Segmental Structure and Complex Segments (Linguistische Arbeiten) 3484303506, 9783484303508

Citation preview

Linguistische Arbeiten

350

Herausgegeben von Hans Altmann, Peter Blumenthal, Herbert E. Brekle, Gerhard Heibig, Hans Jürgen Heringer, Heinz Vater und Richard Wiese

Jeroen van de Weijer

Segmental Structure and Complex Segments

Max Niemeyer Verlag Tübingen 1996

Die Deutsche Bibliothek - CIP-Einheitsaufnahme Weijer, Jeroen van de: Segmental structure and complex segments / Jeroen van de Weijer. - Tübingen : Niemeyer, 1996 (Linguistische Arbeiten ; 350) NE: GT ISBN 3-484-30350-6

ISSN 0344-6727

© Max Niemeyer Verlag GmbH & Co. KG, Tübingen 1996 Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlages unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Printed in Germany. Gedruckt auf alterungsbeständigem Papier. Druck: Weihert-Druck GmbH, Darmstadt Buchbinder: Industriebuchbinderei Hugo Nädele, Nehren

Contents Introduction

1

Ch. 1 Vowel place elements 1. The basic vowels 1.1 Brazilian Portuguese 1.2 The neutralisation of vowel contrasts 1.3 Other languages: segment inventories 1.3.1 Turkish vowel harmony 1.3.2 Morphological considerations 1.3.3 Vowel systems 1.3.4 Patterns in vowel inventories 1.4 Other considerations 2. Processes 3. More mid vowels and additional height contrasts 4. Notation 5. Conclusion

3 4 4 8 13 13 14 15 16 17 18 21 24 24

Ch. 2 Consonant place elements 1. Consonant places of articulation: I, A and U 1.1 Segment inventories 1.2 Other considerations 2. Consonant-vowel interactions 2.1 Interactions involving the element U 2.1.1 Rounding of vowels before or after labial consonants 2.1.2 Labialisation of consonants before or after rounded vowels 2.1.3 Other evidence 2.2 Interactions involving the element I 2.2.1 Fronting of vowels before or after coronal consonants 2.2.2 Coronalisation of consonants before or after front vowels 2.2.3 Other evidence 2.3 Interactions involving the element A 2.3.1 Lowering and/or backing of vowels before or after velar consonants 2.3.2 Dorsalization of consonants before or after back vowels 2.3.3 Other evidence 2.3.4 Discussion 3. Conclusion

27 27 28 30 31 32 32 37 39 42 42 46 47 48 48 49 50 51 51

vi Ch. 3 Complex segments and phonological complexity 1. Complex segments pre-theoretically 2. Phonological complexity 2.1 Assumptions 2.2 Complex segments from a theoretical perspective 2.2.1 Colour mixing 2.2.2 Multiple stricture 2.2.3 Two-rootedness 3. Conclusion

53 53 55 55 60 60 62 65 68

Ch. 4 Head-dependent asymmetries at the segmental level 1. The mismatch condition in metrical and segmental phonology 1.1 Combining two-rootedness and colour mixing 1.1.1 Consonants with secondary articulation 1.1.2 Short diphthongs 1.2 Combining two-rootedness and multiple stricture 1.3 Combining colour mixing and multiple stricture 2. Combining all three types of complexity 3. Conclusion

69 70 73 73 75 76 76 77 77

Ch. 5 Manner primitives 1. Manner 2. Simple segments and natural classes 2.1 Simple segments 2.2 Natural classes 3. Complex segments: multiple stricture 4. Complex segment results 5. Conclusion

79 79 83 83 88 93 100 101

Ch. 6 Colour mixing in consonants 1. Combining the elements 1.1 Combining identical elements 1.1.1 Bilabial vs. labiodental 1.1.2 Dental vs. alveolar 1.1.3 Pharyngeal vs. velar 1.2 Combining different elements 1.2.1 Labial-velars 1.2.2 Labial-coronals 2. Coronal places of articulation

103 105 106 106 109 110 Ill Ill 113 114

vii 3. Uvulars 3.1 Velars and uvulare form a natural class 3.2 Uvulars have secondary velarisation 4. Conclusion

121 122 123 125

Ch. 7 Affricates 1. Spreading [acont] and the Manner-Place dependency in affricates 1.1 Empirical issues 1.2 The Manner-Place dependency in affricates 2. Evidence 3.1 Affricates as single segments 3.2 Segment inventories 3.2.1 Secondary articulation 3.2.2 Place of articulation in fricatives and affricates 3.3 Ordering of [stop] and [cont] in affricates 3. Basque 3.1 Preliminaries 3.1.1 Affricates as single segments 3.1.2 Two monovalent features [stop] and [cont] 3.1.3 The Manner-Place dependency in Basque affricates 3.2 Processes and constraints in Basque 3.2.1 Stop Deletion 3.2.2 Glottal Stop Formation and Aspiration 3.2.3 Sibilant Harmony 3.2.4 Fricativisation 3.2.5 Voice Assimilation 3.2.6 Palatalisation 3.3 Summary concerning Basque 4. Conclusion

127 128 129 133 134 134 135 136 138 139 141 142 143 145 146 147 147 149 151 152 153 155 156 156

Ch. 8 Prenasalised stops 1. Prenasalised stops: three predictions 1.1 Prenasalised stops as unit segments 1.2 Ordering 1.2.1 Ngbaka 1.2.2 The Ganda Law and other rules 1.3 Place attached to [stop] only 2. Alternative representations of prenasalised stops 2.1 Prenasalised stops as contour segments 2.2 Prenasalised stops as two-root complex segments

157 157 162 162 164 167 171 172 172 173

viii

3. Conclusion

175

Ch. 9 / s / plus stop clusters 1. Syllabification 2. Phonological rules and constraints 2.1 Reduplication 2.2 Partial geminates 2.3 Other evidence 3. Ordering 4. Conclusion

177 177 180 181 185 189 191 192

Ch. 10 Two-root complex segments 1. Two consonants under one x-slot 1.1 Clicks 1.2 Labial-coronals 1.3 Postnasalised stops 2. Two vowels under one x-slot: short diphthongs 3. A vowel and a consonant under one x-slot 3.1 Consonants with secondary articulation 3.2 Vowels with consonantal properties 4. Conclusion

193 196 196 202 204 205 208 208 212 217

References

219

Introduction This study offers a theory of segmental structure which pays particular attention to the representation of 'complex segments'. The analysis is limited to manner and place aspects of segmental structure: laryngeal aspects of the segment, such as phonation and tone, are not dealt with. It is also by and large limited to consonants; diphthongs and other vowel types that could be regarded as complex, such as nasal vowels, will receive only limited attention. The study is organised as follows: in the first two chapters I will be concerned with the phonological primitives needed to express place of articulation properties of phonological segments. In these chapters I will illustrate the place elements of Dependency Phonology that I will adopt in this study. These elements will be put to use for the representation of place in vowels as well as in consonants. Chapter 1 illustrates the basic assumptions made with respect to the set of phonological primitives needed to express the place representation of vowels. Evidence will be presented that three vowels in particular, viz. /i a u/, are phonologically basic, which is formally expressed by representing them with a single place element, following Anderson and Ewen (1987). In chapter 2, I argue that the elements used for place in vowels should also be used to represent place in consonants, because the two can affect each other directly. In both chapters emphasis will be on simplex (as opposed to complex) vowels and consonants. The result, then, is a set of primitives that expresses place in phonological segments, and some preliminary ideas about how these primitives can be combined. In chapters 3 and 4 I turn to complex segments in general and discuss their relevance to phonological theory, as well as some problems encountered in earlier attempts to assign adequate representations to such segments. I also give a general formalisation of complexity in segmental phonology, namely as branching in the segmental structure. This is the topic of chapter 3. In chapter 4 I discuss the restricting power of the 'mismatch' condition, which puts a check on possible combinations of instances of branching within the phonological segment. It will be seen that a segment cannot be complex in too many dimensions at the same time. In chapter 5 I will be concerned with the structure of the other phonological dimension that is investigated here, i.e. manner. I will propose a novel geometrical organisation for manner features. In both the proposal for manner and place primitives, a guiding principle will be that simple segments, that is, segments which occur relatively often in sound inventories, should receive simpler representations than complex segments. At the same time, of course, it is imperative to represent all phonological segments, whether simplex or complex, in such a way that their behaviour in phonological processes can be accounted for. Chapter 5 therefore contains, first, a discussion of simplex consonant classes, such as stops, fricatives, and liquids. Second, I will provide evidence that the classes that are expressed by this proposal are in fact natural classes. Finally, I will show that the manner model

2 proposed generates a set of segments which have been regarded as complex in the literature, such as affricates. After having laid out the formalism, I turn to specific studies of segmental complexity on the place and the manner dimension, respectively. In chapter 6 I discuss consonants which have more than one place element in their underlying representation, such as labial-velars, palatals, palato-alveolars and pharyngeals. These places of articulation are different from the 'simple' ones, i.e. labial, coronal and velar. In all cases, the difference can be related to a difference in complexity on the place dimension. From chapter 7 onwards the discussion will focus on segments which are complex for manner, such as affricates and / s / plus stop clusters. Chapter 7 is specifically devoted to affricates. A detailed study of the behaviour of affricates in Basque is presented. Chapter 8 deals with prenasalised stops. In chapter 9 we will deal with / s / plus stop clusters, which are units of a type whose special status has often been noticed, and whose specific representation has been a matter of debate. In chapter 10 I turn to a category of segments that is also complex by the definition of complexity adopted, namely branching within the segmental structure. However, these segments are different in that they are not complex under the root node, but combine two root nodes under a single timing position. Among these are consonants with secondary articulation, but also short diphthongs, and, I will argue, clicks. All these segments will be seen to have phonological characteristics in common, so that the issue of whether or not a segment should be assigned to this class of 'two-root complex segments' is not an arbitrary one.

Chapter 1 Vowel place elements In this chapter I discuss the representation of vowels. The overall approach adopted is that of Dependency Phonology (DP), as described by Anderson and Ewen (1987), in which phonological primitives are assumed to be single-valued (or monovalent, or unary), while constructions in which primitives are combined are assumed to be headed. It is not my objective to reiterate at length the arguments provided by Anderson and Ewen (1987) and others, who argue in favour of single-valued feature theories over binary-feature theories. By Occam's razor, single-valued feature theories are more constrained than binary-feature theories. Work in DP, and in related frameworks such as Government Phonology (Kaye, Lowenstamm and Vergnaud 1985, 1988, 1990; Harris 1990; Kaye 1990; van Lit in preparation) and Particle Phonology (Schane 1984a, 1984b, 1987, 1990; Van Nice 1991) has so far almost exclusively concentrated on the representation of vowels in vowel systems, and of processes concerning vowels such as vowel harmony, umlauting, raising, lowering, and diphthongisation (see the references above, as well as Ewen 1980; Rennison 1983, 1986, 1987a, 1987b; Lass 1984; Ewen and van der Hulst 1985, 1987, 1988; van der Hulst and Smith 1985, 1986, 1987, 1988; van der Hulst, Mous and Smith 1986; van der Hulst 1988a, 1988b, 1989, 1990, 1991, 1995; Durand 1990; van der Hulst and van de Weijer 1991, 1995). Furthermore, Hayes (1990) adopts Particle Phonology primitives to describe diphthongisation processes in various languages. Generally, it seems that single-valued primitives are gaining ground in current phonological theory, whether as Place elements (see Sagey 1986a; Hayes 1990; Selkirk 1991), Manner elements (see Lombardi 1990 for [stop] and [continuant], Itô and Mester 1989; Steriade 1993; Trigo 1993 for [nasal]), or Laryngeal features (Lombardi 1991, 1995) - the same observation is made by Rice (1992: 63). The main topic of this chapter is the representation of vowels. In particular, I will present some arguments which support the claim that the vowels / i a u / are phonologically basic. These three vowels should therefore be represented as maximally simple, which is achieved by identifying them with the phonological primitives I, A, and U: every vowel of the / i a u / triangle possesses one and only one of these primitives in underlying representation. The Manner characteristics of vowels will not be dealt with here. I take up this issue in chapter 3, where it will be argued that vowels have no Manner features like [continuant], for instance because there is never a contrast between vowels which are [continuant] and vowels which are not. I will therefore attach vowel Place elements to the root nodes of vowels (marked as V). Consonant root nodes (marked as C) dominate Manner features and Place features. We will return to the exact status of these root nodes and the labels C and V in chapter 10, where it will be argued that a vocalic root node consists of a unary feature [vocalic], and a consonantal root node of a unary feature [consonantal] (cf. also McCarthy 1988).

4

1. The basic vowels In this section evidence is presented that the vowels /i a u / are basic in vowel systems, and that the elements of which these vowels consist can combine to represent other vowels. The discussion will enable us to introduce some fundamental DP notions. We will focus on data from Brazilian Portuguese (§ 1.1), Russian (§ 1.2), and a number of other languages to illustrate this claim. Finally, some considerations from first-language acquisition and other domains will be presented (§ 1.4).

1.1 Brazilian Portuguese Brazilian Portuguese has different vowel systems according to whether the vowel is stressed or unstressed, and according to whether it appears word-internally or wordfinally. The different vowel systems are as follows (the data are taken from Wetzels 1991, 1993b; see also Da Silva 1992: 66ff.) (1)

ι e ε

u o 0

stressed

U

o

unstressed non-final

unstressed word-final

A full seven-vowel set appears in stressed position, but non-final, unstressed vowels can only come from a set of five. In unstressed word-final position (and, in some varieties, optionally before main stress) a three-vowel system appears, that is, when a final vowel does not bear stress it can only be one of the set / i a u / . This is illustrated in (2) (see Wetzels 1991, 1993b for copious illustration of the fact that the alternations are synchronically active, for instance in verb conjugation): (2)

a.

b.

stressed vowels: s[í]do 'be-PASTPTC' c[é]do 'early' b[é]la 'fair' s[á]co 'bag' unstressed non-final vowels: [i]dáde 'age' [e]rrar 'to mistake' [u]sár 'use'

s[ú]co s[ó]co l[ó]go

'juice' 'blow' 'soon'

h[o]nesto [a]góra

'honest' 'now'

5 c.

unstressed final vowels: sed[i] 'thirst' faç[u] Ί do' faç[a] 'do-lMP'

This situation has parallels in many other languages (compare the case from Russian in the next section, for instance). We ignore nasalisation in the Brazilian Portuguese vowel system, which complicates matters in ways that are irrelevant for the present exposition. For details on the patterning of the nasal vowels, see Wetzels (1991: 79ff.) and Moraes and Wetzels (1993).1 The occurrence of three different vowel inventories can be analysed as three patterns of loss of oppositions between vowels. I will turn to this below. First, however, I wish to focus simply on the representation of the vowels themselves as they occur in the different sets. In unstressed final position the three vowels / i a u / occur. The final position of the word is a position in which the auditory cues must be as salient as possible, especially when the vowel is not stressed, in order to produce a clear signal to the hearer. In other, more prominent positions, finer divisions of vocalism are expected to be permitted. This observation is not new (see, for instance, Trubetzkoy 1939 [1969]: 68). Proponents of single-valued feature systems contend that observations like these can best be expressed in a phonological framework that recognises three basic primitives, which represent these three vowels. This is because the increasingly smaller vowel sets can be related to each other by way of rules of neutralisation, which progressively take more structure away (see below, § 1.2). By implication, the vowels in the smallest set should have the simplest, most basic representations. Thus, the three single-valued features (also referred to as elements or components) proposed in DP express the following properties (see Anderson and Ewen 1987: 28): (3)

I A U

frontness (or 'acuteness' or 'sharpness' or 'palatality') lowness (or 'compactness' or 'sonority') roundness (or 'gravity' or 'flatness' or 'labiality')

The phonetic interpretation of these elements corresponds to a number of (related) acoustic and articulatory properties. The framework therefore seeks to unite insights from acoustically based models like that of Jakobson, Fant and Halle (1952) and articulatory based models like that of Chomsky and Halle (1968). If the elements in (3) occur alone, these elements represent the vowels /i a u/, respectively, which constitute the unstressed word-final vowel system of Brazilian Portuguese. The simplicity of these three vowels is not only expressed by the fact that

6 they have exactly one element in underlying representation, but more precisely by the fact that the relation between segmental material (the elements) and the organising 'root' tier is unmarked, namely, one-to-one. Each vowel of the triangle /i a u / has exactly one Place element, and other vowels diverge from this optimal relationship, either because they have more than one element attached to a single root node, or because they have no element at all (see van der Hulst and van de Weijer 1991: 53 for the same observation). I propose therefore that the one-to-one relation is the formal correlate of simplicity in (segmental) phonology, and that divergence from this optimal relationship results in complex segments (see chapter 3 for further discussion). In (4) I represent the three vowels /i a u/: (4)

V I 1 I

V I 1 A

Ν

/a/

V I 1 u M

The Brazilian Portuguese vowels /i a u/, as in (4), are the only three vowels which are represented by way of a one-to-one relation between root nodes and segmental primitives: in this way their primacy is captured in the phonological representation. The Brazilian Portuguese five-vowel system possesses the vowels / e / and / o / in addition to /i a u/. To characterise the additional two vowels, the elements that make up the three basic vowels are combined: the combination of A and I represents the vowel / e / in such a system, and the combination of the elements A and U represents the vowel /of (the order of the elements is irrelevant). Such simple combinations will henceforth be represented by means of a comma, as in (5): (5)

V ι 1 ΙΑ

V ι I A,U

/e/

/o/

Thus, the vowel / e / is represented as front and low (or 'non-high'), and / o / is represented as round and low (or 'non-high'). The other combination of two elements, that of U and I, which together represent a front rounded vowel, occurs less often in languages. This may be related to a fundamental difference between I and U on the one hand, and A on the other. The former are regarded as 'chromatic' or 'tonality' elements in DP and related frameworks (see Anderson and Ewen 1987: 209ff., 212ff.), while the second element A is regarded as a 'sonority' element (following, for instance, Jakobson 1941, as Anderson and Ewen point out; see also Stampe 1973). In other frameworks, such as Particle Phonology (Schane 1984a, 1984b, among others) and Government Phonology (Kaye, Lowenstamm and Vergnaud 1985; Harris 1990,

7

among others), there are also differences between the vowel elements: in the former, A is the only element that can occur more than once in a vowel representation (which is used to represent vowels at different heights), and in the latter the component A has positive 'charm' while U and I are both charmless (see Harris 1990: 262ff.). That is, Government and Charm Phonology, unlike DP, expresses the difference between tonality and sonority elements formally. I will return below to the question of whether and, if so, how, the difference between elements should be expressed formally (see § 3). Suffice it to note here that the combination of U and I is much less common than that of A with either U or I.2 To represent the Brazilian Portuguese seven-vowel system, i.e. that which appears in stressed position, a crucial DP notion is employed, viz. that of dependency or government. In the mid-vowel series it is necessary to distinguish between / e / and / e / and / o / and /o/, i.e. between a relatively high and a relatively low mid vowel, and the difference is made by assuming that, for the front vowels, both are combinations of the two elements A and I, but that the relation between the elements is different for the two vowels. In the lower of the two vowels, the lowness element A is more prominent, and for the higher of the two the frontness element is more prominent.3 Dependency between elements will henceforth be represented by a vertical line between the elements that enter into this relationship, as in (6):4

2

3

4

V ι 1 A I 1 I

V I 1 I ι 1 A

V I 1 A I 1 υ

ν I I υ I 1 A

M

Ν

M

M

A case in which this relation could be useful is Dutch, which has the vowels /y/, as in fuut 'grebe', and /Y/, as in bus 'bus'. One way in which the contrast between these two vowels could be expressed is by representing the more front vowel lyl with a governing I element and /Y/ with a governing U element. Evidence from stress in Dutch suggests, however, that the former is a long (or 'unchecked') vowel, and the latter is short (or 'checked') (see van der Hulst 1984: 94), so that the difference between the two vowels might be argued to be prosodie, not segmental. In Particle Phonology height gradations are expressed by assigning successively lower vowels more sonority/lowness particles [a] (see Schane 1984a: 39), which is more or less equivalent to an /i-ary feature [high]. A common piece of criticism levelled at the Particle Phonology approach is that there is no principled way of limiting the number of sonority particles a vowel may have (see Clements 1989b, for instance). This is obviously less restrictive than allowing a single type of dependency relation between elements, which predicts that a vowel system may have four heights at most. Note that in (5) the vocalic root nodes are graphically connected to the segmental elements by way of the same kinds of lines as the elements themselves are connected to each other. The former lines represent association lines, while the latter lines express the dependency relationships between the elements. This difference is important, but will not be encoded graphically for convenience of representation. See Ewen (1995) for further discussion of this point.

8 For the contrast between / e / and / e / , as well as for that between / o / and / o / , the lower of the two vowels has a governing A element (referred to as the governor, or 'head'), while in the higher of the two vowels, the same element is governee, or 'dependent'. Note that in a five-vowel system such as the one that has the vowels in (5) above, it is redundant to assume an underlying dependency relationship between the elements. We leave open the question of whether in such a case the dependency relation is assigned at a later stage. Various other notations for expressing the dependency relation between the two elements have been used in the literature. Parallel to notations in metrical phonology, the following representations might also be used (as adopted by van der Hulst 1991, for example): (7)

V

Κ

Α Ι Ν

V

Ν

Ι Α

/e/

In (7) the vertical lines by convention dominate the head elements. The notion of dependency between elements has obvious parallels elsewhere in phonology and other branches of linguistics. For instance, stress may be regarded as a head-dependency relation between syllables (Liberman and Prince 1977), while in diphthongs one of the two vowels is traditionally considered to govern the other to express the fact that diphthongs can be either falling or rising. In both cases headship corresponds to greater salience along some dimension of prominence. For a discussion of the dependency relation in morphology and syntax, see Anderson and Ewen (1987) and Coleman (1992). To conclude, note that the appearance of the three vowel systems in Brazilian Portuguese can be elegantly expressed in the DP framework: in the five-vowel system the head-dependency relation is cancelled, so that the contrast between the mid vowels is neutralised. In the three-vowel system, any combination of elements is disallowed, so that the mid vowels merge with the high vowels. This is not the only possible pattern of neutralisation, however, and in the next section I will discuss such patterns in some more detail.

1.2 The neutralisation of vowel contrasts In this section I discuss a number of neutralisation patterns in vowel inventories. After a brief introduction which argues that neutralisation should be expressed as delinking of a feature, I turn to cases from Russian and Catalan. In all cases known to me the neutralisation tends toward the / i a u / pattern, which lends support to the

9 idea that vowel features are privative and that the elements of which these three vowels are composed are phonologically basic. Trubetzkoy (1939) was one of the first to note that neutralisation, or the suspension of a phonemic contrast, is a common phonological phenomenon. In a neutralisation process or rule, the 'unmarked' member of a two-way opposition is selected in particular circumstances and the 'marked' member is disallowed. That is, when the distinction between two sounds which are exclusively distinguished from each other by the presence vs. absence of a feature is suspended, the 'marked' member of this opposition loses this feature. Perhaps the most often cited example of such a neutralisation is that of syllable- or word-final devoicing in languages like Dutch, German, Polish or Russian: in final position the contrast between voiced and voiceless obstruents is neutralised and only one member, the voiceless one, occurs. The member of the opposition that occurs is considered to be the unmarked member, and the neutralisation is considered to consist of the loss of the mark of the marked member: (8)

Ρ

t

b

d

non-final

final

non-final

g final

non-final

final

Thus, in each case, the voiced member of the opposition is assumed to lose its mark of voicing in final position. (I will not discuss the question of whether [voice] is a binary or a unary feature. For discussion of this point, see Lombardi 1991, 1995): (9)

C

+

[( + )voice]

in syllable-final position

The fact that different vowel systems may appear in stressed or unstressed position (or under other circumstances, see below), may be expressed in the same way as the suspension of consonant contrasts in positions that may also be informally regarded as Sveak' (i.e. syllable-final position above). Let us consider the case of Russian, similar but not identical to that of Brazilian Portuguese above. Under main stress, Russian has a five-vowel system, in which all vowels can be either long or short. In unstressed position the distinction between long and short vowels is neutralised, while the contrast between the vowels / a / and / o / is lost in favour of / a / , and that between the vowels / i / and / e / in favour of / i / . Consider the following examples (taken from Katamba 1989: 142, and Haiman 1972: 371; the transcriptions are slightly adapted for a uniform presentation. Stress is marked by an acute accent):

10 (10)

gorod nog god

[górat] [nok] [got]

'city' 'of the legs' 'year'

goroda noga ν godu

[garadá] [nagá] [ν gadú]

'idem-PL' 'leg' 'in the year'

lies dielo miesto

[l'és] [cPéla] [m'ésta]

'forest' 'thing' 'place'

liesa diela miesta

[ljisá] [d'ili] [m'isti]

'idem-PL' 'idem-PL' 'idem-PL'

Words which have the vowels / i a u / under stress do not change the quality of the vowel when stress shifts. Thus, the following neutralisation patterns obtain: (11)

a

i =>

a

=φ

o stressed

i

e unstressed

stressed

unstressed

This phenomenon can be expressed in completely parallel fashion to final stop devoicing above, if the oppositions between / a / and / o / , and / i / and / e / , respectively, are regarded as privative, i.e. if it is assumed that / o / and / e / have properties which / a / and / i / lack. If the vowel / o / and / e / are represented as combinations of the elements A and U, and A and I, respectively, (as in the five-vowel system of Brazilian Portuguese above), the suspension of the contrast between / a / and / o / can be described as the deletion of the element U, and the suspension of the contrast between / i / and / e / can be described as the deletion of the element A. This is shown in (12) below. When it is irrelevant whether a dependency relation exists between two elements, a combination of elements will henceforth be represented with slanted lines: (12)

V

V

A I

U

A

e -» i

o -* a

Note that Brazilian Portuguese / o / behaves differently from Russian / o / : the former merges with / u / (see above), whereas the latter merges with / a / . Both neutralisation patterns result in an / i a u / pattern. I do not have an explanation for the fact why these patterns should be different. Both can be formally decribed in the same way, namely as the deletion of an element (see also below). Neutralisation of vowel contrasts can also lead to a complete reduction of the vowel quality. Consider the case of Eastern Catalan (Hualde 1992), which has seven

11 underlying vowels which are reduced to three in unstressed position, in the manner shown in (13) (Hualde 1992: 376): (13) o —

stressed

unstressed

Examples (also from Hualde 1992: 376) are given in (14): ull camió cosa menja terra pasta llibre

[ύλ] [kamja] [kóza] [méjiía] [téra] [pásts] [Xíbra]

'eye' 'truck' 'thing' 'he eats' 'land' 'dough' 'book'

ullera camionet coseta menjar terreta pasteta llibreta

[uXérg] [kamjunét] [kuzéta] [majiZá] [t9réta] [pasteta] [Xibréta]

'eye-glasses' 'truck-DIM' 'thing-DIM' 'to eat' 'land-DIM' 'dough-DIM' 'notebook'

This pattern of neutralisation can be analysed as the deletion of an element from vowels which have the element U, but not the element U itself (so that U is retained in / u o o/), the deletion of one element or both elements from vowels which have the element A (the choice between these two options depends on the question of whether the resulting schwa should be analysed as an 'empty vowel' or as an Abearing vowel), and the retention of the element I when this is the only element present. Note that here the behaviour of the mid front vowels is different from that of / e / in Brazilian Portuguese or Russian: in the latter two languages / e / merged with / i / , and in Catalan it becomes / a / . Ultimately, such asymmetries require an explanation, which I will not attempt to provide here. What is important is that the Catalan pattern of neutralisation is also directed towards a triangular vowel system, which can be expressed as the deletion of elements. A pattern of neutralisation that is quite similar to that of Catalan but does not show reduction to schwa is that which occurs in Polish Derived Imperfective formation (Rowicka and van de Weijer 1994), as the examples in (15) show:

12

(15)

Perfective a. zagni[e]áé zapl[e]Sc wymi[e]£c b. zapr[o]sic wyr[o]bié odmí[o]dzicf

Derived Imperfective zagni[a]tac 'to zapl[a]tac 'to wymi[a]tac 'to zapr[a]szac 'to wyr[a]biaé 'to odmí[a]dza¿· 'to

knead' weave' sweep out' invite' shape' rejuvenate'

It is important to note that the position in which neutralisation occurs is not a position which is metrically weak as far as stress is concerned. On the surface Polish has penultimate primary stress, so that the vowels that undergo the change in (15) all bear primary stress (not marked in (15)). This indicates that vocalic neutralisation need not be limited to weak metrical structure, but may also be related to other phonological or morphological properties. Recall that the Brazilian Portuguese pattern of neutralisation discussed in § 1.1 was also partly positional in nature: unaccented vowels in final position could only come from the set /i a u/, while unaccented vowels in other positions could come from a five-vowel set. The important point to observe is that in all cases the neutralisation patterns are toward a triangular three-vowel system. The types of neutralisation patterns found lend support to the idea that oppositions between vowels are privative in nature, if Trubetzkoy's notion of neutralisation as a structure-deletion operation is adopted. Hence, the primitives which should be used to differentiate between vowels are single-valued. In fact, the representation of mid vowels as complex segments predicts that there should be four patterns of neutralisation: either the front mid vowel and the back mid vowel both raise (to [i] and [u], respectively), or they both lower (to [a]), or one raises and one lowers. These four possibilities are presented in (16): (16)

u Î

o

1 I e

u

1

o

e

y Brazilian Portuguese

Russian

\

u t o a ??

\/ Polish

I do not know of any examples of neutralisation patterns in which the front mid vowel lowers to / a / , while the back mid vowel raises to [u]. I will now turn to other considerations that show the primary status of the / i a u / triangle.

13 1.3 Other languages: segment inventories In this section I discuss some cases in which the primacy of the vowels / i a u / or of the elements I, A, U is shown under different circumstances. These particularly concern the appearance of these vowels in segment inventories.

1.3.1 Turkish vowel harmony Prominent among the phonological rules of Turkish is vowel harmony (see Clements and Sezer 1982; van der Hulst and van de Weijer 1991; van de Weijer 1991, among many others). While the analysis of vowel harmony in Turkish has had a profound influence on phonological theory formation in the past century, one interesting aspect has largely gone unnoticed. This point is discussed in Haiman (1972), and briefly repeated in van der Hulst and van de Weijer (1991). Turkish has an eight-vowel system, and differs from a typical five-vowel system in that it has front rounded vowels and a central unrounded (schwa-like) vowel / i / . The representation of such a system is straightforward in a system using single-valued features. All simple combinations of the three elements are employed, so that the head-dependency relation is not put to contrastive use. The eight vowels can be represented as in (17) (see van der Hulst and van de Weijer 1991: 46): V I 1 I Ν

V 1 1 υ Ν

ν I I A Ν

ν ι I ι,υ

V I 1 I,A

V I 1 A,U

/Ü/

M

/o/

V ι I I,U,A /à/

V

Ν

In initial syllables all eight vowels of (17) may appear. 5 However, in non-initial syllables restrictions hold. Because of the vowel harmony rule demanding 'palatal' (or 'frontness') harmony in Turkish, in non-initial syllables the contrasts between / i / and /»/, / e / and / a / , /ii/ and / u / , / ö / and / o / are neutralised, that is, in non-initial syllables "there are only four contrasting, incompletely specified vowels that we may

5

Note lhat in (17) the vowel /»/ is characterised as having no elements at all. This does not mean that it is less marked than vowels like li a u/, because the one-to-one mapping relation is underfulfilled (see page S above). It does suggest, however, that this vowel has a special status in the phonology of Turkish, which is borne out: see van der Hulst and van de Weijer (1991), and van de Weijer (1991b) for discussion. Furthermore, this vowel hardly ever occurs in absolute word-initial position in Turkish, a state of affairs that can also be observed for other languages which have vowels that may be analysed as empty. In Dutch, for instance, schwa does not occur initially except in the unstressed form of the indefinite article een 'a', and in Polish the phoneme /*/ (analysed as empty by Rowicka and van de Weijer 1992) is also excluded from initial position (although other vowels may occur initially in these languages). Finally, in French, loo, schwa cannot occur word-initially (Wetzels 1987).

14 represent as A, I, O, and If' (Haiman 1972: 367). Regular stress in Turkish falls on the final syllable, so that this case is fundamentally different from the one in Russian, where reduction was governed by whether the vowel in question had stress or not. Turkish morphology and phonology seem to conspire to eradicate O from noninitial syllables as well, thus deriving the optimal {A I U} set. First, there are no suffixes which have underlying O and whose vowels would therefore surface phonetically as / o / and / ö / . The only exceptions are the progressive verbal suffix -yor, a once-autonomous root that has only recently degenerated to the status of a suffix, and the diminutive hypocoristic suffix -os. Both of them are invariable, and do not obey the rules of vowel harmony. Second, the other vowel harmony rule, that of rounding harmony, only applies to high vowels. This condition also prevents the occurrence of / o 0 / in non-initial syllables, as it prevents / a e / from being rounded. These two apparently unrelated facts of Turkish ensure that phonetic / o 0 / will never occur in canonical Turkish non-initial syllables. The optimal vowel set can be represented by the elements I, A and U alone, which illustrates the phonological primacy of these elements. The situation in Turkish is comparable to that of Kimatuumbi, a Bantu language of Tanzania described by Odden (1990a, 1990b) and Clements (1991). This language has an underlying seven-vowel system, all of which appear stem-initially. In stemmedial position, however, only / i v a / appear. Thus, in both cases a triangular threevowel system appears in a particular position in the word, underlyingly in Turkish and surface-true in Kimatuumbi.

