o / {a} / a /

Linked to the representation offered in (4), a simple default rule is needed, applying as late as possible in derivations, as in (5):

4 (5)

John Anderson and Jacques Durand Yawelmani default rule {} =» i

The selected default vowel is the same as that suggested by Archangeli (1984), and the same descriptive advantages flow from such a selection. But let us now look more closely at the basis for the selection. Note again that the existence of an unspecified vowel in DP, far from being a theoretical innovation, flows directly from the choice of unary features. Since segments are characterised by the presence vs. absence of components, and since this allows for hierarchies of complexity (e.g. / i / = {i}, / e / = {a,i}, / o / = {a,i,u}), the presence of a segment lacking all features is strongly predicted by the notation (cf. 1 and note 1). Durand & Anderson (in press, to appear) further suggest that the unspecified vowel need not be selected on a one-by-one basis but might follow from geometrical principles which also specify the default realisation. For fourvowel systems such as Yawelmani they put forward principle (6): (6)

System-geometric principle 1 a. b.

System geometry: {X}, {X,a}, {a}, {Y} — Unspecified: {Y} (where X and Y range over i and u) Default rule: { } => Y

In the case of Yawelmani, the instantiation of (6a and b) is given in (7): (7)

a. b.

{u}, {u,a}, {a}, {i} — Unspecified: {i} Default rule: { } => i

The selection of the minimally specified vowel in accordance with the geometry of the system is desirably vulnerable: it is falsified, or at least in deep trouble, if the vowel selected is not the same one as is required by the optimal formulation of phonological rules. Latvian provides an interesting testing-ground for System-geometrical principle 1 since its contrastive inventory, as given in (8), is a kind of mirror-image of Yawelmani: (8)

Fully specified Latvian system {i} / i / {i,a} / e /

{u} / u / {a} / a /

Vowel harmony in Nez Perce

5

This time, principle 1 selects the vowel / u / as the unspecified vowel. Archangeli (1984), on the other hand, suggests that the underspecified vowel of Latvian is / a / . Durand & Anderson (in press) give a battery of arguments motivating the choice of / u / in Latvian. To the extent that they are correct, the extension of their approach to other systems seems warranted. Nez Perce, whose system is briefly outlined in Archangeli (1984), allows us to confront the UT and the DP approaches with a challenging set of data.

3. ARCHANGELI'S ANALYSIS OF THE NEZ PERCE VOWELS

Archangeli (1984: 69-70), basing herself on Aoki (1966), suggests that Nez Perce has the following vowel system: (9)

i high low back round

ae

a

+

-

+

-

+

+

-

-

+

o

+ +

u

+ + +

From the point of view of the Feature Minimisation Principle of Underspecification Theory (UT) matrix (9) is massively overspecified. Feature Minimisation Principle: A grammar is most highly valued when underlying representations include the minimal number of features necessary to make different the phonemes of the language. Within UT, as mentioned above, some features may be omitted altogether (i.e. both values) from underlying representations, and no feature will be assigned more than one value. Thus (some or all) systems will contain (for example) a vowel which lacks any contrastive specification (other than that it is a vowel), a default vowel. As the vowels in Nez Perce divide into two harmonic classes which both contain A / , Archangeli further suggests in relation to that language that / i / , being neutral, might be the default vowel. It will thus be unspecified as to the features characterising the vowel space, and [+high], [-back], [-low] and [-round] will be absent from underlying representations. Only one value of each of [high], [low] and [back] is required to keep apart

6

John Anderson and Jacques Durand

the vowels in (9). The suggested underlying representation of vowels is as follows: (10)

i high low back

ae

a

o

u

+

+ +

+

+

and the missing values are assumed to be filled in by the complement rules in (11): (11)

[ ] [ ] [ ] -

[+high] [ - low] [ - back]

which supply these values to segments not marked for the opposite value. In addition [-high] has been removed on the [+low] vowels since it is universally predictable. Similarly, [round] has been left out since the system conforms to universal markedness predictions for the non-low vowels, and roundness does not distinguish /ae/ from / a / . As a result, the system in (10) would be filled in by redundancies as in (12): (12) high low back round

i +

ae

a

o

-

-

-

+ + -

+

+

+

u +

+

+

At this point, Archangeli posits a language-specific, learned rule which specifies the [+low] vowels as [-round]: (13)

[ ] -

[-round] / [

,+low]

completing the specifications of (9). However, it should be noted that this too is in accord with marking conventions (cf. Chomsky & Halle 1968: 405). The unspecified feature values for the Nez Perce vowels are accordingly all allowed for by complement rules, such as (11), which, though languagedependent, are determined by universal principle, and default rules reflecting universal markedness relations. The major descriptive support for the character of the underspecification proposed is the neutrality of / i / with respect to harmony: we return to this below. Firstly, however, we turn to a characterisation of the vowel system in terms of the notation of DP.

Vowel harmony in Nez Perce

7

4. PRELIMINARY DP ANALYSIS OF THE NEZ PERCE VOWELS

The vowel inventory of Nez Perce adopted by Aoki (1966) is interestingly transcribed therein as / i e a o u / , realized as [i ae a o u/ui ] in terms of phonetic norms, with / o / always rounded and / u / exhibiting variations in the direction of an unrounded segment. It seems to us (in the light of the discussion in 1-2) that given the absence of an / e / in the system, there is no need to specify / e / as [ae] at the phonological level. We leave the / e / > [ae] adjustment to a late level of the phonology, along with the optional unrounding of / u / (which, notice, is left out by Aoki, Archangeli and Chomsky & Halle, 1968: 377-378, who yet feel compelled to equate the phonetic value of [ae] with its phonological value). We thus envisage the superficial realisation rules (14): (14)

{a,i} => ; {IU|> => (3)

the first of which introduces a dependency relation between the a and i features in place of simple combination, so that lowness now preponderates over frontness, and the second of which optionally adds the suppressive 3 component to a segment specified solely as grave, i.e. u (as indicated by the verticals; the suppression is manifested as unrounding (cf. Anderson & Ewen 1987: 6.3).2 Since dependency presupposes combination, the first rule cannot precede the introduction of {a,i}-type segments - which, we shall argue, are not present as such initially. While at some level of derivation we thus posit the existence of a fivevowel system given in (15): (15)

/ i / {i} / e / {i,a}

/ u / {u} / o / {u,a} / a / {a}

we are going to suggest that the underlying treatment of Nez Perce is somewhat different from this. In fact, the analysis that we put forward articulates the Firthian option which is outlined by Aoki side-by-side with a standard generative analysis formulated in terms of Jakobsonian distinctive features (cf. 5 and the reanalysis in Chomsky and Halle 1968: 377-378). Underlyingly, we will assume that there are two specified vowels / i / and / u / (i.e. {i} vs. {u}), an unspecified vowel (i.e. a segment which is categorially {V} but unspecified as to articulatory components) and a prosodic component {a} whose effect we shall detail once illustrative examples have been considered. The unspecified vowel {V}, which cor-

8

John Anderson and Jacques Durand

responds to phonemic / e / , will inherit its articulatory gesture {i,a} by a universally determined default rule. Our analysis rests on the observation, fleshed out in 5, that with extraction of the a-prosody, Nez Perce shows the system of contrasts in (16): (16)

i u e = {i,a}

with / a / and / o / being allowed for by association of / e / and / u / respectively with the a-prosody (see below). Here, clearly, / e / is the isolated vowel, in uniquely involving combination and with a component that does not by itself characterise a (non-prosodic) contrastive element. We tentatively suggest System-geometric principle 2 as appropriate here and elsewhere: (17)

System-geometric principle 2 a. System geometry: {X}, {X,a}, {Y} — Unspecified: {X,a} b. Default rule: { } => X,a

Once again, Principle 2 is appropriate to the extent that it selects vowels whose non-specification optimises the formulation of phonological regularities within the relevant system. Let us now take a closer look at the relevant data from Nez Perce.

5. NEZ PERCE VOWEL HARMONY

The representative examples given by Aoki (1966: 759-760) are as follows (except for the substitution of a geminate vowel for the length mark): {tóot} (i) (ii)

'father' (noun stem) na?toot/ 'my father' /toota?/ 'father!'

{méq} (iii) (iv)

'paternal uncle' (noun stem) /ne?meq/ 'my paternal uncle' /meqe?/ 'paternal uncle!'

{tisqe?} (v) (vi)

'skunk' /tisqe?/ 'skunk' /tisqa?laykin/ 'near a skunk'

{céeqet} (vii)

'raspberry' /ceeqet/ 'raspberry'

Vowel harmony in Nez Perce (viii)

/caqaat'ayn/ 'for a raspberry'

{?aat} (ix) (x) (xi)

'go out' (verb stem) /?aatsa/ '(I) am going out' /?aatsana/ '(I) went out long ago' /?aatsaqa/ '(I) went out recently'

{wéeyik} (xii) (xiii) (xiv) (xv) (xvi) (xvii) (xviii) (xix) (xx)

'go across' (verb stem) /wéeyikse/ '(I) am going across' /wéeyiksene/ '(I) went across long ago' /wàayiksaqa/ '(I) went across recently' /weyewéeyikse/ '(I) am hurrying across' /weyewéeyiksene/ '(I) hurried across long ago' /wayawàayiksaqa/ '(I) hurried across recently' /watwàayiksa/ '(I) am wading across' /watwàayiksana/ '(I) waded across long ago' /watwàayiksaqa/ '(I) waded across recently'

As emerges from the above, a number of forms display alternations between / e / and / a / : /na?/-/ne?/, "my" /weeyik/-/waayik/, "go". In addition, / u / - / o / alternations also occur as shown by verbal sequences such as: (19)

(xxi) (xxii)

/wuulelikepese/ '(I) am riding into bushes' /woolalikapasaqa/ '(I) rode into bushes recently'

The morphological conditioning of the e-a, u-o vowel harmony is, as pointed out by Aoki, quite complex: it can be determined by the stem (e.g. (i), (x)), by the suffix (e.g. (viii)) or by the prefix (e.g. (xviii)). The inclusion of prefixes as harmony-triggerers is debatable, as argued by Hall & Hall (1980: note 2 pp. 227-228); but this issue will not be pursued here. To handle these vowel alternations two classes of morphemes can be set up: a DOMINANT series of morphemes which allow only / i a o / within the domain they govern; a RECESSIVE series which can occur with either /i a o / or / i e u / . As already observed, / i / belongs to both series. Examples of dominant morphemes illustrated above are {toot} 'father', {?aat} 'go out', {wat} 'wade', {laykin} 'near', {'ayn} 'for' and {qa} "recent past". Examples of recessive morphemes are {meq} 'paternal uncle', {tisqe?} 'skunk', {ceeqet} 'raspberry', {weyik} 'hurry', {e?} "vocative" and {ne} "remote past".

10

John Anderson and Jacques Durand

6. THE INTERACTION OF NON-SPECIFICATION AND VOWEL HARMONY

What we want to suggest is that the above facts are easily handled if we make the following assumptions. First of all, we shall assume, as anticipated above, that only three vowels occur underlyingly: {i}, {u} and a {V} unspecified for articulation realised as [e] by default. Secondly, dominant morphemes are marked lexically for an extrasegmental {a} component, or prosody, which is unserialised (see Anderson, Ewen & Staun 1985, Anderson & Durand 1986, Anderson 1987). Thus, the morpheme for 'father' is underlyingly (leaving aside the possibility of further delinearisation advocated in Anderson 1987, to appear): {C;V}

{V} {V} \ I \ I \

I

d

\

\

{C;V}

/ /

I u

The {a} prosody gets associated with any {| V|} segment within the domain that a dominant morpheme controls (including vowels within the dominant morpheme itself). Thus, [o] {u,a} results from association of the {a} prosody with underlying / u / {u}; [a] results from association of the {a} prosody with the unspecified vowel. The {i} vowel is "transparent". The reason we isolated the system of vowel contrasts in (16) as /i, e, u/, that is /{i}, {i,a}, {u}/, was that these occur within recessive morphemes and must be distinguishable independently of dominant morphemes when not controlled by the latter. Once we extract the {a} prosody from dominant morphemes we are left with {i} and {u} as contrastive components within the latter - i.e. a subset of the contrastive components required in recessive morphemes. We therefore get the following mappings: (21)

{u} {u} {i}

{} {}

+ prosodie {a}

± +

{u,a} [o] [u] (=> [9] (cf. 6)) prosodie {a} => [i] prosodie {a} [a] {i,a} [e] => {a;i} [*] (cf. 6)

which involve the effects of harmony, default and low level realisation rules.

Vowel harmony in Nez Perce

11

Thus the verbal sequence /woolalikapasaqa/ '(I) rode into bushes recently' is underlyingly (22) (where only the vowels are considered). The last morpheme {qa}, "recent past" is inherently dominant and its domain extends leftwards to the beginning of the verbal sequence. For the purposes of the presentation, the extra-segmental component can be conceived of as an autosegment but in accordance with the references given earlier, we prefer an alternative based on association with a morphological or prosodic constituent as more restrictive: I

(22)

w {V} {V} 1 {V} 1 {V} k {V} p {V} s {V} [q {V}] vv ' '. < \i 1 u

(with the consonantal specifications abbreviated by the standard phonemic symbol). Here {a} is associated initially with the final (dominant) morpheme. After the {a} prosody has been realised within all the places marked as {V}, we obtain (23), which is the appropriate DP representation of the segments in question: (23)

w {V} {V} 1 {V} \ ; i \ ! 1 u,a a

1 {V} i ! i

k {V} p {V} s {V} q {V} i 1 . I | a a a a

As was noted above, {i} is transparent. Although the {a} prosody could in principle come to be associated with an {i} vowel resulting in {i,a} (i.e. /e/), {i} repulses {a}. This does not seem sufficient, however, within our set of assumptions, to classify / i / as unspecified, as this of itself would not explain why it does not inherit a vowel quality which is spreading: indeed it makes this failure somewhat surprising. Even more interesting is the fact that a morpheme containing only {i} in its vocalic make-up can be marked for the prosody {a} and induce an {a} domain while not inheriting {a} itself phonetically. Consider in this respect the examples in (24): (24)

(i)

{qitti} 'place firmly' (verb stem) /tuleeqittise/ '(I) am putting my foot down firmly' (ii) {cik'il} 'destroy' (verb stem) /tolaack'ilksa/ '(I) am destroying with my foot'

The verb stem meaning 'destroy' is lexically { {cik'il} {a} }, with an {a} prosody which spreads to the other morphemes of this sequence (i.e. underlying /tulVV/ and /sV/, where / V / is the unspecified vowel). However, when /tulVV/ and / s V / are combined with the recessive

12

John Anderson and Jacques Durand

morpheme {qitti}, the latter is not marked lexically for the {a} prosody and the unspecified / V / is ultimately realised as {i,a}. Treating the a-prosody as a property of morphemes, rather than individual segments, is perhaps further supported by the existence of a small class of morphemes which are dominant although they apparently lack recoverable vowels from a synchronic point of view (cf. Rigsby & Silverstein 1969: 55-56). Now the rule which spells out unspecified {|V|} as {i,a} operates at a stage (after spreading of {a}) where the vowel set is as in (25): (25)

{i}

{u} {a,u} {a}

The default rule in (17) does not give a determinate value for X: the default vowel is a combination of either i or u with a. But in Nez Perce, at the point at which the default is implemented, the value u for X is blocked: the system already contains {a,u}; only {a,i} is available. The value for X is thus determined by the geometry of the system at the point at which specification takes place. If this is generally the case, the default (17) can remain in the optimally non-specific form given here. A final characteristic of the principle (6), and of (17), that is worth noting is the special status of a vis-a-vis i and u. This (and e.g. its general greater availability for combination) is consistent with the distinction drawn by Anderson & Ewen (1987: 6.1) and others between i and u as "vowel colours" (the articulatory equivalent of the tone features 1987: 7.5) and a as the articulatory equivalent of V, and thus the unmarked vowel. On the other hand, these principles obviously do not select the unmarked vowel as the default. In this they contrast, as we have noted, with the proposals made by Anderson & Ewen (1981) with respect to non-specification in the English consonant system, which are based on markedness. The trading relation between considerations of type of system geometry and of universal markedness, and between these and other principles (which must be envisaged so long as the domains of the former principles remain indeterminate), is yet to be established, as well as the extent to which there must be recognised exceptions to the principles. All we have tried to establish here is the plausibility of considering the Nez Perce vowels to constitute a further example of an asymmetric system wherein selection of the default vowel is in accord with rather transparent principles based on system geometry.

Vowel harmony in Nez Perce

13

7. ATR OR A-HARMONY?

For reasons of space, we will not attempt to compare the treatment offered here to other well-known studies of Nez Perce (e.g. Kiparsky 1968; Rigsby & Silverstein 1969). There is nevertheless one article that deserves a special mention - namely Hall & Hall (1980). Unlike other studies that we are aware of, Hall & Hall's analysis is based on the feature ATR (Advanced Tongue Root) and claims to account for some of the phonetic variation relegated here and in other treatments to low-phonetic adjustments. Hall & Hall's starting-point is the refusal to phonemicise the [i, ae, a, o, u] system as /i, e, a, o, u / . For them, Nez Perce has an underlying canonical three-vowel system as in (26): (26)

i

u a

with [+/-ATR] providing for a six-vowel inventory split into two harmony sets as in (27):

high back ATR

ii

h

-

-

Ul

u2

+ + + + + + + +

—

-

The ATR harmony rule is as follows: (28)

[+ATR] -

[-ATR] / [-ATR]

where the environment obeys the 'mirror image convention' (Langacker 1969). Thus an example like [na?toot] 'my father' will be derived from /na,?tu 2 u 2 t/ via (28): the dominant morpheme /tu 2 u 2 t/ is [-ATR] and retracts the [+ATR] vowel of / n a / . The [oo] in [na?toot] is the result of a separate lowering process. On the other hand, a form like [nae?maeq] 'my paternal uncle' (earlier represented as [ne?meq]) is /nc^mc^q/ underlyingly, where the two morphemes are recessive, and the manifestation of a [+ATR] /a/ is [ae]. Surface [a] and [ae] result from the filling in by (29) below, of [back], which was left unspecified for the / a / ' s in (27): (29)

[ " - h i g h "I

|_OATRJ

-

[-aback]

14

John Anderson and Jacques Durand

Hall & Hall note that / u / ([+ATR]), while typically pronounced as [u] when long, is often realised as [u] and [ui] when short. The [u] realisation fits in, of course, with a [+ATR] feature-value assignment and Hall & Hall further claim that [ui] 'follow(s) naturally if the tongue root is in advanced position when / u / is articulated' (p. 214). As for the [o] realisation of [-ATR] / u 2 / , the explanation lies, for them, in the fact that the area [in, u, u] is already filled by the [+ATR,+high,+back] vowel, and since there is phonetic space left, the tongue lowers to [o] for the [-ATR,+high,+back] vowel. Finally, they also observe that / i / is typically realised as [i] when short and as [i] when long and account for this by the following allophonic rule: (30)

/-realisation (Hall & Hall's rule (19)) +high -back aATR jSlong

-

[0ATR]

It is not clear to us why this ATR-based analysis should be considered as providing a better account of the phonetic facts than the alternative given above. Hall & Hall do not explain, for instance, why [+ATR] long / u i / is realised as [u:] and not as [u:]. Linked to this problem is the need for a rule of o-adjustment given in (31): (31)

o-adjustment (Hall & Hall's rule (18)) +high +back -ATR

r-high] L-lowJ

While Hall & Hall assert that their set of low-level adjustment rules are 'required by every grammar' (p. 215), we know of no evidence supporting a cross-linguistic rule adjusting u — o in systems using a tense-lax or [ + / -ATR] contrast. In fact, in RP, for example, which has an / u / - / u / (foodfoot) opposition, the realisation of the lax phoneme has been argued to be close to Cardinal Vowel 7 [o] (cf. Gimson 1980: 119), and this may not be untypical. The rule of o-adjustment is therefore an arbitrary lowering; and when compared with the / e / — [ae] posited above, it should be borne in mind that (31) is feature-changing whereas (30) is structure-building (see Anderson & Ewen 1987 on the relationship between ';' and its superordinate',' in DP). Since Hall & Hall recognise the pervasiveness of laxing in Nez Perce,

Vowel harmony in Nez Perce

15

it is unclear why the presence of short [t] and short [u ], for instance, cannot be accounted for by a rule such as (32): (32)

|V| -a

a

(where the association represented by the dot indicates that the articulatory component is linked to one V-gesture only i.e. is not a long vowel). Hall & Hall reject rules of this type because this would 'completely mask the important fact of the neutralisation of tongue-root position in the high front vowel' (p. 216). But the neutralisation of the hypothesised underlying distinction between [+ATR] / i , / and [-ATR] / i 2 / is orthogonal to this problem. As stated above, the [i]-[t] distinction at the phonetic level is correlated with length and never reflects the underlying harmonic distinction between [+ATR] and [-ATR]. The existence of a neutralisation of the ATR-opposition for high front / i / ' s in other systems where ATR seems motivated does not render the proposed merger followed by phonetic split based on length any less arbitrary.

8. CONCLUDING REMARKS

This article has attempted to show the relevance of constructing phonological representations in such a way as to minimise redundancies. The adoption of unary features coupled with (tentative) universal geometrical principles, and a prosodic view of the domain of certain processes, yields analyses which throw new light on old data. In particular, it should be noticed that the typological distinction between feature-switching and feature-specifying harmony processes drawn by Hall & Hall (1980) on the basis of Nez Perce does not receive support from our standpoint. Whatever the differences in the morphological conditioning of harmony processes in Uralic and Altaic systems as opposed to Nez Perce, the fundamental chasm they posit between e.g. Turkish and Nez Perce does not seem strongly motivated. They point out that in Turkish, to select, for instance, one of the plural allomorphs {-ler or -lar) as the underlying form would be arbitrary. We should simply represent the plural as / l V r / where V is specified as [+low] but unspecified for [+/-back]. By contrast, in Nez Perce, recessive vowels have to be specified as [+ATR] and change to [-ATR] under the influence of dominant morphemes. In our approach, it will be recalled, no feature-changes are required. The spread of {a} merely extends the componential make-up of underlying segments. It is also worth noting that Harry van der Hulst (this volume) has independently shown

16

John Anderson and Jacques Durand

that many classical ATR systems can plausibly be reinterpreted as involving {a}-spreading. While ATR (or Retracted Tongue Root) seems well-supported cross-linguistically, it would be interesting if the basic components i, a, u adopted in DP turned out to have in this respect a broader range of application than was initially envisaged.

NOTES 1. This is a simple consequence of a set-theoretical description of the phenomena at hand. Once the set of distinctive components has been established, the set of all and only the (contrastive) subsets of this set is its power set which includes the zero set. 2. In the versions of D P defended in Anderson & Jones (1974, 1977), Anderson & Durand (1986), Anderson & Ewen (1980, 1987), among others, there is no component of lip-rounding (co) as such. In terms of phonetic implementation, and particularly on the articulatory side, we assume it is plausible to envisage a rule having the following effect: (i)

default lip-rounding |V|

u — (0 indicating that {u} in a vowel-segment is partially implemented by lip-rounding. We could then conceive of the suppressive component (s) as an inhibition of rule (i). Of course, the issue as to whether a component of lip-rounding is ultimately needed cannot be claimed to have been settled. Lass (1984) presents a version of D P where a component of lip-rounding is advocated but at the cost of a loss of the direct mirroring of markedness that the notation affords (cf. Anderson & Durand 1986; Anderson & Ewen 1987). We also do not think that examples of languages such as Turkish, usually presented as strong evidence for a [ + / - r o u n d ] contrast among vowels, introduce a particular problem for DP, especially in the light of current work on non-specification in this framework. Assuming that incontrovertible evidence can be found showing the necessity of a roundness component (co), then we could still treat roundness as a marked choice since rule (i), interpreted as a phonological redundancy rule (rather than an implementation rule), would for us be universally available. In that case, the phonetic variation found in Nez Perce would move one step nearer the phonology, but would be explicable in the same vein. 3. Aoki gives [x] instead of [q] in (iii) but notes that these two segments are in complementary distribution and he appears to select / q / as the norm in a number of transcriptions.

REFERENCES Anderson, John M. 1987. The limits of linearity. In Anderson & Durand (eds.). Anderson, John M. To appear. Contrastivity and non-specification in dependency phonology. Anderson, John M. & Jacques Durand. 1986. Dependency Phonology. In D u r a n d (ed.). Anderson, John M. & Jacques Durand (eds.). 1987. Explorations in Dependency Phonology. Dordrecht: Foris.

Vowel harmony in Nez Perce

17

Anderson, John M. & Jacques Durand. In press. Underspecification and dependency phonology. In P.M Bertinetto & M. Loporcaro (eds.) Certamen Phonologicum I, Proceedings of the Cortona Phonology Meeting 1987, Pisa: Scuola Normale Superiore. Anderson, John M. & Jacques Durand. To appear. Unspecified and underspecified segments in dependency phonology: Yawelmani and other dialects of Yokuts. Anderson, John M. & Colin J. Ewen. 1980. Introduction: a sketch of dependency phonology, in J.M. Anderson & C.J. Ewen (eds.). Studies in Dependency Phonology. Ludwigsburg Studies in Language and Linguistics 1. Anderson, John M. & Colin J. Ewen. 1981. The representation of neutralisation in universal phonology. In Dressler et al. (eds.) Phonologica 1980, Innsbruck: Insbrucker Beiträge zur Sprachwissenschaft. Anderson, John M. & Colin J. Ewen. 1987. Principles of dependency phonology. Cambridge: Cambridge University Press. Anderson, John M., Colin J. Ewen & Jergen Staun. 1985. Phonological structure: segmental, suprasegmental and extrasegmental. In Ewen & Anderson (eds.). Phonology Yearbook 2, 203-224. Anderson, John M. & Charles Jones. 1974. Three theses concerning phonological representations, Journal of Linguistics 10, 1-26. Anderson, John Ml & C. Jones. 1977. Phonological structure and the history of English. Amsterdam: North Holland. Aoki, H. 1966. Nez Perce vowel harmony and Proto-Sahaptian vowels, Language 42, 759767. Archangeli, Diana B. 1984. Underspecification in Yawelmani phonology and morphology. Ph.D. dissertation, M.I.T. Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York: Harper & Row. Durand, Jacques (ed.). 1986. Dependency and non-linear phonology. London: Croom Helm. Durand, Jacques. 1987. On the phonological status of glides: the evidence from Malay, in J.M. Anderson and J. Durand (eds.). Gimson, A.C. 1980. An introduction to the pronunciation of English. London: Arnold. Hall, Beatrice L. & R.M.R. Hall. 1980. Nez Perce vowel harmony: an africanist explanation and some theoretical questions. In R.M. Vago (ed.) Issues in vowel harmony. Amsterdam: John Benjamins. Hülst, Harry van der. This volume. The geometry of vocalic features. Kiparsky, Paul. 1968. How abstract is phonology? Indiana University Linguistics Club. Langacker, Ronald W. 1969. Mirror image rules II: lexicon and phonology, Language 45, 844-62. Lass, Roger. 1976. English phonology and phonological theory. Cambridge: Cambridge University Press. Lass, Roger. 1984. Phonology. An introduction to basic concepts. Cambridge: Cambridge University Press. Rigsby, Bruce J. & Micheal Silverstein. 1969. Nez Perce vowels and proto-Sahaptian vowel harmony, Language 45,45-59.

Parasitic Harmony Jennifer Cole and Loren Trigo Yale University andM.I.T.

0. INTRODUCTION

The development of non-linear phonology has lent great depth to our current understanding of long distance phonological processes like harmony. In particular, non-linear theories have provided a formalism in which we can express harmonic assimilation and the situations in which assimilation is blocked. The blocking behavior of opaque segments in harmony systems is attributed to their specification for the harmonic feature at the time harmony takes place. For example, some consonants are underlyingly palatalized in Turkish, and are minimally distinct from their non-palatalized counterparts by an underlying [-back] specification. This [-back] feature prevents the spread of [+back] from root vowel to suffix vowel, as in the following representation of the word idrák-i.1 (1)

+ M.

back plane: dorsal plane: skeleton:

i

i i i

X

X

X

X

X

X

i

d

r

a

k -

i

We will refer to this kind of blocking as direct blocking, since the presence of a specification for the harmony feature (on a non-undergoer) results in the direct prevention of association on the harmony plane. 2 In this paper, we address the question of whether all instances of blocking in harmony systems are to be analyzed as cases of direct blocking. It is argued that in parasitic harmony systems, blocking can derive not only from a specification of the harmony feature, but also from the formal representation of an identity condition on the trigger and target of a harmony rule. Parasitic harmony is described as a harmony process which is dependent on both the trigger and target being multiply linked to some

20

Jennifer Cole and Loren Trigo

contextual feature. Harmony can be said to operate within the domain of the contextual feature, and is therefore called parasitic. We review cases of blocking in parasitic harmony systems in Menomini and Maasai. In each of these two cases, we argue for an analysis in which the segments that block harmony are not specified for the harmony feature when harmony applies.

1. THE NON-LINEAR ANALYSIS OF HARMONY

Several assumptions underlie the analysis of harmony systems in non-linear phonology; we pause to briefly mention a few that play a role in the analyses presented here. 1.1. Underspecification Following Kiparsky (1985), Archangeli (1984), and Steriade (1987), phonological segments are represented underlyingly with only a subset of the feature values that they could bear in a fully specified surface representation. While there are differences in the theories of underspecification referred to above, they are not crucial to our analysis. We adopt Steriade's theory of underspecification in which two kinds of redundant feature specifications are recognized: 1.

2.

R-values are feature values that are redundant for an entire class of segments. Such redundant features are not represented in underlying representation. For example, the class of sonorants in English is redundantly [+voice], and therefore unspecified underlyingly for [voice], D-values are feature values which serve to distinguish segments belonging to one class, and therefore are not redundant for the entire class in question. But even D-features need only be specified for a single value in underlying representation. For example, in a typical five vowel inventory containing /i,e,u,o,a/, the feature [high] distinguishes the mid vowel / e / from the high vowel / i / , but both vowels need not bear a [high] feature in underlying representation. Rather, only [+high] or [-high] needs to be specified underlyingly, with the unspecified value filled in by a redundancy rule for the class of segments in question (the non-low vowels in this case).

While all R-values are systematically absent in underlying representation, Steriade reserves the possibility that D-values are present underlyingly in specific languages.

Parasitic Harmony

21

1.2. Feature Geometry Several researchers have recently enriched the non-linear representation of phonological segments by suggesting that distinctive features are hierarchically organized. Mohanan (1983), Clements (1985), Halle (1986), Archangeli & Pulleyblank (1986) and Sagey (1986) have all proposed models of feature geometry, arguing that a hierarchical representation groups together those features which function as a group in phonologies of human languages. All of these models have their roots in an articulatory model of phonological organization. The model proposed in Sagey (1986) is given below.3 root

(2)

continuant consonantal supralaryngeal

laryngeal constr. glottis ,

soft palate

spread glottis stiff v.c. slack v.c.

labial nasal

coronal

round ant. dist.

dorsal hi lo back

Feature geometry, together with underspecification affords an explanation for why certain classes of segments are systematically ignored in some harmony systems. Consider the Turkish example once more. In Turkish, as in many vowel harmony systems, the harmonic feature spreads from vowel to vowel, skipping over and ignoring most consonants. The explanation for this derives from the fact that for most consonants [back] is a redundant feature (an R-value) and is therefore absent in underlying representation. The consonants that do not bear a [back] specification lexically are simply not seen by Back Harmony, a process which creates multiple associations on the back plane between the spreading [back] feature and the dorsal node of target vowels. Underspecification allows us not to specify the transparant consonants with a [back] feature, and feature geometry allows the association of [back] to proceed on a plane where other feature specifications (in particular, those of the transparent consonants) cannot interfere. The non-linear analysis of harmony provides a way of characterizing blocking segments in harmony systems: all those segments which are

22

Jennifer Cole and Loren Trigo

specified for the harmonic feature at the stage in the derivation where harmony applies will be able to block harmony. If harmony applies early in the derivation, before the redundancy rules have applied, then only segments which are underlyingly specified for the harmonic feature will block harmony. Harmony can also be ordered after some redundancy rules, in which case a larger class of segments could potentially block harmony. 4 But in either case, blocking occurs by virtue of a specification on the harmony plane, and for any harmony system, we can predict the class of segments which can potentially block, by assessing the distinctions between segments in underlying representation. We turn now to consider two cases of indirect blocking in parasitic harmony.

2. PARASITIC HARMONY

Many harmony systems share the property of allowing spreading of the harmonic feature F only when trigger and target are similarly specified [aG], for some contextual feature G. One well-known example of this type is Yokuts Round Harmony, illustrated in (3), which spreads [+round] from [ahigh] to [ahigh] vowels (Archangeli 1984, Kisseberth 1969, Newman 1941). (3)

Yawelmani: gloss 'tangle' 'know about, recognize' 'take care of an infant' 'procure'

fut.pass. xilnit hudnut gopnit maxnit

pass.aor. xilit hudut gopit maxit

prec.ger. xil?as hud?as gop?os max?as

Round Harmony does not apply when the trigger and target vowels are of dissimilar height, as in the following examples: (4)

mo:xil?as suhwa:hin

'grow old-prec.ger.' 'make by means of supernatural powers'

It is possible to represent this condition as in (5), using the formalism of Autosegmental Phonology: (5)

[+R] r—-x

I

[aH]

x

I

[aH]

Parasitic Harmony

23

Yet, such a representation does not capture the significance of the fact that the contexts specified on both target and trigger must be identical. Equally plausible would be a harmony system which spread the harmonic feature only from [aF] triggers onto [-aF] targets, yet no such cases are known to exist.5 An alternate representation of the identity condition on Yokuts Round Harmony is obtained by allowing harmonic spreading of [+round] only when both target and trigger are linked to a single contextual feature [ahigh], as in (6). We will refer to this analysis as the Linked Structure Analysis. (6)

[+R]

r [«H]

The representation of the harmony rule in (6) is simpler than (5), since it refers to fewer features. It also allows a clear expression of the identity condition. The Linked Structure Analysis makes the prediction that Round Harmony will be blocked whenever a segment which is specified as [-ahigh] at the time harmony applies intervenes between trigger and target. The presence of an intervening [-ahigh] segment will prevent the trigger and target from multiply linking to a single [ahigh] feature without creating crossing association lines. This situation is illustrated in (7), where the configuration in (i) does not meet the rule in (6) because the trigger and target are not linked to a single occurrence of [ahigh]. Obviously (ii) is not well-formed. (7)

(i)

[+R] X I [aH]

(ii) X

X

I [-aH]

I [aH]

[aH]

[-aH]

In fact in Yokuts, we can observe that [-ahigh] segments block harmony, and not [-round] segments. The Linked Structure Analysis of harmony expresses the fact that in many languages harmonic spreading of [F] is dependent on prior association of a contextual feature [G], Very clear examples of this sort are found in some of the Uralo-Altaic languages, where the application of one harmony process is dependent on the prior application of another harmony process (Steriade 1981, and references cited there). For example, in Kirghiz, Primary Round Harmony applies freely in words that have undergone

24

Jennifer Cole and Loren Trigo

harmonic spreading of [-back]. Words with [+back] vowels have not undergone Back Harmony (their [+back] specification is provided by redundancy rule), and consequently, do not exhibit uniform Round Harmony. Rather, words with [+back] vowels are subject to a Secondary Round Harmony rule which spreads [+round] onto vowels of similar height. Primary Round Harmony is thus seen to be parasitic on prior application of Back Harmony, which creates the multiply-linked contextual structure. The feature [+round] spreads across all segments linked to the same [-back] feature.

3. MENOMINI HEIGHT HARMONY

Menomini vowel harmony is another example of parasitic harmony, whose peculiar characteristics receive a straightforward explanation using the Linked Structure Analysis.6 Menomini has a system of regressive height harmony which raises long / e , o / when followed by one of the high vowels /i,u/, in the same word. 7 For example, the long / e , o / in the roots in (8i-iii) raise to / I , / u / , respectively, when a suffix is added which contains / i / or /u/: 8 (8)

i.

