Sandhi Phenomena in the Languages of Europe 9783110858532, 9783110098822

Table of contents :
Preface
List of contributors
Introduction: Sandhi
Part I. Descriptive issues
A grammatical hierarchy of joining
A note on Ternes’ paper
Phonological domains
External sandhi rules operating between sentences
The phonological word in Greek and Italian
Stød-sandhi
French external sandhi: the case of liaison
Two cases of external sandhi in French: enchaînement and liaison
Typology of the Celtic mutations
Sandhi in time and space
Les phénomènes de sandhi dans un dialecte bas-francique meridional
Some sandhi-phenomena in the southern Dutch dialects
Some additional data relevant to Stroop’s paper
Moves towards a simpler, binary mutation system in Welsh
On the morphologization of word-final consonant deletion in French
A morphological convergence between consonant liaison and schwa deletion in the Picard and Walloon dialects of French
About Walloon correspondents of French ‘en’ (Lat. inde) : a rule of gemination in Walloon dialects?
Sandhi and prosody: reconstruction and typology
Part II. Germanic
Sandhi im modernen Isländischen
Sandhi in Peninsular Scandinavian
Sandhi phenomena in Frisian
Sandhi im Südniederfränkischen
Sandhi in Swiss German dialects
Celtic
Sandhi phenomena in Irish
Survey of sandhi types in Welsh
A sandhi survey of the Breton language
Romance
Sandhi in Walloon
Catalan sandhi phenomena
Sandhi phenomena in Castilian and related dialects
Portuguese sandhi phenomena
Les phénomènes de sandhi dans le domaine sarde
Sandhi phenomena in Romanian
Slavic
Some sandhi phenomena involving prosodic features (vowel length, stress, tone) in Proto-Slavic, Serbo-Croatian, and Slovenian
Appendix: Guidelines for sandhi surveys
Index of names
Index of languages

Citation preview

Sandhi Phenomena in the Languages of Europe

Trends in Linguistics Studies and Monographs 33

Editor

Werner Winter

Mouton de Gruyter Berlin · New York · Amsterdam

Sandhi Phenomena in the Languages of Europe

edited by

Henning Andersen

Mouton de Gruyter Berlin · New York · Amsterdam

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter & Co., Berlin.

Library of Congress Cataloging in Publication

Data

Sandhi phenomena in the languages of Europe. (Trends in linguistics. Studies and monographs ; 33) Includes index. 1. Phonetics of the sentence. 2. Europe-Languages. I. Andersen, Henning, 1934- . II. Series. P240.S26 1986 414 86-12450 ISBN 0-89925-070-X (alk. paper)

CIP-Kuiztitelaufnahme der Deutschen Bibliothek

Sandhi phenomena in the languages of Europe / ed. by Henning Andersen. - Berlin ; New York ; Amsterdam : Mouton de Gruyter, 1986. (Trends in linguistics : Studies and monographs ; 33) ISBN 3-11-009882-2 NE: Andersen, Henning [Hrsg.]; Trends in lingustics / Studies and monographs

Printed on acid free paper. © Copyright 1986 by Walter de Gruyter & Co., Berlin. All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced in any form — by photoprint, microfilm, or any other means — nor transmitted nor translated into a machine language without written permission from Mouton de Gruyter, a Division of Walter de Gruyter & Co., Berlin. Typesetting: Appl, Wemding. — Printing: Ratzlow-Druck, Berlin. — Binding: Dieter Mikolai, Berlin. - Printed in Germany.

Preface

This volume contains two sets of contributions to the study of sandhi in the languages of Europe. In its first part it presents a series of papers on theoretical issues. Its second part is composed of chapters that survey sandhi phenomena in individual European languages or language areas. The theoretical papers were - with some exceptions - presented in a Workshop on Sandhi Phenomena in the Languages of Europe which was held during the 10th International Congress of Phonetic Sciences in Utrecht, 1-6 August 1983. The workshop was organized by Jadranka Gvozdanovic and the editor of this volume. Their papers could not be presented in the workshop, but are included in this volume. The aim of the workshop was to create a forum in which traditional and recent ideas on the topic could be discussed by specialists who combine an interest in the theoretical questions that are raised by the diverse phenomena called sandhi with expertise in a particular European language or language area. It was the organizers' hope that the discussions in such a forum would serve as a useful review of the questions that have traditionally been asked about sandhi and an evaluation of the answers that have been proposed, and which constitute the received collective understanding of the topic. But beyond that they hoped that a confrontation of different viewpoints and different language specializations represented among the invited participants would lead to the asking of new questions and thus help to focus interest on aspects of sandhi that are not thematicized in current linguistics but cannot in the long run be neglected - particularly questions about the functions of sandhi and about the typology of sandhi rules. Both these kinds of questions arise with particular urgency when one considers sandhi phenomena in different varieties of one and the same language. In this regard, it would seem, the languages of Europe offer a particularly rich fund of documentation since their diachronic development is well known and the study of their diatopic variation well advanced. For this reason it was decided to invite a number of specialists, some of whom were unable to participate in the workshop, to contribute brief surveys of sandhi phenomena in their languages of specialization. These surveys, which compose the second part of this volume, do not

VI

Preface

cover all of Europe. Some major Celtic, Germanic, and Romance areas and one Slavic area are represented. For other areas no specialists were found who would promise a contribution within the time limit that had to be imposed. Perhaps it will be possible in a later volume to fill some of these gaps and, at the same time, to reflect the progress that is to be expected in the understanding of the theoretical issues. In order to promote some measure of uniformity in the survey chapters, the organizers prepared a memorandum on sandhi which was distributed to the authors as part of their guidelines and to the participants in the workshop as well. Since it is referred to in some of the contributions, it is reproduced below as an Appendix (p. 605 f.). In the preparation of this volume I have received encouragement and moral support from Mouton Publishers. It is a pleasant duty to thank them for their patience and guidance. I owe a special debt of gratitude to my friend Dr. Jadranka Gvozdanovic, who helped immeasurably in the planning and organization of the workshop, and to whose enthusiasm and efficiency the smooth running of the arrangements was largely due. It was in no small measure thanks to her that the workshop was a success.

Contents Preface List of contributors Introduction: Sandhi Henning Andersen

V XI 1

Part I Descriptive issues A grammatical hierarchy of joining Elmar Ternes A note on Ternes' paper Hans Basbell Phonological domains Jadranka Gvozdanovic External sandhi rules operating between sentences Irene Vogel The phonological word in Greek and Italian Marina Nespor Stod-sandhi Hans Basbell French external sandhi: the case of liaison Michael Herslund Two cases of external sandhi in French: enchainement and liaison Geert E. Booij Typology of the Celtic mutations Herbert Pilch

11 23 27 55 65 75 85 93 105

Sandhi in time and space Les phenomenes de sandhi dans un dialecte bas-francique meridional Rene Jongen Some sandhi-phenomena in the southern Dutch dialects Jan Stroop Some additional data relevant to Stroop's paper Antonie C. M. Goeman Moves towards a simpler, binary mutation system in Welsh . . . . Gwenllian M.Awbery

117 145 157 161

VIII

Contents

On the morphologization of word-final consonant deletion in French Yves-Charles Morin A morphological convergence between consonant liaison and schwa deletion in the Picard and Walloon dialects of French . . . Yves-Charles Morin About Walloon correspondents of French 'en' (Lat. inde): a rule of gemination in Walloon dialects? Michel Francard Sandhi and prosody: reconstruction and typology Henning Andersen

167

211

223 231

Part II Germanic Sandhi im modernen Isländischen Magntts Petursson Sandhi in Peninsular Scandinavian Stig Eliasson Sandhi phenomena in Frisian Geart van der Meer and Tseard de Graaf Sandhi im Südniederfränkischen Rene Jongen Sandhi in Swiss German dialects William Moulton

251 271 301 329 385

Celtic Sandhi phenomena in Irish Brian 0 Cuiv Survey of sandhi types in Welsh Gwenllian M.Awbery A sandhi survey of the Breton language Jean Le Du

395 415 435

Romance Sandhi in Walloon Michel Francard and Yves-Charles Morin Catalan sandhi phenomena Max W. Wheeler Sandhi phenomena in Castilian and related dialects Ralph Penny

453 475 489

Contents

Portuguese sandhi phenomena

IX

505

Michael Herslund

Les phenomenes de sandhi dans le domaine sarde

519

Michel Contini

Sandhi phenomena in Romanian

551

Andrei Avram

Slavic Some sandhi phenomena involving prosodic features (vowel length, stress, tone) in Proto-Slavic, Serbo-Croatian, and Slovenian

577

Willem Vermeer

Appendix: Guidelines for sandhi surveys

605

Index of names

611

Index of languages

617

List of contributors

Henning Andersen, Institut for Lingvistik, Kabenhavns Universitet, Njalsgade 96, DK-2300 Copenhagen, Denmark. Andrei Avram, Institut de recherches ethnologiques et dialectologiques, Str. Nikos Beloiannis 25, R-70166 Bucarest 22, Romania. Gwenllian M.Awbery, Department of Dialects, Welsh Folk Museum, St. Fagans, Cardiff CF5 6XB, Wales. Hans Basboll, Institut for Nordisk Filologi, Odense Universitet, Niels Bohrs Alle, DK-5000 Odense, Denmark. Geert E. Booij, Vakgroep Taalkunde, Fakulteit der Letteren, Vrije Universiteit, Postbus 7161, NL-1007 MC Amsterdam, Netherlands. Michel Contini, Institut de Phonetique, Universite de Langues et Lettres, Β. P. 25 X, F-38040 Grenoble, France. Tseard de Graaf, Institute of Phonetic Sciences, Groningen University, Groningen, Netherlands. Stig Eliasson, Institutionen för Lingvistik, Uppsala Universitet, Box 513, S-751 20 Uppsala, Sweden. Michel Francard, Lutrebois, B-6652 Villers-la-Bonne-Eau, Belgium. Antonie C.M. Goeman, P.J.Meertens-Instituut, Keizersgracht 569-571, NL-1017 DR Amsterdam, Netherlands. Jadranka Gvozdanovic, Slavisch Seminarium, Spuistraat 210, NL-1012 VT Amsterdam, Netherlands. Michael Herslund, Romansk Institut, Kebenhavns Universitet, Njalsgade 78-80, DK-2300 Copenhagen, Denmark. Reni Jongen, Institut de Linguistique, College Erasme, Place Blaise Pascal, B-1348 Louvain-la-Neuve, Belgium. Jean Le Du, Universite de Bretagne Occidental, Faculte de Lettres et Sciences Sociales de Brest, 20, avenue le Gorgeu, Β. P. 860, F-29279 Brest, France. Yves-Charles Morin, Departement de Linguistique, Universite de Montreal, Case postale 6128, Succursale "A", Montreal, P.Q., H3C 3J7, Canada. William G.Moulton, 318 East Pyne Building, Princeton University, Princeton NJ 08540, U. S.A. Marina Nespor, Italiaans Seminarium, Spuistraat 210, NL-1012 VT Amsterdam, Netherlands. Brian 0 Cuiv, Dublin Institute for Advanced Studies, School of Celtic Studies, 10 Burlington Road, Dublin 4, Ireland.

XII

List of contributors

Ralph Penny, Westfield College, Hampstead NW3 7ST, England. Magnus Pitursson, Institut für Phonetik, Universität Hamburg, Alsterglacis 3, D-2000 Hamburg 36, BRD. Herbert Pilch, Englisches Seminar, Universität Freiburg, Rotteckring 4, D-7800 Freiburg, BRD. Jan P.A.Stroop, Instituut voor Neerlandistiek, Herengracht 330-336, NL-1016 CE Amsterdam, Netherlands. Elmar Ternes, Institut für Phonetik, Universität Hamburg, Alsterglacis 3, D-2000 Hamburg 36, BRD. Geart van der Meer, Department of English, Groningen University, Groningen, Netherlands. Willem Vermeer, Faculteit der Letteren, Rijksuniversiteit te Leiden, Rapenburg 73, Leiden, Netherlands. Irene Vogel, State University of New York, C106 Spaulding Quadrangle (Amherst Campus), Buffalo NY 14261, U.S.A. Max W. Wheeler, Department of Romance Philology, Modern Languages Building, P.O.Box 147, Liverpool L69 3BX, England.

Introduction: Sandhi Henning Andersen In lieu of a definition Ever since its introduction into the terminology of general linguistics a century ago, the term sandhi has had what an anthropologist might call a liminal status. As a technical term without a strict definition it has been used in some kinds of discourse and carefully avoided in others. It has, so to say, been a term, but not quite a term. By coincidence this liminal status of the word is oddly appropriate to its denotation - it refers to liminal phenomena: the junctures between segments, variation and alternations at the boundaries of constituents, or - from another point of view - the interfaces between phonetics and phonemics, and between phonology and morphology, including such truly liminal phenomena as allophones with apparently distinctive function, neutralizations with grammatical function, and so on. It was apparently Georg von der Gabelentz (1891) who first advocated the adoption in general linguistics of the term sandhi as a cover term for all such phenomena, but he did this at a time when the systematic investigation of alternations by Mikolaj Kruszewski (1881) and Jan Baudouin de Courtenay (1895) had already shown that a differentiated conceptual apparatus was a necessity. Sandhi became the general, loose semi-term, a handy label for a diversity of phenomena that individually require more specific names. This is the usage we find in Bloomfield's Language (1935). The term sandhi occurs only a few times in his text. But the lemma sandhi in the index refers to the numerous places in the text where he discusses the plethora of phenomena the term subsumes. These span the gamut from low level phonetics through morphophonemics and lexicalized "included forms" to the expressions of grammatical content in the Celtic initial mutations. This is the extension of the term that has been current since Bloomfield (cf. Crystal 1980, Lewandowski 1975). In this volume two contributions advocate a narrower understanding of sandhi. Ternes would like to limit its meaning to "phonetically motivated alternations, occurring at word boundaries", as in Fr. tasse de cafe

2

Henning Andersen

[zd] (p.16, this volume). Jongen defines sandhi as "all phonological modifications associated with a combination of signs and localized at their boundaries", explicitly excluding discontinuous modification (p. 119 f.). Each of these proposals must be evaluated on its own merits and may well be found useful for language particular purposes. But the general consensus seems to approve of the established usage which includes under the term allophonic variation, neutralization, morphophonemic alternations, however conditioned, as well as internal flection; which recognizes not only segmental, but also prosodic alternations as sandhi, for instance, the Slovak rhythmic law, stress alternations in derivation or inflection, and tone perturbations; which sees no principled difference between continuous and discontinuous conditioning of alternations; and which has no difficulty accommodating vowel harmony or other vowel neutralization phenomena within its compass. Such a broad (and loose) understanding of the term has one advantage over any strict definition, the advantage that has helped the term to survive for so long: it makes the term useful as an informal preliminary label which can be used - unlike any strictly defined term - without prejudging the issues that a given set of data might give rise to. It was in this broad sense the term was used in the call for papers for the sandhi workshop. And it is apparently in this sense it has been understood by most of the contributors to this volume.

Understanding sandhi Like so many other linguistic phenomena, sandhi poses problems of two kinds, problems of description and problems of understanding, which should not be separated, but which - for historical reasons - have in fact been separated for a long time. The strong positivist current in early 20th century linguistics - which was rejected by many European structuralists, first and foremost the linguists associated with the Prague school - continued unchecked through Bloomfield and the post-Bloomfieldians into modern American linguistics. It is reflected in the traditional emphasis on description, and - as far as sandhi is concerned - it is to this current we owe the progress in the development of our descriptive apparatus, from Bloom-

Introduction: Sandhi

3

field's taxonomy of alternations and the later debates over Item & Arrangement vs. Item & Process descriptions, over the rule systems of systematic phonetics to lexical phonology. But the positivist heritage is reflected, as well, in the ways in which we talk about alternations, manners of speaking which involve metaphors that are deeply alien to the nature of language. One obvious example is the appeal to causation in the description of alternations. We continue to speak of segments "causing" other segments to be pronounced thus or so, or to be replaced or deleted - as if we were dealing with natural phenomena. Or we use less crass, but equally causational expressions, such as the late positivist "conditioning" or the machine term "triggering". Nobody, of course, understands these expressions literally in a causal sense. They are accepted because they are useful for descriptive purposes. But one can guess that it is precisely their efficacy in this regard which permits them to present phenomena as adequately accounted for even when they have not been understood. Another example, of more recent vintage, is the "rewrite rule", which presents the picture of an entity being replaced by another entity and thereby, as it were, disappearing and, at least ideally, being unavailable for later processing further down the assembly line. We speak in this connection of "representations" - underlying, intermediate or surface but the notion of replacement inherent in the concept of the rewrite rule obscures the fact that "representation" is derived from a three-place predicate: something represents another thing to somebody. In reality when a thing is assigned a new representation, the thing that is represented does not thereby cease to exist. It seems clear that the emphasis on efficacy of description has tended to promote a reiflcation of representations which belies their true nature of signs. It was precisely the understanding of language as a system of signs that made the European structuralists turn away from the mechanistic descriptions and causal explanations of the positivists. Bühlers' (1934) famous dictum "Alles an der Sprache ist Zeichen" sums up the guiding principle for the Prague school's investigations of the sound aspect of language. The phonological studies of, for instance, Lazicius, Jakobson, and Trubetzkoy show how fruitful was this research project. Every aspect of the speech signal was viewed in semiotic terms, and physiognomic signs, social, expressive, conative, and auxiliary-sociative signs were identified (Jakobson 1962 a, Trubetzkoy 1958). Jakobson showed that distinctive

4

Henning Andersen

features are signs (1962b). And Trubetzkoy's works contain important fragments of a semiotic theory of phonotactics. Prague school phonology remained in many respects a torso. But when in current discussions questions of the function of sandhi phenomena come to the fore - as they do, in several of the contributions in this volume, for instance those by Andersen, Basboll, Contini, Gvozdanovic, Morin, Penny, and Vogel - it seems appropriate to remember that these questions have been addressed before. If the answers that were proposed were not entirely satisfactory and could not be integrated into a coherent theory, this is because the time was not yet ripe, and not because the questions were posed from the wrong point of view. If everything in language is semiotic, an exegetic theory of sandhi germane to the nature of language must explicate sandhi phenomena as signs, that is, determine what kinds of content (or meaning) is encoded in these elements of expression (or form). It may well be that the time for such a theory is still not ripe. But it seems that parts of such a theory can be sketched, and for the sake of argument I will present here some thoughts on what one may call phonological sandhi. I limit my topic by leaving out of consideration first the truly morphological alternations traditionally called internal flection (cf. Appendix, p. 608; and the papers by Ternes, Morin, Penny, Awbery, Le Du, 0 Cuiv below). Secondly I leave out non-automatic alternations, whether conditioned by lexical or grammatical content or by combinations of content features and phonological features. I have offered a survey of the varieties of such alternations, couched in semiotic terms, but viewed in a dynamic perspective, in Andersen (1980). Thirdly I leave out the low-level processes that generate the phonetic texture of utterances, rules of syllabication, foot formation, etc., which are touched on below especially by Gvozdanovic. My topic will thus concern roughly what would traditionally be termed allophonic variation and automatic alternations. These types of sandhi I assume to be strictly phonological in the sense that both allophonic variation and neutralization imply reference to a language particular structure of phonological signs with distinctive (or diacritic) value. The existence of such language particular structures is indicated by the fact that phonological sandhi rules - despite the universal character of "natural" phonetic processes - are in fact language particular (cf. Gvozdanovic below).

Introduction: Sandhi

5

I find that one can define two sorts of function for phonological sandhi rules. They have a systemic function, which is iconic in that they produce distributions of feature values in utterances which reflect (as diagrammatic icons) the distinctive or allophonic value of the features in question and the markedness relations that hold between different values of the same feature (opposition). That distributional facts in this way reflect the structure of a distinctive feature system is nothing new. Linguists have traditionally exploited such facts in their analyses. But one may presume that such distributional facts are important as well in the real life of languages, specifically in language acquisition, where they offer the language learner essential information about the system that has to be inferred. Their function in this regard can be characterized as metaphonological (cf. Andersen 1979). Besides their metaphonological function, however, phonological sandhi rules have a textual function, which I will try to adumbrate here.