1.3.2 Morphological considerations There are also morphological considerations that support the primacy of the vowels / i a u/. For instance, these vowels are the only ones permitted in certain morphemes (Luganda) or surface in the products of reduplication to the exclusion of other vowels in the language (Nupe). One such case was already examined above: in Polish the vowels / a ì u i / are the only vowels that can appear in Derived Imperfective stems, where / i / is demonstrably unspecified for any elements (Rowicka and van de Weijer 1992, 1994). Katamba (1989: 109) observes that "in many Bantu languages with the five vowel phonemes / i e a o u/, only the peripheral vowels / i a u / occur in noun class prefixes which mark the class (or gender) of a noun". This is exemplified for Luganda below, where the noun class prefixes are italicised: (18)

mu-sota mu-ntu ki-ntu ka-tale

'snake' 'person' 'thing' 'market'

mi'-sota ßa-ntu ßi-ntu ßu-tale

'idem-PL' 'idem-PL' 'idem-PL' 'idem-PL'

15

Thus, here the appearance of the two vowel sets /i e a o u/ and / i a u / is governed by morphological factors. A similar example comes from Lelemi, spoken in Central Togo (van der Hulst, Mous and Smith 1986), where the vowels that may occur in prefixes form a subset of the vowels that may occur in stems. In Lelemi, too, the set of prefix vowels consists of a non-high vowel and two non-low vowels, although the realisation of these vowels is /e i u/, rather than /a i u/ as in Luganda. In Likpe, another Togo-remnant language that is analysed by van der Hulst, Mous and Smith, the set of vowels that is allowed in prefixes is /Θ i u/. Finally, in the Cushitic language Iraqw, the vowels /i a u/ "are dominant in Stems; they are the only vowels occurring in Selectors" (Tucker and Bryan 1966: 570). A morphological process where the basic vowel triangle / i a u / might be expected to play a role is reduplication, since it is well known that reduplication may simplify complex phonological structure (see Steriade 1988). We therefore expect that a reduplicated form may tend to neutralise vowel contrasts in the direction of these three vowels.6 This is what occurs in Nupe (Hyman 1970: 67), where root vowels of different qualities surface as / i / or / u / in the prefix in a nominalising reduplication process, namely as / i / when they are front and as / u / when they are rounded (and therefore back) (mid tone unmarked): (19)

/gì/ /ge/ /té/ /tá/ /gú/ /gò/ /tá/

'to 'to 'to 'to 'to 'to 'to

eat' be good' be mild' tell' puncture' receive' trim'

[gigi] tgige] [tité] [titá] [gugú] [gugò] [tuto]

'eating' 'goodness' 'mildness' 'telling' 'puncturing' 'receiving' 'trimming'

In these cases the five-vowel system that occurs in stems is neutralised to a two-way opposition in the reduplicated prefix. Apparently, this takes the form of a rule deleting the element A, even where this occurs alone (as in the reduplicated form of / t á / 'to tell', where default I occurs).

1.3.3 Vowel systems There are numerous vowel systems which have only the vowels / i a u / either in underlying representation or both in underlying representation and at the surface. Katamba (1989: 109) also defends the primacy of /i a u/ on this basis: "languages as 6

For other cases in which reduplication simplifies complex phonological structure, see the discussion of reduplication in Nisgha and Sanskrit in section 2.1 in chapter 9. In these languages, complex segments like affricates and /s/ plus stop clusters reduce to simple segments, which can be formally represented in the same way as the deletion of a vowel element.

16 diverse as Groenlandie Eskimo and Australian Pitta-Pitta have / i a u / as their only vowel phonemes" (see Maddieson 1984 on Groenlandie (language number 900), and Western Desert (number 360), for instance). Other examples include Sanskrit, which has both long and short versions of these vowels, while / e / derives from coalescence of / a / and / i / , and / o / by coalescence of / u / and / a / (Steriade 1988: 93; de Haas 1988: 46, 178ff.). This is not to say, of course, that there are no other three-vowel systems besides those that have /i a u/. Maddieson (1984: 325) gives Nunggubuyu (353), with / i a ui/, which may have developed from an earlier /i a u / system (which all the neighbouring languages have), and Tagalog (414) which has / ι υ a / (cf. also Haida (700) and Jaqaru (820)). Such systems can be analysed with a maximum of one element per vowel. This is not the case for Alabama (759) and Amuesha (824), which have / e o a / , and perhaps Mura (802), with /i o a / . All these systems are also basically 'triangular'. There has also been much discussion of so-called 'vertical' vowel systems, which are claimed to have inventories with just two or three mid vowels, such as / i s a / (see, for instance, Lass 1984: 139; Ladefoged and Maddieson 1990: 97). Below I will briefly discuss the presence of more than one mid vowel in a vowel system, since this presents obvious difficulties for a system of phonological representation in which vowels like schwa and / ¿ / are typically represented as lacking Place elements (see e.g. the Catalan and Turkish cases above).

1.3.4 Patterns in vowel inventories In this section I briefly examine the vocalic inventories of a handful of languages taken from Maddieson (1984), which show /i a u / in particular phonologically defined positions. Such examples include the following: Yurak (language 056 in Maddieson's survey) has short / i e a o u / , and 'overshort' /i a u / (the same situation obtains in Lungchow (406)). Dagbani (110) and Telefol (606) both have a five-vowel system for the long vowels, and /i a u / for the short ones: (20)

short i

long u

a

i: e:

u: o: a:

To account for such patterns, consider the fact that the short vowels are phonetically less salient than the long ones, so that maximum contrast is required in the shortvowel set. This is achieved by allowing them to have only single elements. Since the long vowels are more salient, they are free to employ a larger number of contrasts. Hence, they are permitted to make use of element combination.

17 On the other hand, however, short vowels are normally assumed to have unmarked length, while long vowels constitute a divergence from this norm. Languages typically have more short vowels than long vowels, for instance. Hence, we also expect languages to exist in which / i a u/ occur as long vowels, while a larger set of short vowels occurs. This is the case in Fur (203) (and compare also Neo-Aramaic (255) and Bardi (357)): for both the pharyngealised and the non-pharyngealised vowel sets there are different sets of long and short vowels, and the set of short vowels is the larger: (21)

short i e

long u o

i:

a

u: a:

For both cases the appearance of /i a u / i n a particular set of vowels lends support to the idea that these have a special phonological status. In Chipewyan (703), neutralisation occurs in the nasalised vowel sets: (22)

oral i e

nasalised u o

a

í

ü ä

As expected, the vowel oppositions are neutralised in circumstances where the oral signal is muffled by the nasalisation and must therefore be as salient as possible (see e.g. Fletcher 1972). Obviously, the oral vowels can also be regarded as unmarked as compared to the nasalised ones (cf. for instance Trubetzkoy 1939 [1969]). Hence, the nasalised set is expected to be smaller in all circumstances. The same situation obtains in Nupe (Hyman 1970: 59) and many other African languages, e.g. the Khoisan language / / A n i (Voßen 1986). To conclude, in this section we have seen that the optimal vowel set / i a u / occurs in various phonologically and morphologically defined circumstances: these vowels may occur under stressless position, when they are short, when they are long, when they are nasalised, and when they occur in particular morphemes. Let us now turn to some additional considerations concerning the primacy of the vowels /i a u/.

1.4 Other considerations The vowels /i a u/ are also the most common vowels across all languages. Maddieson (1984: 125) gives the following figures, for a 314-language sample:

18 (23)

/i/ /u/ /a/

occurs in 290 languages (91.5%) occurs in 279 languages (88.0%) occurs in 266 languages (83.9%)

He states: "the three vowels at the corners of the conventional vowel triangle, / i a u / , are the most widespread, but note that there are 24 [sic] fewer languages with / u / than with / i / . These three vowels might be expected to be equally favoured, because they each lie at an acoustic extreme" (Maddieson 1984: 125). If such a preponderance in frequency is to be captured in phonological representations, the DP representations which assign one element to each of these vowels seem eminently suited for the purpose. Note that in Maddieson's figures the vowel / a / is reported as absent from almost fifty languages: it seems unlikely that these languages have no low vowel; rather they would not seem to have a back or mid low vowel (i.e. [a]). For instance, not included among the 266 languages reported to have / a / are languages like Shilha (256), which has an /i ae u / pattern. Perhaps even more pertinent than the statistical preponderance across languages is the observation that the vowels / i a u / are also the ones that are acquired first by children in first-language acquisition (Jakobson 1941 [1968]; Anderson and Ewen 1987: 210; Ingram 1989: 19; Levelt 1994: 126ff.). This is also an indication that they have a more 'simplex' status than other vowel types. In later chapters in this study, we will see that the three elements employed to represent these vowels are also fundamental in that they define the primary types of secondary articulation, viz. palatalisation, labialisation and velarisation (or pharyngealisation). Note that at least the first two of these are also typically the second ('weak') elements of diphthongs, where maximum salience is also required.

2. Processes We now turn to the description of phonological processes to illustrate the framework of vocalic Place representation adopted in this framework. Again, the purpose is not so much to present a comparison with any binary-feature framework, or some version of Underspecification Theory (see Archangeli 1988), but merely to illustrate the formalism, which, as far as basic operations such as spreading and delinking are concerned, is quite similar to any other version of autosegmental phonology. The main focus in this study is on issues of representation. For this purpose, we will consider vowel harmony in Yawelmani, where we present an analysis from Durand (1990: 295ff.), which in turn is based on an analysis by Ewen and van der Hulst (1985 : 40ff.). The vowel system of Yawelmani Yokuts, a Penutian language of central California, is given in (24) (we disregard the length contrast, which is irrelevant for the present discussion):

19 (24)

i

u o a

This vowel system is clearly asymmetrical, since it lacks the vowel / e / . With regard to the representation of these vowels in terms of elements, only the vowel / i / requires the element I. In this language / i / is the epenthetic vowel, which is an argument for regarding it as the underlyingly least specified vowel. On the basis of these facts, Durand (1990: 296) assumes that / i / is the default vowel. H e assumes that empty vowels will receive the I element by default. We might also assume, equivalently for our purposes, that empty vowels are phonetically realised as / i / in this language. Thus the underlying vowel inventory and default rule can be represented as follows: (25)

V

Ν

V I 1 u

V U,A

V I 1 A

M

M

/a/

empty vowel — I With these representations it is possible to give an account of the Yawelmani harmony system which has been a topic of debate in non-linear phonology (Archangeli 1984). The data in (26) below illustrate that vowels become round to the right of a round vowel of the same height. In (26)a there is a perfective suffix whose vowel alternates between [i] and [u], and in the data in (b) there is a conditional suffix whose alternants are [a] and [o]. xat-mi bok'-mi xil-mi dub-mu max-al hud-al giy'-al bok'-ol

'having eaten' 'having found' 'having tangled' 'having led by the hand' 'might procure' 'might recognise' 'might touch' 'might find'

With the representations in (25), rounding harmony can now be simply stated as spreading of the U element to a vowel which either shares an A element with the spreader or which also lacks an A element (see van der Hulst and van de Weijer 1995 for discussion of such 'height' harmony systems). Two sample derivations are given in (27) (where rounding harmony is abbreviated as ' R H ' and default I assignment as 'Def):

20 (27)

U

a.

RH.

Γ - - . . c

ν

I d

c

c

I b

I m

ν

[dubmu]

I

b. Def. C

V

I χ

C

C

V

C

I V C

C

I 1

I m

I

I I 1 m

V

[xilmi] (27)a illustrates regular spreading in [dubmu], namely to the suffix vowel which is underlyingly unspecified for any features. It is also the target for spreading after vowels which do not have A. The derivation in (27)b shows the operation of default Iassignment. 7 To account for the data in (26)b above, it is assumed that the suffix vowel is specified only for A, and will therefore only display roundness harmony after the low round vowel / o / : (28)

U

U RH. C

V

C

V

C

C V

C V

I b

I I I k' I A

I ι I A

I ι

I b

I k'

I I

I I

c I 1

[bok'ol] Other processes such as vowel coalescence also find a straightforward formalisation. In language after language we find processes merging [a] and [i], as well as [a] and [u], resulting in [e] and [o], respectively. Such processes took place in the history of English (as shown by the contrast between spelling and pronunciation of words like main, baud, etc., which were pronounced [main] and [baud] in Chaucer's days, but

7

It is immaterial here whether both vowel slots receive their own I element, or whether this is inserted on one of them and then spreads to the other.

21 had mid-height monophthongs in early modern English), Sanskrit (see above), Tamil (Steever 1987), Ancient Greek (de Haas 1988), and many others. The fact that feature systems - although capable of expressing the change, of course - predict that this coalescence pattern should be no more expected than others must be regarded as a disadvantage of these models. These changes can be expressed entirely straightforwardly in the DP formalism; indeed, an [e] or [as] type mid vowel seems the only possible outcome of a merger of [a] and [i] (see Gorecka 1989: 176ff. for similar observations). For detailed descriptions of these and other processes, such as umlauting and other cases of vowel harmony, see the sources mentioned in the introduction to this chapter. In general, the theories adopted by the reseachers mentioned there restrict the space in which hypotheses about certain phenomena can be postulated, simply by virtue of the fact that there are fewer primitives to spread, delink, etc. For instance, DP, with its three basic primitives {I A U}, predicts that there should be only three basic types of vowel harmony, while a binary-feature theory that recognises, say, five features predicts that there will be ten such harmony systems. Van der Hulst and van de Weijer (1995) show that the match with the actual systems found is much closer for the unary feature approach than for the binary feature approach.

3. More mid vowels and additional height contrasts Although the three elements posited so far characterise a fair number of vowel contrasts, other machinery may have to be invoked to be able to represent less common vowel types. Consider, for instance, the vowel inventory of (RP British) English, which is reproduced in (29) (from Jones and Gimson 1977: xv): (29)

i:

u: ι

υ 3:

3 ε ae

λ

o: υ

α: The triangular shape of this sytem suggests representations like the following for the English short vowels (cf. Ewen 1981):

22 (30)

V I 1 A I 1 I

V I 1 A I 1 u

V

I U

V I 1 I 1 1 A

M

/e/

/¡fi/

M

/a/

V

V

V

I I

I A

Ν

/Λ/

The long vowels may be represented as in (31) (here the vocalic root nodes are all linked to two timing positions, indicated by the x-slots, following for instance Clements and Keyser 1983): X

\ / V I 1 I

/ir/

X

\/ V I 1 υ

/u:/

X

X

\V/ I 1 A,U

M

X

X

X

X

\/ V ι 1 A

/a:/

/a:/

Observe that in English the vowels / ι λ V/ can be identified with the elements I, A and U, respectively, while for instance in Turkish the vowel represented by the element I was / i / (see (14) above): the same phonological representation is used to represent two phonetically different vowels. A question that might be asked is whether this way of representing vowel contrasts should be allowed (see Anderson and Ewen 1987: 219ff. for discussion of this 'normalisation' strategy). In the present study it will be assumed that it is, at least for underlying representations. Just as there can be little argument about the fact that, for instance, Dutch / t / and English / t / are both underlyingly sufficiently specified as a Coronal stop (while they differ in phonetic realisation, the [t] in Dutch thee 'tea' being denti-alveolar, the [t] in English tea being alveolar), the underlying specification of vowels should first of all be determined by the particular system in which they function (see also Lindblom 1990: 136f.). The phonetic realisation does not have to be stipulated in the underlying representation. One type of vowel that I wish to consider in more detail is schwa, as for example in English butter [bAta], where it is not a 'reduction vowel' (Anderson and Ewen 1987: 218), since it does not alternate with a full vowel (such as / a / in photography, which alternates with / a : / in photograph). I will take the position that both underlying schwas and reduction schwas have the same representation, namely of a skeletal point specified for vocalicness, but not for Place (as in (30) above). When schwa occurs together with more mid or back unrounded vowels in a phonological system, the system developed thus far is not able to represent both vowels. Maddieson (1984) lists a number of vowel systems with both high mid and central mid vowels. Consider the system of Kashmiri (018) (whose vowels occur short as well

23

as nasalised; in addition, the language is reported as having an 'overshort' i-lik e vowel, which does not concern us here): (32)

i

i e

u o

a Either the high or the central mid vowel can be represented as an underlyingly empty vowel, but not both. One way of solving this problem is by way of a 'centrality' element | a | , as proposed by Anderson and Ewen (1987). These authors propose that this element is also needed in the representation of high back vowels. A similar solution is proposed within Government and Charm Phonology, which has an extra element called the 'cold vowel' (v°), which on its own represents / i / , and may combine with A to represent schwa. Another way of dealing with vowels like these has been developed by van der Hulst (1990) (see also van der Hulst 1988a, 1988b). He also notes that 'extra notational machinery' is needed to represent a vowel system like that in (32), but points out that the introduction of a new element is problematic, because this increases the combinatorial possibilities and may therefore be too powerful. As an alternative, van der Hulst proposes to formalise the dichotomy between the elements I and U on the one hand, which are 'tonality' elements, and A on the other, which is a 'chromaticity' or 'sonority' element (see also p. 6 above). He proposes to express this difference structurally, such that the former two elements are dominated by a single 'tonality' node, and the latter by a 'sonority' node: (33) o sonority ^ ^ o

tonality

The extra structure postulated may be put to use for the representation of a series of back, unrounded vowels, as van der Hulst shows. Such vowels have sonority or tonality nodes but no elements attached to these nodes. In this way the set of elements does not need to be enlarged. However, note that in that case the tonality and the sonority nodes themselves have element-like properties, which makes it difficult to evaluate this alternative. This is the reason why van der Hulst (1995) proposes a component high, which in his framework has a different status than the other three elements.

24 Another question involves languages which are reported to have more height contrasts than the systems considered so far. For instance, these would have five front vowels / i e e as a/. If such vowel systems exist (a question that is not taken up in this study, but see Lass 1984: 146f.), extra representational tools must also be invoked, either by introducing additional elements or by adopting additional notions such as 'mutual dependency' (see Anderson and Ewen 1987: 26ff.).8 The dependency relation would seem to be suited best to express the gradual character of the vowel height dimension. It should be observed that notions like 'combination' and 'dependency' also add complexity to a phonological system, so that a system which makes use of combination of elements alone is more highly valued (i.e. simpler, and thus predicted to occur more often) than one which makes use of combination of elements as well as dependency, etc. However, since the existence of such five-height systems has not been convincingly demonstrated (see Ladefoged and Maddieson 1990: 94ff.), the extra kind of dependency does not seem required.

4. Notation For convenience, I will recapitulate some of the notational conventions introduced earlier. Vowel and consonant root nodes are abbreviated as V and C, respectively. Place features in vowels are attached directly to the vocalic root node, and, as we will see, to Manner features in consonants. Dependency relations between Place elements are characterised by way of vertical domination ((34)a below), simple combinations by means of a comma ((34)b), and slanted lines appear when the question of whether a dependency relation holds or not is not at stake ((34)c): (34)

a. V I A I I

b. V Α,Ι

c. V / \ A I

5. Conclusion In this chapter I have presented some considerations pertaining to the fundamental premises of DP representations of vowels. I conclude that representing the vowels /i a u/ in terms of single-valued elements I, A, and U is a good starting-point.

8

Where applicable in this study, mutual dependency will be indicated by a colon between the elements that entertain such a relation, i.e. I:A indicates that I and A are mutually dependent on each other.

25 Vowels as such will concern us no further. However, vowels may interact with consonants in that they may affect their primary place of articulation. The evidence that can obtained from such phonological processes will be the topic of chapter 2. Crucially, I have now laid out the set of primitives that will be adopted for the representation of Place. It is my aim to show that the three primitives {I A l l } can also be adopted for the representation of Place in consonants.

Chapter 2

Consonant place elements

In the previous chapter I discussed the representation of vowels and showed that the vowels / ¡ a u / have a special phonological status, which was captured by the fact that their representation consists of only the elements I, A and U, respectively. In this chapter I make a similar proposal for the representation of consonants. In particular, the view is defended that the places of articulation coronal, velar and labial are the consonantal Place interpretations of the elements I, A and U. Representing Place in vowels and consonants by means of the same set of primitives makes the prediction that Place in consonants and vowels may interact. For instance, it is expected that a vowel may spread its Place property or properties to a neighbouring consonant, or vice versa, or that the cooccurrence of a consonant and a vowel in a given domain may be restricted by their places of articulation. Section 2 provides a substantial amount of evidence that this prediction is largely correct. It will be seen that interactions involving the elements I and U are straightforward. However, the evidence regarding the element A (which is used to represent lowness in vowels and dorsality in consonants) is not so easy to interpret. Section 3 contains a brief discussion of the available evidence.

1. Consonant places of articulation: I, A and U In this section I investigate the primary places of consonant articulation. In the previous chapter the fundamental vowel places were found to be high, front, unrounded (/i/), low, unrounded (/a/), and high, back, rounded (/u/), and these vowels were assigned representations containing one element each. In this chapter the primary consonant places of articulation will be shown to be coronal, velar (or dorsal) and labial. These places of articulation will be identified with the same three primitives {I A U}. The question of whether these places of articulation are primary is logically independent of the question of whether they should be identified with the same primitives as those for vowels. Identifying them with the same primitives accounts for a number of interactions between consonant and vowels, and these form the subject of § 2. What would constitute evidence for the phonological primacy of the labial, coronal and velar places of articulation in consonants? First, there is again a statistical argument. Languages typically have stops at three places of articulation, and these places of articulation are typically labial, coronal and dorsal. The situation is not so clear for fricatives, however, and other considerations again play a role for liquids, which are typically coronal. A second piece of evidence, which has not been explored in the literature, may come from patterns of neutralisation in which consonants are involved. Consider a language with velar and uvular stops, which neutralises the contrast between these in

28 specific positions, or is in the process of losing the contrast altogether. In such cases we might also argue that the direction of the neutralisation is evidence for which of the two Place categories, in this case velar or uvular, is primary. Similarly, data from language acquisition may provide evidence for which of a number of places of articulation is primary: we might expect labial, coronal and velar places to be acquired earlier than other places, such as palatal or uvular. Such considerations will be the topic of § 1.2 below. Let us first briefly examine the occurrence of labial, coronal and dorsal places of articulation in phonological systems.

1.1 Segment inventories In this section we draw on data provided by Nartey (1979) and Maddieson (1984). Nartey (1979: 17) points out that, statistically, "languages usually have at least three primary oral stops, while / t / , /k/, and /p/, respectively, are the most frequently occurring primary oral stops in our data" (note that the coronal place of articulation includes dentals, alveolare and segments that were reported as either dental or alveolar; the total sample of languages was 317): (1)

segment: (voiceless)

number of languages:

place of articulation:

t k Ρ

292 284 276

coronal velar labial

q c t

41 52 30

uvular palatal retroflex

190 180 200

coronal velar labial

5 28 24

uvular palatal retroflex

(voiced) d g b G

j

The preponderance of the coronal, velar and labial places of articulation over the uvular, palatal and retroflex ones is clear from these figures. Nartey suggests that the

29 reason for their preponderance may well be the acoustic salience of stop articulations at these places of articulation.1 Nartey's conclusions are confirmed by Maddieson's (1984) study. Languages typically have three places of articulation for stops (53.9% of all languages have exactly these three places, see Maddieson (1984: 31, table 2.4)), while over 99% of the languages in his survey have stops at bilabial, coronal and velar place of articulation (but may in addition have stops at other places as well) (Maddieson 1984: 32, table 2.5). It is clear, then, that there is a tendency for languages to have three places for stops and that there is a very strong tendency to have bilabial, dental or alveolar, and velar stops. If this primacy is to receive any phonological interpretation in terms of the representation of Place, an identification of these places with primary elements seems the only reasonable one. This is the course adopted in most recent work. Consider Sagey's (1986a) feature geometry, where the Place node dominates three nodes, labelled Labial, Coronal and Dorsal:

This proposal is an improvement over that made in SPE, for instance, where two binary features, [± coronal] and [± anterior], defined four primary places of articulation, viz. labial, coronal, palatal and dorsal. In Sagey's proposal, the class nodes Labial, Coronal and Dorsal function like single-valued features: they can be either present or absent. In this respect, Dependency Phonology diverges from Sagey's proposal in that a dependency relation may be posited between elements if there is more than one present in a particular segment (as was the case for vowels). A second difference between the DP approach and that of Sagey is the the use of 'subsidiary' features under each of the separate nodes, such as [± anterior] or

1

Vieregge (1970) showed that /p t k/ are 'quantal' in the sense of Stevens (1972, 1989).

30 [± distributed] under Coronal. DP tends not to use such features, because they do not satisfy the assumption that phonological primes are single-valued. This entails that the distinctions for which these features are responsible should be made in another way. In chapter 6 I will return to this issue.

1.2 Other considerations In this section I will provide some phonological evidence to support the claim that the three places of articulation labial, coronal and dorsal are primary, even in phonological systems which have segments produced at more than just these three places. Consider the case of Labrador Inuttut (Smith 1977). This language has velar and uvular stops ( / k / vs. / q / ) , but the contrast between the two is neutralised in final position. Without going into the representation of uvulars here (see Ch. 6, where it will be argued that uvulars are complex segments, viz. velars with secondary velarisation), we predict in such cases that, ceteris paribus, the opposition will be neutralised in favour of the velar category. This is what happens in Labrador Inuttut. Second, consider the question of a language which has palatals and velars. I will argue that palatals should be analysed as complex segments combining the coronal and the velar place of articulation (see again Ch. 6). This predicts that palatals should be acquired later than either pure coronals or velars. The same goes for a language with velars and uvulars: velars are predicted to be acquired earlier. In the one study in which both predictions can be more or less systematically checked, Straight (1972) on Maya, it appears that during the acquisition process the palato-alveolar and palatal fricatives (Straight 1972: 15) are more frequently subject to deviations from the adult pattern than segments at other places of articulation (see Straight 1972: 207ff.), although "no absolutely consistent incremental developments were discernible" (Straight 1972: 207, 215ff.). Obviously, more research specifically directed towards this issue is needed here. To summarise the first section, we have presented some considerations to establish the primacy of the three places of articulation Labial, Coronal and Dorsal. As with vowels, this primacy can be expressed phonologically by identifying these places of articulation with single elements for Place. In the next section we will examine consonant-vowel interactions which show that these elements appear both in consonants and vowels, so that we will refer to them as U (which occurs in labials), I (which occurs in coronals) and A (which occurs in dorsals). An advantage of Dependency Phonology labels is that no position is taken that these labels are either articulatory or acoustic: whereas 'Labial' is clearly a designation of a place where a segment is articulated, U also corresponds to the acoustic properties associated with such an articulation. A certain amount of 'abstractness' of such primes seems to be called for, as consonant-vowel interactions show.

31

2. Consonant-vowel interactions In recent years a number of reseachers have pointed out phenomena that show that consonants and vowels are represented by the same set of Place primitives, or, at least, that there are certain problems with the SPE proposal in this respect. A point that received a great deal of criticism was that in SPE entirely disparate sets for Place in consonants and vowels were assumed (see Reighard 1972; Campbell 1974; Smith 1988; Pulleyblank 1989; Clements 1989a, 1990b, 1991, 1993; Levelt 1994, who all argue for a (partial) identification of Place features for consonants and vowels). The idea that consonants and vowels should be represented by the same set of primitives is by no means new. For instance, in Jakobson, Fant and Halle's (1952) acoustically based feature framework high, back (rounded) vowels and labials and high front (unrounded) vowels and coronals share certain feature characteristics: the former are both [+grave] and the latter are both [ +diffuse]. This relation was given up by Chomsky and Halle (1968), which Smith (1988: 234), for instance, regards as a "retrograde step" for precisely this reason. In the previous chapter we adopted the Place primitives {I A U} for the representation of vowels. In this section I will review some of the evidence that bears on the representation of Place in consonants. Note that only cases are included where the primary place of articulation of consonants is concerned: the addition of a vowel articulation to a consonant articulation, which results in a secondary articulation on the consonant, is dealt with below (see, for instance, § 3.1 in Ch. 10). The proposal made here diverges from that of Anderson and Ewen (1987). These authors observe that the elements used for the representation of Place in consonants and vowels are "not entirely distinct" (Anderson and Ewen 1987: 206), but limit the identity of the Place elements to the Labial element, U. For Coronals, they introduce a 'Lingual' element, L, while dorsals are regarded as combining the elements U and L. Thus, Anderson and Ewen (1987: 237) propose the following representations of Place in consonants (the notation conforms to the one adopted here): (3)

U labials

L dentals/alveolars

L,U velars

This proposal is appealing for various reasons, but I will not adopt it. Let us first consider the advantages. The claim inherent in this proposal is that dorsals are more complex than either labials or coronals. This accounts for the fact that among the primary Place categories this is the Place category acquired last by children (Jakobson 1941 [1968]). Secondly, this proposal expresses the natural class of labials and dorsals, which share the feature U, the grave consonants. Finally, it expresses the class of coronals and dorsals, which, as is well known, are the consonants typically affected by palatalisation: these share the Lingual element.

32 These advantages do not outweigh the disadvantages, however. Most seriously, this proposal does not take stock of consonant-vowel interactions involving front vowels and coronals. In (3), the only feature used for consonants and vowels alike is U, characterising roundness in vowels and labiality in consonants. We will see below that there is strong motivation for representing front vowels and coronals by way of a common element too. Apart from this, even when Lingual is replaced by Coronal, the representation of velars seems inadequate. That dorsals should be more complex than labials or coronals is not borne out by their frequency of occurrence, for instance, since all three occur at over 99% in segment inventories (see Maddieson 1984: 32). Furthermore, the representations in (3) make it possible to express a number of highly unexpected rules as simple spreading operations. For instance, it is predicted that a labial consonant might become a dorsal after a coronal by spreading of the element L. Similarly, dissimilation processes are predicted that would dissimilate labials (or coronals) from dorsals. Although no phonological process should be declared impossible beforehand, there is no evidence that processes like this are at all common. We therefore conclude that Anderson and Ewen's proposal for the representation of Place in consonants is less attractive than their proposal for the representation of Place in vowels. Now we will turn to some rules and processes to support the idea that consonant Place and vowel Place should be represented by the same elements. The rest of § 2 is divided into three parts: first we will examine the behaviour of the element U, and provide cases in which vowels are rounded next to labial consonants (§ 2.1.1), consonants are labialised next to round vowels (§ 2.1.2), and other types of interactions (§ 2.1.3), respectively. In §§ 2.2 and 2.3 I will turn to interactions involving the element I (Coronal) and A (Dorsal), respectively.

2.1 Interactions involving the element U 2.1.1 Rounding of vowels before or after labial consonants In this section I investigate a number of cases in which vowels become round under the influence of a labial consonant. Let us start with a simple case to illustrate the basic issues and the line of argument. This case is from Finnish (from Skousen 1975: 52ff.), paraphrased as follows: in the Savo dialects of Finnish, there was a historical rule that rounded a stem-final e vowel in the third person singular present indicative. Consider a verb stem like teke 'to do'. The historical suffix representing the third person singular present was vi, and in the environment of the labial consonant the stem-final e was labialised to ö, as for example in tekevi > tekövi (> tekköö). Another example is the verb stem käske 'to command', which was similarly changed: käskevi > käskövi (> käsköö). The phonetically natural rule simply rounds an unstressed e vowel in the

33 environment of the labial consonant v. The rule Skousen draws up to account for the change is the one in (4): (4)

e

[ + round] /

ν

Let us examine this case in some detail. In the SPE feature framework, where labials are represented by the feature combination [-coronal, + anterior] (e.g. Chomsky and Halle 1968: 307), and rounded vowels by way of the feature specification [+round] (e.g. Chomsky and Halle 1968: 309), there is no way of expressing the rule of Finnish as an assimilation rule by way of spreading a feature. However, the spreading operation is the natural way of expressing such assimilatory processes (see Steriade 1987, for instance). To express the rule of Finnish as spreading, we have to assume that the labial place of articulation in consonants and roundness in vowels is expressed by means of the same feature or element. This is obviously the element U. In (5)a below the representation of the vowel / e / is provided, and in (5)b I give the representation of a labiodental fricative / f / (we return to the representation of consonants more fully in later chapters): (5)

a.

V

b.

I Α,Ι

C I [cont] I υ

/e/

/f/

A standard assumption in autosegmental phonology is that identical phonological primitives appear on the same autosegmental tiers (Goldsmith 1976). This assumption is incorporated in (6) by placing the different elements on their own tiers: (6)

a.

V

b.

C I [cont] I U

/e/

[cont] tier U tier I tier A tier

Ν

We can now treat the spreading process in Finnish as the natural rule that it is. In (7), we show that in this process the element U is spread from consonant to vowel,

34

creating a doubly-linked structure, as a result of which the vowel / e / (represented as a combination of the elements A and I) is rounded to [0]:

Ν

- [0]

The Finnish case illustrates the formal identity of labiality in consonants and roundness in vowels, on the basis of the fact that assimilations between consonants and vowels may take place which spread an element from a consonant to a vowel (as in the Finnish case), or from a vowel to a consonant (as we will see below). There are other possible sources of evidence that could establish this identity. For instance, if a labial consonant should change into a round vowel, this could be expressed as a change in the major class specification of the segment in question, without a change in its Place specification, which would be necessary in the SPE framework (i.e. [-coronal, + anterior] ->[ + round]). In the framework adopted here, such a change could be expressed as follows: (8)

C -V I Manner I υ

Unlike assimilation between neighbouring consonants and vowels, which is expressed by spreading, a full change of consonant to vowel is certainly not an expression of a natural phonological phenomenon:2 it is therefore predicted to be rare. If such cases do exist, however, and if rounding is preserved in such cases, these also present evidence for identifying rounding in vowels with labiality in consonants. Below we will see that cases like that exemplified in (8) exist (see § 2.1.3). A third potential source of evidence concerns syllable well-formedness constraints such as ones in which labial consonants are not permitted next to rounded vowels (cases like this are discussed in Yip 1988 and Pulleyblank 1989: see below). In such a

2

This is especially true when the change from consonant to vowel takes place regardless of the position of the consonant: weakening in final or intervocalic position, most usually through an intermediate glide stage, is far more common than processes in initial position.

35 case, the two segments are formally 'too much alike', and therefore not permitted. Consider a potential (but non-existing) onset [pu], for instance: (9)

C

V

I [stop] I U

U

U tier

Here the two U elements occur in sequence, and languages may disallow this as such a sequence violates the Obligatory Contour Principle, of which one formulation (by Mester 1986: 5) is given in (10): (10)

Obligatory Contour Principle: Adjacent autosegments on an autosegmental tier cannot be identical

Thus, if labial stops and round vowels have identical Place elements, constraints are expected that rule out such onsets, unlike onsets like [pi], for instance. No such constraints are predicted if the place of articulation of consonants is described by way of features like [coronal] and [anterior] and Place in vowels is described by way of features like [high], [back] and [round] (as in SPE). As a matter of fact, constraints of the type *[pu] do exist, which again argues for the identity of labiality in consonants with roundness in vowels (see § 2.1.3 below). So far we have uncovered three potential sources of evidence that could establish identity of Place elements for consonants and vowels. Let us now turn to some of the actual evidence, which has been drawn from various sources, prominent among which are Smith (1988) and Clements (1989a, 1990b, 1993). In the Dravidian language Tulu the high unrounded central or back vowel [i] is rounded if preceded either by a labial consonant or by a rounded vowel in the preceding syllable (Bright 1972; Clements 1990b). This is shown in the following examples (taken from Clements): nâçlï ka»! kannï pudarî ugarî ari-n-T

'country' 'bond' 'eye' 'name' 'brackish' 'rice-ACC'

b.

bolpu kappu poppu uccu moroçlu uru-n-u

Svhitener' 'blackness' 'girl' kind of snake 'empty' 'country village-ACC'

In the examples in (ll)(a), the final vowel remains unrounded, and in the (b) examples it is rounded. The rounding of the vowel is predictable: it occurs when the preceding vowel is round, when the preceding consonant is a labial (the one example that is provided is kappu 'blackness'), or both.