/konl-/

ii.

/ateqnohk-/

iii.

/nemU-/

ku-niak 'lumps of snow' (MG p.96) (cf., kon 'snow' (MG p.96)) ateqnuhkuwew 'he tells him a sacred story' (MG p.96) (cf., atqnohkew 'he tells a sacred story' (MG p.96)) nlmit 'when he dances' (MG p.96) (cf., nemow 'he dances' (MG p.96))

There are six underlying vowels in Menomini: /i,e,e,u,o,a/. 9 All six vowels have both short and long variants. In addition to these, there are two diphthongs: / i a / , / u a / . The occurrence of the vowel / « / between the target and trigger of harmony will block harmony from applying, as seen in (8iv,v): (8)

iv. v.

kewaskeplw kewetuaq

'he is drunk' (MG p.96) 'when they go home' (MG p.96)

In contrast, the vowel / a / is transparent to harmony. In the form in (8vi), the root / o / undergoes harmony triggered by the suffixal / i / ; the intervening / a / does not block harmony, as / e / does in examples (8iv,v).10

Parasitic Harmony (8)

vi.

25

moskamUmuskamit 'if he emerges' (ML) (cf., moskamow'he comes up from under water' (ML))

Ignoring for a moment the behavior of transparent / a / and opaque / e / , we can say that Height Harmony is the regressive assimilation of a [+high] feature from a syllable nucleus position to a long (branching) syllable nucleus position. This rule can be formally stated as an autosegmental spreading rule, as in (9). (9)

Height Harmony Dorsal Node:

+H 1

Root Node: Skeleton: \ / N I a

I N I a

In this analysis, Height Harmony is a feature-filling rule which applies before the D-rule has supplied the feature [-high] to the mid vowels. We will first address the problem of transparent / a / in Menomini Height Harmony. Since / a / is specified as [+low], assimilating [+high] to / a / would result in the illicit feature combination [+high, +low]. We suggest that there is a universal filter that prohibits the combination of features [+high, +low], and that this filter acts as a constraint on derivations. In particular, this filter will prohibit Height Harmony from assimilating t+high] onto the vowel /a/. 1 1 However, while the filter prevents / a / from undergoing harmony, it does nothing to stop the harmonic [+high] feature from spreading past / a / . In this view, harmony is not constrained to operate locally, from syllable to syllable, in Menomini.12 Harmony can skip over a segment in the event that the segment is not able to undergo harmony. We assume that non-local application is the unmarked case, and that local rules must be specially constrained, though nothing hinges on this interpretation of markedness. The minimal statement of Height Harmony, together with the strategy of interpreting the filter as a constraint on derivations, results in a system where / a / is transparent. We turn now to the problem of opaque / e / . In order to understand why it is that / e / can block [+high] assimilation, we first characterize this vowel in terms of the vowel inventory of Menomini. The vowel / e / is a lax front vowel which shows more surface variants than any other vowel, alternating between x~ t ~ i . The surface realization of this vowel

26

Jennifer Cole and Loren Trigo

depends on very idiosyncratic properties of Menomini syllables and words.13 It appears that the vowel / e / can be minimally distinguished from the other vowels solely on the basis of the feature [-tense], and perhaps [-back]. 14 We suggest that the height variation of the surface allophones of / e / derives from the fact that / e / bears no underlying specifications for the height features [high] and [low]. Thus, in underlying representation, / e / is specified only as [-tense]. But while this analysis offers the simplest explanation of the allophony of / e / , it leaves us with the puzzling question of why a vowel that is specified only as [-tense] blocks the harmonic spread of [+high]. One solution would be to say that / e / directly blocks harmony by being specified as [-high] when harmony takes place, as in (10). (10)

[-H]

e ...

e

I

[+H]

...

i

[-T] Let us assume that [-high] is the unmarked value for / e / (on the basis of its distributional frequency). The [-high] feature on / e / would be a redundant feature, however, and therefore should not be present in underlying representation. This means that in order to maintain that / e / blocks because of a [-high] specification, there must be a special redundancy rule that assigns [-high] to / e / , before [-high] is assigned to the tense mid and low vowels.15 Allowing a language-particular redundancy rule of this sort weakens underspecification theory, and results in a much weaker theory of harmony systems. If Menomini can employ a special redundancy rule to create a blocking segment, then other languages should be able to do the same thing. It should, in principle, be possible to specify almost any segment as opaque for a given harmony system, by the simple creation of a language-particular redundancy rule. Yet, the fact is that the choice of opaque segments for a given harmony system is not arbitrary. There seems to always be some relation between the opaque segment(s) and the triggers and targets of harmony - a relation which is not captured by allowing the unconstrained use of language-particular redundancy rules. The facts about the allomorphy o f / e / can be taken to support an analysis in which that vowel is unspecified for the feature [high] in underlying representation. As observed above, / e / varies between high, mid and low surface variants. The claim that / e / is [-high] for harmony would require a more complicated statement of the rules that derive the surface forms of / e / . Such rules would have to be formulated as feature-changing rules,

Parasitic Harmony

27

instead of feature-filling or default rules. Moreover, there does not appear to be any other aspect of the phonology that requires / e / to be specified for height at any stage in the derivation. The second way in which opaque segments are sometimes explained in autosegmental treatments of vowel harmony is by the introduction of filters which prohibit the harmonic feature from associating to and skipping over opaque segments (Kiparsky 1981).16 Thus, Menomini could invoke the filter in (lli) to prevent [+high] from linking to [-tense], and the filter in (11 ii) to prohibit the harmonic spread of [+high] from skipping over the opaque [-tense] segment. (11)

(0

+H

(ii)

+H *

-T

/ \

X

X

X

I -T

The filter in (11 ii) can be reinterpreted as a locality condition on harmony, preventing harmony from skipping over non-undergoers (non-undergoers being created by filters like (1 li)). However, this interpretation of harmony blocking in Menomini is at odds with the fact that harmony appears to be a very non-local process; not only does harmony skip over all instances of / a / , it also skips over all short mid vowels, since only long mid vowels undergo. In short, it is a fact that harmony is only blocked by one out of four possible non-undergoers (/e,o,a,e/), and that it is not a local process. Claiming that harmony is strictly local, as part of the explanation for the opacity of / e / , would require that all short mid vowels and /a/ actually undergo height harmony. This would in turn require a neutralizing rule of [+high]-fission to re-derive [-high] /e,o,a/ at the surface. The obvious problem here is that not all instances of short / i , u / would be candidates for this [+high]-fission rule. Some short / i , u / would be underlying - not the product of Height Harmony - and as such must not undergo [+high]fission, in order to surface as [+high]. In the case of Menomini, in order to entertain a Kiparskian "filters" analysis of opacity, it would be necessary to allow a filter like (11 ii) in the grammar, without interpreting it as a general locality condition on spreading rules. Then, the filters in (lli,ii) together would account for the opacity of / e / . The short mid vowels and / a / would not block simply because no filters would prevent Height Harmony from skipping over these vowels. However, even leaving aside the plausibility of filters of the sort in (llii), we observe that there is no motivation for positing the filter (lli) in Menomini; [+high, -tense] segments must be derived in the phonology, since they appear in the surface vowel inventory as allophones o f / « / .

28

Jennifer Cole and Loren Trigo

The opacity of / t / in Menomini Height Harmony can be easily explained by employing the Linked Structure Analysis presented in the discussion of Yokuts, above. Our proposal is that Height Harmony is dependent on structures which are already multiply linked to a single [+tense] feature. The Linked Structure Analysis directly relates the opacity of / e / to its [-tense] specification. The rule of Height Harmony is reformulated in (12).17 (12)

Height Harmony

+T

+H

DORSAL: ROOT:

A

X

X

The formulation of Height Harmony in (12) explains the opacity of / e / in the following way. First, it is necessary to assume that there is a process within Menomini that conflates adjacent, identical [tense] specifications into a single, multiply-linked feature, as in (13):18 (13)

[tense] tier:

+ I x

+ + - + I I I I x x x x

—

^

x

+ K

x

l x x

+ I x

Let us assume that this process applies to all [tense] features before harmony applies. If [+high] assimilation only occurs when both trigger and target are linked to the same [+tense] feature, it follows both that / e / will not undergo harmony, and that it will block harmony. Blocking will occur because the [-tense] specification of / t / , when it intervenes between the trigger and a potential target, will prohibit the trigger and target from becoming multiply linked to a single [+tense] specification, as illustrated in (14). Multiply linking the trigger and target to a single [+tense] feature creates a violation of the "no crossing association lines" principle. (14)

[+T] We have seen that the Linked Structure Analysis offers a simple explanation for the puzzling facts of Menomini, without invoking any ad hoc special rules or filters into the theory. Moreover, as is argued in section 2, harmony

Parasitic Harmony

29

rules which employ the Linked Structure context are well-motivated from general observations about the types of contextual information specified in those rules, as observed in known harmony systems.

4. MAASAI [ATR] HARMONY

The vowels in Maasai can be divided into two partially symmetric classes: a [+ATR] class and a [-ATR] class, as in (15). (15)

+ATR i e

-ATR u o

I E

U O A

The vowels in roots and suffixes alternate between [-ATR] and [+ATR] by two rules of ATR Harmony - General Harmony and Diphthong-induced Harmony, described in the next two sections. 4.1. General Harmony General Harmony has the effect of causing all vowels in a word to surface as [+ATR] in the presence of a morpheme that contains an underlyingly [+ATR] vowel. If no vowel in a word has an underlying [+ATR] specification, then all vowels will surface as [-ATR], We can say that [-ATR] is assigned by a default rule in Maasai. In the examples in (16i), a [+ATR] suffix vowel causes root vowels to surface as [+ATR], while in (16ii), a [+ATR] root vowel causes suffix vowels to surface as [+ATR]. (16iii) illustrates the default application of [-ATR] in words with no [+ATR] vowel.19 (16)

i.

1-tOn-ie — i-ton-ie 2-sit-App. A-Irobl-ju — A-irobi-ju 1-cold-Inc. ii. E-dot-U — e-dot-u 3-pull-MT E-nor-IshO — e-nor-isho 3-hunt-intran. iii. E-jln-U 3-enter-MT A-I-sUf-IshO 1-II-wash-intran.

30

Jennifer Cole and Loren Trigo

As illustrated in (16i), and in the forms in (17) below, the low vowel / A / blocks General Harmony. (17)

i.

O-lE-m-AA-nin MS-Rel-Neg- 1-hear (cf. o-le-m-e-nin from O-lE-m-E-nin) ii. E-nUk-Ar-ie-kl — E-nUk-Ar-ie-ki 3-bury-MA-APP-Pass iii. I-gurAn-U —-i-gurAn-U II-play-MT

Since / A / is not minimally distinguished from any other underlying vowel by the feature [ATR], it will receive its [-ATR] value by a R(edundancy)rule, in Steriade's (1987) theory of underspecification, sketched earlier. The R-rule will assign [-ATR] to all unspecified low vowels at the same time. The non-low vowels are minimally distinguished in underlying representation by the feature [ATR]. If we asume that [+ATR] is the lexical value, then [-ATR] will be filled in on all non-low vowels by a Distinctive)-rule in Steriade's theory. Steriade suggests that R-Rules always follow other redundancy rules (D-rules). If this is true, then before [-ATR] is filled in on the low vowels, it will first be filled in on all unspecified non-low vowels.20 In this case, all vowels would be fully specified for [ATR] when harmony applies. If harmony is feature-filling, then all vowels would block harmony - an absurd situation. We are left with three alternatives at this point: 1.

2.

We could deny the proposed ordering of D-rules before R-rules, and specify [-ATR] on the low vowel as the first redundancy rule. This solution would lead to a weaker theory of underspecification, inasmuch as Steriade (1987) has argued for the opposite ordering on the basis of the facts of harmony systems like Finnish Back Harmony. At best, we would have to allow the ordering of D-rules and R-rules to be determined on a language specific basis. We could accept that D-rules apply before R-rules, and allow harmony to be ordered after both types of redundancy rules have applied. In this case, harmony would have to be of the feature-changing variety. While we accept that feature-changing harmonies do exist (as in Poser's (1982) discussion of Chumash and McCarthy's (1984) discussion of Montañés Pasiego), Maasai simply doesn't bear the hallmarks of a feature-changing system. Most notably, there is no evidence that harmony causes a bidirectional shift of [+ATR] — [-ATR] and [-ATR] — [+ATR], In fact, there is no evidence that both values of [ATR]

Parasitic Harmony

3.

31

are active in the phonology: only [+ATR] needs to be specified in the [ATR] Harmony rules (to be discussed below). We could accept the ordering of D-rules before R-Rules, but specify that General Harmony applies before either class of redundancy rule. This ordering would account for the fact that the mid and high vowels undergo General Harmony; these vowels can all be unspecified for [ATR] when General Harmony applies. But / A / would also be unspecified for [ATR], and therefore we would yet have no explanation for the opacity of / A / .

Adopting the third analysis above, we might attempt to account for the opacity of / A / by assigning it the value [-ATR] in the lexicon. We reject this analysis because it requires a totally redundant feature value in underlying representation for the sole purpose of getting the harmony facts right. Moreover, it would be an ad-hoc solution, since it would remain unexplained why it is / A / that blocks, and not some other vowel. It would be equally plausible to stipulate that / E / is underlyingly [-ATR], in which case it would be the blocking vowel. There is further evidence from a rule of Raising against the analysis that assigns a lexical [-ATR] value to / A / . Raising is a non-local rule that has the effect of changing all occurrences of / A / in a suffix to / o / after a [+ATR] root vowel, as in the following example. (18)

kl-tA-bol-A-kl-t-A - ki-tA-bol-o-ki-to 2-Past-open-Ep-Dat-Pl-Past

If / A / were underlyingly [-ATR], then the rule in question would have to be feature-changing: the [+ATR] value of the root vowel displacing the [-ATR] value of an arbitrary number of low suffix vowels. Assuming / A / to be unspecified for [ATR] allows us to formulate Raising as a rule changing [+low] / A / to [-low] / O / , which would be followed by General Harmony, deriving surface / o / . Since General Harmony is well-motivated for other data, exploiting its application here allows us to maintain the simpler rule of Raising which does not manipulate a [-ATR] feature. Another possible explanation for the opacity of / A / would be to rely on filters to prohibit / A / from acquiring a [+ATR] specification, as in (19): (19)

*

[+ATR] x

I

[+low]

32

Jennifer Cole and Loren Trigo

and to further restrict harmony from skipping over any non-undergoers. This type of analysis was considered and rejected above for Menomini (see discussion of the filters in (lli,ii). The problem with the filter analysis is that in certain environments / A / actually does assimilate [+ATR], surfacing as [a]. Tucker describes a local [+ATR] assimilation process in Maasai that spreads the feature [+ATR] from lexically specified [+ATR] glides onto the immediately preceding and following vowels, irrespective of the height of the preceding vowel. Consider the following examples: (20)

i.

A-I-rOwA - A-I-rowa Inf-II-hot ii. A-I-wAn — A-i-wan Inf-II-evade iii. Ol-owAru — ol-owaru MS-beast

We can see from the form in (20i) that this local assimilation rule follows General Harmony, since the derived root vowel / o / does not cause the preceding prefix to surface as [+ATR] / i - / . If we were to adopt the filter analysis, we would have to say that the filter in (19) is somehow deactivated before local [ATR] assimilation takes place. A third alternative to explaining the opacity o f / A / is to say that General Harmony is parasitic on structures linked to the feature [-low]. This accounts for the fact that all high and mid vowels undergo harmony, and that the vowel / A / is the only blocker. Parasitic General Harmony is formulated in (21). A sample derivation is given in (22). (21)

Parasitic General Harmony (bidirectional) +ATR X

X

-low (22)

+A E - nUk - Ar - ie - ki V / I -lo +lo -lo

The analysis of General Harmony parasitic on a linked [-low] configuration is consistent with the analysis of Raising suggested above. Recall that Raising

Parasitic Harmony

33

changes the [+low] feature of / A / to [-low] after [-low, +ATR] vowels (/i,u,e,o/). We can view this as a [-low] assimilation rule if we allow the redundancy rule specifying [-low] to apply before Raising. We also argued that the simplest analysis of Raising was to assume that it does not directly affect the [ATR] specification of / A / , but rather is followed by General Harmony, which will always specify the output of Raising ( / O / ) as [+ATR] / o / . This analysis imposes the following rule ordering: 1. 2. 3.

0 — [-low] (Redundancy Rule) Raising General Harmony

Since [-low] is specified before General Harmony takes place, it is possible to exploit the presence of this feature, as in the analysis of Parasitic General Harmony in (21). While the analysis of General Harmony as parasitic on a linked [-low] configuration does explain the opacity of / A / , we concede that it is not yet very strongly motivated. In particular, the analysis which relies on the filter in (19) and a locality condition on harmony is also plausible if we allow filters to apply only to some levels of derivation. However, a better argument for a parasitic harmony system in Maasai is presented in the following discussion of Diphthong-induced Harmony. 4.2. Diphthong-induced Harmony Another source of [+ATR] vowels in Maasai is the rule of Diphthonginduced Harmony. With the exception of a few lexically specified onset glides, glides do not generally play a role in General Harmony. In contrast, all glides that are the first member of the diphthongs / y A / and / w A / trigger [+ATR] harmony on preceeding high vowels, as in (23). (23)

i.

I-tU-pUnU-t-U-A — i-tu-punu-t-w-A 2-Past-come-Pl-MA-Past ii. kl-tl-blrl-A — ki-ti-biry-A 1 P-Past-come-Pl-M A-Past iii. ImArlrl-A [—] ImAriry-A look up to-Past

Mid vowels do not undergo Diphthong-induced Harmony; rather, they block this harmony process, as shown in (24). (24)

i.

A-I-nOr-U-A - A-I-nOr-w-A 1 -II-look-MT-Past

34 (24)

Jennifer Cole and Loren Trigo ii.

k-I-nOr-U-t-U-A - k-I-nOr-u-t-w-A lp-II-look-Pl-MT-Past iii. En-k-ItOrE-A - In-k-ItOry-A FS-II-command-Nom. iv. Il-nOjlnE-AA - Il-nOjiny-AA MP-hyena-Pl

Diphthong-induced Harmony must follow General Harmony, since otherwise the blocking effects of the mid vowels would never be realized. The reverse ordering would result in the following ill-formed derivation of the word in (24ii). (25)

k-I-nOr-U-t-U-A k-I-nOr-u-t-w-A Diphthong-induced Harmony k-i-nor-u-t-w-A General Harmony

*

We propose that the blocking behavior of mid vowels in Diphthong-induced Harmony is explained by stipulating that Diphthong-induced Harmony is parasitic on structures linked to a [+high] contextual feature, as in (26). (26)

Parasitic Diphthong-induced Harmony: -HAIR X

...

-i-high^'

X

X

+lo

The analysis of Parasitic Diphthong-induced Harmony requires ordering the D-Rule specifying [-high] on the mid vowels before harmony. Then the mid vowels will bear a [-high] specification when harmony applies, and will block the formation of the necessary linked [+high] contextual feature. As illustrated in (27), the presence of a mid vowel in between a glide and a high vowel target prevents the glide and the high vowel from being linked to the same [+high] feature. (27)

+A

A

Il-nO j iny-AA

I

I V

+H -H +H One might suggest that mid vowels block Diphthong-induced Harmony

Parasitic Harmony

35

by virtue of being specified [-ATR] at the time harmony applies. If this were the case, then the derivation of (24iv) could proceed as in (28). (28)

-A +A Il-nOjmy-AA

The problem with this analysis is that it requires a special redundancy rule that will specify the mid vowels as [-ATR] before the high vowels (recall that all vowels surface as [-ATR] in words that have no [+ATR] vowel). In the theory of underspecification adopted here, the D-Rule for [ATR] will specify [-ATR] on mid and high vowels at the same time (since these are the vowels for which the feature [ATR] is minimally distinctive in underlying representation). Therefore, the analysis which relies on a [-ATR] specification on mid vowels to explain their opacity would require weakening the theory by allowing languages to tailor their redundancy rules to their own needs. Needless to say, such a theory is lacking in predictive power. Of course, as we saw in the discussion of Menomini, it would be possible to explain the behavior of the blocking segments by ordering harmony after the redundancy rules specify the blockers for the harmonic feature. In this case it would mean allowing Diphthong-induced Harmony to apply after both high and mid vowels are specified as [-ATR]. In this case, harmony would be a feature-changing rule that transforms [-ATR] high vowels into [+ATR] high vowels. As we saw for Menomini, there is no motivation for calling this harmony a feature-changing process, since only one value arguably spreads. Moreover, under the feature-changing analysis, we would still need some way of explaining why mid vowels do not undergo harmony, and the linked structure analysis in (26) seems the best explanation.

5. CONCLUSION

Menomini and Maasai illustrate that not all blocking phenomena can be explained by referring to specifications of the harmonic feature alone. In these two cases, the class of blocking segments is characterized by referring to some contextual feature: the blocking segments form the complement of the class of segments that trigger and undergo harmony with respect to the contextual feature. In Menomini Height Harmony, the [+tense] vowels comprise the triggers and undergoers, while the [-tense] vowel blocks. In Maasai Diphthong-induced ATR Harmony, the [+high] vowels and certain glides are triggers and undergoers, while the [-high] vowels

36

Jennifer Cole and Loren Trigo

block. What is important is that the unusual blocking phenomena in both of these languages can be explained by referring only to properties of the representations to which harmony applies. By adopting the non-linear theory of phonological representation, together with the theory of underspecification, we can explain blocking in parasitic harmony systems as well as the more familiar type of direct blocking, as seen in Turkish, without introducing powerful filters or diacritic devices into the theory.

NOTES •The research presented here was supported by graduate student fellowships at the Massachusetts Institute of Technology. In addition, Cole's research was supported by graduate student fellowships from the National Science Foundation and the IBM Company. We are also grateful for helpful comments from Donca Steriade, Morris Halle, and G.N. Clements. Parts of this research have been presented at the LSA Annual Meeting in 1986, and at Cornell University. We are grateful for comments from both audiences. 1. Example and analysis from Clements & Sezer (1982). Note that in this particular case, the analysis would not change significantly if we were to adopt [-back] as the spreading value, and specify [+back] by default. In such an analysis idrak-i would not exhibit harmony blocking; however, analogous cases exist in which a non-palatalized velar consonant appears root-finally after front vowel roots, yet conditions back vowel suffixes: Sevk-i. In the analysis that takes [-back] as the lexical (spreading) value, such consonants would have to be represented lexically with the feature [+back]. This lexical value would prohibit the spread of [-back], as in the following derivation: (i)

-B

+B

2. The terms "plane" and "tier" are used somewhat ambiguously in the literature to refer to the level of representation in which a feature associates to segments independent of the association of other features. We adopt the term "plane", and use it throughout this paper. The particular analysis of feature geometry we assume is spelled out in the next section. 3. Sagey emphasizes a distinction between the terminal nodes in hierarchical representation and non-terminal class nodes. Only the former have binary values ( " + " and " - " ) ; class nodes are either present or not - but there is no " - " value of a class feature like coronal in her model. 4. Steriade (1987) argues that some harmony processes apply after redundancy rules have already supplied the targets of harmony with a specification for the harmonic feature. Such harmony processes would be feature-changing. See also McCarthy (1984) and Poser (1982) for a discussion of feature-changing harmony in Pasiego and Chumash, respectively. 5. In fact, there is only one example cited in the literature of a harmony rule which specifies a context on both target and trigger, yet where the contexts are not identical. This is the case of Sanskrit n-Retroflexion (nati), which requires that the targets be [+nasal], while the triggers are [+cont, -ant, -dist]. But even in this case, the trigger and target are both coronal. See Whitney (1889), Schein and Steriade (1986). 6. All Menomini data are obtained from Bloomfleld (1962, 1975). For a more lengthy presentation of the facts relating to harmony, and other vowel alternations, see Cole (1986).

Parasitic Harmony

37

7. The glides /y,w/ in onset and coda positions do not trigger harmony, although they derive from the same underlying segments - / I , U / - as the high vowels /i,u/. These facts can be explained by constraining the triggers to [+high] segments which belong to a syllable nucleus (i.e., high vowels). 8. The abbreviations MG and ML are used to indicate references in Bloomfield (1962) and (1975), respectively. 9. In this paper we are ignoring the interesting problem posed by the complementary distribution of the vowels u,o. In Cole (1986a), it is argued that both vowels derive from underlying U, a [+high, +back] vowel, which is lowered in certain environments to o. Since U-Lowering must precede Height Harmony, the derived segment inventory at the time harmony takes place will include the vowel o. For clarity, we will abstract away from their underlying source, and treat u,o as distinct vowels. 10. In this example, Vowel Harmony is seen to apply after suffixation and subsequent loss of root final / U / . For a discussion of the coalescence rules that precede harmony, see Cole (1986a). 11. For the use of filters in constraining derivations, see Kiparsky (1981), McCarthy (1984), and Archangeli & Pulleyblank (1986). 12. For a different perspective on the locality of association rules, see Archangeli and Pulleyblank (1986). 13. Short e is realized as [i] in the personal prefixes before AC. In all other words, [«] is realized as [ae] before h or q, and as [e] elsewhere. In rapid speech, t, when not preceding h or q, is realized as [i]- Long c ranges over [e],[ae] and even [a], although Bloomfield does not state the environments for these alternations. 14. We do not present a complete analysis of the underlying underspecified vowel inventory here. In fact, it is possible to distinguish / e / solely on the basis of the feature [-tense], even though the low, back vowel / a / is most likely also [-tense]. The [-tense] feature on / a / is predictable, given the vowel inventory, on the basis of its [+low] feature. 15. Note that such a special redundancy rule would qualify neither as a D(istinctive)-rule or Redundancy)-rule in the underspecification theory of Steriade (1987). 16. This type of analysis is also discussed in Pulleyblank (1985) and Van der Hulst & Smith (1986). 17. Although in (12) it appears that the association lines linking [tense] and [high] cross, in true three-dimensional structures both of these features occupy independent planes. Also, we are assuming here that the feature [tense] links to the dorsal node, but this analysis is completely consistent with a theory of representation in which [tense] links to a seperate tongue root articulator node. 18. An obvious candidate here is the OCP. Also, McCarthy (1986) argues that feature merging is a reflex of Plane Conflation. For discussion of his proposal see Cole (1987). 19. The sources for these forms are Levergood (1984) and Tucker & Mpaayei (1955). 20. Even in Archangeli's (1984) theory of underspecification, the [-ATR] value on low vowels would be filled in after or at the same time as the [-ATR] value on non-low vowels.

REFERENCES Archangeli, D. 1984. Underspecification in Yawelmani Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Massachusetts. Archangeli, D. and D. Pulleyblank. 1986. "The Content and Structure of Phonological Representations," ms., University of Arizona, Tucson, and University of Southern California, Los Angeles.

38

Jennifer Cole and Loren Trigo

Bloomfield, L. 1962. The Menomini Language, Yale University Press, New Haven, Connecticut. Bloomfield, L. 1975. The Menomini Lexicon, Milwaukee Public Museum Press, New Haven, Connecticut. Clements, G.N. 1985. "The Geometry of Phonological Features," Phonology Yearbook 2, 223-252. Clements, G.N. and E. Sezer. 1982. "Vowel and Consonant Disharmony in Turkish," in H. van der Hülst and N. Smith, eds. (1982). Cole, J. 1986. "Vowel Harmony and Coalescence in Menomini," ms., MIT, Cambridge, Massachusetts. Cole, J. 1987. Planar Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Massachusets. Halle, M. 1986. "Speech Sounds and their Immanent Structure," ms., MIT, Cambridge, Massachusetts. Hülst, H. van der and N. Smith (eds.). 1982. The Structure of Phonological Representations, Vols. I-II, Foris, Dordrecht. Hülst, H. van der and N. Smith. 1986. "On Neutral Vowels," in K. Bogers, H. van der Hülst and M. Mous (eds.), The Phonological Representation of Suprasegmentals: Studies on African Language Offered to John M. Stewart on his 60th Birthday, Foris, Dordrecht, 233-279. Kiparsky, P. 1981. "Vowel Harmony," unpublished ms., MIT, Cambridge, Massachusetts. Kiparsky, P. 1985. "Some Consequences of Lexical Phonology," Phonology Yearbook 2, 85-138. Kisseberth, G.W. 1969. Theoretical Implications of Yawelmani Phonology, Doctoral dissertation, University of Illinois, Urbana, Illinois. Levergood, B. 1984. "Rule Governed Vowel Harmony and the Strict Cycle," NELS 14. McCarthy, J. 1984 "Theoretical Consequences of Montañés Vowel Harmony," Linguistic Inquiry 15, 291-318. McCarthy, J. 1986. "OCP Effects: Gemination and Antigemination," Linguistic Inquiry 17, 207-263. Mohanan, K.P. 1983. "The Structure of the Melody," ms., MIT, Cambridge, Massachusetts, and National University of Singapore. Newman, S. 1941. The Yokuts Language of California, Viking Fund Publications in Anthropology 2, New York. Poser, W. 1982. "Phonological Representations and Action-at-a-Distance," in H. van der Hülst and N. Smith, eds., Vol. II. Pulleyblank, D. 1985. Tonal and Vocalic Redundancy Rules, Paper presented at the Colloque phonologie pluri-lineaire, Lyon. Lexical Phonology, Reidel, Dordrecht. Sagey, E. 1986. The Representation of Features and Relations in Non-Linear Phonology, Doctoral dissertation, MIT, Cambridge, Massachusetts. Schein, B. and D. Steriade. 1986. "On Geminates," Linguistic Inquiry 17, 691-744. Steriade, D. 1981. "Parameters of Metrical Harmony Rules," ms., MIT, Cambridge, Massachusetts. Steriade, D. 1987. "Redundant Features," Papers from the Parasession on Autosegmental and Metrical Phonology, CLS 23. Tucker, A.N. and J.T.O. Mpaayei. 1955. A Maasai Grammar, Longmans Green and Co. Whitney, W.D. 1889. Sanskrit Grammar, Harvard University Press, Cambridge, Massachusetts.

Transparent Vowels Hamida Demirdache M.I.T.

1. WHAT IS A TRANSPARENT VOWEL?

Consider the following vowel systems with transparent vowels, (1) Harmonic feature

Finnish

Hungarian1

Montañés2

back a ä o ö u ü

back a e = [ae] o ö u ü

high i e u o a

e

é

As we can see in (1), a transparent vowel is an unpaired vowel, asymmetrical in a phonological system: it does not alternate, having no harmonic counterpart. Opaque vowels also lack a harmonic counterpart and thus also fail to alternate. But unlike transparent (or neutral) vowels, they also block spreading of the harmonic feature. The low [-ATR]vowel in Akan, which has a dominant ATR harmony, is such a vowel: (2)

i.

ii.

Harmonic feature

o-be-je-1 o- be-je-i o- bisa-1

Akan ATR u e

0) e

o

3

a 'he came and did it' 'he came and removed it' 'he asked (it)'

Opaque vowels have two distinctive characteristics, that of being ungovernable and that of bounding a harmonic domain (or in terms of Clements

40

Hamida Demirdache

and Sezer (1982), they are non-undergoers and blockers). The occurrence of a transparent vowel on the other hand never bounds a harmonic domain: vowels on either side of the non-alternating vowel harmonize. Neutral vowels in Hungarian or Finnish occur in both back vowel stems and front vowel stems, (3)

Hungarian: Finnish:

radir-nak müvésc-nek pari-ko várttiná

'eraser (dat.)' 'artist (dat.)' 'a pair' 'spinning wheel'

The aim of this article is to explain the transparency of these non-alternating vowels. Our analysis will be based on a very simple and natural hypothesis: if a transparent vowel, though it is invariant, 'lets harmony through', then it must be included in the harmonic domain, governed by its head as is any harmonizing vowel. Contrary to the prevalent assumption that neutral vowels are exempt from harmony (excluded from the set of 'P bearing units'), that they must be skipped over when the association conventions apply and that harmonic domains containing neutral vowels are discontinuous, we assume the null hypothesis: transparent vowels as opposed to opaque vowels, are governed and enclosed in the harmonic domain. A further aim of this article is to understand the relation between the feature composition of transparent vowels ([Iround] in Finnish and Hungarian, [+low] in Montañés) and the nature of the harmonic feature ([back] and [high] respectively). Why are ü and ó but not a, a, u or o optionally transparent in Finnish vowel harmony? Why is low e harmonic, and mid é transparent in Hungarian vowel harmony? Finally, we demonstrate that the setting of a single parameter is responsible for the presence vs. absence of transparent vowels in harmonic systems. Before presenting our analysis, we will review briefly the treatment of transparent and opaque vowels within autosegmental phonology and demonstrate why we reject these analyses.

2. ON THE AUTOSEGMENTAL ANALYSIS OF TRANSPARENT AND OPAQUE VOWELS

In order to account for (dis)harmonic structures - that is, for the behaviour of opaque and transparent vowels in languages where sequences of vowels are subject to harmonic constraints - the autosegmental framework provides the two options given in (4):

Transparent Vowels (4)

i. ii.

41

lexical prelinking of a redundant feature specification filters

The filters in (4ii) are not like syntactic filters, filters on the output of rules3, but negative conditions on rule application that proscribe the linking of an autosegment to a skeletal position if such an association generates impossible feature combinations, that is, is not structure preserving. What is remarkable is that given (4i-ii), it is possible to describe the contradictory behaviour of two types of fundamentally different vowels - opaque and transparent vowels - in terms of either option. If association conventions do not differentiate between floating and linked autosegments, spreading taking place in both cases, (4i) will describe an opaque vowel (Clements 1976, 1981, Clements and Sezer 1982): (5)

Opacity in Akan O - blsa - I

>

o

- bisa

- i

+ATR -ATR +ATR -ATR (Capitals represent unspecified vowels) In (5) the Akan low vowel a is invariably [-ATR] and consequently is bound to a [-ATR] autosegment. Prelinking will block spreading of the floating [+ATR] autosegment, and entail a [-ATR] harmony. There are variants of this analysis. For Halle and Vergnaud (1981) spreading is restricted to free autosegments. That the final vowel in (5) surfaces as [-ATR] is explained not in terms of spreading but of the surfacing of the segmentally specified unmarked feature value ([-ATR]). The latter point is irrelevant. What is of importance is that opacity is accounted for in terms of prelinking. If on the other hand spreading is restricted to floating autosegments, (4i) will describe a neutral vowel (Van der Hülst and Smith 1986): (6)

Transparency in Hungarian i. back vowel root with a neutral vowel rAdir nAk > radirnak -B ii.

'eraser (dat.)'

-B

front (neutral) vowel root viz - nAk > viznek

water (dat.)'

-B hid

'bridge (dat.)'

-B

nAk

>

-B hidnak -B

42

Hamida Demirdache

In (6) the Hungarian non-low unrounded vowels are redundantly [-back] and hence are bound to a [-B] autosegment. Lexical linking secures the absence of spreading and the consequent surfacing of the default value ([+back]). In (6ii) we have given the representation of a neutral vowel root inducing front harmony ('viz') and that of one inducing back harmony ('hid'). Thus, as the analyses in (5)-(6) demonstrate, lexical prelinking can describe either a neutral vowel or a transparent vowel. Let us consider (4ii). If one assumes (7) a language well-formedness condition prohibiting discontinuous association, (7)

*a F

vcvcv then (4ii) entails opacity (Pulleyblank 1985, Van der Hülst and Smith 1986). On the other hand if one rejects (7) and allows discontinuous association (i.e. assumes that the structure in (7) is well-formed), then option (4ii) entails transparency (Kiparsky 1981, Van der Hülst 1985). (8)

Opacity in Akan i. * +ATR V [+low] ii.