Textual function

Sandhi operates on different levels of representation, but its textual function on each level is to signal text cohesion. This is illustrated very nicely in the studies by Vogel and Jongen in this volume. As Vogel shows, the linking-/· in r-less dialects of English may occur between sentences and, when it does, signals that they are pragmatically, semantically and/or syntactically connected (p.60ff.). Also Avram's survey is relevant in this regard; (cf.p. 565). Jongen shows that sandhi rules may apply utterance-initially in an answer, to signal its cohesion with the interlocutor's question (p. 127 f.). At the opposite end of the scale, Timberlake (1978) has shown in an analysis of some Polish dialect data that the introduction of allophonic palatalization of velars before i and e is more consistent and more advanced when velar and vowel belong to the same morpheme than when they are separated by a desinence boundary, and more consistent and more advanced when the vowel after the velar is stable than when it alternates with zero. Here is a rule which is in the process of being established, and which by its sensitivity to the difference between uniform and alternating environments illustrates how an allophonic feature

6

Henning Andersen

which links one segment to the following segment can signal different degrees of cohesion within the (morphological) word. Between the two extremes illustrated by the English linking-/· and the Polish "linking palatalization" we find a gamut of phonological sandhi rules which produce signs of text cohesion of different kinds. They produce these signs by applying to environments which include reference to the boundaries of phonological or syntactic constituents. Any rule of the general form X->- Ύ/Ζ establishes a sign, Y, which at one and the same time represents X and Z. One may say that it stands for X, but that it points to Ζ (regardless of the overt manifestation of Z) by virtue of the reference to Ζ in the rule. It is an index of Z. Basboll's contribution below offers an account of the amount of information carried by such indexes in Danish stod sandhi (p. 82 f.). Eliasson's paper explicitly discusses the retrievability of distinctive feature information in Swedish and Norwegian sandhi (p.284ff.). But by and large one must acknowledge that phonological sandhi operates to distribute redundant features - either allophonic features or realizations of neutralized oppositions - in such a way that they index phonological or syntactic domains. In the case of neutralizations it is clear that redundancies in underlying representations are sacrificed and exploited for the creation of signs of textual cohesion. In my paper below (p. 245 ff.) I mention three aspects of cohesion that seem to be signaled by phonological sandhi. Rules that apply within a domain irrespective of boundaries within this domain serve an integrative function. They produce signs of the internal cohesion of the given domain. Rules that apply at boundaries may serve a concatenative function, if they produce signs that link elements together across the given boundaries. Or they may have a delimitative function if they produce signs that do not. In the paper below I illustrate these different aspects of phonological cohesion with types of "voicing sandhi" found in the Slavic languages. But other, perhaps more obvious examples are not difficult to find. Russian vowel reduction, by which vowel distinctions are neutralized in unstressed syllables within the phonological word irrespective of internal boundaries, is a clear example of a complex of phonological sandhi rules with an integrative function. Turkish vowel harmony, by contrast, which links morpheme to morpheme within the phonological word, is produced by sandhi rules with a concatenative function. It is evident that sandhi rules operating on different phonological features in the same language may have different cohesive function. Their

Introduction: Sandhi

7

combined effect is to create phonological texture in a sense comparable to the one in which the term texture is used in reference to discourse (cf. Halliday - Hasan 1976). Saying that phonological sandhi serves to signal text cohesion is tantamount to claiming that it is the phonological counterpart of the text grammatical phenomena usually referred to by this term. For those who are interested in homologies between different levels of language it may be interesting to note that the three aspects of phonological cohesion identified in the course of an analysis of Slavic sandhi phenomena seem to correspond to the three main devices for participant tracking in discourse, reflexivization, anaphora, and switch reference. Reflexivization applies within the domain of the sentence and is integrative. Anaphora and switch reference apply to mark continuities or discontinuities of participants between sentences and serve functions that may be called concatenative and delimitative. This parallelism may strike some as a coincidence. But the deeper correspondence between the devices of the two levels of language gives pause to thought. In discourse, cohesion is achieved, inter alia, by reducing the descriptive content of a co-referring expression - whether by lexical substitution, by pronominalization or by ellipsis. Thus emptied of (part of) its content, it becomes a two-fold representation. It stands for the full content of the term it replaces and - provided it is thus interpreted - it points back or forward to the contiguous environment relative to which this substitution is licensed or obligatory. Similarly in neutralization, when, say, the final k in Pol. jgzyk ainu [-ga-] 'the Ainu language' loses its specification [ — voice] and is assigned the realization [-(-voice] before initial vowel. The reduced final segment of jpzyk stands for a k and - if it is thus interpreted - points to the neighboring word boundary. Admittedly, in comparisons of this kind there are always a great number of mutanda, and the account given here is sketchy and informal. But to the reader who recognizes the importance of elucidating the similarities between the different planes of language it will suggest that sandhi rules are not just a peculiar encumbrance which obscures the underlying representations of morphemes and is counterproductive from a communicative point of view. It is hardly the case - as one participant in the sandhi workshop exclaimed during the discussion - that languages "would be much better off without them". Sandhi rules may be understood as a way of utilizing redundancies in lexical representations for the production of subsidiary signs of cohesion. And they may some day

8

Henning Andersen

- perhaps on a par with the lexical and grammatical devices for text cohesion - be shown to conform to more general strategies in human semiotic behavior.

References Andersen, Henning 1979

'Phonology as semiotic', in: A semiotic landscape. Proceedings of the First International Congress of the International Association of Semiotic Sciences, ed. by Seymour Chatman (The Hague: Mouton), 377-381. 1980 'Morphological change: towards a typology', in: Historical morphology, ed. by Jacek Fisiak (The Hague: Mouton), 1-50. Baudouin de Courtenay, Jan I. 1895 Versuch einer Theorie phonetischer Alternationen. Ein Kapitel aus der Psychophonetik (Straßburg: Trübner). Bloomfleld, Leonard 1935 Language (London: Allen & Unwin). Bühler, Karl 1978 Sprachtheorie. Die Darstellungsfunktion der Sprache (Stuttgart: Fischer). Crystal, David 1980 A first dictionary of linguistics and phonetics (London: Deutsch). Gabelentz, Georg von der 1891 Die Sprachwissenschaft. Ihre Aufgaben, Methoden und bisherigen Ergebnisse (Leipzig: Tauchnitz). Halliday, Μ. Α. K. - Ruqaiya Hasan 1976 Cohesion in English (London: Longman). Jakobson, Roman (with Morris Halle) 1962 a 'Phonology and phonetics', in: Jakobson (1962 c: 464-504). 1962b 'Zur Struktur des Phonems', in: Jakobson (1962c: 280-310). 1962 c Selected writings, I. Phonological studies (The Hague: Mouton). Lewandowski, Theodor 1973-1975 Linguistisches Wörterbuch, 1-3 (Heidelberg: Quelle & Meyer). Kruszewski, Mikolaj 1881 Über die Lautabwechslung (Kasan: Universitätsbuchdruckerei). Timberlake, Alan 1978 'Uniform and alternating environments in phonological change', Folia Slavica 2: 312-328. Trubetzkoy, N. S. 1958 Grundzüge der Phonologie (Göttingen: Vandenhoeck & Ruprecht).

Parti Descriptive issues

A grammatical hierarchy of joining Elmar Temes

"· · · f a c t u a · simplicity has often been achieved at the price of conceptual confusion." (W.S.Allen 1972:5)

0. Introduction The word 'sandhi' comes from the Sanskrit samdhi and simply means 'putting together' (Bloomfield 1935: 186, fn.l) or 'joining'. In Sanskrit grammar, two types of sandhi had already been distinguished, namely 'internal sandhi', which occurs at morpheme boundaries and results from the juxtaposition of morphemes within words, and 'external sandhi', which occurs at word boundaries and results from the juxtaposition of words within phrases or sentences. In 19th and 20th century linguistics, including present-day usage, 'sandhi' has become a cover-term for a veritable host of the most divergent phenomena, among them assimilations or dissimilations of all kinds, allomorphic or morphophonemic alternations, atonic forms and proclitics of various sorts, elision (as in French I'homme), synaloepha (as in Spanish la amiga), French liaison, French enchainement, Celtic mutations, Italian raddoppiamento sintattico, Tuscan gorgia, English short forms (as in we're), Notker's Anlautgesetze in Old High German, and many others. Most of these phenomena have, at some stage of the history of their linguistic processing, vaguely been attributed to an endeavor of achieving 'ease of articulation' (e.g. Lewandowski 1976: 589), or even, horribile dictu, 'euphony' (e.g. Marouzeau 1961: 91, with reference to French t in va-t-il). A glance at some glossaries of linguistic terminology (e.g. Crystal 1980, Dubois et al. 1973, Lewandowski 1976, Marouzeau 1961, Pei 1966) confirms that the term 'sandhi' is indeed understood in a very broad sense. One may even quote in support, no less a person than Bloomfield (1935:186ff.) who defines sandhi as "[f]eatures of modulation and of phonetic modification ... in many syntactic constructions" (186) and cites a number of rather divergent examples. As a typical specimen from a recent 'dictionary of linguistics and phonetics', one may ad-

12

Elmar Tentes

duce Crystal (1980: 311), who defines sandhi as follows (capitalization in the original): "[a] term used in SYNTAX and MORPHOLOGY, to refer to the PHONOLOGICAL MODIFICATION of GRAMMATICAL FORMS which have been juxtaposed ... ASSIMILATION and DISSIMILATION are two widespread tendencies which could be classified under this heading. The merit of the sandhi notion is that it can be used as a very general term within which can be placed a wide range of structural tendencies that otherwise it would be difficult to inter-relate." In accordance with Allen (1972: 5) however, whose statement has been chosen as a motto for this paper, the present writer is inclined to regard the very generality of the sandhi notion in linguistic writing as a disadvantage rather than as a 'merit', if indeed conceptual clarity is a desirable goal in linguistic description. It therefore seems to be a mandatory prerequisite for a 'workshop on sandhi phenomena' to try to achieve more conceptual clarity by delimiting various types of 'joining'-phenomena, by arranging them with respect to their position within the hierarchy of grammatical description, and by assigning a non-ambiguous term to every one of those types. This will be the goal of the following pages. 'Joining', which is the literal translation of 'sandhi', implies that in fact any kind of syntagmatic (as opposed to paradigmatic) feature could be classified under this heading. We shall take up for discussion various syntagmatic alternations, beginning on the lowest rank of grammatical hierarchy and gradually working up the scale.

1. Allophonic variation On the lowest level of joining is the phenomenon known in phonetics as 'coarticulation'. The articulations of any two (or more) succeeding phones usually overlap in such a way that the following phone retains an articulatory feature of the preceding one and/or the preceding phone anticipates an articulatory feature of the following one. This is an inevitable consequence of the inertia of the articulatory organs. It may have more or less clear effects. In many cases, it goes without even being noticed by the average speaker or hearer. But in other cases, it produces easily noticeable allophones, as in German Nacht [naxt] 'night', as op-

A grammatical hierarchy of joining

13

posed to nicht [ni^t] 'not', where [x] retains the feature of velarity of the preceding [a], and [?] the feature of palatality of the preceding [i]. Both [x] and [9] represent positionally conditioned allophones of one and the same phoneme /x/ in German. This kind of joining, which takes place at the subphonemic level, has hardly ever been subsumed under the notion of sandhi. It will suffice to retain the terms 'coarticulation' for purely phonetic purposes, and (positionally conditioned) 'allophonic variation' for phonemic description. Whereas the variation in German between [x] and [5] is readily explicable in terms of articulatory conditions, a variation such as the one of British English /I/ in feel[fl:l] vs./ee/wg[fi:liq] (i.e. β] in word-final position, [1] in intervocalic position) has no immediate articulatory motivation. It may therefore be useful to distinguish between phonetically transparent and non-transparent allophonic variations.

2. Phoneme distribution The next step on the scale of grammatical hierarchy leads on to the phonemic level. Every language has specific rules governing the stringing together of phonemes within words. It is important to note that at this stage, we consider only strings of phonemes that do not contain a morpheme boundary. In Spanish, among the consonant clusters permitted in word-internal position are /mb/ as in ambos /ämbos/ 'both', and /nd/ as in onda /onda/ 'wave'. The clusters */md/ or */nb/ are not permitted. This is again immediately explicable in terms of articulatory conditions: the clusters /mb, nd/ are homorganic, whereas */md, nb/ are not. Restrictions of this kind are of course language-specific, as can be seen from the German Hemd /hemt/ 'shirt', pi. Hemden /hemdan/, where /md/ is permitted as an internal cluster. This phenomenon as well has hardly ever been associated with the notion of sandhi. It is satisfactorily treated under the heading phonotactics or phoneme distribution. Restrictions as to the occurrence of phoneme clusters may again be phonetically transparent or not. Whereas Spanish /mb, nd/ and nonpermitted */md, nb/ are readily explicable in articulatory terms, there is no immediate motivation for the occurrence of German initial /Jtr/ as

14

Elmar Tentes

in Straße/Jtra:s9/ 'street' to the exclusion of */str/, and exactly the reverse situation in English, where one has initial /str/ as in street /stri :t/ to the exclusion of */ftr/. It may therefore be useful to make a distinction between transparent and non-transparent restrictions in phoneme distribution.

3. Morphological variation At the next stage, it is again sequences of phonemes within words that are considered, but this time sequences containing a morpheme boundary, i. e. resulting from morphological processes. In English, the plural suffix {-s} appears as / - s / after voiceless consonants as in cats /kasts/, and as I-τ! after voiced consonants as in dogs /dogz/. Consonant clusters such as */tz, gs/ are not permitted in English. In Spanish, the verbal prefix {en-} appears as /en-/ before dentals as in endurecer /endureGer/ 'to harden', and as / e m - / before labials as in enviar /embiär/ 'to send'. As already noted under (2), the consonant clusters */md, nb/ are not permitted in Spanish. Processes of this kind are sometimes referred to as 'sandhi' or, more specifically, as 'internal sandhi' (in German also 'Wortsandhi', as opposed to 'Satzsandhi' (4)). This is in conformity with usage in the grammar of Sanskrit. On the other hand, it is probably more customary to refer to these very processes as morphological (allomorphic) variation or, under specific conditions, as morphophonemic variation. Since the term allomorphic variation (which includes morphophonemic variation) is very common and well established indeed, there is no need for a competitive term such as (internal) sandhi. It is therefore suggested not to make use of the notion of sandhi at all, in connection with purely morphological processes within word boundaries. We thereby avoid the competition of two terms for one and the same phenomenon. At the same time, we avoid the necessity for the awkward distinction of internal vs. external sandhi. And we gain in conceptual clarity by reserving the term 'sandhi' for one specific and clearly defined phenomenon only, viz. the one described under (4). Case (3) is basically different from cases (1) and (2) because, at this level of description, both phonetic/phonemic structure and grammatical (in this case morphological) structure of the language in question are

A grammatical hierarchy of joining

15

involved. A sequence like English /ks/ in rocks /roks/ has both phonetic shape and grammatical information, whereas the same sequence in box /boks/ has phonetic shape only. It is therefore important to make a distinction here between phonetically motivated1 and non-motivated allomorphic variation. Both examples at the beginning of this section (English plural {-s}, Spanish verbal prefix {en-}) have an obvious phonetic motivation: in the first, there is a harmonization of the feature [voice], in the second a tendency towards homorganic consonant clusters. An example of phonetically non-motivated morphological variation is consonant gradation in Finnish: Nom.

kukka /kukka/ 'flower' luku /luku/ 'number' jalka /jalka/ 'foot'

Gen.

kukan /kukan/ luvun /luvun/ jalan /jalan/

Another example is from Breton. One type of inflectional prepositions has alternation of / d / and / t / , as shown in the following partial paradigm of evid/evi:d-/ 'for': Sg. 1. evidon /evi:don/ 'forme' - 3. m. evitan /evitä/ 'forhim' 2. evidout /evi:dut/ 'for thee' f. eviti /eviti/ 'for her' It does not seem to be customary to distinguish consistently between the two types of morphological variation. Although it would certainly be useful to have a terminological distinction, we do not venture to innovate at the moment. To some extent, it seems that 'morphophonemic variation' is preferably, but not consistently, used for non-motivated morphological variations. For the time being, it may suffice to distinguish simply between motivated and non-motivated morphological alternation. The notion of sandhi however should be avoided in either case.

4. Sandhi Moving further up the hierarchy of grammatical description, we now leave the domain of the word and come to processes that occur when two or more words follow each other in a phrase or sentence. In this case, in many languages, certain phonetic or phonemic alternations take place, mainly at word boundaries. The following two examples are taken from French:

16

Elmar Tentes

une fasse /tas/ + de /da/ cafe—nine lasse de /tazda/ cafe 'a cup of coffee'. pas de /da ~ d / + chance /Jäs/—•/»as ife chance /pat/äs/ 'no luck'. In both examples, there is regressive assimilation of the feature [voice] across word boundaries. Assimilations of this kind may also work in a forward direction (i. e. progressive assimilation), as in the following example from Breton: bennoz /benos/ 'blessing' + Doue /du:e/ 'Go d'-^bennoz Doue /benos tu:e/ 'God's blessing'. Among the languages that make extensive use of alternations of this kind are the Romance languages (especially Portuguese) and the Celtic languages (especially Breton, see Ternes 1970:68-110). German, on the other hand, lacks such features to a great extent.2 This phenomenon is known as 'sandhi' or, more specifically and in accordance with the grammar of Sanskrit, as 'external sandhi' (in German also 'Satzsandhi', as opposed to Wortsandhi (3)). Since, as suggested under (3), the term 'sandhi' should be avoided in connection with morphological processes, the concept of sandhi (reference to 'external' being redundant) should be restricted to phonetically motivated3 alternations occurring at word boundaries, when two or more words are being pronounced in succession rather than in isolation (cf. the definition in Ternes 1970: 68). Indeed, the term 'sandhi' serves best here, because no other specific term has developed in linguistic description for this very phenomenon, except perhaps the rather vague heading 'sentence phonetics' which also embraces other features.

5. Initial mutations The next step leads on to the famous 'initial mutations' of Celtic. There is no room here for describing the phenomenon in detail. We refer instead to Ternes (1977). A typical example is taken from standard Breton. The lexical entry penn /pen/ 'head' changes its initial consonant as shown in the following paradigm of the possessive phrase: Sg.l.

vafenn /vafen/ 'my head' PI. 1. horpenn /horpen/ 'our head'

A grammatical hierarchy of joining

2.

da benn /da ben/ 'thy head'

3 . m . / e b e n / f. hefenn4 / e fen/

'hishead' 'her head'

2. hopenri4 / o pen/ 3.ofenn

/ofen/

17

'your head' 'their head'

Thus, / p / alternates with / f / and / b / . It should especially be noted that alternating consonants in Celtic mutations always have the status of phonemes in the respective language, not of allophones. Minimal pairs demonstrate the phonemic status of / p / , / f / , and / b / in Breton: plac'h /pla:x/ 'girl' vs.flac*h /fla:x/ 'palm of hand' per /pe:r/ 'pears (coll.)' vs.ber /be:r/ 'spit (for roasting meat)' fank /fank/ 'mud' vs. bank /bank/'bench' The above example makes clear that there is no phonetic or phonemic motivation whatsoever for the alternation of the initial consonant. Conditioning is of a purely grammatical nature. Within that frame, it may be morphological, syntactical, or lexical. Generally speaking grammatical characteristics (such as person, number, gender etc.), not phonetic or phonemic shape, determine whether a word undergoes a mutation or not and, in the former case, which type of mutation applies. Diachronie excursus The origin of the Celtic mutations is well known from historical and comparative observations and reconstructions. They originated from sandhi phenomena as defined under (4). Since it is the initial of the second element that is affected, assimilation in sandhi must have been progressive, as in the Breton example under (4). In the course of the historical development of the language in question, the final phoneme(s) of the preceding word dropped due to regular sound changes. But - and this is the crucial point - the changes that the disappeared phoneme(s) had induced on the initial of the following word remained. From this moment on, the conditioning of these changes was not phonetic any more, but grammatical. An example of this development is described in Ternes (1977: 27-28). Another, more straightforward, example is the following one taken from Welsh. In Welsh, the conjunction a 'and' causes a mutation by which initial /p, t, k/ are changed into /f, Θ, x/ respectively, e.g. tad/ta:d/ 'father', but mam a thad /mam a 9a:d/ 'mother and father'. From historical reconstruction, it is evident that modern Welsh α 'and' derives from ProtoWelsh *ak. The phrase as a whole may be reconstructed as *mammä ak

18

Elmar Ternes

tatos. The original consonant cluster *kt developed regularly into modern Welsh /Θ/, both within words (cf. Welsh wyth /ui0/ 'eight' vs. Latin octo) and across word boundaries. Then, final *-k dropped. But /Θ/ remained as the initial of the following word, as if final *-k were still present. Since final *-k is not recoverable syntagmatically in modern Welsh, the alternation Λ / - / Θ / is not phonetically conditioned any more, but grammatically. - End of diachronic excursus.