36

On the basis of examples such as those in (11), Clements (1989a, 1990b, 1993) also assumes that consonants and vowels are represented by way of the same primitives (which in his model are binary features). Let us briefly indicate the differences between his approach and the one adopted here. In Clements' view, the Place features are connected to different nodes according to whether they occur in consonants or in vowels: the Place features can be connected to a node for 'V(owel)Place' and the same set of features can also be connected to a node for 'C(onsonant)Place'. The representation in (12)a is taken from Clements (1989a), the representation in (12)b is from Clements (1990b, 1993) (irrelevant nodes omitted; {P} is the set of Place primitives adopted): (12) a.

root

V-Place I {P}¡

b. C-Place I {P}¡

This model stipulates that the sets for consonant Place and vowel Place are identical, whereas the one developed here derives this, since in the latter Place elements are interpreted according to whether they appear in a consonantal or vocalic segment: (13)

root I {P}

In (13) the root may be either C(onsonantal) or V(ocalic), and the interpretation of the Place element takes place accordingly. Clements' model in (12) has a number of further consequences that we will not go into in detail here. Clements assumes that it is only possible to spread features between nodes of the same kind, that is, from V-Place to V-Place or from C-Place to C-Place.3 It is therefore necessary that within the segment a particular feature that shows interaction between a vowel and a consonant is first moved (or spread segment-internally) to the node (C-Place or V-Place) that is also present in the target. Apart from the fact that this is not an elegant expression of these processes, it is not clear what triggers this segment-internal feature movement, as Lahiri and Evers (1991) point out. I refer to these authors (as well as to Jacobs and van de Weijer 1992 and Levelt 1994) for a critical assessment of Clements' formalisation of processes of consonant-vowel interaction.

3

This assumption is abandoned in Hume (1992) and Clements and Hume (1995).

37

In the present model, the tenet that features spread between nodes of the same kind is given up. In (7) above, for instance, the element U spreads from the feature it is attached to (the feature [continuant]) to the target vocalic root node. I assume that as long as the 'docking sites' of segmental elements are well-defined (i.e. root nodes for vowels and Manner features for consonants) there is no arbitrariness as to where elements may link. A final example that shows rounding of vowels next to labial consonants comes from Burmese languages (Thurgood and Javkin 1975). In the history of Burmese, final / - a p / changed to a rounded mid vowel, varying from / o / to / o / . This change was part of a larger group of changes in which final / - a t / changed to / e / or / e / (see § 2.2.1 below), and final / - a k / to / a / or /as/ (see § 2.3.1 below). The correspondence between reconstructed / - a p / and present-day rounded mid vowels in several Burmese languages is presented in (14) (cf. Thurgood and Javkin 1985: Table 1):

reconstructed *kyap *rap *tsap *C-kap *k-nyap •pyap

written Burmese kyap ~

khap hnyap ~

Lahu -

•yò? chô? -

nô? -

Lisu -

yò tshó khò nyó phó

Akha co •γο tSD

xo -

pyo

'to stick into' 'needle' 'to stick into' 'to draw water' 'pinch' 'to turn over'

The other changes in Burmese mentioned above will be discussed below.

2.1.2 Labialisation of consonants before or after rounded vowels So far, we have seen a number of cases in which vowels became rounded due to the effect of a labial consonant. The opposite process, i.e. the acquisition of a labial primary place of articulation of consonants (here referred to as 'labialisation'), due to the effect of round vowels, appears to be not at all uncommon either. It occurred frequently in the history of the Bantu languages, for instance (Guthrie 1970; Clements 1990b). Clements (1990b) points out that consonants at various places of articulation changed to labials before the high back rounded vowel / u / : (15)

pu, bu, tu, du, ku, gu > fu pu, tu, ku > fu, bu, du; gu > vu tu, du, ku, gu > fu tu > vu; du > bvu ku > fu ku > ßu

38 In each of these cases a non-labial consonant changes into a labial. Clements points out that labialisation took place in widely scattered zones throughout the Bantu-speaking area, and must therefore have arisen independently. He observes that this kind of labialisation may be therefore be regarded as a natural phonological process. In the Bantu process the U element of the vowel replaces the original consonantal Place feature. Another example of this type comes from the development of the velar fricative [x] in Late Middle English (Gussenhoven and van de Weijer 1990). Some of the relevant facts are provided in (16) (etymology from the Concise Oxford Dictionary·, LG is Low German): a.

OE trog ME coghe

/trox/ /koxa/

> trough > cough

b.

OE rüh LG sluwe OE swogan OE genog

/ru:x/ /slu :x/ /swo :x/ /jenorx/

> rough > slough > sough > enough

c.

OE dohtor ME slahter OE dohtig OE thurh

/doxtor/ /slaxtar/ /duxti/ /θηιχ/ ( daughter > slaughter > doughty > through

/doita/ /sbits/ /danti/ /Gru:/

d.

OE OE OE OE OE OE OE OE

/9o:xt/ /bo:x/ /aixta/ /swoi-yan/ /9e:ax/ /da:x/ /plo:x/ /slo:x/

> thought > bough > ought > sough > though > dough > plough > slough

/9o:t/ /bat)/ /o:t/ /sau/

thöht bog ähte swogan thëah dag plóh slöh

/traf/ /kof/ /rAf/ /slAf/ /SAf/ /lïlAÎ/

/Ö3V/

/dai»/ /plav/ /slau/

In (16)a and (b), syllable-final [x] corresponds to [f] in the modern forms, while in (16)c and (d) [x] is deleted and a long vowel or a diphthong with a rounded second element appears. The former change can be referred to as the 'consonantal' development, and the latter as the Vocalic' one. Note that the vowel quantity distinctions that existed before the change were neutralised after it. Regardless of whether the input form had a short vowel ((16)a and c) or a long vowel ((16)b and d), a short vowel appears in the consonantal development and a long vowel in the vocalic development. This correlation is exceptionless. In some cases, other processes took place and for some words there

39

exists variation between the vocalic and the consonantal development up to the present day (see Gussenhoven and van de Weijer 1990: 323f.). One question that these data raise is how [x] could change into [f]. One possible answer that Gussenhoven and van de Weijer consider but reject is that [x] changed into [f] as a result of a perceptual error, which was then reproduced in production. Phonological changes of this kind are amply exemplified in Ohala (1989, 1990). Such an account would not explain why nowadays only short vowels appear before [f], while before the change both long and short vowels appeared before [x]. Gussenhoven and van de Weijer argue that the key to an explanation of both changes is a labial glide which developed between the back vowel and [x]. The latter consonant then lost its dorsal place of articulation, retaining, however, its consonantal Manner properties. In the consonantal development, the labial glide filled the vacated slot, moving out of the nucleus and thus shortening it. This part of the analysis is relevant here, since it presents another instance of consonant-vowel interaction: a vocalic U element links to a consonantal slot and receives a consonantal interpretation. If the vacated slot was not filled, on the other hand, the nucleus remained long and the labial glide was interpreted as the second part of a diphthong or combined with the preceding vowel to form a long vowel. I refer to Gussenhoven and van de Weijer (1990: 325-30) for the technicalities of this analysis. A final example of a round vowel giving rise to a labial consonant comes from Prague Czech, where an epenthetic labial fricative /v/ developed before the mid back vowel / o / (de Bray 1951: 446, 451; Hickey 1984: 352): (17)

okno otee on

> vokno > votec > von

Vindow' 'father' 'he'

2.L3 Other evidence The second source of evidence for the identity of vowel Place elements and consonant Place elements that was pointed out above is the (infrequent) change of vowels into consonants. One example of this is provided by Smith (1988: 220): in Limbum (Elias 1979), a Bantu language spoken in Cameroon, the vowels /i u/ developed into the syllabic consonants /z v/ in certain cases: (18)

Proto-Northern Mbam-Nkam *bid *(N)dib *Ngud *(N)bun

>

Limbum bzr Nvar' mdzp 'water' mgvr 'oil' bv 'hip'

40 A second example of the same kind comes from Lendu, an African non-Bantu language (classified as a language from the Moru-Mangbetu group by Tucker and Bryan 1966), which also has syllabic ζ and r, corresponding to (front) vowels in other languages of the same group. The following examples are taken from Tucker and Bryan (1966: 30) (transcription retained): (19)

Lendu tsz dzz drr Tí

Ma'di tj dzè kjdry ky-érá

'cow' 'to buy' Vein' 'medicine'

Another similar case, noted by Smith (1988), is that of Usila Chinantec, where / ü / alternates with syllabic / m / synchronically (Skinner 1962). Smith assumes that such changes can be formally expressed as a change in the feature [consonantal], leaving the Place features unaffected. Whether the change can be expressed as spreading, or involves the replacement of one major class feature by another, will not concern us here. The question of how major class is represented is discussed in chapter 10. Smith notes that, while a change into consonants is rare for true vowels, it is not at all uncommon for the glides / j w/ to turn into obstruents (Smith 1988: 221; see also Anderson 1976b: 24ff.). This has happened in some varieties of Dutch, for instance, where wreed 'cruel' (which in some varieties still has a labial glide) has become homophonous with vreet 'devour-IMP' (which has a labiodental fricative): (20)

wreed 'cruel' earlier [wre:t] historical change:

[w] -*• [v]

cf. vreet

'devour-IMP'

I adopt the proposal that glides should be analysed as vowels in non-nuclear position (see e.g. Steriade 1984). The change of glides to true consonants may then be regarded as an instance of consonant-vowel interaction. Formally, it would involve the same sort of change as the change of vowels into consonants (see above), but the syllabic position of the target segment is different: for the former change the vocalic root node is within the nucleus, in the latter case it is outside. This may be argued to be the reason for the higher frequency of the latter type of changes. Smith (1988: 227) also presents evidence of another sort. The short vowel * / o / in Scots became unrounded in the 16th century when it was adjacent to a labial consonant: compare RP English /top, p u t / with Scots / t a p , p a t / for top, pot. This may be analysed as a case of dissimilation: two segments that are 'too much alike' are no longer permitted to occur together, on the assumption that the rounding in the vowel and the labiality of the consonant are expressed by the same feature or element. The Obligatory Contour Principle (see p. 35 above) formalises this notion.

41 Similarly, constraints may rule out round vowels next to labial consonants, testifying to their identical Place specifications. This is the situation in varieties of Chinese (Yip 1988; Pulleyblank 1989). In Cantonese (Yip 1988), vowels whose roundness must be underlyingly specified cannot occur after labial consonants. In Mandarin Chinese, rounded glides are never found after labial initials (Pulleyblank 1989). Furthermore, in Ayutla Mixtee rounded vowels do not occur following labial consonants (Pankraz and Pike 1967: 289). Such constraints may also be analysed as dissimilation effects which have been grammaticalised as morpheme structure conditions. 4 They, too, indicate that the roundness of round vowels and the labiality of labial consonants should be expressed by the same phonological primitive. A final source of evidence is the failure of rules to apply: in the Middle English standard variety [u] unrounded to [λ] in most contexts, but was generally preserved after labial consonants and sometimes before the velar [k] (Jespersen 1909; Clements 1990b). The change is assumed to have taken place in the seventeenth century (Jespersen 1909: 330). The following examples are from Clements: (21)

unrounding applies in:

unrounding fails to apply next to labials, as in:

as well as pre-dorsally in:

cut sun

hug husband

gum us, etc.

wool bull

bush pull

push full

cook

look

took

book wolf, etc.

There are several possible formal accounts for the failure of unrounding. One is that in items such as bull, the labiality of the consonant and the rounding of the vowel are expressed by one and the same element, that is, one that is doubly linked. The representation in (22) is intended as an informal characterisation: (22)

b

u V U

1 I 1

Unrounding cannot apply to this linked structure, if the rule is formalised as applying to singly linked structures only (see Hayes 1986; Schein and Steriade 1986 for this 'inalterability' phenomenon). Another possible formal account is that unrounding did in fact apply to vowels such as those in bull, but that the neighbouring labial con-

4

English has similar prohibitions: syllable-initial clusters of labial or labiodental consonants (including [m]) followed by [w]) are not allowed (McCarthy 1988; Pulleyblank 1989). Pulleyblank (1989) notes furthermore that while English does not exclude combinations of rounded vowels and labial consonants in syllable-final position, it does exclude labial consonants after the diphthong [au], which has a rounded second element.

42 sonant caused it to become rounded again. I will not attempt to choose between these analyses; both hinge on the assumption that vowel Place elements and consonant Place elements are an identical set.5 Finally, note that neither analysis straightforwardly accounts for the fact that unrounding also (more sporadically) failed to take place before dorsals. Facts that indicate that labials and dorsals form a natural class will not be investigated here (see Hyman 1973; Vago 1976; Odden 1978 for a presentation and discussion of such evidence). To summarise this subsection, there is a good deal of evidence that indicates that rounding in vowels and labiality in consonants is represented by the same phonological primitive. Following Anderson and Ewen (1987), I propose that this is the element U. Finally, notice that there is no evidence that non-round vowels and nonlabial consonants have a feature or element in common. That is, there is no evidence that the feature [round] is binary, or, put differently, that the negative value of this feature should play a role in the phonology of languages. There are no processes, as far as I know, in which velars and/or coronals unround vowels, for instance, or processes in which non-round vowels cause labials to shift to a coronal or velar place of articulation. Until the existence of such processes is firmly established, the adoption of a single-valued feature [round] or an element U is fully warranted.

2.2 Interactions involving the element I 2.2.1 Fronting of vowels before or after coronal consonants In the previous section the element U was examined. The various kinds of evidence that were adduced there can be repeated for the element I, which has been proposed to represent frontness in vowels and coronality in consonants, diverging from Anderson and Ewen (1987), but following Clements (1976), for instance. Clements (1976) points out that front vowels should be classed as Coronal, even by the standard definition of the feature [± coronal] in SPE, and provides several cases in which coronal consonants interact with vowels. I will repeat and discuss some of these below. The first case is comparable to that of Tulu (§ 2.1.1 above), and involves high vowel reduplication in Fe?fe?-Bamileke, a language of Western Cameroon, the phonology of which is described by Hyman (1972). The consonants of this language fall into two classes with regard to certain phonological processes. These classes are given as A and Β below:

5

Dobson (1957: 508) mentions that [u] underwent "unrounding and lowering to [Λ] (unless protected by the influence of rounding consonants, as perhaps in book)". This implies that no change took place (as in the first scenario).

43

(23)

A Β

{pbfvmw} j t d s ζ η 1}

The principal phonological justification for the classification into two sets comes from the process of high vowel reduplication. The dialect of Petit Diboum is of particular interest in this respect. In this dialect, the normal reduplicative vowel is the high unrounded central vowel [i] (see (24)a below). This vowel is rounded to [u] when (i) followed by [u] in the next syllable, or (ii) followed by a consonant of set A and a rounded vowel of any height (see (24)b below). Furthermore, it appears as the front vowel [i] under parallel conditions, that is, when (i) followed by [i] in the next syllable, or (ii) preceding a consonant of class Β and a front vowel of any height (see (24)c): za to keen ben kuum ko boh siim teen cen

ziza tito kikee bipen kukuu kuko pupoh sisii titee ricen

'to 'to 'to 'to 'to 'to 'to 'to 'to 'to

eat' punch' refuse' accept' carve' take' be afraid' spoil' remove' moan'

As Hyman (1972: 105) points out, the SPE feature system does not allow any simple statement of these alternations, while if it is assumed that front vowels and glides share the feature Coronal, or the element I, with Coronal consonants, these assimilations are quite natural. The case of Maltese Arabic (Hume 1991; Clements 1990b) parallels that of Bamileke. In Maltese Arabic the imperfective prefix vowel is normally a copy of the following stem vowel (25 )a, but it is realised as [i] in case the first consonant of the stem is a coronal obstruent (25)b (note that the coronal liquids do not cause the same assimilation): (25) a.

perfective forok kotor rifed laha? ?asam

imperfective yo-frok yo-ktor yi-rfed yi-lha? ya-?sam

'to 'to 'to 'to 'to

limp' abound' support' /rifid/ reach' /lihi?/ break'

44 b.

dahal siket talab ">abar dalam cahad

yi-dhol yi-skot yi-tlob yi-^bor yi-dlam yi-chad

'to 'to 'to 'to 'to 'to

enter' be silent' pray' collect' grow dark' deity'

A rather different case involves a sound shift described for Tibetan by Michailowsky (1975), Clements (1976), Ohala (1979) and Pulleyblank (1989), among others. A comparison between Written Tibetan and the modern dialect of Lhasa reveals that the following changes have taken place: (26)

Written Tibetan a o u

Lhasa Tibetan e œ y

before Written Tibetan /d η 1 s/

Consonants which were final in Written Tibetan have dropped in Lhasa Tibetan.6 Coronal consonants left traces, however: they left a tone (omitted below), and lengthened and fronted the preceding vowel, while nasals also nasalised the vowel: brgyad bod bdud sman ston bdun bal thol yul

[ke:] [phce:] [ty:] [me:] [tos:] [ty:] [phe:] [thee:] üy:]

'eight' 'Tibet' 'demon' 'medicine' 'autumn' 'seven' Svool' 'extra' 'country'

The above changes can be regarded as the change of back (non-coronal) vowels to front (Coronal) vowels before coronal consonants. In other words, the vowel comes to anticipate the following consonant with respect to one of its characteristics, the raising of the tongue from neutral (or lower-than-neutral) position. Formally, it can be expressed as association to the vowel of the Place element of the consonant that has been deleted:

6

See also Pulleyblank (1989: 382), who states that "in Lhasa Tibetan non-front vowels have undergone fronting before coronals, except -/. These include -n, as well as -d, and -s, which have since dropped. Orthographic zan is [sen], kun is [kyn]". Whether /II participated and whether /n/ was also lost finally does not affect the basic point.

45 (28)

V

\

C

\

ι

Manner \ +

In Burmese the same kind of changes took place (cf. also § 2.1.1 above). The correspondences between reconstructed / - a t / and present-day front mid vowels in a number of Burmese languages are presented in (29) (cf. Thurgood and Javkin 1975: Table 2):

reconstructed *k-r-WQt

*cat *C-sat *S3-wat

*tsat *kyat

written Burmese krwat tshat sat wat tshat - -

Lahu vè? chê?

Lisu wè chwé 8è

-

Si-vê? -- -

si-we tshé ché

Akha ye tse se ye tse ce

'leech' 'to break in two' 'to kill' 'flower' 'deer' 'to run'

The cases above involve assimilatory effects.7 Dissimilation effects form another source of evidence. These also indicate that front vowels and coronals have an element in common, and may therefore be subject to OCP effects. In English, for instance, the coronal glide [j] was lost in American English after coronals: RP [tjuin] [dju:n] [ljuid] [nju:]

AmE [turn] [du:n] [lu:d] [nu:]

tune dune lewd new

The pattern of loss of / j / is sensitive to the place of articulation of the preceding consonant: if this identical to the place of the glide, the glide is lost. Cases like these suggest that coronal consonants have an effect on vocalic place of articulation, whether directly (by fronting them) or indirectly (as in the last case). Let us now turn to cases that show that the opposite effect, that of front vowels on consonants, also exist.

7

Other such assimilatory cases are provided by Pulleyblank (1989) from the history and synchronic phonology of various Chinese and other Asian languages. In Cantonese, for instance, non-low rounded vowels are fronted after coronals: [ty:n] 'end, tip' < Late Middle Chinese /tuan/, Mandarin [twan] and in Pekingese underlying /jan/ is realised as [jen], where the only source for the fronting is the coronal nasal ending.

46 2.2.2 Coronalisation of consonants before or after front vowels One process in which non-coronal consonants (i.e. labials or velars) become coronal is palatalisation (see Lahiri and Evers 1991; Jacobs and van de Weijer 1992, among others). One typical palatalisation change is that from / k / to / t s / or / t j / before front vowels. Such changes can be regarded as examples of coronalisation of consonants in the context of front vowels. Consider the case of Ancient Greek (Allen 1957; Pulleyblank 1989). Among other things, the labialised velars */k w / and */k wh / became [t] and [th] before the front vowels /i e/, */g w / became [d] before / e / (but not before [i]), */k w j/ became [t] in syllable-initial position and [tt] intervocalically. Curiously, plain */k/ was not affected. In other languages with palatalisation rules, however, plain / k / does change into a coronal. Bhat (1978: 58f.) presents a number of such languages: (31)

a. b. c.

In Slavic, /k g χ / became /c ζ s/, respectively, before / a j / and / o j / (Chomsky and Halle 1968).8 In most of the Greek dialects of Lesbos, / k / is changed into / t s / before front vowels (Newton 1971). In Spanish, / k / becomes / s / (through / t s / ) before front vowels (Harris 1969).

Labials have occasionally been noted to change into a coronal under the influence of a front vowel. This is especially common for the glide / w / , which, however, is often not only labial, but labial-velar phonetically (see Bhat 1978: 69). It is also possible for / w / to change into a labiodental / f / under the influence of front vowels, which retains the labial place of articulation, and may be compared to the change of / w / to [v] in Dutch (see above). At least three cases exist in which palatalised consonants are known to have changed into labiodentals or coronals:

This case is especially interesting since the conditioning glide /]/ is separated from the target velar by the vowel /a/: this may be regarded as evidence that the vowel /a/ is not specified for the element I (i.e. that it is back /a/) and that the spreading element I can spread from /)/ to the velar across this vowel without affecting it. Alternatively, it might be regarded as evidence that the two parts of the diphthong are phonologically unordered, so that this would be an 'anti-edge effect' in diphthongs (see chapter 7 for such effects in affricates). The structure of diphthongs will not concern us in detail here, so we leave this issue at this point.

47 (32)

In Bushooq, spoken in the Kasai province in the Congo Republic, palatalised / b / becomes /v y / (van Spaandonck 1964). b. In some localities in the Litomysl area of northeastern Bohemia, the traditional dialects as still spoken toward the end of the 19th century had apico-alveolar /t d n / where Proto-Slavic had */p b m/, in a small and diminishing number of very common lexemes (Jakobson 1938; Andersen 1973: 765). Andersen argues that palatalisation on the labials developed into full coronals, replacing the original primary place of articulation. In Sino-Vietnamese the labial stops / p / and / p l i / of Late Middle Chinese are sometimes represented by dental [t]. The conditioning factor was a following high front glide / j / , as in ti 'low' for Late Middle Chinese pji (Pulleyblank 1970, 1971, 1984a, 1989: 383). Pulleyblank shows that there is no evidence for an intermediate stage in which the labial was a palatal, since palatals have not changed into dentals in Sino-Vietnamese. a.

In the first of these cases, primary place of articulation is not affected, however, and therefore does not count as 'coronalisation'. In the second, the palatalisation that was present on the labial is 'promoted' to primary articulation. It is structurally the same change as that from /k w / to / p / , which has occurred in a number of languages, such as Romanian and Proto-Indo-European (see Campbell 1974: 53). The third case seems to be the best example of a primary labial changing into a coronal, without intervention of a secondary articulation stage or an intermediate palatal stage. To conclude, there are examples of changes in which labials and velars shift to plain coronals, where the only conditioning factor is a front vowel. Such processes indicate that front vowels and coronals have an element in common, here proposed to be the element I. The representations adopted in chapter 1, where / i / is typically represented by only an element I, and other vowels have additional elements, suggest that this vowel should be the most likely trigger of such changes. The same holds for the effects of round vowels on consonants: it would seem that 'unpolluted' / u / has a greater effect than mixed vowels like / o / . This was the case of the Bantu changes discussed on p. 37 above, for instance, where the shift is only noted to occur before / u / .

2.2.3 Other evidence Constraints in which front vowels are disallowed after coronals are not as common as constraints ruling out labials adjacent to round vowels. However, cases can be found in the literature. For instance, the glide / j / is disallowed after coronals and the palatal stop in Korean (Martin 1951; Clements 1990b): "the phoneme / j / does not occur in syllables beginning with / t e s / , or including the vowel /i/" (Martin 1951: 526). The same constraint can be postulated for American English (see above), which do not allow / j / after the coronals /t d η 1 s/.

48

To conclude this subsection, the evidence strongly indicates that front vowels and coronals have a Place element in common, since coronal consonants can front back vowels, front vowels can coronalise velars and labials, and there is some evidence that constraints exist that rule out coronal consonants and the glide / j / in the same syllable. We have postulated that this element is I. Let us now turn to the third element of the set A. 2.3 Interactions involving the element A In the framework of vowel representation developed in chapter 1, front as well as back non-high vowels have an element A. In the introduction to the present chapter (§ 1) I proposed to represent the velar place of consonant articulation by the same element (following Smith 1988). This predicts that velars should be able to lower vowels, and that low vowels should be able to retract consonants. These predictions are not fully borne out, or rather, the evidence concerning the element A is generally much less clear than for the other two elements, I and U. In the following three subsections we will present the available evidence that can be found in the literature, and in § 2.3.4 we will discuss its theoretical implications.

2.3.1 Lowering and/or backing of vowels before or after velar consonants A number of examples exist that show that velar consonants seem to have the ability to back vowels, rather than to lower them. Since in chapter 1 we assumed that the element A in vowels correlates with [ + low] (or [-high]), and not with [+back], the assumptions might have to be reconsidered. We discuss this issue in § 2.3.4 below. Let us first turn to a review of the available evidence. Herzallah (1990) argues that certain exceptions to Ablaut patterns in Palestinian Arabic are due to the presence in roots of velar, uvular, or 'emphatic' consonants. In this case, the regular ablaut vowel is replaced by [u] in the imperfective. Herzallah argues that this is due to the fact that the back consonants spread their feature [ +dorsal] to the vowel, after which it is subject to a redundancy rule providing the feature [round] (or [labial]). Kiparsky (1992) argues that velars in certain dialects of Finnish may impose a [back] value on following vowels, interrupting the vowel harmony pattern (examples taken from van der Hulst and van de Weijer 1995): (33)

a.

b.

itikka etikka tiirikka nimi neljä heinä

'mosquito' 'vinegar' 'lock pick' 'name' 'four' 'hay'

nimikko nelikko heinikko

'namesake' 'a quarter, barrel' 'hay field'

49 In the examples in (33)a the velar consonants apparently force a back vowel to surface, in an otherwise front-harmonic word. This is an idiosyncratic effect, since fully front-harmonic words with velars also exist (e.g. Helsinki). In (33)b the velars are part of a suffix, which, although attached to a front-vowel word, has a back vowel. The evidence about velars is contradictory, however. Consider the behaviour of the velar in the Burmese sound changes discussed above (cf. §§ 2.1.1 and 2.2.1). The changes with a final velar are more diverse than the changes involving final labials and coronals: "the *-ak rhyme either remains an -a in Lisu or becomes -ae. In Sani (another Loloish language) all *-ak rhymes go to -e. In Modern Burmese, the reflex of *-ak is -e followed by a glottal stop. At any rate, the vowel fronts in all those cases in which it changes" (Thurgood and Javkin 1975: 162; my italics, JW). In Lahu, back / a / changes to front / a / . Some examples of the changes involving the velar are given in (34) (cf. Thurgood and Javkin 1975: Table 3):

reconstructed *zak *wak *nak *?gak *lak

written Burmese sak wak m9nak a-khak lak

Lahu yà? và? nà? qá là?

Lisu yàe á-váe née sí-ká lè

Sani ze ve ne qe le

'to descend' 'Pig' 'early morning' 'branch' 'hand'

Thurgood and Javkin (1975) point out that these changes are more complex than those involving labials and coronals and offer a tentative acoustic explanation for all three patterns. It is clear, however, that it is peculiar that vowels back before velars in one language (e.g. Estonian, cf. also !Xóó below) and front before velars in another (Burmese). The effect of velars on vowels therefore needs further research (see § 2.3.4 below for discussion).

2.3.2 Dorsalization of consonants before or after back vowels Let us now examine whether low or back vowels can affect consonants, in the sense that they should shift their primary place of articulation to a dorsal (or perhaps pharyngeal) site. A complex set of changes involving vowels, glides and consonants occurs in Maxakali, an indigenous language of Brazil (Gudschinsky, Popovich and Popovich 1970; Reighard 1972; Clements 1990b; van der Hulst 1991; Wetzels and Sluyters 1995). Maxakali has the vowel inventory / i i e o a/. There are two processes that are relevant. First, in tautosyllabic VC sequences, an extra-short unstressed vowel is inserted before the consonant, agreeing with it in nasality. Before velar consonants, the vowel is back. Second, under certain circumstances vowels can turn into sequences

50 of vowels plus glide (or consonant) plus vowel. The nature of the resulting consonant or glide is determined by the quality and nasality/orality of the vowel, as follows: (35)

V i î o ö

C/glide j j w w

V » í a ä

C/glide Ύ q γ q

The Place properties of the vowel are retained in this process. Both mid central unrounded [i] and [a] result in velar fricatives. If the former is analysed as an empty vowel, we must assume a default rule to provide the consonant with an A default Place feature. Axininca Campa (Yip 1983; Black 1991) has a three-vowel system /i a o/. It also has three glides, namely at bilabial (not labial-velar), palatal and velar places of articulation. The last segment, the velar glide, has been the source of some controversy. Yip (1983) proposes that it "is the vowel / a / associated to a C skeletal slot" (see Black 1991: 189). Black points out some problems for this analysis: however, there does seem to be a case for consonant-vowel interaction effect here, since "the velar glide only surfaces when it is surrounded by low vowels" (i.e. / a / ) (Black 1991: 188). There is obviously a relation, then, between the velars and the vowel / a / . In this language, this vowel is low as well as back (Black 1991: 186), so that it is impossible to establish whether it is the property of lowness or of backness that is responsible for the interaction.

2.3.3 Other evidence Reighard (1972) provides extensive evidence that round vowels and labial consonants, as well as back vowels and dorsal consonants, are related in terms of their featural composition. Consider the following deletion data, from the history of French: (36)

Latin secururuga tabula pavore

French sûr rue tôle peur

'sure' 'street' 'steelplate' 'fear'

Reighard points out that "what appears to be going on is that intervocalic labials are deleted in the environment of a round vowel and intervocalic velars are deleted in the environment of a back vowel. Since round vowels are redundantly back, the two environments are the same" (see also Clements 1990b: 15).

51 2.3.4 Discussion The cases examined in the preceding paragraphs indicate that our approach of identifying A (=non-high) for vowels with A (=dorsal) for consonants might have to be reconsidered. Modifying the proposal for Place in consonants and vowels might go in different directions. One possibility might be to advocate the adoption of a fourth Place element, corresponding to backness in vowels and dorsality in consonants. This element might then be used in languages which have a contrast between schwa and /»/, for instance (see § 3 in Ch. 1). However, in my view it has not been sufficiently motivated that such an element is necessary in the representation of vowels. It would also lead to problems since the combination of this element with the element for frontness (I) cannot be interpreted. Furthermore, the introduction of such an element would predict the occurrence of a type of vowel harmony, namely backness harmony, which seldom if ever occurs (see van der Hulst and van de Weijer 1995). I will not try to weigh the advantages and disadvantages of the introduction of an element for backness. I will keep to the representation of vowel Place by way of the elements I, A and U, and represent consonant Place by way of the same elements because this seems the most economical solution, with the proviso that A in vowels and A in consonants seem to correspond to different phonological and phonetic parameters.

3. Conclusion In this chapter we have attempted two things. First, we argued, primarily on the basis of frequency of occurrence but also on the basis of other evidence, that the primary places of articulation can be identified as Labial, Coronal and Dorsal. It is reasonable to assign these places simple types of representations. For example, SPE represents Labial as a combination of features, which does not bring out its inherent simplicity. I have also argued that these representations should be (partly) formally identified with those for vowel Place, that is, they should be in terms of {I A U } (and not, for instance, { L C D } for Labial, Coronal, Dorsal), because consonant Place interacts with vowel Place in a wide range of phonological phenomena. The evidence was straightforward for the labial and coronal places of articulation. In the case of non-high vowels (represented with an element A), however, things are not simple. We found no evidence for the assumption that lowness in vowels corresponds to dorsality in consonants. Rather, dorsal consonants have a backing effect on vowels, while clearcut cases of dorsalisation of consonants in the neighbourhood of (low or back) vowels are absent.

Chapter 3

Complex segments and phonological complexity

In the previous two chapters, some motivation was provided for adopting the primitives {I A U } for the representation of Place in vowels as well as in consonants. These chapters were focused on the representation of 'simple' Place categories, such as 'high, front' for vowels, or 'velar' for consonants. This chapter offers a first exploration of the representation of complex segments, that is, segments that diverge from the one-to-one relation that characterises such simple segments. Place is not the only dimension that plays a role, however: the category Manner also enters into the discussion. A full treatment of Manner primitives is presented in chapter 5. The emphasis in this chapter is on a theory of complex segments. That is, just as syntactic theory seeks to generate all possible sentences of a language and only possible sentences, a theory of phonological segment representation should characterise all possible segments and only possible segments, including complex ones. In such an approach, 'complexity' is a theoretical notion which must be defined, and not an intuitive notion that follows from tradition, as we will see. The chapter starts with a general discussion of complex segments and their relevance to phonological theory (§ 1). Then I propose a definition of complexity in phonological theory, and give a preliminary overview of the kinds of segments that are expected to occur, given the assumptions made in the previous chapters and this one concerning Place and Manner primitives, and other segmental and suprasegmental structure (§ 2). Particular segment types will then be considered in more detail in the later chapters. The final section presents conclusions and a further outlook.

1. Complex segments pre-theoretically What is a 'complex segment'? 1 Traditionally, a number of segment types have been described as such, for example consonants with secondary articulation, such as /ρ*/ or /k w /, affricates ( / p f / , /ts/), labial-velar stops ( / 1 φ / ) and pre- or postnasalised stops ( / n d / , / b m / ) . Such segments have a number of characteristics in common. First, they have a non-homogeneous internal phonetic structure: on the assumption that they are one segment, one part of the segment is realised differently from another part (specifically, in such a way that the parts into which a complex segment in a given language can be divided can also occur as independent phonemes in the same language or in some other language). For instance, the first part of an affricate is realised as a stop, while the second part is realised as a (homorganic) fricative. Consonants with secondary articulation such as / p ' / have an initial stop articulation 1

Part of this chapter appeared in a slightly different form as van de Weijer (1991a).