O - blsA - I

>

+ATR (9)

o - bisa - I

>

+ATR

o -bisa - i +ATR

Transparency in Hungarian i. • -l-Back V -low -round

[ ii.

rAdlr - nAk +B

>

radlrnak +B

>

radirnak +B

In (8), the filter (8i) blocks association of the autosegmentalized feature to the low vowel, and (7) blocks all further association. Consequently the default value surfaces. In (9), the filter (9i) entails skipping the neutral

Transparent

Vowels

43

vowel when the association of the [+B] autosegment takes place. These non-low unrounded vowels will acquire a value for backness at a later stage of the derivation by means of a redundancy rule4. In similar analyses (Booij 1984), discontinuous association follows not from a filter prohibiting association but from the fact that neutral vowels are lexically specified as [-back] in the segmental core. That either filters or prelinking are formal devices enabling us to generate correctly disharmonic forms, is a fact. We do not question the descriptive adequacy of these options. However, that they do not reach a level of explanatory adequacy follows straightforwardly from the fact that either option describes either vowel - these options are interchangeable. It appears arbitrary to select one option over the other to describe what is the subject of this paper: transparent vowels. We can evaluate the explanatory adequacy of a formalism in terms of its predictive value. That is, in terms not only of the class of possibilities it entails, but more significantly in terms of the class of possibilities it excludes. Thus, we should require of a formalism that it be as restrictive as possible in terms of the class of possible analysis it allows. Explanation can only proceed from a maximally constrained theory, and in the case at hand (disharmony) a maximally constrained analysis. A phonological model should only allow a single unequivocal analysis of respectively transparent and opaque vowels. As (5) and (6) demonstrate, given lexical prelinking the behaviour of a vowel is predictable, but transparency is not the only available outcome. Furthermore if in some model the behaviour of a vowel is predictable, then as a corollary, both the set of non-alternating vowels and the set of invariably (or optionally) transparent vowels should be predictable. In (5)-(9) transparent vowels are identified either by a filter or a stipulation (an associaton line in the lexicon). In other words the feature content of these vowels is not predictable. Why is it that in ATR harmony the low vowel is opaque, or what is the relation between the conjunction [l'r°und] and the spreading of the feature [back]?

3. THEORETICAL PREMISES

The analysis of transparent vowels we will develop is based on the following three fundamental theses of the theory of segmental representations presented in Kaye, Lowenstamm and Vergnaud 1985 (henceforth KL&V). 3.1. Elements

The primes of phonological theory are not features but elements. Elements are fully specified matrices. Thus vowels are either elements in isolation,5

44

Hamida Demirdache - round -BACK +high - low I

+ROUND +back +high - low U

- round +back -HIGH +low A

combinations of elements or merely the absence of elements. The mid vowels e and o for example are compounds involving (respectively) the sets of elements (I, A) and (U, A). The insight that the basic phonological unit is not the feature but a higher-level unit is also found in Dependency Phonology (Anderson and Jones 1974, Anderson, Ewen and Staun 1985), where vowels are formed from the various combinations of the three components: | i| 'frontness',|u| 'roundness', and | a| 'lowness'; and in Particle Phonology (Schane 1984) where the primitives of phonology are the elementary particles palatality, labiality and aperture. This insight could also be related to the recent work within autosegmental phonology on the sub-grouping of primes, either tiers or features. Clements (1985) proposes a model of segmental representations in terms of simultaneous lower-level groupings of features and higher-level groupings of tiers, thus allowing sets of features to behave as a unit with respect to phonological processes such as assimilation. 3.2. Underspecification This theory of phonological representations is a version of what is known as 'underspecification theory'. Underspecification is used in this framework as a means for encoding markedness into segmental representations without resorting to linking conventions. An element is defined as a fully specified feature matrix with one (and only one) marked feature value, that which is underlined in (10). Elements are arrayed on autosegmental tiers identified in terms of the features for which these elements are respectively marked: U reposes on the round line, I on the back line and A on the high line. The absence of an element on a line has an interpretation: an unmarked element, that is, a fully specified matrice with no marked feature 6 , surfaces: (11)

- round +back +high - low V

Transparent Vowels (12)

45

round

u

V --

V

back

V

I -—

high

V u

v 1 1

11

11 v-— i

A a

u —11 v A o

V --

V

I1 I — V _. 11 -- A -— V _. i e

The default vowel given in (11) is in fact a real vowel: its feature matrix is that of i, but it is realized only if all lines are empty. Accordingly, i is the least marked vowel, the 3 basic vowels u, i and a, which are elements themselves are equally marked, and the markedness value of compound vowels is a function of the number of elements involved. Markedness is built into phonological representations. Ranking of vowel inventories in terms of markedness is also determined by the line conflation parameter: line fusion excludes compounds segments unattested in a vowel system (mid vowel series if all lines are fused, front rounded vowel series if the back and round lines are fused). Thus, the geometry of representations is explicitly related to the complexity of vowels in a given system. This system is similar to single-valued feature systems (Dependency Phonology, Goldsmith 1985, Rennison 1983, Van der Hulst and Smith 1985, Ewen and Van der Hulst 1985) where the marked value of a feature is represented by the occurrence of this (single-valued) feature and its unmarked value by the absence of this feature. If for instance the marked value of the feature round is [+round], then in either of these systems the absence of U (be it an element, a dependency component, a particle or merely an abbreviation for the feature specification [+round]) has the interpretation [-round] and is unmarked. Likewise in Archangeli's model of underspecification, though features are binary valued, marked feature values are lexical and unmarked feature values are supplied by either default or complement rules. But the KL&V framework differs from all the above systems in one major respect: elements are fully specified feature matrices and consequently no redundancy, complement rules or linking conventions are required to supply feature values other than round. U is specified for all feature values. The cold vowel is a fully specified matrix with no marked feature. 3.3. Matrix calculus To generate the matrices of compound vowels, KL&V provide a procedure that derives a single matrix from the combination of various elements. This operation is called 'fusion', its symbol is '•':

46

Hamida Demirdache - round +back -HIGH +low A

•

+ROUND +back +high - low U

—

~+ROUND +back -HIGH - low o

The resulting matrix in (13) has all the feature values of U but one: the marked feature value of A. The computation of a matrix consists in commuting one and only one marked feature specification. The element which contributes only its marked feature value is the operator, the element which contributes all other feature values is the kernel. By convention the kernel is underlined. In (14) we give the analysis of a vowel involving the compounding of 3 elements. The underlined element behaves as the kernel, the two others as operators contributing respectively their marked feature. - round -BACK +high - low I

•

- round +back -HIGH +low A

•

+ROUND +back +high - low u

—

+ROUND -BACK -HIGH - low Ö

Given that in the fusion procedure, one element (the kernel) is dominant with respect to the other(s), compounding of the same set of elements can generate different vowels as we can see in (15) where we give the analyses of several complex vowels. (15) (16)

i.

(A*U) = o (U*A) = a (V*A) = a (A*V) = a

(A*I) = e (I* A) = ae

but

ii.

(U*I) = u (I*U) = u (V'l) (I»V)

= i = i

As (16) demonstrates, the unmarked vowel having no marked feature by definition, can only contribute a feature value (i.e. [-low]) if it is compounded with A and if it is a kernel. This system differs from others in that the relation of dominance 7 holding between the elements of a compound is formally defined in terms of an operation: fusion. We will see that this point is crucial to our analysis of transparency. But before examining neutral vowels, let us consider what harmony is and how best to express it. We will do so in terms of another relation of dominance: government.

Transparent Vowels

47

4. HARMONY

Government is the relation holding between a head and its dependents or domain. Consider (17), (17)

I

back round high

— I »

—U I 1

A 2

U— I 3

»

> —U I

I » !

1

»

I — U —

A 2

In (17) we have given the derivation of a harmonic front vowel sequence: I spreads along the back line. In terms of the framework explicated in the preceding section where segments, are defined as primitive units of the form 'X' or compounds of the form 'X*Y', spreading of an element along a line will entail a succession of fusion operations. Spreading along the back line in (17) can be represented by the following set of derivations, where the arch ^ is the symbol for concatenation: (18)

1 2 3 4

I I- (U) I« (U~A) I* (U~A~U)

In the standard autosegmental model, the output of harmony in (17) would be an unbounded left headed tree structure. We can view the derivation in (18) as some kind of grid representation of this unbounded head initial constituent. Just as in (18'), the head is the leftmost element and the parenthesized substring, its domain: (18')

1 2 3 4

x x(x) x (xx) x (xxx)

Government is thus, the relation 'X(Y)' holding between a head X and its dependents, or domain Y. Step 2 in (18) is an instance of paradigmatic government: fusion. Steps 3 and 4 are instances of syntagmatic government: harmony. In an autosegmental representation, the relation of government holding between a head and its dependents, is expressed in terms of a tree structure.

48

Hamida Demirdache

An autosegment binds a sequence of positions, (18")

back

I 1

round

2

U-

high u

3 U-

-A ä

u

In (18"), spreading of I to all vowel positions is an instance of government, where the governor is the element I, and its domain the string (1 2 3). In our framework, the geometrical relation 1(1 2 3) entails the set of fusion operations given in (18). Directional harmony can be stated in terms of the government relation 'X(Y)', where the governor X is some element belonging to the set (I, U, A). The domain Y is, i. a sequence of positions (l...n), and X(Y) = X binds Y ii. the sequence of elements respectively identified by (l...n), and X(Y) = X* Y In unbounded harmony, the domain of the governor is maximal: the governor must spread along its line until all free positions are in its scope. Thus, setting the values of the parameters given in (19) will secure directional and unbounded harmony: (19)

i. ii.

Identity of governor X, where X e (I, U, A) Government to the right/Government to the left

(19i) identifies the element which is licensed to spread. (19ii) secures spreading and identifies the direction of spreading. Assume the governor is I, and government is to the right. Consider (20). (20)

back round

U

high 1 u

U A 2 a

3 u

In (20), the back line is empty. In this system, the absence of an element

Transparent Vowels

49

on a line has an interpretation: the default vowel surfaces. In (20), it will surface all along the back line. It has the unmarked feature value [+back], a back vowel sequence is derived. On a line we have a binary choice: presence/absence of an element. The value of the governor on the harmonic line will be X / ~ X . Before considering what is the statement of harmony in a language with transparent vowels, let us first try and understand what exactly a transparent vowel is, and answer the questions: why are nonlow unrounded vowels transparent in backness harmony? Why is the low vowel a transparent in Montañés height harmony? Are vowels predictibly transparent?

5. DERIVING TRANSPARENT VOWELS

In section 2, we have seen that a transparent vowel, though invariant, does not impede the spreading of the harmonic feature whatever its value ( + / - ) , but 'lets it through'. We have assumed the null hypothesis, these vowels are enclosed in the harmonic domain, governed by its head as are all harmonic vowels. Let us see if such a vowel -invariant but governed - is theoretically derivable given the matrix calculus. We have defined government as the relation 'X(Y)' holding between a head and its domain. In our framework, X(Y) = X • Y. A vowel will be transparent if the output of fusion with the governor whatever its value ( X / ~ X , where X is some element belonging to the set (I, U, A)) is invariant. Recall that in this system, ~ X = V. This default vowel, having no marked features, behaves like the identity element in the fusion operation '(V*Y)' (i.e. (V*Y) = Y). Thus for an element Y (Ye(I, U, A, V)) to be transparent, it must satisfy (21), (21) or

(X* Y) (X*Y)

= =

(~X* Y) ( VY)

The equation given in (21) will only be satisfied if X = Y. Consider for example which vowels will behave transparently in a language where the governor is I, or equivalently which elements will satisfy (21) if X = I. (22)

but

(I-U) ( I . A) (I*V) (1*1)

# # * —

(V'U) (V'A) (V'V) (V'D

50

Hamida Demirdache

These fusion operations demonstrate that, if we consider the set of vowels which are defined as elements themselves (respectively in (22), u, a, i, i), only the vowel i will be transparent in a backness harmony. Let us now turn to vowels which are analysed as combinations of elements. For a compound (Z • Y) (Z, Y e (I, U, A, V)) to be transparent, it must satisfy (23). (23) or or

X • (Z*Y) X • (Z*Y) X • (Z*Y)

= = =

~ X • (Z*Y) V • (Z* Y) (Z-Y)

Again the equation in (23) will only be satisfied if X = Y, or equivalently if X = (... *Y). That is, a compound will be transparent if and only if its kernel bears as its kernel, the harmonic feature. The fusion procedures in (24) where X = I, confirm this prediction. (24) but

I*(A«U) I • (U*A) I • (A»V) I • (A*I)

* # —

(A-U) (U-A) (A*V) (A'D

Thus, the only compound vowel that will behave transparently in a backness harmony is the compound whose kernel is the harmonic element: e. If we admit that all segments are compounds, i.e. expressions involving necessarily a relation between a kernel and an operator, then the segments i, u, a will be compounds involving the fusion of the elements I, U, A (respectively) and of the unmarked element V. The transparent vowels in a backness harmony will be analysed as the following expressions, (25)

(VI) = i (A«D = e

These vowels are the only vowels satisfying the equations given in (2123); they are the only vowels which are analysed as expressions of which the harmonic element is the kernel. We have shown that if we assume the theory of segmental representations presented in section 2, the set of transparent vowels is predictable. A transparent vowel is a vowel whose kernel is the harmonic element. Such is the case in Finnish and Hungarian backness harmony, where the vowels i and e are transparent. What of a language where sequences of vowels must agree with respect to the feature high? In this language the harmonic element is A. The only possible transparent vowels will be those vowels whose kernel is A. Such is the case in Montañés, whose vowel system

Transparent Vowels

51

was given in (1) (see also footnote 2): (26)

[+high] words lubúkus 'young wolves' öipiidus 'hunchbacks' but arína 'flour'

[-high] words kolór 'color' beberé 'to drink (lsg fut)' kalór 'heat'

In Montañés height harmony, the low vowel a is neutral. 8 This vowel is analysed as the expression (V* A). It is the only vowel in Montañés whose kernel is the governor.

6. THE KERNEL/OPERATOR PARAMETER

In section 4 we defined harmony in terms of the relation of government 'X(Y)': if X, Y are elements belonging to the set (I, U, A, V), then X governs Y = (X*Y). This characterization of harmony entails two logical possibilities: (27)

i.

Government by a kernel: X(Y) = (X• Y) When X spreads along the harmonic line, it behaves as a kernel, back I round

- U

U

high 2 3 1 ii e ü All skeletal positions are in the scope of a kernel on the back line ii.

Government by an operator: X(Y) = (X • Y) When X spreads along the harmonic line, it behaves as an operator, back I U high

i 1 ü

\

A 2 ä

i 3 ü

All skeletal positions are in the scope of an operator on the back line.

52

Hamida Demirdache

We must assume that in the successive fusion operations that harmony entails, the governor behaves uniformly as either an operator or a kernel. If not, the output of harmony would be unpredictable. Government in (28) could generate either of the vowel sequences given in (28'). (28)

back round high

(28')

i. ii. iii. iv.

I I U I 1 ii

A 2 a

A 3 a

ii ii ii ii

a a e e

a e e a

If for instance the correct output of harmony were (28'ii), to generate this sequence we would have to stipulate that I behaves as an operator in the compounding of the second vowel, and as a kernel in the compounding of the third vowel. Note that I behaves asymmetrically only when it is compounded with A, but the latter is asymmetrical in all fusion procedures ((A«U)^(U*A), (A«I)^(I«A), (A*V) t^ (V*A) ). Thus if A were to spread along its line, we would have 2n possible outputs to harmony (where n represents the number of skeletal positions). If we look at the vowel analyses listed in (21-24), we see that we have assumed all along in order to derive theoretically neutral vowels, that the governor behaves uniformly as an operator. We have assumed in fact that harmony is government by an operator. Consider now what happens if harmony is government by a kernel, if in the sequence of fusion operations entailed by the spreading of an element, this element behaves as a kernel. Would it be theoretically possible to generate transparent vowels? The answer is no: no vowel will satisfy either of the following equations, where the governor X is a kernel. (Recall that a vowel will be transparent iff the output of fusion with the governor, whatever its value X / ~ X , is invariant.) (29)

i. or a. or

(V-X) (V.X) (Y'X) (Y'X)

(V—X) (V. V) (Y—29 (Y* V)

Transparent Vowels

53

(29)

iii.

Z*(Y*X) = or Z*(Y«X) = where X, Y , Z e ( U , I, A)

(29')

i. ii. iii.

(V-D (A-D A*(U*I)

* #

Z*(Y—X) (Z* Y)

(V'V) (A*V) (A*U)

When the governor behaves as a kernel, all vowels alternate. In (29') where the governor is I, we have (respectively) the following alternations: i ~ i, e~a and d ~ o . Thus government by a kernel entails the alternation of vowels, whereas government by an operator entails the transparency of certain vowels. In terms of this distinction, the following vowel alternations obtain: (30) value of the governor I in the fusion operation:

i

ii

KERNEL

OPERATOR

(u • ! ) — ii (V «U) - u

(I (V

U) = ii U) = u

(V . ¡ ) — i i (V *V)

(I (V

D D

(A • ! ) = e (V • A) — a

(I (V

A) = à A) = a

A*(U • I) — ò (A *U) - o

I* (A (A

U) = ò U) = o

I» (A (A

D D

= i = i

= e = e

The vowel system in (30i) is the vowel system of Turkish: each vowel has a harmonic counterpart. (30ii) is the vowel system of both Finnish and Hungarian: the unpaired vowels (respectively) i, e and i, i, e are transparent. Thus by merely setting the value of parameter (32), we derive these respective vowel systems and predict transparency. (32)

Government by a kernel/Government by an operator

54

Hamida Demirdache

In Montañés height harmony (see (1), (26) and footnote 8), the value of the above parameter is government by an operator. The low vowel a is predictably transparent. (33)

Montañés i. Governor: A ii. Government by an operator Vowel alternations (A*U) = o (A*I) = (V'U) = u (V*I) =

e i

(A*A) (V*A)

= =

a a

Assuming parameter (32) has allowed us to identify the set of neutral (or non-alternating) vowels as opposed to the set of harmonic (or alternating) vowels. As we have demonstrated at length in (21-24), the matrix calculus could not generate the hypothetical vowel system (34) with the unpaired vowels behaving transparently, (34)

Governor u 0 1

I ü ó á

If in a language vowel sequences are subject to harmonic constraints, then the values of parameters (35) must be set. (35)

i. Identity of governor X, where X e (U, I, A) ii. Government by a kernel/Government by an operator iii. Government to the right/Government to the left

When the value of (35ii) is government by a kernel, all vowels are harmonic. When its value is government by an operator, vowels whose kernel is the governor are transparent.

7. THE STRUCTURE OF REPRESENTATIONS'

A lexical representation will consist of, i. A sequence of positions defining linear ordering (precedence): the skeleton. Following KL&V, we will represent the skeleton as a sequence of numbers. Each position is linked to a set (possibly null) of elements. ii. Autosegmental lines on which elements repose. As for the identification of these lines, we will consider two options.

Transparent Vowels

55

7.1. Option 1 Autosegmental lines are labelled in terms of the marked features round, back, high, which respectively identify the elements U, I, A. Each element reposes on its line: U on the round line, I on the back line, A on the high line. The autosegmental representation of Finnish backness harmony is given in (36-36'). In Finnish, front and back vowels do not co-occur in native non-compound words. Neutral vowels violate this restriction on the distribution of vowels: they co-occur freely with either back or front vowels. The vowel system of Finnish was given in (1). (36)

Finnish backness harmony i. Governor: I ii. Government by an operator iii. Government to the right

(36')

i.

Front vowel word pòutà+stà 'table (elative)' back I» » I» ! II ! round U U high

A 1

V 2 ii

p ò ii.

»

t

I» i! V A 3 à

V

round

V

high

A 1 p o

U

V

I

;

j V A+ 4 + sta

Back vowel word pouta+sta 'fine weather (elative)' back V V V U

»

A V A 2 3 + 4 u t a + s t a

iii. Mixed vowel words varttina 'spinning wheel' back — I — I — V — — I » » I » » I — I I I > I i ! high — A — V — A — —A V A — 1 2 3 1 2 3 v a r t t i n a v a r t t i n a

56

Hamida Demirdache iv. palttina 'linen cloth' back V I high A 1 p a ltt v.

I V F I V A 2 3 i n a

Neutral vowel stem + suffix el+o 'life' back I V F I V U I I A A 1 + 2 e l + o

round high

The values of the parameters that secure harmony are given in (36). I will qualify as a governor only if it is an operator. This condition is not met in (36'iv, v): spreading is excluded. The unmarked vowel surfaces along the back line. In front vowel words (36'i, iii), there is an operator on the back line. It spreads to all vowels in the stem, and to any suffixes, fronting them. The representation of harmony is straightforward. The vowel e is transparent in a neutral vowel stem because it does not trigger front harmony in a suffix (36'v). The vowel i is transparent in front and back vowel words because the governor (whatever its value, I/V) spreads along the back line 'through it' 10 : the output of fusion along the syntagmatic axis is invariant, (36")

(I • I) (VI)

=1 =1

7.2. Option 2 We assume that autosegmental lines are identified in terms of the kernel/ operator distinction. We obtain the following lexical representations,

Transparent Vowels (37)

i.

Back vowel word 1 kernel

57 U

I

2 operator — A

3 operator ii.

Front vowel word 1 kernel 2 operator — A

3 operator The autosegmental planes in (37) analyse vowels in terms of the distinction kernel/operators. That is, in terms of the asymetrical relation holding between elements in fusion operations. They identify the kernel and the various operators involved in fusion. We have three autosegmental planes in (37), because a vowels is always a combination of three elements, a kernel and two operators. The vowel o is the combination of the kernel U and the operators A, I. For certain vowels, an operator line (or both) is empty. The unmarked vowel surfaces and these vowels are analysed as expressions involving a relation between a kernel and two operators, though one of these operators (or both) contributes no feature value in the fusion procedure. The vowel e for instance, is the combination of the kernel I, and the operators A, V. In (37), we have in fact two important lines of analysis: line 1 which identifies the respective kernels of the compounds, and line 3 which we will call the harmonic line. In Finnish, an I on an operator line must spread along it to all vowel positions. The harmonic line will be the line on which we represent the spreading/absence of spreading of the harmonic element. In (37ii), the harmonic line contains a governor. It must spread to all free positions.

58

Hamida Demirdache

(37') 1 kernel

U I 2 operator — A — |

I I V—|

I

—A —|

_-3

I o

U

V—|

\J

1 3 operator

A—

i

>

\J

I o

a

V I I V—|

\1

i

A— I V -|

\l

a

(37') is the representation of a front vowel sequence. The vowel i is transparent because given the matrix calculus, (1*1) = (I*V) = i. (37i) is the representation of a back vowel sequence: the harmonic line is empty. The vowel i is transparent because it does not contribute an overt operator on this line. Below we give the autosegmental representation of harmonic words in Finnish.

(38)

i.

Back vowel word tuhma + sta 'naughty (el.)' kernel U A1 operator t

ii.

2

+

V V uhma

V +sta

Front vowel word tuhma + sta 'stupid (el.)' kernel U

t

iihma

iii. Mixed vowel words kernel A I 1 I operator V p a 111

I I 2 I V i

+sta

n

A I 3 I V a

Transparent Vowels iv. kernel

59 —A — I —A — 1

2

3

operator — I v a rtt i n a v.

Neutral vowel stem + suffix kernel I operator — .A \ operator

U

A 1 + I V e 1+

\

2 I vo

If autosegmental lines are labelled in terms of the distinction kernel/ operators, then there is no difference between back vowel words with or without neutral vowels, and between front vowel words with or without neutral vowels. In the former, the harmonic line is empty. In the latter, I spreads along the harmonic line to all vowel positions. We will select this mode of identification of autosegmental lines on the basis of the following assumption: it is because vowels are analysed in terms of the paradigmatic kernel/operator relation that certain vowels are transparent. To explain the transparency of non-low unrounded vowels in backness harmony, we have assumed that these vowels are analysed in terms of this distinction. Thus the lines in (37-38) have no theoretical status other then that of analyzing compounds, in terms of the distinction relevant to explaining their harmonic behaviour. A phonological representation is a set of concomitant analyses (autosegmental planes) mapped onto a skeleton: the stress plane, the syllable structure plane, the vowel and consonant melody planes, and the kernel/operator planes. Solely on the basis of the above considerations have we selected option 2. Given that harmony in Finnish is government by an operator, option 1 is descriptively adequate. We cannot be accused of resorting to some kind of tier duplication to account for transparency: an account in terms of option 1 is available.

8. THE TRANSPARENCY OF V AND O IN FINNISH VOWEL HARMONY

In Finnish, the neutral vowels (i, e) co-occur freely with either back vowels

60

Hamida Demirdache

or front vowels. Back vowels (u, o, a) may not co-occur with front vowels (ii, o, à) in native non-compound words. However, in disharmonie stems (unassimilated loans), all front vowels, excepting à, are optionally (and in certain Finnish dialects necessarily) transparent. (39)

i.

careful speech klorofulli-sta sutenoori-sta but *miljonari-sta miljonari-sta

ii.

casual speech klorofiilli-sta sutenoori-sta miljonari-sta

'chlorophyl (elative)' 'pimp (el.)' 'millionaire (el.)'

The vowels ii and o can be analysed as the following expressions: (40) or

(I *U)= (U*D =

A*(I *U)= or A*(U *I) = but I*(U • A ) # U*(I » A ) *

ü ü

ö ö ö ö

Significantly, these vowels can be analysed as expressions of which the element I is the kernel: the fusion operation (U*I) is commutative. (41)

- round -BACK +high _- low I u.

+ROUND +back +high - low U

•

•

+ROUND +back +high _- low U - round -BACK -high +low I

=

—

+ROUND -BACK +high _- low _ Ü +ROUND -BACK -high +low ü

In the fusion operations which yield u and o, the I component is not necessary, but only optionally selected as the kernel. Given that a transparent vowel is an expression whose kernel is the harmonic element, ii and o are only optionally transparent. Whereas i and e are necessarily transparent: these vowels are necessarily analysed as expressions of which I is the kernel. The vowel i being the I element itself, and the fusion procedure yield e not being commutative.

Transparent Vowels (42)

u.

61

- round +back -HIGH +low A

- round -BACK +high - low I

+round -BACK -HIGH _- low _ e

- round -BACK +high - low I

- round +back -HIGH _+low A

- round -BACK -HIGH +low ä

We submit that in dialects or idiolects where u and 6 are transparent, these vowels are analysed as expressions of which the I component is the kernel, whereas in dialects (or idiolects) where u and o are opaque (trigger front harmony in suffixes), these vowels are analysed as expressions of which the I component is an operator: (43)

Transparent (U*I) = u A*(U*I) = o

Opaque (I*U) = u A#(I*U) = o

The fact that transparent and opaque ii and 6 are analysed as asymmetrical expressions in the lexicon is not ad hoc: the occurrence of a transparent ii/o is not predictable but idiosyncratic and therefore must be accounted for in terms of the lexical analysis of these vowels. Different dialects or idiolects select different analyses of these vowels. What is of importance is that these vowels are analysable in terms of the matrix calculus as either (U*I) or (I*U), whereas a is not analysable as equivalently (I*A) or (A*I), (see 42). The vowel a is predictably always opaque, never transparent. Summarizing, the 3 back vowels (u, o, a) which lack the I component in their representation, and the front vowel a can never behave transparently. (43')

Disharmonie stems with final front vowels i. Transparency of ü kernel U A operator — A -

operator

1 I _v [ o

y — 2 I V a

U —

V3 + I V ü ]+

4 I V — a ]

62

Hamida Demirdache ii.

Opacity of ii kernel

U A —

operator

\

operator

A

U

A

y —-V

V-

\

1 I —V [ o

2 I V a

u ]+

U

A

A

iii. Opacity of a kernel operator — A —

operator

[

[

+

y — 1 I V — o

> a

]

V

2

+

\|

I

] +

a]

The front vowel a must be analysed as an expression in which I is an operator. When this element is on an operator line, it must spread to all vowel positions. J.R. Vergnaud has pointed out an interesting prediction made by our analysis of transparent vowels. Given that, i. ii.

a transparent vowel is a vowel whose kernel is the governor government is by an operator

then, a vowel will be transparent even if its harmonic counterpart exists. This prediction is empirically verified. In Finnish, u/d are transparent even though they have [+back] counterparts. As Anderson (1980: 31-33) remarks, "We do not know of any instances in which this sort of behaviour is instantiated in the fully productive part of a harmony system, but Finnish provides at least a marginal example of exactly this t y p e . . . For a word in which y (=[u]) thus functions as a neutral vowel, the harmony process must assign the feature [ ± back] without regard to its presence: (17)

[+back] /mArttUUri+Us/

Transparent Vowels

63

If left unaltered, however, this representation would yield *[marttuurius], A rule of absolute neutralization can properly change all instances of /»/ in [+back] domains to [i], but a corresponding rule for incorrect instances of / u / in such forms would incorrectly apply to genuine instances of / u / . There is no way to distinguish between representation (17) and that given below for the form partureja 'barber (partitive plural)': (18)

[+back] /pArtUre+j+A]

The rule correcting */marttuurius/ to [marttyyrius] would apparently also change /partureja/ to *[partyreja]. This shows that neutral vowels cannot actually be integral parts of harmonic domains with respect to [±back] in Finnish, but must rather be skipped over, much as consonants are disregarded in assigning harmonic features."

The autosegmental analyses we discussed briefly in section 2, must assume a neutral vowel is a vowel without a harmonic counterpart. There is no difference between an absolute neutralization rule and the fronting rule (44i), or the redundancy rule (44ii) which specifies neutral segments either in the segmental core, or at a later stage of the derivation. (44)

i.

+B

-

-B

-low -round n.

-low -round

— -Back

iii. * +B i __ iiI -low -round These rules and the filter (44iii) which secures discontinuous association, could not be formulated if neutral vowels had a harmonic counterpart. This is not the case with neutral (/', e) in Hungarian or Finnish. But such vowels do exist, neutral ii/6 have a [+back] counterpart. Our analysis differs from previous analyses in that, - a neutral vowel need not be without a harmonic counterpart - a neutral vowel is a governed vowel, harmonic domains with neutral vowels are not discontinuous.

64

Hamida Demirdache

9. THE STATUS OF LOW e IN HUNGARIAN VOWEL HARMONY

In Hungarian, the co-occurrence of back and front vowels in stems is excluded; suffixes alternate depending on the harmonic class of the stem. There are two classes of exceptions to this statement: disharmonic stems (foreign words and compounds) and invariant suffixes. Neutral vowels co-occur freely with back or front vowels. Hungarian has a 14 vowel system (7 short vowels/7 long vowels), given in (45) (using standard orthography, in which the acute accent marks length, the diaeresis fronting, and the double accent length and fronting). Short a is phonetically [+round], and short e [+low]. (45)

Vowel system Harmonic vowels front ii ii o o e (=[«])

Neutral vowels back u o a(=M)

u o a

i

i e

We have classified e, which represents the low front vowel [ae], as a harmonic vowel. However its status is controversial. Certain authors classify this vowel as neutral, others as a harmonic front vowel. If a vowel is transparent in a backness harmony if and only if it can be analysed as an expression whose kernel is I, then low e is necessarily harmonic. For fusion to yield a [+low] vowel in the compounding of A and I, the former element must be the kernel. Mid e and low e in Hungarian (or e and a in Finnish) involve compounding of the same set of elements. They are differentiated only in terms of the asymmetrical kernel/operator relation: (46)

(A*I) = e (I *A) = ae

We have shown in (42i, ii), that these fusion procedures are not commutative. Low e must be analysed as an expression whose kernel is A. It cannot be transparent in backness harmony. The data adducing evidence for the harmonic status of e is given below. (47)

i.

e alternates with the back vowel a: nak ~ nek (suffix, dative)

ii.

Certain monosyllabic stems with a neutral vowel trigger back harmony in suffixes. There is no such stem with e:

Transparent Vowels (47)

hid + nak cél + nak

65 'bridge (dat.)' 'goal (dat.)'

*kert + nak kert + nek

'garden (dat.)'

iii. In a suffix, it always undergoes regular harmony. There are no invariant suffixes with e. All neutral vowel suffixes contain either i, i or e\ Vajda + ne + nak 'wife of Vajda (dat.)' Szoke + ne + nek 'wife of Szoke (dat.)' iv. Disharmonie stems with a final e trigger front harmony: oktober + nek "oktober + nak 'october (dat.)' Jozsef + nek * Jozsef + nak 'Joseph (dat.)' iv'. If e were neutral, it should let i or é: *taxi + nek taxi *kâvé + nek kâvé Jozsef + nek but [Jozs+i]

back harmony through like + nak + nak + nak

'taxi (dat.)' 'coffee (dat.)' 'little Joseph (dat.)'

The disagreement as to the status of low e stems from its irregular behaviour in disharmonie stems. There is a class of exceptions to (47iv): certain nonnative words can take either front or back harmony. (47)

v.

Âgnes + nek kôdex + nek

or or

Âgnes + nak kôdex + nak

'Agnes (dat.)' 'codex (dat.)'

In these disharmonie stems, which are called doublets, final e is optionally transparent. However, according to Jensen (1984): "In fact, orthographic e represents two originally distinct vowels, neutralized in the standard language but retained in many dialects. In dialect studies these are distinguished as mid e and low e, and I will use this notation also, when the distinction is relevant to the discussion."

Significantly, Jensen distinguishes the stems in (47iv) from those in (47v) in terms of low e and mid e (respectively): "Jozsef "Agnes

Jozsef Agnes

+ nek + nek

* Jozsef -I- n a k " Agnes + n a k "

If neutral e was originally (and in certain dialects still is) a mid vowel, then the class of stems containing a neutral e does not invalidate our claim that only vowels whose kernel is the harmonic element are predictably

66

Hamida Demirdache

transparent, but confirms it. For fusion to yield a mid vowel, I must be the kernel and A an operator. To explain the synchronic vacillations of low e, we assume that this vowel is still analysed as a mid vowel in certain idiolects or dialects. In other wordt, the sound change e > e , which is merely the reversal of the kernel/operator relation in the compounding of I and A, has not entailed (systematic) restructuring of lexical representations. To account for the [+low] surface vowel, we assume that at a lower level of representation, the kernel/operator reversal takes place mirroring the sound change.

10. THE COLD VOWEL AS THE KERNEL OF AN EXPRESSION

To explain the transparency of low e in Hungarian, we have assumed that in certain dialects, this vowel is still analysed as the mid vowel it originally was. In other dialects, restructuring of lexical representations has taken place and e governs front harmony. Thus e in Jozsef triggers front harmony, whereas e in Agnes > Agnes can either trigger front harmony or let back harmony through. This appears to be a general property of neutral vowels in backness harmonies. Consider the data given below. (48)

Hungarian i. hid + nak eel + nak ii.

viz ver

+ nek + nek

'bridge (dat.)' 'goal (dat.)' 'water 'blood

(dat.)' (dat.)'

iii. Doublets analizis rôkamié Agnes but *József

+ + + +

nak nak nak nak

iv. Vacillating suffixes - ne - 'wife of elnok 'president' tanar 'teacher'

or or or

analizis rokamie Agnes József

+ + + +

nek nek nek nek

elnok + ne + nek tanar + ne + nek

'analysis 'sofa 'Agnes 'Joseph

(dat.)' (dat.)' (dat.)' (dat.)'

or tanar + ne + nak

In the regular case, neutral vowels do not participate in backness harmony. On the contrary, these transparent vowels let harmony through. However, in the cases illustrated in (48), they can control the value of backness

Transparent Vowels

67

in suffix vowels, they can govern front harmony. Thus, neutral vowels in a backness harmony system are not always governed, they can also be governing vowels, vowels controlling frontness in suffixes. This is clearly a lexical property of the stem in which they occur: in terms of their vowel quality, the stems in (48i, ii) are identical (hid/viz, cel/ver). In (48iii, iv) both options (governed/governing vowel) are possible. The analysis of transparent vowels we have presented, is an attempt to relate the internal structure and composition of segments to their harmonic behaviour, to explain the latter in terms of the former. Thus, we will try to explain the dual behaviour (transparent/harmonic) of neutral vowels, in terms of their internal structure. In section 7.2, we said that vowels are always a combination of three elements. That is, they are expressions involving a relation between a kernel and two operators, though one of these operators (or both) may contribute no feature value in the fusion operation. The vowel e is the combination of the kernel I, and the operators A, V; the vowel /', the combination of the kernel I, and the operators V. This is very clear in a phonological representation where lines are labelled in terms of the marked features back/round/high, (49)

i.

back round high

\ I V I A e

ii.