Mutations are very common in all Celtic languages. Within the grammatical system of every single language, the mutations form a highly complex system. But as mentioned in Ternes (1977), the Celtic languages are by no means unique in this respect. Similar phenomena are to be found in other languages as well, especially in West African languages such as Fula. It has been shown likewise in Ternes (1977) that French liaison and Italian raddoppiamento sintattico can be interpreted along similar lines as the Celtic mutations.5 Celtic mutations are sometimes misinterpreted by linguists who have no first-hand knowledge of the Celtic languages. Features similar to Celtic mutations, which do exist in other languages, have not always been investigated thoroughly enough as yet. Those features which have been investigated are usually seen isolated within one specific language family. Celtic scholars as a rule have no knowledge of the grammatical structure of West African languages and vice-versa. It is therefore difficult to suggest a terminology that would be acceptable for all languages involved. There is no doubt however that phenomena of this kind have become famous among linguists mainly from their occurrence in the Celtic languages. It therefore does not seem unreasonable to use the term that has already been common in Celtic studies for a long time, for other languages as well. We consequently suggest calling 'mutations' or, more precisely, 'initial mutations' any kind of grammatically and/or lexically conditioned alternation of word initials. It is particularly unfortunate to extend the notion of sandhi in such a way as to include Celtic mutations, as has frequently been done, since this conceals the very nature of the mutations. It is particularly important to make a clear distinction between phonetically conditioned alternations at word boundaries (i.e. sandhi as under (4)), and grammatically conditioned alternations (i. e. mutations).

A grammatical hierarchy of joining

19

6. Incorporated mutations That 'sandhi' is indeed not a suitable designation for the Celtic mutations also appears from their further historical development. Some mutations may even go one step further in the process of grammaticalization. Although they are in no way conditioned phonetically, it may be said that most mutations are so to speak 'triggered' by specific morphological or syntactical forms or constructions. These forms or constructions in most cases immediately precede the word which undergoes mutation. Some mutations of Celtic however may also occur on their own. In this case, the mutation by itself transforms one grammatical form into another one, without any triggering element being present. Two examples of this are given in Ternes (1977: 25). In the spoken Celtic dialects, this tendency is much more wide-spread than orthographical representations would suggest. In standard Breton, there is a regular mutation ki /ki:/ 'dog' vs. ar c{hi /ar xi:/ 'the dog', triggered by the definite article when used before a masculine noun in the singular. The corresponding forms in the Breton dialect of Scaer (Sud-Finistere) are /ki/ 'dog' vs. / ς ϊ / 'the dog'.6 The article itself has dropped and is not recoverable synchronically in the dialect. Absence or presence of the mutation alone determines absence or presence of 'definiteness'. Similar examples are - I repeat - much more common in the spoken Celtic dialects than one seems to have hitherto recognized. In cases like this, mutation has even left the domain of 'joining' altogether, because there is nothing preceding or following that could in any way syntagmatically be related to it. It therefore seems justified to use a different term for mutations of this kind. In Celtic studies, it has been suggested by Oftedal (1962) to use 'projected mutations' for the 'triggered' type of mutations, and 'incorporated mutations' for the latter type (cf. also Ternes 1977:25). Thus, a joining phenomenon which no doubt began by allophonic variation has gradually worked up the grammatical scale by becoming phonemic in the first place, then - losing its phonetic conditioning - becoming grammatically conditioned, and eventually leaving the domain of 'joining' altogether by forming a grammatical process on its own resources.

20

Elmar Ternes

7. Final remarks Going further up the scale of grammatical hierarchy, one finally comes to the traditional domain of syntax, where there are joining phenomena such as 'concord' and 'government'. But these are beyond the scope of the present paper. We recall to mind that it has been advanced as an advantage of the sandhi notion "that it can be used as a very general term within which can be placed a wide range of structural tendencies ..." (Crystal 1980: 311). Some of the different categories of joining phenomena we have been trying to delimit, do indeed overlap in certain languages. In Spanish, for example, voiced spirants as realizations for voiced stop phonemes (i. e. [β] for / b / , [d] for / d / , and [γ] for /g/) occur within morphs, at morpheme boundaries, and also at word boundaries. As a consequence, this could be seen as an instance of allophonic variation (1), but also of sandhi (4). Again in Spanish, the restrictions for the occurrence of consonant clusters, as described above, are identical both within morphs and at morpheme boundaries, and may therefore be seen as an instance of phonotactics (2), but also of allomorphic variation (3). A similar situation led to the use of 'sandhi' in Sanskrit grammar for both allomorphic variation (3) and sandhi proper (4). Overlappings of this kind are, however, strictly language-specific. In other languages, the categories that overlap may be different, or there may be no overlapping at all. The case of Spanish voiced stops means that, in this case, there is no phonetic word-juncture. This is a typological feature of Spanish (as indeed of most Romance languages). Germanic languages, on the contrary, are usually different in this respect. A counter-example is provided by Icelandic: The phone [ö] is an allophone of /Θ/ occurring, among others, in intervocalic position, e.g. pyding /QvMrfti/ [6i:öii]k] 'significance'. In connected speech, word initial /Θ/ remains [Θ], also after a preceding word ending in a vowel.7 The occurrence of intervocalic [Θ] therefore indicates word initial position in Icelandic and serves as a marker of word-juncture. The occurrence of various overlappings should be no alibi for lumping together under one heading the most diverse kinds of joining phenomena. For the sake of conceptual clarity (see motto), the six categories of 'joining' treated in the body of this paper should in principle be kept apart as a minimum. This does not exclude that, for the needs of the description of specific languages, it may be appropriate to make use of various cover-terms.

A grammatical hierarchy of joining

21

Notes 1. Whereas under (1) and (2), the term 'transparent' is used, 'motivated' is preferred under (3). This is in line with the superposition of phonetic shape and grammatical information in the latter case. There may well be detected, at closer investigation, a phonetic motivation for the occurrence of P] in word-final position in British English, although it is not immediately transparent. On the other hand, the alternation /kk/ ~ / k / in the Finnish example below has no phonetic motivation on the synchronic level, although it may well be transparent what has happened diachronically (in this case an influence of the opening and closing of syllables). The alternation /kk/ ~ / k / is grammatically motivated (nom. vs. gen. in specific inflectional paradigms). 2. This applies to the standard language. Some regional dialects are different in this respect. 3. See note 1. 4. h in he, ho is merely orthographical to avoid homography. Thus e and he, ο and ho are perfectly homophonous. In these cases, the distinction of person is shown by the type of initial mutation alone. 5. For the basically different nature of Tuscan gorgia, see ibid. 6. The author's observation in the field. 7. This applies to words under sentence stress, such as nouns and verbs. Words which do not normally receive sentence stress, such as adverbs and pronouns (e.g. pad 'it, that'), may be used enclitically and then have [d] for initial p.gerdu£a re layv (Here 'ω' denotes the prosodic word, 'Σ' denotes the foot, and 'σ' the syllable).

38

Jadranka Gvozdanovic

Similar cases of morphologically conditioned change of the prosodic structure have been discussed by Kiparsky (1983), who proposes viewing the English affixes causing an accent shift onto the stem as ordered at a different, higher, level than those not causing it. An example of the first type is the suffix -al in English, and of the second type, -ize. (5) Foot assignment according to Kiparsky (1983): ω ω ω I I F Fw Fs F A \ A /X s w \ s w s w parent+al - parental ; standard+Ize - standardize (Kiparsky omits the syllable symbol 'σ', and denotes the foot by means of 'F'). It follows from the accounts of the English accent system mentioned above that in addition to a large amount of predictability, given the tense vs. lax vowel specification, the syllable structure and the boundaries involved, the crucial points of the prosodic structure still have to be specified in the lexical representation. The lexical representation must contain information on whether the feet of a prosodic word are mutually arranged as strong - weak or weak - strong (cf. relayN vs. relayv, and ally vs. ally, but Dundee vs. nightingale, with (') denoting primary, and (') secondary accent). And not only that. In Selkirk's approach, the lexical representation must also contain information on whether the feet are monosyllabic or bisyllabic (cf. gymndst vs. modest). Let us examine Selkirk's statement that both modest and gymndst contain a lax vowel in the second, closed, syllable (henceforth: a lax closed second syllable). It is the weak syllable of a bisyllabic foot in modest, and a monosyllabic foot in gymndst. Let us examine this statement by comparing gymndst with reldy, the latter, according to Selkirk, with an underlying tense vowel in the first syllable, because of its being tense in rildy. Why not then compare gymndst with gymndsium, the latter with a tense vowel in the second syllable? If we compare gymndst with gymnasium, then the second syllable of gymndst can be viewed as unspecified for [±tense] due to the, lexically specified, accent shift. According to a general markedness convention, a vowel unspecified for [±tense] shows up in such cases as lax, and phonetically lengthened in a closed syllable at a prosodic word boundary. This is, then, the difference between the second syllable of modest and that of gymndst: the former is [—tense],

Phonological domains

39

whereas the latter is unspecified for [±tense]. There is apparently a regularity in English by which a syllable which is unspecified for [±tense] due to a prosodic shift forms a foot of its own at a prosodic word boundary. This can account for the secondary accent in gymndst, due to the monosyllabic foot there, but why can it not account for reläy, where the so-called 'defooting' occurs? In my opinion, the reason for this can be seen in the restrictions on prominence in terms of the syllable structure formulated by Selkirk (1980b). The first syllable of reläy is an open syllable - and an open syllable with a lax vowel cannot have any prominence in English, as correctly stated in Selkirk's general remarks. In the approach I am proposing here, lexical representation contains maximal feature specifications of the segments pertaining to a form, on the condition that each maximally specified set of distinctive features must occur in at least one of the corresponding surface forms. In other forms, distinctive feature specifications may become predictable due to the newly surrounding segments, or due to higher-level phenomena. One of the latter is the prosodic structure, by which a [+tense] vowel becomes unspecified for [±tense] if it occurs in a form where it is neither an accent foot nor the strong syllable of a binary foot. According to a general markedness convention, it shows up as lax in the phonetic, or surface, representation. If occurring in a closed syllable at a prosodic word boundary, it is phonetically implemented as lengthened and perceived as having a secondary accent. Accent alternations in English are relatively restricted, and there is a large amount of predictability, to be captured in terms of the strong weak alternations within binaty feet, and a possibility of having monosyllabic feet at word boundaries. The only characteristic which is not simply computable is the position of the main accent on stems and the capacity of affixes to affect it or not. This crucial point of the prosodic structure forms part of the lexical representation. All of the remaining regularities are predictable at the level of the phonetic representation. They can be formulated in terms of general rules, which can as such be kept out of the individual lexical representations. This statement is not restricted to English. Vogel - Scalise (1982) have shown that an analysis of secondary accent in Italian does not necessitate reference to any foot assignment in the underlying representation, but can be stated rather by means of general rules regulating the strong weak alternations, in addition to lexical marking of the main accent. In a different way, Malikouti-Drachman and Drachman's paper in this work-shop discusses regularities of the strong - weak alternations in

40

Jadranka Gvozdanovic

Dimotiki Greek in relation to the prosodic word and what they call 'the prosodic intonation', showing the general and automatic character of foot assignment in Greek as well.1

5. The prosodic word and its relation to the phonological word The prosodic word has been defined by Selkirk (1980b: 570) as being constituted by a sequence of one or more feet or superfeet (a superfoot being a strong foot followed by a weak syllable in English, i.e. s aw), joined in a right-branching structure. The domain of the prosodic word in English equals the simple non-branching stem and any stem affixes. There is a distinction between cohering and non-cohering affixes, and only the former form part of the prosodic word: the cohering affixes syllabify together with the stem, whereas the non-cohering ones do not. In English, -ic is a cohering suffix, and e.g. the prosodic word rhythmic is pronounced with a syllable-onset [m], whereas -y is an example of a non-cohering suffix, and e. g. rhythmy is pronounced with a syllable nucleus [iji], as the prosodic word boundary directly following it, and preceding the -y, prevents the former from becoming a syllable onset. Syllabification is consequently viewed as indicative of the domain of the prosodic word, within which prominence relations are operative. This statement, valid for English, does not hold for all languages, though. In Serbo-Croatian, for example, prominence relations have a domain which does not coincide with that indicated by means of syllabification. Syllabification treats prepositions and other proclitics (cf. below) as non-cohering, whereas prominence relations treat them as cohering. This can be seen from the occurrence of the falling tonal accents, in the standard language, bound to the initial syllable of the prosodic word (for a discussion of variation in this respect, see Gvozdanovic 1985 a), which in such cases occur on the proclitic, as illustrated below. (6) Syllabification and prosodic word boundaries in Standard Serbo-Croatian: ovcu 'a/the sheep, accusative', [ovcu], is syllabified as [ov-cu]; nad ovcu 'above a/the sheep, accusative', [nädövcu], is syllabified as [näd-öv-cu]

Phonological domains

41

(with Ο denoting the long falling tonal accent, (") denoting the short falling tonal accent, (") denoting vowel length, and (-), syllable boundaries). In Serbo-Croatian, syllabification is consequently indicative of major morphological boundaries. It is indicative of the boundary between a proclitic and the following morphological word, but it is not indicative of the boundary between a morphological word and the following enclitic, as in the following example, where the proclitic does not syllabify with the following morphological word (and the sequence [ts] occurs with a boundary between its constitutive parts), whereas the enclitic does syllabify with the preceding morphological word (and a syllable onset [c] occurs corresponding to [ts] at a proclitic boundary). (7) Syllabification and morphological word boundaries in SerboCroatian: od sad se 'from now on itself, [otsace], is syllabified as [ot-sa-ce] (with ( Λ ) denoting the short rising tonal accent). True, some speakers of Serbo-Croatian simplify the situation by not syllabifying enclitics with the preceding morphological word either, but this cannot be considered the normal pronunciation rule. If syllabification is not necessarily indicative of the prosodic word in every language under consideration, how can the prosodic word be defined then? It obviously need not equal the morphological word, which can be isolated or separated from the remaining parts of the sentence without a change of the syntactic structure. As a language-specific matter, the prosodic word can comprise either more or less than a morphological word, or be equal to it. A prosodic word comprises more than a morphological word if it includes clitics, which may be called 'bound morphological words', since they do not occur in isolation (i.e. they cannot form an utterance on their own), and can be separated from the remaining parts of the sentence only in a limited number of cases. As a language-specific matter, clitics are adpositions, short pronominal forms, auxiliaries, particles, and determiners at various levels. They have in common that they have no lexical accent marking of their own. Some languages, such as Serbo-Croatian, distinguish between proclitics, which precede a morphological word with a lexical (i.e. inherent) accent marking, and enclitics, which follow the first constituent in a sen-

42

Jadranka Gvozdanovic

tence. This can be illustrated by example 7 above, where od is a proclitic, and se an enclitic. The Serbo-Croatian proclitics have less restricted placement possibilities, i. e. they are relatively free, in comparison with the enclitics. The proclitics are distinguished from the enclitics prosodically in that only the proclitics are accentable, for they may attract the accent from a following morphological word with a lexical marking for a tonal accent. In Serbo-Croatian, the tonal accent is a feature of morphemes, which can be erased when different morphemes are put together. This erasure means that a distinctive tone becomes predictable, i. e. unmarked, as a morpheme feature in combination with another morpheme. Due to a general markedness convention, the tone which is unmarked for [±rising] shows up as phonetically falling (for a justification of tone contours rather than levels, see Gvozdanovic 1980:24ff.). I have argued elsewhere (cf. Gvozdanovic 1983 c: 79-80) that an initial falling tone which alternates with the rising tone in the lexical marking can in such cases be seen as unmarked for [±rising]. This absence of distinctive tone is equivalent, phonetically, to an initial falling accent, for this implements a preceding prosodic word boundary, which is always either high or rising in Serbo-Croatian, cf. Gvozdanovic 1980: 99). If now a morphological word with an alternating falling tonal accent is combined with a proclitic, the prosodic word boundary is shifted to the left of the proclitic, which thus automatically acquires an initial falling tonal accent itself, due to the initial prosodic word boundary. As distinguished from the alternating falling tone, the nonalternating falling tone in Serbo-Croatian must be seen as distinctively [—rising]. This is responsible for a different retraction onto the proclitics, as illustrated below. (8) The Serbo-Croatian tonal accents on the proclitics: casa 'a/the glass, nominative' [öäsa] co.su 'a/the glass, accusative' [6äsu] u casu 'into a/the glass' [uöasu] glava 'a/the head, nominative' [gläva] glavu 'a/the head, accusative' [glavu] u glavu 'into the head, into one's mind' [üglävu] (") denotes the short falling tonal accent, Ο denotes the short rising tonal accent, (") denotes the long falling tonal accent, (') denotes the long rising tonal accent, and (") denotes unaccented vowel length. The proclitics differ from the enclitics also in that only the proclitics can be used contrastively and consequently bear the sentence accent

Phonological domains

43

(whereas for the enclitics, only their corresponding full forms can be used contrastively). Whenever either a proclitic or the following morphological word is used contrastively, there is a prosodic word boundary between the proclitic and the following morphological word, so that we can speak of two prosodic words, as illustrated by (9 b) below. (9) Contrastive usage treats proclitics as separate prosodic words: a. bez case je 'lit. without glass is, i. e. it is without glass' [besöaseje]; b. bez case je 'lit. without glass is, i. e. it is without glass that it is' [bes cäseje] (Here (') denotes the sentence accent). We can see that the neutral division into prosodic words can be overruled by the higher-level domain of sentence intonation and accent as expressing contrast. If a morphological word with a lexical tonal accent marking of its own is not surrounded by clitics, then it either equals the prosodic word, or is larger than the corresponding prosodic word. As a language-specific matter, the prosodic word can comprise less than a morphological word in derived words and in compounds. In Serbo-Croatian, a compound consisting of a determinans and a determinatum forms a single prosodic word, whereas in other cases each stem with its suffixes, if any, forms a separate prosodic word, as illustrated by example (10) below. (10) Prosodic analysis of Serbo-Croatian compounds: crvena 'red, definite, feminine' [crvena] kapica 'hood, diminutive, feminine' [käpica] Crvenfcapica 'Little Red Riding-hood' [crvenkapica] (with a falling tonal accent of the determinatum retracted onto the determinans) Jadran 'Adriatic' [jadrän] plastika 'plastic' [plästika] Jadranplastika 'name of a plastic factory, i.e. not a sort of plastic' [jadrämplastika] radio 'radio' [rädiio] tehnika 'technics' [tehnika] radiotehnika 'radiotechnics' [rädiiotehnika]. On the basis of examples (9) and (10) above we can conclude that the phonetic implementation rules specifiying voice and place assimilation in consonants are operative throughout a unit which can be defined as

44

Jadranka Gvozdanovic

equaling a morphological word with any surrounding clitics, irrespectively of its prosodic analysis. For Serbo-Croatian, the morphological word with the surrounding clitics can be called 'the phonological word'. It has a constant definition and it is independent of the prosodic word, which must be defined as the unit within which the prosodic phenomena are operative. Serbo-Croatian is not an isolated case of a language distinguishing the prosodic word from the phonological one. A comparable example is Savo Finnish, where vowel harmony is operative throughout the phonological word and thus indicative of its domain (comparable to that in Serbo-Croatian), whereas the prosodic word, equaling the domain of the tonal phenomena there, is restricted to the syllable or syllables containing the first two vowels of the word (cf. Chelimskij 1977: 19). Another example of non-coincidence of the prosodic and phonological word is found in Turkish. Like Savo Finnish, Turkish has vowel harmony which is operative throughout the phonological word. Prosodically, Turkish has bound accent at the end of the prosodic word, implemented by means of a pitch rise. The bound accent precedes certain morphemes, such as personal endings, even though they do undergo the vowel harmony rules, which make them belong to the same phonological word as the preceding morpheme, as illustrated by the examples given in (11) below. (11) Turkish vowel harmony indicates the domain of the phonological word: - if the vowel of the first syllable of a word is a back vowel, so are the vowels of subsequent syllables; - if the vowel of the first syllable of a word is a front vowel, so are the vowels of the subsequent syllables; - unrounded vowels are followed by unrounded vowels within the same word; and - rounded vowels are followed by low unrounded or high rounded vowels. gelirim Ί come' geliyorum 'lamcoming' gelecegim 'pertaining to my future coming' gelecegim Ί shall come' adam idi/adamdi 'it was the man' deniz 'sea'

Phonological domains

45

alti 'its underside' denizalti 'submarine' The bound accent is indicated by means of underlining. In adamdi, we have one prosodic word and one phonological word, whereas in denizalti, we have one prosodic word and two phonological words. We can conclude that the prosodic word is a level of the prosodic structure which is autonomous with respect to the grammatical structure, and must be defined for each language under investigation. The phonological word, on the other hand, is as a rule fully derivable from the grammatical structure (in the way outlined by Chomsky - Halle (1968: 368) already, but with language-specific elaborations). The prosodic and the phonological word may coincide as a language-specific matter.