54 (in this case produced at the labial place of articulation) and a following off-glide (in this case produced with a high front tongue position), although the glide part may overlap the primary articulation to a certain degree. Complex segments may also show a combination of places of articulation: a labial-velar stop like /kp>/ combines a stop articulation at the labial place of articulation with a simultaneous velar stop articulation, in contradistinction to either a simple labial or a simple velar stop. Another characteristic of complex segments is that they may violate certain tendencies that segments normally observe. As is well-known, an important argument for regarding prenasalised stops, for instance, as single segments is that when these segments occur syllable-initially, as in [mba], they violate the sonority hierarchy, or rather the sonority sequencing generalisation (Selkirk 1982), which demands that segments with a higher sonority value occur closer to the syllable nucleus (i.e. the vowel) than segments with a lower sonority value. That is, a nasal is expected to occur closer to the syllable nucleus than an obstruent stop (and indeed many languages have stop plus nasal onsets), because nasals are more sonorous than obstruent stops. As prenasalised stops violate this generalisation, an explanation may be sought in some (as yet undefined) special status that these segments might have. The same goes for affricates, in particular when they occur word-finally. Here the more sonorous fricative part occurs after the stop part. In recent years the internal structure of complex segments such as these has become a topic of debate. It is important to investigate these segments for different reasons. First, complex segments have been represented as segments which have two phonological specifications where simple segments have one. That is, soon after the SPE proposal that affricates were represented as different from stops in that the former were represented as [-delayed release] and the latter as [ +delayed release], proposals were made to represent an affricate as being made up of a specification for a stop (i.e. [-continuant]) and a specification for a fricative (i.e. [ +continuant]) (see e.g. Campbell 1974; Anderson 1976a). The simple fact that segments are represented by means of two specifications suggests that they form a potentially powerful source of evidence with respect to the set of phonological primitives, and to the question of how these primitives may combine. The fullest statement of this line of investigation appears in Sagey (1986a). Let us consider her representation of affricates, given in (1) (irrelevant nodes omitted): (1)

C

/\ [ - cont]

[ + cont]

While simple stops have a [-continuant] feature, and fricatives a feature [ +continuant], affricates have both. Representations like these have come under attack in recent years on theoretical and empirical grounds. Empirical arguments will be considered later (see Ch. 7). Here I will focus on the theoretical side. The assumptions adopted in this study, one

55 of which is that phonological primitives are single-valued, would not allow for a representation of affricates as in (1). Even if all binary features are replaced by pairs of unary features, a theory like Sagey's faces the question of whether segments which combine the elements corresponding to the opposite values of features other than [ ± continuant] also exist, and whether it is a coincidence that the minus value of the feature appears on the left and the plus value on the right (and if the feature values could also be sequenced in the opposite direction). Sagey's theory as it stands does not provide answers to these questions, and, at worst, is therefore not a theory about complex segments: it does not provide a calculus which generates all possible segments and only possible segments, although this is the goal that current phonological theories in segment representation set out to attain. The model developed here is intended to be a step on the way towards such a theory. In such a theory, what counts as a 'complex' segment is based on theoretical considerations, and is not defined by what the tradition dictates. Of course, it would be surprising if the set of segments that the tradition has labelled 'complex' turned out to be very different from the set of segments generated by a given theory. Fortunately, the present theory does not make very different predictions about the set of possible complex segments; there are, however, a number of clear divergences, which will be discussed in detail. Let us now turn to a formal perspective on phonological complexity, and devise a theory that allows us to generate a set of complex segments.

2. Phonological complexity This section is organised as follows: in § 2.1 I lay out the basic assumptions that underlie the model that I argue for and propose a general definition of complexity in phonology. The model proposed allows for three types of complexity. I will call these types colour mixing, multiple stricture and two-rootedness\ these terms will be explained and illustrated in some detail in § 2.2. All three are instantiations of the general concept of segmental complexity, namely branching within the segmental structure.

2.1 Assumptions In this section I will make explicit some more of the assumptions that are directly relevant to the issue of complex segments. First, let us discuss the units of phonological representation. In the previous two chapters I have discussed the set of Place primitives, for vowels as well as consonants. Besides these, there is obviously a set of Manner primitives. These two dimensions, Place and Manner, are the primary two dimensions on which a segment can be complex. Henceforth, we will refer to complexity on the Place dimension as 'colour mixing' and to complexity on the Manner dimension as 'multiple stricture'.

56 I propose that all complexity in segmental phonology is the result of branching. All types of complex segments mentioned in the introduction involve some type of branching.2 Branching is formally defined as the presence of more than one entity under the same node, and may be represented as follows: (2) a

ß

In the phonological model sketched so far, a and ß will either be Place or Manner features. There is one other dimension (or 'level') on which segments can be complex. I assume that the timing characteristics of segments are expressed by timing units, or x-slots, which appear on a separate tier, as is widely assumed (see, for instance, Sagey 1986a: 19, and the references cited there). This tier mediates between segmental and suprasegmental structure. To express the melodic unity of segments, I assume that timing units dominate root nodes, which in turn dominate all other segmental material within the segment. The structure of the segment is therefore roughly as follows: (3)

χ I C/V I Manner/Place elements

timing tier root node tier segmental tiers

The usual relation between timing units and root nodes is one-to-one (resulting in a 'simplex' segment). When a root node is linked to two timing units the whole segment is interpreted as a geminate consonant or a long vowel (see Clements and Keyser 1983). I assume that root nodes are actually specifications of major class features (following McCarthy 1988, for instance). There are two possible specifications, one for a consonantal root node (abbreviated as 'C'), and one for a vocalic root node (abbreviated as 'V'). The relation between timing slots and root nodes makes it possible for phonological complexity also to occur on the level between timing tier and root nodes. In fact three different types of segments with two root nodes are predicted to exist: one with two C-nodes, one with two V-nodes, and one with a C- as well as a V-node. These are given in (4) (in (4)c, the C and V roots occur on

2

Branching has of course always been used in metrical phonology as a determining factor for labelling certain syllables as 'heavy'. This heaviness is therefore the metrical parallel of complexity in segmental phonology. Note that the two do not have the same status: while binary branching in metrical phonology is the norm (in the light of the fact that feet consisting of one or three syllables - or other units - are marked), binary branching units (i.e. complex segments) in segmental phonology are marked.

57 different phonological tiers and are only presented on the same graphical tier for convenience): (4)

a.

χ

b.

χ

/\ C

c.

χ

/\ C

V

/\ V

C

V

We will return to complex segments of these types later, and refer to the type of complexity in these segments as 'two-rootedness'. Note that in the last of these three representations the consonantal and the vocalic root do not appear on the same tier, and are therefore not phonologically ordered with respect to each other. Hence, there cannot be a contrast between a representation with C on the left and V on the right, and one with the two reversed. We return to the implications of this 'unorderedness' in chapter 10. Let us briefly contrast the model sketched so far with that of Sagey (1986a). This is given below:

In Sagey's model, Manner primitives such as [continuant], [nasal] and [lateral] (see Sagey 1986a: 281) are attached to a number of different nodes, but there is no essential difference between their definition in Sagey (1986a) and those assumed here. Note that the nodes labelled Labial, Coronal and Dorsal function like singlevalued features, in that they are either absent or present, just like the Place features adopted in the present study. One of the differences between Sagey's model and mine, then, is the presence of intermediate nodes like 'Supralaryngeal' and 'Place' in her model and the absence of such nodes in the present one. It is clear that a phonological model that has such nodes predicts the existence of segments which branch at these levels, that is segments with two Place nodes and segments with two

58 Supralaryngeal nodes. These kinds of segments are predicted to exist in addition to segments which have two different Place features. This is admitted by Sagey (1986a: 50) and she notes that "it would be preferable if our feature representation did not predict a distinction between them". As a means of restricting the possible segment types which combine two nodes, she proposes that branching may only take place at the feature level. However, this restriction is largely ad hoc, and others have suggested that branching may take place only at the root node level (see Piggott 1988: 134). There is a separate point which should be made with respect to the status of nodes such as Place (see also van der Hulst 1991 for discussion of this point). Compare two class nodes typically postulated in feature geometry, namely the root node and the Place node: (6)

Root Laryngeal

Place Supralaryngeal

LabiaT

Coronal

Dorsal

The function of the root node is to combine two independent properties of phonological segments, i.e. laryngeal and supralaryngeal properties. The unity of the combination of these properties is expressed by the fact that there is a dominating node (segments which combine a supralaryngeal and a laryngeal node are therefore not complex). This is fundamentally different for the Place node, however: segments do not contain specifications for all of the nodes that the Place node dominates (segments which contain two specifications under the Place node are therefore complex). Rather, the presence of nodes like Labial or Coronal specify the 'value' of the Place node (cf. also McCarthy 1988). Hence, the status of the root node and that of the Place node is completely different. However, within feature geometry these nodes have been treated in the same way, which has led to the overgeneration problems sketched out above. I will take the position that no such nodes as Supralaryngeal or Place are required for the representation of complex segments. The position that Supralaryngeal is redundant has been argued for by McCarthy (1988) and Iverson (1989) (see also Padgett 1991). Let us therefore investigate the type of empirical evidence that has been advanced in favour of nodes such as Place. Processes such as Place assimilation have been standardly adduced as a source of evidence for the Place node (see e.g. Sagey 1986a: 37). These processes are then analysed as operations on a superordinate Place node, rather than on individual features. Expressing assimilation as spreading of a single node explains why this is more common than assimilation of a number of unrelated Place features. An assimilation process in which all features spread is more expected than a process in which, for instance, only Labial and Coronal assimilation takes place, and Dorsal assimilation does not.

59 However, an alternative notation for such operations which does justice to the naturalness of across-the-board Place spreading and which does not refer to the Place node is not hard to find. An alternative way that I propose is to postulate a universal set of Place primitives (which can be recognised as such, for instance, because dependency relations can hold between its members). Equivalently, we might say that a Place element can be recognised as such by way of an index that it bears. To illustrate this alternative approach, consider a process such as that in Kpelle (Sagey 1986a: 37, Weimers 1973), in which a nasal consonant assimilates in place of articulation features to a following obstruent. The language in question has / p t k kp/, whose places of articulation can be analysed as /U I A/ and a combination of U and A, respectively. The result of spreading from the labial-velar /kp·/ to the nasal is a doubly articulated nasal [rrfrf]. The assimilation rule can be expressed as follows (where Ρ is the set of Place primitives): (7)

Spread α (where a e {Ρ}) from a stop leftward to a nasal, in the appropriate domain.

Equivalently, if Place elements can be recognised as such by way of an index, say P: (8)

Spread a ? from a stop leftward to a nasal, in the appropriate domain.

Both formalisations retain the spreading format, and express Place assimilation as a single rule. Spreading of a single Place feature must be expressed differently. A process spreading just the element U in the same language can be characterised as in (9): (9)

Spread U from a stop leftward to a nasal, in the appropriate domain.

The rule in (7) is more general than the rule in (9), which expresses the relative naturalness of the two rules. Similarly, a debuccalisation rule in which final /p t k/ are realised as a glottal stop can be formalised as follows: (10)

Delete a (where α ε {Ρ}) from a stop, in the appropriate domain.

We now have the tools to look at different kinds of complex segments from a theoretical perspective. With regard to Place primitives, we have adopted {I A U} and assumed that this is an unordered set. With regard to Manner, I will propose that Manner elements do not form an unordered set, but rather are organised in a tree structure (see Ch. 5). Finally, the relation between Manner and Place primitives assumed here needs to be made explicit. Wetzels (1989, 1991, in preparation) proposes that the feature [continuant] dominates the Place features directly. We will

60 adopt his proposal here and provide further discussion of and some motivation for this assumption in chapter 7.

22 Complex segments from a theoretical perspective In this section I present a rough investigation of what kinds of segments the model outlined so far allows. In §§ 2.2.1 and 2.2.2 I discuss complexity on the Place and Manner dimension, respectively, and in § 2.2.3 I discuss the third type of complexity, namely that which characterises segments with two root nodes and a single timing position. All these segment types will be discussed in more detail in later chapters. 2 2.1 Colour mixing In chapter 1 we have already seen a few examples of vowels with more than one Place element. Vowels such as / e / and / o / consist of two Place elements. In this section we examine this type of complexity further. It occurs when more than one Place primitive occurs under a consonantal or vocalic root node. Following nomenclature in Natural Phonology (Hooper 1972; Stampe 1972, 1973; Vennemann 1974a,b), I will refer to this as colour mixing. For vowels, colour mixing results in vowels other than /i a u/. Colour mixing in consonants results in consonants other than those at the primary places of articulation coronal, labial and dorsal. Colour mixing refers to the presence of more than one Place primitive in one segment. The segment can be a consonant or vowel, but colour mixing is far more common in vowels than in consonants. As far as vowels are concerned, I refer to DP sources for arguments that vowels like / e o o / are composed of two or three of the elements {I A U} (see also Gorecka 1989 for the same insight).3 In this sense, vowels like / e / or /of are complex segments (see also Ch. 1): (11)

V / \ A I /e/

V / \ A U /o/

If two elements combine in a vowel, a head-dependent relation will hold between them, either by default (as in a five-vowel system where it is redundant to postulate such a relation underlyingly), or to create phonemic contrasts, as in languages which 3

For example, Gorecka (1989: 153) states /e/ to be: "a vowel which I propose to represent as a complex segment with a palatal and [a] pharyngeal [constriction]", and about /e/: "the pharyngeal constriction appears to be more pronounced in /ε/, the lax counterpart of /εΛ This reconstructs the DP proposal perfectly.

61 have a four-height vowel system (recall the discussion of Brazilian Portuguese in chapter 1). Note that a vowel like / o / is far more common cross-linguistically than a consonant like / k p / . The question might be asked whether the following contrast should be allowed, and, if so, what interpretation it must be assigned: (12)

V I I

V I

I

Van der Hulst (1988a, 1988b) proposes that a dependent I element signifies an ATRdistinction, that is, the representation with single I in (12) may characterise / i / , and the representation with two elements I may characterise / i / , in languages which have both vowels.4 This entails that / i / is analysed as being less complex than / i / in such a language. With regard to Place in consonants, we have assumed that the presence of I, A and U results in coronal, dorsal and labial consonants, respectively. Palatal consonants can be analysed as complexes of coronal and dorsal articulations, as Keating (1988b) and Jacobs (1989) have shown on phonetic and phonological grounds. That is, these segments consist of the elements A and I. If a language has both palatal and palatoalveolar sounds, the head-dependent relation may be invoked to distinguish between the two: in palatals the element A would be head, in palato-alveolars it would be I (see also § 2 in Ch. 6 and § 3.1 in Ch. 7). I propose that the presence of the elements A and U results in labial-velar articulations like the labial-velar stop / 1 φ / , which occurs in African languages, and the labial-velar glide / w / . Note that it may not always be easy to show that / w / must have the two elements underlyingly, however: languages do not usually have contrasts between labial, velar and labial-velar glides. On the other hand, there are languages that have labial stops, velar stops and labial-velar stops (such as Kpelle and Dan — see Maddieson 1984: 286f.). For the third combination, of the elements U and I, a number of interpretations are possible. One is that it represents labiodental sounds in languages which have both bilabial and labiodental fricatives, such as Irish or Twi (Maddieson 1984). However, it is also possible that this contrast should be expressed differently, for example as the presence of a single U element vs. two U elements, or as a Manner difference, perhaps involving stridency. Another interpretation of the combination of U and I is that it represents a linguo-labial place of articulation (Ladefoged and Maddieson 1986). See chapter 6 for further discussion of this issue.

4

In some languages other analyses might be possible: it could easily be imagined that either vowel has the default status in the language and should therefore be analysed as underlyingly empty. The other vowel could then be analysed as consisting of the element L

62 The presence of two elements A might be invoked to represent the contrast between uvular and pharyngeal Place in languages which have a contrast between velars and uvulars or pharyngeals. In fact, I will show that uvulars can be represented as velars with secondary velarisation (see § 3 in Ch. 6). First, this fills an otherwise unexplained gap: although contrastively labialised labials can be found (although they are few), and while palatalised coronals are not at all rare, velarised velars have not been reported (although pharyngealised velars and other 'back' consonants occur in Arabic). Uvulars, if they are velarised velars, fill this gap. In many languages velars and uvulars alternate according to backness or lowness of surrounding vowels, and this would also be an instance of a consonant acquiring secondary articulation, just like a coronal consonant acquiring labialisation next to a rounded vowel. The presence of two I elements, finally, could well represent dentals in languages which have dentals as well as alveolars. This is intended to obviate the need for features like [strident] and [distributed]. Colour mixing in consonants deserves much more discussion than these preliminary notes. This is offered in chapter 6.

2.2.2 Multiple stricture A segment is also complex if it has more than one Manner feature in underlying representation. This is what I will call multiple stricture. I assume that vowels do not have Manner features, except that they may have the feature [nasal]. As Manner features are exclusively consonantal features, multiple stricture is only possible with consonants. Let us consider affricates and prenasalised stops, both of which will be seen to fall under this heading. In Sagey (1986a), both these segment types are analysed as contour segments, i.e. segments in which there is sequencing of opposite feature values. The feature involved is [ ± continuant] for affricates, and [± nasal] for prenasalised stops: 5 (13)

o

o

/\

/\

[-cont]

[ + cont]

affricates

[ +nasal]

[-nasal]

prenasalised segments

Theoretical considerations and empirical evidence suggest that the representations in (13) cannot be maintained. The presence of opposite feature specifications within a single segment in this proposal seems limited to the features [ ± continuant] and

In Sagey (1986a), the feature [continuant] is located under the root node, and [nasal] is under a 'soft palate' node. This difference is irrelevant here.

63 [± nasal]. Apart from being both Manner features, these features seem to have little in common, and the split of the feature set into two groups, the first of which contains [± continuant] and [± nasal], and the second all other features, is not motivated by Sagey. The two values of a feature like [± voice], [± lateral] or [± strident] do not occur within a single segment. There are other arguments against the representations in (13). One is provided by underspecification theory (Archangeli 1984, 1988; Kiparsky 1982a,b, 1985, and others). This would presumably reject the presence of both feature values within one phonological system, and therefore also within a single segment. The representations in (13) are therefore odd if not impossible within an underspecification approach. To this can be added that the sequencing of the two values of [ ± continuant] is completely predictable (the [-continuant] part of the affricate invariably appears before the [ +continuant] part in phonetic realisation), so the information that these features are ordered should arguably not be present in the underlying representation (as observed by Kaye 1985). The representation of affricates in (13) has also been criticised on empirical grounds, most specifically by Hualde (1987, 1988a, 1988b, 1991) and Lombardi (1990), and their work provides the key to the solution for the problems that contour segments pose for both binary and unary feature theories. Hualde and Lombardi show that the [-continuant] and [ +continuant] parts of an affricate cannot be ordered in underlying representation. In the process of Basque pre-stop stop deletion, for instance, stops are deleted before other stops (where nasals and / l / also count as stops, i.e. are [-continuant]). The stop part of an affricate is also deleted before a stop. This process can be analysed as an OCP effect on [-continuant], which entails that the [-continuant] parts of trigger and target should be adjacent. From this in turn it follows that the [ + continuant] and [ - continuant] parts of the affricate cannot be on the same tier (as they were in (13)), and hence the features [-continuant] and [ +continuant] are best re-analysed as single-valued features [stop] and [cont], respectively. The revised representation of affricates is given in (14): (14)

C

/\

[stop] [cont] affricates The representation in (14) conforms to the general definition of complexity, and I will adopt it, or rather derive it from the general model of Manner offered in the next chapter. 6 It predicts that [stop] and [cont] are not ordered in underlying representa-

6

The monovalent features [stop] and [cont] represent the two values of the formerly binary feature [continuant]. This is done on the basis of evidence that both [stop] and [cont] define recurrent classes of sounds, namely stops and fricatives. The case is totally unlike that of a feature [nasal] or [round], where there is large skewing in favour of the perceptually salient pole, i.e. [+nasal] and

64 tion, and that phonological rules therefore cannot refer to the 'edges' of an affricate. This accounts for the Basque 'anti-edge effects' referred to above. At some point, however, the features [stop] and [cont] will have to be ordered to allow for phonetic realisation, and after this the edges can be referred to. It is therefore expected that low-level phonetic rules can address specific edges of affricates, and Lombardi (1990) shows that this prediction is indeed borne out. An interesting issue is whether a head-dependent relation can hold contrastively between the two Manner features in (14). I propose that the head Manner feature dominates the Place features in a segment which is complex for Manner. That is, affricates are represented as in (15): (15)

C

/\

[stop] [cont] I Place Such a representation accounts for a number of facts in a straightforward way. In Basque affricates, the place of articulation of affricates corresponds to the place of articulation of (coronal) fricatives. Basque has apico-alveolar, predorso-alveolar and prepalatal fricatives (Hualde 1991), and affricates at the same places. In every affricate the stop part can be represented as an unmarked coronal / t / . Hence, the [cont] part of a Basque affricate is more closely related to its Place specification than its [stop] part (see also § 3 in chapter 7, which deals with the Basque affricates in detail). The theory forces us to consider the question of what it would mean for [stop] to be head in a structure like (14). A suggestion that I would like to make is that this is a way of representing / s / plus stop clusters in languages in which these behave phonologically like unitary segments. For example, these clusters are not split up in languages that do not normally allow initial clusters (e.g. in Turkish such clusters receive a prothetic vowel, according to van der Hulst and van de Weijer 1991). They also behave like units in syllabification (Fudge 1969) and alliteration (Kurylowicz 1966, cited by Ewen 1982). In the complex / s / plus stop units, then, the place specification is more closely related to the [stop] part, and the continuant part is the unmarked coronal / s / : exactly the reverse of what is the case for the Basque affricates. In both cases the head feature also determines phonetic sequencing, as it is

[+round]. I therefore assume that the existence of these two monovalent features is not contrary to the spirit of Dependency Phonology, in which primitives are monovalent. In fact, Anderson and Ewen (1987) propose to analyse Manner distinctions in terms of two elements | V | and |C|, which correspond to 'relatively high sonority' and 'relatively low sonority', respectively (see Anderson and Ewen 1987: 151ff.). Hence, they also use elements which correspond to endpoints of a single dimension.

65 final in both cases. See chapter 9 for a fuller discussion of the representation of / s / plus stop clusters. The elimination of affricates as contour segments raises the question of whether the category is phonologically well-motivated at all, i.e. whether prenasalised stops can be similarly re-analysed. Sagey (1986a) goes to great lengths to show that these are unitary segments. I differ from Sagey as to the internal structure of prenasalised stops, however. I propose that prenasalised stops differ from normal stops and normal nasals in that they are underlyingly specified for both [stop] and [nasal]. In this way there is no need to refer to the minus value of a binary feature [nasal]. Thus, the following underlying representation might be proposed: (16)

C

/\

[nasal] [stop] prenasalised stop Again, the two features are predicted not to be ordered in underlying representation. Evidence bearing on this issue is hard to come by, however. Prenasalised stops will be examined further in chapters 5 and 8, where I propose a slightly different representation from that in (16). In the chapter on Manner features, I will propose that the representations as they have been proposed for affricates and prenasalised stops are not isolated occurrences of two Manner features lumped together in a single segment. Rather, the Manner set is structured in such a way that segments like lateral affricates and prenasalised fricatives are also derived as part of the set of possible complex segments.

2.2.3 Two-rootedness Above I assumed that phonological structure includes both a level of timing representation and a level at which the melodic unity of the segment is expressed. Levels of representation above the timing tier will not concern us here. Given these two levels, an asymmetry may occur between the two levels. In Clements and Keyser (1983), a case is made for representing long consonants and vowels with two positions on the timing tier, both connected to the same root node on the melodic tier. The opposite situation, two root nodes connected to a single timing position, is also predicted to be a well-formed phonological representation. In fact, as pointed out above, three types are expected: one combining two C-type root nodes, one combining two V-type root nodes, and one combining a C-type with a V-type root node. The three are repeated in (17):

66 (17)

a.

χ

b.

/\ C

C

χ

c.

/\ V

V

χ

/\ C

V

We will not discuss the first type of two-root complex segment here (see Ch. 10 for possible interpretations). The second type, with two vowels combined under a single timing slot, can straightforwardly be interpreted as a short diphthong. The headdependency relation may then be employed to make a distinction between falling and rising diphthongs (if such contrasts exist). The structure in (17)c is interesting, since it combines two different entities, a consonantal root node with a vocalic one. If the consonantal root is head, the interpretation is that of a consonant with secondary articulation. If the vocalic root is head, the interpretation should be of a vowel with some kind of consonantal property. A number of interpretations come to mind in this connection. First, retroflexed vowels (e.g. in Tarascan or Mandarin; see Maddieson 1984) could be described as vowels that have an element I on a separate, consonantal root. In Badaga (see Clements 1990b), a number of 'retracted' vowels occur which could be described as having A in the same position. Rhotacised schwa, as in American or Scottish English bird, may have a representation like the one in (17)c. The element U on the consonantal root of a two-root vowel could serve to distinguish the well-known Swedish in-rounded from out-rounded vowels, which have in fact been described as diphthongs with a consonantal, fricative element (see Clements 1990b). These suggestions are preliminary and will be further discussed in chapter 10. To illustrate a segment with a consonantal root node as head, consider two examples of consonants with secondary articulation, / p i / and / k w / : χ / \

(18)

χ / \

C

V

C

V

I [stop]

I I

I [stop]

I U

I υ

I A /PV

A7

What restrictions exist with respect to secondary articulations? There appear to be four predominant types of secondary articulation, namely labialisation, palatalisation, velarisation, and pharyngealisation (Ruhlen 1976; Ladefoged 1982; Maddieson 1984). These four types correspond straightforwardly to the three primitives U, I and A, in that order, of Dependency Phonology, if velarisation and pharyngealisation can both be phonologically expressed as the presence of A. This is what is assumed here. It makes the prediction that the two kinds of secondary articulation cannot contrast, which is borne out for the languages surveyed in Maddieson (1984) (see also Keating

67 7

1988a). Uvularisation may also be included under the heading of an A-type secondary articulation. The variety of interpretations is probably related to the fact that the effect of velars on vowels is also varied (see § 2.3 in chapter 3). Note, finally, that the fact that the three types of secondary articulation correspond to the three basic primitives of DP provides strong confirmation for this theory: on the basis of secondary articulation, it is not necessary to allow for more primitives in the theory. There are two alternatives to this strategy. The first is to have a fourth primitive, say R (for 'Radical') to represent pharyngeal segments and pharyngealisation. This is the option taken by Clements (1990), and others, but I will not adopt it, because the introduction of a fourth segmental primitive will generate far more combinatorial possibilities than are instantiated. Nor does this option make the prediction that was referred to above. The second is to assume that the element A represents pharyngeality rather than velarity, and that velarity corresponds to the absence of primitives. Thus, velar Place is represented as defective by van der Hulst (1991), since in some cases this can be shown to behave as a default place of articulation, just like Coronal does in other cases. However, I hesitate to attach far-reaching theoretical consequences to this fact, since in the previous chapter it appeared that the phonological primacy of the vowels /i a u / is parallelled by the primacy of consonant Place / p t k/ (rather than / p t q/). I will assume that pharyngeal segments, and all other consonant types, can be described as combinations of the primitives {I A l l } . Vowels occurring by themselves of course also involve the structure of a vocalic root node dominating segmental material. Secondary articulation is thus closely linked to vocalic structure, and secondary articulation in fact usually appears to derive from assimilation of a consonant to a neighbouring vowel. Secondary articulation on consonants also often interacts with vowels in terms of processes and constraints. We will return to phenomena like these in the chapter on two-root complex segments, which includes secondary articulation (see chapter 10). A final word is appropriate in this place about the relation between syllable structure and phonetic realisation. It has become quite standard in nonlinear phonology to assume glides (or semivowels) are vowels in non-peak syllable position (see § 2.1.3 in Ch. 2 above). Thus, in our framework they are represented under a vocalic root node. It is therefore predicted that these segments cannot bear secondary articulation. This prediction is correct for the glides in the corpus of Maddieson (1984).8

7

8

N o languages are known to contrast the secondary articulation types labialisation and velarisation (see Jakobson, Fant and Halle 1952: 31, for example). This has been taken as an argument for identifying both with a single primitive, for example the feature specification [-grave]. N o t e that laryngeals such as /h/ and 111 are predicted to be capable of bearing secondary articulation. Labialised / h w / is reported for a. number of languages in Maddieson (1984), such as Igbo, Amharic and Hupa, and palatalised /hi/ is reported for Irish by Ni Chiosáin (1992). A labialised glottal stop / ? w / has been postulated for Kabardian (Kuipers 1960).

68

3. Conclusion In this chapter I have proposed a model for segmental structure which incorporates aspects of feature geometry models like that of Sagey (1986a), but which uses the elements I, A and U for its Place features. Complexity in phonology was proposed to be a function of branching. The model proposed allows for three types: colour mixing, multiple stricture and two-rootedness. Simplicity is the formal correlate of a one-to-one relation between Manner features and Place features. The notion of 'contour segments' (Sagey 1986a), anomalous within phonological theory, was abandoned in favour of a representation of these segments in which more than one Manner feature was present underlyingly. A priori, combinations of different types of complexity are not ruled out, so that we expect that affricates with secondary articulation may occur, prenasalised palatals, prenasalised stops with secondary articulation, and so on. Such combinations require a more detailed study, and this is carried out in the next chapter.

Chapter 4

Head-dependent asymmetries at the segmental level

In the previous chapter I proposed a formal definition of how 'complex' segments deviate from simple segments. The branching relation which is responsible for complexity manifests itself in three ways in segmental phonology, giving rise to 'mixed' vowels and consonants, consonants with more than one Manner feature, and consonants with secondary articulation. These types of complex segments are represented in (1) below, where / e / is an example of a complex vowel, the affricate /tí/ is an example of a segment with more than one Manner feature, and a labialised /k w / is an example of a two-root complex segment: X 1 V

X 1 c

/\ A

/\ I

M

[stop] [cont] I 1 Place /tí/

X / \ C 1 [stop] I1 A

V 1 U

A7

Before discussing these representations in more detail, and examining particular complex segment types (in the following chapters), it is necessary to investigate one more theoretical issue. This is to examine if (and, if so, to what extent) these three types of complexity can be combined within a single phonological segment. In principle, it should be possible to combine the different types and it is not a priori clear what restrictions could play a role. This will lead us to examine segments which are (even) more complex than those in (1). It will become clear that there is one condition that plays a crucial role when different types of complexity are combined. This is the Mismatch condition, which was originally proposed for metrical phonology (see below), but a version of which appears to hold in segmental phonology as well. The Mismatch condition stipulates that segments with two roots or two Manner features cannot be complex for Place under the dependent root or Manner feature, respectively. Thus, in this chapter we will further investigate the issues concerned when a head-dependency relation holds between two phonological entities gathered under a single node, such as two Manner features (as in affricates) or two root nodes (as in consonants with secondary articulation). To the extent that asymmetries are found between the head node and the dependent node, for example in terms of what kinds of Place specifications are allowed, this is support for the assumption that head-dependency relations play a role in segmental phonology. It would indeed be surprising if there were no intrasegmental

70 notion of phonological strength, considering its wide acceptance in suprasegmental phonology. Dresher and van der Hulst (1994) offer a discussion of head-dependent asymmetries at various levels of phonology, focusing mainly on suprasegmental levels. They point out that, while head-dependent asymmetries can be found at all levels, the actual manifestation of the asymmetries may differ from level to level. This chapter, then, forms an investigation into the shape of head-dependent asymmetries at the segmental level. This chapter is organised as follows: first, in § 1, I introduce the Mismatch condition in metrical phonology and discuss possible ways of applying it to segmental phonology. I present all logically possible combinations of two of the types of complexity that were mentioned, and point out that certain ones appear to be impossible. The non-attested cases can all be put down to the Mismatch condition. In § 2, I examine possible combinations of all three types of complexity.

1. The mismatch condition in metrical and segmental phonology In metrical phonology, it is often observed that particular syllable types are ruled out in certain positions within the stress foot. In particular, in certain languages the strong part of the foot (which corresponds to the stressed syllable) must be occupied by a 'strong' syllable, i.e. one which is heavy as a result of having a long vowel or a coda. This generalisation has been expressed differently for different languages. It might be possible to say that there are different degrees of restrictiveness. The most restrictive stipulation is that in the strong branch of the foot only a heavy syllable, and in the weak branch only a light syllable may appear. This has been referred to as the Obligatorily Branching' (OB) foot (for discussion see Hammond 1986). This restriction on footing may be relaxed by allowing two syllables that are both light to be labelled (s w). This is a weaker restriction than the OB restriction, since it only disallows heavy syllables in the weak part of the foot, and does not demand a branching syllable in the head position. This is the situation in a quantity-sensitive (QS) system (see Hayes 1980 for examples). The weakest constraint is the third, namely that the dependent part of the foot may not be stronger than the head part. This also allows two heavy syllables to be labelled (s w), which the previous two conditions on feet excluded. The only foot that is ruled out as (s w) then is a light syllable followed by a heavy syllable (labelled 'the mismatch foot' (Mis) by van der Hulst 1984: 211), while the three other logical possibilities are well-formed as (s w). Finally, in a quantity-insensitive system (QI) there are no restrictions on heaviness; we might say that in such a system there are no asymmetry effects. The four restrictions described above allow different foot types to be labelled (s w), and can be summarised as follows ( / indicates that labelling (s w) is allowed, absence of a mark indicates that it is not allowed):

71

σ Η

(2) OB QS Mis QI

σ L / • / •

σ L

σ L • • •

σ Η

σ Η

• •

σ L

σ Η

•

In the scheme above the four foot types OB to QI are different parameter settings, which have been motivated for different languages. In this chapter I provide evidence for the claim that similar effects can be seen at work in segmental phonology. A straightforward parallel between segmental and metrical phonology is not available, however. It will be seen that in certain cases dependents do not show the full array of Place contrasts that heads allow. Furthermore, it is not the case that in a complex segment the head node must branch (in some sense), so that the OB setting would be too strong in any event. I hypothesise that the QS option in metrical phonology comes closest to the condition at work in segmental phonology. The asymmetry effects concern the relation between Manner and Place. Recall from the previous chapters that in a phonological construct composed of (at least) two entities of the same type, a head-dependency relation is postulated to hold. Thus, in a segment involving 'multiple stricture', one of the Manner features present is the head. As I will try to show in the following chapters, this is the Manner feature that dominates Place in underlying representation. In a segment that has two root nodes under one timing slot ('two-rootedness'), one of the root nodes is likewise the head. The condition I propose constrains the Place specification of the non-head in such complex constructions. This condition is given in (3) (termed 'Mismatch', although as observed above the QS parameter setting is the closest metrical resemblance to the constraint proposed for segmental phonology): (3)

The Mismatch Condition in segmental phonology Non-head nodes may not branch

The generalisation that non-head nodes may not branch can be instantiated in various ways. It is satisfied when in a non-head there is only a single Place element, and also when there are no Place elements. Both ways of observing the constraint are in fact found: in consonants with secondary articulation the secondary articulation must be of a simple kind, that is, the vocalic root node in such a consonant may dominate only one Place element. This accounts for the predominance of labialisation, palatalisation and velarisation as types of secondary articulation. In multiple-stricture complex segments such as affricates, on the other hand, which are hypothesised to have two Manner features, and therefore branch at another level, the dependent Manner feature may not dominate any Place material at all. I will return to both points in later chapters.