I I V I V i

The cold vowel has by definition no marked feature values. Its occurrence on the round line in (49i) and on the round and high lines in (49ii) means that e is [-round] and i, [-round,+high]. We have differentiated front harmonic vowels and transparent vowels in terms of the kernel/operator relation. When I is a kernel, a front vowel is transparent; when it is an operator, a front vowel governs front harmony. Thus, neutral vowels will trigger front harmony, only if they are analysable as expressions of which the I component is an operator. If such is the case, which element is the kernel? Clearly, it can not be A. There is no A in the representation of the vowel i, and (I*A) does not yield a mid but a low front unrounded vowel (see (42)). If both I and A must be operators, then the cold vowel must be the kernel. Transparent vowels in a backness harmony system are analysable as expressions in which the cold vowel is the kernel, and the I component an operator: the fusion operation (V*I) is commutative.

68

Hamida Demirdache

(50)

u.

- round +back +high - low V

- round -BACK +high - low " I

- round -BACK +high - low

- round -BACK +high - low I

- round +back +high - low V

- round -BACK +high j- low i

In the data given in (49), neutral vowels can either let harmony through or govern front harmony. We can explain this property of neutral vowels in backness harmonies in terms of a property of the fusion operation that yields these vowels: the fusion operation is commutative, (V*I) = (I*V). The kernel/operator relation can be reversed. We will assume that when a neutral vowel is a governed vowel, it is analysed as an expression in which I is a kernel; when it is a governing vowel, it is analysed as an expression in which I is an operator. (51)

Transparency of i and e (VI) = i A-(VI) = e kernel operator

—V — — V —

operator (51')

Opacity of i and e (i'Y) = i A*(I*V) = e

1 2 I I V V h i d + nak

—A

— V1 I

v

2 I

V -

cé1 + nak

Transparent Vowels kernel

69 V 1 — V—( \ l 1

operator

operator

I viz

A

V

1 V—| \ l 2

—A—|

A— I I V—|— \ l 1 2

I v é r + n e k

+ nek

The vowels in hid/viz and cél/vér are identical. To explain their respective harmonic behaviour, we have assumed that they are analysed as asymmetrical expressions in the lexicon. As for vacillating stems (doublets) and suffixes, which can take both front and back vowel suffixes indifferently, we can assume either, that the option 'reverse the kernel/operator relation (I/V)' is a lexical property of these items; or that different idiolects select different analyses of these vowels. Likewise, to explain the harmonic behaviour of ù/ô in disharmonie stems (note that doublets also, are disharmonie stems), we assumed that they were analysed, in different dialects, as asymmetrical expressions, expressions in which the kernel/ operator relation (I/U) is reversed. This reversal of the kernel/operator relation is not always a lexical property of an item. It can also follow from a universal principle governing phonological representations and structures. A major characteristic of backness harmony in Finnish and Hungarian, is that it is directional and unbounded. I must spread along the operator line until all vowel positions are in its scope. In other words, (52)

Government domains are exhaustive where the governor X e (I, U, A)

(52) is a well-formedness condition. In a language where the relation 'X(Y)' holds, it rules out the following ill-formed representations: (52')

i.

* —X 1

ii. 2

* —X

3

Consider now the data given in (53), (53)

Finnish Monosyllabic neutral vowel stems i. [el] + o 'life' [mitt] + u

1

2

3

70

Hamida Demirdache ii.

+ + +

[el + el] [el] [mitt + el]

ii io ii

'living' 'organism' 'match'

Compounds iii. [ [volta] + metri] + sta [ [baro ] + metri] + sta [ [ [plus] + kvam] +perfekti] + sta iv. [ [polii] + teismi] + sta or [ [ polii] + teismi] + sta [ [ polii] + teekki] + sta or [ [ polii] + teekki] + sta Hungarian v. hid + nak cél + nak vi. oxigen + nek

*hid + nek *cél + nek *oxigen + nak

'from the voltameter' 'from the barometer' 'from a verb tense' 'from polytheism' 'from the polytechnic school'

'oxygen (dat.)'

In (53), we see that native stems with a single neutral vowel normally require [+back] suffixes; whereas stems with a sequence of neutral vowels require [-back] suffixes. In Finnish compounds (loanwords), where the last harmonic vowel is followed by several neutral vowels, we find the kind of vacillation already seen in Hungarian (49iii, iv). What is the lexical representation of a stem containing a sequence of neutral vowels? (54)

kernel

I

[ operator

U

2

1

V [ m i t t

+

]

+

3

v e 1 ]

+

u

v -

In (54), the identity of adjacent elements on the line identifying the respective kernels of compounds, has triggered the OCP (Obligatory Contour Principle). The OCP is a lexical constraint to which phonological representations are subject. It is a condition of distinctness prohibiting identity of adjacent elements at any level of representation (back/round/high lines, kernel/ operators lines). We interpret it as securing deletion (under identity) and concomitant spreading. The output of OCP in (54) is a branching structure. That is, a harmonic structure: an element binds a sequence of skeletal positions. In (54), the relation 'X(Y)' holds, where the governor is I and

Transparent Vowels

71

its domain the substring (1 2). The OCP, like harmony, is an instance of government. Thus, the representation in (54) is subject to, (52")

Government domains are exhaustive, where the governor is I

The data in ((49), (53)) illustrates the harmonic behaviour of neutral vowels: they normally let back harmony through, but can also govern front harmony. We have explained this property of neutral vowels in backness harmonies, in terms of a property of the fusion operation which yields these vowels. The kernel/operator relation (I/V or I/U) is reversable. We are suggesting that this reversal is either the lexical property of an item (hid+nak, viz+nek), or follows from a general principle governing phonological representations, (52'). The representation in (54) is ill-formed. To satisfy (52'), I must spread to all free vowel positions. In Finnish backness harmony, only operators are licensed to spread. In (54), I can not spread along the kernel line. The kernel/operator reversal takes place to secure spreading to all free positions. (54')

kernel operator

operator [ m i t t

+ e1 ]

+

u

Thus, whenever an autosegment I (lexically) binds a sequence of skeletal positions, (52') triggers spreading of the operator I to all suffix vowels.11 Polysyllabic neutral vowel stems will take front suffixes. (55)

Finnish piene + sta 'little (elative)' kernel I A—

V

operator— V — + 4 operator

I V—V—V pi e n e

V— +sta

>

V—V

A—

72

Hamida Demirdache

The words listed in (49iii) and (53iii, iv) take both front and back suffixes. When they take back suffixes, they violate principle (52'). Perhaps we can relate this vacillation to the fact that these stems are borrowings or compounds that do not respect vowel harmony restrictions. That is, a stem such as 'rokamie' is itself a violation of principle (52'). In the regular case, neutral vowels are expressions of which I is the kernel, they are transparent. However, they can also govern front harmony. We have explained this property of neutral vowels in backness harmony, in terms of a property of the fusion operation which yields these vowels. The reversal of the kernel/operator relation I/V or I/U can be a lexical property of an item. It can also follow from a general principle governing phonological representations, principle (52). We have seen that both the OCP and harmony are instances of government. Structures generated by the OCP, where the governor is I, will thus be subject to principle (52).

11. CONCLUDING REMARKS

If we were to identify the class of transparent vowels (as opposed to the class of harmonic vowels) in terms of feature specifications, we would identify the class of transparent vowels in Montañés in terms of the feature [+low], and in Finnish or Hungarian, in terms of the conjunction [-low, -round]. In our framework, we can not manipulate features. Thus, only one means of identification is available: in terms of a paradigmatic relation of dominance between a head (the kernel) and its dependents (operators). This relation of dominance is directly encoded into the autosegmental representation of segments, and formally characterized by an operation: fusion. The syntagmatic dependency relation holding between a head and its domain, has been sufficiently motivated both in phonology, where it is the basis of metrical phonology, and in syntax. We have attempted to demonstrate that this relation is crucial in explaining the transparent behaviour of vowels: in Finnish and Hungarian backness harmony, the set of transparent vowels is the set of vowels whose kernel is I: (/, e), and optionally (it, d) in Finnish. In Montañés height harmony, the set of transparent vowels is the set of vowels whose kernel is A: (a/A). Thus, analysing vowels in terms of this kernel/operator asymmetry predicts which vowels are transparent, and which vowels alternate. To account for harmonic systems with/without neutral vowels, we merely assume the null hypothesis: though fusion may be asymmetrical, the value (kernel/operator) of the governor is constant. When it spreads along a line, it behaves uniformly as either an operator or a kernel. When it is an operator, certain vowels are transparent. When it is a kernel, all vowels alternate. To generate (dis)harmonic sequences with transparent vowels,

73

Transparent Vowels

we neither posit abstract lexical representations with vowels which do not exist phonetically, formulate rules or filters so as to skip their occurrences, nor exempt these vowels from harmony.

NOTES •This paper was written in September 1986 while I was at the University of Paris 7, Jussieu. I am very indebted to Jean-Roger Vergnaud for long and insightful discussions of the work presented here. I wish to thank Francois Dell for his comments and careful reviewing of a previous draft. 1. Hungarian has a 14 vowel system (7 short/7 long vowels) as given below. We are following standard Hungarian orthography, the acute accent marks length, the diaeresis marks fronting, and the double acute accent, fronting and length. Short a is phonetically [+round], and short e [+low]. We have classified this vowel as a harmonic vowel, its status however is controversial. It will be discussed in section 9. Harmonic vowels front short long ü ü ö ö e = [ae] 2.

Neutral vowels back short long u ü o ó a = [o] â

short i

long i e

Montañés Spanish (data from McCarthy 1984) has the following vowel system: tense lax

i I

e

a o u A O U

It has both height harmony and ATR harmony. Both tense and lax [+low] vowels (a/A) are transparent in height harmony; e is transparent in ATR harmony. 3. Should not Alters of this type be barred from the grammar, as are all 'looking ahead' rules? For these filters to prohibit linkings, they must 'know' that these linkings will generate impossible feature combinations. 4. There seems to be some redundancy in having in the grammar, both filters of the form (i) and redundancy rules of the form (ii): (i)

*+ATR i i V [+low]

ii.

[+low]

-

[-ATR]

5. In KL&V 1985, there are in fact, four basic elements relevant for vowel systems: (I, U, A, !•). í is the ATR element. We have not taken it into consideration, because it is irrelevant here. 6. KL&V call this unmarked vowel, the cold vowel (and marked features, hot features).

74

Hamida Demirdache

7. In Dependency Phonology, this relation of dominance (of 'dependency') is expressed in terms of the notation: i.

/e/=i I

ii.

/e/=a I

a

i

and in Particle Phonology, in terms of particle multiplication: iii.

[e] = ai

iv.

[e] = aai

8. See footnote 2. If both a/A are transparent in Montañés height harmony, then the set of transparent vowels in Montañés is the set of vowels whose kernel is the harmonic element. We have ignored the transparency of A, because we did not wish to consider here the question of ATR harmonies. 9. In the autosegmental analyses of neutral vowels, nothing hinges on the geometry of representations. Steriade (1986) for instance assumes, (a) a multi-tiered model of feature organization similar to that of Clements (1985), (b) underspecification theory, but it is the Alters (4ii) discussed in section 2, which secure the transparent behaviour of vowels. 10. A few remarks are in order concerning rounding harmony in Khalkha Mongolian. In this language, we have a backness harmony and a parasitic rounding harmony, in which respectively (i.e) and (i) are transparent. The transparency of i in rounding harmony does not invalidate our analysis of transparent vowels. It follows from the structure of autosegmental representations. In Khalkha, we have three lines (or half-planes anchored onto the skeleton), back

I 1

high

2

3

i

a/ o

A o

In the above representation, the vowel / is transparent in rounding harmony, because it is never the target of spreading. Rounding harmony involves only non-high vowels: o/o are triggers, and a/e are targets. The vowel / is disregarded in rounding harmony, just like intervening consonants. The representation of this type of transparency is straightforward, (see Goldsmith 1985). 11. In fact, we could derive this reversal of the kernel/operator relation from the OCP. In Finnish, the values of parameters (35) are i. ii. iii.

Governor I Government by an operator Government to the right

If the OCP secures deletion under identity and concomitant spreading, then I in (iv) cannot spread to the free position, unless the kernel/operator relation I/V is reversed,

Transparent Vowels iv.

I I 1 I V

75

I I 2 I V

We must assume that this reversal takes place. Recall that we explained the alternations a/a and a/e in terms of the value of the parameter government by an operator/government by a kernel. Whatever the correct solution is, (52) is still part of the grammar of a language where harmony is unbounded, and the OCP and (52) account for front suffixes after polysyllabic neutral vowel stems.

REFERENCES Anderson, J.M. and C. Jones. 1974. Three theses concerning phonological representations. JL 10, 1-26. Anderson, J.M., C. Ewen and J. Staun. 1985. Phonological structure: segmental, suprasegmental and extrasegmental. In: Phonology Yearbook 2. Cambridge University Press. Anderson, S.R. 1980. Problems and perspectives in the description of vowel harmony. In: R. Vago (1980) (ed.), 1-48. Archangeli, D. 1984. Underspeciflcation in Yawelmani phonology and morphology. PhD dissertation, MIT. Booij, G. 1984. Neutral vowels: an autosegmental analysis of vowel harmony. Linguistics 22,629-641. Campbell, L. 1980. The psychological and sociological reality of Finnish vowel harmony. In: R. Vago, 1980 (ed.), 245-270. Chomsky, N. and M. Halle. 1968. The sound pattern of English. New York: Harper and Row. Clements, G.N. 1976. Vowel harmony in nonlinear generative phonology: an autosegmental model. (Distributed by Indiana University Linguistics Club). Clements, G.N. 1981. Akan vowel harmony: a nonlinear analysis. In: G.N. Clements (ed.), Harvard studies in Phonology. Vol II, 108-177. (Distributed by Indiana University Linguistics Club). Clements, G.N. and E. Sezer. 1982. Vowel and consonant disharmony in Turkish. In: Van der Hülst and Smith, 1982 (eds.), Part II, 213-255. Clements, G.N. 1985. The geometry of phonological features. In: Phonology Yearbook 2. Cambridge University Press. Ewen, C. and H.G. van der Hülst. 1985. Single-valued features and the nonlinear analysis of vowel harmony. In: Bennis and Beukema (eds.) Linguistics in the Netherlands 1985. Dordrecht: Foris Publications, 39-48. Goldsmith, J. 1985. Vowel harmony in Khalkha Mongolian, Yaka, Finnish and Hungarian. In: Phonology Yearbook 2. Cambridge University Press. Halle, M. and J.R. Vergnaud. 1981. Harmony Processes. In: W. Klein and W. Levelt (eds.), Crossing the boundaries in linguistics, Reidel, 1-22. Hülst, H.G. van der. 1985. Vowel harmony in Hungarian. A comparison of segmental and autosegmental analyses. In: H. van der Hülst and N. Smith, 1985 (eds), 267-303. Hülst, H.G. van der and N. Smith. 1986. On neutral vowels. In: K. Bogers, H. van der Hülst and M. Mous (eds), The representation of suprasegmental in African Languages. Dordrecht: Foris Publications. 223-279.

76

Hamida Demirdache

Jensen, J. 1984. A lexical phonology treatment of Hungarian vowel harmony. Linguistic Analysis, Vol. 14, Number 2-3, 231-253. Kaye, J., J. Lowenstamm and J.R. Vergnaud. 1985. The internal structure of phonological elements: A theory of charm and government. In: Phonological Yearbook 2. Cambridge University Press. Kiparsky, P. 1973. Phonological representations. In: Three dimensions of linguistic theory, TEC Corp., Tokyo. Kiparsky, P. 1981. Vowel harmony. Unpublished Ms MIT. McCarthy, J.J. 1984. Theoretical consequences of Montañés vowel harmony. LI, Vol. 15, Number 2, 291-318. Pulleyblank, D. 1985. Tonal and vocalic redundancy rules. Paper presented at the Colloque phonologie pluri-linéaire, Lyon 1985. Rennison, J. 1983. On tridirectional feature systems for vowels. Wiener Linguistische Gazette 33-34, 69-94. Ringen, C. 1980. A concrete analysis of Hungarian vowel harmony. In: R. Vago, 1980 (ed.), 133-154. Schane, S.A. 1984. The fundamentals of particle phonology. In: Phonology Yearbook 1. Cambridge University Press. Skousen, R. 1972. Finnish vowel harmony: rules and conditions. In: M. Kenstowicz and C. Kisseberth (eds.); Issues in phonological theory. The Hague: Mouton, 118-129. Steriade, D. 1986. Non-underlying feature specifications. Paper presented at the Wassenaar Workshop 'On features', June 1986. Vago, R. 1980. A critique of suprasegmental theories. In: R. Vago, 1980 (ed.), 155-181. Vago, R. 1980. (ed.). Issues in vowel harmony. Proceedings of the CUNY linguistics conference on vowel harmony, 14th May 1977. Amsterdam: Benjamins.

The Geometry of Vocalic Features Harry van der Hülst University of Leiden

0. INTRODUCTION

Recent years have shown an increased interest in simplex or single-valued phonological features', for an early statement, cf. Sanders (1972). In my view, a single-valued feature system represents the logical end point of what is called Radical Underspecification (Kiparsky 1982, Archangeli 1984, 1988). The central goal is to express that the two members of a binary opposition do not play the same role in the phonology. Radical Underspecification Theory expresses this by using binary features, while stating that only one value may be specified in the underlying representations; the other value, often called the default value, is added during or at the end of the derivation. In a single-valued approach a stronger claim is made. The default value is eliminated as a phonological entity altogether, so that the feature is single-valued at all stages of the phonological derivation. Current single-valued theories are also stronger than current underspecification theories in another respect: in a single-valued theory it is always the same member of the opposition that represents the marked phonological property. Underspecification approaches allow the possibility that either of the two values of any feature is marked. In this article I propose a particular single-valued feature system for vowels, which differs from other current single-valued systems, although it shares fundamental insights with some of them. In section 1, I will outline the model and concentrate on providing typological and phonetic justification. In sections 2 and 3 I will turn to the phonological argumentation and suggest how the system can be put to use in the analysis of vowel harmony processes, and in section 4 I offer a discussion of some potential arguments against a single-valued system. Section 5 addresses the issue of transparency and opacity in harmony systems.

1. THE PROPOSAL: A SINGLE-VALUED FEATURE SYSTEM FOR VOWELS

My proposal can be seen as a development of Dependency Phonology (DP,

78

Harry van der Hülst

cf. Anderson & Ewen 1987) and Government-based Phonology (GBP, cf. Kaye, Lowenstamm & Vergnaud 1985) and much work inspired by (aspects of) these approaches (e.g. Schane 1984, Goldsmith 1985, 1987, Rennison 1986, 1987a, 1987b, Van der Hulst & Smith 1985, 1986a). For brief summaries of both DP and GBP and a comparison to the present proposal I refer to Van der Hulst (1987, forthc. a, b). For a more extensive and somewhat critical discussion of these approaches, as well as of binary underspeciflcation approaches, I refer to Den Dikken & Van der Hulst (1988). The argumentation in favour of the single-valued feature system is complex, depending on a variety of empirical and conceptual considerations. By way of introduction, let me draw attention to the fact that studies of typological surveys of phonological systems, e.g. those of Crothers (1978) or Maddieson (1984), reveal that, in general, there are more oppositions amongst high vowels than amongst low vowels, so that languages having two low vowels differing only in the front-back dimension, and not in length, are typologically marked. A feature system such as that of SPE, however, treats the vowel space as basically rectangular, as in (1):

(1)

[-back] /i/ /e/ /ae/

[+high] [-high] [-high]

[+back] /u/ /o/ /o/

[ - low] [ - low] [+low]

On the basis of typological considerations, however, we would prefer rather to have a feature system which characterizes vowel systems as basically triangular (cf. Fisher-Jorgensen 1984a,b and Basboll 1984 for a discussion of these matters). We can do this, following DP and GBP, by assuming a different feature system which takes the three points of the vowel triangle as primitives. This gives us three features, which will be represented in terms of the letters | i |, | u | and | a | between vertical strokes, as in DP. A simple three vowel system consisting of [i], [u] and [a] can now be represented in the following way: (2)

/i/ |i|

/a/ Ia|

/u/ | u|

In DP the term component is used, instead of feature, and in GBP the primitives are referred to as elements. Here, I ignore potential substantial issues which are involved in the choice of terminology, and continue to use the term single-valued or unary feature. (Instead of single-valued, others have used terms like simplex (Sanders 1972), monovalent (Archangeli 1988) or singulary (Chafe 1970) in the same sense.)

The Geometry of Vocalic Features

79

Not all vowel systems have the form in (2), of course. Vowel systems may include, for example, mid vowels, front rounded, back unrounded, central vowels, advanced vs. retracted vowels, etc. Within the present approach, and again I share this with DP and GBP, additional vowels are represented by combinations of the three features. For example, mid front vowels can be represented as combinations of the features | i | and |a|. But vowel systems can contain more than one mid front vowel. Assuming that we want to characterize all of them as combinations of | i | and | a |, we then need some way of stating the difference. This is where the dependency or government relation comes in. A characteristic feature of the present approach, also shared with both DP and GBP, is that the features which make up the content of segments are organized in terms of an asymmetrical relation indicating, in intuitive terms, which of them represents the more salient property. Hence, given that we have two features | i | and | a |, there are two possible feature structures: (3)

a.

/e/ i I a

b.

/e/ a I i

(governor) (dependent)

Using representations as in (3) raises the question how we interpret them phonetically and how we can actually tell that combining | i | and | a | gives us a front mid unrounded vowel rather than a voiceless lateral velar fricative. This is where the phonetic interpretation of the features becomes relevant. In my approach the status of each feature as either governor or dependent is reflected by a distinct phonetic interpretation, corresponding to what would be a separate phonological feature in feature systems which do not make use of the government-dependency relation (henceforth Grelation). The interpretation I propose for |u| is the following: (4)

Interpretation of Governor: Dependent:

|u| Velar constriction I Rounding

These two aspects of |u | correspond to different articulatory gestures which naturally go together in the sense that lip-rounding enhances the acoustic effect of velar constriction (cf. Stevens, Keyser & Kawasaki 1987). It is, therefore, far from arbitrary to give formal expression to the intimate relation between roundness and backness in the way proposed here; although different from an articulatory point of view, they are the same

80

Harry van der Hülst

in acoustic terms. I would like to go further and state that it must be deducible from any theory of features that the features [velar] (or [back]) and [labial] (or [round]) are intimately related. What I propose here reflects the same idea that has prompted our choice for the single-valuedness of features in the first place, viz. the idea that markedness considerations should be expressed directly in the primitives and principles of the theory. The use of single-valued features represents an attempt to deduce markedness from the basic structure of the theory. However, markedness not only involves the "context-free" phenomenon that one value of every feature is marked, it also involves "contextdependent" phenomena such as the fact that backness and roundness tend to go together. The present proposal, then, represents an attempt to render superfluous two types of SPE "markedness rules" by building their content directly into the phonological formalism. In (5), the first rule type is replaced by the claim that features are single-valued, while the second type is replaced by the dual interpretation of the phonological primitives: (5)

a. b.

[Uround] — [-round] [Uround] — [around] / [ - , aback]

The advantage we gain is that the "content" of the expressions in (5) is no longer arbitrary from the perspective of our theory. For the features | i | and | a |, I suggest the following dual status: (6)

a.

Interpretation of | i | G: Palatal constriction I D: Advanced tongue root

b.

Interpretation of | a | G: Pharyngeal constriction I D: Openness

Can we justify the grouping of properties in these cases, too? We know that palatal constriction results from advancing the tongue root, as indicated for example in the studies of Wood (1982). It is therefore not arbitrary to suggest that [Advanced tongue root] is closely linked to [Palatal constriction]. Similarly, it is also clear that a-type vowels (i.e. vowels produced with pharyngeal constriction) are produced with a jaw opening which is wider than that for u-type vowels, thus showing that jaw opening and pharyngeal constriction tend to cooccur as well. The use of | i|, | u | and | a | , instead of the SPE features in (1), with their interpretation as linguistically relevant constriction locations, ties in rather well with findings and proposals regarding vowel systems, presented in Wood (1982), who shows in some detail that the traditional tongue arch model in which vowels are characterized according to the location of the highest point of the tongue simply cannot be maintained in the

The Geometry of Vocalic Features

81

light of X-ray recordings of actual vowel production. Wood's feature system is briefly discussed in Van der Hulst (1987, forthc. a). The typological and phonetic considerations just mentioned seem to suggest an initial plausibility for the proposal under consideration, but are not intended as a substitute for phonological justification, which will be supplied in the following sections. Let us first look, however, at some further aspects of the model. Consider the representations in (7) of simple three or five vowel systems: (7)

a.

/i//u//a/ i u a

b.

/i//e//u//o//a/ i i u u a I I a a

I will adopt a universal redundancy rule which assigns a dependent feature identical to the governor, unless its absence is distinctive in the system. In this way I account for the fact that, for example, a back vowel is naturally rounded: (8)

Universal Redundancy Rule: f f I f

Furthermore, I assume that the phonetic interpretation of features is absolute, in the sense of non-gradual. From this it follows that a single feature can occur at most twice in the representation of a single segment, once as a governor and once as a dependent; there is no point in stating twice that a segment is rounded. This means that for any pair of features we can generate maximally 8 feature structures; take |i| and |a| as an example: (9)

i I i

i

i I a,i

i I a

a I i

a I a,i

a

a I a

I also assume that the phonological representation of a particular segment is, in part at least, system-dependent. For example, the way in which a particular mid front unrounded vowel is represented in a given system will depend on its phonological behaviour and on the overall structure of the vowel system. Thus, I will allow an e-type vowel to be represented in terms of different structures, e.g.:

82 (10)

Harry van der Hülst a.

i

b.

i I a,i

c.

i I a

We will see various examples of the many-to-one relations between feature structures and segments. On this basis, I conclude that this flexibility, although perhaps suspect at first sight, is actually necessary. It is perhaps the case that we allow more combinations of two features than are allowed in DP (cf. on this Den Dikken & Van der Hulst 1988) and in any event we allow more than are allowed in GBP, because every feature can occur twice in the representation of a single segment. The total number of potential contrasts is not disturbingly great, even though we do generate more contrasts than any individual language will allow at the level of lexically distinctive segments. This is not the kind of expressive power one is after, but I see no way of avoiding it completely. In this system, there is no need for an independent feature [ATR], since ATR is identified with the feature | i| in dependent position. I stress this, because both DP and GBP have a separate feature [ATR]. Also, DP and GBP use a fifth component (i.e. | "centrality") or element (i.e. [v], the so-called "cold vowel"), which plays an essential role in the characterization of central and back unrounded vowels. In the present proposal we can characterize central and back unrounded vowels without the use of extra features. Consider the following representations: (11)

a.

/ui / u

/u/ u I u

b.

/i/ i I i

/i/ i

Given the phonetic interpretation of bur features, (11a) represents the distinction between a back unrounded and a back rounded vowel. The feature specifications in ( l i b ) represent a distinction between an advanced high front vowel and its non-advanced or slightly less fronted counterpart. It seems to me that these are the kinds of distinctions we can use for systems having central or back unrounded vowels. One could object to the representations in (11), saying that rounded back vowels are formally more complex than their unrounded counterparts, whereas they are less marked typologically. One should not fail to notice, however, that it is the presence of the back unrounded vowel in a system which causes this complexity. If such a vowel is lacking, / u / is simply represented in terms of a governor |u| without |u| as a dependent feature. It is not obvious that "system-complexity" should be reflected in the representation of the sounds which tend to only occur in more complex

The Geometry of Vocalic Features

83

systems. To be explicit: I will reject this correlation. In so far as fully specified feature structures in isolation represent complexity, this is phonetic (in particular, articulatory) complexity rather than typological complexity. The latter type of complexity can be derived from looking at the vowel system as a whole. In the next section, I will introduce some further aspects of the present model. Firstly, it will be argued that the three features | i|, | u| and | a| are organized in a binary geometry, which leads us to postulate a node dominating | i| and | u|. This node, referred to as | y|, will turn out to share some properties with | s \ and [v] mentioned above. A second important point will be that the desire to discriminate between distinctive and nondistinctive information will lead us to adopt underspecification within our model. It will be shown that such forms of underspecification are not only economical but also lead to more optimal analyses. Before we turn to the next two sections, which offer a discussion of some vowel harmony systems, there is one other line of argumentation in support of the present approach that I want to mention here. In order to explain why certain segments fail to initiate or to interrupt a particular spreading process (even though, phonetically at least, they bear the relevant property), it is proposed in the current literature that such segments lack a specification on the relevant tier. Steriade (1987) argues that the absence of such specifications either results from the fact that the relevant feature is single-valued (universally or in the language at issue), in which case she speaks of trivial underspecification, or from the fact that one of the values of the binary feature has been left unspecified because it is not contrastive. In that case, Steriade speaks of nontrivial underspecification. Here my concern is to suggest, without going into full detail, that certain cases mentioned by Steriade as examples of nontrivial underspecification turn out to involve trivial underspecification if we make use of the feature system adopted in this paper. If this is correct, it would show that our single-valued feature system is superior to the underspecification approach because it explains rather than stipulates why certain types of segments fail to trigger, or fail to interfere with, certain processes. Steriade discusses certain dissimilation phenomena in Ngbaka and Ainu, involving the features [high] and [back], arguing that / a / is crucially (but not trivially) unspecified for the feature [high] in Ngbaka and [back] in Ainu. Thus, the argument goes, we can explain that / a / does not participate in certain processes that make reference to these features. Both cases will be analysed in section 3. Anticipating this more detailed discussion, I would like to point out now that the behaviour of / a / follows straightforwardly in our model from the fact that it is exhaustively specified in terms of the feature | a|. The fact that this vowel does not interfere

84

Harry van der Hülst

with processes which refer to height or backness, then, follows from trivial underspecification. From these facts we may draw the conclusion that the cases Steriade uses to exemplify nontrivial underspecification theory in fact provide strong arguments in favour of a single-valued approach in that most - if not all cases of so-called nontrivial underspecification turn out to be instances of trivial underspecification. In this regard, however, the choice of the feature system plays an essential role. In the examples just given / a / "does not get in the way" if reference is made to "high" because the height dimension is represented by a single-valued feature "non-high" (i.e. | a|). In another example discussed by Steriade (from Tamil) the point is that / a / has no specification for [back]. Here we derive the "irrelevance" of / a / , because the low vowel has no specification whatsoever indicating its position along the front-back axis.