6. Higher-level prosodic domains Against the background of Chomsky - Halle's (1968: 366-367) discussion of the boundaries which characterize syntactic clusters created around a major lexical item, Selkirk (1980b) has offered the following definition of the phonological phrase: (i) an item which is the specifier of a syntactic phrase joins with the head of the phrase; (ii) an item belonging to a nonlexical category, such as Det, Prep, Comp, Verbaux, Conjunction joins with its sister constituent. Besides the nonlexical items, which show up as clitics, the phonological phrase comprises also the phrase specifiers, and it is here that the distinction between the phonological word and the phonological phrase comes in - in the languages where it is relevant. In French, for example, 'liaison', which is a case of morphologized sandhi reflected in the syllabification properties, appears to be operative within a phonological phrase (cf. Morin - Kaye 1982, and Booij in this volume) which is not distinguished from the phonological word. In Dutch, on the other hand, consonant voice assimilation occurs in a significantly higher number of cases within a phonological word than across word boundaries within a

46

Jadranka Gvozdanovic

phonological phrase (cf. Van Hooff - Van den Broecke 1983), though it remains to be seen whether the two domains are in fact different. The phonological phrase is directly derivable from the grammatical structure, and the prosodic phrase is isomorphous with it, unless it is dictated by a contrastive sentence accent in a way parallel to the case of the prosodic word delimitation discussed above (see 5). Besides the phonological phrase (φ), a derived phonological phrase (φ') has been defined by Nespor - Vogel (1980) in order to capture both segmental and prosodic Italian sandhi phenomena, which occur obligatorily within φ and optionally within φ': φ construction: join into a φ any lexical head (X) with all items on its non-recursive side with the maximal projection and with any other nonlexical items on the same side; φ' restructuring: a nonbranching φ which is the first complement of X on its recursive side loses its label and is joined to the φ containing X under a new node labeled φ'. In example (12), the word initial consonant is doubled following the final accent of the preceding word if both words occur within a φ οτ φ', but not across φ'. (12) Italian sandhi ('raddoppiamento sintattico') according to Nespor-Vogel (1980): Φ' Φ

φ

φ

Luca inviterä [dijamele e (Jhiara inviterä [kji vuole. 'Luca will invite Daniele and Chiara will invite whoever she wants.' The optional relevance of φ' to Italian 'raddoppiamento sintattico' is paralleled by its optional relevance to French 'liaison', which is conditioned sociolinguistically: 'liaison' can occur within the entire φ' in elevated French (cf. Morin - Kaye 1982: 294). Italian optionality must be investigated further. The concept of 'restructuring' in the case of φ' as compared with φ, which was taken by Nespor - Vogel to correspond structurally to the observed optionality, must be investigated further in regard to its triggering factors. This 'restructuring' may be conditioned either sociolinguistically or systematically, as due to the information structure and its contras-

Phonological domains

47

tive value in the sentence, which is reflected in the sentence intonation. This systematic possibility is always present in a language, but it may be revealed either by means of intonation alone, as in non-elevated French (cf. Morin - Kaye 1982: 303), or additionally by means of sandhi, as in Italian. (13) Intonation contours on a lexical head followed by its nonbranching first complement on the recursive side in French (from Morin-Kaye 1982: 303): apporte le lit 'bring the bed' ##apporte##le#lit## Co Q

Co

( # # ) denotes a phonological word boundary, and (#), a morphological word boundary. Intonationally, there would be φ' 'restructuring' in the upper example in (13) with a unifying intonation contour, and none in the lower example lacking a unifying contour. As Morin - Kaye point out, each phonological word in French can function as an intonational unit. This, however, has no bearing on 'liaison'. We can conclude that in French, too, prosodic domains are autonomous with respect to the grammatical structure - though restricted by it. The prosodic domains in French have no bearing on 'liaison', which is a matter of morphological sandhi, but they do appear to coincide with the domain for 'enchainement', which is a matter of phonological sandhi reflected by syllabification, which is operative throughout the φ'. An example of phonological sandhi operative throughout an intonation unit - even across the boundaries of φ' - is found in the place assimilation of nasals to following consonants in Dimotiki Greek, which is discussed by Malikouti-Drachman - Drachman in this workshop (they refer to it as the 'prosodic intonation'). What is the intonation unit and how does it relate to the utterance? Selkirk (1980b) defines the intonation unit (I) as the domain over which an intonation contour is spread, and the highest level, the utterance (U), as the span between two pauses in connected speech. Both I and U are right branching. Nespor - Vogel (1982: 231 etc.) have elaborated on the syntactic factors which are relevant to I and U construction as follows:

48

Jadranka Gvozdanovic

I construction (i) any displaced syntactic constituents, parentheticals and non-restrictive relative clauses obligatorily form at least one I; (ii) starting with the first φ/φ' of a sentence, join as many φ/φ' as possible into an I until either a) the end of the maximal projection of an Ν is reached, or b) another S begins; once such an I is formed, proceed in the same way until the end of the main sentence is reached; join any remaining φ's at the end of a sentence into an I; U construction join all I's in a root sentence (most generally the highest category of syntactic structure) into a U. Following Wheeler's (1981) Branching and Prominence Constraint, by which right branching trees are labelled weak - strong and left branching trees strong - weak, independently of whether or not their nodes branch, Nespor - Vogel have established that prominence relations within I and U are predictable given the type of syntactic branching. In spite of these correspondences, I and U are still not simply derivable from the syntactic structure, as can be concluded from their formulation of I restructuring, which runs as follows: I restructuring a) eliminate very short I's by joining them with adjacent I's; b) eliminate very long I's by breaking them down into shorter I's. Nespor - Vogel (1982: 234) write that they are not yet sure exactly what factors determine restructuring, and that these may include "other than strictly linguistic considerations (e.g. physiological limitations, perceptual strategies, stylistic considerations)". It is our task, though, to sort out the considerations which are linguistic in the sense that their presence invariably invokes a unifying intonation contour, whereas their absence invokes a break of the intonation contour, from other considerations which are either automatic or nonsystematic. This is a task for the future investigation of intonation units. Urestructuring has been formulated by Vogel (this volume) as partly, though - according to the author - presumably not fully, dependent on the following factors: U restructuring adjacent U ' s may be joined into a single U when a) they are produced by the same speaker, b) they are directed to the same addres-

Phonological domains

49

see(s), and c) there is a syntactic relation (ellipsis, anaphora), or a semantic relation (and, therefore, becausej, or a pragmatic relation between the U 's in question. We can see that the term 'restructuring' in the context of φ I , and U, refers to the relation between prosody and syntax, as these prosodic domains have originally been defined on the basis of syntactic units, but are apparently not in all cases derivable from them in a straightforward way. An alternative approach to the higher-level prosodic domains would be to investigate their phonetic correlates in intonation and any other possible clues, such as pauses, and to relate these phonetic correlates to any semantic, syntactic or pragmatic factors which are associated with them in a constant way, and which do not occur in the absence of these clues. In doing this, one must keep in mind that syntactic factors are a formal relation means in a way comparable to intonation which is a prosodic relation means, and that linguistic investigation must relate these units of form to the corresponding semantic and pragmatic units of meaning. A linguistic analysis which is based on the unity of form and meaning would speak of restructuring only if it can be shown that factors of the same type are involved in the construction of the higher-level prosodic domains and in their restructuring, so that the latter can be derived from the former. Further linguistic research is needed in order to show whether this is the case in the prosodic domains discussed above. Concerning the relation between I and U, it can be seen that U is defined as a higher-level domain with respect to I. Further phonetic investigation of U must show whether it is in fact a higher-level I, characterized by a corresponding intonation contour. Recent investigations of intonation show that intonation units can in fact be hierarchically structured. Martin's (1978) investigation of French shows that this is done by means of rises and falls of different amplitude. Termination of a major intonation unit is signalled by a [-(-amplitude] fall, whereas embedded intonation units within that major unit are terminated by a non-amplitude fall. The elements of French intonation established by Martin (these are contours, their amplitude and timing) have parallels in other languages. In Dutch (cf. Collier - ' t Hart 1978), amplitude seems to signal the hierarchy of intonation units, and timing with respect to the lexically accented syllable signals both predictable prominence given the syntactic

50

Jadranka Gvozdanovic

branching (of the type discussed by Nespor - Vogel 1982) and distinctive prominence on an I if the intonation rise or fall is early in the accented syllable, or it signals prosodic boundaries if the rise or fall is late. Further investigation must establish meaningful correlates of the prosodic intonation units described above, which are hierarchically structured and as such related to the prosodic hierarchy which proceeds from I to U as the highest-ranking prosodic domains. The relevant formal intonation elements are obviously rises and falls, which open and close intonation units, respectively, and are further distinguished by their timing and/or amplitude. In addition to this, there is a phonetically predictable declination line showing what the extent of the U is (which is restricted by the sameness of the speaker and the addressee, but not, for instance, by sentence boundaries, as shown by Vogel in this volume).

7. Conclusions Some of the prosodic domains distinguished by Kiparsky (1981: 245) and extensively discussed in this paper are fully derivable from the phonological and the grammatical structure, whereas others are autonomous. I propose to reserve the term 'level' only for the autonomous prosodic domains. The prosodic domains which are derivable from non-prosodic levels, but which must be defined for each language under investigation, are: the syllable, the foot, and the prosodic phrase. The syllable is derivable by means of rules stating possible distinctive feature sequences provided grammatical boundaries are taken into account. The foot is derivable from the lexical accent marking, the segmental features of the syllable nuclei involved, and the prosodic word boundaries. Finally, the prosodic phrase is derivable from the grammatical structure. The derivation rules of these domains which are fully derivable in terms of other levels of the language structure must be stated for each language under investigation. In other words, 'derivable' means 'analysable in terms of other units by means of a general rule'. The segment, as a set of distinctive features, is a level of the phonological structure, referred to by the remaining phonological and prosodic domains. The phonological domains are defined as the domains within which

Phonological domains

51

various types of rules involving distinctive features are operative in such a way that they affect either locally neighboring segments or segments which are neighboring in a sequence based on the same dominating distinctive feature. An example of the latter is found in vowel harmony, which is linear within a distinctive feature dimension in the sense that it affects an entire vowel sequence (such that so-called 'neutral' vowels, due to a lack of feature specification, are transparent to such a generalization without stopping it). The phonological word and higher-level phonological domains are derivable from the grammatical structure. The autonomous prosodic domains, or levels, are the prosodic word, the intonation unit (at least in some languages comprising the derived phrase φ'), and the utterance, which participate in a hierarchical structure. The phonological domains are formally expressed by means of segmental sandhi phenomena (involving vowel harmony and various types of assimilations, or feature neutralizations). The prosodic domains are expressed by means of prosodic features and their sandhi phenomena (cf. Basboll's contribution in this volume), which may be paralleled by segmental sandhi, depending on the language under investigation. Sandhi thus formally indicates domains within which sets of semantic and relational features are put together in a language, either due to a grammatical unit with its possible hierarchy, or due to pragmatic organization of the information units in a speech situation.

Note 1. Unfortunately the workshop paper by Angeliki Malikouti-Drachman and Gaberell Drachman could not be included in this volume. (Editor's note.)

References Andersen, Henning 1972 'Diphthongization', Language48:11-50. 1975 'Variance and invariance in phonological typology', in: Phonologica 1972, ed. by Wolfgang U. Dressier - F. V. Mares (München: Salzburg, Wilhelm Finck Verlag), 67-79. 1978 'Vocalic and consonantal languages', in: Studia Linguistica Alexandre Vasiliifilio Issatschenko a Collegis Amicisque oblata, ed. by Henrik Birnbaum et al. (Lisse: The Peter de Ridder Press), 1-12.

52

Jadranka Gvozdanovic

Anderson, Stephen R. 1978 'Syllables, segments, and the northwest Caucasian languages', in: Syllables and segments, ed. by Alan Bell - Joan B. Hooper (Amsterdam: The North Holland Publishing Company), 47-58. Basbell, Hans 1986 'Stod-sandhi', In this volume, 75-83. Bell, Alan - Joan B. Hooper 1978 'Issues and evidence in syllabic phonology', in: Syllables and segments, ed. by Alan Bell - Joan B. Hooper (Amsterdam: The North Holland Publishing company), 3-22. Booij, Geert E. 1986 'Two cases of external sandhi in French: enchancement and liaison', In this volume, 93-103. Chelimskij, E. 1977 'Tonal'nye oppozicii ν ural'skich jazykach', Nyelvtudomänyi Közlemenyek 1-2:7-55. Chomsky, Noam - Morris Halle 1968 The sound pattern of English (New York: Harper and Row). Collier, R - J . ' t Hart 1978 Cursus Nederlandse intonatie (Diepenbeek: Wetenschappelijk onderwijs Limburg). Cook Τ. L. 1985 An integrated phonology of Efik, Vol. I. (Univ. of Leiden dissertation). Ebeling, Carl L. 1960 Linguistic units (The Hague: Mouton). Farmer, Alvirda 1979 'Phonological markedness and the sonority hierarchy', in: Papers on syllable structure, metrical structure and harmony processes, ed. by K. Safir (M.I.T. working papers in linguistics 1: Cambridge), 172-177. Gvozdanovic, Jadranka 1980 Tone and accent in Standard Serbo-Croatian, with a synopsis of SerboCroatian phonology (Vienna: Österreichische Akademie der Wissenschaften). 1983 a 'Patterning of distinctive features in relation to variability', in: Proceedings of the 13th International Congress of Linguists, ed. by Sh. Hattorf ICInoue (Tokyo), 611-617. 1983b Distinctive feature constraints on speech errors', in: Sound structures, Studies for Antonie Cohen, ed. by M. P. R van den Broecke et al. (Dordrecht: Foris), 129-140. 1983 c 'Typological characteristics of Slavic and non-Slavic languages with distinctive tonal accents', in: Dutch contributions to the 9th International Congress of Slavists (Amsterdam: Rodopi), 55-108. 1985 a Language system and its change. On theory and testability (Trends in Linguistics. Studies and Monographs 30) (Berlin, New York, Amsterdam: Mouton de Gruyter). 1985b 'Markedness and ordering at various levels', in: Studio linguistica diachronica et synchronica. Werner Winter sexagenario ed. by U. Pieper G. Stickel, 285-302.

Phonological domains

53

Hockett, Charles F. 1955 Manual of phonology (Bloomington: Indiana University Press). Hooff, C. M.C. van - M. P. R. van den Broecke 1983 'Assimilation of voice in Dutch at three types of boundaries', Progress Report (Institute of Phonetics, University of Utrecht, 8) (Utrecht), 31-40. Isaöenko, A.V. 1940 'Versuch einer Typologie der slavischen Sprachen', Linguistica Slovaca 1/2:64-76. Jakobson, Roman - E.Colin Cherry - Morris Halle 1962 'Towards the logical description of languages in their phonemic aspect', in: Roman Jakobson, Selected writings /(The Hague: Mouton), 449-463. Jakobson, Roman - C. Gunnar M. Fant - Morris Halle 1951 Preliminaries to speech analysis (Cambridge: The Μ. I. T. Press). Kiparsky, Paul 1981 'Remarks on the metrical structure of the syllable', in : Phonologica 1980, ed. by Wolfgang U. Dressier et al. (Innsbruck: Innsbrucker Beiträge zur Sprachwissenschaft), 245-256. 1982 'From cyclic phonology to lexical phonology', in: The structure of phonological representations, ed. by Harry van der Hulst - Norval Smith (Dordrecht: Foris), 131-175. Kuipers, Aert H. 1960 Phoneme and morpheme in Kabardian (The Hague: Mouton). Malikouti-Drachman, Angeliki - Gaberell Drachman MS. 'Greek sandhi rules and refooting', presented in the Sandhi Workshop. Martin, Phillipe 1978 'Questions de phonosyntaxe et de phonosemantique en fran^ais', Linguistic Investigations2:93-125. Morin, Yves-Charles - Jonathan D. Kaye 1982 'The syntactic bases for French liaison', Journal of Linguistics 18: 291-330. Nespor, Marina - Irene Vogel 1980 'Prosodic hierarchy and speech perception', Paper presented at the Conference on the Perception of Language, Florence. 1982 ' Prosodic domains of external sandhi rules', in: The structure ofphonological representations, ed. by Harry van der Hulst - Norval Smith (Dordrecht: Foris), 225-263. Selkirk, Elizabeth O. 1980 O n prosodic structure and its relation to syntactic structure', (Indiana University Linguistics Club: Bloomington, Indiana). 1980 b "The role of prosodic categories in English word stress', Linguistic Inquiry 11,3:563-605. 1980c 'Prosodic domains in phonology: Sanskrit revisited', in: Juncture, ed. by Mark Aronoff - M.-L. Kean (Saratoga: Anma Libri), 107-129. 1982 On the major class features and syllable theory (Unpublished ms. Μ. I. T. and University of Massachusetts, Amherst). Steriade, D. 1982 Greek prosodies and the nature ofsyllabification (Μ. I. T. dissertation).

54

Jadranka Gvozdanovic

Vogel, Irene - Sergio Scalise 1982 'Secondary stress in Italian', Lingua 58:213-242. Vogel, Irene 1986 'External sandhi rules operating between sentences', In this volume, 55-64. Ward, Ida Caroline 1933 The phonetic and tonal structure of (Cambridge: HefFer). Wheeler, Deirdre W. 1981 Aspects ofa categorial theory ofphonology (Amherst dissertation).

External sandhi rules operating between sentences Irene Vogel

1. Introduction The linking-r and the intrusive-r are two well known phenomena in English Received Pronunciation (RP)1 that account for the appearance of an [r] in certain contexts in connected speech. It is generally recognized that both linking and intrusive-r apply within words and between words in a sentence; I will thus refer to them simply as sandhi rules, without specifying 'internal' or 'external'. In this paper, I will demonstrate that linking and intrusive-r are, in fact, a special type of sandhi rule in that they are not limited to applying within a sentence, but can also apply, under certain circumstances, between words in different sentences. While the focus in traditional generative phonology, or the SPE framework, was on word level phenomena, linking and intrusive-r clearly show the need for phonological theory to be able to handle entities much larger than the word, and in this case, larger even than the sentence. In the analysis of linking and intrusive-r presented here, I will demonstrate that the recently developed prosodic theory (cf. Selkirk 1980; Nespor - Vogel 1982,1983), in fact, responds to this need. Specifically, it will be proposed that the domain of application of linking and intrusive-r is the largest phonological unit in the prosodic hierarchy, the phonological utterance, and, furthermore, that the definition of this unit must take into consideration certain aspects of discourse structure. It will thus be shown that the application of sandhi rules across sentences is not an odd type of phonological phenomenon, but rather one that directly follows from the nature of the phonological utterance in the prosodic hierarchy.

56

Irene Vogel

2. Two r phenomena While linking-r and instrusive-r have in common the fact that they account for the pronunciation of an Y sound in certain contexts, they are usually treated as separate rules on the basis of whether or not an r appears in the orthography. Specifically, according to standard descriptions of English (cf. Jones 1966), linking-r can apply whenever a word ending in orthographic r(and optionally a 'silent e") is followed by a suffix or another word beginning with a vowel (e.g. neaf vs. nea[r]est, anothef vs. anothe[r]egg, where '/'is an unpronounced r), and intrusive-r can apply whenever a word ending in [a], [a], or [d] is followed by a suffix or another word beginning with a vowel (e. g. draw vs. draw[rfing, saw vs. sawfrJEd). While it may seem unlikely that such disparate types of contexts as an orthographic rand [a, a, d] can form part of a single phenomenon, a closer examination of the first of these contexts provides insight into the basic process underlying both linking and intrusive-r. Let us first consider which vowels may be followed by an orthographic rin word final position. These are: [i] (e.g .fear), [ε] (e.g. bare), [u] (e.g. sure), [o] (e.g. sore), [a] (e.g. car), [a] (e.g. fir),2 and the diphthongs [ei] (e.g. layer), [ai] (e.g. tire), [oi] (e.g .foyer), [au] (e.g. hour), [ou] (e.g. mower). Phonetically, what we find is that a schwa appears in the position of the orthographic r following [i, ε, u] and the diphthongs ending in [i] and [u].3 Thus, the endings of fear and mower are the same as those of idea and boa, respectively, i. e. [ia] and [oua]. After the remaining vowels, there is no sound in the position of the orthographic r, so, for example, the endings of sore and car are the same as those of saw and spa, i. e. [d] and [a], respectively. What this means is that the contexts in which the so-called linking-r can appear are exactly the same as those in which the intrusive-r can occur, namely following [a, a, o] and before another vowel. In other words, both linking and intrusive-r serve the same function, that of breaking up sequences of two vowels in which the first one is [a], [a], or [o]. We can thus formulate a single rule of r insertion that accounts for both linking and intrusive-r as follows:

It was mentioned above that both linking and intrusive-r can apply before a suffix or another word beginning with a vowel. The r-insertion

External sandhi rules operating between sentences

57

rule in (1) that expresses these two phenomena is, therefore, incomplete, since it does not include any information about its domain of application. Before we can propose a solution to this problem, however, it is necessary to examine systematically the cases in which an r may and may not be inserted. Although it has just been shown that linking and intrusive·/· are, in fact, a single process, at least as far as their segmental contexts are concerned, they will be illustrated separately here, in order to demonstrate that their domains of application are also the same, and that from this point of view, too, they represent a single process. It should be noted that neither linking nor intrusive-ris obligatory, and the discussion that follows is aimed at accounting for the environments in which they may apply and those in which they may not, rather than predicting exactly when a linking or intrusive ris actually pronounced.4 As mentioned above, linking and intrusive-rboth apply within words. Since their segmental contexts5 do not occur in monomorphemic words, the rules apply only in complex (i. e. derived and compound) words, in which the appropriate context is created by the addition of a suffix or another word. The two rules are illustrated in (2), where 'f indicates an unpronounced r. (2 a) linking-/· b) intrusive-/· wintef- + ize—• winte[r]ize claw + ing—• claw[r]ing eaf + ache—*• ea[x]ache awe + inspiring—• awe[r]inspiring It was mentioned, in addition, that linking and intrusive-r can apply between two words in a sentence, and, as the examples in (3) show, they can apply between any two words, regardless of their syntactic categories or the relation that holds between them.6 (3 a)

linking-r a poof example —• a poor[r]example a jaf of honey —• a ja[r]of... covef up errors —• cove[r]up... The doctof examined him. —• ... docto[r]examined...

b) intrusive-/· law and order —• law[r]and saw everything —* saw[r]everything gnaw incessantly —• gnaw[r]incessantly Amanda always lies. —• Amanda[r]always ...