72

That there should be different ways of satisfying the Mismatch condition is not necessarily damaging to the idea that head-dependent asymmetries play a role in segmental phonology. In metrical phonology, too, there are different ways of making a syllable heavy (both a long vowel and a consonant-final syllable involve a branching rhyme), and there may also be different ways of restricting the segmental material in weak parts of the foot. In Russian (see § 1.2 in Ch. 1 above) the vowels permitted in the weak part of the foot are restricted to /i a u/. This forms a parallel to consonants with secondary articulation. A language which allowed only schwa to occur in such positions would parallel the situation for affricates. Let us now turn to an investigation of the types of complex segments found and the way in which the Mismatch condition rules out segment types that are not found. First, let us turn to affricates and other segments which have more than one Manner feature (the full set is explored in the next chapter). In (1) above it is assumed that Place features in affricates are attached to the feature [cont] only. This is intended to express the generalisation that affricates typically appear at places of articulation where fricatives also occur. Thus, the feature [cont] is a type of 'head' in affricates, and [stop] is the non-head.1 If we consider the relation between these two features to be identical to the (two) syllables in a foot, it is immediately clear that one of the two features should be the head (and therefore have some special status - of course it cannot be 'stressed' as the head syllable in a foot is), and that there can only be one head. My interpretation of headedness in affricates is that the feature which bears the Place features is the head. With respect to the non-head feature in affricates, [stop], the Mismatch condition comes into play. There are a number of possible ways in which the condition could be satisfied. In affricates, I assume, the Mismatch condition is satisfied because the non-head feature, that is, the weak node of the affricate, does not dominate any Place features at all, so that the condition that weak nodes do not branch is satisfied. Consonants with secondary articulation function as consonants. Hence, I assume that the consonantal root node in such two-root complex segments is the head, and that the vocalic root is the non-head. The Mismatch condition in (3) above stipulates that this non-head must also be simplex. This condition is satisfied because the nonhead root node is only allowed to bear one of the Place features that we have assumed, I, A or U. This results in a formal account of the observation that there are

1

This is a different claim from that made by Steriade (1989), who observes that affricates typically occur at places of articulation different from those at which fricatives occur, and therefore claims that they are underlying stops. While this leaves certain problems unsolved (such as the fact that affricates may occur at places of articulation where both a stop and a fricative also occur), and therefore needs further study, it accounts for the fact that affricates typically (although not exclusively) pattern with stops in phonological rules. A different proposal to account for the latter asymmetry is made by Shafer (1995), who proposes that in affricates the feature [stop] geometrically dominates [cont]. This leads to problems when the [stop] feature is delinked, however, as in Basque or Luiseño (see Chen 1993, as well as Ch. 7 below), since the deletion of [stop] entails the deletion of [cont].

73 three main types of secondary articulation, namely palatalisation, velarisation and labialisation, as was briefly discussed in the previous chapter. In the rest of this chapter I will investigate consonant types which combine different sorts of branching. Consonants can be complex for Place and for Manner, and they can have two root nodes, as has been assumed so far. Let us see what combinations are predicted to exist, and whether this prediction fits the data. In § 1.1 I discuss consonants which have two roots and which have colour mixing. In § 1.2 segments with two roots as well as multiple stricture are examined, and in § 1.3 I discuss combinations of colour mixing with multiple stricture. On the basis of what is found in these sections, in § 2 I examine whether segments are also predicted that would have all three kinds of complexity. In fact, such segments are predicted to occur, and a segment which has this structure in fact occurs in at least one of the world's languages.

1.1 Combining two-rootedness and colour mixing Running ahead of a fuller discussion of segments with two roots, recall that in the previous chapter it was pointed out that in a theory which recognises a timing tier as well as root nodes it is predicted that three types of combinations of two root nodes under a single timing slot are predicted to exist: one with two consonantal root nodes, one with two vocalic root nodes, and one with one of each. Here we will investigate the latter two kinds of complex segments, because their interpretation is most straightforward: a segment with a C as well as a V node represents a consonant with secondary articulation (discussed in § 1.1.1 below) 2 and a segment with two V nodes under a single timing slot represents a short diphthong (discussed in § 1.1.2). In the following two subsections we investigate how colour mixing, that is, the presence of more than one Place primitive in a single segment, combines with the two-rootedness of such segments.

1.1.1 Consonants with secondary articulation Above we saw that colour mixing is not allowed under the vocalic root node of a consonant with secondary articulation, because this violates the Mismatch condition. Such complex secondary articulation types have been reported, however, and in this section I offer a short discussion. An example comes from Irish, where Maddieson

2

The same segment with the vocalic root node as head instead of the consonantal one is interpreted as a vowel with a consonant-like constriction, such as the retroflexed and retracted vowels of Badaga (see sc. 3.2.1 in Ch. 10) or the short rounded diphthongs of Swedish, which have been described as ending in a consonantal glide [/?]. These are not discussed here, since the paucity of information on such vowels does not permit a full investigation. The information that is available about such vowels indicates that they also conform to the Mismatch condition as formalised here.

74 (1984) reports a simultaneously labialised and velarised bilabial fricative /b w /. However, there is no three-way contrast between labialised, velarised, and labialvelarised bilabials in this language. Hence, the labial-velarised segment may be analysed as either only labialised or only velarised underlyingly, with the other secondary articulation added as a phonetic enhancement factor (see Stevens, Keyser and Kawasaki 1986). The same is true for ordinary [w]. Another example of reported labial-velarisation comes from Ethiopian languages. Keating (1988a: 10, fn. 4) cites information from Maddieson (p.c.) who claims that Ethiopian languages may contrast labial-velarisation with pharyngealisation. She makes the same suggestion as the one made above for Irish, namely that velarisation may be phonologically redundant for the labial-velarised segments. A second type of secondary articulation which at first sight might appear to be complex is uvularisation, because it might be analysed as being different from velarisation. However, no cases are known to me where uvularisation and velarisation contrast within a single language, so that uvularisation may, just like velarisation, be analysed phonologically as the presence under the V node of the element A. This follows Keating (1988a: 10), who claims that velarisation, uvularisation and pharyngealisation never contrast in languages. Within a language like Arabic the three types may vary from dialect to dialect: "Pharyngealization and uvularization are at best rare (fn: What is described as pharyngealization in Arabic varies phonetically across dialects from velarization to uvularization to true pharyngealization, with Tunisian clearly having uvularization (...), and Syrian, Jordanian, and related dialects supposedly more likely to have true pharyngealization.)". As no further examples of potentially complex types of secondary articulation are known to me, the conclusion may be drawn that the Mismatch condition makes the correct (phonological) predictions, in that it is never necessary to recognise complex secondary articulation types underlyingly (see § 3 in Ch. 6 below on uvulars with secondary articulation). Colour mixing is allowed under the C node of such a consonant, however. This results in the following kind of structure: (4)

χ V I Ύ

Examples of colour-mixed stops are labial-velar stops /1φ/, but also palatals (see Ch. 6). The representation in (4), with the extra V root, therefore characterises segments like these with secondary articulation. Reports of such segments with distinctive secondary articulation are rare, however, and Maddieson (1984: 35) finds no examples in his survey of plosive

75 segments with complex Place: "none of the retroflex, palatal, palato-alveolar or labialvelar plosives in [the survey] is reported as having any secondary articulation. This may be due to the comparative rarity of plosives at these positions. Since stops with secondary articulation are less common than their simple counterparts, [the] survey may omit them by chance". Note that this absence refers to plosives only, and that Maddieson does report palatal affricates and fricatives with secondary articulation, so the suggestion that chance might play a role is well taken. This is also what Keating (1988a: 11) suggests. Distinctively palatalised palatal stops have been reported elsewhere, however. Campbell (1974: 58) reports that c and d contrast in Mordvin. Kuipers (1963: 62f., 65) reports a distinction between plain and palatalised palatals in Bêdux and Temirgoj, both Western Circassian languages, giving examples of the contrast for Bêdux plosives (Kuipers 1963: 65). Hence, secondary articulation on colour-mixed consonants seems to be phonologically possible.

1.1.2 Short diphthongs Consider now short diphthongs, which were proposed to consist of two root nodes under a single timing slot. It is well known that the two parts of a diphthong are not equally prominent. In 'falling' diphthongs, the first part of the diphthong is more prominent, and in 'rising' diphthongs the second part is more prominent. The more prominent part of the diphthong can be interpreted as the head, and the falling part as the non-head. The Mismatch condition predicts the non-head to be simple, that is, to consist of a single Place element, whereas the head part is predicted to be either simple or complex. That is, structures such as (5)a below are acceptable, while structures like (5)b are ruled out for a falling short diphthong (the vertical line between x-slot and root node indicates headship by convention):

A short diphthong like that in (a) has a complex head, and a simplex non-head (recall from Ch. 1 that / e o / are complex vowels, and /i u / are simplex ones). Examples of such diphthongs would be / e i / or / o u / . These diphthongs are among the commonest of the short diphthongs that are found. They are also among the most common long diphthongs, in fact (where it is well known that the second element may be phonetically lower than / i / or / u / ; however, this does not concern the phonological representation).

76 The kind of short diphthong in (b) would be / a e / or /io/, etc. Such diphthongs have not been found in the literature available to me (e.g. Hayes 1990; Sluyters 1992), and they are also rare among long diphthongs. Hence, in these cases the predictions made by the Mismatch condition are borne out.

1.2 Combining two-rootedness and multiple stricture A segment with a consonantal and a vocalic root node, in which the former displays multiple stricture, may be represented as in (6) below (again an affricate serves as an example of a multiple stricture-segment):

Affricates (and other multiple-stricture segments, such as prenasalised stops) with secondary articulation exist. For instance, affricates with distinctive palatalisation occur in Bulgarian and Kashmiri, and affricates with distinctive labialisation occur in Lak, Hupa and Gä. Labialised prenasalised stops / I g " / occur in Ngizim, Washkuk, Wantoat and Nambakaengo. A palatalised / n dy occurs in Paez (Maddieson 1984).

1.3 Combining colour mixing and multiple stricture We have hypothesised that a consonant with more than one Manner feature only specifies one of these features for Place in underlying representation. In principle, this 'head' Manner feature can also be specified for two Place features. The representations of alveolar and palato-alveolar affricates are thus different in that alveolar affricates have only Coronal (I) under [cont], while a palato-alveolar affricate is doubly complex: in addition to multiple stricture it has colour mixing since it has Coronal and Dorsal under [cont]. The latter is given below: (7)

C [stop]

[cont] / \ A

I

77 Of course, such affricates are attested, as are palatal prenasalised stops. Maddieson (1984) also reports dental affricates, as well as a uvular one (in Kabardian). All these places of articulation are outside the primary ones (labial, coronal, dorsal), so we hypothesise them to be complex in their Place representation.

2. Combining all three types of complexity A segment with two roots, let us say a consonant with secondary articulation, two Manner features, let us say an affricate, and two Place features, let us say a palatal, looks schematically as follows: (8)

[stop]

[cont]

a

/ \ A

I

The segment in (8) is a palatal affricate with a simplex kind of secondary articulation (i.e. palatalisation, labialisation or velarisation). Maddieson (1984) reports exactly one such segment: a palatal affricate with distinctive labialisation in Akan. Such a segment is the most complex segment representable in the present framework.

3. Conclusion In this chapter I have examined the ways in which the three different types of complexity introduced in the previous chapter can combine in single segments. We have seen that this combination is unconstrained, except for the fact that branching may not take place under a root node or Manner feature that can be regarded as *weak'. This was related to the Mismatch condition, originally exclusive to the domain of prosodie phonology. Weak nodes were identified with the offglide part of a consonant with secondary articulation, the part of a short diphthong which is the less prominent, and the part of multiple-stricture segments which do not bear Place features. In the next chapter I will examine multiple stricture segments in more detail.

Chapter 5

Manner primitives

In this chapter I investigate the organisation of the Manner features. After a short introduction to the Manner dimension in phonology, I will provide representations of consonants which are not complex on this dimension, such as simple obstruent stops and fricatives, and of nasals and other sonorants. These representations are cast in a geometrical model of Manner. I will also provide evidence for a number of natural classes that this model defines. After that, we turn to a discussion of the complex segments that the model allows for. These segment types are introduced here only briefly; some of them will receive a fuller discussion in later chapters. First, the Manner model that I propose is introduced, together with the representations of the simple segments that this theory allows (§ 1). Then I discuss processes and other phonological phenomena that can be captured within this proposal (§ 2). Finally, the proposal is examined with specific reference to complex segments (§ 3).

1. Manner Traditionally, consonants are characterised in terms of three-term labels, e.g. / p / is described as a voiceless bilabial stop. The first of these labels refers to laryngeal aspects (voiced, voiceless, aspirated, etc.), the second to place of articulation (bilabial, labiodental, dental, velar, uvular, etc.), and the third to the Manner (or Stricture) type the segment has (stop, fricative, nasal, lateral, rhotic, approximant, etc.).1 All three aspects of segmental structure appear to correspond to independent entities. This is insightfully captured in Clements' (1985: 248) proposal for the organisation of distinctive features, where each of these entities corresponds to a separate 'class node' (details omitted, class nodes written with a capital letter): (1)

Root

[nasal] [strid] [cont] Labial

1

Place Coronal

Dorsal

In this study it is assumed without discussion that laryngeal features are attached to the root node. The exact content of the set of laryngeal features does not concern us here.

80 The three class nodes Laryngeal, Stricture and Place correspond to the three dimensions expressed in the three-term labels. Each of these three class nodes can be shown to be affected by phonological processes, such as spreading and delinking, except for the class node 'Stricture'. A phonological process that would be evidence for a Stricture node would affect the Stricture characteristics of a class of segments in the same way, parallel to a process that affects the Place characteristics of a class of segments in the same way. For instance, if a class of segments loses their Place feature(s), regardless of which features these are, this is evidence for an abstract node Place, or, alternatively, for a set of Place elements. Such processes do exist, such as nasal assimilation to stops. We have already seen that it is not necessary to recognise a class of Place features by postulating an abstract Place node, but that this can be equally easily achieved by allowing for a set of Place features (see § 2.1 in Ch. 3). The question of whether similar processes exist for Stricture cannot be answered so easily. A possible source of evidence is the phenomenon of 'coda weakening', which refers to the well-known observation that codas typically allow a smaller set of phonemic distinctions than onsets. This can be expressed as the suspension of Manner contrasts in this position. Alternatively, however, it can be expressed as the suspension of whole-phoneme contrasts, or as the suspension of Place contrasts alone (see Goldsmith 1989). It has also been argued that in specific cases spreading of all Manner features does take place. A candidate process comes from Yakut (Wetzels 1989). In this language the plural possessive suffix alternates between a bilabial obstruent plosive (voiced or voiceless) and a nasal. Examples are given below: (2)

tiinniik ohoq kötör aam tiiq

'window' 'stove' 'bird' 'door' 'squirrel'

tünniik-püt ohoq-put kötör-büt aam-mit tiiq-mit

'our 'our 'our 'our 'our

window' stove' bird' door' squirrel'

Such a suffix can be represented underlyingly as an obstruent stop only, with [nasal] spreading after nasals, much as in Dakota (see below in § 2.1). In the suffix alternations provided by Wetzels (1989: 174f.), the behaviour of suffixes after / l / varies: in some forms the first suffix consonant is changed to [1] (e.g. kiil-le 'ashes-PART', and in others it remains an obstruent (e.g. uol-ga 'son-DAT'). Finally, suffixes that begin with / l / also change this consonant to an obstruent stop after obstruents (e.g. at-tar 'horsePL'). In such cases the Manner representation is completely delinked (including Place), and a default Manner ( = [stop]) and Place (=Coronal) specification is inserted. Such cases offer support for the idea that Manner is an independently addressable phonological category. However, the Yakut case does not in my view show that there is a need for a separate 'Stricture' node, which dominates all and only stricture features.

81 Since the most important evidence that features or nodes should be incorporated in the feature tree is usually assumed to be the fact that such feature or nodes can be shown to spread (see Sagey 1986a; McCarthy 1988), this observation has led to a rejection of the class node 'Stricture' in almost all current feature geometry proposals. This has led to a wide proliferation of proposals with regard to the location of features that were under the 'Stricture' node in Clements (1985). Consider, for instance, Sagey's (1986a) proposal, in which the features that were under Clements' Stricture node are italicised: Root

(3)

[continuant] Laryngeal ...

Supralaryngeal Soft Palate [nasal]

Place Labial

Coronal

Dorsal

[anterior] [distributed] The Stricture node has been abandoned, and the features [nasal] and [continuant] are attached to nodes at different places in the geometry. Sagey does not commit herself to the place of the feature [strident], (see Sagey 1986a: 281). Other proposals differ from that of Sagey, in that they locate the Stricture features again under other nodes (see den Dikken and van der Hulst 1989 for an overview), or even assume that they may be attached at different places on a language-to-language basis (as in Yip 1990). This proliferation of ideas obscures the fact that Stricture features (or Manner features, as we have referred to them) form a single dimension within the segment. In a structure like (3), the features [continuant] and [nasal] would seem to have as little in common as the features [spread glottis] and [round], or any other feature combination. However, the fact that Manner features characterise a single dimension is clear from the fact that Manner, and not Place, or a combination of Manner and Place, to a great extent determines the linear order of segments within the syllable. That is, nasals usually occur closer to the syllabic than stops, rhotics occur closer than fricatives, fricatives closer than stops, etc. (Kurylowicz 1948; see, among others, Clements 1990a for general discussion of specific proposals). That is, the Manner features determine to a great extent the sonority value of a segment, and the sonority value of a segment determines its place in the syllable. Place of articulation is largely irrelevant in this respect, although this may obviously play a role in language-specific syllable structure constraints (such as the fact that in English labial consonants are not permitted after the diphthong / a v / ) . Feature geometry proposals such as Sagey's

82 (unlike Clements') do not express the fact that the Manner features behave as a group in this respect (see also van der Hulst 1991 for discussion of this point). The situation is quite different in Dependency Phonology, in which the class nodes Stricture and Place of Clements (1985) correspond to distinct gestures within the segmental structure. That is, there is a categorial gesture which (roughly) expresses the Manner characteristics (laryngeal characteristics are also conceived of as part of the categorial gesture, see Anderson and Ewen 1987: 149), and an articulatory gesture for the representation of Place. In this study, it has been proposed that there is no separate Place class node, or gesture, but rather that Universal Grammar provides a set of Place primitives. These primitives can enter into relationships with each other, in which the dependency relation may play a role. For Manner, however, I propose a geometrical model, which I will now develop. 2 To begin with, I assume a division between segments produced with a complete closure, which are represented with a [stop] feature, and segments produced without such a complete closure, which are represented as [cont].3 When the oral airstream is blocked, it can obviously no longer be modified. The only possibility is to add a second closure (or, equivalently, to lower the soft palate, so as to allow the air to escape nasally), necessarily at the velum, which will result in a nasal segment. On the other hand, if the airstream is not completely blocked, further modification is possible. I suggest that such modifications result in lateral segments (which have a blockage of the airstream) and rhotics (which do not). The model which expresses this is given in (4) below. Notice that in such a model the name of the feature [nasal] is completely redundant, and might be replaced by [stop] (see van de Weijer 1992a). (4)

C [stop]

[cont]

obstruent level

I [nasal]

[stop] [cont]

sonorant level

In such a model, traditional Manner features such as [nasal] and [lateral] correspond to dependents; in other words, these features cannot occur on their own in segments, but have to be dominated by an 'obstruent-level' feature. Obstruents are represented as segments which only have features on the highest level, while sonorants require further expansion. This will be further illustrated below. Finally, note that the model does not have a second contour of [stop] and [cont] under [stop]: the reason for this absence is grounded in phonetics: since the airstream is completely blocked by the higher-level [stop] feature, it can no longer be modified.

2 3

Part of what follows appeared in a slightly different form as van de Weijer (1992a). The features [stop] and [cont] are my abbreviations for unary features corresponding to the binary feature values [-continuant] and [+continuant], respectively.

83

2. Simple segments and natural classes In this section I illustrate the (simple) segment types that the Manner proposal made in the previous section can represent. First, we will look at the representation of individual segments (§ 2.1), and then we will look at processes that affect a number of segment types at the same time, that is, in which classes of segments that share the same Manner type are involved (§ 2.2).

2.1 Simple segments In this section I show how segments which do not involve internal branching of Manner features are represented. The discussion is limited to consonants. I assume that vowels (and glides) differ in a fundamental way from consonants: they lack Manner features altogether (except for [nasal]), and are distinct from consonants because of their Major Class specification (see chapter 3). These segments therefore lack any constriction. It is predicted that vowels (or glides) cannot form a natural class together with fricatives.4 Obstruent stops and obstruent continuants are represented as single expansions of the 'Manner tree': (5)

C

C

I [stop]

I [cont]

stops

continuants

The Manner contrast between stops and continuants might also be expressed in other ways. One other possibility is given in (6), where [stop] is underspecified (a proposal like this is equivalent in spirit to that of Rice 1992): (6)

C

C I [cont]

stops

continuants

I assume that if the underlying contrast between stops and continuants is to be represented in terms of absence ra. presence of a Manner feature, this should be

4

However, nasal vowels do involve a constriction, namely at the velum. On the internal structure of nasal vowels, see section 3.2 in Ch. 10.

84

done as in (6), and not the other way around (with [cont] underspecified). Stops are clearly less marked consonants than continuants, which is evident from facts about segmental inventories (the presence of fricatives implies the presence of stops) and child language (stops are learned first, and fricatives are typically replaced by stops in certain positions of the syllable; see Fikkert 1994: 56ff.). This might argue for treating [stop] as literally 'unmarked', that is, unspecified for Manner. However, as stops clearly also form a natural class, I assume that it is necessary that have some feature in their representation by way of which this class can be referred to.5 I therefore propose that the contrast is as in (5), and assume that [stop] is 'unmarked' relative to [cont]. However, the difference in status between the two features does not follow from the model as it stands, which in this respect therefore leaves room for improvement. I do not take a position with respect to the question of whether underspecification of Manner features in general should be disallowed. Sonorant segments, such as nasals, laterals and rhotics, are marked off as a class which is structurally different from obstruents: they need a further expansion under either [stop] or [cont]. Although this permits us to retain the concept of sonority, it makes the feature [sonorant] redundant. This explains why this feature does not spread (see McCarthy 1988). In this model the sonority of a segment increases as its segmental structure becomes more elaborate, while, ceteris paribus, [stop] segments are less sonorous than [cont] ones.6 Representations of sonorants are given in (7): c ι 1 [stop] 1 [nasal]

C I 1 [cont]

nasals

laterals

1 [stop]

C I 1 [cont] 1 [cont] rhotics/ trills

C I 1 [cont] / \ [stop] [cont]

obstruent level sonorant level

flaps

In this model laterals and rhotics have a similar structure, since both have a feature under [cont] on the obstruent level, and therefore form a natural class. The difference

5

6

This assumption is not made by Rice (1992: 88), who assumes a structure like (6). The natural class of obstruent stops, in such a model, must be expressed by referring to the complete absence of all Manner features. We would like to avoid reference to the absence of features in this case since it makes it impossible to express an OCP effect like Stop Deletion in Basque, which (among other things) deletes one obstruent stop before another (see Hualde 1988b, 1991, as well as Ch. 7 below). The effect is in principle inexpressible in Rice's framework. It is also generally considered unattractive to refer to the absence of features (see, for instance, Hall 1994). See Rice (1992) for a similar idea, namely to derive sonority from the internal structure of segments, as well as for general discussion in this respect. Dogil (1988), Clements (1990a) and Harris (1990) are others who have sought to derive this notion from internal segment structure; see Rice (1992) for discussion.

85 between them is one of continuancy. Following recent research (e.g. Halle and Clements 1983; see also Clements 1987; Jacobs 1989: 146), I assume that laterals are [ - c o n t i n u a n t ] (or [stop]), while rhotics are [+continuant] (or [cont]). The further distinction in laterality of these segments is then phonologically redundant. Above I pointed out that the relation between [stop] and [cont] is such that stops are unmarked relative to continuants: the same goes for laterals versus rhotics, where the former are less marked. For instance, Maddieson (1984: 73) reports that laterals are more common cross-linguistically than rhotics. With regard to the proposal that they have a representation with a dual stricture type, this is also brought out in the Dependency Phonology proposal by Anderson and Ewen (1987: 163). The proposal in (7) is intended as a variation on the same theme. What is the representation of lateral fricatives? One possibility is that the relation between the features characterising laterality is different in approximant laterals from that in fricative laterals. This is suggested in van de Weijer (1992a). Another possibility is to recognise a separate node or feature to distinguish between obstruents and sonorants (see Brown 1994, for instance). In the model in (4) the latter approach is redundant in the light of the fact that obstruents and sonorants are already structurally different. I leave this (important) topic for future research. The last expansion in (7) specifies [cont] for both [stop] and [cont]. I suggest that this representation is adequate for flapped types of rhotic, which may contrast phonemically with trills (resulting in a maximal three-way contrast among the liquids). 7 A 'contour' representation of the flap has been suggested by Banner (1989), on the basis of phonetic considerations and the rule of flapping in English. She observes that a flap in a spectogram looks like a somewhat shorter, sometimes "fricated stop" (Banner 1989: 65), but makes a rather different final proposal from that in (7), in which a flap has three branches, two of which dominate specifications for [ - c o n s ] . More information about the phonological behaviour of flaps is needed to test the representation of flaps proposed here. Relevant in this respect may be an observation for Catalan, made by Hualde (1992), that affricates and flaps are the only two segment types that are disallowed in word-initial position: in the current representational system, these are both complex segments (see Ch. 7 below for a more detailed discussion of affricates: like flaps, they have both a [stop] and a [cont] feature). At any rate, the languages surveyed in Maddieson (1984) bear out the prediction that a three-way Manner contrast between liquids is the maximum. However, the model does not distinguish between flaps and taps, which Maddieson assigns to one and the same category, but between which Ladefoged (1971) makes a distinction on phonetic grounds (see also Bhat 1974b: 85, who notes that "the published phonologies generally do not distinguish between the above (...) varieties of sound").

7

See Bhat (1974b: 82ff.) for discussion of the "somewhat ambiguous" use of the term flap in phonological studies.

86 Some discussion is in order with respect to the relation between the [stop] and [cont] features under the root node, and the same two features under [cont]. Although these are the same features, so that segments that have, for instance, a [stop] feature form a genuine natural class, under this model it is predicted that the continuancy specification of an obstruent cannot affect the continuancy specification of a liquid directly by spreading, because there is no plane in which to spread the relevant features. That this may well be correct is suggested by the fact that in languages in which [1] and [r] are allophones of the same phoneme, the selection of the allophone never seems to be determined by the continuancy specification of neighbouring obstruents (see Maddieson 1984: 83, Bhat 1978: 71f. and van de Weijer 1992c for case studies and further discussion). A final point that should be made here is that this model predicts that nasals, which have a [stop] feature on the obstruent level, should be able to cause hardening of fricatives into stops, which takes place in African languages like Kikuyu (Armstrong 1967).8 It is predicted that / I / cannot have the same effect, which to the best of my knowledge is correct. The representation of nasals as segments with a two-level structure deserves some more discussion. It leads us to expect processes in which a nasal changes into a stop by deleting the lower [nasal] feature ((8)a below), as well as processes in which a stop becomes a full nasal by spreading the same feature from a neighbouring nasal (whether a consonant or a vowel) ((8)b below).9 Note that it is predicted that fricatives can never be subject to such rules. (8) a.

[stop] +

b.

[stop]

[nasal] nasal - * stop

[stop] [nasal]

stop

nasal

Both kinds of processes are attested. Trubetzkoy (1939 [1969]: 165) draws attention to "the Taz dialect of Ostyak-Samoyed (Selkup): here the opposition between occlusive and nasal is phonologically irrelevant in final position. This means that one and the same word in final position sometimes has a voiceless oral occlusive and sometimes the corresponding nasal. Accordingly m and ρ (or η and t, or η and k respectively) in

8

9

The same change takes place synchronically in Eromangan, the language spoken on Eromanga, an island in the New Hebrides. Capell (1972: 54) states: "Synchronically, the variation is phonetically simple: a change of ν (=[β]) into ρ before a labial nasal consonant". Capell makes it clear that the process takes place after a nasal as well as before, e.g. vai 'take' > ya yampai Ί shall take'. In all cases provided, the nasal and fricative share the same place of articulation, although it is not clear from the description whether identity of place of articulation is required. Identity of place of articulation is required for the occasional pronunciation of words like seven [sebm], and business [bidnis] in English, where a similar rule operates sporadically. The representation of nasal vowels and vowel nasalisation is discussed in more detail in section 3 of chapter 10.

87 that position are optional variants of one archiphoneme, while in all other positions m, n, and η on the one hand, and p, t, and k on the other, are distinguished as independent phonemes." Some examples of the neutralisation patterns are presented in (9), with / m η q / respectively, from McNaughton (1976: 201ff.): a.

b.

c.

nom su: ram tam qaim nätän man soNman l'akcan oq toq nätäq otänaq

— nop ~su:rap — tap -qaip — nätät — mat — soNmat - l'akcat -ok — tok — nätäk — otänak

'heaven' 'beast' demonstrative pronoun Svhither-ACC' 'of a girl; girls' Τ 'prolative sg. of soq' 'heel' 'mouth' 'rear' 'girl-NOM-SG' 'reindeer-DAT-SG'

This type of (variable) neutralisation can be expressed as delinking of the [nasal] feature (see Ch. 1 for discussion of neutralisation of vowel contrasts, formalised in the same way), and suggests that the difference in representation between stops and nasals is privative: nasals have a feature that stops lack. The process in (8)b above occurs in Dakota (Shaw 1980: 97), where vowel nasality changes certain consonants into nasals at the same place of articulation. In the Teton dialect of Dakota b is changed to m and / to n before nasal segments. For example, when an ablauting verb such as ya 'to go' changes its final a to ì (...), the preceding consonants assimilate with respect to nasalization, e.g. (accent suppressed, JW): ya la bla gla higla

'he goes' "you go' 'Igo' 'he goes home' 'he suddenly becomes'

yî-kta ni-kta mnl-kta gnl-kta hïgnï-kta

'he will go' *you will go' Ί will go' 'he will go home' 'he will suddenly become'

Similarly, if the demonstrative le 'this' undergoes vowel deletion such that the I is juxtaposed to a nasal vowel, the I surfaces as n: (11)

le-isko

[nïsko]

'as large as this'

The change of b to m is a typical example of the rule type in (8)b. The change of I to n must be expressed in a different way, which is not straightforward in the present framework.

88 A parallel to the Dakota rule can be found in the Nilotic languages Dinka and Shilluk (Tucker and Bryan 1966), where the choice between stop and nasal is morphologically conditioned, as the following examples show: (12)

Dinka tl*k h yòl Shilluk $5k ló-t

'woman' 'house'

tiq γόη

Svoman-DET' 'house-DET*

'mouth' 'stick'

daq

'mouth-DEM' 'stick-DEM'

All five articulation places have stop and nasal variants. Finally, a similar kind of rule appears to have operated historically in the Zapotee languages. Proto-Zapotec */p/ changed into / m / "through the influence of another nasal in the same word" (Swadesh 1947: 222), as shown by examples like *[pênné?] > Cuixtla Zapotee [mèn] 'person', *[pecì?na?] > [mzìn] 'deer' (Swadesh 1947: 223). The same kind of process as predicted for nasals is also predicted for laterals and rhotics. Both are predicted to be able to turn into fricatives by loss of the sonorant Manner feature. This prediction is much less easily illustrated. One possible example comes from Ciyao, a Bantu language from Mozambique (Hyman and Ngunga 1993), in which I and s alternate productively, as shown in the examples in (13) below. An underlying /!/ shows up as [s] in causative morphology: (13)

ku-míl-á ku-díl-á ku-lól-á ku-gwáál-á ku-kámúl-a kw-iinjíl-a

'to 'to 'to 'to 'to 'to

swallow' cry' look at' escape' hold' enter'

ku-mis-y-a ku-dis-y-a ku-lós-y-a ku-gwáas-y-a ku-kámús-y-a kw-iinjís-y-a

'to 'to 'to 'to 'to 'to

make make show' make help' make

swallow' cry' off with' enter'

For the other liquid, r, the situation is less clear. The model in (4) would lead us to expect that rhotics may show alternations with fricatives. As it happens, rhotics are sometimes phonetically fricatives, rather than sonorants, as for example in Dutch and or other Germanic languages (cf. also Hall 1993).