2. HARMONY SYSTEMS

In this section I intend to illustrate the characteristics of the feature system proposed here by offering analyses of a number of harmony systems. Because I have chosen here to give a "broad" illustration, these analyses are rather schematic and incomplete, and are presented here in a largely "data free" manner. In almost all cases, however, I discuss fairly familiar data (for which references will be given) or I summarize results which are discussed in more detail elsewhere. The reader who consults this section looking for an introduction to the phenomenon of vowel harmony will end up frustrated. (S)he is referred to the various studies in Vago (1980), in particular the introduction and Anderson's contribution to that volume. A typological overview of vowel harmony is offered in Ultan (1973). 2.1. | i | -systems Since dependent | i | represents ATR, it seems as if we cannot make a distinction at the phonological level between ATR-harmony and palatal harmony, as for example in Finnish and Hungarian. Indeed, I want to suggest that the two types of systems are closely linked, in that both involve the spreading of | i|. This is precisely what we want. Firstly, it has been claimed that there can be a smooth diachronic development from one into the other (cf. Svantesson 1985 on Mongolian), which suggests that the two are closely related, and, secondly, there seem to be restrictions on the possibility of combining palatal harmony and ATR-harmony in a single system, which suggests that the two do not involve independent features. The question as to whether the spreading | i| has dependent status

The Geometry of Vocalic Features

85

in both types of systems (as suggested in Van der Hulst 1987) will be discussed below. 2.1.1. ATR-systems Typical ATR-systems have two sets of vowels, a [+ATR] set and a [-ATR] set. Williamson (1973, 1984) discusses a whole array of such systems, suggesting that a 2 x 5 system is prototypical. I will assume that such a "full" ten-vowel ATR-system is characterized as follows: (12)

/i/ /u/ /e/ /o/ /a/ i u i u a I I I I I i i a,i a,i i

| | I |

A / Ao/ /e/ /o/ / a / i u i u a I I a a

/ e / is represented as | i—-a | rather than | a—i | to ensure that only vowels in the [+ATR] set have a dependent | i | . In conjunction with this, the claim has to be that ATR-harmony involves spreading of dependent | i| only. It is not clear from the representation that / a / characterizes the most marked combination, as is generally assumed. In fact, our framework has had very little to say about preferred and dispreferred feature structures, so far. We only know that the presence of a dependent | f| is most natural for segments which have | f| as a governor. This unmarked cooccurrence has been expressed in terms of the universal redundancy rule in (8). From this point of view, / u / and / o / should be as marked as / a / , although one would presumably want to obtain the ranking which reflects the usual pattern of decay in ATR-systems (cf. Williamson 1973, 1984, Lindau 1975, Van der Hulst, Mous & Smith 1986a,b): (13)

/i/ Preferred — |

/e/ /u/

/o/

/a/ |—> dispreferred

This could be obtained if we express formally that the features | i| and | u | have something in common, as opposed to | a |. This is not an unexpected manoeuvre. The idea that the properties denoted by | i | and | u | (when interpreted as dependents) can be grouped together as the "tonality" or "colour" features can be traced back to the work of Trubetzkoy (1939) and Jakobson (1941) (cf. also Donegan 1978). Furthermore, when interpreted as governors, these features are closely linked in that both involve tongue-body constriction. Let us assume, therefore, that | i| and | u | (disregarding their specific interpretations) form a structural unit which

86

Harry van der Hülst

is opposed to | a|. For reasons independent of the present discussion, Ewen & Van der Hulst (1988) make a proposal of this kind, which they express as follows: (14)

o / |a|

\ I y| /

\

I will assume that (14) expresses the geometry of vocalic (or rather: place) features. There is the question as to what exactly the status of | y | is. This issue has not been fully explored at present and is discussed in more detail in Van der Hulst (forthc. b). Here, I will simply assume that | y| can occur without either | i | or | u |. As such it represents an incomplete specification which is usually provided with | u| or | i| to make it complete. Below, however, we will see that in certain cases we are forced to say that the incomplete | y| surfaces. I will interpret this as a parametric choice. Since | y | is not a "real" feature it has no interpretation independently of | i| or | u|. Occurring alone it represents the complete absence of any positive property. Occurring as a dependent it adds nothing to the interpretation of a feature structure; being "nothing" it can never govern | a | when incomplete. Given (14), we have a basis for saying that dependent | i| combines more readily with | u | than with | a | , because | i| and | u | are closer to each other to begin with. The generalization, then, would be that the closer a dependent to its governor, the more likely the combination. I.e., | i| is closest to itself - hence (8) expresses the most likely combination and closer to | u | than to | a |. According to this / e / is better than / o / , because its non-ATR counterpart already contains | i |. This reasoning gives us the scale of (dis)preferred combinations in (15). This way of thinking about feature combinations, i.e. in terms of government restrictions, will be exploited further in this paper (cf. the discussion of Kpokolo below and of Finnish in section 2.1.2). But not all "segment structure conditions" can be interpreted as such. As just indicated, whether or not bare | y| is allowed to surface is a different type of restriction. We will also encounter "complexity restrictions", i.e. restrictions limiting the number of dependents that a feature structure can contain. Such complexity restrictions could possibly be reduced to government restrictions. A comprehensive theory of "segment structure conditions" is not offered at this stage, however. As pointed out in Den Dikken (1987), the alternations which are usually identified with ATR-harmony can be analyzed in terms of | a |-spreading

The Geometry of Vocalic Features

87

in a standard DP-feature framework. This is also true within the present framework. The array of possibilities we have for representing vowels along the dimension of height allows us to characterize two sets of four pairs. In one set the members of each pair differ in terms of | i|, whereas in the other set the difference involves | a|: (15)

| | i| vs. 0

I i I i |

Ovs.|a|

1 i

I i I a,i 1

1 i I a

I a I i |

1 1 a

a I a,i 1

a I a |—|

Given a tenvowel system, we therefore have, in principle, two ways of characterizing the ATR-alternations; I take the front/low vowels as examples: (16)

i I i /i/ i I i

i

-

A/ i I a,i

i a,i /e/ i

i I -

a /e/ i I a

I

a I i /a/ a

a

-

/a/ a I a

Ambiguity arises if we allow A / - / e / and /co/ - / o / to "switch" representations. It is not at all unexpected that A / and / e / can both get the same two representations (in different systems of course). The phonetic interpretations of | i — i,a| and | i| are close, due to the fact that the two dependents neutralize each other's effect. The presence of | i| implies a narrowing of the constriction, whereas the presence of | a| has the opposite effect due to the lowering of the jaw. This is also in accordance with the fact that it is very difficult to keep apart A / and / e / (as well as /co/ and / o / ) in ATR-systems. I would like to suggest that a ten-vowel system with root-controlled harmony is indeed phonologically ambiguous to some degree. However, I will also assume that due to certain characteristics, specific systems can always be "identified" as either | i | -spreading or | a | -spreading. For example, in so-called dominant harmony systems a decision can be made depending on which feature appears to be dominant. Hence in Kalenjin (Hall et al. 1974) or Tunen (Van der Hulst, Mous & Smith 1986a) dominant morphemes are [+ATR], indicating that | i| is the spreader. However, in Chukchee

88

Harry van der Hülst

or Nez Perce (cf. section 2.3), dominant morphemes have low vowels, implying that| a| is spreading. However, in root-controlled systems evidence might also be available pointing one way or the other, for example if contexts can be found in which for some reason or other the default value has to appear. Another interesting result arising from the above ambiguity is that languages may start to show signs of both harmonies independently. A number of such cases are discussed in Van der Hulst, Mous & Smith (1986b) and in Van der Hulst & Smith (1987); an example is also given in section 4.2.1. The typical situation seems to be that affix vowels show a three-way alternation, e.g. between / i / , / i / and / e / , where the second alternant results from | i|-spreading and the third from | a |-spreading. I will not go into these cases here. Rather, I would like to discuss another ATR-system, which is more complex than the tenvowel system discussed above, i.e. that of Kpokolo (discussed in Kaye, Lowenstamm & Vergnaud 1985). This system comes out as follows in the present feature system: (17)

/ i / / e / /¿/ / a / / u / / o / i i u u u u I I I I I I i a,i i a,i u,i a,u,i

| | I |

A / / e / / I / / a / /z/ /co/ i i u a u u I I I a a u

/d/ u I a,u

Observe that the central vowels / i / , / a / , / i / and /z/ are represented with a governing | u|, whereas back rounded vowels have |u — u | (cf. 1 la). The advanced counterpart of / a / in Kpokolo is / a / , which is also the regular counterpart of /z/. Apparently, then, the result of harmonizing / a / has to be merged with the harmonic congener of /z/. To deal with this, I will assume that, in Kpokolo, | a | cannot govern | i | . (In fact, it does not govern | u| either, so that we will say that it cannot govern | y|; this can be interpreted as a parametric choice. Cf. Van der Hulst forthc. b.) When | i| spreads to / a / an illformed representation will therefore arise. I assume (and again this has to be a parametric choice) that in such cases spreading is blocked unless a repair principle has been set in the grammar. In Kpokolo the repair principle has been set such that | a| is "demoted" to a dependent role, resulting in the "appearance" of | y | . The surface representation is derived by adding the governor | u | , which apparently acts as a default governor. Presumably, another possibility would have been to insert | i |; both possibilities lead to a neutralization of the harmonic congener of / a / with another underlying segment: (18)

a I i

-

y I a,i

-

u I a,i

The Geometry of Vocalic Features

89

An additional fact about Kpokolo is that rounded vowels alternate with central vowels. Given our featural representation, this involves the presence or absence of dependent | u|. At this point I close the discussion of ATR systems. Systems which have been analysed as involving [-ATR] will be discussed in section 2.3 and in section 4.1.2. 2.1.2. Palatal systems Prime examples of palatal systems are the Finugric languages, of which Finnish and Hungarian have been rather well studied. In this section, Finnish will be discussed (Campbell 1980, Anderson 1980, Skousen 1970). Hungarian, which has a limited labial harmony system alongside palatal harmony, is discussed in the next section. Given the phonetic interpretation of our features, it would seem that (19) gives the most straightforward representation of the Finnish vowel system: (19)

A / / e / / y / / u / / o / / o / /ae/ / a / i i i u i u a a I I I I I a u a,u a i

If harmony involves the spreading of dependent |i however, the representation of the front rounded vowels has to be different: (20)

/ i / / e / / y / / u / / o / / o / /a;/ / a / i i u u u u a a a

i

a,i

a

i

In Van der Hulst (1987, forthc. a), I adopt the representation in (20), which implies that there is no difference, in terms of the phonological analysis, between ATR-harmony and palatal harmony: both involve dependent | i | -spreading. It is probably the case, however, that palatal harmony is different phonetically in that the harmonic congeners differ more in their constriction location than in the position of the tongue root. Therefore, I would now like to propose that what differentiates the two systems is the inability of | u | to govern | i | in systems of the Finnish type. This implies that we have to accept that in such systems | i| spreads, both as a governor and as a dependent. Whether a spreading | i| comes out as governor or dependent on the target is determined by the governing restriction, i.e. spreading | i | to | u| will always result in | i —u|. In this respect, note that palatal harmony is also different from ATR-harmony in that |a| can govern | i|. Hence spreading | i| to / a / results in | a —i|.

90

Harry van der Hülst

In section 3.3 I will give additional support for analyzing! u| as a systematic dependent in palatal harmony systems. / i / and / e / do not have harmonic counterparts in the system; they are neutral. For this reason it is questionable whether these vowels should contain | i| in their representations, rather than being specified as | y| and |y —a|, respectively. This would in any event be the appropriate representation for those instances o f / i / and / e / that fail to initiate | i|-spreading themselves, in which case they act "transparently". | y | could be completed with | i | , the only governor which does not appear underlyingly, by a "complement" rule. The issue of transparency is discussed separately in section 5. As suggested in Ewen & Van der Hulst (1986), the addition of governors reflects a "complement principle" reminiscent of the complement rules of Radical Underspecification Theory. It may either be the case that the governor is not used underlyingly at all or that it is not used governing a particular other feature. The latter situation holds in Finnish, the former in the case of Chamorro (discussed below). If this is correct, then it will not be necessary to formulate explicitly a fill-in rule in such cases. In Kpokolo, where we also needed the addition of | u | , the situation is different, because in that case |u|-addition has the status of a language-specific repair strategy, which could also have inserted We also encounter palatal systems which have harmonic alternations of the following type: (21)

/i/

-

/u/

/e/

-

/o/

/at/

-

/a/

We find this for example in Chamorro (Poser 1982). At first sight we would like to assign the following representation to a six vowel system of this type: (22)

i

u

i I a

u I a

a I i

a

But, given (22), we cannot account for the harmonic pairs without a featurechanging operation. I therefore suggest (23) as the appropriate representation: (23)

i

y

i I a

y I a

a I i

a

Here | y| is completed by | u|, which is the governor not used underlyingly.

The Geometry of Vocalic Features

91

2.2. |i| and |u|-systems Various systems, especially in the Altaic family, have both palatal harmony and labial harmony. In Turkish, labial harmony is limited to high vowel targets. Due to the fact that, at least for high vowels, two harmonies operate together there is a four-way alternation: /i, y, i, u / (Clements & Sezer 1982). We must represent the vowel system as follows: (24)

/i/ /y/ /i/ /u/ /e/ /o/ /a/ /o/ i i y u a a a a I I I I i i,u u u

Thus, as in the case of Finnish, | u| is not a possible governor, / i / is not specified with either | i| or | u|, giving | y|, which is also the underlying representation of a high suffix vowel. Observe that | y| has to surface as such. Adding | i | or | u| would neutralize the contrast with the front unrounded and back rounded vowels. There is no "complement" feature, then, and no "repair rule". Turkish allows "the incomplete vowel". The fact that low vowels do not undergo | u|-spreading has to be accounted for in terms of a "negative condition" of some kind. We cannot, as in the binary framework, assign [-round] to these vowels, nor is there a basis for invoking a government restriction, since the two low unrounded vowels do have rounded counterparts in the vowel system. In sections 4.1.1 and 5 I will come back to the form and role of such conditions. Ultimately, of course, we would like to find a less stipulative account for the limitations on rounding harmony in Turkish, e.g. in terms of government restrictions, especially since this kind of restriction is recurrent (cf. section 3.2.3). Given that | u | is subjoined to | y |, which characterizes high vowels, it is to be expected that | u | associates less easily with low vowels. So, we could say that although | a | governs | u | in the underlying representations it does come to govern | u | by a spreading process. Another type of reduced rounding harmony occurs in Uygur and Hungarian (cf. Sezer & Wetzels 1986). In Uygur the following two vocalic alternations involving the high vowels exist: (25)

a.

sfx

y

u

stem

i,y,e,o,ae

u,o,a

b.

sfx

i

y

u

stem

i,e

y,o

u,o,a

The surface vowel system of Uygur is the same as that of Finnish (cf. 19). The three-way alternating suffix can be handled by including in the underlying inventory a vowel specified as | y|. Uygur does not allow this

92

Harry van der Hülst

bare | y| to surface and assigns | u| to it if neither | i| nor | u| spreads onto it; this is the case if an / a / precedes. | u|-addition is a languagespecific statement, as in Kpokolo. The high two-way alternating vowel has the | u| underlyingly. The underlying system of Uygur, then, is: (26)

/ i / / y / / i / / u / / e / / o / / o / /«,/ / a / i i y u i u u a a I I I I I u a a,i a i

Let us now turn to Hungarian. The vowel system is analyzed as follows (Van der Hulst 1985, Ringen 1988a): (27)

/i/ /y/ /u/ /e/ /o/ /o/ /a/ y i u y i u a u

a

a,u

a

Hungarian has palatal harmony and limited rounding harmony. The palatal harmony is straightforward and affects all vowels. It involves spreading of | i|. / i / is neutral and transparent; its representation is simply | y |, because the presence of | i | is predictable, / e / also shows transparent behaviour, but it is not neutral: / e / is the harmonic counterpart of / a / . Long and short / e / differ phonetically - short / e / is lower than long / e / - and, in conjunction with that, long / e / is practically always transparent, whereas short / e / typically is not. I will account for this as follows. I will assume that | a| can never govern | i| in long vowels. Spreading | i| to long | a| will therefore result in | i — a|. Most short / e / s have the G-relation reversed: | a —i|. These vowels, then, trigger palatal harmony. Only those instances of short / e / which are transparent get the representation in (27), which is also valid for all long underlying /e/s. As in Finnish, | y | will be completed by | i|. Hence, we build some abstractness into our analysis in that | y/i —-a| and | a —i| are mapped onto the same segment if associated with a single vowel slot. The amount of abstractness could actually be reduced if we assumed that the transparent / e / is represented as | a,y|, i.e. with the G-relation unspecified. We have not suggested this possibility earlier, but it is in fact the case that the G-relation is non-distinctive, and thus redundant, if there is just one series of mid vowels. I suggest, tentatively, that we make use of the possibility that the G-relation can (or must?) be left unspecified. Rounding harmony only involves mid (short) vowels. I will now discuss

The Geometry of Vocalic Features

93

three alternation types which are particularly fascinating because each involves the low vowel / a / . The following vocalic alternations are involved: (28)

a.

c.

sfx

e

a

b.

stem

i,e,y,o u,o,a i,e (t.)

sfx

e

o

o

stem

i,e

y,o

u,o,a i,e (t.)

sfx i,e

y ,6

a u,o i,e (t.)

("t." means "transparent"; transparent vowels do not trigger | i|-spread; cf. above.) The interesting challenge is to represent the difference between the above 2, 3 and 4-way alternating suffixes. A two-way (short) alternating suffix vowel is derived in (29): (29)

a

/e/ after /i,e,y,o/ (| i |-spreading)

/a/ elsewhere Since two-way alternators do not undergo rounding harmony, they have to be provided with a negative condition. (All high vowels have this condition too.) To derive three-way alternators, I allow an incomplete feature structure which cannot surface as such and which is minimally distinct from the representation of / a / . This has to be | y —a|. A repair rule filling in | u| will prevent | y —a| from surfacing (i.e. after / a / , from which | i| nor | u| spreads). The careful reader will object that | y —a| is already in "use" as the underlying representation of transparent / e / . Interestingly, however, all three-way alternators are short and, in addition, transparent short /e/s, rare as they are, never occur in suffixes. Suffixal / e / is always involved in one of the alternation types discussed here. We will have to say, then, that the repair rule filling in | u| only applies to suffix vowels, bleeding the rule which fills in | i| in the transparent short / e / s occurring in stems. The three-way alternation, then, is derived as in (30):

94

Harry van der Hülst

(30)

a

/e/ after /i,e/ (| i| spreading) /o/ after / y , o / (| i| and | u| spreading)

/o/ after / u , o / after / a / (| u| spreading) (| u| by default) Hungarian also has suffixes showing the four-way alternation e / o / a / o . In this analysis the relevant vowel is represented as | a| without the negative condition. Hence the four-way alternation is derived as follows: (31)

a

/e/ after /i,e/ (| i| spreading) /o/ after / y , o / (| i| and | u| spreading)

/o/ after / u , o / a (| u| spreading) In (30) and (31) I have assumed that the spreading of | i| and | u|, as well as default | u | assignment, are "structure preserving" in the sense that the resulting feature structures are the same as the underlying segments in terms of what governs what. Of course many aspects of the Hungarian vowel harmony system have been left unconsidered here. Yet, I believe that the above captures the basics of the system. A much more extensive analysis of the suffixes in Hungarian, using an i/u/a-system without dependency, is offered in Kornai (1987). 2.3. | a | -systems In Van der Hulst & Den Dikken (1987), it is shown that the harmony systems of Nez Perce, Chukchee and Middle Korean can be understood in terms of |a|-spreading. Consider Nez Perce (Zwicky 1971, Hall & Hall

The Geometry of Vocalic Features

95

1980). The harmonic relations have been indicated by arrows. The lowering is conditioned by the low vowels / a / and / o / : (32)

/ i / /ae/ / a / / o / / u / y a a u u

The vowel /ae/ is represented as | a|. A similar analysis, with a completely empty specification for /ae/, is proposed in Anderson & Durand (1988). I do not reject the option of leaving one of the vowels incomplete, but I see no purpose for it here. The representation of / i / , which has no harmonic counterpart, deserves some attention. Some morphemes with / i / trigger lowering. These should be specified as |y—-a|. Interestingly, the other / i / s are transparent to | a |-spreading, which suggests that they are compatible with the spreading value. I will come back to the representation of / i / in section 5. Middle Korean is highly similar, but has central vowels (Hayata 1975, Kim-Renaud 1986): (33)

/i/ /a/ /i/ /e/ /a/ /o/ /u/ y u u a a u u I I I I a a a,u u l l i [ I •

As in Nez Perce, the vowels specified with | a| are both the triggers and the "output" of harmonic spreading of | a |. Again / i / has no counterpart, and it appears to be opaque this time (cf. section 5). Let us now consider Chukchee, which has the most complicated system (Krause 1980, Kenstowicz 1983): (34)

/ i / [e] i i i I

i.a i

- / e / [ae] / e / / a / / o / / a / [o] / u / i i a a a a u u I a i u a a I * 4 i

Three remarks are in order. First, note that harmony is not structure preserving, in the sense that the derived vowels, between square brackets, are different from all underlying vowels of Chukchee. Secondly, some "schwas" appear to be non-triggers and transparent. We could represent

96

Harry van der Hülst

them as plain | y | , which becomes | a — y| by a language-specific default rule. This representation is equivalent to plain | a| on the assumption that a dependent incomplete | y| literally means nothing. In Coeur d'Alene we also find | a |-harmony. The following is based on the analysis presented in Cole (1987), who uses standard SPE features. An interesting aspect of this harmony, ignored here, is that a leftward | a |-spread is triggered by faucal consonants. It brings about the following shifts: (35a)

/i/

/u/ I

/a:/

/o/ f

/a/

I i

I I

(a) (b) (c) (d)

i 1

If all alternations are seen in terms of | a |-spreading, we have a problem in cases (c) and (d). To solve this I will postulate a double underlying source for / i / (as Cole does): | i | and | y | ; the latter is later completed as | i|. In addition we have two representations for both / a / and /ae/, one underlying and one derived: (35b)

/ i / [ae] = /ae/ / a / i i a a I I I i_J

LJ'1

=

[a] / i / / u / / o / a y u u I

- A i»

u o u o

C0u CDo C0o C0u

The following examples (from Newman 1944) illustrate harmonic alternations in suffixes. The effects of two postlexical rules, Lowering and Shortening, have been suppressed in these forms in the interest of clarity; capital letters stand for retroflex consonants; page numbers refer to Newman (1944).

138 R. Armin Mester (29)

a.

b.

d.

Aorist ?ugun-h«n c'uum-hwn

'drank' 'devoured'

151 122

t'uy-hwn duyduy-hwn duulul-hun ?ohyoo-hin c'ow-h/n lihim-hin Sil'-hm ninii-hm xayaa-h/n wan-h/'n caw-h/n ?agay-h/n tan-h/'n yawal-h/'n

'shot' 'stung rep.' 'climbed' 'search' 'touched' 'ran' 'saw' 'became quiet' 'placed' 'gave' 'shouted' 'pulled' 'went' 'followed'

163 122 122 163 164 137 145 122 166 166 135 134 134 122

120 ( - [ o o ] by Lowering)

di?S-al xat-al

'might pull out' 'might take the scent' 'might make' 'might eat'

120 120 120

Imperative t'uy-k'a yolow-k'o xoSxoS-k'o t'oyix-k'a lan-k'a

'shoot!' 'assemble!' 'rub repeatedly!' 'give medicine' 'hear!'

118 118 118 118 118

Passive Verbal Noun lox-honoo 'being eaten' gob-honoo 'being taken in' luk'ul-hanaa 'being buried' hudhud-hanaa 'being repeatedly recognized'

149 22 149 149

Dubitative Suug-al hoTn-ol

*[o] by Lowering and Shortening) *[oo] by Lowering)

*[ee] by Lowering)

The interesting point here is that rounding disharmony is permitted if the vowels involved differ in height, but not if they agree in height. The analysis in dependent ordering terms is quite clear: two tiers in the vowel

Dependent Tier Ordering and the OCP

139

feature plane, occupied by the features [high] and [round], are aligned as in (30) (following Archangeli 1985). (30)

[round] [high] v

The [round] tier is dependent on the [high] tier and has no immediate access to the skeleton, all connections are mediated through the [high] tier. I will assume, following Kuroda (1967) and in particular Archangeli (1984, 1985), that suffix vowels in Yokuts are unspecified for rounding. The unmarked value [-round] is assigned by default to all vowels still unspecified for roundness at the end of the derivation. The only way in which vowels underlyingly unspecified for [round] can receive the specification [+round] is through the action of Tier Conflation. Let us consider one of the examples given in (29) above, the aorist form t'uy-hun 'shot'. The underlying representation after affixation of the nonfuture suffix is (31), with the root vowel specified as [+round], but without roundingspeciflcation on the suffix vowel. (31)

[+rd] [+hi]

t+hi]

t'uy +

h in

Through the action of (morphemic) Tier Conflation, the vocalic melodies of root and suffix come to occupy the same tier. McCarthy (1986) suggests that, as part of Tier Conflation, heteromorphemic adjacent identicals are fused. This has interesting consequences for a melody plane with dependent tiers. Consider again the representation in (31): both root and suffix vowel are [+high], and fusing the two [+high] autosegments into one will result in the representation given in (32). (32)

[+rd] t+hi] t'u y h u n

140 R. Armin Mester Although only the feature [+high] is directly involved in the fusion effected by Tier Conflation, the hierarchical tier structure has the consequence that the second vowel is derivatively also specified as [+round] in this case. The form hoTn-ol 'might take the scent' (from underlying /hoTn-al/) is derived in an entirely parallel fashion, the only difference being that [-high] is fused instead of [+high]. (33)

[+rd] I [-hi] [-hi] I I hoTn+al

[+rd] I [-hi]

>

hoTno1

The crucial advantage of this approach over spreading analyses is found in the case of vowels which differ in height, as in c'ow-hirt 'touched' (34a) and t'uy-k'a 'shoot!' (34b). The representations of these forms do not induce any fusion on the [high] tier when Tier Conflation takes place, because the two height values are different. (34)

a.

[+rd] [-hi] I c' o w

b. [+hi] I h in

[+rd] [+hi][-hi] I I t'uyka

Since in this analysis rounding harmony is a derivative effect of the fusion of identical [high] autosegments by Tier Conflation and not a process in itself, we have an immediate explanation for its absence in vowel sequences which differ in height. For a form like c'owhin the only respect in which underlying and surface representation differ is due to the operation of default [-round] insertion, which specifies the suffix vowels as unrounded (see (35)). As a result, 'rounding disharmony' is permitted in vowels differing in height. (35)

[+rd] I [-hi] I c' o w

[-rd] I [+hi] I h in

- b y default

Dependent Tier Ordering and the OCP

141

2.2. Theoretical Consequences As a first result we note that there is no special rule of rounding harmony in Yokuts which would have to be constrained so as to spread [+round] in a height-stratified manner, i.e. from [+high] to [+high] and from [-high] to [-high]. Rather, Yokuts is characterized by a particular tier structure, with a more central height tier and a tier for the feature [round] which is anchored in the height tier. Tier Conflation fuses identicals on the two tiers. Since only [+round] is underlyingly specified and since suffixes are obligatorily unspecified for rounding, the crucial action of Tier Conflation takes place on the central tier, where the feature [high] is represented. Given the tier arrangement with rounding dependent on height, this results in derivative rounding harmony effects. The precise nature of the harmony effects follows from the fact that Tier Conflation only fuses adjacent identicals on the height tier, therefore rounding harmony, as a derivative effect of fusion, will only appear between vowels of like height. Looking beyond Yokuts, we see that this approach provides a principled explanation for the existence of height-stratified harmony systems and the absence of harmony systems governed by height polarity (where only vowels opposite in height would show harmonic interaction). If the height-stratified character of Yokuts harmony were encoded in a special harmony rule stipulating height identity by means of a coefficient variable, we could just as well imagine a rounding harmony rule spreading rounding from [a high] to [ - a high]. Such systems have not been reported, and the tier structure approach pursued here provides a principled explanation for this. It would be logically possible to encode such a restriction in Universal Grammar by proscribing the polarity use of the a-notation. This would still be an inferior theory, one which needs a special exclusion clause where the theory based on dependent tier ordering and Tier Conflation makes the correct predictions from the start. Furthermore we can tighten phonological theory by disallowing the use of descriptively powerful coefficient variables in (language-particular) phonological rules. Since Tier Conflation (ultimately, the OCP) detects occurrences of adjacent identical autosegments on a tier, no variable over feature coefficients is necessary to capture the fact that rounding is transmitted only from [a high] to [a high] (earlier analyses had to have recourse to such a device, see e.g. Archangeli (1985:350)). As has been argued by a number of researchers in Nonlinear Phonology (Halle and Vergnaud 1980, Goldsmith 1981, Steriade 1982, Hayes 1986, and others), assimilation processes are best understood as involving autosegmental spreading and not copying of features, so their statement requires no anotation. In the same vein, the analyses of Yokuts Rounding Harmony and of Ainu Backness Dissimilation (section 1.2) rely on operative principles

142

R. Armin Mester

which are intrinsic to the autosegmental framework (tier ordering, Tier Conflation, OCP) and do not make use of the powerful algebraic device of coefficient variables.

NOTES •I would like to thank Mark Feinstein, Junko ltd, John McCarthy, Alan Prince, Lisa Selkirk, and the participants in a phonology seminar at UT Austin for helpful comments and criticisms. 1. It should be noted that the most central feature tier is linked to a melodic core (or root node, in terms of Clements (1985) represented by lower-case c and v, which is in turn linked to a prosodic skeleton in the sense of McCarthy and Prince (forthcoming). 2. Segment-internal dependency relations have also been posited, in a different way and with different motivations, in the framework of Dependency Phonology (see e.g. Anderson and Jones 1974,77 and Ewen 1982) and in the more recent developments of Feature Geometry (see e.g. Clements 1985 and Sagey 1986). 3. While the mid and high vowels receive the feature [-low] by redundancy rule, I am assuming that they are both lexically specified for the feature [high]. This assumption is necessary for the analysis presented below. 4. The importance of this fact for an explanatory account of such cooccurrence restrictions was pointed out by ltd (1984) (see Prince (1984) for general phonotactic considerations along similar lines). 5. The distribution of final glides in roots is restricted by a dissimilation requirement exactly parallel to the one governing the choice of suffixal vowels, see ltd (1984) for discussion. 6. There is one local assimilation process causing labials adjacent across morpheme boundary to agree in velarization (as determined by the second labial), see ltd (1986) for discussion. 7. I am assuming that plain labials are underspecified for backness and also that nonlabials do not carry the feature specification [-labial]. Since the positive feature values are the only ones present in morpheme structure, I will write e.g. [labial] instead of [+labial]. For recent proposals on Underspecification theory see Kiparsky (1982) and Archangeli (1984).

REFERENCES Anderson, J. M. and C. Jones. 1974. "Three Theses concerning Phonological Representations," Journal of Linguistics 10, 1-26. Anderson, J. M. and C. Jones. 1977. Phonological Structure and the History of English, NorthHolland Publishing Company, Amsterdam. Archangeli, D. 1984. Underspecification in Yawelmani Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Mass. Archangeli, D. 1985. "Yokuts Harmony: Evidence for Coplanar Representation in Nonlinear Phonology," Linguistic Inquiry 16, 335-72. Burssens, A. 1969. Problemen en Inventarisatie van de Verbale Strukturen in het Dho Alur (Noordoost-Kongo), Brussels. Chiri, M. 1952. "Ainugo ni okeru boin chowa (Vowel Harmony in the Ainu Language)," Hokkaido University Bulletin in Humanities 1, 101-18. Reprinted in Chiri Mashio Zenshuu [Collected Works of Mashio Chiri], Vol. 4, Heibonsha, Tokyo (1984) 201-225. Chomsky, N. and M. Halle. 1968. The Sound Pattern of English, Harper and Row, New York.

Dependent Tier Ordering and the OCP

143

Churma, D. 1984. "On Explaining Morpheme Structure," Ohio State University Working Papers in Linguistics 29, 12-29. Clements, G. N. 1982. "A Remark on the Elsewhere Condition," Linguistic Inquiry 13, 6825. Clements, G. N. 1985. "The Geometry of Phonological Features," in Phonology Yearbook 2, 225-52. Ewen, C. J. 1982. "The Phonological Representation of the Welsh Mutations," in J. Anderson, ed., Language Form and Linguistic Variation, Benjamins B.V., Amsterdam, 75-95. Goldsmith, J. 1976. Autosegmental Phonology, Doctoral dissertation, MIT, Cambridge, Massachusetts. [Distributed by Indiana University Linguistics Club, Bloomington.] Goldsmith, J. 1981. "Subsegmentals in Spanish Phonology: An Autosegmental Approach," in W. W. Cressey and D. J. Napoli, eds., Linguistic Studies in the Romance Languages 9, Georgetown University Press, Washington, D.C. Halle, M. and J.-R. Vergnaud. 1980. "Three Dimensional Phonology," Linguistic Research 1, 83-105. Hayes, B. 1986. "Assimilation as Spreading in Toba Batak," Linguistic Inquiry 17, 46799. Itö, J. 1984. "Melodic Dissimilation in Ainu," Linguisitic Inquiry 15, 505-13. Itö, J. 1986. Syllable Theory in Prosodic Phonology, Doctoral dissertation, UMass/Amherst. [Distributed by GLSA, Dept. of Linguistics, UMass/Amherst; (forthcoming) Garland Publ.] Kenstowicz, M. and C. Kisseberth. 1979. Generative Phonology, Academic Press, New York. Kiparsky, P. 1982. "Lexical Phonology and Morphology," in I.-S. Yang, ed., Linguistics in the Morning Calm, Linguistic Society of Korea, Hanshin, Seoul, 3-91. Kuroda, S.-Y. 1967. Yawelmani Phonology, MIT Press, Cambridge, Massachusetts. Leben, W. 1973. Suprasegmental Phonology, Doctoral dissertation, MIT, Cambridge, Massachusetts. [Published 1979 by Garland Press, New York. ] McCarthy, J.J. 1979. Formal Problems in Semitic Phonology and Morphology, Doctoral dissertation, MIT, Cambridge, Massachusetts. McCarthy, J.J. 1981. "A Prosodic Theory of Nonconcatenative Morphology," Linguistic Inquiry 12, 373-413. McCarthy, J.J. 1983. "Consonant Morphology in the Chaha Verb," in M. Barlow, D. Flickinger, and M. Wescoat, eds., Proceedings of the West Coast Conference on Formal Linguistics, Stanford Linguistics Association, Stanford, California. McCarthy, J.J. 1985. "Features and Tiers: Semitic Root Structure Constraints Revisited," talk delivered at the University of Illinois at Urbana, Oct. 1985. McCarthy, J.J. 1986. "OCP Effects: Gemination and Antigemination," Linguistic Inquiry 17, 207-63. McCarthy, J.J. and A. Prince (forthcoming) Prosodic Morphology, MIT Press. Mester, R. Armin. 1986. Studies in Tier Structure, Doctoral dissertation, UMass/Amherst. [Distributed by GLSA, Dept. of Linguistics, UMass/Amherst; (forthcoming) Garland Publ.] Newman, S. 1944. Yokuts Language of California, Viking Fund Publications in Anthropology No. 2, New York. Prince, A. 1984. "Phonology with Tiers," in M. Aronoff and R. Oehrle, eds., Language Sound Structure, MIT Press, Cambridge, Massachusetts, 234-244. Prince, A. 1987. "Planes and Copying," Linguistic Inquiry 18,491-509. Rehg, K. L. and D. G. Sohl. 1981. Ponapean Reference Grammar, The University Press of Hawaii, Honolulu, Hawaii. Sagey, E. 1986. The Representation of Features and Relations in non-Linear Phonology. Doctoral dissertation, MIT. Steriade, D. 1982. Greek Prosodies and the Nature of Syllabification, Doctoral dissertation, MIT, Cambridge, Massachusetts.

144 R. Armin Mester Thomas, J. M. C. 1963. Le parler ngbaka de Bokanga: Phonologie, morphologie, syntax. Mouton and Co, Paris and The Hague. Thomas, J. M. C. 1970. Contes Ngbaka - Ma'bo. Editions Klincksieck, Paris. Tucker, A.N. 1969. Review of Burssens, A. 1969. Problemen en inventarisatie van de verbale strukturen in het dho alur (Noordoost-Kongo), Brussels, Journal of African Languages 8, 125-6. Wescott, R. W. 1965. Review of Thomas. (1963), Language 41, 346-7. Younes, R. 1983. "The Representation of Geminate Consonants," ms., University of Texas, Austin.

Underspecification Theory and Binary Features Catherine O. Ringen University of Iowa

1. INTRODUCTION

Underspecification theory, as developed in the recent work of Kiparsky (1981, 1982a, 1982b), Archangeli (1984), and Pulleyblank (1986a) provides a framework which appears to overcome many of the problems encountered by earlier autosegmental approaches to vowel harmony. This paper examines vowel harmony involving the feature Advanced Tongue Root ([ATR]) in two languages, Kalenjin and Igbo. It is shown that certain non-alternating suffixes in these languages pose problems for underspecification theory. A solution which makes use of [+ATR], [-ATR], and [ATR] (unspecified for [ATR]) in lexical representations is sketched. It is suggested that such an analysis does not entail abandoning the claim that features are binary.1

2. UNSPECIFIED VALUES

Until recently most phonologists believed that if phonological features are to be binary rather than ternary, unspecified values for features must be disallowed. Obviously, given some feature F, if representations contain [+F], [-F], and [ F] (unspecified for F), and if it is possible to refer in rules to [ F] (unspecified for F), as well as to [+F] and [-F], the system has three v a l u e s , ' + ' , a n d ' ' (unspecified). Lightner (1963) and Stanley (1967) have shown that even if reference to the unspecified value is not permitted, it is possible for an unspecified value to function as a third value, distinct from ' + ' or ' - ' . The argument is as follows: Given a rule R, [+A] — [+B] and a representation unspecified for a feature mentioned in the structural description of the rule, e.g., [_b], it must be decided whether R applies to the representation or not. The assumption that such a rule does not apply to such a representation has been called the submatrix convention of rule application; the assumption that R does apply to such a representation has been called the distinctness convention of rule application (Stanley 1967:16-17; Chomsky and Halle 1968:382). Under either

146

Catherine O. Ringen

the submatrix or distinctness convention of rule application, blanks can function as a third value, distinct from both ' + ' and '-'. Specifically, consider the segments: (1)

"+a b c

"+a~ -b c

"+a" +b c

and the rules (2)

(i)

[-b]

(Ü) (iii) (iv)

[+a] [+b] [-c]

— —

—

"-a +c_ -c "+c -b:

and assume the submatrix convention of rule application. If the rules of (2) are applied in the order given, the following derivations result: (3)

"+a" b c.

-fa" +b _ c.

(ii) J . ~+a~ b

(Ü) I >a~ +b

(iv)

(iii)

"+a" -b

~+a~ +b +c

"+a" -b c_ (i) * "-a" -b +c_

Since the first segment emerges as distinct from both the second and third segments, and since the only difference between the three segments is their specification for one feature, the feature [b], the unspecified value of [b] has functioned as a third value, distinct from both plus and minus. Alternatively, consider the distinctness convention of rule application, the same underlying segments, and the rules

Underspecification Theory and Binary Features (V)

[-b]

—

(vi) (vii)

[+b] [-a]

—

147

[+c] [+b]

Again, initially non-distinct segments can be rendered distinct by applying the rules in (4) in the order given: +a b c_ 1 1 (V) I -a b -c_

|

(vi) i -a" b

|

+a +b c_

+a -b c_

(vi) 1 +a +b +c

(V) 1 -a -b _—c_ 1 (vii) 1 "-a" +b _-c_

|

(vii) J "-a" +b +c If such derivations are allowed, the argument goes, the result is specious simplification, since an unspecified value is actually functioning as a value distinct from both ' + ' and Therefore, some way of preventing this possibility must be devised. Until quite recently phonologists assumed, following Stanley, that the only viable way to prevent an unspecified value from functioning as a third value was to disallow representations with unspecified feature values altogether. Various ways of allowing unspecified values without having ternary features have been suggested. For example Ringen (1975) points out that blanks do not function as distinct from + and - in the examples of Stanley and Lightner if the rules are not ordered or if feature changing applications to non-derived forms are prohibited. More recently, Archangeli (1984) and Pulleyblank (1986a) have

148

Catherine O. Ringen

argued that it is possible to maintain the position that features are binary in underspecification theory, which allows (requires) that phonological rules apply to representations which are not fully specified. They accomplish this by adopting a set of constraints to guarantee that at no point in the derivation are representations with [ F], [-F], and [+F] available. Rather a derivation is divided into two parts: first [ F] and [+F] (or [-F]) occur in representations and second only [+F] and [-F] occur (see Pulleyblank 1986a). In this way the binarity of features is preserved while unspecified representations are permitted. Consider now how an asymmetric vowel harmony system can be analyzed within the framework of underspeciflcation theory.