Furthermore, although it may seem strange at first glance, linking and intrusive-/· can also apply between two words that belong to different sentences in certain cases, as illustrated in (4).

58

Irene Vogel

(4a) (linking-/·): Close the doof. I'm freezing. —•.. .doo[t\I'm... b) (intrusive-r): Let's go to Canada. It's cheaper this year. —• .. .Canada[r]It's... The application of linking and intrusive-r across sentences is not free, however, as the sentences in (5) show. That is, although the segmental environments in (5 a) and (5 b) are identical to those in (4 a) and (4 b), respectively, the rules in question apply in the examples in (4) but not in those in (5). (5 a) (linking-r): Close the doof. I'm Frances. —• *. .door[r\I'm... b) (intrusive-r): Let's go to Canada. It's Chip at the door. —• * . .Canada[r]It's... At this point, it is worthwhile to note one more fact in relation to the application of linking and intrusive-r, before we proceed to the analysis of their domains. While all of the examples considered thus far are relatively short, length is not the crucial factor in determining the application of the rules, as can be seen in the longer examples in (6). (6 a) (linking-r): When I go to visit my mothef, I usually end up staying over night. —•.. .mothe[r]I... b) (intrusive-r): Sit on the other sofa. It's much more comfortable. —* .. .sofa[r]It's... What the data show, first of all, is that linking and intrusive-r apply in the same domains, and that we are thus further justified in treating them as a single rule of r-insertion. They show, in addition, that as far as the domain of application is concerned, r-insertion applies across all types of boundaries, to use traditional generative terminology. If the rule were limited to applying within the sentence, the issue of the domain would be easily handled by making reference to boundaries; it would be sufficient to say that r-insertion may apply across all boundaries weaker than ' # # # # ' , where the two sets of ' # # ' represent the point at which two sentences come together. The situation is not this simple, however, since we have also seen that r-insertion is not limited to the domain of the sentence, but sometimes applies across sentences as well. At this point, it seems that the traditional generative model is at a loss. We cannot simply say that r-insertion may also apply across sentence boundaries, since there are cases in which this is not true. Furthermore, simply stating that the rule applies across some sentences and not others is not particularly interesting since there are no boundaries, or other mechanisms, in the

External sandhi rules operating between sentences

59

theory that allow us to represent differences among sentence types, assuming it were clear which differences had to be represented. Essentially, what the problem boils down to is the fact that there is no unit that can include more than one sentence, such that it can be said that this unit is the domain of application of /--insertion, and all other sandhi rules that operate between sentences.

3. The prosodic domain of r- insertion In an SPE-type model of phonology, the only way of representing the domains of a phonological rule is in terms of morphosyntactic constituents, the implicit claim being that such constituents are, in fact, the only domains in which phonological rules may apply. Recent work has shown, however, that morphosyntactic constituents are not the appropriate domains for at least some phonological rules, and in particular, those operating above the word level. Instead, it has been demonstrated that the appropriate domains are hierarchically arranged prosodic (phonological) units, the largest of which is the phonological utterance (cf. Selkirk 1980; Nespor - Vogel 1982,1983). Although the phonological utterance (U) is usually considered to be coextensive with the root sentence in syntax, there is no reason that this has to be the case. In fact, one of the fundamental principles of prosodic phonology is that while reference is made to syntactic notions in the construction of prosodic categories, the resulting categories are not themselves necessarily isomorphic to any syntactic constituents. Thus, within the prosodic framework, it is possible for a unit to consist of more than one sentence, even though such a combination does not correspond to a constituent in syntax. Such a unit is created by restructuring into a single (phonological) utterance certain smaller U's that were originally built to be coextensive with root sentences by the basic U-construction rule (cf. Nespor - Vogel 1982). What this means in the case of /--insertion, and all sandhi rules operating across sentences, is that the domain of this type of rule is the phonological utterance, which may consist either of a single root sentence, or of more root sentences in the case that U-restructuring has taken place. What is needed now is some principle that determines which U's may be restructured. Without such a principle, we must specify separately for

60

Irene Vogel

each pair of sentences, or U's, whether or not they may undergo U-restruduring, in which case the prosodic analysis, aside from the fact that it includes a unit that may be larger than a sentence, is no more revealing than a traditional analysis. Since I have not come across any principles directly applicable to U-restructuring in the (syntax, speech act, or discourse analysis) literature, what follows is a first attempt at identifying the principles underlying U-restructuring. As a starting point, let us consider two general hypotheses: (i) that Urestructuring depends on the nature of the individual sentences involved, and (ii) that it depends on some relationship between the sentences involved. As far as hypothesis (i) is concerned, it is easy to demonstrate that Urestructuring does not depend on the types of the individual sentences, either in terms of their grammatical structure or their illocutionary force. That is, U-restructuring can occur with combinations of the same types of sentences as well as with combinations of different types, in different orders. The examples of /--insertion (as a result of U-restructuring) in (7) illustrate various combinations of declarative, imperative, and interrogative sentences. (7 a) (DEC-DEC): It's notfat. I'll walk. .. .fa[r]I'll... b) (INT-INT): Where's Anna? Is she late again? —• 4nna[r]Is... c) (INT-DEC): Where's Victop. I'm leaving. ... Victor]I'm... d) (IMP-DEC): Try this soda. It's good. .. .soda[r] It's... e) (DEC-IMP): There's Petef. Ask him to come in. —• .. .Pete[T]Ask... f) (IMP-INT): Take the cat. Isn't itfaster?—* .. .CO[T] Isn't... In order to evaluate hypothesis (ii), we can no longer consider individual sentences; instead, we must consider sequences of sentences, and this moves the issue into the field of discourse analysis. In fact, if one imagines each of the pairs of sentences in (7) as a sequence uttered as part of a conversation, at least at an intuitive level, it seems that there is some type of relationship between the two members of each pair, thus, 'intuitively' confirming hypothesis (ii). The question to ask at this point is obviously what type of relationship must exist for two sentences to be able to form a single linguistic unit. While there is quite a bit of discussion in the discourse literature about possible sequences of sentences, the focus is primarily on sentences uttered by different speakers, in contrast with the type of phe-

External sandhi rules operating between sentences

61

nomena we are concerned with here.7 This leads us to formulate a first general condition on U-restructuring, and that is, that the two U's to be restructured must be uttered by the same speaker, since no matter how fast two sentences follow each other, they cannot form a single unit for the purposes of phonology if they are uttered by different speakers. This is not sufficient, though; it seems that a second general condition must also be met, and that is, that the U's to be restructured must also be directed to the same addressee(s). Thus, if the sentences in (8) are addressed to the same person, r-insertion would normally apply, as in (8 a); if they are addressed to different people, intrusive-r cannot apply, as in (8 b). (8) Where's Petef. I need help. 3 er υ ε (4 V3

.κ S « ε a -S '•Ρ υ 3 ·£ ο. £ υ ,ο je

S

νο * I t l c ο § s "β ·° cn •a tu υ υ ε 3 σ" •5 « «J — ΐ5 & 6 5 ο Ιπ ρ 2 Ε C/3 C

suoipujssj jeoixsi jo ]boi}buiuibj8 suoipujsaj jo ad/ίχ

79

80

Hans Basbell

6. The different kinds of linguistic restrictions applying to the main types of st0d-sandhi are summarized in the table on p. 79. 'Environment' (i.e. Relevant Environment) is the morpheme to the right and/or left of the domain which is relevant for the occurrence or non-occurrence of stod in the domain. Notice that the environment follows the domain in I-III, but precedes it in IV. As has been said above, all cases of stod-sandhi are subject to two general restrictions, which have therefore not been included in the table: (i) only syllables with "stod-basis" can have stod; and (ii) the morphemes are relevant units in all cases, i.e. there are both phonological and morphological restrictions which apply to any type of stod-sandhi. It is immediately clear from the table that IV is of quite a different nature than the other three types of st0d-sandhi. The answer to questions 3 and 5 in column IV are similar, but unrelated to anything we have looked at so far. They can be paraphrased thus: "the lighter the first part of a compound (or a derivative), the more probable is stod-addition in the second part (or stem) if it is a verb, or a verbal derivative, or an adjective formed by suffixation". "Light" both refers to "phonological weight", cf. that unstressed first parts generally condition stod-addition (e.g. trolove [tKo'b?v9] 'betroth'); and to "lexical weight" in the sense that so-called "small words" like prepositions, adverbs etc. also condition stod-addition (e.g. päsmore ['pA,since?b] 'smear on, apply'). If the first part is a stressed polysyllabic "lexical category", stod-addition is rare (e.g. friturestege [firi'tyiBiSdaib] 'deep-fry'), whereas it is more of a possibility with a monosyllabic first part like terkoge ['toe^koiua/ •toe^kofya] 'cook dry'. Whereas the stod sandhi types (just like rules of vowel quantity) are kinds of internal sandhi, whose relevant environment does not exceed the word, including clitics, other types of sandhi in Modern Danish typically extend across word-boundaries. The rule which assimilates the place of articulation of a nasal (which must be dental or unspecified with respect to place of articulation) to a following stop, for example, is obligatory within morphemes (e.g. banke$>ar$gQ>\ 'bank'), whereas it can apply across grammatical boundaries to an increasing degree with decreasing formality or increasing speed, even across the strong word- and phrase-boundaries. Nothing like this pattern is found in the cases of stod-sandhi.

St0d-sandhi gl 7. Let us now, finally, turn to what appears to be a crucial problem of the present paper, viz. the relation between stod-sandhi and other stod rules. There are three main types of native monomorphemic words in Danish, viz. monosyllables, disyllables ending in / a / , and disyllables ending in / a / plus a sonorant consonant. The stod-rules which apply to these three types will be mentioned separately (cf. Hansen 1943: 23-42, Basboll 1972: 8-13 and Heger 1980). The simplest case is that of disyllables ending in / a / : they (practically) never have stod; cf. the fact that the ending / s / - in any of its grammatical functions - nearly always conditions stod-loss of a preceding monosyllable. Monomorphemic disyllables in / a / are thus stod-less just like bimorphemic ones, and in this case stod-sandhi has contributed to the phonological fading out of morpheme boundaries. The pattern is thus very strong here. Monosyllables fall into several different types as far as stod is concerned: (i) if they contain a long vowel, or a short vowel followed by a sonorant consonant followed by at least one more consonant, they have st0d (e.g. Ids, folk [lo?s, faPg] 'lock', 'people'); (ii) if they end in a short vowel followed by just one sonorant consonant, they can either have stod or be stod-less (e.g. vend, ven [νεη\ νεη] 'turn', 'friend'); and (iii) if they have a final short vowel, or contain a short vowel immediately followed by an obstruent, they are stod-less (e.g. nu, hest[nu, hesd] 'now', 'horse'). That such forms are stod-less is a simple consequence of the fact that they lack "stod-basis". What happens when monosyllables occur before the stod-adding endings? All those with "stod-basis" get stod. Group (iii) splits up, so that those with a final short vowel get vowel lengthening, and thus "stod-basis", and stod, whereas the others (with short vowel plus an obstruent) remain stod-less. This difference in stod behavior is a consequence of the fact, I would say, that obstruents do not belong to the class of consonants which permit vowel lengthening (they have stable vowel quantity before them), whereas vowels before vowels occur in a position with no possible commutation of vowel quantity, and thus in a phonologically unstable quantity position. Finally, monomorphemic words which are disyllables ending in / a / plus a sonorant most often have stod, but a fair number are stod-less, and the picture is, as a whole, very complicated, just like the one for bimorphemic disyllables.

82

Hans Basbell

8. It should be emphasized that the general case, according to the present analysis, is that no stad-sandhi takes place, which helps to simplify the identification of the morphemes in a given phonological string. To end the paper, let me try, by means of a few examples, to show how the sted-sandhi-rules and other stod-rules may have an invaluable function in the perception and interpretation of speech, whereas they may appear rather arbitrary, and to a large extent non-functional or even dysfunctional, from the point of view of production. (1) [tals, haPs]. The stod rules (cf. (i)) show unambiguously that / s / must be a separate morpheme in the former case; it is part of the stem in the latter (tals, 'of number' (gen.), hals 'neck', /tal+s/, /hals/). (2) [taqgan, taq^gan]. It could not be excluded a priori that both words are monomorphemic (but such words in / a n / are rare). If there is a morpheme boundary, however, the stod-rules unambiguously point towards the analyses tanke plus η and tank plus en, respectively. Furthermore, since tank ends in -nk, it must have stod when occurring in isolation (tanke [taqgs] 'thought' (sb.), tank[taq?g] 'tank', both def. sg.). (3) [νεηθη, νεη^η]. Same type as above, i. e. the possible morphological decomposition is [vena] plus [n] and [νεη?] plus [an]. Here is a further ambiguity, however, since the "basic form" of the latter could be either with or without stod (venden 'turning', from vende [vena] 'turn', and vennen 'the friend' from ven [νεη] 'friend'). Note that this analysis correctly gives three possibilities on this level (cf. manden [man?an] 'the man' from mand [man*]), and that stod-occurrence in the monomorphemic forms vend!, ven [νεη\ νεη] is exactly the opposite of that found in venden, vennen. (4) [kuiB, ku?B]. The stod-rules point to a morphological decomposition into kue and r versus ku and er, which is correct (kuer 'cows (vb.)' from kue [ku:a], q'er, plural of q [ku?] (name of letter)). The stod-full form could also, phonologically speaking, have been derived from a stod-less "basic form" as in nuer[nu?e] (pi.) from nu [nu] 'now'. (5) [kte:da, kk?da]. The first form could either be monomorphemic, or be decomposed into a stem plus a stod-removing ending. The second form must be polymorphemic because of the stod. Its stem is clearly something here manifested [kte?], and the ending could either be the stod-neutral / d a / (cf. hojde, hoj [hApda, IiaP] 'height', 'high'), or - as happens to be the case - a non-syllabic and hence stod-neutral flexive (/t/) plus / a / which cannot remove the stod from the stem because another morpheme intervenes) (klsedte [kteida] 'dressed' (past) /kleiö+ta/, klsedte [kl£?da] 'dressed' (past participle, def. or plural form) /kle:ö -I-1 +a/).

Sted-sandhi

83

(6) [uöitaeita, 'u0,tae?b]. The words are easily analysed as being composed from ud and tale, ['ud—] is ambiguous as to whether its form will contain a stod vowel or not (cf. ud, bud[u?d, bud] 'out', 'command; messenger'). The stod-rules unambiguously tell us that [—,tae?b] is a verb as opposed to [—,tae:ta]. Notice the diversity of the above examples with respect to the stod: in (1) and (2), only the stod-rules for monomorphemic words apply; in the rest of the examples, stod-sandhi is essential, but in several ways and in different degrees, for helping the hearer to detect the distinction in morphological analysis of the forms in question. I conclude from such examples, which could be multiplied ad infinitum, that an important function of the whole complex of stod-rules in Modem Danish, including sted-sandhi as a crucial part, is to facilitate the morphological analysis of phonological strings, i.e. a perceptual function.

References Basboll, Hans 1972

'Some remarks concerning the sted in a generative grammar of Danish', Derivational processes, ed. by Ferenc Kiefer (Stockholm: KVAL), 5-30. 1985 'St0d in Modern Danish', Folia Linguistica 19:1 -50. Brink, Lars - Jörn Lund 1975 Dansk rigsmil 1-2 (Copenhagen: Gyldendal). Hansen, Aage 1943 Stedet i dansk (Det kgl. danske Videnskabernes Selskab, Historisk-filologiske Meddelelser 29, 5) (Copenhagen: Munksgaard). Heger, Steffen 1980 'Stedregler i dansk', Danske Studier (Copenhagen: Akademisk), 78-99. OreSnik, Janez 1979 'Stodet i rigsdansk morfologi', Danske Studier (Copenhagen: Akademisk), 123-129. Spore, Palle 1965 La langue danoise (Copenhagen: Akademisk).

French external sandhi: the case of liaison Michael Herslund

1. Introduction French liaison has undoubtedly been one of the favorite topics of generative phonology. No single subject seems to have attracted more attention or caused more ink to flow. So why not restage this classic for another performance in the framework of a general discussion of sandhi phenomena in the European languages. The classical generative approach to French liaison, i.e. Schane (1968) and subsequent works in that vein, is to treat the well-known facts of(l): (1 a) les copains [le ko 'pe] un petit copain [e pa ti ko'pe] b) les amis [le za 'mi] un petit ami [e pa ti ta 'mi] by assuming the same underlying forms for the prenominal determiners and adjectives of a. and b., viz. /lez/ and /patit/. The a. case is then derived by a general rule which deletes a consonant before another consonant, whereas, in the b. case, the underlying word-final consonant remains prevocalically and surfaces as a syllable initial consonant of the following word. I assume that all this is well-known, not only from the rich specialized literature (Schane 1968, etc.), but also from extensive use of French liaison as a model example in phonology textbooks. The immediate interest of the description is of course that it allows a unified treatment of e.g. adjectives, in so far as all inflected and derived forms can be obtained by applying general phonological rules to a single common underlying form, cf. (2) :

(rules)

/patit (J/ /patit V/ /patita/

Surface: [pati]

[patit]

[patit]

/patit+esa/

[patites]

86

Michael Herslund

This is of course a highly desirable outcome, but unfortunately, it is not without problems. The most serious problem for the classical approach is the fact that the assumed rule of consonant deletion cannot be a general rule of French phonology since many words in fact end in a consonant, e. g. concept, avec, sens, club, autobus, relatif, etc. To provide such words with an underlying final schwa is of course only an arbitrary way of deciding that the problem does not really exist. The fact that final consonant deletion could not be a general rule and additional facts about language learning and development have in recent years led to a thorough reconsideration of liaison, mainly within what has come to be known as "natural phonology" or "concrete (morpho)phonology" (Klausenburger 1974, 1978). But before discussing those proposals, in 2.2,1 would like, in 2.1, to consider briefly the proposal that liaison is a kind of mutation.

2. The alternatives 2.1 Liaison as mutation The proposal of Ternes (1977) is that French liaison should be treated on a par with the mutation phenomena known from the Celtic languages, i.e. as a case of initial modification of certain morphemes in certain environments. In the case of French, however, we have a very simple case of mutation, because the phonemic alternation characteristic of mutation in this case is an alternation between 0 and /z, t, n/ and a couple of other consonants. The decisive factor which leads to this interpretation is the fact that the liaison consonant of e. g. [pa ti ta 'mi], [le za 'mi], belongs to the following word, not to the word where orthography would lead one to place it. This observation, Ternes claims, is contrary to traditional assumptions. I must confess that I am not aware of the tradition to which Ternes refers here. It has in fact always been acknowledged that liaison consonants phonetically belong to the following word, just as the general phenomenon of enchainement is one of the basic wellknown facts of French phonetics. But it is hard to see why this simple phonetic fact should be decisive in setting up underlying word boundaries. It is in fact perfectly natural, and also explicitly stated in the classical approach mentioned in sec.l, that the underlying final consonants

French external sandhi: the case of liaison

87

of petit and les become surface initial consonants of the following words, cf. (3): (3) Underlying form: /patit ami/ Surface form: [pa ti ta 'mi]

/lezami(z)/ [le za 'mi]

But the crucial point of Ternes' analysis, and that is where it goes wrong, as I shall suggest in 3. below, is that liaison is a question of initial modification of morphemes. This initial modification is triggered by certain morphemes, which carry an appropriate marking to that effect, cf. (4): (4) /le z ami/ —• /le zami/ —• [le za 'mi] /pati1 ami/ —• /pati tami/ —• [pa ti ta 'mi] The small raised letters are the marking, or instruction to prefix a certain phoneme to a following vowel initial morpheme. Certain details apart, this approach does not look very different from the classical generative analysis, except in one respect: under the mutation analysis, all vowel initial morphemes in French are to be represented as consonant initial, i.e. with a zero consonant in order for the alternation to take place. For this is the very definition of mutation: a non-phonetically conditioned alternation of phonemes (Ternes 1977:34). So zero is a consonantal phoneme in French, and one might start wondering whether this would not lead to some difficulties in the formulation of vowel elision, which is no less a central feature of French phonology than liaison is. But this is probably where the discussion degenerates into mere questions of terminology, so let me go on to the other proposal I would like to discuss.