22 Natural classes The issue of the internal structure of segments is closely linked to the question of what natural classes occur, that is, which segments pattern together in phonological rules. In this section we make explicit the predictions that the model for Manner features outlined above makes in this respect, and examine if these predictions fit the

89 empirical data. Let us first list the relevant natural classes with reference to the Manner model, repeated in (14): (14) [stop] I [nasal]

[cont]

obstruent level

[stop] [cont]

sonorant level

The following natural classes are defined either by virtue of the fact that they have an identical geometrical structure, or by virtue of the fact that they have a feature in common. (15)

Natural classes: o obstruent level: o sonorant level:

obstruents sonorants

o left-hand branch: o right-hand branch:

stops and nasals fricatives and liquids

o all [stop] segments: o all sonorant [stop] segments:

obstruent stops, nasals and laterals nasals and laterals

o all [cont] segments to the exclusion of [stop] segments:

fricatives and / r /

o all segments under [cont]:

liquids

Non-natural classes are fricatives and nasals, or rhotics and nasals, or obstruent stops and rhotics. Some of the natural classes in this model are amply motivated in phonology, and will only be briefly discussed here. For example, obstruents are typically those segments that can have a voicing contrast. The class of obstruents can be expressed as follows (where a ¿ 0): (16)

C I a

90 In some languages sonorants are the only segments that can close syllables, they may be the only segments permitted to bear tone or be syllabic, etc.10 This class can also be expressed configurationally (where α, β ¿ 0): (17)

C I a I β

Liquids typically show identical behaviour in syllabification and clearly act as a class in dissimilation. The class of liquids is expressed as follows (where α φ 0): (18)

C I [cont] I a

I will discuss the following classes in more detail: (i) stops and nasals, (ii) fricatives and liquids (possibly excluding [-cont] liquids), and (iii) nasals and laterals (possibly including, in addition, the obstruent stops). The natural class of stops and nasals is expressed as follows (where a is either present or not): (19)

C I [stop] I a

Some examples that show that nasals and stops can form a natural class are provided by Anderson (1976a) and Poser (1981).11 Anderson (1976a) points out that in the Brythonic dialects of Celtic the morphological rule of lenition had the following effects:

10

11

In a language in which both vowels and sonorants can bear tone, the class will have to be referred by way of a disjunction. The formal 'cost' of this disjunction seems small, especially when it is taken into account that in a framework with the feature [sonorant], the group of segments that is [+sonorant] is larger than the group of segments that is [+sonorant, -consonantal]. If anything, this predicts that it should be more usual for a language to allow tone on all sonorants than on only vowels. Our model makes the opposite prediction, namely that it is more typical for vowels alone to bear tone than for vowels and sonorants to bear tone, which to the best of my knowledge is correct. I would like to thank William J. Poser for making his unpublished work available to me.

91 (20)

a. b. c.

voiceless stops became voiced voiced stops became spirants / m / became [ß]

Fricatives were unaffected. Poser notes that if nasals are treated as being [-continuant] (c) may be treated as a special case of (b). If nasals are [+continuant], the lenition of / m / will require a separate rule of denasalization. Ceteris paribus, the former treatment is superior. Another case of nasals and oral stops forming a class as the undergoers of a rule is a rule stated by Safir (1979) on the basis of a reanalysis of Loos' Capanahua data (Loos 1969). In Capanahua nasals and oral stops, but not oral fricatives, are deleted in syllable-final position. In several other cases nasals and oral stops form a class as triggers of a rule. Anderson (1976a) cites the case of Finnish, which has a rule that converts / k / to / h / before stops. This rule applies both before oral stops and before nasals, but nowhere else, for instance not before / s / . A second case of this type is due to Foley (1977), who observes that in Classical Latin the long thematic vowel / a / of First Conjugation verbs shortens when it occurs in the final syllable and is followed by an oral stop or a nasal, but not when followed by /s/. This may be seen in the following paradigms of the present and imperfect tenses of the verb amare 'to love'.

(21) 1 2 3

SG

PL

SG

PL

amò amäs amat

amämus amätis amant

amäbam amabas amäbat

amäbämus amäbätis amäbant

The thematic vowel is seen to shorten before / t / , / n t / and / m / but not before the / s / of the second person singular present. A final example is found in Luiseño (Davis 1976), which has the following rule:12 (22)

c— g / _

# [-cont]

This rule applies before nasals, as in the example below, but not before oral fricatives: (23)

12

[néSmal]

'old woman'

from

nécu

'become an old woman'

I return to this rule in more detail in section 2.3 of chapter 7. Note that the leftmost [-cont] part of the affricate is deleted before a [-cont] feature to the right of the affricate. This rule is an example of an 'anti-edge effect' and is therefore evidence for the position that in affricates the Manner features are phonologically unordered.

92 These cases amply illustrate the fact that stops and nasals form a natural class. Maddieson and Precoda (1992: 57) also observe that these segments typically occur at the same places of articulation in phonological inventories. We will see below that Place features are attached to the same Manner feature, [stop], in both stops and nasals. The class oí fricatives and liquids can be shown to behave naturally in Frisian (Sipma 1913; Tiersma 1985; Gussenhoven 1991). If / n / is followed by a fricative or liquid, it deletes and the preceding vowel is nasalised. This process is shown in the data in (24)a below. The data in (24)b provide examples in which the consonant following the /n/ is an obstruent stop or a nasal, which context does not trigger lengthening. ynjean ynswart oanstean ynfalle ynwenje oanroppe oanfleane ynlizze ynpakke yngean oantrekke oannimme oankomme

[ija:n] [i :swat] [o:5stian] [ifals] [C vêjia] [oârops] [osfliana] [C ltza] [impaka] [iqgisn] [oantreks] [oannima] [oaqkoms]

'to give in* Very black' 'to please' 'to fall in' 'to live with one's parents' 'to call' 'to attack' 'to preserve' 'to wrap up' 'to enter' 'to take to heart' 'to accept' 'to arrive'

We also expect rules that would treat fricatives and /r/ alike, to the exclusion of / I / . These would be the segments that have [cont] somewhere in their representation, but not [stop]. A rule which targets this class of segments is the Scottish vowel length rule (Aitken's Law), in which vowels are lengthened before voiced fricatives and /r/, but remain short for /l/. The rule is formalised as follows by McMahon (1991: 33): (25) I X - X X/ I \ / V V [ +tense]

3 ö

Examples that show the length/quality alternation appear in (26) (partly from McMahon 1991: 33):

93 (26)

/i/

[i] [i:]

beat beer

wreath wreathe

leaf leave

bean agree

feel bee

/ΛΪ/

[ΛΪ] [a:i]

fight fire

life live

lice lies

line tie

mile dry

Morphological factors also play a role, but these will not concern us here. Finally, the sonorants are divided into [stop] sonorants (nasals and / l / ) and [cont] sonorants (i.e. /r/). It is therefore predicted that it should be possible for nasals and laterals to behave as a natural class. Some examples that show that these segments can function as a natural class appear in (27): (27)

a. b. c. d.

In English, nasals and / l / form a context for intrusive stop formation (sen[tjse, fal[t]se) in some dialects (Clements 1987) In Basque, the nasals and laterals spread their voicing to following stops (see Hualde 1991: 135ff.) In Gallo-Romance, palatalisation affected /lj/ and / n j / in the same way as it affected /kj/, and differently from /sj/ and / r j / (Jacobs 1989: 146) In the history of Latin to Portuguese and Galician, intervocalic [1 n] were lost (Lipski 1973: 49)

Apart from the fact that nasals and / l / form a natural class, these segments may form a natural class together with obstruent stops as well. In Basque, precisely these segments trigger Stop Deletion (see § 3.2.1 in Ch. 7). Another case in which this (uncommon) class occurs is in a similar sort of process in Wayana (Jackson 1972), namely 'aspiration' (i.e. realisation as [h]) of / t / (see van de Weijer 1992c for details). We conclude that the predictions made by our proposal are borne out. Other groupings of segments can also be expressed, namely by way of disjunctions (see also note ? above). However, such classes are predicted to be less common. In the next section we turn to the representation of segments which are complex on the Manner dimension.

3. Complex segments: multiple stricture The organisation of Manner features proposed above embodies a calculus which generates a set of segments which have been described as 'complex' in the literature. An algorithm of this sort is completely absent in current work in feature geometry. That is, current phonological theory does not predict which complex segments may occur and which may not. In this section we show how the model proposed here is an improvement. It will be seen that increasingly complex representations denote increasingly rare segments. As before (see Ch. 3), the formal correlate of complexity

94 is branching: all combinations of branches within the model result in possible complex segment types. First, it is necessary to make explicit some assumptions concerning the relation between Manner primitives and Place elements in the model advanced here. I adopt an idea of Wetzels (1989, 1991, in preparation), who attaches the Place node to the feature [ ± continuant] (this is discussed in more detail in § 1 of Ch. 7). Translated into a unary-feature model, this means that Place may be attached to either the topmost [stop] or the topmost [cont] feature (the Manner-Place dependency assumed here is also explored in van de Weijer 1992a, 1992b, 1993a; see also Rowicka and van de Weijer 1992). Following the idea adopted in chapter 2 that Place features are not attached to an abstract node, but rather form a natural set of features, Place features are attached immediately to the feature [stop] or [cont]. Consider for instance, the representation of a velar nasal / η / and that of a labial fricative / f / (for instance in English): (28)

C

C

I [stop]

I [cont]

[nasal]

ν

A

I U

m

In such structures, the features [stop] and [cont] cannot spread independently of the Place element A and U, respectively, while in the case of / q / [nasal] can spread on its own (permitting nasal spreading, as in vowel nasalisation, or in a process like Dakota nasalisation, discussed above). In a complex segment combining the two features [stop] and [cont], this implies that Place may be attached to either of these Manner features. In such a case, the Manner feature that dominates Place will be referred to as the head. In line with the Mismatch condition, discussed in chapter 4, the non-head may not be complex for Place. In fact, I will assume that it is not specified for Place underlyingly (i.e. it does not bear any Place elements), and that is always realised as Coronal in phonetic interpretation, possibly subject to further assimilation. The relevant issues are also discussed in van de Weijer (1993a). Note that the part of the complex segment that is specified for Place in underlying representation is phonetically realised after the part that is not specified for Place in underlying representation. Complex segments, then, involve a combination of expansions of the Manner tree. There are five basic (i.e. non-complex) expansions given in (5) and (7) above, and repeated in (29):

95 (29)

'Manner types' obs stops

obs fries

nasals

laterals

rhotics

C I [stop]

C I [cont]

c

c

c

I [stop]

I [cont]

I [cont]

I [nasal]

I [stop]

I [cont]

I hypothesise that only two (not three) of these expansions can be combined into a segment complex. The typical branching relation involves two elements (see Ch. 3). In Table 1 below, the five 'Manner types' are cross-classified. This leads to 25 possible combinations, some of which will be seen not to correspond to possible segments, however. By surveying the existing segments, we will be able to postulate a restriction which limits the 25 possible complex segments to the actually occurring ones. stop stop

fricative

nasal

lateral

rhotic

1

2

3

4

5

6

7

8

9

fricative

X

nasal

X

X

lateral

X

X

X

rhotic

X

X

X

10 X

Table 1 Cross-classifying the five 'Manner types' These twenty-five combinations can be reduced by making two natural assumptions. First, we hypothesise that segments combining identical Manner types, i.e. [t] plus [t], [s] plus [s], etc., are not meaningful. This is why the NW-SE axis in Table 1 is shaded. Second, I assume that ordering of the two Manner types in a segment is not distinctive (this is why 'x' symbols appear in the bottom left-hand half of Table 1). Both assumptions are stated in (30): (30) a.

*C / \ [stop] [stop]

b.

C / \ [stop] [cont]

C = / \ [cont] [stop]

The second assumption follows from the basic premiss in autosegmental phonology that different primitives appear on different tiers. It is therefore not a stipulation. The first assumption does have to be stipulated, however, and may appear suspect in the

96 light of the fact that identical Place elements may be combined within a single segment (see Ch. 6 below). However, I will retain this hypothesis as a working assumption. The result is that only the cells numbered (1-10) in Table 1 are predicted to be meaningful complex segments. We will see below that at least one more assumption is necessary to arrive at a set of complex segment types that forms a reasonable match to the set of segments actually attested. As we pointed out above, in a segment which combines two Manner features, Place can in principle be attached to either Manner feature. However, there is an overriding condition: in liquids, only Coronal may be attached, that is, in segments which have one of the liquid expansions (which consists of [cont] dominating either [stop] or [cont]), Coronal must be attached to the liquid expansion, and must only be attached there (a segment with Place attached to two Manner features would have two heads, which is ruled out). In (31) the representations of the liquids /I r / are given: (31) a.

C I [cont] [stop]

b.

C I [cont]

Coronal

[cont]

Coronal

M

Ν

After these preliminaries, we can turn to the cells 1-10. Cell 1 represents affricates ((32)a) or, I propose, / s / plus stop clusters ((32)b), depending on which of the Manner features Place is attached to: (32) a.

C

b.

C

/ - I [stop] [cont] I Place

/ I [cont] [stop] I Place

affricates

/s/ plus stop clusters

I will not discuss these segment classes any further here. It turns out that the part of the segment linked to Place in underlying representation is realised phonetically on the right: in affricates the fricative part determines the place of articulation, and in / s / plus stop clusters the stop part has the unpredictable Place specification, while the underlyingly unspecified parts are realised as Coronal. For clarity, I will therefore represent the branch which carries Place on the right. For further discussion of

97 affricates, see chapter 7. The behaviour and representation of / s / plus stop clusters is discussed in more detail in chapter 9. With respect to the combination of [stop] and [stop] dominating [nasal] (cell 2 in Table 1) there are again two possibilities: (33)

a.

b.

C

*

C

[stop] [stop]

I [stop] [stop]

I I [nasal] Place

[nasal]

prenasalised stops

prestopped nasals

Place

In (33)a the representation is that of prenasalised stops, such as [nd mb]. In chapter 8 we will return to the fact that Place in these segments is underlyingly just attached to the rightmost [stop]: this reflects the insight that Place is underlyingly just specified on the stop part of a prenasalised stop. Hence, the nasal part of the segment must receive its Place specification by a default mechanism. If Place is attached to the leftmost [stop] ((33)b), the representation is that of prestopped nasals (e.g. [dn]). Such segments do not, as far as I know, function as independent phonemes in any language. They will therefore tentatively have to be excluded from the set of possible phonemic segments. This is achieved by a condition that will be discussed below (§ 4). With respect to the combination of [stop] and [cont] dominating [stop] or [cont] (cells 3 and 4 in Table 1), we have assumed that in liquids the Place feature Coronal must be attached to the 'liquid complex' [cont] dominating [stop] or [cont] (see (31) above). This makes these combinations straightforward: represented are lateral affricates ((34)a) and rhotic affricates ((34)b): (34)

a.

C

b.

^

C

.

[stop] [cont] [stop]

I

[stop] [cont] Coronal

[cont]

Coronal

lateral affricates rhotic affricates In (34)a the representation of lateral affricates such as / t i / is given, which occur in native American languages such as Navaho and Tlingit, as well as others (Maddieson 1984). Again, the [cont] feature may dominate [cont] instead of [stop] (in either case the segment will receive a Coronal specification). The resulting segment is a rhotic affricate ((34)b). Phonological evidence about such segments is rare. However, in

98 Araucanian (a language spoken in Chile and Argentina), described by Echeverría and Contreras (1965), a rhotic retroflex affricate occurs, which alternates with a single retroflex stop. The language allows no clusters except for ordinary affricates and these rhotic affricates, which is why both are regarded as single segments (see also Suárez 1959: 178; Key 1978: 284). Similar evidence about rhotic affricates can be obtained from Melanesian languages. The same sort of segment has also been reported for Malagasy, by Dyen (1971), following the phonemic analysis by Dahl (1952) (see also Maddieson 1984: 86), and for Avokaya (Dimmendaal 1986). Finally, Dunstan (1963: 235) lists a tr phoneme for the Ngwe language, spoken in West Cameroon, but does not offer further discussion. Note, finally, that the stop parts of these affricates must be coronal. There are no /pT/-like lateral affricates. This is an effect of the Mismatch Condition (see Ch. 4), which dictates that the weak branch of a complex segment must be simplex. Since the liquid part of the affricate is specified for Place, this is the head, and the other, [stop] part must be left unspecified. In this study, these classes of complex segments will not be further discussed, since lateral affricates are not controversial, and, apart from the 'sightings' mentioned above I know of no other reports of rhotic affricates. With respects to the combination of [cont] and [stop] dominating [nasal] (cell 5) there are again two possibilities, provided in (35): (35)

a.

C

b.

C

[stop] [cont]

[cont] [stop]

I I [nasal] Place

[nasal]

prenasalised fricatives/ prenasalised affricates

Place

prefricativised nasals

The representation in (35)a may be reasonably interpreted as that of a prenasalised affricate or a prenasalised fricative: the structure may either be conceived of as that of an affricate with an additional nasal feature, or as a fricative with preceding nasality. No contrast can be expressed between the two categories. This may be a good result, because, as Steriade (1993) points out, (a) there are no languages which contrast the two, (b) there are languages in which the two occur in free variation, and (c) there are languages in which prenasalisation of a fricative results in a prenasalised affricate. The hypothesis in (35) that the two have the same phonological representation is therefore a desirable one: whether it is pronounced as a fricative or an affricate is left to the phonetic interpretation rules of the language. A problem arises when the other possibility for attaching Place is considered ((35)b). This would result in prefricativised nasals (e.g. [sn]), which have not been reported as phonemic segments. Like prestopped nasals, they will have to (tentatively)

99 excluded from the set of possible complex segments. However, since their representation is very similar to that of / s / plus stop clusters, an interpretation as / s / plus nasal clusters may be possible, if these may be shown to behave as unit segments. We now turn to the combination of [cont] and [cont] dominating [stop] and that of [cont] and [cont] dominating [cont] (cells 6 and 7 in Table 1). Both are given in (36): (36)

a.

« C

b.

[cont] [cont] [stop]

*

C

[cont] [cont] Place

[cont]

Place

Both these combinations seem uninterpretable, or more precisely, the kind of segments that are represented (prefricativised laterals [sl] and prefricativised rhotics [ s r]) seem to be excluded as single, complex segments. Below (§ 4), I will make a proposal as to how such segments should be ruled out. With respect to the same combinations, but with an additional [nasal] under the [stop] feature (representing cells 8 and 9 in Table 1), the interpretations are straightforward: represented are prenasalised laterals ((37)a) and prenasalised rhotics ((37)b). Again, Coronal must be attached to only the liquid part. (37)

a.

C [stop] [cont]

ι

r

[nasal] [stop]

b.

\

Coronal

prenasalised laterals

C [stop] [stop]

ι

[nasal] [cont]

Coronal

prenasalised rhotics

In this model prenasalised liquids are the only kinds of prenasalised sonorants permitted. In Fiji, described by Maddieson (1989), the prenasalised post-alveolar stop is sometimes realised as a prenasalised trill, so a prion we should not rule out this kind of segment (for the Fiji phoneme, see also Hockett 1955: 124 and Schütz 1963). As such a segment is the most elaborate of the complex segments generated here, with expansion on both Manner sides, we may expect it to be among the rarest. Alongside the prenasalised rhotic in (37), a prenasalised lateral is also predicted to exist. The existence of such a segment has not been proven, although it does not seem as inconceivable as the 'prefricativised rhotic' above, for instance. The Katla group of northeast African languages is reported to have "[nasal] compounds with liquids" (Tucker and Bryan 1966: 263), but it is not clear whether single segments or sequences are involved. I propose that prenasalised liquids should tentatively be included in the class of possible complex segments.

100 Note furthermore that (37)b may also represent a prenasalised rhotic affricate, just as (35)a could represent either a prenasalised fricative or affricate. A prenasalised rhotic affricate [ndr] occurs in the Nigerian language Logo (Goyvaerts 1983), where it is explicitly claimed to be a single segment on the basis of syllabic structure (but see also Dimmendaal 1986: 6f., who challenges Goyvaerts' analysis). The final representations (cell 10) involve two fully specified branches, as in (38): (38)

a.

* C

b.

/ - I [cont] [cont]

I

[cont] [stop]

*

C

/ I [cont] [cont] Coronal

ι

[stop] [cont]

Coronal

Represented are 'prelateralised rhotics' or 'prerhoticised laterals', which seem uninterpretable (or rather, again, the kinds of segments represented have not been reported, either as phonemes or otherwise). We might point out that this absence is technically already accounted for by the assumptions made so far: since Coronal must be attached to the liquid complex, both sides of the complex segment would have to be represented with a Place feature, the segment would be double-headed, and therefore ruled out.

4. Complex segment results If we cross out the combinations that have been found to be uninterpretable from Table 1, a generalisation emerges. Consider Table 2 below, where all ill-formed or meaningless representations are noted by 'x', and the felicitous representations are filled in by the corresponding segments:

stop

fric

nasal

lat

rhotic

stop

X

affr/ssc

pns/x

laf

rhof

fric

X

X

pnf/x

X

X

nasal

X

X

X

pnl/x

pnr/x

lat

X

X

X

X

X

rhotic

X

X

X

X

X

Table 2 Results (affr=affricates; ssc=/s/ plus stop clusters; pns=prenasalised stops; laf=lateral affricates; rhof=rhotic affricates; pnf=prenasalised fricatives or affricates; pnl=prenasalised rhotics; pnr=prenasalised rhotics)

101 The generalisation that arises is that only complex segments that involve a complete constriction on the obstruent level are allowed. Affricates have an oral [stop] feature, prenasalised segments have this feature as well as [nasal]. In terms of the model advanced here, this may be posited as a restriction on the class of possible complex segments: (39)

A complex segment must have [stop] on the obstruent level

The restriction may be related to the fact that [stop] is the unmarked Manner feature. With the condition in (39), all cases can be accounted for, except for the fact that it allows for the prestopped and prefricativised nasals. Their absence must be stipulated: (40)

In a complex segment which has [nasal], Place may not be attached to the [stop] feature that dominates [nasal]

With these two restrictions we have represented all possible complex segments and only the complex segment types that occur in the world's languages. A list of the existing complex segment types is given below: (41)

with oral stop:

with nasal stop:

unclear:

affricates / s / plus stop clusters lateral affricates rhotic affricates prenasalised stops prenasalised liquids prenasalised affricates/fricatives prestopped nasals prefricativised nasals

5. Conclusion The geometry of Manner features developed here is able to capture a number of facts about Manner features and about segment types. It affords a non-arbitrary interpretation of natural classes like stops, continuants, etc., and marks off obstruents from sonorants in a structural way. It also provides an explicit calculus for the generation of complex consonants and complex segment types like / s / plus stop clusters. In the following chapters we will examine affricates (Ch. 7), prenasalised stops (Ch. 8) and / s / plus stop clusters (Ch. 9) in more detail, to see if the representations derived here can be maintained in the light of phonological rules and other relevant phenomena. First, however, we will turn to complexity on the Place dimension.

Chapter 6 Colour mixing in consonants In this chapter I investigate double Place specifications in consonants. Such a branching Place specification results in the type of phonological complexity I referred to as 'colour mixing' in chapter 3. The simplex Place representations (i.e. those in coronals, velars and labials) were dealt with in chapter 2; now we will turn to 'complex' places of articulation, such as palatals, palato-alveolars and uvulars. In this chapter I will start from the strongest hypothesis possible, namely that all places of articulation other than labial, coronal and velar involve some type of phonological complexity. This hypothesis is to some extent also inherent in Sagey's (1986a) feature geometry, which has three articulator nodes under the class node Place, as shown in (1) (only the Place node and the structure immediately underneath are represented): Place

(1) Labial

Dorsal In a language with three places of articulation, just the three articulator nodes need to be posited. These nodes act as unary features: a labial consonant is only specified for a Labial node, and never acquires the minus specification of 'Coronal' or 'Dorsal', simply because there are no nodes that correspond to those values. This is much the same as in the Dependency Phonology approach to vowel representation, where the elements I, A and U are single-valued features (recall that the DP approach to consonant place of articulation advanced in Anderson and Ewen 1987 is different, since it argues that velars are complex segments, consisting of the elements for labials and coronals; see § 2 in chapter 2). Representations of stops with simplex places of articulation in Sagey's model are given in (2) (Manner and Laryngeal specifications omitted): (2)

C

C

C

Place

Place

Place

Labial

Coronal

Dorsal

/P/

Ν

M

If a language contrasts more than three places of articulation (for instance because it has palatals or labial-velars), the contrast can be made in two ways in

104

Sagey's model. Place elements may be combined, in which case a complex place of articulation results, or a feature may be present that subclassifies the primary place of articulation. In the latter case, features like [± anterior], [± distributed] and [± strident] are invoked, all of which are under the Coronal node: (3)

Coronal

[ ± strid] Note that in Sagey's model, common places of articulation are represented in a simple way, while divergences from this pattern require postulating extra structure. The two ways in which extra structure can be added can be referred to as 'combination' and 'subclassification'. Both these strategies are shown in (4): (4) a.

o / \ a ß

b.

combination

o I a I β subclassification

Sagey (1986a) makes use of both strategies in (4) to express Place contrasts. For instance, labial-velar stops are represented as in (5)a below and [ +distributed] stops (as opposed to [-distributed] stops) are represented as in (5)b: (5) a.

C I Place / \ Labial Velar

/ m

b.

C I Place I Coronal I [+distr] Ν

In effect, Sagey makes use of two different sets of Place features: the first of these sets consists of the 'primary' Place features {Labial, Coronal, Dorsal} and the 'secondary' set of Place elements consists of features like [distributed], [strident], etc. The secondary features can only come into play when a primary Place feature is present.

105 In this study it is assumed that only a single set of Place elements exists, namely {I A U}. Hence, only the first way of combining elements is available. In this chapter it is examined to what extent this strategy can successfully characterise all and only existing segments. Of course, it is not a priori excluded that the second strategy is adopted in addition to the first, but this would involve a serious weakening of the model advanced and should therefore be avoided as much as possible. 1 This radical proposal is not unprecedented in the literature. At least for contrasts in the coronal region there has been a shift from using the second strategy, i.e. subclassification, to the first: Keating (1988b) and Jacobs (1989) argue that palatals and palato-alveolars essentially involve a combination of the Coronal and Dorsal nodes, rather than being characterised as Coronals with an extra secondary feature (see below for further discussion). Such a strategy enables us to do away, at least partly, with subclassifying features such as [± anterior], which were previously used to make distinctions in this region. This chapter is organised as follows: in § 1 I investigate possible combinations of the elements that were adopted for the representation of Place in consonants. Of particular interest is the representation of different places of articulation in the coronal area, and the representation of the uvular place of articulation. These will form the topics of §§ 2 and 3, respectively. As we will see, only the first of these involves colour mixing proper, that is, the simultaneous presence of two Place elements in one segment, while the second will be argued to be the result of secondary articulation, and thus involves what I called 'two-rootedness'. Specifically, the uvular place of articulation is represented as an A-type vocalic articulation superimposed on an A-type (velar) consonantal articulation. It will nevertheless be discussed here, since phonetically it involves a single articulation, rather than the more obvious secondary articulation types like those in [k*], [p>], etc.

1. Combining the elements In (6) I repeat the Place elements, together with their consonantal interpretations: (6)

I A U

Coronal Dorsal Labial

The possible combinations of these elements are given in (7), with headship (which may of course be redundant in specific cases) indicated as before (see Ch. 1) (note that these are (4)a-type combinations, since elements of the 'primary' set {A I U} are

1

Another way of combining more than one Place element in a single segment is by allowing two root nodes, both of which can dominate separate Place features. Below (section 3) I argue that uvulars have this kind of complexity.

106 combined - it is an unfortunate coincidence that the expression of the head-dependency relation as used in this study is the same as that used to express the type of combination that was glossed as subclassification in (4)b): A I 1 A

I 1 1 A

υ I 1 A

A I 1 I

I I 1 I

u 1 1 I

A I 1 u

I 1 1 υ

υ 1 1 υ

There are three combinations of two identical elements (i.e. I and I, U and U, A and A), and six combinations of two different elements. These two groups will form the topic of §§ 1.1 and 1.2, respectively. At the end of § 1 a table will be provided with the various interpretations proposed.

1.1 Combining identical elements 1.1.1 Bilabial v& labiodental At first glance it might appear odd for two identical Place elements to occur in a single segment. There are even well-motivated phonological conditions, such as the OCP (see § 2.1.1 in Ch. 2), that might rule out such combinations. At the same time, however, combinations of identical elements are not rare at all in phonology: a metrical foot, for instance, typically consists of two syllables, and reduplication consists of the doubling of segmental material in a base. On the basis of the fact that structures in which a head dominates two identical daughters exist elsewhere in phonology, I will tentatively assume that, although it may not be common, identical Place elements may also appear in single phonological segments. An alternative to this approach might use a subclassificatory feature. In combinations of identical elements no head-dependency relation can hold contrastively between the two elements that are combined.2 Let us first turn to the combination of two U elements.

2

See section 2.2.1 in chapter 3 for a brief discussion of the combination of identical elements in vowels.

107

A long-standing problem in phonological segment representation is the way in which the contrast between bilabial and labiodental fricatives should be represented. In Chomsky and Halle (1968: 314), bilabials are [ +distributed] and labiodentals are [-distributed]. This approach was rejected by Sagey, who proposes that the feature [± distributed] should only be used for making contrasts among coronals. She points out that [-distributed] labials and [-distributed] coronals do not act as a natural class in phonological rules. To make a distinction between labials and labiodentals, she proposes that a combination of [+cont] and the articulator node Labial is phonetically interpreted as labiodental (Sagey 1986a: 278). Thus, a labial stop is interpreted as bilabial, and a labial fricative is interpreted as labiodental. However, this does not solve the question of how the contrast between labial and labiodental fricatives should be represented. This is not a common contrast, but it is generally accepted that it occurs in Ewe, where voiced as well as voiceless labial fricatives contrast with labiodental ones (see Westermann and Bryan 1970: 90; Maddieson 1984). According to Maddieson, Irish and lai also have a phonemic contrast between labial and labiodental fricatives. The contrast is therefore possible, and phonological theory should be able to express it. To express the contrast between labial and labiodental fricatives, I propose that a double- vs. single-U representation could be used, rather than a feature like [± distributed]. The bilabial fricative is phonetically obviously 'more labial' than the labiodental one, so that a representation in which the former is represented with two U elements, and the latter with one U element, is intuitively justified. This way of representing the two segments entails that bilabial fricatives are complex segments as compared to labiodental ones (in languages which have both kinds of fricatives), since the former have a double Place specification. No such implication exists in a binaryfeature representation of the contrast. In languages which have only bilabial fricatives, such as Hausa (Maddieson 1984), there is no contrast among fricatives in the Labial region, so there is no need to posit a complex segment representation: Hausa /Φ/ can be represented just with a single U element. Hence, it is the system in which two kinds of labial fricatives occur that is complex, and not necessarily the representations of the segments themselves. Bilabial and labiodental fricatives, then, are proposed to be represented as in (8): C

c

[cont] / \

[cont]

(8)

U Ν

U

U m

The relative complexity between the two classes of segments is clearly expressed in (8). Further evidence for making the contrast in this way might come from languages

108 where bilabial fricatives show a greater labialisation effect on vowels (or on other consonants) than labiodental ones. Such evidence is not available. Let us now investigate the representation of bilabial and labiodental stops and nasals. In this case, the bilabials are the most common, and the labiodentals are rare. Labiodental stops, if they exist at all, can be ruled out phonetically, since the acoustic signal they produce is not robust.3 However, labiodental nasals are not so rare, [irj] exists in English as an allophone of /n m/ when these occur before labiodental fricatives (as in infant, symphony and invent). Since this concerns an allophonic effect we will not deal with this case here. A phonemic contrast between /n]/ and / m / exists in the Congo language Teke (Paulian 1975; Maddieson 1984). Consider the following minimal pairs in this language (from Paulian 1975: 36): (9)

mii (ki)màâlà

'urine' 'to complete'

njfl (kì)n]ààlà

'eyes' 'to mock'

I suggest representing the difference between /rrj/ and / m / as follows: (10)

c I [stop] / U

C I [stop]

[nasal]

A

[nasal]

U U M

M

That is, the bilabial nasal should be represented as the more labial segment of the two. There is some evidence within the language that the contrast can be expressed in this way. In Teke, /iq/ is not allowed before the rounded vowels / u / and / o / : only / m / is found in this position (Paulian 1975: 41). The language has a five-vowel system /i e a o u/. Data that show that / m / does occur before these vowels appear below: (11)

mù (mù)nkònòmò kì.tshùùmù

class prefix 'a stroke' 'arboreal rat'

These data can be expressed by assuming an underlying /rq/ which becomes more labial (i.e. bilabial) before the labial vowels / u o/. This is expressible as an assimilation rule if the above representations are adopted: 3

We might also speculate that a labiodental stop is interpreted as a labiodental affricate [pf], following Steriade (1989). I know of no evidence that would indicate a difference in representation between coronal affricates (with [stop] and [cont]) and labiodental affricates (with just [stop]). Hence, I will not pursue this approach here.

109 (12)

C

V

I [stop] [nasal]

U

The rule in (12) turns the labiodental nasal into a bilabial one in the specified environment.

1.1.2 Dental vs. alveolar I propose to represent the contrast between dental and alveolar consonants by way of a double- vs. single-I-representation. These places of articulation contrast in more languages than those in which bilabials and labiodentals contrast (see Maddieson 1984). For instance, dental and alveolar fricatives contrast in English ( / s / vs. /%/). The contrast for stops exists in many Australian languages, as well as the Hokan languages Pomo and Diegueño, and others. Consider for instance the (partial) consonant inventory of Western Desert (360) (Maddieson 1984): (13)

voiceless plosive voiced nasal vd. lateral approximant

ρ m

J; ji

t η i

l

t ρ l

k q

In this language dental and alveolar stops, nasals and laterals contrast. I propose that dentals, which are articulated more to the front than alveolare, have a combination of I elements, while alveolare are represented with one I element in languages which have both. The proposed representations are given in (14) below. c ι 1 [stop]

/\

I /t/

I

c ι 1 [stop] 1 I

Ν

Another way of distinguishing between dentals and alveolare is made by Pulleyblank (1989: 384f.), who proposes that dentals are [+front] and alveolare are [-front] (both are coronals, i.e. have a Coronal node). The propoeal made by Pulleyblank is akin in spirit to the one advanced here. Both predict that dentals should be more likely to front vowels, for instance, or that coronals should have dental allophones before front vowels and alveolar ones elsewhere. One difference between Pulleyblank's proposal

110 and that advanced here is that in the former two binary features are assumed, [± front] and [±back], as well as Coronal and Dorsal articulator nodes. In my proposal, on the other hand, there are only single-valued elements (comparable to the articulator nodes), which may combine with each other or with themselves.