3. KALENJIN 2

Kalenjin, a southern Nilotic language, has asymmetric or dominantrecessive vowel harmony. 3 The vowels of Kalenjin can be divided into two sets, those that are [+ATR] ([+Advanced Tongue Root]) and those that are [-ATR]. There are no neutral vowels: (6)

[+ATR]

[-ATR]

I

U

I

e

o

e

D

CO

o a

Morphemes fall into two classes, those that alternate and those that do not. For the most part, morphemes that do not alternate have [+ATR] vowels, whereas alternating morphemes have either [+ATR] or [-ATR] vowels. When a word is made up entirely of alternating morphemes, the vowels of the word are [-ATR]. When a non-alternating morpheme with [+ATR] vowels occurs anywhere in a word, whether prefix, suffix or root, all vowels in the word are [+ATR], The example in (7a) is made up entirely of alternating morphemes and all vowels are [-ATR]: (7)

a.

kidistant past

a - ger I shut

0 3rd sg. obj.

b.

ki distant past

o - ge:r - in I see you(sg.) obj.

'I shut it'

4

I saw you'

Underspecification Theory and Binary Features c.

ki o - ger e distant I shut noncompast pletive

149 0 'I was shutting it' 3rd sg. obj.

The form in (7b), in contrast, contains the non-alternating root 'see' and all vowels are [+ATR]. The form in (7c) contains the non-alternating noncompletive suffix -e and all vowels in the word are [+ATR]. The Kalenjin data can be analyzed within the framework of underspecification theory by assuming that non-alternating morphemes have [+ATR] autosegments while alternating morphemes are unspecified for the ATR feature. Vowel harmony can be assumed to be a rule that spreads [+ATR] bidirectionally: (8)

Vowel Harmony [+ATR] V

V

V

Vowels which remain unspecified for [ATR] are specified as [-ATR] by a Redundancy Rule which applies as late as possible in a derivation:4 (9)

Redundancy Rule [] -

[-ATR]

Thus, the underlying representations for ge:r 'see' and -e, the noncompletive marker, would have [+ATR] autosegments, the forms ger/ger 'shut', ki/ ki 'dist. past., and a/n 'I' would be unspecified for [+ATR] (capital letters represent segments not specified as [+ATR]): (10) gEr 'shut'

[+ATR]

[+ATR]

gE:r 'see'

E noncompletive

Derivations would proceed as in (ll): 5

ki 'distant' past

A 'I'

150

Catherine O. Ringen

(11) a.

[-ATR] kl

- A - gEr

Default — ki

- a - ger6

[ATR] b.

kl

- A - gE:r

[+ATR] - In

UAC VH

— ki

- o - ge:r

[+ATR]

[+ATR] c.

kl

- A - gEr

- in

UAC VH

- E

ki

-

D

- ger

I am assuming that initial association is accomplished by the Universal Association Convention (UAC) on the first cycle and that Vowel Harmony applies on subsequent cycles to associate the remaining vowels. There are three non-alternating morphemes with [-ATR] vowels that never have [+ATR] vowels, even in a word with [+ATR] vowels. These are the negative prefix ma - the perfectivizer ka ~ ga, and the reflexive suffix ke: ~ -ge:. Examples are given in (12):

(12) a.

ki distant past

-

D

un

b.

ma neg.

-

c.

ka - ma recent neg. past

I

wash

ti imper.

un - ge: wash self D

I

- ge:

'I washed myself

self

-

ge:r see

'don't wash yourself ok you (pi.) obj.

'I didn't see you (pi.)'

Underspecification Theory and Binary Features d.

ka

- ma

recent past

- ga

neg.

- go

perfeetivizer

151

- ge:r

sequential def. 3rd subj.

see

-

'and he hadn't me seen me' obj.

D

The forms in (12c) and (12d) show that not only do the non-alternating morphemes fail to become [+ATR], they also block the spread of [+ATR] to the regularly alternating suffix vowels to their left. In earlier autosegmental accounts of vowel harmony, such opaque vowels were analyzed as having lexically associated [-ATR] autosegments: (13)

a.

[+ATR] kl

b.

-

A

-

Un

[-ATR] kA

-

ma

t-ATR] -

ge:

'I washed myself

[+ATR] -

A

-

gE:r

-

Ak

'I didn't see you' (pl-)'

On such an analysis, [+ATR] would spread until blocked by a lexically associated [-ATR] vowel. However, such a treatment is not possible if, as assumed by Archangeli and Pulleyblank, [+ATR], [-ATR] and [ ATR] cannot all occur in lexical representations.7 How then are these forms to be handled?8 One possibility, the one that will be defended here, is to relax the requirement that only one feature value may be present in lexical representations and to allow the exceptional affixes to have (floating) [-ATR] autosegments. A second is to claim that the morphological structure of the words containing the exceptional morphemes accounts for the failure of these morphemes to undergo harmony (e.g. they are syntactically derived clitics and hence not present when spreading applies). A third possibility is to assume some sort of exception marker which prevents them from being associated with [+ATR] by the VH rule (i.e. they are marked as extra-harmonic or with a negative rule feature, [-VH]). The rule feature solution seems to be the least desirable alternative. One of the advantages of the autosegmental treatment of vowel harmony has been claimed to be that non-alternating segments which block the spread

152

Catherine O. Ringen

of harmony can be handled with devices provided by the autosegmental notation without the necessity of diacritic (rule exception) features (Clements and Sezer 1982). The second possibility, that the blocking is explained by the morphological structure of the words containing the exceptional morphemes, has been suggested by Hall et al. (1974). Such an explanation has some support in the case of -ke, which also fails to trigger the otherwise general rule of nasal assimilation (kipconge:). In the case of the other suffixes, however, no such analysis seems possible. Since regularly alternating affixes occur following these suffixes, it is not possible to claim that these suffixes are derived in the syntactic component and VH applies only lexically. Marking these suffixes as extra-harmonic does not appear to be a possibility either. The non-alternating suffixes do not undergo vowel harmony whether peripheral or not. Pulleyblank (1986a) argues convincingly that certain clitics in Tiv are marked as extra-tonal. Crucially, however, these clitics behave differently depending on whether they are peripheral or not, losing their extra-tonality when not peripheral (i.e. when followed by another affix). Unlike the Tiv clitics, the Kalenjin suffixes behave identically whether peripheral or not. A treatment claiming that the problematic suffixes in Kalenjin are extra-harmonic, similar to Pulleyblank's extra-tonal clitics in Tiv, would be unable to explain why these suffixes fail to undergo harmony when not peripheral. 9 1 am suggesting, then, that the derivation of (12c) would proceed as follows: (14)

cycle 1

[+ATR] by UAC [ge:r]

cycle 2

[+ATR] \

\

by VH \

[[ge:r] ok]] cycle 3

[+ATR] VH [D

cycle 4

[ge:r ok]]

[-ATR]

[+ATR] by UAC

[ma

[D

ge:r

Dk]]

Underspeciflcation Theory and Binary Features cycle 5

153

[-ATR] [+ATR] no rule applicable [kA

post cyclic10

[ma

[-ATR]

d

ge:r

ok]]

[-ATR]

[ka

ma

[+ATR] ge:r

o

ok]

4. IGBO

Consider next the cases of Igbo which also has [+ATR] harmony. The vowels of Igbo can be divided into two sets:11 (15)

-ATR

+ATR

to

In Igbo the harmonic quality of root vowels determines the harmonic quality of affix vowels as illustrated by the examples in (16)12 (roots are underlined): (16)

a.

a a e -

b.

- là - ta i - vu

-

zòì tà kè

- la - là - lè

'don't buy' 'don't eat' 'don't share'

'to marry' 'to eat' 'to carry'

ò b ò

(p. 25) (P- 1) (p. 25) l& ta vu

'manier' 'eater' 'carrier'

(p. 27)

In underspeciflcation theory these facts can be described if we assume a rule that spreads [+ATR], identical to the one proposed above for Kalenjin, (17)

Igbo Vowel Harmony [+ATR]

154

Catherine O. Ringen

and a Redundancy Rule, also identical to the Kalenjin rule, which assigns [-ATR], Unlike Kalenjin, Igbo does not have dominant-recessive or asymmetric harmony. Thus, as noted in Ringen (1979), an analysis involving the bidirectional spread of the feature [+ATR] feature depends crucially on there being no affixes with underlying [+ATR] vowels. If such an affix existed, it would (incorrectly) cause unspecified root vowels to become [+ATR] as happens in Kalenjin. In fact, Green and Igwe (1963) list two non-alternating affixes with [+ATR] vowels which they classify as suffixes: -be 'yet, ever' and -mo 'lest'. As argued in Ringen (1979), however, there is good evidence that these two morphemes are incorrectly classified as suffixes by Green and Igwe. Rather, they are base formatives (Welmers 1970) which form compound-like verb bases, but do not occur as independent verbs. Base formatives, like verbs in compounds, do not harmonize with the other compounding element. Here, it seems, the failure to harmonize can be explained by the morphological structure. As in Kalenjin, there are a number of [-ATR] suffixes in Igbo that do not alternate. These include the partitive suffix too, the distributive suffix si, and the emphatic suffix di. Examples are given in (18) (roots are underlined): (18) a.

nye ri vii

- too - tco - si

(m mnu)

'give me a little salt' 'eat something' 'carry' (distributive)

b.

aba anyi e - bho basket we pref. put load on head

- si distr.

(p. 62) (p. 85) (p. 90)

- ghi emphatic

ha then

'our not helping them to put the baskets on their heads' (p. 137) c.

abo basket

m I

bho put load on head

- we begin

-

ghi emphatic

Ekwe

'my not beginning to help Ekwe put the basket on his head' (p. 137)13 The example in (18a) shows that the exceptional suffix -si does not become [+ATR] when the root vowels are [+ATR], The example in (18b) shows that the vowels in suffixes following a non-alternating suffix are also [-ATR]; in other words the vowels in the non-alternating suffixes are

Underspecification Theory and Binary Features

155

opaque. The form in (18c) shows that the emphatic marker -ghi/ghi alternates regularly when preceded by a [+ATR] vowel.14 The opaque vowels in these non-alternating affixes can be handled if we assume they have (floating) [-ATR] autosegments as illustrated in (18):15 (19)

[-ATR]

[-ATR]

si

[-ATR]

tU

dl

A form such as k - bho - si - ghf in (18b) would be derived as follows: (20)

cycle 1

[+ATR] byUAC [bho] [+ATR]

cycle 2

by VH

/ /

/

/

[e [bho]] cycle 3

[+ATR]

/A [[e

cycle 4

bho]

[+ATR] [[e bho

post cyclically

[+ATR]

[-ATR] I I I

byUAC

si] [-ATR] si] [-ATR]

no rule applicable

ghl [-ATR] by default

e^blio

si

ghl

5. BINARY FEATURES

As noted above, it is possible to maintain that features are binary even if [+F], [-F], and [ F] occur in lexical representations. For example, if it is assumed, as proposed in Ringen (1977), that feature changing applications are permitted only to derived forms, and furthermore that

156

Catherine O. Ringen

no rule may refer to an unspecified value of some feature, then it is not possible to use a single feature to make a three way distinction as in the Lightner-Stanley example cited above. In the framework of lexical phonology, as developed in the recent work of Kiparsky, the requirement that feature (and structure) changing applications are permitted only to derived forms is imposed by the SCC.16 It might seem that it is necessary in underspecification theory to refer to the lack of specification (or association) of a feature. As far as I can tell, however, this is never actually necessary. For example, redundancy rules in underspecification theory are written as [ ] — [-ATR] where [ ] is taken to mean 'unspecified for ATR'. But the rule can be written as V — [-ATR] as long as feature (or structure) changing applications are permitted only to derived forms. Since redundancy rules never change features, an assumption which Archangeli (1984) explicitly adopts, this change seems unproblematic. The possibility of using an unspecified value as a third value arises if unspecified values remain in the post-lexical component, however, because here the SCC does not hold. As with the lexical rules and redundancy rules it seems that the only well-motivated rules either apply to non-derived forms and fill-in features (e.g. the Default and Redundancy Rules of Archangeli (1984)) or apply in derived environments (e.g. voicing assimilation in English contracted forms, Kiparsky (1982a)). It would seem, therefore, that if we assume that feature (or structure) changing applications are permitted only to derived forms in the post-lexical component, then the problem of unspecified values functioning as distinct from both ' + ' and does not arise.17 An alternative way of maintaining binarity is suggested in Pulleyblank (1986a). He shows that a constraint on rule application such as (21) below prevents unspecified values from functioning as third values, distinct from plus and minus: (21)

A rule must not refer to [aF] in its structural description before a redundancy rule assigns [aF]. 18

Returning to consider the segments in (1), either [+b] or [-b] will be assigned by a redundancy rule in underspecification theory. This means that either rule (i) or (iii) in (2) cannot apply until that redundancy rule has applied. But if this is so, then the three segments cannot emerge as distinct and the features will be binary, not ternary.

Underspecification Theory and Binary Features

157

6. CONCLUSION

It has been shown that the non-alternating morphemes in Kalenjin and Igbo which contain [ - A T R ] vowels are not easily described in underspecification theory as outlined in the recent work of Archangeli, Kiparsky, and Pulleyblank. It is suggested here that these non-alternating forms can be analyzed as having (floating) [-ATR], that vowel harmony in these languages spreads [+ATR], and that some vowels are unspecified for [ATR] in their underlying representations, but that this does not mean abandoning the claim that features are binary. 19 However, the analysis sketched here is inconsistent with Pulleyblank's (1986b) suggestion that if [aF] occurs in a lexical representation it triggers the application of a redundancy rule introducing [aF]. Therefore, if the analyses sketched here are adopted, Pulleyblank's (1986b) analysis of Okpe must be reconsidered. 20

NOTES * I am grateful to D. Archangeli, M. Hammond, and D. Pulleyblank for their helpful comments on an earlier version of this paper. Discussions of some of the issues and data with R. Chametzky, J. Goldsmith, K. Houlihan, G. Iverson, C. Kesterson, C. Ulrich, and V. Urkewich have also been useful. The usual disclaimers apply. 1. Some linguists have acknowledged that they adopt a ternary feature system. See for example Goldsmith (1985). 2. Data are from Hall et. al. (1974). 3. Aoki (1968) distinguishes between symmetric and asymmetric vowel harmony. In a symmetric system any vowel in a certain position can determine the quality of the vowels for the entire word, whereas in an asymmetric system, one set of vowels dominates the other so that the presence of a dominant vowel anywhere in the word changes the vowels of the non-dominant set. He classifies languages such as Finnish, Hungarian, Turkish and Igbo as having symmetric vowel harmony and languages such as Kalenjin and Diola-Fogny as having asymmetric harmony. As argued in Ringen (1975,1979) however, the same harmony rule operates in both Igbo and Diola-Fogny, suggesting that it is incorrect to classify these as distinct harmony types; the superficial differences result from a different distribution of [+ATR] vowels. Note that Halle and Vergnaud (1981) use the term "dominant harmony" to refer to bidirectional harmony. The only bidirectional harmony they discuss is Kalenjin, an asymmetric or "dominant-recessive" harmony. These are, however, bidirectional harmonies that are not dominant-recessive (e.g. Igbo) and hence their equation of "bidirectional" and "dominant" is confusing. See Ringen (1987) for discussion of Clements' (1977) and Halle and Vergnaud's (1981) treatments of asymmetric vowel harmony. 4. Archangeli (1984) proposes two types of redundancy rules, Complement and Default Rules. This distinction is not crucial here. 5. The analysis proposed here for Kalenjin is identical to the analysis of Diola-Fogny, another dominant recessive harmony system, proposed in Ringen (1975) except that the analysis of Diola-Fogny assumed a segmental rather than autosegmental rule of [+ATR] harmony.

158

Catherine O. Ringen

6. This representation assumes the OCP has applied. Nothing rests on this assumption. Association resulting from the UAC is indicated by a solid line. Association resulting from application of VH or the Default Rule is indicated by a dashed line. 7. Actually, the three specifications would be possible if they were in different environments. 8. After his paper was written I learned that Dresher (1985) also discusses the problem of these non-alternating Kalenjin morphemes for underspecification theory. 9. It might appear that this could be accomplished as follows. A non-alternating suffix is marked as extra-harmonic and hence the association of [+ATR] with the vowels of this suffix is blocked on the cycle in which the suffix is added. If on the next cycle another suffix is added, then the association of [+ATR] would be blocked, not by the extra-harmonic marking, which would be eliminated when the suffix in question ceased to be peripheral (see Pulleyblank 1986a: 208), but by the Strict Cycle Condition (SCC) in just the way that Archangeli and Pulleyblank (to appear) argue the SCC blocks the spread of harmony in Yoruba disharmonic roots. This solution will not work, however, because the vowels of the non-alternating suffix will loose their extra-harmonic markings when the suffix ceases to be peripheral (Peripherality Condition, Pulleyblank 1986a:208) and hence be eligible for the application of VH by the SCC because they would have become subject to VH on that cycle. 10. For clarity I have shown the representation after the Default Rule applies, ignoring the issue of whether the OCP would modify this representation. 11. Ladefoged (1964) shows that the basis of the distinction between the two sets of Igbo vowels is relative advancing of the tongue root. For further discussion of [ATR] see Halle and Stevens (1969), Stewart (1967), and Lindau et. al. (1972). J2. All examples are from Green and Igwe (1963); page numbers refer to that source. V = low tone, V = down-stepped high tone, V = high-low glide, and V (unmarked) = high tone. 13. Negation is indicated in these examples by a tone pattern which is distinct from that used in the affirmative. 14. It might seem that these data could equally well be analyzed by assuming that all affixes are unspecified and VH spreads [-ATR] rather than [+ATR] and that [+ATR] is supplied by redundancy rule. Such an analysis is not possible, however, because the (necessarily) bidirectional [-ATR] spreading rule would incorrectly predict that non-alternating [-ATR] suffixes cause preceding [+ATR] roots to become [-ATR]. 15. The alternative of special markings (extra-harmonic or rule feature) are rejected in Igbo for the same reasons as for Kalenjin. 16. Note, for example, that the SCC of Kiparsky (1985) would block the application rule (i) in the derivation in (3). It might seem that I have not avoided the possibility of ternary vaules because the representations in (1) could be assumed to be derived and then the application of the rules in (2) would not be blocked. If the forms are derived, then the problematic applications are not avoided, but this is beside the point because the Lightner-Stanley argument is about underlying, not derived segments. Moreover, it is unclear what the rules and underlying segments could have been to yield these derived segments. 17. The type of rule application excluded would be that of velar softening to non-alternating, non-derived forms such as /kankiv/ to derive [kansiv] 'conceive'. 18. Pulleyblank uses default rule, not redundancy rule in his formulation of the constraint (p. 135). 19. If the suggestions of this paper are correct, it means that there is an alternative to the account of disharmonic loanwords proposed in Archangeli and Pulleyblank (to appear) and adopted in Ringen (1988). 20. It seems to me that Pulleyblank's (1986b) analysis of Okpe does not, contrary to his claims, depend on the assumption that the occurence [-low] in lexical representations triggers

Underspecifîcation Theory and Binary Features

159

the application of all redundancy rules inserting [-low]. Rather the correct ordering of the redundancy rules seems to follow from the Elsewhere Condition (Kiparsky 1973).

REFERENCES Anderson, S. R. and P. Kiparksy. 1973. A Festschrift for Morris Halle, N.Y.: Holt, Rinehart, and Winston. Aoki, H. 1968. "Toward a typology of vowel harmony," IJAL 34, 142-145. Archangeli, D. 1984. Underspeciflcation in Yawelmani phonology and morphology, PhD dissertation, MIT, to be published by Garland Publishing. Archangeli, D. and D. Pulleyblank (to appear) "Yoruba vowel harmony," LI. Chomsky, N. and M. Halle. 1968. The Sound Pattern of English , N.Y.: Harper and Row. Clements, G. 1977. "The autosegmental treatment of vowel harmony," Phonologica 1976, W. Dressier and O. Pfeiffer (1977), 111-119. Clements, G. and E. Sezer. 1982. "Vowel and consonant disharmony in Turkish," in Van der Hülst and Smith (1982), 213-255. Dresher, E. 1985. "Constraints on empty positions in tiered phonology," Cahiers linguistiques if Ottawa 14, 1-15. Dressler, W. and O. Pfeiffer (eds.). 1977. Phonological 1976, Innsbruck: Innsbrucker Beiträge zur Sprachwissenschaft. Dressier, W., H. C. Luschützky, O. E. Pfeiffer and J. Rennison (eds.). 1987. Phonologica 1984, Cambridge: Cambridge University Press. Goldsmith, J. 1985. "Vowel harmony in Khalkha Mongolian, Yaka, Finnish, and Hungarian," Phonology Yearbook 2, 251-275. Green, M. and G. Igwe. 1963. A Descriptive Grammar of Igbo, Berlin: Akademie-Verlag. Hall, B. L., R. M. R. Hall, M. P. Pam, A. Myers, S. Anteil, and G. Cherono. 1974. "African vowel harmony from the vantage point of Kalenjin," Afrika und Übersee 57, 241-267. Halle, M. and J.-R. Vergnaud. 1981. "Harmony processes," Crossing the Boundaries in Linguistics, W. Klein and W. Levelt (eds.), Dordrecht: Reidel, 1-22. Halle, M. and Stevens. 1969. "On the feature 'Advanced Tongue Root'," Quarterly Progress Report No. 94, Research Lab of Electronics, MIT, 209-215. Hülst, H. van der and N. Smith (eds.). 1982. The Structure of Phonological Representations, Vol. 2. Dordrecht: Foris. Kiparksy, P. 1973. "Elsewhere in Phonology," in S. R. Anderson and P. Kiparsky (1973), 93-106. Kiparsky, P. 1981. "Vowel harmony," MS. Kiparsky, P. 1982a. "Lexical morphology and phonology," in I.-S. Yang (1982), 3/91. Kiparsky, P. 1982b. "From cyclic to Lexical Phonology," in Van der Hülst and Smith. 1982,131-265. Kiparsky, P. 1985. "Some Consequences of Lexical Phonology," PY2, 85-138. Ladefoged, P. 1964. A Phonetic Study of West African Languages. West African Language Monographs, 1, Cambridge: The University Press. Lightner, T. 1963. "A note on the formulation of phonological rules," Quarterly Progress Report of the Research Laboratory of Electronics, MIT 68, 187-189. Lindau, M., L. Jacobson, and P. Ladefoged. 1972. "The feature Advanced Tongue Root," in Working Papers in Phonetics No. 22, Univ. of California, Los Angeles, 76-94. Pulleyblank, D. 1986a. Tone in Lexical Phonology, Dordrecht: Reidel. Pulleyblank, D. 1986b. "Underspeciflcation and low vowel harmony in Okpe," SAL 17, 119-153.

160

Catherine O. Ringen

Ringen, C. 1975. Vowel Harmony: Theoretical Implications, PhD dissertation, Indiana University, published by Garland Publishing, 1988. Ringen, C. 1977. "Vowel harmony: Implications for the Alternation Condition," in W. Dressier and O. Pfeiffer (1977), 127-132. Ringen, C. 1979. "Vowel harmony in Igbo and Diola Fogny," SAL 10, 247-259. Ringen, C. 1987. "Vowel harmony: Linear or non-linear?" Dressler et al. (1987), 247-251. Ringen, C. 1988. "Transparency in Hungarian vowel harmony," Phonology 5. Stanley, R. 1967. "Redundancy rules in phonology," Language, 41:2, 393-436. Stewart, J. M. 1967. "Tongue root position in Akan vowel harmony," Phonetica 16, 185204. Weimers, W. 1970. "The derivation of Igbo verb bases," SAL 1,49-59. Yang, I.-S. 1982. ed. Linguistics in the Early Morning Calm. Seoul: Hanshin.

Vowel harmony, rule formats and underspeciflcation: the dialect of Francavilla-Fontana Willebrord Sluyters University of Nijmegen 0. INTRODUCTION

Vowel harmony in the Apulian (Southern Italy) dialect of FrancavillaFontana has recently attracted attention because of the rather complex alternations resulting from it. In Calabrese (1985) an attempt was made to explain these harmony processes by way of underspeciflcation and a new rule format. The purpose of the present paper is to show that the rule format, although interesting in itself, is inadequate from a theoretical as well as from an empirical point of view. Furthermore, it will be argued that once a model of Lexical Phonology along the lines of Kiparsky (1985) has been assumed, the data can be derived by way of quite simple autosegmental spreading operations. Third, given the relation between vowel harmony and diphthongization in the dialect, we will investigate the nature and representation of diphthongs, a relatively unexplored field of research. The dialect referred to exhibits vowel harmony with respect to three features: [HIGH], [BACK] and [ROUND]. Height harmony is traditionally called "metaphony", a term which will be used here as well. The interaction between the harmonic features and the special status of diphthongs in this respect, a problem treated in Sezer and Wetzels (1986a,b) for Hungarian, Uygur and Yakut, is one of the topics of the present paper. A second problem is the strange behavior of the vowel [a] with respect to [+HIGH] harmony. In an underspeciflcation model one would expect the [a], being [+LOW] and therefore redundantly [-HIGH], to behave as the neutral vowel. However, underlying [a] is only in part neutral, and can be opaque as well. This state of affairs raises serious theoretical questions. A third interesting aspect of the dialect is the result of [+HIGH] harmony. The stressed vowels [e] and [o], in metaphonic environments, develop into their high counterparts [i] and [u]. Stressed [e] and [a] in harmonic environments give rise to the diphthongs [ie] and [ue]. Unstressed [e] and [a] on the other hand are raised to [i] and [u]. First, we must establish the exact conditions under which these vowels diphthongize. Next, the

162

Willebrord Sluyters

nature of diphthongs in the dialect has to be investigated, in particular whether they are mono- or bipositional, why the metaphonic diphthongs are rising, and how the second element [i] is to be accounted for ([ue] suggests a kind of epenthesis). In section 1, data on vowel harmony in Francavilla-Fontana will be presented. These data are taken from Calabrese (1985), Ribezzo (1911'12) and Rohlfs (1956-'61). We will formalize the relevant rules and investigate the environments of their application. Section 2 contains our analysis of metaphony, and especially of metaphonic diphthongization. It will be shown that diphthongs in the dialect have various sources, but are all subject to one constraint, which is partly responsible for [ie] and [ue] as metaphonic products of [e] and [o]. The analysis presented in Calabrese (1985) is shown to be inadequate with regard to several aspects of the diphthongs. Vowel harmony phenomena like metaphony are not merely restricted to this dialect, but appear in many variants of Italian. The data however, and therefore the exact rules responsible for them, differ very much, even between neighboring villages. On several occasions we will mention data from other dialects and we will indicate how our analysis might be extended to cover these too. For example, two dialects might have the same rule of metaphony, but different conditions on its application. By formalizing these differences we can arrive at a more general picture of the phenomenon in the Italian linguistic area as a whole.

1. DATA

The data on the vocalic system of the dialect are not unambiguous. As to the existence of the vowels [i], [e], [a], [o] and [u], all sources agree. Calabrese (1985) proposes surface [e] and [o] as well, while Rohlfs (1956'61) presents the relevant items with [e] and [a]. Ribezzo (1911-'12) is unclear. The oppositions [e] vs. [e] and [o] vs. [d] are crucial to Calabrese's analysis. They exist in most Italian dialects. We will suppose with Calabrese that the oppositions are relevant to Francavilla-Fontana as well. Perhaps the difference between the various sources is due to different diachronic stadia. We will not go into that problem here.

[i]: [vivu] [kuniggm] [vita] [«]: [pekura] [denti] [tjedda]

gloss alive rabbit screw sheep tooth cell

[e]: [pessi] [nei] [fiteli] [a]: [man] [käni] [sänd3i]

gloss fish snow fidelity sea dog blood

Vowel harmony in the dialect of Francavilla-Fontana [o]: [fòrti] [mòrti] [mònuku] [u]: [múlu] [lúna] [núdda] (Rohlfs 1956-'61)

strong dead monk mule moon nothing

163 tail crown tower

[o]: [kóta] [króna] [tórri]

The seven vowel system is restricted to stressed syllables. In unstressed position the mid vowels are raised to [i] and [u]. This can be demonstrated by way of the following alternations: (2)

gloss *[métuku] physician (male) [péssi] fish [mónuku] monk [kóta] tail

vs.

[séntu] vs. [sintímu] [kréu] [kritiámu] [tróu] [truámu] [kanósku] [kanujjímu] (Rohlfs 1956-'61, Calabrese

[mitikessa] [piskatritjl] [munatjieddu] [kutedda]

gloss feel, believe, find, know, 1985)

gloss physician (female) kind of frog monklike ghost neck

l.pers. sing./plur. pres.ind. id. id. id.

Addition of a stressed suffix causes destressing of the originally stressed vowels of the lexical stem in the examples in (2). The destressed mid vowels surface as their high counterparts [i] and [u]. We formulate the following rule. (3)

V

—

[+HIGH]/

["

"I

L-STRESS J [aBACK]

I

[aRND] This is a specific case of a rather general phenomenon of unstressed vowel raising, observable in many Romance languages and dialects (cf. Standard Italian and Brazilian Portugese, where seven stressed vowels are reduced to five: [e] and [o] become [e] and [o]).

164

Willebrord Sluyters

Alternations resulting from metaphony, the main topic of this paper, are given in (4). (4)

fem.sing. male sing./plur. fem.plur. [lènta] vs. [liéntu], [lienti] but: [lènti] [frédda] [friddu], [friddi] [fréddi] [gròssa] [gruéssu], [gruéssi] [gròssi] [pilósa] [pilusu], [pillisi] [pilósi] (Rohlfs 1956-'61)

gloss slow cold big hairy

(adj.) id. id. id.

It is clear that metaphony is a rule of partial, regressive vowel harmony, triggered by a high unstressed vowel. A preceding unstressed high vowel, like in [pilósa], does not cause raising of the following vowels. The target vowels [e] and [o] are raised to [i] and [u], respectively (note the difference with the previously discussed raising rule (3), which applies to unstressed vowels only). The target vowels [e] and [o] diphthongize into [ie] and [ut]. The fourth column in (4) displays exceptional forms, high vowels but no metaphony of the preceding stressed vowels. We will return to this later. The data in (5) suggest that the triggering high vowel must be wordfinal as well (taken from Calabrese 1985:55). (5) [pérsika] not [fémmina] [tólika] [duménika]

*[piérsika] *[fimmina] *[tuélika] *[duminika]

morphology [pérsik]+[a] [fémmin]+[a] [tólik]+[a] [duménik]+[a]

gloss peach woman bean Sunday

Of course we might stipulate that metaphony is triggered only by wordfinal high vowels, or else, that the penultimate high vowels in (5) are underlying nonhigh vowels, raised by rule (3), as proposed by Calabrese (1985:56). These, however, would be rather trivial solutions. The morphology of the examples in (5) points in another direction. The high vowels are part of the lexical stem ([persik] etc.), while the final high vowels we considered in (4) are case morphemes, the addition of which creates a derived environment. Metaphony might be a cyclic rule, subject to the Strict Cycle Condition. To check this hypothesis we consider the data in (6) (taken from Rohlfs 1956-'61). (6) [pejju] [spetji] [kretu] [tjekku]

not

*[piej"ju] *[spietji] "[krietu] *[tjiekku]

gloss worse (adv.) special (adv.) behind except

Vowel harmony in the dialect of Francavilla-Fontana

165

Italian in general has very few underived surface items. Nouns, adjectives and verbs are always derived, many adverbs as well. One syllable words, like articles, most prepositions and pronomina, are irrelevant, because metaphony requires at least two vowels. The scarce data gathered in (6) are an indication of the relevance of the cycle: final high vowels and yet no metaphony in underived items. Additional evidence comes from numerals, which can be derived and underived. gloss morphology ten [détji] [ditjìtóttu] eighteen [détjì] + [óttu] [duditji] twelve [dóe] + [détjl] [trititjl] thirteen [tré] + [détjl] sixteen [sèi] + [détjl] [sititji] [tiersu] third [térs] + [u] (Rohlfs 1956-'61, Ribezzo 1911-M2)1

gloss ten + eight two 4- ten three + ten six + ten

In the underived item [detjl] metaphony does not apply, [ditjitottu] is derived, but the domain of the metaphony rule is still the underived item [5ttu], so the rule cannot apply here either. On the other hand, in the examples [duditji], [trititjl] and [sititji], the domain of the metaphony rule is derived. The first vowel of these words is stressed. The second vowel, which is underlyingly stressed (see [detjl]), is destressed because of the stress on the first. As predicted by the Strict Cycle Condition (SCC), metaphony applies. The ordinal [tiersu] is derived as well because of the casus morpheme. Here as well, metaphony applies. Another check on cyclicity is the behavior of loan words. (8)

etymology [arkéstu] [orkéstra] [ddirramótu] [terremoto] [ritjèttu] [ritjètta] [sótu] [sodo] [fangóttu] [fagóte] [ddjileppu] [gulàb] (Rohlfs 1956-'61)

gloss orchestra earthquake recipe solid bassoon fruit syrup

origin ital., learned ital., learned ital., learned late latin mediev. french arabic

The examples in (8) are readily explained if metaphony is subject to strict cyclicity. Loan words typically behave as underived items with respect to cyclic rules. Although the evidence is limited, we think it suffices to propose that metaphony is a cyclic rule. We adopt Kiparsky's 1985 definition of the SCC.

166 (9)

Willebrord Sluyters "Strict Cycle Condition (SCC) If W is derived from a lexical entry W', where W' is nondistinct from XPAQY and distinct from XPBQY, then a rule A —B / XP QY cannot apply until the word level." (Kiparsky 1985:89)

Kiparsky proposes this formula to exclude the possibility, left open by the 1982 definition, of metrical and syllable structure assignment creating a derived environment. It is obvious that we need this restriction in the present case as well. Assignment of metrical structure to, for example, [detjl] shouldn't create a derived environment, otherwise metaphony would still apply, yielding the incorrect *[dietji]. Nonetheless, items like [[persik]a] remain problematic. The SCC allows for cyclic rules to apply at the word level in nonderived items. We still predict *[pitrsika]. What might be at issue here is whether the rule is neutralizing or not. According to Mascaro's original definition of the cycle (Mascaro 1976), a rule can be cyclic only if it is a neutralization rule. Kiparsky (1982:171) claimed that it is exactly the class of obligatory neutralization rules that tends to become cyclic. Corroborating evidence on the relevance of neutralization for cyclicity has been provided by Halle and Mohanan (1985) on Vedic, and by Hermans (pers. comm.) on Lithuanian stress. From their observations emerges a model in which cyclic rules can apply to underived items, but in which the application of neutralizing cyclic rules is subject to the derived environment constraint. With this proviso the data in (6) can be explained. As stated before, metaphony is a cyclic rule. Furthermore, it can be shown to be neutralizing. In the case of metaphonic raising, this is quite clear: the oppositions [i]~[e] and [u]~[o] are neutralized. As to metaphonic diphthongization, we must show that the diphthongs [ie] and [ue] exist independently of metaphony, that is, are not in all cases derivable. This is indeed the case. (10)

gloss [liggitra] square pillow [sguéjja] ugly woman -[ieri] nominal, suffix, sing. (plur. identical) -[ienti] nominal, suffix, sing. (plur. identical) (Rohlfs 1956-'61)

The diphthongs in (10) cannot be metaphonic. Final [a] is incapable of triggering diphthongization. The final [i]'s of the two suffixes shouldn't trigger diphthongization either, because, as will be shown, nouns having identical singular and plural forms never display metaphony in the singular.