2.2 Liaison as epenthesis The opposite of the classical solution mentioned in sec. 1 above has recently been advocated in a number of works starting with Klausenburger (1974). The proposal is simply stated this. Liaison constitutes, historically, a case of rule inversion. Instead of the rule of final consonant deletion, which is essentially a synchronic repetition of the diachronic process, it is assumed that the preconsonantal forms of e. g. les and petit, [le], [psti], are underlying and that the longer, prevocalic forms, [lez], [patit], are formed by a rule of consonant insertion. This rule is then an inverted form, (5) a., of the rule of consonant deletion, (5) b.:

88

Michael Herslund

(5 a) 0—»-C/ b) C—>-0/

V C

The question which immediately comes to one's mind and which undoubtedly presents the most serious problem for the insertion analysis is of course how one would know which consonant to insert, and how one can block the rule from inserting consonants where no epenthesis should occur. In brief, how do speakers know that a ί is to be inserted in petit ami, a ζ in les amis, but nothing in joli ami /Obviously, the rule (5) a. won't do as stated. The most comprehensive answer to such questions is found in Tranel (1981). The core of Tranel's proposal is this (I shall have to limit myself to a discussion of adjective liaison, the most complex case). A general linking rule inserts an unspecified consonant between any adjective and a following vowel initial noun. The lexical entries of adjectives contain information as to which consonant, if any, is to be inserted when the linking rule has applied, as in (6) (the proposal is actually extended to cover also the gender consonants inserted to form feminine adjectives, but I am forced to ignore that aspect here): (6) joli /zoli/ petit /pati/

(/t/)

grand /grä/ gros /gro/

(/t/) (/z/)

"With respect to this lexical determination, it is important to emphasize that the consonants given in parentheses in (27) (=my (6)) are not part of the phonological representations of the adjectives; rather, they are idiosyncratic phonological markings which are simply part of the lexical entries" (Tranel 1981: 238). But these "diacritic consonants", if they are not themselves part of the phonological representation of the adjective, are phonological entities to the extent that they specify the phonological manifestation of the inserted unspecified consonant; i.e. to make the proposal explicit, these markings serve as triggers to as many different phonological rules as there are different linking consonants. Furthermore, in the case of joli, the general linking rule also inserts an unspecified consonant; but since joli carries no marking as to which consonant is to be inserted, the unspecified linking consonant is deleted. The importance of the question how to implement these proposals seems however greatly minimized by Tranel: "Adjectives are lexically specified as to which linking consonant, if any, they require; by convention, the features of this consonant fill in the unspecified features of the inserted / C / ; if an adjective has no such lexical consonant, then

French external sandhi: the case of liaison

89

/ C / automatically receives a null phonetic realization" (Tranel 1981: 237). It is unclear to me which convention Tranel has in mind. In any event, the grammar has to be complicated by such a convention, and it is hard to see how it should be put to work if not by a set of phonological rules specifying / C / as / t / , / z / , etc. To say that the unspecified / C / is deleted automatically is also to say that there is a (perhaps general) phonological rule which deletes unspecified consonants, and that rule too must be part of the grammar. So we have, paradoxically, in this "anti-deletion" approach, the following situation: where the classical approach does not delete a word-final consonant, Tranel inserts a / C / which is converted to / t / , / z / , etc. Where nothing at all happens in the classical approach, i.e. in the jolicase, Tranel deletes the / C / which has not received any phonological specification. But note that Tranel, cleverly, avoids the verb 'delete' by saying "receives a null phonetic realization" ! To sum up then, Tranel's proposal seems to constitute but a notational variant of Ternes' mutation analysis. I do think, however, that Tranel is correct in treating adjective liaison as a lexically determined phenomenon, but wrong in treating it as a consonant inserting process.

3. Liaison in the morphophonological pattern of French 3.1 Final or initial modification of morphemes? The two approaches I have reviewed briefly have some inherent merit, but I believe that they both fail because they treat liaison either as initial modification of morphemes (mutation) or as "inter-morphemic" insertion of linking consonants. Contrary to these assumptions I would maintain that one crucial feature of French morphophonemics is the systematic distinction between prevocalic and preconsonantal (and prepausal) forms. That this is the case is demonstrated not only by elision phenomena, but also by the existence, and preservation over centuries, of special suppletive prevocalic forms in determiners and adjectives: where "regular" adjectives like petit wad grand have consonant-final liaison forms prevocalically, a small group of adjectives have special suppletive forms, viz. beau - bei, nouveau - nouvel, vieux - vieil, etc. These forms cannot of course be described either by mutation or by epenthe-

90

Michael Herslund

sis. On the other hand, I believe that suppletion is an important phenomenon and that the suppletive forms should be handled the same way as liaison forms. Suppletive forms normally follow the regular morphophonological pattern, i. e. if there is suppletion in one paradigm there is at least one other paradigm where the same alternation is implemented by a regular morphophonological process. Put otherwise, the very reason for identifying and acknowledging suppletion is the existence of regular morphophonemic processes in environments parallel to those where suppletion is found: if e.g. suppletion is acknowledged in the Danish verb vsere 'be', er 'is', var 'was', the reason is that other verbs form these three forms by regular means, e.g. leve 'live', lever 'lives', levede 'lived'. But one does not expect to find a suppletive morph in a paradigmatic position not otherwise provided for by the morphological pattern of the language; so one does not expect to find, say, a suppletive aorist passive in a language which does not otherwise have aorist passive forms (if such a form is actually found, one would rather speak of a relic than of a suppletive form). Now, returning to French, the very existence of prevocalic suppletive forms such as bei, nouvel, vieil, etc., points to the conclusion that the morphophonemic pattern of French has some regular means of distinguishing preconsonantal from prevocalic allomorphs. The distinction is implemented in the adjectives regularly by liaison, irregularly by suppletion, by regular morphophonemic processes in nouns and verbs, and by both in the determiners (plur. les is regular, demonstrative ce - cet suppletive). The conclusion to be drawn from this discussion is that the same distinction which is relevant for suppletion is also relevant for liaison and that the two cases should really be treated together.

3.2 The lexical solution: latent consonants How then can the description capture the two facts which I have argued are important, viz. that liaison is a question of final modification of morphemes, and that this modification is lexically determined since a general rule of consonant deletion is not viable? This is where the notion of latent consonants comes in. How could that notion be made meaningful? As most often understood it refers to the written language, cf. Heger (1968) or Ternes (1977); understood in that way, the notion is of course somewhat paradoxical:

French external sandhi: the case of liaison

91

"... da es keine „stummen Konsonanten" gibt und sie daher auch nicht in bestimmten Fällen „hörbar" werden können" (Ternes 1977: n.41). But I think that the notion could be meaningful, phonologically speaking, in the light of the preceding discussion of suppletion. A latent consonant is a consonant which is pronounced under certain circumstances (prevocalically), silent under others (preconsonantally or prepausally). This amounts of course to saying that certain lexical units really have two underlying forms, one with a final consonant, the other without (cf. Flydal 1979). Just like nouveau, petit has two underlying forms in the masculine then: a short preconsonantal (or prepausal) form, /pati/, and a longer prevocalic form, /patit/. The difference between such regular adjectives and the suppletive adjectives is simply that the two forms of the former can be collapsed by the well-known parenthesis notation while the latter cannot: (7) Regular: Suppletive:

petit les beau - bei ce - cet

/pati(t)/ /le(z)/ / b o / - /bei/ /sa/ - /set/

The consonant which appears in the longer liaison form in the adjective is the same as the one appearing in the feminine and derived forms, or a similar consonant obtained by the well-known adjustments of / d / to / t / , / s / to / z / , which any approach has to account for. So the underlying unity of adjective stems is kept, and it is not really important for my proposal whether feminine forms have an underlying final schwa or not. The important thing is that feminine forms have a final consonant and that the same consonant has the status of a latent consonant in the masculine, i.e. the masculine really has two forms: (8) masculine petit /p9ti(t)/

feminine petite /p3tit(a)/

4. Conclusion In this approach then, the underlying unity of adjective stems is preserved, and the "true" nature of liaison, a lexically (or morphologically) determined final modification of certain morphemes is acknowledged.

92

Michael Herslund

Suppletion in adjectives and determiners, which proved a valuable clue, is integrated into the general framework of liaison, the systematic distinction between prevocalic and preconsonantal (or prepausal) positions, which constitutes, I believe, the crucial feature of French external sandhi. There are, under this proposal, no rules of deletion and no rules of insertion. Speakers simply have, with certain morphemes, the choice between two underlying forms. That choice is dictated by position, syntactic structure and style; but all speakers do not necessarily have the same underlying forms; that they sometimes choose the wrong form, i. e. do not make an obligatory liaison, un gros avantage [e gro a vä ta^], or an erroneous form, i. e. make a false liaison, quat ζ enfants peat zä 'fa], should not surprise anybody, and certainly does not constitute an argument for the epenthesis proposals. The price which my approach has to pay, i. e. a certain complication of lexical representations, is, as seen, common to this and Tranel's approach. I think however that it is a price which any serious description of liaison should consider paying, one way or the other.

References Flydal, Leiv 1979

Heger, Klaus 1968

'Latences et liaisons en franfais - systemes coexistants ou un seul?', in: Estudios ofrecidos a Emilio Alarcos Llorach, 4 (Oviedo: Univ. de Oviedo), 43-68.

'Die liaison als phonologisches Problem', in: Festschrift Walther von Wartburg zum 80. Geburtstag, ed. by Κ. Baldinger, 467-484. Klausenburger, Jürgen 1974 'Rule inversion, opacity, conspiracies: French liaison and elision', Lingua 34:167-179. 1978 'French linking phenomena', Language 54:21 -40. Schane, Sanford A. 1968 French phonology and morphology (Cambridge, Mass.: Μ. I. T. Press). Ternes, Elmar 1977 'Konsonantische Anlautveränderungen in den keltischen und romanischen Sprachen', Romanistisches Jahrbuch 28:19-53. Tranel, Bernard 1981 Concreteness in generative phonology. Evidence from French (Berkeley Los Angeles: University of California Press).

Two cases of external sandhi in French: enchainement and liaison Geert E. Booij

In French, word-final consonants can be syllabified with the vowel of the following vowel-initial word. This phenomenon, called 'enchainement', is illustrated by the examples in (1), where σ == 'syllable' i1 (1)

syllabification: cinq heures 'five hours' pour arriver 'to arrive' avec eile 'with her' une amie 'a friend (fern.)' cette Anglaise 'this English woman' cognac exquis 'exquisite cognac' bloc enorme 'enormous block'

(se)a (koer)a (pu)a (ra)a (π)σ (νβ)σ (a)a (νε)σ (kel)a (y)a (na)a (mi)a (se)CT (tä)a (gleiz)«, (ko)a (jia)a (ke)a (ski)a (bb) a (ke)a (ηοπη)σ

The phenomenon of enchainement can be qualified as a natural phonological process, since it results in optimal syllable patterns in which the onsets of syllables are not empty. Yet, it is a disadvantage as well, since the boundaries between words are obliterated by this process, and this may impede the recognition of the individual words. The rule of enchainement can be formulated as follows: (2) Enchainement Q) . . . W Q V . . . where ω=phonological word, and Q is a possible syllable-initial cluster. Note that more than one consonant may be shifted, as is illustrated by pauvre ami 'poor friend' with the syllabification pattern (ρο)σ (vra)CT (ιηί)σ. If more than one consonant is shifted, the resulting word-initial cluster must be a possible syllable-initial cluster of French: in e. g. petites amies [ptitzami] 'little friends (fem.)' only the [z] is syllabified with the following vowel [a], because [tz] is not a possible syllable-initial cluster. Enchainement also affects another rule of French, Liquid Deletion. This rule optionally deletes a word-final liquid ([1] or [r]) after an obstruent (cf. Dell 1981). Compare the examples in (3):

94

Geert Ε. Booij (3) quelle table? 'which table?' quelle arbre? 'which tree?'

[keltabl] or [kelarbr] or

[keltab] [kelarb]

However, Liquid Deletion cannot apply to pauvre ami:[povami] is an illformed phonetic representation. This follows from the fact that Enchainement has shifted the [r] of pauvre to the following word, and thus the [r] is no longer word-final.2 The second, and related kind of external sandhi in French that I will discuss in this paper is liaison. The crucial difference between liaison and enchainement is that in liaison latent consonants, which are not pronounced in word-final position, are syllabified with the first vowel of a following vowel-initial word, and then pronounced (cf. Grammont 1938: 129, Malmberg 1972:141, Carton 1974: 87). The examples in (4) illustrate liaison (indicated by w ): (4i) petit^ami 'little friend' ii) dansjune heure 'in one hour' iii) les^enfants 'the children' les petits^animaux 'the small animals' iv) il vit^en France 'he lives in France' Liaison applies to the final consonants of adjectives (4i), some adverbs and prepositions (4ii), the plural suffix / z / for adjectives, nouns and determiners (4iii), and the final consonants of verbal inflectional suffixes (4iv). There are at least three different types of analysis which account for this surfacing of latent consonants before vowel-initial words. The approach of standard generative phonology, exemplified by Dell (1980) is to assume that these consonants are present underlyingly, and that they are deleted by a rule of Consonant Truncation unless they are followed by a vowel separated from the consonant by at most one word boundary (5) Consonant Truncation [- son]

0 / ##

This rule presupposes that words in liaison context are separated by at most one word boundary. For instance, petit and ami in the phrase petit ami are supposed to be separated by one # .

Two cases of external sandhi in French

95

The second type of approach is to insert consonants in liaison contexts, and is advocated by Tranel (1981). The obvious problem for such a theory is how to predict which consonant must be inserted. Moreover, it should be expressed somehow that the same consonant turns up in both the masculine form of an adjective in liaison context, in its feminine form, and in words derived from that adjective. Compare: (6) petit ami [ptitami] (eile est) petite [ptit] petitesse [ptites]

'little friend' 'she is small' 'smallness'

Tranel therefore assumes that in liaison contexts empty consonant-slots are inserted which are filled by consonants which occur as optional consonants in lexical representations. For instance, the word petit is assigned the following lexical representation: (7) /pati/

(/t/)

However, the nature of these consonantal slots and the formal procedure for filling these slots is not specified in any detail. The analysis furthermore implies that such empty C's must also be inserted before each vowel-initial derivational suffix (cf. complex words like petit-esse 'smallness' and soldat-esque 'soldier-like' in which the latent consonant of the stem surfaces). Another problem is that in certain words two consonants are latent, as in respect [respe] 'respect' (cf. respecter [respekte] 'to respect'). Here, two empty C's would have to be inserted. Yet, it is unpredictable in which words. A variant of this theory can be found in Morin - Kaye (1982) who assume a set of epenthesis rules which insert various consonants in liaison contexts. This implies that for each word it has to be learnt which of the epenthesis rules has to apply. With respect to the plural suffix / z / they moreover assume that this morpheme has been reanalyzed as a plural prefix which is attached to post-nominal adjectives in phrases such as des amis agreables 'nice friends'. The third theory of latent consonants, and the one that I will assume here, is the theory of extrasyllabic consonants proposed by Clements Keyser (1981).3 They propose that in French word-final consonants can be extrasyllabic, i. e. not linked to a syllable node. Such consonants will therefore not be realized phonetically unless they are linked to a syllable node by some rule. For instance, petit ami 'little friend' and petit gargon 'little boy' will have the following representations respectively:4

96

Geert Ε. Booij

(8) ρ 91 i t σ σ

ami σ

patit

σ

gars δ

σ σ

σ

σ

Clements - Keyser now formulate liaison as a rule which links an extrasyllabic consonant (C) to the first syllable - with empty onset - of the following word: (9) Liaison (Clements - Keyser 1981: 80):

[- sylir

+ syll]

The dotted line indicates the linking process. The domain of the rule will be discussed in the next section. Rule (9) will link the extrasyllabic t of petit in petit ami to the σ-node dominating the a of ami, but cannot apply to the final t of petit in petit garfon. As was pointed out above, extrasyllabic consonants also show up in derivational morphology and inflectional morphology. For instance, the final t of petit surfaces in petitesse and petite. This follows from the Peripherally Condition for extrametrical constituents proposed by Hayes (1982: 270): (10)

X [+ ex]

[- e x ] / - Y ] D

where Y=£0 and D is the domain of the stress rules if we subsume the notions 'extrametricality' and 'extrasyllabicity' under one notion 'extraprosodicality' and generalize the Peripherality Condition to extraprosodic constituents.5 Thus in the case of petitesse the final t of petit will lose its extrasyllabicity because it is no longer word-final and hence will be linked to the syllable node dominating the first vowel of -esse by means of the process of automatic resyllabification that applies after each application of a morphological or phonological rule. The same condition accounts for the surfacing of two extrasyllabic consonants in a phrase like mon circonspect ami 'my cautious friend' where the final -fo-cluster of circonspect is extrasyllabic. The rule of Liaison applies to the final t of circonspect and thus is no longer extrasyllabic. Consequently, the A: loses its extrasyllabicity by (10). When there is no rule which effects the attachment of the extrasyllabic consonant, it is finally deleted by the following convention:

Two cases of external sandhi in French

97

(11) Non-syllabified segment deletion Delete from phonetic representation any segment that is not dominated by σ (cf. Cairns - Feinstein 1982: 219) This convention can be assumed to be universal (cf. Booij, to appear). The striking similarity between liaison and enchainement is explained by the fact that they are related historically (Malmberg 1972:140): enchainement saved the word-final consonants from truncation by shifting them to word-initial position. The main advantages of the theory of extrasyllabic consonants with respect to liaison can be summed up as follows: (i) The representation of exceptions to consonant truncation (words such as type 'type', pilote 'pilot', chef1 chief, honnete 'honest', vague 'vague' is very straightforward now: they simply lack extrasyllabic consonants. (ii) Minor C/0-alternations such as the [r]/0-alternation in words in -ier and -er (e. g. dernier 'final', premier 'first', leger 'light') can be covered by the same rule of Liaison. The majority of words in -r do not show this alternation (eg. rare 'rare', bizarre 'bizarre') and thus simply do not have an extrasyllabic r.6 (iii) The relation between the surfacing of the latent consonants and the concomitant resyllabification is directly expressed in the rule of Liaison.7

The domain of external sandhi In traditional descriptions the domains of application of enchainement and liaison are referred to as "breath groups" (Pulgram 1964: 131) or "rhythmical units" (Grammont 1938: 129, Malmberg 1972: 141), but these units did not receive any formal definition. In Selkirk (1978: 25) it is claimed that the phonological phrase (φ) is the domain of obligatory liaison. Referring to this proposal, Nespor - Vogel (1982: 241) also claim that in colloquial French liaison applies within φ. They define φ as follows: (12) φ-construction Join into a φ any lexical head (X) with all items on its non-recursive side within the maximal projection, and with any other

98

Geert Ε. Booij

non-lexical items on the same side (e.g. prepositions, complementizers, conjunctions, copulas ...) (Nespor - Vogel 1982: 228-229) This rule of φ-construction will qualify the phrases in (4i)-(4iii) as φ, and thus liaison is correctly predicted. Example (4iv), however, consists of two 213

The effect of this rule is that the metathesized consonants are exempted from Consonant Truncation, which truncates consonants before word boundary. Selkirk (1972: 228) pointed out that this analysis runs into problems when two latent consonants surface in liaison context, as in mon circonspect ami. Selkirk argues that this problem cannot be solved by metathesizing two consonants, because in petits amis only the / z / surfaces although both / t / and / z / are latent. In my analysis this problem is solved (cf. the main text). Cf. Booij (to appear) for some other advantages of the analysis in terms of extrasyllabicity and a comparison with the theory of Liaison proposed in Anderson (1982). 8. Nespor - Vogel (1982) do not mention this, and explicitly state that Liaison does not apply in domains larger than φ. 9. The / t / of petits cannot be deleted by convention (11), because, after the linking of the / s / to the following syllable (due to Liaison), the / t / will lose its extrasyllabicity, analogously to the / k / of circonspect in mort circonspect ami. 10. Rule (14) also accounts for the well known difference between: (i) un [savanflN [anglais^ 'an English scientist1 (ii) un[savant]A [anglais]N 'a learned Englishman' where we find liaison only in (ii). 11. Bernard Al (pers. comm.) provided this example. 12. Yves-Charles Morin (pers. comm.)