1.1.3 Pharyngeal vs. velar Finally, the combination of two A elements may be used to represent pharyngeal segments. The representation with two A elements again expresses a more 'extreme' articulation than one with only a single A element: pharyngeals are further back than velars. Secondly, pharyngeals are represented as complex relative to velars, which is appropriate as they are so much rarer. However, it might be argued that other places of articulation could also be candidates for having representations with two A elements, specifically, the uvulars. I will argue below that uvulars should be represented by way of a vowel-like secondary articulation, rather than with an extra consonantal element (see § 3 below). This entails that velars, uvulars and pharyngeals form a natural class, since all have at least one A element. Maddieson (1984) reports one pharyngeal stop in Iraqw (Whiteley 1958).4 Pharyngeal fricatives are rather more widely reported, for instance in a range of AfroAsiatic languages, such as Somali, Arabic and Tigre, and in Caucasian languages such as Kabardian. The representations of pharyngeal and velar fricatives that I propose are given in (15): C

(15) [cont] A

Ν

A

[cont] I A

M

The representations in (15) predict that pharyngeals should have a greater backing and/or lowering effect on vowels than velars. This is indeed the case in Somali (Armstrong 1934), where the pharyngeal and the glottal fricatives "have an opening effect on vowels" (Armstrong 1934: 134). For the velars no such effect is reported. A similar phenomenon occurs in Kabardian (Kuipers 1960), which also has voiceless and voiced pharyngeal fricatives (although the latter occurs only in Arabic loanwords, 4

That this language has pharyngeal consonants is confirmed by Mous's (1992) study (see also van der Hulst and Mous 1992). However, Mous (1992: 18) describes /V/ as "a pharyngeal constriction (not a stop) produced with creaky voice". Thus, it would seem to be the case that there are no true pharyngeal stops in the languages surveyed by Maddieson.

Ill where it is usually replaced by its voiceless counterpart; see Kuipers 1960: 22). Kuipers states that the phonetic realisation of the short vowels 3 and a depends on the surrounding consonants: "front variants (i e) are found after laterals, palatalized palatoalveolars and j, back variants (y a) after plain uvulars, pharyngeals and h, ? (after the latter two consonants central vowels are also heard), back rounded variants (u o) after labialized palatoalveolars, uvulare and laryngeals, central variants (9 œ) after other consonants". Thus, velar consonants have no effect on the realisation of the vowel, while pharyngeals cause the central vowels to become back. A final, well-documented, case of pharyngeals that cause lowering comes from Palestinian Arabic (Herzallah 1990; Clements 1990b). Here the vowel [a] shows up instead of an expected [i] when the vowel occurs next to one of the segments / χ γ fi î h ?/: Perfective

Imperfective

fatali manaY nahab sa?al farax mara7

yi-ftafi yi-mnaî yi-nhab yi-s?al yi-frax yi-mra7

'to 'to 'to 'to 'to 'to

open' deprive' rob' ask' scream' stain'

Clements points out that the imperfectives in (16) would normally have an ablaut pattern with [i], but that the presence of the pharyngeal causes the vowel to be [a] instead.

12 Combining different elements There are six possible combinations of two different elements, if we take into account the fact that in a combination of two elements a head-dependency relation may hold. First I will devote two short paragraphs to combinations of U with either other element. Section 2 will deal with combinations of I and A.

12.1 Labial-velars Let us first turn to combinations involving the element U. Obviously, this will be interpreted as a segment with a labial articulation. The combination of U with the element A can be interpreted as a labial-velar. For instance, a voiceless labial-velar stop /kp/, widely reported in African languages, is represented with both these features under the feature [stop], as in (17):

112 (17)

C I [stop]

/\

A

U

m/ Labial-velar stops are commonly cited as examples of segments that have a complex Place specification (see e.g. Sagey 1986a). That they are single segments is uncontroversial. For instance, when nasals assimilate in Place to labial-velar stops both elements may spread from the stop to the nasal, resulting in a doubly articulated labial-velar nasal /nfi}"/. This would be completely unexpected if [kp] were a cluster. A second argument comes from languages such as Ewe, where single consonant reduplication simplifies onset clusters, but leaves the labial-velar stop /1φ/ intact (see Sagey 1986a: 86; Ansre 1963). Besides labial-velar stops, labial-velar nasals and fricatives have also been reported. Labial-velar nasals may result from nasal assimilation to labial-velar stops. The labial-velar fricatives /xw/ and /7w/ occur in the Nigerian language Urhobo (both are truly labial-velar, according to Ladefoged 1968: 26 (Table 5), see also Dunstan 1969: 154f.).5 All of these labial-velar consonant types are rather rare. A category of segment that is not at all uncommon and that might also at first glance be categorised as a labial-velar is the approximant /w/. The segment is listed as such in Maddieson (1984). A language in which the labial-velar glide shows its double Place is Tenagno Otomi, where a preceding glottal fricative becomes bilabial when it precedes /w/, while a preceding nasal becomes velar in the same position (Ohala 1993). However, I will adopt the claim that the labial-velar approximant belongs to the class of complex vowels, syllabified into non-nuclear position. This view of the relation between vowels and glides has been widely adopted, although it is not uncontroversial (e.g. Steriade 1984; Guerssel 1986; van der Hulst and van Engelenhoven 1994). Finally, the question might be raised if a head-dependency relation exists between the two elements A and U in labial-velars. There do not seem to be two phonetically different segments that both combine a labial and a velar articulation. However, a number of phenomena indicate that headship does play a role in labial-velars, as

5

Greenberg (1983) discusses the occurrence of labial-velars as an areal characteristic of African languages, and reports that "most commonly languages have either an unvoiced or a voiced stop or both. The corresponding nasal is far less common and is almost always found in languages which have at least one non-nasal stop. The labiovelar stop has occasionally been reported as implosive. Labiovelar fricatives also occur, though only rarely" (Greenberg 1983: 4). On languages outside Africa, he notes that "the Kate-Ono group of non-Austronesian (Indo-Pacific) languages in northeastern New Guinea and some Austronesian languages of Melanesia have these sounds" (Greenberg 1983: 5).

113

pointed out by Anderson (1976b). For instance, asymmetries in segmental inventories indicate that in some languages the labial-velar functions as a velar stop, while in others it functions as a labial stop (see Anderson 1976b: 22-23). Nasals may also assimilate to either the labial or the velar part, Anderson maintains. Such facts can naturally be expressed as differences in headship. This would entail that two segments which differ solely in headship can be realised the same phonetically.

1.2.2 Labial-coronals Let us now turn to possible interpretations of the combination of the I (Coronal) and U (Labial) elements under a single Place node. This representation can be used to represent labial-coronal stops like [φ], for instance, which have been reported for West African languages by Ladefoged (1968: 11): "There are (...) languages which have labial alveolar stops such as jfT and bei. In Gur languages such as Dagbani these sounds are allophones of Ι φ and gTî which occur before front vowels. But in the Chadic languages Bura and Margi the labial alveolar mode of articulation is clearly on an equal standing with all other points of articulation". These segments have also been reported for Yeletnye, spoken in New Guinea (Ladefoged and Maddieson 1986). A labial-coronal stop might, then, be represented as in (18): (18)

C I [stop]

/\

U

I /Φ/

More recently, labial-coronal fricatives / S v / (with voiceless, murmured and prenasalised phonemes) and affricates (which occur voiceless and murmured) 6 have also been reported, namely for Shona by Myers (1991). These are "produced with two simultaneous fricative occlusions: a bilabial one and an apico-alveopalatal one (Doke 1931: 87)" (Myers 1991: 4). Myers emphasises that these are not sequences (in spite of what the orthography suggests), since the two articulations are simultaneous. The language does not allow initial consonant clusters except for these purported complex segments. Myers shows that a unit interpretation of these segments also allows an account of consonant mutation in this language. To the extent that his analysis is successful, it supports the representation of labial-coronals as single segments. Since both places of articulation occur with the same Manner type, a representation such as

6

That is, this language conforms to the generalisation that a language cannot have both prenasalised fricatives and prenasalised affricates (see sc. 3 in Ch. 5).

114

that proposed above seems suitable. A representation of labial-coronals with two Manner features, or even two root nodes, leaves this fact unexplained. The analysis of certain other labial-coronal segments as single units is controversial. Maddieson (1983) argues that the labial-coronals in Bura are consonant clusters rather than complex segments, as Ladefoged assumed (see above). Maddieson argues that these segments are phonetic sequences of a labial followed by a coronal, on the basis of the fact that they have a longer duration than single consonants. However, Sagey (1986a) presents convincing evidence that, phonologically, these labial-coronals behave as unit segments (see further § 1.2 in Ch. 10). It is possible that in some cases labial-coronals should not be represented as in (18), but as segments with two roots. I will assume on the basis of studies like that of Myers (1991) that labial-coronals do exist as single segments, and therefore that phonological theory should be able to represent such segments. It may not be possible to make out whether this should be achieved as in (18) or by way of two-root segments (see chapter 10), on the basis of the available evidence. After this discussion of labial-velars and labial-coronals, I turn to the last possible combination of elements, that of I and A. I will suggest that, unlike the combinations already discussed (which both involved U), there is more articulatory 'room for compromise' between the two elements that occur in this combination. That is, a segment combining I and A does not necessarily involve two constrictions at coronal and dorsal places of articulation (in fact, coronal-dorsal stops (like [fk]) or fricatives have not been reported at all, as far as I know),7 but may give rise to what is produced as and sounds like one articulation. This may also be the reason why in the combination of I and A the head-dependency relation may be used contrastively. The next section is devoted to a discussion of the array of places of articulation in the coronal area. Section 3 will deal with the representation of uvulars.

2. Coronal places of articulation The idea that palatals and palato-alveolars combine the elements I (Coronal) and A (Dorsal) is not new. It can be found in Keating (1988b) and Jacobs (1989), among

7

See Lahiri and Evers (1991: 85), who argue that corono-dorsal segments are clicks. I will argue, however, that a coronal-dorsal stop is in fact a palatal stop, so that this apparent gap in the system is not a real gap. Clicks will be discussed in chapter 10, where it will be argued that these segments also involve a coronal and a dorsal articulation, but that in clicks these are divided over separate root nodes (see section 1.1 in Ch. 10). In palatals, they occur under the same root node.

115 others. 8 The present discussion is intended to embed these proposals in a theory of possible complex segments. In this section I examine the phonological representation of a range of fricatives in the area of articulation between the upper teeth and the soft palate. The representations assigned to these fricatives are also assumed to hold for stops (and other segments) in the same region, but here the issue is more difficult, because the status of palato-alveolar stops is unclear (see Lahiri and Blumstein 1984: 139ff.). It is doubtful whether such segments exist at all. The fricatives therefore have the largest number of possibly contrastive places of articulation and present the greatest challenge. The fricatives I examine are listed in (19), arranged in order of increasing anteriority (only voiceless fricatives are given): (19)

/s/ /$/ /p/ /ç/ /x/

alveolar palato-alveolar (or 'post-alveolar') alveolo-palatal (or 'alveopalatal', 'pre-palatal') palatal dorsal

I will refer to the consonants that are not plain alveolars or plain dorsals as 'palatoids'. With respect to the anteriority of / ç / , Pullum and Laduslaw (1986: 31) state that it is "articulated further forward than [ç] (true palatal) but not as far forward as [J] (palato-alveolar)". The palato-alveolar / j / occurs in English (e.g. ship [Jip], and the alveolo-palatal fricative / p / occurs in Polish (e.g. Kasia [kapa] (a name)). The German 'Ich-laut' is / ç / (e.g. Mädchen [meitçan] 'girl'). There are a number of languages which contrast two of the three palatoid fricatives. Polish contrasts / j / and / p / , and Basque contrasts a palato-alveolar fricative and a fricative which has been described both as a retroflex and as a pre-palatal (see § 3.1 in Ch. 7 below). A number of languages (e.g. Irish, Norwegian, Komi, Margi, Paez; see Maddieson 1984) contrast / î / and / ç / . However, no language that I know of contrasts all three fricatives in the region (Maddieson 1984). The retroflex fricative [ç] should also be included among the coronal fricatives. However, it may be misleading to assign this segment a separate place in the range of fricatives in (19), since the retroflex fricative is articulatorily identical or very nearly identical to the alveolo-palatal fricative. This has been observed for the retroflex fricatives found in Indian languages; see Keating (1991: 35) and, as pointed out above, Basque χ has been described both as a prepalatal (Hualde 1991) and as a

8

Hall (1993: 101, fn. 15) notes that in German the palatal fricative does not pattern with the coronals, although previous studies - such as that by Keating - had indicated that palatals should be represented as coronals. He observes that the first linguist to argue that palatals pattern in some languages with coronals and in other languages with dorsals was Trubetzkoy (see Trubetzkoy 1939 [1969]: 126). A representation of palatals as consisting both of the I element (which is present in coronals) and the A element (which is present in dorsals) straightforwardly reflects such facts.

116

retroflex (Maddieson 1984, following his source). There are no languages that contrast retroflexes and alveolo-palatals. Hence, I propose to assign the retroflex and alveolopalatal fricatives the same phonological representation. Any phonetic differences may then be entrusted to rules of phonetic interpretation. Let us now consider the Place representation of the palatoid fricatives. Consider a language such as Polish, which has / î / and /ç/, as well as ordinary alveolare and velars. I propose to assign the following representations to these fricatives:9 (20)

I

Ν

I I A Ν

A I I Ν

A

Ν

In a language like Paez, which has /s J ç γ / , the representations in (21) might be appropriate. That is, the phonetic interpretations of the representations differs from the interpretation of the same representations in Polish: (21)

I

Ν

I I A Ν

A I I

A

/ç/

h!

In these representations the vertical line indicates a dependency relation. If a language with more than these four fricatives can be found, the additional contrast could be achieved by using a relation such as mutual dependency (cf. Ch. 1 for discussion), or by the introduction of other elements (which might then subclassify an element like I or A in the sense of (4) above). However, as far as I am aware no such extension of the model is necessary. Let us now turn to the evidence for the representation of palatoids as complex segments consisting of two Place elements. After a brief review of the relevant literature, I will turn to a discussion of the Ich-Laut allophony in German and palatalisation in Polish. Keating (1988b) argues that all true palatals are complex segments involving the articulator nodes Coronal and Dorsal (see also Keating 1991: 45). She bases her conclusion on the examination of X-ray tracings of palatal consonants like the Czech nasal / j i / . These show two constrictions: "palatals (...) do not involve a single part of

9

Another approach which may be pursued is to represent Polish /p/ as a segment with underlying secondary articulation, specifically as /sty. Polish does seem to allow a contrast between [ç] and [si], however, as in Ka[ç]a (name) vs. Äo[s)]a 'Russia'.

117

the tongue. They use the very back of the blade, and the large front part of the dorsum" (Keating 1988b: 81). She proposes the representation in (22): (22)

Place Coronal [-ant]

[+distr]

Tongue Body [-back]

[+high]

Keating concludes that palatals are "therefore complex segments, equivalent to a double articulation, as found, for example, in labial-velars" (Keating 1988b: 89) and that this representation predicts their rarity relative to other places of articulation, at least for obstruents. Gussenhoven (1991: 114) argues that the palato-alveolars (such as /$/) are also complex, on the basis of palatalisation facts from English and Dutch. We will discuss similar palatalisation processes below. Let us investigate two possible sources that could inform us about the representation of palatals and palato-alveolars. The first of these is the German Ich-Laut allophony, and the second is palatalisation in Polish. Both are complex phenomena, and the discussion here is limited to the most essential aspects. In German, [ç] and [x] have often be argued to be allophones of the same phoneme (Hall 1989, among others). The velar (or even uvulair) allophone occurs after back vowels, and [ç] occurs elsewhere. Examples that illustrate this pattern are given in (23): (23)

a. nicht echt tüchtig mächtig

[niçt] [eçt] [tyçtik] [meçtik]

'not' 'real' 'able' 'powerful'

b. Sprache kochen rauchen Spruch

[îpra:x9] [koxan] [rauxsn] [Jprux]

'language' 'to boil' 'to smoke' 'saying'

The allophony pattern plays a role in derivation and inflection: mächtig, for instance, is derived from Macht 'power', which has [x], and the plural of Spruch is Sprüche, which has [ç]. Note that both allophones are spelled the same, namely as ch.10

10

Initial eft occurs in loanwords, where it may be pronounced as [s] (as in Champignon 'mushroom', Chemise 'chemise', and Chauvinismus 'chauvinism'), as [tj] (as in Chip 'chip' and Charter 'charter'), or as [k] (as in Chrom 'chromium', Chlor 'chloride'. A number of lexical items have [ç] (regardless of the vowel that follows), such as Charts 'grace', Chorde 'chord', Chirurg 'surgeon', Chinin 'quinine' and Chemie 'chemistry'.

118

To account for the forms in (23), an underlying velar fricative might be posited which becomes fronted after a front vowel:11 (24)

V

C I [cont]

I

A

To account for a handful of items with initial [ç] (cf. note 10), either a separate rule must be assumed or, alternatively, in these forms an underlying / ç / must be postulated. In the latter approach German is simply assumed to have / ç / as a phoneme. The difference between these two options is of some relevance if we consider the output of the rule in (24). German has a palato-alveolar fricative [J], as in schief [îi:f] 'slanted', Maschine [majkns] 'machine' and falsch [falj] 'wrong'. It is not subject to alternations similar to those of [x]. The point is that the output of the rule in (24) does not merge with the palato-alveolar fricative: the two segments must therefore be represented differently. The representations in (25) seem adequate for this purpose: c ι 1 [cont] I1 I 11 A

C I 1 [cont]

M

[ç]

A I 1 I

If [ç] is assumed to be an underlying segment, the head-dependency relation that contrasts [î] and [ç] is underlyingly present. If all instances of [ç] are derived from underlying /x/, then the head-dependency relation must be introduced when the rule in (24) applies. A second source of evidence on palato-alveolars and alveopalatals comes from Polish, which is well-known for the complexity of its palatalisation processes. In fact, Polish has a system with four of the five fricatives in (19): it has /s J ç χ/ (as well as affricates at the corresponding places of articulation for the coronal fricatives). The data in (26) show that the four are different phonemes in Polish:

11

Note that in the Dependency Phonology framework it would appear to be impossible to posit an underlying coronal fricative and assign an A element to it if it occurs after the vowels /u a o/, because the vowel /u/ does not have the A element.

119

Ν

Ν

Ν

M

ka[s]a 'cash register'

ka[s]a 'gruel'

Ka[e]a name

fla[x]a 'bottle-AUGM'

kie[s]a 'sack'

wie[j]a 'he hangs'

wie[x]a obwie[p]a 'scoundrel-GEN-SG' Svhisp-AUGM'

no[s]y 'nose-PL'

no[s]y 'stretcher-GEN-PL

no[e]i 'he carries'

Wio[x]y 'Italy'

As suggested above, these fricatives can be assigned the following representations: C I 1 [cont] I 1 I

A

c ι 1 [cont] I 1 A I 1 I

Ν

M

c I 1 [cont] I

M

Palatalisation in Polish regularly shifts / s / to / e / and / x / to /$/, as shown by the data in (28)a and (28)b, respectively. Both changes play a role in derivation and inflection, as the data also show: nió[s]í w?[s] o[s]a mu[x]a ci[x]a ia[x]a

'he carried' 'moustache' Svasp-NOM-SG' 'fly-NOM-SG' 'silent-FEM' 'sandbank-NOM-SG'

no[ç]i w§[ç]ik o[ç]ie mu[j]e ci[í]a ia[î]e

'he carries' 'idem-DlM' 'idem-DAT-SG' 'idem-GEN-SG' 'silence-NOM-SG' 'idem-DAT-SG'

It is interesting that Polish does not shift the fricatives to the 'next' place of articulation, that is, / s / is not shifted one place to [î], but two places to [p]. Similarly, [x] does not shift forward one place to [ç], but two places to [j]. 12 If the representations in (27) are adopted, both processes have the same effect: they create a complex,

12

Polish also has other palatalisation phenomena, which are often morphologised. It is beyond the scope of this study to deal with the complexities of these effects here, however (see e.g. Rubach 1984 for a description in a Lexical Phonology framework).

120 palatoid segment in which the original element (I in the case of / s / , A in the case of / x / ) becomes the dependent in a combination of I and A. With respect to the shift of the velars, the introduction of the I element may be ascribed to the influence of a front vowel or / j / , which forms the typical environment of palatalisation. With respect to the shift of the alveolars, the source of the A element that turns these into palatals is not so clear. For this shift, the environment is also typically formed by front vowels (or /j/), which do not have an A element. There are a number of possible approaches to this problem. In the first place, it might be argued that all vowels are dorsals, since all are articulated with the tongue body. In Sagey's (1986a) framework all vowels are in fact represented with a Dorsal node, which dominates a number of traditional vocalic features such as [high], [low] and [back]. Following this idea, it might be possible to argue that all vowels have or acquire a Dorsal element, which may be equivalent to the A element of Dependency Phonology. This element could then play a role in the phonology. A second approach to the palatalisation of alveolars could be to regard this as a dissimilation effect, rather than as a type of assimilation: since both alveolars and front vowels have an I element in common, effects are expected that make them dissimilar. The introduction of the A element on the consonant is one way of achieving this. Finally, the palatalisation of alveolars could be achieved in two stages: first a consonant with secondary articulation could be created (i.e. [sj] in the case at hand), which would be phonetically interpreted as a palatoid. However, Polish does not automatically interpret [sJ] as a palatoid, see note 9. In the latter two cases, no new elements need to be introduced. However, they have the disadvantage that they either do not allow palatalisation to be described as a spreading process, although this phenomenon is traditionally regarded as a type of assimilation, or need to rely on a hypothetical intermediate stage. Above we have discussed various places of articulation in the coronal region. It was proposed that it was not necessary in any language to make use of elements other than the familiar {I A U } set. A final phonetic parameter that needs to be discussed is the distinction between apical and laminai articulations. These terms refer to different manners of articulation. Apical sounds are produced with the tongue tip and laminai articulations are made with the tongue blade. It appears to be the case that no language contrasts apical and laminai manners of articulations at the same place of articulation (see Chomsky and Halle 1968: 313, who follow Ladefoged 1968). Although one language may have a laminai articulation at the front of the teeth-ridge and another language an apical articulation, no language contrasts apical and laminai articulations at the same site. Chomsky and Halle conclude that it is not necessary to postulate an additional feature to describe the contrast - even though it is a systematic phonetic one - and suggest leaving the precise interpretation to "low-level phonetic rules that assign numerical values to the different features" (Chomsky and Halle 1968: 313). This suggestion will be adopted here. To conclude, I have discussed various places of articulation that could all be classified as coronal. These range from dental and alveolar to retroflex, alveopalatal,

121 palato-alveolar and palatal. It would seem possible to describe all places of articulation by way of the elements I and A, if it is correct that in specific languages only three of the latter four places of articulation may occur. If this is the case, no new elements have to be added to the {I A U} set, nor is it necessary to introduce a subcategorisation of such elements on the existing ones (as in (4)b above). I end this section with a table of the different interpretations proposed so far in this chapter. (29)

A

I

I A

I

pharyngeals e.g. /h/

e.g.

A

U A

I

palatals /$/

labial-velars m i

A

I

U

I ι

I ι

I ι

palatals /ÇÇ/

dentals /t/

labial-coronals /φ/

A

I

U

I υ

I υ

I υ

labial-coronals ftp/

bilabials M

labial-velars e.g. / I Φ /

3. Uvulars In this section I discuss the representation of the uvular place of articulation. There is a gap in the phonological system, noted by Gussenhoven (1991), among others, that uvulars might fill. Gussenhoven notes that while labialised labials and palatalised coronals are attested as independent phonemes, although rarely so, velarised velars have not been reported at all. I propose that uvulars fill this gap. In chapter 10 I will argue that segments with secondary articulation are represented as two-root segments (see Jacobs and van de Weijer 1992 for a proposal to represent secondary palatalisation in this way). Specifically, / q / is proposed to have the structure in (30):

122 (30)

χ / \ C I [stop] I A

V I A

/q/ If uvulare have the structure in (30), a number of predictions are made. First, it is expected that uvulare form a natural class with velars (and with pharyngeals), since these have a feature in common, namely the element A. This predicts that velars and uvulare may be subject to dissimilation, or that other constraints may apply to them. Second, it is expected that velars may shift to uvulare in specific vocalic environments, particularly in the environment of low or lower-mid vowels. These two predictions will be discussed in two separate subsections.

3.1 Velars and uvulars form a natural class The evidence that velars and uvulars form a natural class is overwhelming. Some of it is presented in Elorrieta (1991), on which this section draws.13 Keating (1988a) mentions that in !Xóó velars and uvulars pattern with back vowels in phonological rules. It is well known that some of the Arabic morpheme structure constraints refer to place of articulation, in which velars and uvulars both count as back consonants (Greenberg 1950). These cases show that velars and uvulars form a natural class, by virtue of the fact that they have a Place property in common in their underlying representation. The same morpheme structure condition holds in Eskimo, which does not allow velars and uvulars in the same word (Dresher 1991). In Bella Coola (Newman 1971), only velars and uvulars are affected by two phonological process accompanying the formation of the diminutive and continuative. These are spirantisation and labialisation, exemplified in (31) and (32), respectively: (31)

Spirantisation: /k k' kw k,w q q w / — /x x' xw x'w χ χ 7 /q w als/ 'fir tree needles' /qwa:x™ls-i/ 'fir tree needles-DIM' /nik'-/ 'cut' /nixnik'/

13

'cut-CONT'

I would like to thank Jabier Elorrieta for making his unpublished work available to me.

123 (32)

Labialisation: / k /4-uq'ait/ /4-u4-q'wa:4-t-i/ /tuk'-/ /tutk' w /

k' q'/ - /k w k'w q' w / 'cedarbark' 'cedarbark-DIM' 'stretch' 'stretch-CONT'

These consonant changes involve the class of velars and uvulars only. These segments can straightforwardly be referred to as a natural class if the representation of uvulars as in (30) is adopted. In Coast Tsimshian (Dunn and Hays 1983), finally, it can be shown that all palatal, velar and uvular spirants / ç χ xw χ / have been lost in word-final position. All these segments, including the palatals, have an element A in their underlying representation. To conclude, there is good evidence that velars and uvulars form a natural class. In the framework developed in this study this is expressed in the representation of velars and uvulars: both classes of segments have the element A under a consonantal root node.

3.2 Uvulars have secondary velarisation The representation of uvulars as in (30) above, with a vocalic root node dominating A, makes a second prediction, concerning alternations of velars with uvulars in specific vocalic contexts, and concerning the effect of underlying uvulars on surrounding vowels. Velars can be expected to change to uvulars in low- or mid-vowel contexts, while uvulars can be expected to have a lowering effect on vowels. These predictions are also borne out by the facts. It is important to note that the effect is of lowering, not of backing, that is, it seems that an element A is involved here, not a feature (or element) [back]. The vowel [u], for instance, never has a uvularising effect on velars in the languages known to have such an allophony rule. Let us inspect some of the evidence, again mainly from Elorietta's work. In Chemehuevi (Press 1979), a Shoshonean (Uto-Aztecan) language, there is a rule creating an allophone [q] from an underlying / k / , when preceded by one of the vowels / o a / . The vowel system of this language is /i i u o a / . In a binary feature system, the velar to uvular consonant shift could be analysed as the addition of a feature [+low], if / o / is also specified as such. In the DP framework, the change can only be expressed as the addition of an A element. Similarly, in Siuslaw (Hymes 1966), a Penutian language, / k / surfaces as [q] when preceding / a / . Here the vowel system consists of /i a a u/. This latter rule can be formalised as follows:

124 (33)

X

X

ι

C

"V

A

A

—

c A

Α

X

X

ι

V

V A

In the Papuan language Boaki (Voorhoeve 1970), the uvular segments /q Nq R / only occur preceding or following / a of, where they contrast with non-backed velars. In the related language Zimakami, the opposition is also an allophonic one. These languages have a richer vowel system, also including / e e/. Still, the change of / k / to [q] does not take place preceding or following the mid vowels, which might be expected since these vowels also have A in their representation. Trigo (1990) mentions Turkana and Akha as two other languages in which uvulars are created as allophones of velars next to the vowels / a o o/, but not next to front mid vowels. In none of the languages examined does the allophonic change take place next to / u / . There is only one report of this change taking place, for the Arizonan language Yaqui (Crumrine 1961), which is claimed to have uvulars as allophones of velars next to all back vowels. However, Elorietta (1991: 15) points out that none of the examples provided involve the vowel /u/. Wetzels (1989) observes a similar phenomenon in Yakut (Böhtlingk 1851), a northern Turkic language spoken in eastern Siberia: "in Yakut velar and uvular obstruents are in complementary distribution, although the complementary sets of contexts are not identical for voiced and voiceless back obstruents: q occurs before / a o/, after low vowels and diphthongs which are low at their right side: / a o e ö ia uo ie üö/, and after q. Elsewhere k is found. Voiced G is obligatory before all low vowels / a o e 0/ and after /a o e ö ia uo ie üö/. Obviously, the surface shape of the back consonant is determined by the quality of the surrounding vowel(s)" (Wetzels 1989: 174). As pointed out above, it is expected that uvulars can spread their vocalic A articulation to neighbouring vowels. Indeed, this is the effect these segments have in Spokane Salish (Carlson 1967), where / u / is realised as [o], and / e / as [a], next to uvulars.14,15

14

15

A rare case in which both uvulars and velars are reported to have this effect is Bolivian Quechea (see Hart-González 1982: 477): of the velar segments in this language only /y/ is claimed to have the lowering effect, however (Herrero and Sánchez de Lozada 1978: 10). A case like this would support the identification of A with 'low* or 'lower-mid' in vowels and Velar' in consonants (see Ch. 2). Relevant are also the phenomena of 'flattening', which has been reported for Eskimo languages (Cook 1983, 1987), and 'emphasis', which has been the traditional name for the same phenomenon in Semitic languages. Both should probably be analysed as velarisation or pharyngealisation, both involving the element A in the present framework. This flattening or emphasis can also extend to vowels. Such pharyngealised vowels are discussed in chapter 10.

125 With respect to the representation of uvulars, a problem remains. Uvulars can have a secondary articulation, such as labialisation or palatalisation, and the representation of /q w / might be the following: (34)

χ

[stop] I A

A

U

However, the structure in (34) involves a 'Mismatch violation', since the weak node (here the vocalic root node) branches. One possible way out of this problem is to argue that these uvulars are in fact pharyngeals, i.e. have two A elements under the consonantal root, and no A element under the vocalic root.

4. Conclusion In this chapter I have examined possible interpretations of combinations of Place elements in a single (consonantal) segment. A good match seems available between the possibilities generated by the model and possible interpretations in terms of place of articulation. Where such a match does not seem straightforward, as in the case of coronals, a number of available strategies was examined. It was concluded that it might not be necessary to include extra consonantal Place elements to describe dental, palatal or retroflex consonants. Uvulars were argued to involve a secondary articulation type, rather than a combination of two elements under the same Manner feature. This accounts for their frequent interaction with vowels: first, velars often become uvulars in specific vocalic contexts and, second, uvulars often cause vowels to lower.

Chapter 7 Affricates In this chapter I present a further analysis of the structure of affricates. In chapter 3 some arguments were given for the following combination of Manner features in affricates: (1)

C / \

[stop] [cont] The Manner features [stop] and [cont] appear on different autosegmental tiers, and are therefore unordered in underlying representation. We return to the evidence for the hypothesis that the Manner features in affricates are unordered in this chapter more fully. The structure in (1) does not include a Place specification. Let us examine some ways in which Place might be specified in affricates. In Sagey (1986a), Place primitives are attached directly to the root node, via a Supralaryngeal node. That is, there is no geometrical relation between Place and Manner features. If we leave out of consideration the Supralaryngeal node (see Ch. 3 above), this means that Sagey represents affricates as follows: (2)

C [+cont] Place

This structure expresses the fact that affricates have only a single place of articulation. In other words, it accounts for the fact that sequences like [px] have never been proven to be monosegmental in any language. However, there is no reason why the following representation should not also be possible in Sagey's framework: (3)

r

.

[-cont]

.

[+cont]

Place / \ Labial Dorsal It is not clear how such a representation should be interpreted. In a framework in which the relation between Manner and Place features is more explicit, such a problem does not occur. The representation in (3) does not observe the dependence of Place features on Manner features that I have assumed. The proposal that we derived from the theory

128 developed in chapter 5, and that we will consider in more detail in this chapter, is given in (4): (4)

C / \ [stop] [cont] I Place

This structure will be examined in the light of the complex facts of Basque, a language which has played an important role in the discussion of the structure of affricates. It will be shown that the representation in (4) has important theoretical and empirical advantages over other proposals made in the literature. Other languages suggest the same, as we will see. Specifically, this chapter is organised as follows: in § 1 I will examine the generalisation that Manner does not spread independently of Place in more detail, and conclude that it is well-founded. I will then examine in what ways the constraints that this generalisation imposes can be incorporated in the structure of affricates. In § 2 I will examine evidence from a number of languages that suggests that the representation in (4) is appropriate. Here we will present considerations lending support to the idea that affricates are single segments, beyond the familiar ones that affricates violate sonority sequencing (§ 2.1), facts from segment inventories that show that affricates and fricatives typically occur at the same places of articulation (§ 2.2), and I present data from Luiseño, which provides support for the hypothesis that the Manner features in affricates are not ordered (§ 2.3). In § 3 I turn to facts of Basque relating to affricate structure. This will include theoretical discussion (§ 3.1) as well as an investigation of processes and constraints in Basque that bear on the structure of affricates (§ 3.2).