Vowel harmony in the dialect of Francavilla-Fontana

167

Metaphonic diphthongization neutralizes the underlying opposition between [e], [o] and [ie], [ue]. The rule, being cyclic and neutralizing, cannot apply to underived items like [persik]. Interesting alternations arise also in words with proparoxytonic stress, viz. where trigger and stressed target vowel are separated from each other by another vowel (examples taken from Calabrese 1985:54). (11)

a. b.

plur.2 sing. underlying form gloss [miétitjl] [miétuku] [métik] + [u]/[i] physician [siékuli] [siékulu] [sékul] + [u]/[i] century [muénitjì] [mónak] + [u]/[i] [mónuku] monk [stuémitjl] [stómak] + [u]/[i] stomach [stómuku] [kuéflni] [kófunu] [kófan] + [u]/[i] barrel, coffin ([u] and [i] are the male sing./plur. case morphemes)

The examples in (11a) show that metaphonic diphthongization applies regularly if the stressed vowel is separated from the trigger by a high vowel. On the other hand, in ( l i b ) we note that an underlying intermediate [a] blocks metaphony in the singular and allows metaphony in the plural forms. The underlying [a] is motivated by morphologically related words (examples from Calabrese 1985:57). (12) [m5nuku] [kofunu]

gloss monk barrel, coffin

[munatjieddu] [nkufanitjare]

gloss monklike ghost to put in a barrel

Alternations are attested as well in the verbal paradigms. gloss vs. [[tró]i] find, [[tró]unu] [[siént]i] [[sént]i] feel, [[sént]unu] [[kanó/Jli] [[kanufj]i] know, [[kanósk]unu] [[krit]i] [[krét]i] believe, [[krét]unu] (Calabrese 1985:10-11, Ribezzo 1911-'12)

[[true]i]

2./3.pers.sing.pres.ind. 3.pers.plur.pres.ind. 2./3.pers.sing.pres.ind. 3.pers.plur.pres.ind. 2./3.pers.sing.pres.ind. 3.pers.plur.pres.ind. 2./3.pers.sing.pres.ind. 3.pers.plur.pres.ind.

A comment on [tróunu] is in order. Calabrese gives [troanu], probably a reconstruction based on a neighboring dialect. Ribezzo gives [tróunu] for Francavilla. Nonetheless, an underlying [a] might be motivated because of the infinitive [[tru]ari]. We might suppose that the suffix [anu], which

168

Willebrord Sluyters

originally belonged only to one conjugation, has been extended to the other two conjugations. In that case the problem of the 3.pers.plur. forms in (13) is identical to that of [monuku] from [m5nak]+[u], for which we will put forward a solution. Exceptional also are forms like [senti], as opposed to [sienti]. Similar examples can be attested in noun and adjective paradigms (data from Calabrese 1985:2, 14-15 and Ribezzo 1911-'12). (14)

a.

b.

male/fem.sing. [[mel]i] [[paret]i] [[stajonji] [[rikkettsja] fem.plur. [[bon]i] [[ross]i]

male/fem.plur. vs. [[miel]i] [[parit]i] [[stajunji] [[rikkittsji] male plur. [[buen]i] [[russ]i]

gloss honey wall season richness good red

The irregular forms (final high vowels, yet no metaphony) in the first column in (14) are not incidental. Nouns and adjectives with identical final vowels ([i], see a) for the singular and the plural, systematically never display metaphony in the singular and always in the plural. Female nouns and adjectives of the first declension (see b) never display metaphony in the plural. Once more, we could solve this problem by postulating an underlying word-final [e] (to be raised afterwards by rule (3)), as proposed in Calabrese (1985:13). Likewise, however, there are no alternations motivating an [e]. The argument is somewhat circular: no metaphony, therefore a final nonhigh vowel, a final nonhigh vowel, therefore no metaphony. It seems more plausible to us that the metaphony rule, lexicalized as we have argued already, is even more restricted so as to apply only within certain morphological categories in which it has a morphological function as well. For example, the metaphonic diphthong in [sienti] identifies this form as the 2. pers.sing., while the absence of metaphony in [senti] identifies it as the 3.pers.sing. The same goes for minimal pairs like [meli] and [mieli], where metaphony indicates the plural, and its absence the singular. In the pair [boni] vs. [bueni], metaphony marks the male plural, its absence the female plural. We could equally extend this line of argument to pairs like [sonu] ("I ring") vs. [suenu] ("sound" noun). Metaphony marks the noun, its absence the verbal 1.pers.sing. There is still another group of systematic exceptions to metaphony, but these seem to have a phonological motivation. Metaphony does not apply if trigger and target are separated from each other by an intervening [k?]

Vowel harmony in the dialect of Francavilla-Fontana

169

or [gy], the only two consonants, which apparently block the rule. Observe the data in (15), where we have included examples with other high consonants as well. sing. [vékkm] [spikkm] [kupérkm] b. [uikkyu] [tjirkyu] [finukkyu] c. [luengu] [liédd3u] [kruéttjì] [priéfju] [nuéffu] (Rohlfs 1956-'61) a.

plur. [vékkyi] [spékkyi] [kupérkyi] [uékkyi] [tjlrk y i] [finukkyi] [luéngi]

morphology [vékki] + [spékki] + [kupérki] + [uékki] + [tjlrki] + [finukki] + [long] + [liédd3] + [kruéttj] + + [prié/J] + [nó/J]

[u]/[i] [u]/[i] [u]/[i] [u]/[i] [u]/[i] [u]/[i] [u]/[i] [u] H [u] [u]

gloss old mirror table-cloth eye circle fennel long light well hook pleasure our (possess.)

Because of [vekkm], etc., Calabrese (1985:27) supposes that metaphony is blocked by an intervening high consonant. The forms in (15b,c) show that this cannot be the correct generalization. Metaphony is blocked only by a palatalized velar, and even then only in the case of the target vowel [e]. Target vowels [o], [e] and [o] regularly undergo the rule and surface as [ue], [i] and [u], respectively. Metaphony, as such, is not blocked by [k>] or [gy]; rather, the development of the diphthong [ie] is problematical in this environment. We propose that a rule of dissimilation is responsible for the examples in (15a). Surface palatalized velars are derived from underlying [ki] sequences. When the case vowel [u] or [i] is added to stemflnal [ki], the unstressed prevocalic [i] consonantalizes into [j] and palatalizes the preceding velar consonant. (16)

[

ki] + [u]/[i] — [

kju/i] — [

kyu/i]

After adding the case morpheme we obtain a diphthong [iu] or [ii], the initial part of which is identical to the initial part of the metaphonic diphthong [ie]. The following rule of diphthong dissimilation reduces the first of the two diphthongs.

170

Willebrord Sluyters

(17)

N

N

A

A

[+STRESS]

[+STRESS] / \ [+HIGH]

[+HIGH]

/ \V

V +

[+HIGH]

[+LOW]

[+HIGH]

[+LOW]

[«BACK]

[aBACK]

[aBACK]

[aBACK]

[aROUND]

[aROUND]

[aROUND]

[aROUND]

For (17) to operate correctly, it is necessary that the second element of the metaphonic diphthong be, somehow, invisible. The rule must be able to establish the exact identity of the two glides. The feature [+HIGH] does not pose a problem because the [e] (whether specified [-TENSE], as Calabrese supposes, or [+LOW], as we presume) will be unspecified for [HIGH]: on the [HIGH] tier both glides are adjacent. On the tiers of the features [BACK] and [ROUND], both glides are adjacent only if the specifications for the second part of the metaphonic diphthong are absent. We will consider this second part epenthetic. The V-position will be underspecified until the application of the following rule, which fills in the relevant values. (18)

N

A

N

A

[+STRESS] / \ V V I [+HIGH] [+LOW]

[+STRESS] / \ V V I [+HIGH] [+LOW]

[aBACK]

[aBACK]

I [aROUND]

[aROUND] [-ROUND]

[-BACK]

Vowel harmony in the dialect of Francavilla-Fontana

171

Metaphony is not the only type of vowel harmony in the dialect of Francavilla-Fontana. There is harmony as well for the features [BACK] and [ROUND]. (19) a. b.

sing. [monuku] [stomuku] [mittuku] [and3ulu]

c.

[siekulu]

plur. [muenitji] [stutmitji] [mietitji] [andjili] [siekuli]

morph. [monak] + [u]/[i] [st5mak] + [u]/[i] [metik] + [u]/[i] [and3il] + [u]/[i] [sekul] + [u]/[i]

gloss monk stomach physician angel century

Both [-BACK] and [+ROUND] harmony apply regressively like metaphony, but unlike metaphony their application is restricted to unstressed vowels. Vowels preceding the stressed vowel are not affected, so the domain of the rule is strictly posttonic. There is a striking difference between these rules as well. Examples (19a,b) show that rounding harmony applies regardless of the underlying specification of the target vowel: [a], being [+BACK], is affected as well as [i], which is [-BACK], The [-BACK] specification of [i] is, so to speak, "overruled". On the contrary, backness harmony respects the original specification of the target segment, as shown by (19c): the rule does not apply to underlying [+ROUND] vowels (recall that rounded front vowels do not exist in the dialect). In the present analysis we adopt a restricted way of underspecifying segments. Only information which is contextually redundant may be omitted from underlying specifications. Furthermore, we follow Stanley (1967), in requiring redundancy rules to be "surface true" that is to be true generalizations about the phonological system of a language or dialect. The seven vowel system has the following underlying specification. (20) HIGH LOW BACK ROUND

[i] +

[e]

[e]

[a]

[o]

+

+ +

+ + +

[o] -

[u] +

+

+

Unspecified values are filled in by the redundancy rules in (21). (21)

a.

[aBACK] ** [aROUND] /

I" 1 L-LOWJ

172

Willebrord Sluyters b.

V

I

—

[-LOW]

c.

[+HIGH]

V

I

-*• [-HIGH]

[+LOW]

If rounding harmony, triggered by [u], applies to [i] and results in [u], then we must suppose that spreading of the feature [+BACK] is simultaneously applied, otherwise [ii] would surface. The rule must refer to two features. (22)

[-STRESS]

I

V

[-BACK]

[-STRESS]

[-STRESS]

I —»• VI I [+ROUND] =: I T [+BACK] [-BACK]

[-STRESS]

V

^

IV -J [+ROUND] [+BACK]

The feature [+BACK] is redundant to [+ROUND] on [u]. As proposed by Kiparsky (1985:92-93), redundant values may not be assigned in the cyclic component of derivations. So, if rounding harmony makes crucial use of [+BACK], we can identify the rule immediately as postcyclic. This in turn means that rounding harmony is automatically ordered after metaphony, since metaphony is a cyclic rule. By contrast, backness harmony does not apply to [+ROUND] segments, that is, it does not make use of the redundant feature [-ROUND] on the trigger [i]. We will consider this rule to be cyclic, ordered in the same component as metaphony. Because backness harmony can feed metaphony, the first rule will apply prior to the second. Like rounding harmony, backness harmony is a common autosegmental spreading rule. (23)

[-STRESS]

[-STRESS]

V

V

[+BACK]

[-BACK]

I

I

[-STRESS] —

V

[-STRESS] V

f----J [-BACK]

[+BACK]

The independently motivated ordering of backness and rounding harmony in different components provides us with an explanation of the proparoxytonic words in (11), notably the alternation between [monuku] and [muenitjl] from underlying [m5nak]+[u]/[i]. If the intervening [a] can act as a blocker, we must assume that metaphony does not skip V-positions: it applies to adjacent vowels. The fact that consonants do not intervene

Vowel harmony in the dialect of Francavilla-Fontana

173

(except for the palatalized velars, which we have treated already) must be due to their lack of specification for the feature [HIGH]. [HIGH] on consonants in Italian is entirely predictable by means of other features on the same segment or by way of the phonological environment. Now, if we try to spread the feature [+HIGH] from word final [u] to the preceding [a] in [[monak]u], we would create the segment [+]. This segment does not exist in the dialect. Its creation by a cyclic rule is excluded because of the SCC, which requires cyclic rules to be structure preserving. So, metaphony cannot apply to [[m5nak]u]. In the postcyclic component, rounding harmony (22) will apply, turning the [a] into an [o], which subsequently will be raised to [u] by the raising rule (3), for all unstressed vowels. In [[m5nak]i] we can first apply backness harmony (23), changing [a] to [e]. To this [e], metaphony applies, yielding [i]. This [i], being a derived high vowel, will trigger metaphony of the preceding stressed [a]. The result is a diphthong [ue]. Schematically, we get the following sample derivations (on the formalization of diphthongization, see section 2). (24)

input: cyclic component: backness harmony (23) metaphony postcyclic component: rounding harmony (22) unstressed raising (3) output: input: cyclic component: backness harmony (23) metaphony postcyclic component: rounding harmony (22) unstressed raising (3) output: input: cyclic component: backness harmony (23) metaphony postcyclic component: rounding harmony (22) unstressed raising (3) output:

[[monak]u]

[[m5nak]i] [[mónttj]i] [[muenitj]i]

[[mónok]u] [[mónuk]u] [mónuku] [[sékul]u]

— — [muenitjl] [[sekul]i]

[[siékul]u]

[[siékul]i]

[siékulu] [[métik]u]

[siékuli] [[métik]i]

[[miétik]u]

[[miétitj]i]

[[miétuk]u] [mietuku]

[miétitjì]

174

Willebrord Sluyters

2. MET APHONY: RAISING VERSUS DIPHTHONGIZATION

Metaphony does not pose serious problems if it is applied to the vowels [e] and [o]. It has the effect of raising them to [i] and [u], We can consider the rule to be a standard autosegmental spreading operation on the feature [HIGH]. (25) [ ROUND] I [ BACK] I [-LOW] I [-HIGH] ([e]/[o])

[+HIGH]

[ ROUND] I [ BACK] I [-LOW] •F-[-HIGH]

[+HIGH]

([u]/[i])

([i]/[u])

([u]/[i])

Delinking cum spread turns [e] and [o] into [i] and [u] respectively. Metaphonic diphthongization of [e] and [o] is less straightforward. Of course we would like to analyze this case as essentially the same phenomenon of [+HIGH] spread. The different output must be due to the different underlying specification. Calabrese (1985) has advanced the following formalization. Considering [e] and [o] to be nonlow lax vowels, he proposes that the application of [HIGH] spread yields the high lax vowels [I] and [U]. (26)

V [-LOW] I [-TENSE] I [-HIGH]

V

—

V

V

[-LOW] I [-TENSE] f — -• [+HIGH]

([e]/[o]) ([u]/[i]) (Calabrese 1985:34-35)

[-HIGH]

[+HIGH]

([I]/[U])

([u]/[i])

These segments do not exist in the dialect; a kind of repair mechanism,

Vowel harmony in the dialect of Francavilla-Fontana

175

called "clean up" has to be applied to them. The clean up in the case of the output of (26) consists of "linearization". (27)

V

—

V

V

[-TENSE]

/

[+HIGH]

[+HIGH]

[-TENSE]

[+TENSE]

[-TENSE]

[+HIGH]

[-HIGH]

The two conflicting features, [-TENSE] and [+HIGH], are linearized. After filling in the missing values, we obtain the diphthongs [ie] and [uo]. [uo] is subsequently changed to [ue] by a rule of rounding dissimilation. In itself, Calabrese's analysis is attractive because it enables us to describe metaphony as a simple rule of autosegmental spreading, not only in the case of the target vowels [e] and [o], but also in the case of [e] and [o]. Nonetheless, it raises questions. In the first place, we observed earlier that metaphony is a cyclic rule. The SCC, because of structure preservation, prohibits the creation of the segments [I] and [U] in the cyclic component. Furthermore, one would expect the repair mechanism to be the expression of a universal format, a kind of automatic rule. This is certainly not the case. Even within the group of Italian dialects a large number of them do not resort to diphthongization. The format is dialect specific. Third, the ordering of the conflicting features within the diphthong ([+HIGH] [-TENSE] and not [-TENSE] [+HIGH]) is not explained. The analysis contains a good deal of arbitrariness as well. This becomes clear when we consider the proparoxytones, [m5nak]+[u]/[i] etc. The derivation proceeds as follows (cf. Calabrese 1985:57-60). (28)

[[monak]

+

u]

[-TENSE] [+LOW] [-HIGH]

[-TENSE] [+LOW] [+HIGH]

r

[-HIGH]

[+HIGH]

After this first step we obtain a problematic configuration: [+HIGH] conflicts with both [+LOW] on universal and [-TENSE] on language specific grounds. The solution in this case is not linearization, but a second

176

Willebrord Sluyters

type of clean up, called "negation": the conflicting feature values are turned into their opposites. (29)

V I [-TENSE] [+LOW] [+HIGH]

V I ' [-TENSE] \ [+LOW] ) v [+HIGH] /

[+TENSE] [-LOW] [-HIGH]

With concomitant spreading of [+ROUND] we arrive at the intermediate representation *[mônoku]. The [o], being [-HIGH], will not trigger metaphony of the preceding stressed vowel. In the case of [[monak]i], we reach the same problematical configuration. However, here only two features are reversed. (30)

V [-TENSE] [+LOW] [+HIGH]

/[-TENSE] \ - V [+LOW] / [+HIGH]

[+TENSE] [-LOW] [+HIGH]

Simultaneous spreading of [-BACK] leads to the intermediate form [monitji], in which the derived penultimate [i] will cause diphthongization of the preceding [o]. Observe that no less than three clean up mechanisms and three stipulations on their application are needed. First, we must stipulate linearization or negation. Then, if negation, we must mention the number of features. Still, we cannot explain all data. Consider [duditji] from [d6]+[detjl]. After the first [+HIGH] spread, from the final [i] to the [e], the result is a high lax vowel. According to the analysis, we would get linearization, that is, diphthongization: [dudietji]. Not a single Italian dialect, however, displays diphthongization of unstressed vowels. The relevant generalization is exactly that diphthongization is limited to stressed vowels. This necessitates a third stipulation, mentioning [STRESS]. In this way the analysis is robbed of most of its explanatory power. In our view, a reanalysis of diphthongization is in order. Metaphony is not the only source of diphthongs in the dialect of Francavilla-Fontana. Lexical items can contain diphthongs, which may be rising or falling. All pairs of vowels may combine to form a diphthong, provided that at least one of them is high. The high vowel will be interpreted as a glide. In the case of two high vowels, usually the first one becomes a glide, while the second remains a full vowel. Note that in this case preference is given to, for example, the rising diphthong [iu] over its falling

Vowel harmony in the dialect of Francavilla-Fontana

111

counterpart [iu]. Unattested are the combinations [UD] and [DU] (again there seems to be an argument against Calabrese's analysis in which [uo] is created in the course of the derivation). (31)

gloss [lauru] bat [krai] tomorrow [uarra] war [iarta] high [5i] eggs [ionnula] catapult (Rohlfs 1956-'61, Ribezzo

[kutjeuli] [beiri] [kruedda] [liggiera] [iuttsa] [fuitia] 1911-'12)

gloss easy to cook drink (verb) straw basket square pillow faeces thursday

In addition to these lexical diphthongs, there are diphthongs created by two rules of [1] vocalization. We distinguish onset [1] and coda [1]. Onset [1] vocalizes after a tautosyllabic consonant; the result is [j]. This process is known in virtually all Italian dialects. Since the rule does not depend on the quality of the surrounding segments, synchronic alternations are not attestable. (32)

Vulgar Latin Franc.-Fontana [oklus] [uekkm] [blarjku] [¡¿Tjku] (Rohlfs 1956-'61)

gloss eye white

Whether the rule is synchronic or diachronic is irrelevant to our analysis. Relevant is the observation that the on-glide diphthongs it creates surface as such. Vocalization of coda [1] produces [w], but now look at the data in (33). (33)

Vulgar Latin Franc.-Fontana [falke(m)] [fuatji] [falsu(s)] [fuazu] [kelsu(s)] [dzuezu] [kalke(m)] [kuatji] [kalk(s)] [kuatfi] (Rohlfs 1956-'61)

gloss 'Orion' false mulberry tree lime kick (noun)

»[fautji] *[f£uzu] *[dz«uzu] •[kautji] *[kautjl]

Instead of the expected falling diphthongs (see the last column) we note rising [ua] and [ue]. These diphthongs cannot be attributed to a rule of [1] metathesis, say [al] -*• [la], because onset [1] always vocalizes into [j]. There must have been a rule of diphthong internal metathesis, applying after vocalization [1] [w].

178

Willebrord Sluyters

Coda [1] vocalization is conditioned by the following consonant, which must be a [tj] or [s]. If it is [k] or [d], coda [1] develops into [r]. Therefore we get the following, admittedly scarce, synchronic alternations. (34)

stem [kalk] [kalk]

[kuatjl] [kuatjl]

gloss lime kick

vs. [karkara] [karka^rju]

gloss limekiln heel

This evidence, together with the observation that metaphonic diphthongs are, counter to our expectations, rising, brings us to the generalization that all derived diphthongs in the dialect are rising. The same kind of metathesis is known in other dialects. It is also possible for metaphonic diphthongs to be falling, like in the dialect of Bisceglie (De Gregorio 1939). From this state of affairs it can be inferred that the rising character of the metaphonic diphthongs in Francavilla must not be attributed to metaphony, but is independently motivated. We might think it to be a consequence of the very reduced syllable inventory of the dialect. In the coda only one consonant is allowed. Since the high vowel of a falling diphthong will be interpreted as a glide, and subsequently considered to be part of the coda, falling diphthongs in closed syllables would create coda's of two segments. Moreover, for whatever reason, there has been a (diachronic) tendency to avoid falling diphthongs by way of monophthongization or consonantalization of the glide. For example, Latin [tauru] in Italian dialects becomes [toro] or [tavuru]. The data in (32) and (33) furthermore indicate that diphthongs, at least in this dialect, are linked to two timing slots on the CV tier. There is no reason to suppose that [1] vocalization has any effect on the number of timing slots. If diphthongs resulting from [1] vocalization and those resulting from metaphony are subject to one and the same rule, or constraint (namely that they must be rising), it becomes highly probable that metaphonic diphthongs are bipositional as well. We also observe that metaphonic diphthongs appear only in stressed syllables. Compare metaphony with unstressed raising (3). (35)

stem [monak] [monuku] [muenitjl] [p^t] [peti] [pieti] [detji]

gloss gloss monk monks vs. [munatjieddu] monklike ghost foot feet ten vs. [duditji] twelve

Both rules add a feature [+HIGH] to their target vowels. Still, the result is entirely dependent on stress. If stressed, [e] and [o] diphthongize, if

Vowel harmony in the dialect of Francavilla-Fontana

179

unstressed, they are raised to [i] and [u]. Considering a diphthong to be linked to two timing slots, we are able to explain this distribution. In most Italian dialects, the possibility of two tautosyllabic V-positions is limited to stressed syllables. Including stress causes no complication of the rule, because we must stipulate anyhow that metaphony does not affect pretonic vowels. Cross-dialectically, this hypothesis implies two correct predictions. The double V is marked in closed syllables, so we predict that when dialects reduce diphthongs to monophthongs, this will happen first in the closed syllable. As a matter of fact, there are dialects with metaphonic diphthongs in open and closed syllables (the one examined here, for example). There are dialects with metaphonic diphthongs exclusively in open syllables. In the dialects of the border zone between Tuscany and Umbria, metaphonic [uo] from [D] appears only in open syllables (Schiirr 1970:40). There is no dialect which has metaphonic diphthongs only in closed syllables. The second prediction is that generalization of originally metaphonic diphthongs will primarily affect the open syllables. This too is correct. There are a few dialects which have generalized the diphthongs independently of syllable structure (the Palermo dialect, for example). Elsewhere, originally metaphonic diphthongs appear also before final nonhigh vowels, but only in open syllables, like in Northern Apulia (see Stehl 1980:242243). Once again, there is not a single dialect that has generalized the diphthongs in closed syllables. These predictions cannot be made by an analysis like that of Calabrese, in which diphthongs are considered monopositional - a kind of complex vowel. Complex segments are marked independently of syllable structure. Instead, we propose that diphthongs are complex nuclei: a diphthong consists of two vowels linked to one and the same syllable nucleus. We believe we have adduced convincing enough evidence to conclude that the particular nature of metaphonic diphthongs is only in a very limited way related to metaphony itself. The two positions result from the presence of stress. The rising character follows from a rule or constraint applying to all derived diphthongs. We must choose between stipulating this as a constraint or formulating a rule of diphthong internal metathesis. In view of the evidence presented in (34), we propose a metathesis rule. Nothing crucially hinges on this, however. If one were to avoid metathesis because of its inherent complications, the analysis would essentially remain the same. A third aspect of the metaphonic diphthongs, the second segment [«], has already been touched upon. We may add that neither the quality of the second segment, nor its epenthetic nature as such, is completely unexpected. We think the quality of the second part to be the result of universal considerations: [-BACK] and [-ROUND] are the universally

180

Willebrord Sluyters

unmarked values for both features. Cross-dialectically too, this choice is motivated. Dialects display [it] and [uo] as metaphonic diphthongs, or [ie] and [ue]. The combination [io] and [uo] never occurs. That is to say, the [+BACK], [+ROUND] second part is never generalized. Furthermore, certain Italian dialects have postlexical diphthongization rules, leading to alternations between [o] and [eu] (see Rohlfs (1938) for examples). So [eu], or [ue] for that matter, as a product of [o] is possible even in the case of a synchronically fully productive rule. The same goes for the Spanish diphthong [ue] from [o] (Harris (1985) and Garcia-Bellido (1986)). The preceding observations lead us to the following analysis. The second V position of diphthongs is contextually determined by stress. The feature values it takes are predictable as well. We propose first a rule introducing the empty V. (36)

[+STRESS] 0 — V /

/ II

V

After application of (36) to the underlying segments [e] and [o], we obtain a double V, of which the second one is empty. To this empty V, metaphony, that is, spreading of [+HIGH], applies. (37)

[+STRESS] / \ V V V I [+LOW] [«BACK] [aROUND] [+HIGH]

[+STRESS] / \ V V

I

[+LOW] [aBACK] [aROUND]

V \ \ \ [+HIGH]

The output of (37) is subject to the rule of diphthong internal metathesis (38). The features [+LOW] and [+HIGH] are delinked from their V slots and are subsequently spread to the other V slot of the complex nucleus. We suppose that this kind of metathesis is possible because [LOW] and [HIGH] are features having their own tiers. They are related by way of another tier, to which we will refer as the 'Vertical Position' (VP) tier, following Wetzels (1986:339).

Vowel harmony in the dialect of Francavilla-Fontana (38)

[+STRESS]

[+STRESS]

/ V

V

\ V [+HIGH]

181

V t+HIGH]

VP [+LOW]

[+LOW]

The output of (38) is subject to rule (18) that fills in [-BACK] and [-ROUND] on the second V. Of course, the underlying specification of the first V is maintained, in the case of [a], [+BACK +ROUND]; in the case of [e], [-BACK -ROUND]. If the configuration resulting after V insertion (36) does not find itself in a metaphonic context, then the second V will remain empty. At the end of the derivation, the feature values of the first V will be spread by virtue of a mechanism which, in De Haas (1988), has been put forward to account for vowel coalescence. This mechanism is an elaboration of the Nuclear Fusion Principle (NuFuP) proposed in Wetzels (1986). The mechanism collapses the matrices of all tautosyllabic vowels, provided that they do not contain conflicting feature specifications. The result is a long homorganic vowel, to be shortened in closed syllables. Metathesis (38) presupposes another mechanism as well. The association line of the feature [+HIGH], between trigger and target of the metaphony rule, must be erased, leaving both segments [+HIGH]. This operation can be seen as a consequence of the Nuclear Fission Principle (NuFiP, see Wetzels 1986). (39)

N

N

I

I

V

V

N 1 1 Vj

N 1 1 V|

[aF] [aF] [aF] By (39) the joint matrix of two heteronuclear vowels is split up into two identical nonlinked matrices. We need a principle like (39) to be able to metathesize the feature [+HIGH] in (38) from the second V to the first (recall that [+HIGH] originated from a spreading rule). Of course, if NuFuP and NuFiP are universal formats, they do not bring any extra costs to our analysis. Two sample derivations are given in (40).

182

Willebrord Sluyters

(40)

[gross] + [u]/[i]

[lent] + [u]/[i] V insertion (36)

eV

metaphony (37)

eV

DV V

oV

' " - - — J

metathesis (38)

V

I

IV

t+HIGH]

[+HIGH]

iVI [+LOW] feature filling (18) output

[+HIGH]

t+HIGH]

NuFiP (39)

ie [lient] + [u]/[i]

V

V I t+HIGH]

V I [+HIGH]

uVI t+LOW] ue [grutss] + [u]/[i]

We conclude that the relatively complex data of vowel harmony in Francavilla-Fontana can be explained by the normal autosegmental spreading rules. What matters is the interaction of the relevant rules and the way in which they are conditioned. Once it is clear that metaphony and backness harmony are cyclic, and rounding harmony is postcyclic, the apparently strange alternations presented in (11) follow automatically. No complex machinery is needed to derive them, provided that the SCC is amended in the sense of Halle and Mohanan (1986) and Hermans (forthcoming). More complicated is the account of metaphonic diphthongization. We have demonstrated the need for three language specific rules (apart from the spreading rules referring to the features [+HIGH], [-BACK] and [-ROUND]); V insertion (36), metathesis (38) and feature filling (18), of which the last one has universal justification as well. Rules (36) and (38) were shown to be independently motivated in the dialect. Our analysis does not make use of intermediate representations, in which segments appear that neither exist underlyingly, nor in surface items. It remains within the limits dictated by the amended SCC, and therefore is able to explain much more data than covered by Calabrese's analysis. With regard to special rule formats, the highly language specific formats of Calabrese (1985) can be dispensed with. We only need two principles, NuFuP and NuFiP, the relevance of which has been demonstrated independently of the dialect studied in this paper.

Vowel harmony in the dialect of Francavilla-Fontana

183

At several points we have indicated how the analysis might be extended to other dialects displaying this kind of phenomena. For example, one and the same rule may be differently conditioned. In many Calabrian dialects metaphony is not a cyclic but a postcyclic rule. Elsewhere, metaphonic diphthongs are falling, which means that the metathesis rule is not present there. The reduction of unstressed vowels is dialect specific too. In Francavilla-Fontana there are only three unstressed vowels, in Tuscany five. Other dialects admit only one or two unstressed vowels. Here again, the phenomenon, unstressed vowel reduction, is extremely common, but its elaboration is dialect specific.

NOTES * The research for this paper is part of project number 300-164-011 of the Dutch Foundation for the Advancement of Pure Scientific Research (ZWO), entitled 'An investigation into the origin and representation of complex vowels', carried out under the guidance of Leo Wetzels. I'm very grateful to him for careful reading of and commenting on earlier drafts of this paper. Thanks also to Wim de Haas and Haike Jacobs for comments and suggestions. 1. The only exception to the rule is underived [uettu], where we would expect [ottu]. Given the existence of forms like [ditjitittu], we propose to consider [uettu] an exception, probably a loan from a neighboring dialect. 2. We choose not to treat the rule of affricatization/palatalization, which is responsible for the consonant alternations [k]~[tj]. It does not intervene in the harmony processes discussed in this paper.

REFERENCES Calabrese, A. 1985. Metaphony in Salentino. (ms.) Cambridge Mass., MIT. Garcia-Bellido, P. 1986. Lexical Diphthongization and High-Mid Alternations in Spanish. Linguistic Analysis 16, 61-92. De Gregorio, I. 1939. Contributo alla conoscenza del dialetto di Bisceglie (Bari). L'Italia Dialettale 15, 31-52. De Haas, W. 1988, forthcoming. Vowel coalescence. Halle, M. and K. Mohanan. 1985. Segmental Phonology of Modem English. Linguistic Inquiry 16, 57-116. Harris, J. 1985. Spanish Diphthongization and stress: a paradox resolved. Phonology Yearbook 2, 31-45. Kiparsky, P. 1982. From Cyclic Phonology to Lexical Phonology. H. van der Hulst and N. Smith (eds.). The Structure of Phonological Representations, (part 1). DordrechtCinnaminson, Foris, 131-175. Kiparsky, P. 1985. Some consequences of Lexical Phonology. Phonology Yearbook 2, 85137. Mascaro, J. 1976. Catalan Phonology and the Phonological Cycle. Bloomington Ind., I.U.L.C. Ribezzo, F. 1911-'12. Il dialetto apulo-salentino di Francavilla- Fontana. Martina Franca.

184

Willebrord Sluyters

Rohlfs, G. 1938. Der Einfluss des Satzakzentes auf den Lautwandel. Archiv für das Studium der neueren Sprachen 174, 54-56. Rohlfs, G. 1956-'61. Vocabolario dei dialetti salentini (Terra d'Otranto). (3 vol.) München, Beck'schen Verlagsbuchhandlung. Rohlfs, G. 1966. Grammatica storica della lingua italiana e dei suoi dialetti. I. Fonetica. II. Morfologia. III. Sintassi e formazione delle parole. Torino, Einaudi. Schiirr, F. 1970. La Diphtongaison Romane. Tübingen. Sezer, E. and L. Wetzels. 1986a. On the interaction of backness and rounding harmony. F. Beukema and A. Hulk (eds.). Linguistics in the Netherlands 2, 1986. Dordrecht-Riverton, Foris, 209-216. Sezer, E. and L. Wetzels. 1986b. Derived opacity in Yakut Vowel Harmony. Gramma 2, 201-224. Stanley, R. 1967. Redundancy Rules in Phonology. Language 43, 393-436. Stehl, Th. 1980. Die Mundarten Apuliens. Historische und strukturelle Beiträge. Münster, Aschendorffsche Verlagsbuchhandlung. Wetzels, L. 1986. Phonological Timing in Ancient Greek. L. Wetzels and E. Sezer (eds.). Studies in Compensatory Lengthening. Dordrecht-Riverton, Foris, 297-344.

Vowel harmony in Finnish word games Robert M. Vago City University of New York

1. INTRODUCTION

There is a growing body of research in the phonological literature that seeks to bring external evidence derived from word games, secret languages, speech disguises, and the like (called ludlings by Laycock (1969, 1972)) to bear on the various non-linear theoretical frameworks that have emerged in the recent past; McCarthy (1979, 1982), Yip (1982), Mohanan (1982), Davis (1985), Vago (1985b), Clements (1986), McCarthy and Prince (1986), Bagemihl (1987a, 1987b), Lefkowitz (1987), and Lefkowitz and Weinberger (1987) is a partial but representative list. In line with this research program, the present contribution has a twofold purpose: (a) to provide a formal account of the patterning of vowel harmony in three Finnish word games, and (b) to draw conclusions for underspecification theory. The data are drawn primarily from a series of studies by Campbell (1980, 1981, 1986), and, when noted, from Seppánen (1982). General familiarity with the basic tenets of underspecification theory is assumed; see in particular the seminal works of Kiparsky (1982b, 1985), Archangeli (1984), Pulleyblank (1986), and Archangeli and Pulleyblank (1986). Two further assumptions will be made: (a) as claimed by the framework of lexical phonology (cf. Kiparsky 1982a, 1982b, 1985; Mohanan 1982, 1986; Pulleyblank 1986, among many others), phonological rules may apply in the lexical component and/or the postlexical component of grammar; (b) as suggested by Mohanan (1982) and Bagemihl (1987b), language games are played in a separate component, here identified as the game component, mediating between the lexical and postlexical components, to which both rules of the natural language and rules specific to word games may belong. Thus, phonological rules will be assigned to one or more of the following, sequentially ordered, components: lexical, game, postlexical.1 Section 2 lays the groundwork with a brief overview of the underspecification analysis of Finnish vowel harmony assumed for the body of the paper. Section 3 describes the first word game and develops the major rules relating to vowel harmony suggested for the game component of

186 Robert M. Vago Finnish. Section 4 is concerned with the second word game and consequences for the analysis of transparent, or neutral, vowels. Section 5 provides an account of disharmonic words in the third word game. Finally, section 6 summarizes the most salient results.