References Anderson, Stephen R 1982 'The analysis of French shwa\ Language 58:534-573. Booij, Geert E. To appear Extrasyllabicity and French C/0-altemations. Cairns, Charles E. - Mark H. Feinstein 1982 'Markedness and the theory of syllable structure', Linguistic Inquiry 13: 193-226. Carton, F. 1974 Introduction ά la phonitique frangaise (Paris/Bruxelles/Montreal: Bordas). Clements, G.N. - S.J.Keyser 1981 A three-tiered theory of the syllable (MIT, Center for Cognitive Science Occasional Paper 19) (Cambridge, Mass.). Dell, Francis 1970 Les regies phonologiques tardives et la morphologie derivationelle du franfais (MIT dissertation). 1980 Generative phonology and French phonology (Cambridge: Cambridge University Press) [Revised English version of Les regies et les sons (Paris: Hermann), 1973].

Two cases of external sandhi in French 1981

103

O n the learnability of optional phonological rules', Linguistic Inquiry 12: 31-38.

Fouche, Pierre 1956 Tratte de prononciation frangaise (Paris: Klincksieck). Grammont, Maurice 1938 La prononciation frangaise (Paris: Delgrave). Hayes, Bruce 1982 'Extrametricality and English stress', Linguistic Inquiry 13:227-276. Jong, D. de - Ε. Roll - W.Woudman 1981 La Liaison. L'influence sociale et stylistique sur l'emploi de la liaison dans le frangais parle a Tours (Groningen: Universiteit van Groningen). Malmberg, Bertil 1972 Phonitiquefrangaise (Malmö: Hermods). Morin, Yves-Charles - Jonathan D. Kaye 1982 'The syntactic bases for French liaison', Journal of Linguistics 18: 291-330. Nespor, Marina - Irene Vogel 1982 'Prosodic domains of external sandhi rules', in: The structure ofphonological representations, Part I, ed. by Harry van der Hulst - Norval Smith, (Dordrecht: Foris), 225-256. Pulgram, Ernst 1964 'Prosodic systems: French', Lingua 13:125-144. Rubach, Jerzy 1981 Cyclic phonology and palatalization in Polish and English (Warszawa: University of Warsaw). Selkirk, Elizabeth 1974 'French liaison and the X' notation', Linguistic Inquiry 5:573-590. 1978 On prosodic structure and its relation to syntactic structure [Unpublished ms] (University of Amherst, Massachusetts). 1980 The phrase phonology of English and French (Indiana University Linguistics Club: Bloomington, Indiana). Sten, Holger 1962 Manuel de phonetiquefranfaise (Kebenhavn: Ejnar Munksgaard). Tranel, Bernhard 1981 Concreteness in generative phonology. Evidencefrom French (Berkeley/Los Angeles/London: University of California Press).

Typology of the Celtic mutations Herbert Pilch

1. Focus The Celtic languages have their share of phonologically conditioned sandhi just like many other languages. An example is the aspiration of the stress-initial resonants / r m η q / of Welsh when preceded by proclitic material. Within the domain of the lexical word, the non-aspirates / r m n q / accordingly alternate with the aspirates /r 4 m' n' q'/, as in: aros 'stay' 'cymwys 'suitable' 'cynneddf 'constitution' cyngor 'council'

~ arhos-af'l stay' ~ cymhwys-iad 'adjustment' ~ cynhiddf-au1 (pi.) ~ cynghor-wr 'councillor'

This alternation operates only over part (not the whole) of the lexicon. A phonologically conditioned sandhi which operates over the whole of the lexicon is the productive nasal mutation of Colloquial Welsh. Phonologically, it consists in the (obligatory) co-articulation of a final nasal with an adjacent initial stop. In the co-articulated complex, the feature 'nasal' dominates over the feature 'oral', thus: nasal + / p ' / ('aspirated, labial, oral')

—• / m ' / ('aspirated, labial, nasal') nasal + / p / ('unaspirated, labial, oral') —• / m / ('unaspirated, labial, nasal'), and so on for nasal + / t ' t k' k/. The domain of this alternation is the initial consonant of any syntactic group determined by the first singular pronoun / a n / : 1 ps. sg. pron. —• syntactic group2, as in: pen 'head' ~ /am'en/ 'my head', pen öl 'bottom' ~ /am'enol/ 'my bottom', bygythio 'threaten' ~ /amakap'jo/ 'threaten me' bola-haelo 'bask in the sun' ~ /amolahailo/ 'my basking in the sun'.

106

Herbert Pilch

What is exceptional about the Celtic languages from the point of view of sandhi typology is, however, their aphonological sandhi, i. e. syntactically conditioned sandhi. In the Celtic languages, there is (obligatory) alternation between different initial consonant phonemes. Which consonant is chosen depends, in every particular instance, on the syntactic relations involved. For instance, initial / k ' / alternates with initial / k / in Welsh. In particular instances, / k / (spelt g) is chosen for the direct object of a finite verb, / k ' / (spelt c) for the direct object of an infinite verb, as in: gig gi gath gar gerbyd

'John bought 'John bought 'John bought 'John bought 'John bought

meat' a dog' a cat' a car' a cart'

~ y mae John cig wediprynu ci cath car cerbyd

Similarly, initial / f ' / (spelt ph) is chosen for the verb (rather than / p ' / ) when the construction is negated, as in phrynodd John fawr iawn ο gig 'John did not buy very much meat'. This type of alternation is known as the Initial Mutations. Their aphonological conditioning is so unusual, typologically speaking, that they have been placed "outside the realm of sandhi phenomena. But since they develop from sandhi phenomena, they naturally belong in a survey of sandhi types" (cf. below, p. 608). As this survey is primarily addressed not to the Celtic specialist, but to the linguistic typologist, it will present the initial mutations as a phenomenon even at the expense of restating some of the information which is routinely given in textbooks. The Celtic specialist will, however, note a number of new facts which we have observed in colloquial Welsh and which modify the typological picture, notably: (i) The productivity of the nasal mutation under specific phonological (not syntactic) conditions, as pointed out above. (ii) The genetic relationship between the residual (not the productive) nasal mutation of Welsh and the eclipsis of Irish. (iii) The genetic relationship between the spirant mutation of Welsh and the prothetic / h / of Irish. (iv) The productivity of the lenis and spirant mutations under specific syntactic conditions.

Typology of the Celtic mutations

107

2. The morphophonemic paradigms The Welsh consonant system features: (i) The three usual places of articulation 'labial', 'alveolar', 'dorsal' for the stops, nasals and glides, but two more for the spirants. (ii) An aspirate correlation not only for the stops and spirants, but also for the liquids and nasals. (iii) Technically, /P1/ are not liquids but spirants in Welsh (Pilch 1972: 208). Thus we can present the Welsh system conventionally as a table: Table 1 labial

alveolar

prepalatal

lateral

dorsal

aspirate

P'

t'

k'

non-aspirate

Ρ

t

k

aspirate

f'

V

non-aspirate

f

Ρ

aspirate

m'

η'

non-aspirate

m

η

stops s'

s'

Γ

χ'

spirants 1

nasals aspirate

r'

non-aspirate

r

Ό

liquids glides

w

j

h

The gap for the non-aspirate / χ / is, by an idiosyncratic array of speakers, filled with an uvular / R / , under which / r ' / and / r / coalesce. Such speakers are found in all areas and in all social strata. My impression is that about one speaker in twenty practices the uvular / R / , thereby removing the isolated row of liquids. The aspirate correlation is usually presented as one of voice. However, this is unrealistic at least for North Welsh. The voicing of medial non-aspirates can be realistically considered a redundant peculiarity of some (chiefly Mid-Welsh) dialects. The initial mutations of Welsh can be adequately presented as a paradigm of consonant alternations:

108

Herbert Pilch

Table 2 Radical /pV /tv /kV

/p/ /t/ /k/ /IV /TV

/m/ /n/

Lenition

Spirantization

Nasalization

/p/ /t/ /k/ /f/ /p/

/fV /pv /χν

/mV /nV /η'/ /M/

/n/ /r,/

zero

/I/ /r/ /v/

zero

/mV /nV /h/

The Irish paradigm is cognate with the Welsh paradigm3, but the phonological details differ to the point that the label 'nasalisation' appears absurd for the fourth column in Irish, being replaced by 'eclipsis': Table 3 Radical

Lenition

/p/ /t/ /k/ /b/ /d/ /g/

/{/

/{/ /s/

/m/ zero

Spirantization

Eclipsis /b/ /d/ /g/ /m/ /n/ /I)/ /v/

/h/ /%/

/v/ /γ/ /γ/ zero /h/ /V/ /h/

Each row of tables 2 and 3 presents one set of alternating consonants. Each consonant is enclosed in slant lines, as the alternation is between different phonemes, not between different allophones of the same phoneme. Each column in the two tables presents a set of consonants which will be chosen under particular syntactic conditions. For instance, those conditions under which / k ' / alternates with / k / (as in Welsh cig ~ gig above) are the same as for / ρ ' / ~ / p / , / t ' / ~ / t / etc. (disregarding certain exceptions, see fn. 5 below). The mutations are productive in the sense that they operate over the

Typology of the Celtic mutations

109

whole of the lexicon apart from specific exceptions (Pilch 1975: 97). In addition, there are certain residual alternations. They involve the same morphophonemic paradigms, but their operation is limited to a closed set of words. Such is the case of: (i) The residual nasal mutation of Welsh. Its domain is the word blynedd 'years' determined by one of the cardinal numerals from chwe 'six' onwards, as in chwe mlynedd 'six years'. In addition, it is found in the word dydd 'day' determined by the historical pronoun beu- 'every' in beunydd 'every day'. Irish eclipsis occurs, inter alia, under the same syntactic conditions (even though eclipsis affects an open set of words). This is one major reason why the two alternation paradigms appear to be cognate. (ii) The residual spirant mutation of Welsh. It affects a few words governed by the morphemes ar, dy-, gor-, tri 'three': ugain 'twenty' cän 'song' cuddio 'hide' cant 'a hundred'

~ ~ ~ ~

un ar hugain '21' dy-chan 'satire' gor-chuddio 'hide' tri chant 'three hundred'

3. Phonetic coherence Nobody will be surprised that, under given conditions, / p ' / should alternate with / p / in Welsh just the same as / t ' / with / t / , / k ' / with / k / , but why, under the same conditions, / p / with / f / , / k / with zero? And why not / ρ / ~ / f / under the same conditions as / ρ ' / ~ / f ' / ? The difficulty hails from the distinctive feature framework which, for all its sophistication, remains squarely based on the time-honoured concepts of place-and-manner of articulation. For the second column, the difficulty disappears as soon as we use the concept of 'air flow energy' (Petersen 1957) to interpret the vague label of 'lenition'. In each instance, the radical consonant has more air flow energy than its lenited correspondent, for instance / p ' / has more than / p / , / p / more than / f / . True, we have not actually measured the airflow energy, but we are judging it by imitation label (Pike 1943:16-23). This is surely a drawback. But, then, even the conventional distinctive features are no more than just imitation labels, not measured quantities (Pilch 1974: XVIII-XIX).

110

Herbert Pilch

For the third column, the difficulty is resolved by the concept of 'channel noise'. What we segment as an initial / h / is phonetically a "voiceless vowel" (Pike 1943 : 71-72), i.e. an articulation with channel noise rather than the usual channel resonance. By the same token, the nasals / m n / can be articulated with channel noise rather than the usual channel resonance. The result is segmented either as a voiceless, aspirate nasal /m' η'/ or as a sequence nasal + / h / , i. e. /mh nh/ - depending on the preference of the analyst. Again, we can prefix initial channel noise to the aspirate stop /pV. The initial channel noise then dominates over the usual initial closure phase. At the same time, the local (labial) noise dominates over the channel noise. The result is local (labial) noise with no initial closure. This is what we segment as / f ' / in Welsh and Breton - and so on for / t ' / ~ /pV, /kV ~ / χ ' / . The phonetic coherence of the fourth column of Irish is usually interpreted in the same way as in Welsh (see section 1 above), as far as the (radical) voiced stops / b d g/are concerned. As for the voiceless consonants / p t k f / , the preceding nasal with which they are co-articulated at one stage imparted to them its voicing (rather than its nasalization), while its nasalization passed to the preceding vowel. The vowel nasalization was, at a later stage, lost before the stops (the nasal consonant survives to this day before the vowels), thus -Vmp- > -Vb- > -b- (Martinet 1956:290-291). The major difference is that the co-articulation is a historical reconstruction for Irish and also (with a few differences of the phonological table) for Literary Welsh, but has remained a synchronic process of Colloquial Welsh to this day (Pilch 1958).

4. The hierarchy of conditioning factors The syntactic conditions are largely (not fully) the same in Irish and in Welsh for the respective columns to the point that these conditions can be stated within the framework of comparative Celtic Grammar (Lewis-Pedersen 1937: sections 186-207, 217-244). Generally speaking, any given syntagm has its initial consonant (if any) determined by its direct tactic relation with a preceding syntagm. For instance, the object is in direct tactic relation with the preceding verb (even when the subject is inserted between them), as in prynodd John gig (see section 1 above):

Typology of the Celtic mutations

111

sbj. obj.

vb.

Even though the subject John is inserted between the verb and the object, it is only in indirect relation with the object via the verb, and indirect relations do not govern mutations. Now, if the word cig is not at the head of the object syntagm, but in a later position, it is still the initial consonant of the object syntagm as a whole that is determined by the preceding verb, as of fawr (radical mawr), ddim /p/ (radical dim /t/) 4 , bwer/p/ 'a lot' (radical pwer/pV): fawr iawn o—• gig John ddim ο —*• 'ichig John

phrynodd

• phrynodd

bwer o—• gig John

prynodd

At the same time, the initial consonant of the noun cig is determined by its direct tactic relation with the preceding determiner, e. g. by ο in the first and third examples above, by the article y in y cig 'the meat', by different personal pronouns in 'i gig 'his meat', i chig 'her meat', 'i cig 'their meat', e. g. prynodd John y cig 'John bought the meat': det.

noun sbj.

vb.

In this construction, the relation between the noun cig and the verb prynodd is not tactic, as the tactic relation is between the whole object syntagm y cig and the preceding verb prynodd (Mulder 1980: 52, 163-164). As the word y which is initial in the object syntagm starts with a vowel (not a consonant), there is no initial mutation of the object syntagm in this example. The conditioning factors are not quite the same in Colloquial Welsh as in Literary Welsh. This phenomenon is usually handled by describing Colloquial Welsh as 'faulty', but we do not accept this as adequate. Notably, the spirant mutation occurs in Colloquial Welsh in two specific constructions:

112

(i)

Herbert Pilch

The initial stop of the finite verb is replaced by the corresponding spirant5 in the negated predicative syntagm, as in phrynodd John fawr iawn ο gig 'John did not buy very much meat'. The direct tactic relation is, in this instance, with the preceding negative particle. This analytical view is, we believe, adequate, even though this particle is usually represented by zero in Colloquial Welsh: (zero «— phrynodd) •*— (fawr iawn o—• gig)

(ii) Any noun (including verbal nouns) has the spirant mutation when in direct tactic relation with the preceding determiner / 'her', as in ? chig 'her meat', iphrynu 'buy her'. In this construction (but not in the former), even the nasals / m n / undergo the spirant mutation in Colloquial Welsh, as in /im'oini/ 'fetch her', /in'ain/ 'her grandmother', but not is in Literary Welsh. This goes to show that Colloquial Welsh offers not simply a reduced version of the mutations (as compared to Literary Welsh), nor vice versa. For instance, the spirant mutation is not governed by the numeral tri in Colloquial Welsh (as it is in Literary Welsh) - apart from a few residual instances such as tri chant above. The fact that Irish has its "prosthetic h" in both these constructions suggests that - contrary to current views (Martinet 1956: 265, fn.19) the third columns of Welsh and Irish are genetically related.

Notes 1. This is phonetic spelling. The orthographic spelling disregards the aspiration (Pilch 1975: 97-98). 2. Using Mulder (1980)'s notation. 3. As far as lenition is concerned, the genetic relationship has been explained by Martinet (1956: 257-296). 4. The negative particles fawr iawn, ddim etc. determine the object in the Welsh construction, not the verb as in the English translation. This is easily seen as those particles cluster with the object (not with the verb) even under the word order transformation, as in phrynodd John fawr iawn ο gig—* fawr iawn ο gig brynodd John 'very little meat John bought', phrynodd John ddim cig—• dim cig brynodd John 'John did not buy any meat*. 5. This alternation cuts across the columns of table 2 above, as both the aspirate and nonaspirate stops alternate with the corresponding spirants, and even / m / with / { / .

Typology of the Celtic mutations

113

References Lewis, Henry - Holger Pedersen 1937 Concise comparative Celtic grammar (Göttingen: Vandenhoeck & Ruprecht). Martinet, Andre 1956 Economie des changements phonetiques (Bern: Francke). Mulder, Jan 1980 The strategy of linguistics (Edinburgh: Scottish Academic Press). Petersen, Gordon 1957 'Breath stream dynamics', in: Manual of phonetics, ed. by Louise Kaiser (Amsterdam: North Holland Publishing Company), 139-148. Pike, Kenneth L. 1943 Phonetics (Ann Arbor: University of Michigan Press). Pilch, Herbert 1958 'Morphologie der Nasalmutation in der Umgangssprache von Nordcardiganshire', Lingua 7: 269-273. 1972 'Aphasia in Welsh', Word2%: 207-229. 1974 Phonemtheorie (3rd ed). (Basel: Karger). 1975 'Advanced Welsh phonemics', Zeitschrift für celtische Philologie 34: 60-102.

Sandhi in time and space

Les phenomenes de sandhi dans un dialecte bas-francique meridional Rene Jongen

0. Nous adopterons pour l'essentiel un point de vue monographique. Nous indiquerons tout d'abord en quel sens il nous parait souhaitable de restreindre le domaine des phenomenes de sandhi. Ensuite, nous passerons en revue un certain nombre de regies ou de proces sandhi, tels que nous avons pu les deceler dans le dialecte bas-francique meridional (Südniederfränkisch) de Moresnet. Notre but est modeste: nous voudrions montrer que les mecanismes de structuration phonologique d'un parier - en l'occurrence les phenomenes de sandhi - peuvent presenter un degre important de diversite et de complexite. Cette complexite n'est nullement inherente - ni necessairement ni exclusivement - au caractere moφhonologique de certains de nos exemples: nous avons enregistre d'une part l'existence d'un systeme assez complexe de regies non-morphonologiques (deux niveaux distincts, l'un preallegro, l'autre post-allegro), et d'autre part l'existence de regies morphonologiques generalisantes (p. ex. R. 3.6.2.) 1.1. Toute description monographique comporte presque inevitablement une composante contrastive interne. Cela est tout particulierement le cas pour l'etat de langue decrit ici (1). II s'agit en realite d'un diasysteme, d'un ensemble de sous-systemes differents en relation de diversite relative parce que motivee (c.-a-d. de diversification, recente ou en cours d'elaboration, a partir d'une identite genetique). Nous nous contenterons ici de distinguer deux sous-groupes d'äge et ne prendrons en compte que les collisions phonematiques qui, d'une fa5on ou d'une autre, donnent lieu ä un traitement sandhi particulier (pour plus de details, voir Jongen 1972 a). 1.2. Moresnet est situe dans l'extreme Nord-Est de la province (beige) de Liege. Son dialecte appartient ä la zone de transition entre le limbourgeois et le ripuarien. L'isoglosse de la 2e mutation consonantique du Hoch- et Mitteldeutsch (Benrather Linie) le separe du dialecte ripuarien d'Aix-la-Chapelle (Aachen), p. ex. Mor. u.t/ ά.ρ / mä. ks = rip. üs

118

Rene Jongen

(fr. / all. ; / ; / ). Le lecteur pourra utilement se reporter a la carte en annexe. II y trouvera en outre des informations relatives ä la distribution geographique de quelques regies sandhi importantes (2). Pour la lecture des symboles, voir ci-dessous; voir aussi Goossens (1965) et la contribution de J. Stroop dans le present ouvrage. / äf / mäxd

1.3. Les systemes phonematiques (voyelles/consonnes) (voir schemas). II s'agit de systemes maximaux (nous negligeons les nombreuses neutralisations intersegmentales et les restrictions distributionnees). Les potentialites distinctives des voyelles pleines sont ä multiplier par quatre, car chacune est susceptible de se combiner paradigmatiquement avec l'un des deux prosodemes binaires suivants: duree vocalique (/V/ vs / V / ) et accentuation rhenane (opposition Schärfung (accent frappe) (/V:/) et Trägheitsakzent (accent traine) (/V./) (voir 1 et 2 dans le schema et exemples de paires minimales) (3). Relevons pour les consonnes quelques particularites importantes (voir numerotation dans schema): 3 a 5 : existence de trois collisions orientees (voir fleches): le deuxieme groupe d'age (jeune generation) utilise, en lieu et place des allophones palataux/velaires du phoneme fricatif voise / γ / , des allophones qui se confondent avec ceux d'autres phonemes: a) avec / g / ä l'initiale devant consonne ([gC-]) (mais / g / a par ailleurs une distribution tres restreinte (/-g/, intervocalique et final (voir cependant R. 3.1. a); b) avec la semivoyelle / j / ä l'initiale (allophone palatal [γ]); c) et avec / R / (fricative uvulaire voisee, qui est par ailleurs Γ allophone usuel de / r / ) (allophone velaire [γ]) (voir les exemples). 6 et 7: situation analogue au niveau de la fricative velaire sourde 8: equivalence phonologique entre / s - / (initial) et /ts-/ En outre: c'est bien l'opposition de voisement (et non de tension) qui constitue la dimension pertinente dans les oppositions b/p, d/t etc.