2. FINNISH VOWEL HARMONY AND UNDERSPECIFICATION THEORY

As is well-known, the Finnish vowel harmony system is root-controlled, based on the feature [back].2 With respect to vowel harmony, the Finnish vowel qualities are classified as follows, using standard Finnish orthography: (1)

a. b. c.

High y u i

Front harmonic: Back harmonic: Transparent (neutral):

Mid ö o e

Low ä a

In general, the vowels of a root may belong to either set (la) or set (lb); transparent vowels may occur with either set. Typically, harmonic vowels in suffixes alternate, based on the harmonic category of the root: back harmonic roots take back harmonic suffixes, front harmonic roots and those containing only transparent vowels take front vowel suffixes.3 The transparency of / i / and / e / is established by the fact that a preceding back vowel determines back harmony for a following harmonic vowel. Some examples from Steriade (1987b) are provided in (2). (2) a. b.

talo mykà lume Pariisi

'In' talossa mykàssà lumessa Pariisissa

'house' 'mute' 'snow' 'Paris'

The harmonic alternation of the vowel of the inessive case suffix is evident in (2a); the examples of (2b) reveal the patterning of the two transparent vowels. Loanwords may violate the usual cooccurrence restrictions imposed by vowel harmony: they may contain both front and back harmonic vowels. Steriade (1987b) mentions the following examples of disharmony: (3)

a.

martyyri jonglôôri analyysi

'martyr' 'juggler' 'analysis'

b.

syntaksi tyranni fôljetongi

'syntax' 'tyrant' 'feuilleton'

Finnish word games

187

The suffix harmony of disharmonic roots is determined by the last harmonic vowel of the root. Accordingly, disharmonic roots like those in (3a) take front harmonic suffixes, those in (3b) back harmonic ones.4 The word games that will be examined impact directly on the three aspects of the Finnish vowel harmony system whose essential and relevant facts have just been exposed: regular harmony, transparent vowels, and disharmony. The underspecification analyses assumed for these topics are discussed next. 2.1. Regular Harmony Under the strong interpretation of underspecification that is adhered to in most current works (cf. the references cited in section 1), only one value of a distinctive feature is specified underlyingly.5 With respect to the harmonic feature [back] in Finnish, some analyses argue for [+back] to be the lexical feature (e.g. Kiparsky 1981), others opt for [-back] (e.g. Goldsmith 1985). The present study follows the position that [+back] is specified and spreads lexically and that [-back] is filled in by default; the opposite position appears to be tenable as well, with necessary modifications, to be sure. Ramifications for the lexical value of the feature [back] in Finnish will not be explored here; ultimately, the choice will rest on language internal evidence. One of the fundamental axioms of underspecification theory is that distinctive features and their class nodes are arrayed on autonomous tiers within a universally defined segment architecture. For some specific proposals on the geometry of phonological features, see Clements (1985), Sagey (1986), Archangeli (1987), Archangeli and Pulleyblank (1986, 1987a, 1987b), and Steriade (1987a) among several others. As regards the Finnish vowel system, it is uncontroversially distinguished by the features [back], [high], [low], and [round]. The hierarchical organization of these features is assumed to be as in (4).

188 (4)

Robert M. Vago Nucleus

N

Skeleton

X

Root Node

o

Supralaryngeal Node

o

Place Node

o

Labial Node Dorsal Node

[low] For the argument, given in the next section, concerning the representation of transparent vowels to hold, it is crucial to accept the position that the feature [round] does not dock onto the same class node tier with the other three features. Given the (incomplete) segmental structure in (4), the vowel harmony rule of Finnish is formalized in the following terms: (5)

Back Harmony Dorsal Node Back

(BH) ° /. 0 +B

Back Harmony spreads the feature [+back] to a following dorsal node. Since harmonic vowels are targets of the spreading process, they must be represented on the dorsal node tier. On the assumption that class nodes cannot be terminal tiers in the feature hierarchy, harmonic vowels will be characterized by at least one of the features that is dominated by the dorsal node, namely [back], [high], or [low]; see section 2.4. The operation of Back Harmony on the suffix cycle is observed in the sample derivation given in (6).

189

Finnish word games (6)

t

a

1

o

s

s

a

x

x

x

x

x

x

x

Supralaryngeal Node

I O

I O

I O

I O

Place Node

l o

Skeleton Root Node

I

I

i

o

I

l

o

I

i

I

Dorsal Node

\ /

I

o

I

Note that the dorsal node tier is irrelevant for consonants. Tiers not in the path connecting the skeleton and the [back] tier are omitted from consideration. 2.2. Transparent Vowels Underspecification theory explains transparent behavior in geometric terms; see in particular Archangeli and Pulleyblank (1986) and Vago (1988, in preparation). Since transparent vowels (and consonants in general) are neither triggers nor targets of the Back Harmony rule, they will lack representation on the dorsal node tier. This analysis is forced by the Locality Condition, a constraint of underspecification theory proposed by Archangeli and Pulleyblank (1986, 1987a), according to which the trigger and target nodes must be adjacent to each other on the tier where spreading takes place. To see this, consider the application of Back Harmony shown in (7), where the transparent vowel / e / is represented on the dorsal node tier: (7)

1

u

m

e

s

s

a

x

x

x

x

x

x

x

Supralaryngeal Node

I O

I O

I O

I O

Place Node

°l

Skeleton Root Node

Dorsal Node

I

I

i

I

l I °

L_

+B

I

\ /

I

i I °

__

I

190 Robert M. Vago Since the trigger and target nodes are not adjacent on the dorsal node tier, where spreading takes place, the above representation is rendered illicit by the Locality Condition. 6 If the Finnish transparent vowels have no dorsal nodes, i.e. are not specified for the features [back], [high], and [low], then two possible analyses are left open: (a) they are specified only for the feature [round], which is the only distinctive feature of the Finnish vowel system that does not anchor onto the dorsal node tier; (b) they are completely underspecified for the set of features relevant for vowels and thus have no representation below the skeleton. Since Finnish has two transparent vowels, both possibilities are utilized: one is analyzed in terms of [-round] and no other feature, while the other is specified for no feature at all. This author is aware of no language internal arguments that would force a particular choice. However, on the view that extreme underspecification is the preferred analysis of transparency, and taking into account the fact that in [back] harmony systems / i / is the quintessential transparent vowel (see for instance Anderson 1980), considering / i / to be the unspecifed vowel would seem to make sense. As a consequence of the geometrically construed analysis of transparency, the spreading of [+back] onto the dorsal node tier will operate on adjacent trigger and target nodes. E.g: (8)

l u m e s s a

P a r i i s i s s a

Root Node

xxxxxxx I I I I \/ I o o o o o o

x x x x x x x x x x III I \/ I ooo o o o

Supralaryngeal Node

o| o| | o oI oI oI

III

I

I I

Place Node

o o o o o o

ooo

o

o o

Skeleton

Labial Node

i I I I

Dorsal Node

° +B

l l ° I I -R

I

i I I

I

I ° +B

—°

191

Finnish word games 2.3. Disharmony

In the case of disharmonie roots, back vowels are lexically linked to [+back]; front vowels, as before, are not specified for backness. E.g: (9)

a Skeleton

n

a

l

y

y

s

i

s

s

à

Root Node

x x x x x x x x x x x I I I I \ / I \ / I o o o o

Supralaryngeal Node

I o

I I o o o

Place Node

o

o

I

Dorsal Node

I o o

I

I

I +B

The Strict Cycle Condition (Kiparsky 1982a) prevents the spreading of the lexically associated [+back] feature root internally to the front vowel /yy/. The Locality Condition in turn prevents spreading to the suffix vowel, since on the dorsal node tier / y y / intervenes between the trigger back root vowel and the target suffix vowel.7 Rather, front harmonic / y y / and / a / , as well as transparent / i / , become [-back] by default. The analysis of disharmonic roots whose final harmonic vowel is [+back], as in /syntaksi/, is unproblematical. The [+back] feature of the opaque back vowel does not link to the front root vowel, for either of two reasons: (a) strict cyclicity, as discussed above; (b) Back Harmony does not spread leftward. The Strict Cycle Condition is inoperative on suffix cycles, so [+back] can spread to suffix vowels: cf. for example the inessive inflection /syntaksissa/. 2.4. Summary We are now in position to set up the underspecified vowel system of Finnish: (10)

i [back] [round] [high] [low]

e

ii

o

a

-

-

-

u +

o +

a +

-

-

-

192

Robert M. Vago

The lexical values [+back] and [-round] force [-high] and [-low] to be lexical as well: either lexical values [+high], [+low] would yield nondistinct representations in the vowel system.8 The features [round], [high], and [low] are specified minimally, allowing for a maximally underspecified vowel system. No vowel is analyzed solely in terms of [+back]; this allows for the possibility to capture root harmony by means of a floating morphemesized [+back] feature. This feature becomes associated with the vowels / u o a / first by a universal linking convention ("one-to-one, left-to-right"; cf. Pulleyblank 1986) and then by the Back Harmony rule given above in (5).» The feature values missing in (10) are supplied by the following default rules:10 (11)

Default Rules a. [-round] b. [-round] c. [-low] d. [-high] e. [] f. [] g- [] h. []

—

— — —

— — — —

[-low] [-high] [+round] [+low] [-low] [+high] [-round] [-back]

Rules (1 la,b) are applicable only to the vowel / e / , since at the stage when they apply only / e / is specified as [-round]. Default rules supply predictable feature values, and as such cannot change an already specified value. This characteristic is evident in the fact that rules (11c,d) do not apply to the vowel / e / ; i.e., (11c) does not change the lexical value [-round] and ( l i d ) does not affect the [-low] value that was assigned to / e / by rule (11a). Rules (lle-h) fill in the remaining predictable feature values. Default rules are subject to strong constraints, such as prohibiting extrinsic ordering within components and disallowing ternary feature values; see Pulleyblank (1986) for an excellent synthesis of these issues. For immediate purposes we need not be concerned with the specific ordering and component assignments of the default rules in (11), except, of course, for the [-back] default rule (llh). In the next section it will be suggested that rule ( l l h ) applies as early as the game component. 11 In summary, this section highlighted the assumed analyses of those aspects of the Finnish vowel harmony system for which the word games to be described in the next three sections have direct relevance. To facilitate the exposition of Finnish forms in derivations, segmental structure will be compressed into two tiers only: the [back] tier, represented with the autosegment B, and all other tiers combined, represented with phonemic symbols. E.g:

193

Finnish word games (12)

+B r lumessa

The above representation indicates the application of the Back Harmony rule. In interpreting such abbreviations, care should be taken to bear in mind the suppressed feature geometry and segmental underspecifications, especially as regards the transparent vowels. 3. /tà/ GAME To play the first game, the consonant / t / and a low vowel are inserted after the initial syllable of a word. In the case of regular (harmonic) roots, the low vowel is back / a / if the initial syllable contains a back vowel, front / a / otherwise. E.g: (13)

kala kevàt kâdessâ

'fish' 'spring' 'in hand'

katala ketâvât kàtâdessâ

To account for these facts, the following rule is assigned to the game component of Finnish:12 (14)

/ t a / Game Rule Insert / t a / after the initial syllable of a word.

In effect, the vowel / a / derived by rule (14) is unspecified for the feature [back]. To account for the fact that / a / undergoes harmony in back harmonic contexts (cf. /katala/), we simply extend the domain of the Back Harmony rule to include not only the lexical phonology but the game component as well. Otherwise, [-back] is assigned to / a / by default rule (11 h). Of course, other default rules are applicable as well, but we will not be concerned with these. A further consequence of the / t a / game is that the / t a / variant induces front harmony in a following back vowel. This is observed in the case of disharmonic loan words whose initial front vowel is followed by a back vowel. E.g: (15)

dosa fyssa

'bus' 'physics class'

— —

dotasa fytassa

The facts of disharmonic loans that contain a back vowel followed by a front vowel will be discussed in section 5.

194 Robert M. Vago The harmonic adjustment of back vowels observed in (15) is derived in two steps. First, the [-back] default rule is active in the game component. And second, the insertion of [-back] default values is followed by a featurechanging Front Harmony rule: (16)

Front Harmony Dorsal Node Back

(FH) 0

-B

0

+B

If the Redundancy Rule Ordering Constraint of underspecification theory is accepted (cf. Archangeli 1984; Pulleyblank 1986; Archangeli and Pulleyblank 1986), then default rule (llh) is automatically ordered before Front Harmony: the default rule derives the [-back] feature to which Front Harmony refers. The sample derivations in (17) illustrate the organization of the game component as developed so far: (17)

Output of

+B

lexical phonology

kala

kevàt

+B / t à / game rule (14)

katàla

+B I dosa

ketàvàt

+B I dotasa

N/A

N/A

+B BH (5)

A

katala

-B Default ( l l h )

N/A

A

ketàvàt

- B +B

K I

dotasa

FH (16)

N/A

N/A

/K

-B dotasa

It may be noted that Back Harmony is exclusively left to right directional: cf. */dotasa/. Also, whenever appropriate, representations will reflect the effect of the Obligatory Contour Principle (OCP); 13 cf. for example the output of default rule (1 lh).

Finnish word games

195

That back vowels following the inserted / t a / undergo reharmony is also evident in the partitive singular inflections of the roots /meri/ 'sea' and /veri/ 'blood'. These roots are predictably front harmonic, except in the partitive singular, where they take the back harmonic variant of the regularly alternating partitive suffix: /merta/, /verta/. In the / t a / game, these irregularities are removed: /mertata/, /vertata/. The explanation of these cases parallels that o f / d o t a s a / in (17) above.

4. PIG GERMAN

Another Finnish word game is played according to the following rule: (18)

Pig German Game Rule Interchange the onset (if any) and nucleus of the initial syllable in each succeeding pair of words.

Campbell gives no specific name to this game; Seppanen refers to it as Siansaksa 'Pig German'. Some examples are listed in (19). (19)

a. b.

Saksalaisia hatyytettiin 'the Germans were attacked' —•haksalaisia satuutettiin tykkaan urheilusta 'I like sports' —• ukkaan tyrheilysta

It is readily observed that following the interchange, harmonic vowels in non-initial syllables assimilate in backness to the transposed harmonic vowel in the initial syllable. This consequence follows automatically if the Pig German Game Rule precedes the set of rules which spread or specify values for the feaure [back] in the game component; the sequencing of these rules has already been established in the preceding section. Rule (1 lh) supplies the default [-back] value to the front harmonic vowels / y o a / as well as to the transparent vowels / i e/. Ordering this rule before Front Harmony makes the prediction that a transposed / i / or / e / will induce the fronting of a back vowel in a succeeding syllable, just as a transposed front harmonic vowel does. Apparently, this claim is correct for some dialects, but not others. In Seppanen's description of Pig German, reharmony does in fact obtain: e.g. /leipa taikina/ 'bread dough' becomes /taipa leikina/. However, Campbell's discussion of the facts indicates that displaced transparent vowels fail to condition reharmony. Cf. the following examples: (20)

a. b.

otsansa hiessa 'in the sweat of his brow' — hitsansa oessa pitaa kalasta 'likes fish' — kataa pilasta

196 Robert M. Vago Since in the dialect described by Campbell front harmonic and transparent vowels pattern differently with respect to reharmony, their default specifications should not be derived simultaneously. In particular, front harmonic vowels become linked to [-back] before the rule of Front Harmony and thus induce reharmony, but transparent vowels receive their [-back] default values after Front Harmony and therefore block reharmony. This analysis, if correct, lends support to Steriade's (1987b) suggestion to distinguish between two kinds of predictable feature values: distinctive vs. redundant. In the present instance, [-back] is distinctive for the front harmonic vowels, redundant for the transparent vowels. In terms of a formal account, harmonic front vowels undergo the following rule: (21)

Distinctive Default (DD) Dorsal Node Back

° i i -B

The above default rule associates [-back] with a dorsal node which is not linked to the [back] tier. Since it introduces a feature which serves as input to Front Harmony, it is intrinsically ordered before Front Harmony by the Redundancy Rule Ordering Constraint. At the stage when rule (21) applies, only /y o a / have dorsal nodes which are not associated on the [back] tier: / u o a / are tied to [+back], and ex hypothesi, the transparent vowels / i / and / e / have no representation on the dorsal node tier. Following the application of Front Harmony, rule (1 lh) applies and supplies the [-back] default value of the transparent vowels. In the context of the present discussion, it might serve well to rename this rule: (22)

Redundant Default (RD) [ ]--B

Note that if Redundant Default and Front Harmony were assigned to the same component, the Redundancy Rule Ordering Constraint would force Redundant Default to precede Front Harmony. To account for the behavior of transparent vowels in Campbell's dialect, we assign Front Harmony to the game component and Redundant Default to the postlexical component; 14 in Seppanen's dialect, both reside in the game component. (23) summarizes the component assignments of the rules in the two dialects under consideration. The default rules for the features [high], [low], and [round] are excluded from the list in (23). It should also be noted that in Seppanen's dialect, Distinctive Default precedes Redundant Default by the Elsewhere Condition (Kiparsky 1973a, 1982).»s

Finnish word games

197

(23) Component

Campbell's dialect

Lexical

Game

BH

Seppänen's dialect

(5)

BH

Game rules (14; 18) BH (5) DD (21) FH (16)

Postlexical

RD

(5)

Game rules (14; 18) BH (5) DD (21) RD (22) FH (16)

(22)

(24) below contains two pertinent derivations in Campbell's dialect of the Pig German word game; derivations in the less interesting variant described by Seppanen should be obvious. (24)

Output of lexical component

Game rule (14)

+B tykkään +B I ukkään

urheilusta +B \ tyrheilusta +B \

+B BH (5)

+B

tyrheilusta +B

DD (21)

-B

I

+B

pitää +B I

+B

katää

pilasta

+B

+B pilasta

N

IV

kataa

\

K

N/A

tyrheilusta +B FH (16)

IV

+B RD (22); OCP

-B

ukkaan

(V

ukkaan

N/A tyrheilystä -B

+B

- B +B

tyrheilysta

kataa

pilasta

198

Robert M. Vago

5. / k o n t t i / GAME

Another Finnish word game impacting on vowel harmony is called kontti kieli or kontin kieli 'knapsack language': (25)

/kontti/ Game Rule Place /kontti/ 'knapsack' after a word.

Rule (25) is followed in turn by the Pig German game rule stated above in (18): the onset and nucleus of the initial syllable of /kontti/ and the preceding word are interchanged. The preposed back vowel of /kontti/ then triggers the application of Back Harmony. For instance, the sentence / H a n asuu Helsingissa/ 'he lives in Helsinki' becomes / H a n kontti kon hantti asuu kontti kosuu antti Helsingissa kontti kolsingissa hentti/. The facts of the /kontti/ game parallel those of Pig German and will therefore receive similar treatment. Some expressions demonstrating the application of Back Harmony and the transparency of / i / and / e / are listed in (26). kày mita nàhnyt sikiô pysàhtyà menetelmà

'he visits' 'what' 'seen' 'embryo' 'to stop' 'system'

kontti kontti kontti kontti kontti kontti

—

— — — — —

kou kântti kota mintti kohnut nàntti kokio sintti kosahtua pyntti konetelma mentti

The source materials on the /kontti/ game include information on the patterning of disharmonie words, with interesting results: the exceptional nature of disharmony with respect to the basic restrictions imposed by vowel harmony is preserved. That is, the non-initial front harmonic vowel of a disharmonie word does not harmonize to the preposed back harmonic vowel of/kontti/, as one might be led to expect on the basis of the patterning of harmonic roots. Cf. the examples in (27). (27)

manòòveri jonglôôri klorofylli hydrosfaari

'manoeuvre' 'juggler' 'chlorophyl' 'hydrosphere'

kontti kontti kontti kontti

konòòveri konglòòri korofylli kodrosfaari

mantti jontti kontti hyntti

The salient fact to be noted concerning disharmonic roots with an initial back vowel is that their root internal exceptionality survives through the game component. Although this fact is not evidenced in the Pig German game, where the sources include no relevant information, we may safely

Finnish word games

199

assume this to be the case in view of the fact that in other essential details, such as the patterning of diphthongs and long vowels, the /kontti/ game and the Pig German game have the same phonological consequences.16 As discussed in section 2, the back vowels of disharmonic roots are prelinked to [+back]; the Strict Cycle Condition blocks the application of the Back Harmony rule root internally. The facts in (27) clearly show that disharmonic roots do not undergo Back Harmony in the game component either. This fact is further evidenced in the / t a / game. It will be recalled from section 3 that in this game a disharmonic root of the sort / d o s a / undergoes the rule of Front Harmony and becomes /dotasa/. However, a back vowel in the initial syllable of a disharmonic root does not induce spreading on the [back] tier. Thus, for instance, /jongloori/ 'juggler' becomes /jongtaloori/; note that [+back] does not spread to the inserted / t a / either. In terms of a formal account, we can prevent the spreading of [+back] from the initial syllable of a disharmonic root by supposing that the following front harmonic vowel is linked to [-back] prior to the application of Back Harmony; recall that this rule targets harmonic vowels which are not specified for the feature [back]. It is possible to achieve this in the game component via the following rule: (28)

Disharmony Dorsal Node

°

I

+B

°

i

-B

Disharmony states that an unassociated (on the [back] tier) harmonic vowel which follows a back vowel becomes front. Following the lexical application of Back Harmony, a back vowel will be succeeded by a dorsal node that is not associated with [+back] only in the case of disharmonic roots; transparent vowels, it will be remembered, have no representation on the dorsal node tier. As regards the ordering of Disharmony, the following relations are noteworthy: it bleeds Back Harmony; it precedes the /kontti/ and Pig German game rules, since [+back] spreads in the case of harmonic roots; it follows the / t a / game rule - this explains the frontness of / t a / in cases like /jongtaloori/; 17 it precedes the Distinctive Default rule (21) within the game component by the Elsewhere Condition. The derivations in (29) should suffice to exemplify the account of disharmonic roots advocated here. They will also serve to illustrate the final list of rules proposed, as required for the three word games considered.

200 (29)

Robert M. Vago GAME COMPONENT Output of lexical component

+B I jonglööri

+B I jonglööri

+B I

/ t ä / game (14)

jongtälööri +BI - B I jongtälööri

Disharmony (28)

+B

-B

I /I

jonglööri

I -B /I

+B /kontti/ game (25)

jonglööri

I -B/I

+B Pig German game (18) BH (5)

DD (21); OCP

konglööri N/A +B I

POSTLEXICAL

I

+B jontti

N/A N/A

N/A COMPONENT +B

RD (22); OCP

kontti

N/A -B IV

jongtälööri

FH (16)

I

+B

-B

I ^ jongtaloori

+B

-B

I A

konglööri

+B -B

I I

jontti

It will be noted that in the current analysis there are three sources for the feature [-back]: Disharmony, Distinctive Default, and Redundant Default. Each rule is active in the derivation of /jongtaloori/ above. Finally, it should be clear that the entire game component is optional for the grammar of Finnish. Of its contents, only the Back Harmony and Distinctive Default rules belong to the natural language; the former is relegated to the lexical component, the latter to the postlexical (or phonetic)

Finnish word games

201

component, to precede the Redundant Default rule. Presumably, the / t a / game on the one hand and the /kontti/ and Pig German games on the other are mutually exclusive. Further, playing the /kontti/ game entails playing the Pig German game; it makes no difference whether this fact is expressed in the form of incorporating the Pig German game rule into the /kontti/ game rule, or by means of rule ordering, as practised here.18

6. CONCLUSION

This paper was concerned with an underspecification analysis of the patterning of vowel harmony in three Finnish word games. The facts were described within an independent game component presumed for the grammar of Finnish. This component was found to contain special rules added to a set of natural language rules. Transparency presented the most interesting cases for rule interaction. The relevant facts strongly suggest that transparent vowels and harmonic vowels receive their default specifications separately.

NOTES 1. Pulleyblank (1986) suggests that the postlexical component serves as input to a separate phonetic component. The distinction between the postlexical and phonetic components will, for the most part, be ignored here. 2. For detailed accounts of the facts of Finnish vowel harmony, including discussion of problems posed for pre-autosegmental, linear phonological theory, see especially Kiparsky (1973b), Anderson (1975, 1980), Ringen (1975, 1980), and Campbell (1980, 1981), among others. 3. However, certain derivational suffixes have back vowels following roots containing only neutral vowels; see for instance Rardin (1969). 4. In prestigious speech styles disharmonic roots of the sort in (3a) can take back harmonic suffixes. In these situations the front vowels / y / and / 6 / , and perhaps / a / as well, seem to behave as if they were transparent; see Campbell (1980, 1981) and Kiparsky (1981). The treatment of such cases is not straightforward and will not be considered here. For some suggestions, in widely different autosegmental models, see Kiparsky (1981), Halle and Vergnaud (1981), Vago (1984), and Steriade (1987b). 5. A notable exception is Steriade (1987b), who allows for the possibility of specifying both values lexically. 6. One might suppose that transparent vowels are in fact targets of harmonic spreading and undergo late neutralization. However, this "abstract" analysis of transparency is not available universally. For discussion, see Vago (1985a, in preparation). 7. The above analysis of disharmony follows that proposed by Kiparsky (1981), with one exception: Kiparsky accounts for the failure of spreading to suffixes in cases like

202

Robert M. Vago

/analyysissa/ by an opacity filter. Steriade (1987b), on the other hand, fills in the [-back] value of front harmonic vowels before spreading [+back] to suffixes; spreading is featurechanging in cases like /syntaksissa/. Invoking the Locality Condition would seem to accord a more principled explanation here. 8. Specifically, under the assumption that [+back] and [-round] are lexical values, and keeping in mind that / i / and / e / are unspecified for [high] and [low], postulating [+high] and/or [+low] lexically has the following consequences: [+high], [-low] leaves / i / and / a / nondistinct; [+high], [+low] leaves A / and / o / nondistinct; and [-high], [+low] leaves / i/ and / i i / nondistinct. 9. Detailed arguments in favor of considering harmonic features to be the property of morphemes (— floating features) can be found in such works as Archangeli and Pulleyblank (1987b) and Vago (1988, in preparation). 10. The term "default" is used here in a generic sense as a cover for Archangeli's (1984) default/complement distinction. 11. Default rules have one further characteristic that is worth attention: they automatically trigger node generation (cf. Sagey 1986). Thus for instance, rules (lle-h) supply the default values of the unspecified vowel / i / as well, even though A / lacks the class nodes which dominate the output features. 12. The rules of the word games discussed in this paper are stated informally. 13. This well-known constraint was originally proposed by Leben (1973). For recent discussion, see McCarthy (1986) and Odden (1986). 14. We might follow the position advocated by Halle and Mohanan (1985) that, unless evidence exists to the contrary, default/redundancy rules (as well as phonological rules) are assigned to the last component of grammar. Accordingly, in the absence of contradictory evidence, the Redundant Default rule (22) would apply in the phonetic rather than postlexical component of Campbell's dialect. 15. An alternative view would be that Seppanen's dialect lacks Distinctive Default, Campbell's dialect Redundant Default. In the latter case, Distinctive Default would reapply automatically in the postlexical component, once a dorsal node is generated for transparent vowels via the application of the default rules in (11). See Pulleyblank (1986) and Archangeli and Pulleyblank (1986, 1987b) for motivating the claim that default rules reapply whenever their structural description is met subsequent to their initial application. 16. See Vago (1985b) for a formal treatment of diphthongs and long vowels in the /kontti/ word game. 17. In the case of back harmonic roots like /kala/, the structural description of Disharmony is met subsequent to the insertion of / t a / . However, the prohibition against crossing association lines appears to be sufficient to prevent the application of Disharmony: +B

N

kala

+B katàìa

+B-B katala

Rather, correct /katala/ is obtained as a result of Back Harmony. Alternatively, a constraint is built into Disharmony to the effect that the [+back] feature is not linked to the right. 18. The analysis of disharmonic roots developed above makes a number of predictions which cannot be tested, due to incomplete data in the source descriptions. For instance, it is expected that / t a / would show up following a monosyllabic root with a back harmonic vowel: here Disharmony is applicable to / t a / . Similarly, in the /kontti/ and Pig German games the front harmonic vowel of a disharmonic root having the vowel sequence /{i, e}...{y, 6, a)...{u, o, a}/ is expected to harmonize to a preposed back harmonic vowel in the initial

Finnish word games

203

syllable: here Disharmony is inapplicable. Should these predictions turn out not to be correct, some form of exception devices, such as extraprosodicity or root-sized exception features, might need to be resorted to.

REFERENCES Anderson, L.B. 1975. Phonetic and Psychological Explanations for Vowel Harmony, Especially in Finnish. Unpublished Ph.D. dissertation, University of Chicago. Anderson, L.B. 1980. Using Asymmetrical and Gradient Data in the Study of Vowel Harmony. In Vago (ed.), 271-340. Archangeli, D. 1984. Underspecification in Yawelmani Phonology and Morphology. Unpublished Ph.D. dissertation, MIT. To be published by Garland Press, New York. Archangeli, D. 1987. Feature Organization: Implications for the Maximal/Minimal Parameter. Unpublished MS, University of Arizona. Archangeli, D. and D. Pulleyblank. 1986. The Content and Structure of Phonological Representations. Unpublished MS, University of Arizona and University of Southern California. To be published by MIT Press, Cambridge, MA. Archangeli, D. and D. Pulleyblank. 1987a. Maximal and Minimal Rules: Effects of Tier Scansion. In J. McDonough and B. Plunkett (eds.), 16-35. Archangeli, D. and D. Pulleyblank. 1987b. Yoruba Vowel Harmony. Unpublished MS, University of Arizona and University of Southern California. To appear in Linguistic Inquiry. Bagemihl, B. 1987a. The Crossing Constraint and 'Backwards Languages.' Paper read at the annual meeting of the Linguistic Society of America, San Francisco, CA. Bagemihl, B. 1987b. Tigrinya Speech Disguise and Constraints on Spreading Rules. WCCFL 6. Campbell, L. 1980. The Psychological and Sociological Reality of Finnish Vowel Harmony. . In Vago (ed.), 245-270. Campbell, L. 1981. Generative Phonology vs. Finnish Phonology: Retrospect and Prospect. In D.L. Goyvaerts (ed.), Phonology in the 1980's. Ghent: Story-Scientia, 147-182. Campbell, L. 1986. Testing Phonology in the Field. In Ohala, J.J. and J.J. Jaeger (eds.), Experimental Phonology. Orlando, FL: Academic Press, 163-173. Clements, G.N. 1985. The Geometry of Phonological Features. Phonology Yearbook 2,225-252. Clements, G.N. 1986. Compensatory Lengthening and Consonant Gemination in LuGanda. In L. Wetzels and E. Sezer (eds.), Studies in Compensatory Lengthening. Dordrecht: Foris, 37-77. Davis, S.M. 1985. Topics in Syllable Geometry. Unpublished Ph.D. dissertation, University of Arizona. To be published by Garland Press, New York. Goldsmith, J. 1985. Vowel Harmony in Khalkha Mongolian, Yaka, Finnish and Hungarian. Phonology Yearbook 2, 253-275. Halle, M. and K.P. Mohanan. 1985. Segmental Phonology of Modern English. Linguistic Inquiry 16, 57-116. Halle, M. and J.-R. Vergnaud. 1981. Harmony Processes. In W. Klein and W. Levelt (eds.), Crossing the Boundaries of Linguistics. Dordrecht: Reidel, 1-22. Hülst, H. van der and N. Smith (eds.). 1982. The Structure of Phonological Representations (Part I), Dordrecht: Foris. Kiparsky, P. 1973a. "Elsewhere" in Phonology. In S.R. Anderson and P. Kiparsky (eds.), Festschrift for Morris Halle. New York: Holt, Rinehart and Winston, 93-106. Kiparsky, P. 1973b. How Abstract is Phonology? In O. Fujimura (ed.), Three Dimensions in Linguistic Theory. Tokyo: TEC, 5-56. Kiparsky, P. 1981. Vowel Harmony. Unpublished MS. Kiparsky, P. 1982a. From Cyclic Phonology to Lexical Phonology. In Van der Hülst and Smith (eds.), 131-175.

204

Robert M. Vago

Kiparsky, P. 1982b. Lexical Morphology and Phonology. In I.-S. Yang (ed.) Linguistics in the Morning Calm. Seoul: Hanshin, 3-91. Kiparsky, P. 1985. Some Consequences of Lexical Phonology. Phonology Yearbook 2, 85-138. Laycock, D. 1969. Sublanguages in Buin: Play, Poetry, and Preservation. Pacific Linguistics (Series A) 22, 1-23. Laycock, D. 1972. Towards a Typology of Ludlings, or Play-Languages. Linguistic Communications (Working Papers of the Linguistic Society of Australia) 6, 61-113. Leben, W. 1973. Suprasegmental Phonology. Unpublished Ph.D. dissertation, MIT. Lefkowitz, N.J. 1987. Talking Backwards, Looking Forwards: The French Language Game Verlan. Unpublished Ph.D. dissertation, University of Washington. Lefkowitz, N.J. and S.H. Weinberger. 1987. The First Branch Principle and Parameter Setting in Language Games: The Case of Verlan. Paper read at the annual meeting of the Linguistic Society of America, San Francisco, CA. McCarthy, J.J. 1979. Formal Problems in Semitic Phonology and Morphology. Unpublished Ph.D. dissertation, MIT. McCarthy, J.J. 1982. Prosodic Templates, Morphemic Templates, and Morphemic Tiers. In Van der Hülst and Smith (eds.), 191-223. McCarthy, J.J. 1986. OCP Effects: Gemination and Antigemination. Linguistic Inquiry 17, 207-263. McCarthy, J.J. and A. Prince. 1986. Prosodic Morphology. Unpublished MS, University of Massachusetts, Amherst and Brandeis University. McDonough, J. and B. Plunkett (eds.). 1987. Proceedings ofNELS 17. Amherst: University of Massachusetts. Mohanan, K.P. 1982. Lexical Phonology. Unpublished Ph.D. dissertation, MIT. Mohanan, K.P. 1986. The Theory of Lexical Phonology. Dordrecht: Reidel. Odden, D. 1986. On the Role of the Obligatory Contour Principle in Phonological Theory. Language 63, 353-383. Pulleyblank, D. 1986. Tone in Lexical Phonology. Dordrecht: Reidel. Rardin, R.B. 1969. On Finnish Vowel Harmony. Quarterly Progress Report (MIT Research Laboratory of Electronics) 94, 226-231. Ringen, C.O. 1975. Vowel Harmony: Theoretical Implications. Unpublished Ph.D. dissertation, Indiana University. To be published by Garland Press, New York. Ringen, C.O. 1980. Finnish Vowel Harmony: A Closer Look. Paper read at the Fourth International Conference of Nordic and General Linguistics, Oslo. Sagey, E. 1986. The Representation of Features and Relations in Autosegmental Phonology. Unpublished Ph.D. dissertation, MIT. Seppänen, J. 1982. Computing Families of Natural Secret Languages: An Exercise in Functional Linguistics. (Research Report No. 23) Helsinki: Helsinki University of Technology Computing Centre. Steriade, D. 1987a. Locality Conditions and Feature Geometry. In McDonough and Plunkett (eds.), 595-617. Steriade, D. 1987b. Redundant Values. In A. Bosch, B. Need and E. Schiller (eds.), CLS 23, 339-362. Vago, R.M. (ed.). 1980. Issues in Vowel Harmony. Amsterdam: Benjamins. Vago, R.M. 1984. Morpheme Level Harmony in a Multi-Leveled Autosegmental Framework. Unpublished MS, Tel Aviv University. Vago, R.M. 1985a. The Autosegmental Analysis of Neutral Segments in Harmony Systems. CUNYForum 11, 103-120. Vago, R.M. 1985b. The Treatment of Long Vowels in Word Games. Phonology Yearbook 2, 329-342. Vago, R.M. 1988. Underspecification in the Height Harmony System of Pasiego. Phonology 5.2.

Finnish word games

205

Vago, R.M. In preparation. Vowel Harmony in Underspecification Theory. Yip, M. 1982. Reduplication and C-V Skeleta in Chinese Secret Languages. Linguistic Inquiry 13,637-661.