Sandhi darts un dialecte bas-francique meridional

119

Schemas: 1) i

u

y

e 0 0 e 0e 9 a d

1. / V / vs / V /

*

2. /V./ vs /V:/

2) m b p

ν

f

η d

-η /-gk ~[gC-]/ 3

/

/ y / y ~ f / 4

/ 9~ = 6 / - ζ / 7S/ ' VR/ / z / /s ~(t)s—/ / ! / / w / / h - /

Exemples: l e t 2 : za.i (/) vs za.i (~/ ~) vs ( / ); hd.t (/) vs hö:t (/) 3: Yru3:t~jru9:t~gru9:t (/) 4: yö:t~jö:t (/) vs jy:t, pi. jy.dd (/)

5:

(/); drä.yd~drä:r9 ter>/); vs wq:rg (/) 6: 9n vro:£ ~vro:R (/): lä.£ (/) vro^[9~vrq:r9

7: e.f~e.s(/);

( d) et disparition du / j / (J> 0). 3 a) Phonologisation des produits de la R. 1 (p. ex. / u ~ y / W u / vs / y / ) et lexicalisation des items [o] (voir 2.2.): p. ex. >-*[br0ggi]=> */breggd/(Mor. /bre:g/) b) Moφhonologisation des cas [—]: existence d'un nombre important de suffixes germaniques/westgermaniques contenant l'un des elements definitoires du contexte originel (p. ex. pluriels nominaux en / = i / ou /=ir/; suffixes verbaux 2e et 3e p.sg. en / = i s / et / = i t / ; derivatifs des verbes causatifs en /=jan/, etc.). Les allophones se phonologisent sans cependant se lexicaliser: le changement structural se conserve comme tel (existence d'alternances), mais au facteur contextuel originel (qui s'est perdu, voir 2) a ete substitue par reinterpretation abductive un contexte nouveau, (partiellement) gram-

Sandhi dans un dialecte bas-francique meridional

121

matical: p. ex. certaines suffixations declenchent, automatiquement ou en combinaison avec certaines classes ou entites lexicales seulement, l'application d'une regie (nu^honologique) du Umlaut: p. ex. Verbes [—reg] —• Umlaut / (C) = X (voir 3.8.6) Γ+voc _—cons Cond.: X est le flexif de la 2e ou 3e p.sg. ind. pres. Exemple: /slo:p=9/(dormir/schlafen)=>2e p.sg. [§loe:ps], 3e p.sg. [sl0e:pt] C)

Moφhologisation de l'alternance (qui de phenomene concomitant devient le support meme de l'operation nu^hologique): p. ex. Mor. /za.k/ (/)=>pl. /za.k=Umlt/-^{ze.k]

Ni la regle phonologique originelle, ni ses derives morphonologiques ne relevent des phenomenes de sandhi. Sans doute, des contraintes phonotactiques peuvent-elles se manifester par-delä certains types de frontieres (p. ex. le Umlaut originel s'applique aussi ä Finterieur du mot derivateme ou flecteme), sans doute les regies moφhonologiques sont-elles par definition des parties integrantes d'operations moφhologiques (qu'elles accompagnent et signalent), mais les changements structuraux qu'elles declenchent ou specifient ne sont ä considerer comme phenomenes de sandhi que si les segments affectes sont localises aux frontieres des signes combines. 2.2 Symboles: I. (Frontieres)

II. (par ailleurs)

1)[o]:

sequence phonologique ne contenant aucune frontiere 2) [—]: frontiere moφhologique, soit flexionnelle (H), soit derivationnelle ( H ) 3) [ J : frontiere entre Constituante de mots composes 4) [#]: frontiere syntaxique (clitiques) 5) [##]: frontiere syntaxique (sans pause phonatoire) 6) [ I ]: pause phonatoire 7) [#T#]:les frontieres 2 ä 5 (etc) 1)V (voyelle); C (consonne); (cons, voisee); C (cons, sourde); Q (cons, tendue); ^ (cons, [-tendue]; Τ (d et t);? (coup de glotte)

122

Rene Jongen

2)

5)-:

/

: assimilation regressive/progressive pas d'assimilation en variance libre avec (actualisation phonetique); => (operation morphologique); > (changement diachronique)

3. Les regies: pourchaque regie, nous proposons au moins un exemple et Tun ou l'autre commentaire. Certaines regies sont formulees en termes de contraintes du type: si A, alors B. 3.1. a. Devoisement des obstruentes devant frontiere syllabique (et done aussi en finale absolue) (Silbenauslautverhärtung): Si

[-son] $ 1 Alors [—vse] (+tdu) II s'agit d'une contrainte phonotactique generale. A vrai dire, eile ne releve pas des regies sandhi (absence de combinaison de signes), mais elle en constitue un prealable important. Exemples: /bro:g, pi. bro:g=3 / (/) —• [bra; A, broiga]; /du:v, pi. du:v=a / (/ ) —• [duif, duiva] 3.1.b. Les sonorantes peuvent subir en finale un devoisement partiel ([4-vse—vse]) Exemple: /vo.l/ (/)—•[vo.J] 3.1.c. Particularite: la regie 3.1 .a. distingue toujours les deux antecedents des collisions mentionnees en 1.3. Exemples: 1) /le:v=3~le:j=9 / (/) => l e p.sg. ind. pr. /e.x # le:y/ —• [(e.)ile:9~le:s] vs /vre:j=a/ (/ ) => /e.x # vre:j/ —• [(e.) 5fre:j~9fre:j] 2) /zä:y=3~zä:r=9 / (/) => /e.x #za:y/ —>· [5sä:jf~9säiR~9sä:R] vs /va:r=a/ (/) —• [(e.) 9fä:R] (comp. /jsno:x/ (/) —• [janoDf—janoiR]) OA

Sandhi dans un dialecte bas-francique meridional

123

Cela implique que dans les sous-systemes ou ces collisions sont un fait accompli, certains / r / et / j / doivent etre affectes d'un trait diacritique susceptible de declencher une desonorisation morphonologique (obligatoire dans le cas du / j / ) : {/R/, /j/}

^l9/s] ηιοφίιο—[—son] phon.

[R~R]

Cette analyse est corroboree par l'apparition de formes simplifiees (perte du trait diacritique), p. ex. /kri:j=3/ (/) —* (e.9 kri:j~kri:j] 3.2.a. Assimilation progressive.

!)[##] Si

[—son]

Λΐ Alors

n

2)[##] -vse ,. . 4-tdu

"

(+cont.)J

d —vse 4-tdu [

4.

Exemples: / o . p ##de:z##d0.js=a/ (//) [hea.wQi/Wa.t] 3.3.a. Contexte: devant obstruente occlusive voisee (complement des contextes en 3.2.) 1) contexte: C [ # T # ] b

2) contexte: C

3) contexte:

[#Ta

'g

C [—] d

+vse (—tdu)

+vse —tdu

ou

—vse +tdu

—vse +tdu

ou

+vs6 -tdu

~+vse -tdu

—vs^__^vse -tdu -tdu

-tdu

-tdu

L'assimilation y est au moins partiellement regressive (soit de voisement, soit de tension), sauf devant /*gC—/ (voir 1.3.) Exemples: 1) /a:v w bre9.k=9/ (/) -• [aiv^rea.ks— äivZvea.ka]; /me.t # # brua:d/ (/) -*• [me.i/ftruait ~ me.i/fcruait] 2) /ze.s##yru9:t=9##b6.m=pl. (Umlt) / (5) (/) [ze.s^ruaitabecm ~ zc.skr.. ~ ze.z^r..] 3 a) Suffixe preterit verbes reguliere (voir 3.8.b.) :/=da(n9)/ p. ex. /e.x#ma.k=da(na) / (/) [e.9ma.gi/3(n3)]; pret.de /py.ts=s/ (/rip. ) py.dzd^(ns) b) Suffixe part.passe verbes I et III (voir 3.8.b.) :/=d=/ p.ex. /e.=j9=ma.k=d/ (/) [e.jama.kt] (R.3.1.) => forme flechie [e.jsma.gi/a]; /pl0.k=s/ (/) => pret.le p.pl. /v0e.r##pl0.k=da/ (classel) ou /v0e.r##plp3.=t=3/ (cl. Ill) [voe.rple.gds ~plQ9Ja] => part, passe flechi /j9=pl0.k=d=3~j9=plp9.

Sandhi dans un dialecte bas-francique meridional

125

=d=3##a.p9l=pl. (Umlt)/ -• [japla.gds ~ jeplpa.i/a e.pal] (/ [hya: yds] (/), [kre.qfgJJa] (/), [deibdo] (/)... d) Suffixe des ordinaux (/ = da/) et des noms de fractions de l'unite (/ = dal/), p. ex. / v o . f ^ d a / (/) [vo.v*/a]; / v o . f = d a l / (/ /we.t=t#dar#da.t/ (/) [we.i/arda.t]; /Q.f#dar##kQ:m=t/ (/) [Q.vi/arkpimt]; /var=st0:=t#dar##da.t#dar##da.t # # d ü a . # # m o : t = t / (/) [varstoidar da. I) / *er#racine verbale =t /=> inversion: /r.v.=t#er/ fr.v.=d#erj (R.3.4.a.) => II) changement abductif: /r.v.(=t)#der/, mais le produit de la regle sandhi initiale ( = d#er) s'est maintenu! 2) Dans un petit nombre de composes (etymologiques, mais partiellement demotives (p.ex. ravales au rang de derivatemes)): a) les noms de fractions de l'unite (3.3.a.3.d.): / = d a l / < de.l (/) < germ. *dailab) /de.s=dax/ (/) [de.zdax] (mais p.ex. [doinasiax] (/)); [jp.dzdal] ( (toponyme)/) ( < *jo.tswdä.l) c) /kre.s^dä.x/ (/), /pp:s w dä.x/ (/) ->- [kre.zdä.x], [pQ:zi/ä.x]; /hä.nd w dö:k/ (/) [hä.ndoik] (mais p.ex. /dö.js^döik/ (/) [d0.js/ö:k]); /hü.zjdoeir/ (/) ->· [hu.zi/oeir]. Pour les exemples 2 c, on a egalement la sequence (C^C] (R.3.2.a.2) 3.3.c. Les exemples en 3.3.b.2) semblent etre des vestiges d'un passe revolu. La regle 3.2.a.2) (devoisement d'un / d / initial non-morphemique) serait alors le resultat d'une innovation plus recente. II existe plusieurs arguments en faveur de cette hypothese: 1) / d / est le produit d'une collision des phonemes germ. * / d / et */P>d/ (fricative). On peut supposer qu'avant la collision le comportement sandhi etait conforme aux tendances generates: voisement devant occlusive voisee ((7 d), devoisement de la fricative initiale (C P). Apres la collision, ce comportement differencie peut s'etre maintenu pendant un certain temps (les exemples 3.3.b.2) (tous ont ä l'initiale un * / d / germanique!) et les exemples en 3.3.a.3) en seraient les temoins fossilises), pour ensuite ceder ä une generalisation du principe d'assimilation progressive. 2) Apres cette meme collision, les dialectes limbourgeois les plus proches (designee par le symbole Δ sur la carte en annexe) ont d'abord opere une reinterpretation abductive du contexte de la regle sandhi ( d-initial etaient des mots grammaticaux (6), c'est cette information «mot grammatical» qui est devenue le trait definitoire du contexte determinant de la regle. Par ailleurs, il y a eu generalisation du principe d'assimilation regressive (C d). Schematiquement: 1)

2) a) reinterpretation du, contexte (/P~d=>/mots grammaticaux) b) generalisation [ Ο ] voir carte: Δ

a) generalisation

[

Sandhi dans un dialecte bas-francique meridional

127

3.3.d. S'il n'y avait 3.2.a.2,3.3.b. et les quelques lexemes avec / d z / , nos regies 3.2.a et 3.3.a. se laisseraient unifier en une seule regie, p.ex.:

cont

[—son]

[A] r[_avse

1)-

L

2) [avse]

Γ# 9 #Ί

J

(regressif)

[avse] +cont

(progressif) (7)

Par ailleurs, il semble y avoir un echelonnement geographique, allant d'est en ouest, vers une actualisation de plus en plus proche de l'ideal theorique que represente une telle regie unifiee: A Δ > (voir carte). La situation designee par le symbole A^est^lus proche de cet ideal (que p. ex. Δ), parce que la composante «T#(#)t» a une motivation exclusivement phonetique (assimilation progressive lors de la rencontre de deux obstruentes occlusives dentales) (voir notre contribution: «Sandhi im Südniederfränkischen» ici-meme, pp. 329-383). 3.3.e. II existe une autre exception ä la regie 3.2.a.2), dont la motivation est purement fonctionnelle: on a (facultativement) assimilation regressive devant des lexemes (avec / d / initial) en emploi metalinguistique ou autonyme. La regle est obligatoire lorsqu'il y a risque d'homonymie. Bien sür, on peut toujours aussi recourir ä une pause phonatoire (qui est l'indice par excellence de la mention!). Exemples: /ke:n=s#d3##9t##WQ3.d##do:ndar/ ( (=)?/< Kennst du das Wort «Donner» ?>) -• [ke:nst3(a)tWQ9.t I do:nd3r~3twps.t I to:ndar (voir 3.3.f) ~ ..WQ3.io:nd9r ~ ..WQ9.iio:nd3r]; /9t##wp3.t##dö.nk9l/ (/) [(9)twQ3.tö.nk3l] vs /3t##wp9.d##dö.nk9l/ (/») [(a)twp3.i/ö.nk9l]; /da.t##sri: v=t# ze.x##me.t## de:/ (/) pö./>9S9Ü9.]... Β) Assimilation regressive obligatoire devant mot enclitique, p. ex. /9#mä.k=t#9t#9m/ (/) [ama.gi/öi/am] C) Dans les autres cas (sans pause): a) assimilation regressive de voisement (et de tension); b) assimilation regressive de tension uniquement et coup de glotte; c) absence d'assimilation et coup de glotte Exemples: /me.t##0.x/ (/) ->• [me.i/0.5 ~ me.^0.9]; /do.w##e9.t=s##a:las##o.p/ (/ ) [do.(w)e.^za:tazo.p~~..e.(^?a:l95?o.p]; /o.p w ea.t=3/ (/) [o.£>e9.t9~o./? ?e9.t9]; /at#a.p=ka (Umlt)/ ( (dim.)/) [(9)de.pk9~ (9)(/?e.pk9] 3.4.b. Dans certains composes ou groupes (originels), 3.4.a.2) et 3) sont exclus. II s'agit soit de sequences en voie de (plus grande) idiomatisation, soit de cas d'homonymophobie.

Sandhi dans un dialecte bas-francique meridional

129

Exemples: /ze.s##ü:r/ (/) [ze.züir]; les adverbes pronominaux avec /r/ intermediate, p. ex. /dgi^r^o.p/ (/) [dpiro.p]; les pronoms reciproques prepositionnels, p. ex. /o.pJ./ (/) [o.6e.] (vs /o.p##e./ (/ (neutre)) [o.be.~ o./>?e.]); /a:ne.na:/ (demotive) (/ ) [a:/ie./ia:]... 3.4.c. Devant la conjonction /e.n/ (allegro: /(a)n/) (/), on a aussi: retention de la consonne sourde (sans coup de glotte), p. ex. /dar#hö:t##e.n##d3#he.s=o/ (/) -»- [dar höide.ndahe.sa— hö: pl. /ky:z=a/ [ky:za] Comparer avec 3.1.c. et 3.2.b.: dans les trois cas, on constate un «dephasage» chronologique entre le niveau phonematique (collision) et le niveau des proces phonologiques (qui continuent ä operer sur l'etat anterieur). Si l'on ( = l e descripteur) prend la collision phonematique au serieux, on est oblige de postuler une complexification de la grammaire (p.ex. certains /s/ doivent recevoir un trait diacritique susceptible de declencher un voisement accompagne d'une semi-vocalisation) 3.4.e. La regle 3.4.a. ne s'applique pas aux produits de deletions de /a/ ou de /h-/ (style allegro), p.ex.: 1) /o.pJie:v=3/ (/) -»· [o.phe:va~ o.peiva]; /vo.f##hü.z=ar/ (/) -> [vo./(h)ü:zar] 2) /da.t##ha:=t# za#o.x/ (/) -»- [da.ta] La regie sandhi ignore la disparition (definitive) du */h-/ initial. II n'en va pas de meme pour les formes faibles des autres pronoms avec / h - / initial, p. ex. /h0e:m ~ am/ (/~), /hde:r~ar/ (~/~), /h0e.n~an/ (~/~), p. ex. /za##je:v=t#am##a:las/ (/) [zaje:fv,)/) [de:e.j(a)n t(h)e.j~..e.z(9)nt..] 2) les formes reduites de l'article defini prepositionnel (8): /ana/, / a n / , / a / et / a n / (masculin, feminin, neutre et pluriel) (formes non reduites: /ja(n)(a)/ et /ya(n)(a)/), p. ex. /o.p#ana##ko.p/ (/) [o./^ana ~ o.ftana ko.p] 3.5. Lorsque deux consonnes identiques (ou devenues identiques par application de l'une des regies mentionnees ci-dessus) se rencontrent, il peut y avoir degemination (consonne longue) ou abregement (consonne simple). La tableau ci-dessous montre les possibilites de realisation aux differents types de frontieres (9): -X

X-

+

Si Alors

#X X#

C

+

I c

+

+

+

+

+

+

##

#|#

+

+

+

+

+

Exemples: /we.t=t#dar##da.t/ (voir 3.3.b.l)) *we.d=d#dar [we.i/ar]; /e.s#za##vu:t/ (est-elle partie?/ist sie fort?) [e.javuit];

Sandhi dans un dialecte bas-francique meridional

/o.p^pa.s=a/ (/)

131

[o.ppa.ss~o.

3.6.1. Insertion d'un / s / devant les formes enclitiques du pronom personnel 2 e p. sg.: /dö.w ~ do.(w) ~ da/. Cette regie morphonologique est (devenue) obligatoire (seules les personnes ägees utilisent encore des formes sans / s / apres conjonction se terminant par une sonorante, p. ex. /we:n# dö.w/ (/) -»- [weιndö.w~we:nstö. w]). Elle est le resultat d'une fausse segmentation de la sequence «radical verbal=s#da» (-•[r.v.=s#ta] => /r.v. (=s)#sta/): dans ce contexte (apres verbe), qui est de loin le plus frequent, la realisation superficielle reste ambigue. Cependant, on trouve la manifestation (deductive) de la reinterpretation dans le contexte «apres conjonction». Exemples: /dö.w — da##we.t=s#da.t/ -»• [dö.w~da we.tsta.tl /we.t=s#dö.w ~ da#da.t/ (/) [we.iJtöda.t]; /da.t#do.w~da/ (/) [da.(t)itö.w ~ da.(t)jta] (3.9.C.); /p.f#d9/ (/) [Q.fits] 3.6.2.a. Insertion (facultative) d'un / t / devant /=%../ (moφheme flexionnel avec / s / initial). Cette regie, bien que morphonologique (puisque liee au contexte [/=s..]), decrit neanmoins un proces productif (de generalisation) et dont les motivations abductives sont multiples (voir 3.6.2.g.). Par ailleurs la regie est en principe facultative, mais il semble y avoir des facteurs contextuels qui en favorisent l'application. Manifestement, il existe une relation d'exclusion mutuelle (et peut-etre de complementarite) entre d'une part des «amplifications» de ce genre en position phonatoire forte (10) et d'autre part les reductions en position phonatoire faible (style allegro). 3.6.2.b. Exemples: /dö.w# ma.k=s/ (/) [dö.w ma.ks ~ ma.k/s]; /dö.w#zö:x=s/ (/) -> [dö.wzö:R(0s] 3.6.2.C. La regie ne s'applique pas devant [o] ou devant toute frontiere autre que [=], p. ex. /sna.ps/ (/) [sna./>i]; /ätf^zeij^/ (/) [ä:^e:ja]; /Q.f#za/ (/) - [Q./i3]

132

Rene Jongen

3.6.2.d. La regle est obligatoire dans certains cas, p. ex.: 1) apres sifflantes ou chuintantes en position forte (11), p.ex. /dö.w#pä.s=s##o.p/ (/) [pä.z: /s],

[adblQi/sta— srp:/sta]

3.6.2.e. Contexte favorable: apres semi-voyelle, p.ex. /sre:j=a/ (/), /ho:w=a/ (/), /d0:j=9/ (