1. Spreading [acont] and the Manner-Place dependency in affricates In Wetzels (1991: 89), the observation is made that rules of the type given in (5), where the only spreading feature would be [ + cont] or [-cont], are conspicuously absent from the literature. (5)

*apsa -*• afsa *apsa apta

(spreading of [ + cont] from s to p) (spreading of [ - cont] from ρ to s)

Wetzels proposes to capture this observation in the feature geometry, so as to make the phonological feature model more restrictive, and to achieve this he proposes that Place features are directly associated to the feature [± continuant] (see Wetzels 1989: 176, fn. 2, 1991: 82, 1993a, and Jacobs 1989: 59ff. for brief discussion of the same

129 idea). The model proposed by Wetzels is given in (6) below. It in turn modifies ideas by Clements (1985, 1987), in which Place and Manner features are both dominated by a third node, namely Supralaryngeal in Clements (1985) and Oral Cavity in Clements (1987). The relation posited by Wetzels is a more restrictive variation on the same theme. (6) I [±cont] Place

In this way the Place and Manner features may function as a unit, because although Place can still spread alone, the feature [±cont] can only spread if Place also spreads. It predicts that rules of the type given in (5) above, which would spread only [+cont] or [-cont], without affecting Place, are impossible. Wetzels (in preparation) examines apparent counterexamples to this claim, and shows that these do not constitute actual counterevidence, i.e. that these cases can be re-analysed. If this is correct, then it seems best to capture this directly in the feature hierarchy, which is best suited to express facts about possible and impossible assimilation types. The same dependency of Place on Manner appears in independent work by Selkirk (1988a, 1990) (see Padgett 1991: 306ff. for discussion) and the same concept is worked out in a slightly different framework in van der Hulst (1995) and in Steriade (1993), who assumes that [ ± continuant] is a major class feature that is present in the root node. Various aspects of this proposal must be carefully considered, however. In § 1.1 I discuss possible counterexamples to Wetzels' claim, and in § 1.2 I turn to the question of how Wetzels' generalisation relates to the structure of affricates.

1.1 Empirical issues In this subsection I will discuss a number of phonological rules in which [cont] spreading appears to be involved, and show that a more insightful analysis is available in all cases. First we will look briefly at lenition, and then at a historical process that left alternations among Germanic languages. Consider first the phenomenon of lenition. This has been used as an umbrella term for a number of processes in which intervocalic consonants change into other consonants or are deleted. For instance, intervocalic geminates may degeminate, voiceless stops may become voiced, and stops may change into fricatives (see e.g. Lass 1984: 177ff.). The latter kind of lenition process is at stake here. A lenition phenomenon that gave rise to a correspondence between medial fricatives in English and

130 stops in languages such as Dutch can be illustrated on the basis of words like the following:1 (7)

weather feather heather leather father

cf. Dutch

weder veder heide leder vader

The original stop corresponds to a fricative in intervocalic position. Let us first discuss the disadvantages of an analysis in which [+cont] can be spread independently of Place. After that, I will give a number of possible re-analyses which are not subject to these disadvantages. The first analysis to consider is one in which the feature value [ + cont] is spread from the vowel to the stop in a framework where Manner features are independent of Place features: (8)

Place I ν

Place I c

Place I ν

[ + cont]

[-cont]

[ + cont]

The [ + cont] could be spread from either the vowel that precedes or from the vowel that follows the stop. In a framework such as Wetzels', where Place features are dominated by [acont], spreading of [ + cont] from the vowel would incorrectly result in delinking of the Place features of the stop: (9)

V

C

V

[ + cont] I Place

[-cont] I Place

[ + cont] I Place

There are two points to make about this apparently incorrect prediction: first, the representation of vowels in (9) is not likely to be correct. Vowels are redundantly [+cont] (if the definition of continuancy applies to vowels at all, since the friction in

1

Thanks to Carlos Gussenhoven for pointing out to me that some of the Dutch words in (7) had a fricative at some point in their history. It is also noteworthy that the lenition demands a following [r] in the word (cf. Dobson 1957: 956), which also points to a positional, rather than assimilatory analysis of the phenomenon.

131 fricatives is totally unlike the continuancy in vowels) and it is expected that features that are introduced later should not dominate features that are distinctively present. 2 The second point to make is that if a language has a lenition rule, i.e. shows the transition between stop and fricative in a particular environment, the distribution of the two in that position becomes predictable. In that case the continuancy specification of an obstruent in intervocalic position could be left unspecified. This specification could then be derived from a filling-in rule such as the following: (10)

[Ucont] — [ + cont] / V _ V

In such an approach, there is no violation of the Manner-Place dependency as proposed by Wetzels. Another approach that avoids spreading of [ + cont] by focusing on the position of weakened segments in lenition is that advanced by Harris (1990). Harris points out that consonants in final and intervocalic position are especially prone to lenition, while consonants in the onset are not. There is no reason to expect such an asymmetry if lenition is mere spreading of [ + cont] from a neighbouring vowel. In other words, Harris proposes to depart from the original idea that lenition should be regarded as an assimilation process. An additional argument not to regard lenition as an assimilation process is that degemination is often part of lenition processes. Obviously, degemination cannot in any way be described as assimilation. In Harris' Government Phonology framework, lenition can be expressed as a loss of complexity, fricatives being less complex than stops. In the framework developed here, lenition cannot be described as simplification, since stops and fricatives both have a Manner feature, but the point about the positional character of the process is well taken. I will therefore adopt this aspect of Harris' analysis. Let us now turn to another case study that could challenge the dependency of Place on Manner. 3 Consider the following triplets, where the English cognates of the Dutch items all have /sk/ and the German cognates all have [î]: (11)

Dutch [sx] schaal school schaden schedel

English [sk] scale school scathe skull

German [j] Schale Schule Schaden Schädel

Present-day Dutch /sx/ derives from historical /sk/. There are two further facts which are important to note: German has [i] before all stops (Stein [itain] 'stone', Spiel

This indicates that lenition processes are typically late phonological or phonetic processes. An analysis of sprirantisation as a phonetic, rather than a phonological, process is offered by Mascará (1991). The following discussion draws on Wetzels (in preparation).

132

[Spil] 'game'), and Dutch has spirantised only the velar (cf. stelen [stella] 'to steal', spelen [speils] 'to play'). The Dutch facts might be analysed, at first glance, as a case of spreading [ + cont] from the / s / to the stop (regardless of whether these / s / plus stop clusters are single segments or clusters). An analysis along these lines has been offered by Goman (1981), cited in Churma (1990). However, this analysis does not explain why /sp/ and / s t / clusters have not been subject to a similar change, and neither does it account for the German data with [j]. An alternative analysis (which does observe the Manner-Place dependency) is available, however. This analysis has the advantage that it also derives the German facts unproblematically. The first stage of the development from /sk/ to [sx] consisted of the change from /sk/ to [Sk], i.e. palatalisation of the coronal fricative before the velar. This rule did not apply in the dialects from which the English words were borrowed (so that these remained [sk]). The rule can either be formalised as spreading of the feature Dorsal (i.e. the element A) from the stop to the continuant ((12)a), or as the introduction of a Dorsal element on the stop ((12)b): (12) a.

C

C

b.

C

I I [cont] [stop] r ^ . ι I ^A

I [cont] r-^ I A

C I [stop] ι A

The reason to assume that (12)b may be the correct approach rather than (12)a (which might a priori seem more straightforward) is that there were also dialects in which the [s] was palatalised before all stops. In these dialects a non-spreading palatalisation rule must therefore be assumed (equivalent to (12)b, without the Place specification on the stop). Alternatively, it might be assumed that the rule in (12)a was the original rule, which later became generalised). Such dialects ultimately gave rise to languages like German, which has [$]piel 'game', [i]fufc 'piece', etc., i.e. palatals before all stops. However, for Dutch it must be assumed that palatalisation only took place before the dorsal, so that we can either assume spreading or introduction of A, as in (12) above. The second stage in the development involved spreading of [cont], i.e. [Jk] was turned into [îç]. This step is shown in (13) (where the consonantal roots have their own timing slot, to show that a 'serial' interpretation is forced): (13)

C

C

[cont]

[stop]

I

A

C

C

[cont] I

A

133 The resulting complex segment could be simplified in two ways: either by fusion, which resulted in the German cognate [J] (irrespective of the place of articulation of the stop that followed; see (14)a below), or by splitting up the complex matrix into two continuants. The latter possibility is given in (14)b, and reflects the course of events for Dutch: (14) a.

geminate simplification: German c

\X

c

c I [cont] / \ I A

[cont] / \ I A b.

cluster simplification: Dutch c

V

c

[cont] I

A

c -*

c

ι

ι

[cont] 1 I

[cont] ,

1 A

Thus at all stages of the development from /sk/ to /sx/ the Manner-Place dependency was observed. Spreading of [cont] is dependent on previous Place assimilation. Only in this way can it be explained why /sp/ and /st/ were not amenable to changes like these, and the various directions that the process took.

1.2 The Manner-Place dependency in affricates If the idea that Manner features such as [stop] and [cont] dominate Place is correct, the question should be considered how this insight should be incorporated into affricate structure, since affricates have both [stop] and [cont]. There are at least three ways, given in (15) below.4 (15) a. C / \ [stop] [cont] Place

4

b.

c.

C / \

[stop] [cont] I αΡΙ

I aPl

C / \

[stop] [cont] I Place

A fourth logical possibility is to attach Place to [stop] only; this representation was interpreted as that of an /s/ plus stop cluster in sc. 3 of chapter 5 above. It will be taken up again in chapter 9.

134 The representation in (15)a seems most straightforward: since Place is dominated by both [stop] and [cont], the fact that in affricates both 'sides' are produced at the same place of articulation is expressed directly in the representation. However, we will not adopt it, partly for theoretical and partly for empirical reasons. We return to both below (§ 3.1.3). The representation in (15)b is less attractive. The stipulation that the Place specifications under [stop] and [cont] should be identical begs the question if, and if so how, a segment in which they are not identical should be ruled out. Finally, the representation in (15)c is the one adopted here. To rule out the representations in (15)a and (15)b, we will assume that the Manner feature that bears Place is the head of the segment (see § 3 of Ch. 5). If it is assumed that a single segment must have one and only one head, the representations in (15)a and (15)b are ill-formed. Furthermore, the facts of Basque (see § 3.2 below) will be seen only to be compatible with the representation in (15)c.

2. Evidence In this section I present some facts that shed light on the internal structure of affricates. Specifically, I will examine some facts that show that affricates are single segments (§ 2.1), and also adduce some evidence concerning segment inventories that bear on their representation (§ 2.2), specifically, the dependency of Place on [cont]. Finally, facts from Luiseño bear on the ordering of [stop] and [cont] in affricates (§ 2.3). These topics will also be discussed with specific reference to Basque in § 3.

2.1 Affricates as single segments In this section I provide some cases that bring out well the single segment status of affricates. These are taken from Coptic, Ewe, Upper Chehalis, and English. I will not adduce evidence from syllable structure such as that which will be discussed for Basque (§ 3.1.1), but rather concentrate on other types of evidence that might qualify as showing the 'psychological reality' of affricates as single segments. Consider, first, the selection of the definite article in Coptic. Coptic is a late form of Egyptian written in the Greek alphabet with the addition of certain letters adopted from Demotic. It became extinct as a spoken language probably in the seventeenth century. Sahidic was the main literary dialect of Upper Egypt, probably originally the dialect of Thebes (Greenberg 1962: 22). Hintze (1950: 22f.) provides an interesting piece of evidence for the single-segment status of affricates in this dialect, concerning the selection of the article: Im sahidischen Dialekt des Koptischen (...) hat der Artikel zwei Formen: p(masc.), t- (fem.) vor einfachem Konsonanten und Vokal; pe-, te- vor Doppelkonsonanz (vgl. psëre 'der Sohn', pto 'die Erde', tsöne 'die Schwester', gegenüber

135 pesklm 'das Greisenhaar', tehbür 'die Linke', pethab 'der Sauerteig'). Vor c hat der Artikel aber die Form p-, t-, und nicht etwa pe-, te- (vgl. pco 'die Achsel', tcamë 'die Ruhe' usw.). Dies zeigt deutlich, dass c hier in der Sprache als Einzelphonem behandelt wird. In the Coptic case affricates act on a par with single segments for the selection of the article. In Ewe, there are a number of reduplication processes in which only a single consonant of an initial cluster reduplicates (Ansre 1963; Sagey 1986a: 86). The reduplication processes have a variety of semantic functions. Ordinary clusters are simplified, as shown in (16)a below. However, clusters with an initial affricate reduplicate the affricate, and if there is only an initial affricate, this reduplicates as a whole. Both facts indicate that affricates are treated as single segments (16)b:5 (16)

a.

jira - jiajiralá fle - fefle + é b. dzrá - dzádzrá tsi - tsitsi + í tso - tsotso + ede

'to rave - raver' 'to buy - bought' 'sell - selling' 'to grow - grown' 'to be fast - fast'

The same kind of evidence is adduced by Kinkade (1963: 187ff.) for the Salishan language Upper Chehalis, where /ts tj t i / are treated as single segments in reduplication. Finally, Fromkin (1971: 33) notes that in a large corpus of English speech errors there is "not a single example [that] shows a splitting of [tj] or [d3] into sequences of stop plus fricative". She concludes this is supporting evidence for an analysis of affricates as single segments.

22 Segment inventories In this section I would like to investigate the place that affricates occupy in segment inventories. There are two aspects that will be considered in some detail: first, in § 2.2.1 I examine the relation between secondary articulation on affricates and secondary articulation on other segments, and in § 2.2.2 I examine the typical places of articulation of affricates, especially in comparison to the place of articulation of other segments in particular inventories.

5

In this reduplication process, /kfT/ is also treated as a single segment, as forms like Ifpla - kpalipia, 'lead - leading' show.

136

2.2.1 Secondary articulation It is often the case that particular types of secondary articulation appear on specific señes of segments in a language. For instance, palatalisation or labialisation as types of secondary articulation usually do not appear on only one stop or fricative, but are typically employed to make a contrast on an entire series of stops or fricatives. 6 If a type of secondary articulation appears on a set of segments rather than a single one, we would be missing a generalisation if the set of segments was stated as a list. Rather, such a generalisation calls for an explanation in terms of a rule. Such rules may then be used to determine the feature composition of affricates, which are expected to pattern with certain sets of other segments. The term 'secondary articulation' here must be understood in a wider sense than just referring to the imposition of a vowel-like articulation on a consonant. Laryngeal secondary articulations, such as aspiration or laryngealisation, are also included, as well as ejectivity and length contrasts. Consider, for instance, the partial inventory of Chontal (language 710 in Maddieson 1984): (17)

voiceless plosive voiced plosive vl. ejective stop vl. sibilant arícate vl. sib. eject, affric.

p b p'

i

k ji t'

k' ts ts'

tj tj'

The feature of ejectivity can occur contrastively on all voiceless stops, and also on the voiceless affricates. We can express this regularity as an underlying relation between the feature [stop] and the feature [ejective] (or some equivalent thereof). If a segment is [stop], it may bear the feature [ejective] contrastively. I will call this a 'possibility rule', and to avoid any confusion with other types of rules use a '=? notation: (18)

Chontal possibility rule: [stop] =t [ejective]

Since affricates have the feature [stop] in underlying representation, they share with stops the possibility of bearing [ejective] contrastively. The same situation (with regard to ejectivity) obtains in a number of related and unrelated languages, such as Zulu (126), Tigre (251), Amharic (252), Kullo (262), Kefa (264), Tzeltal (712) and K'ekchi (714). With respect to aspiration, compare Korean (070), which displays a three-way laryngeal contrast:

6

The material dealt with here is discussed in more detail in Hinskens and van de Weijer (in preparation).

137 (19)

voiceless plosive. vi. aspirated plosive vl. laryngealized plosive vl. sibilant affricate vl. aspirated sib. affric. vl. laryngealized sib. aff.

Ρ ph Ρ

k kh k

t th I tj t$h W

Other examples include Beembe (123), which has an aspiration contrast on voiceless plosives, and also on the two affricates in the language, /pf/ and /ts/. Compare also Mundari (300) and Kharia (301): in both languages voiceless plosives can be contrastively aspirated (and the voiceless palato-alveolar affricates can too), and voiced plosives can be contrastively produced with breathy voice (and the voiced affricate can too). Simple cases include also Sui (403), Saek (404) and other languages in this group. Javanese (409) can contrastively release its plosives breathily, and the same goes for its affricate. Changchow (503), like Thai, has a three-way contrast (voiceless, voiceless aspirated, and with breathy release), realised on plosives as well as affricates. Ojibwa (750) uses contrastive pre-aspiration on its plosives and affricate (cf. also Guajiro (826)). Kan (506) uses the contrastive aspiration on stops and both its dental/alveolar and palatal affricates. Chipewyan (703) has ejectivity and aspiration as possible secondary articulations on its dental, dental/alveolar, palatoalveolar, and lateral affricates. Some of the relevant possibility rules are stated in

(20): (20)

Korean: Japanese: Mundari:

[stop] [stop] [stop] [stop, [stop,

=t [spread glottis] =t [laryngealisation] =t [breathy voice] -voice] =t [spread glottis] + voice] =t [breathy voice]

Stops may also have a contrastive length contrast, which then also may be employed for making a distinction between long and short affricates: Yakut (061) is of this type: (21)

voiceless plosive long voiceless plosive vl. sibilant affricate long vl. sib. affric.

ρ p:

t t: tí tj:

k k:

Other examples here are Japanese (071) and Delaware (751). Capturing this in the possibility format is less straightforward since length is not a feature (following Clements and Keyser 1983). The statement in (22) is intended as a suggestion:

138 (22)

Yakut possibility relation χ [stop] =t

χ \ /

Finally, systems are found where stops as well as affricates are prenasalised. Such systems abound in Maddieson (1984), e.g. Luvale (125), Washkuk (602), Mazatec (727), and Siriono (829). It may roughly be expressed as in (23): (23)

[stop] =t [stop, nasal]

Above, only cases were presented in which the secondary articulation on stops also appeared on affricates. This accords well with a suggestion made by Steriade (1989) that affricates 'are' stops at particular places of articulation. For instance, a labial affricate / p f / would be a labiodental stop underlyingly. In an approach such as ours, where affricates are [stop] and [cont] from the underlying representation onwards, and which incorporates possibility rules, we would expect to find rules that refer to the [stop] value (such as those above), as well as to the [cont] value. Do regularities concerning secondary articulation on fricatives and affricates occur in a corpus such as Maddieson (1984)? First, one should note in this respect that secondary articulation on fricatives is rarer in any event than secondary articulation on stops. Second, there are a few cases in Maddieson (1984) in which a particular type of secondary articulation that occurs on fricatives also occurs on affricates, but not on the stops in the language, which therefore provide support for the assumption that fricatives and affricates have a feature [cont] in common. These cases include Kashmiri (018), which has / j j / as well as /tf à j / but no secondary palatalisation on stops, Akan (115), which has /ç w / and also /cç w jj w / but no secondary labialisation on stops, and Gä (117), which has /s w / and distinctively labialised affricates, but no distinctively labialised stops. To conclude, it is often the case that affricates have a type of secondary articulation in common with stops (although there are counterexamples to this generalisation: see Hinskens and van de Weijer in preparation). However, among the small number of cases where fricatives have secondary articulation, examples can be found where the secondary articulation on fricatives matches that on affricates. This lends plausibility to the claim that affricates form natural classes with both categories, and therefore share formal characteristics with both.

2.2.2 Place of articulation in fricatives and affricates In § 3 of this chapter we will see that Basque has three different places of articulation of coronal fricatives, and the same three places of articulation in affricates. In

139 Maddieson (1984) more systems of the same type can be found. For instance, the situation is exactly parallel in Güniina-Kena (821), which has dental, palato-alveolar and palatal sibilants, and also three affricates at the same places. Similar situations also obtain in Burushashki (915), which has dental/alveolar, palato-alveolar and palatal sibilants and affricates. In Komi (052), dental/alveolar, palato-alveolar and palatal affricates correspond to fricatives at the same places of articulation. In Chipewyan (703) the situation is even more striking with dental, dental/alveolar, palato-alveolar and velar affricates occurring in the same system with fricatives at these places. For Aztecan languages, Goller, Goller and Waterhouse (1974: 126) point out that typically in these languages there is matching of "affricate stops ts, tj, with fricatives s, J". Such cross-linguistic parallels demand a structural explanation. The generalisation can be expressed by the fact that in affricates as well as in fricatives Place is attached to [cont] directly (see (24)a below). This formalises the general tendency that they occur at the same places of articulation. In stops and nasals Place is attached to the same feature [stop], which are therefore also predicted to be typically articulated at the same places of articulation ((24)b): 7 (24)

a.

C

C / \

I [cont]

[stop] [cont]

I Place

Place

affricates

fricatives b.

C I [stop]

C I [stop] [nasal] nasals

I Place

Place stops

2.3 Ordering of [stop] and [cont] in affricates Evidence from Basque and Maya (Lombardi 1990; Hualde 1991) suggests that the Manner features [stop] and [cont] are not phonologically ordered in affricates. One

7

See also Maddieson and Precoda (1992: 57) for the observation that nasals and stops are typically articulated at the same places of articulation. They regard this as a kind of phonetic efficiency. Here we see that this fact can also be expressed phonologically.

140 example is the rule of pre-stop stop deletion in Basque, which also applies to affricates (see § 3.2.1 below). This case is not unprecedented. In the Southern Californian Uto-Aztecan language Luiseño (Bright 1968; Hyde 1971; Munro and Benson 1973; Davis 1976), a very similar rule occurs which turns an affricate into a fricative word-finally as well as before obstruent stops, nasals and laterals (see also van de Weijer 1993a). The ordering of the rule has been the topic of some debate in the literature (see Anderson 1974b), but its content has received little attention. 8 Some morphologically related forms, taken from these sources, are given in (25): (25)

tjapómkat tjikwi:la ?é:tsi mótji tàçkitîil púitjil nétju

'liar' 'to be sad' 'above' 'to weave' 'shadow-OBJ' 'eye, fruit-ACC' 'to become an old woman'

tjajpumkatum tjikwijkwi ?éjkawis mójlat tàçkijla pújla néjmal

'liars' 'to suffer' 'upper lip' 'belt' 'idem-NOM' 'idem-NOM' 'old woman'

Cases that show that affricates do appear before fricatives and the glide / w / are provided in (26): (26)

?étjvaî pátjxam wátjxat wátjxavat pátjxayal qawitjxal pu-mútjvi tjátswumal

'left hand' 'to launder' 'shoe' a type of plant 'Scirpus microcarpus' 'Bloomeria crocea aurea' 'its top, summit'9 'Gilia capitata staminea'

The behaviour of the affricate before r is unclear. Davis (1976: 199) gives the one alternation in (27): (27)

tjuró?a-

'to level'

pujró?ax

'his levelling'

The form for 'his levelling' does not appear in Bright (1968). In fact, this dictionary does not contain any example of words with / r / preceded by either an affricate or a fricative. Hence, it may be the case that only obstruent stops, nasals and / act as triggers of the rule, which is the same class of segments as that in Basque Stop

8

9

Poser (1981: 4) adduces this rule to show that nasals and stops form a natural class of [-cont] segments. However, it is not completely clear that stops and nasals exclusively form the set of trigger segments, since the lateral [1] also seems to trigger it. This gloss according to Davis (1976: 199); Bright (1968: 25) gives 'person who burns clothes at mourning ceremonies'.

141

Deletion (see § 3.2.1 below). The effect of the rule is also the same: the occlusion part of an affricate is deleted before another occlusion. Sagey (1986a: 93ff.) discusses a number of examples in which affricates behave as [-cont] with respect to processes sensitive to their left edge and as [ + cont] with respect to processes sensitive to their right edge. This shows that at some point in the derivation the features making up the affricate contour must be(come) ordered. However, as pointed out by Lombardi (1990), cases such as Basque Stop Deletion (and Maya morpheme structure constraints) show that these features are not ordered in the underlying representation. This predicts that true edge effects can only occur at a late level of the derivation. Morpheme structure constraints, for instance, are predicted never to be sensitive to the internal ordering of [stop] and [cont].

3. Basque In this section I present an examination of the behaviour of affricates in Basque.10 The phonology of Basque has played an important role in the discussion of the representation of affricates. Hualde (1987, 1988a,b, 1991), Lombardi (1990), A r c a n geli and Pulleyblank (1987) discuss rules and constraints in this language which point to a representation of Basque affricates in which the Manner features corresponding to the stop part of the affricate and the continuant part are not ordered, contra Sagey (1986a), thus accounting for a range of 'anti-edge effects'. A representation with [stop] and [cont], two independent unary features, accounts for and predicts such effects. In this section I also show that the representation of affricates with Place dependent on only [cont] is more appropriate than one in which Place is attached to the root node. Three aspects of the underlying representation of Basque affricates that we will focus on specifically are given in (28): (28)

a. b. c.

affricates have a single root node and therefore behave as monopositional segments. the continuancy specifications of the affricate are expressed by two monovalent features [stop] and [cont], which are unordered phonologically. Place features in the affricate are dependent on its [cont] part.

Thus, the representation incorporates the insight from Hualde (references cited above) and Lombardi (1990) that the Manner features in an affricate are unordered phonologically. The representation attempts to motivate a relation of direct domi-

lo

This section appeared in a slightly different form as van de Weijer (1992a). See also van de Weijer (1994b).

142 nance between the feature [cont] and the Place features. 11 Note that the representation of affricates in (4) does not contain intermediate nodes such as Supralaryngeal or Place: instead, individual Place features are attached directly to the Manner feature [cont]. In the phonology of Basque, there is no evidence that nodes like Supralaryngeal or Place are needed. The rest of this section is organised as follows: in §§ 3.1.1 to 3.1.3 I will discuss the general aspects of (4) in some more detail. These form the preliminaries to § 3.2, which presents a study of the phonological structure and behaviour of affricates in Basque, based on extensive discussion in Hualde (1991: Ch. 5). Section 3.3 is a brief conclusion for Basque.

3.1 Preliminaries Basque has fricatives and affricates at the following places of articulation: apicoalveolar, predorso-alveolar and prepalatal (Hualde 1991).12 Basque only has a single coronal stop, / t / . The realisation and spelling of the fricatives and affricates is as follows: (29) apico-alveolar predorso-alveolar prepalatal

fricatives

affricates

[s] s [s] ζ [ç] χ

[ts] ts [tj] tz [tç] tx

The phonological representation of these places of articulation may be expressed in other ways. It is possible to represent them using binary features, as Hualde (1991) does, as in (30):

11

This develops and modifies a suggestion made by Hualde (1987: 88, fn. 2), who assumes that in affricates, which he represents with two root nodes attached to a single x-position, the root node which is specified for [+cont] has a Place node in underlying representation. See also Hualde (1991: 129): "If affricates are represented as having two root nodes, as proposed also in Clements [1989b], it would be the root node dominating the feature [+cont] that also dominates the place node of the segment". Below, I argue against representing affricates by way of two root nodes. Moreover, the representation of affricates in which [cont] dominates the Place features directly seems an even more direct expression of the tighter relation between the two than the two-root representation of Hualde.

12

Maddieson (1984: 419) gives the three places of articulation as dental/alveolar, palato-alveolar and retroflex, following his source. The basic points made in this section remain unaffected if the exact places of articulation should turn out to be phonetically different. Hualde (personal communication) describes the three places of articulation as apical and laminai [s], and palato-alveolar [s].

143

(30)

[coronal] [anterior] [apical]

+ + +

+ +

+

s/ts

í/tj

ç/tç

Alternatively, we can express these places of articulation in terms of unary primitives, such as I and A. In such a conception, the (apico-)alveolar affricate has the element I alone under the [cont] feature and the predorso-alveolar and prepalatal have headed combinations of the two elements, as was discussed in chapter 6. (31)

I

s/ts

I

A

I A

I I

s/ti

ç/tç

I will use the elements I and A in the representations below, but the structural properties regarding the relation between Manner and Place may be expressed in either a binary or a unary feature framework. Notice that all three places of articulation are coronal places of articulation. It is not a priori clear whether the representation of Basque affricates can be extended to affricates at other places of articulation, such as German [pT] or Swiss German [Kx~], Let us now proceed with a brief discussion of the three aspects mentioned in (28) above. These will be the topics of three subsections.

3.1.1 Affricates as single segments In (4) above affricates are represented with one (consonantal) root node, while they are also assumed to have one skeletal point, or timing slot. The fact that they have one root node reflects the insight that they behave as single segments for purposes of syllabification and timing. That affricates are single segments is not controversial. However, I would still like to examine the relation between segmental structure and timing structure in affricates in some more detail. Apart from the position advocated here, there are three other options which are logically possible: i. affricates have two timing positions and two root nodes. ii. affricates have two timing positions but one root node. iii. affricates have one timing position but two root nodes. These three alternatives, which are all well-formed phonological representations, are given in (32):

144 (32)

(i)

(iii)

(ii)

χ

χ

C

c

χ

\/ c

χ

χ

timing tier

/\ C

C

root nodes

The representation in (32)i can be immediately rejected. It is equivalent to representing affricates as two segments. Hualde (1991) shows on distributional grounds that this position cannot be maintained for Basque. If affricates are sequences of segments, they violate the sonority hierarchy whenever they appear in the coda of the syllable. However, affricates are allowed in the coda in Basque, a language which does not show any other apparent sonority sequencing violations. Basque also allows two consonants in coda clusters, the first of which must be a sonorant, and the second an obstruent. Basque also allows sonorant-affricate clusters in the coda, which again indicates that affricates behave as single obstruent segments. The representation in (32)ii is like that which has been proposed for long vowels in Clements and Keyser (1983). The sonority considerations and clustering possibilities mentioned above are not counterevidence to this representation, if both of these are defined with respect to the root node level. However, there is no evidence in Basque that affricates behave like long segments. Hence, this representation must also be rejected. This leaves us with the representation in (32)iii as the only real alternative to the representation in (4) (which had one root node and is assumed to be attached to one x-slot). A representation like this was proposed by Clements (1989b), and adopted by Hualde (1991). However, this representation must also be rejected, for two reasons, the first of which is a general theoretical reason and the second an empirical one. First, the representation with two root nodes is much less restrictive than one in which only one root node is used. There are no constraints inherent in the geometry which prohibit putting completely different segments under the two root nodes, possibly differing for voicing, place of articulation, as well as major class and manner features. Without further constraints, the phonological model would then have the freedom to recognise [ga], for instance, as a single segment. Below, I hope to show that such a non-restrictive representation is simply not necessary in the case of Basque affricates. Ceteris paribus, the representation with one root node should therefore be retained as long as the evidence permits. At the same time, segments with two root nodes are expected to exist in a model with both a timing tier and root nodes. In chapter 10 I investigate the question what kinds of segments could correspond to such a representation. Second, there is an empirical consideration. 13 This comes from templatic systems like that of Arabic, where vocalic and consonantal segments spread onto syllabic

13

I would like to thank Douglas Pulleyblank (personal communication) for bringing this point to my attention.

145

templates (McCarthy and Prince 1986, 1990). This has been expressed as left-to-right association of root nodes onto positions in the template. If affricates consist of two root nodes, they are predicted to associate to two positions on the template, rather than to one, in dialects of Arabic which have affricates; association to the template would result in a stop and a fricative, as shown in (33)a: (33) a.

*

χ ' ·"

b.

χ

χ

χ

...

/ /

c' Ι

οI

[stop]

[cont]

χ y — C A· ' • " _ —

"

~

[stop] [cont] As McCarthy and Prince point out, however, affricates link onto the template as whole segments, as in (33)b.14 This is therefore evidence that affricates in languages which have templatic morphology should be represented as having one root node. As there is no evidence that representations with two root nodes are necessary in Basque, there is no good reason to assume that the structure of Basque affricates should be different from affricates in these languages. I conclude from this that the representation in (4) with one root node and one xslot is superior to the representations in (32)i-iii on general theoretical and independent empirical grounds. I hope to show that the behaviour of affricates in Basque is also completely compatible with the one-position-and-one-root representation in (4).

3.12 Two monovalent features [stop] and [cont] Hualde (1987, 1988a,b, 1991) and Lombardi (1990) show in detail that the stop and the continuant part of affricates in various languages are phonologically unordered. Both adduce evidence from Basque, some of which will also be examined below. Lombardi concludes from this that it is misleading to represent affricates with two values of a binary feature [ ± continuant], because then the values of this feature would be on the same tier. In that case, the fact that in a variety of cases these

14

I would like to thank Samir Khalaily for bringing the following Hebrew words to my attention which are relevant here: pilsets Ί ο make something explode', nitsels 'the light shines' and katsats 'to cut', which all seem to involve roots in which affricates occupy single skeletal points.

146

features appear to be unordered phonologically becomes impossible to express. The two rival representations are given in (34): (34)

Ordered:

Unordered:

C

C

[ - c o n t ] [ + cont]

A

[stop]

\ [cont]

In the unordered representation of the Manner features in affricates the features [stop] and [cont] appear on different tiers. When in representations below these features are placed at the same height, this is done for typographical convenience.

3.1.3 The Manner-Place dependency in Basque affricates In this section I wish to motivate that in Basque affricates only one of the Manner features dominates the Place specifications. Above I pointed out that there is more than one way in which the idea that [ ± continuant] dominates Place could be extended to affricates. First, it would be possible to make the Place features doubly dependent on both [stop] and [cont], in the sense that Place is linked by an association line to both Manner features. This possibility, which we will reject, is repeated in (35): (35)

C

/\ [stop]

[cont]

V / Place First, this structure does not reflect the generalisation that in Basque (as in many other languages, as we saw in § 2.2.2) the place of articulation of the set of affricates is identical to the place of articulation of the set of fricatives (see (29) above). I express this by postulating a closer phonological relation between the feature [cont] in each of these affricates and their Place specifications. This is why I propose that in affricates Place is only associated to the feature value [cont]. Secondly, we will see below that the representation in (35) is also empirically inadequate. Briefly, a representation such as (35) predicts that when [cont] is delinked from the root node, Place will still be attached to [stop], and this is incorrect in Basque (see § 3.2.4). Second, a rule of Basque which targets coronal stops, namely Palatalisation, does not apply to coronal affricates (see § 3.2.6). That this should not happen is unexpected under the assumption that Place is doubly linked, as in (35).

147

In underlying representation the feature [stop] in affricates is not attached to Place features in (4). For phonetic interpretation, however, it would appear to be necessary to supply the Manner feature with Place features (see also van de Weijer 1993a for discussion). In principle, this could be achieved in two ways. Either the feature [stop] could get a default Place specification, uncontroversially Coronal, here expressed as the element I, or the Place primitives under [cont] could spread, segment-internally, to the feature [stop]. I have no phonological or instrumental evidence that could decide which of these two options should be preferred. If, for instance, the stop part of each of the three Basque affricates is produced at a relatively stable place of articulation, the first option might be preferred. Below we will see that the first strategy (insertion of I) is employed in Basque when the [cont] feature comes to be delinked under circumstances to be described, leaving the segment in effect without Place specifications.

3.2 Processes and constraints in Basque In this section I discuss a number of Basque processes involving affricates. The analysis presented is based on facts provided in chapter 5 of Hualde (1991). The purpose of the exercise is to show that the processes discussed there can be described using a simpler, more restrictive model of representation.

3.2.1 Stop Deletion The process of Stop Deletion removes the first of two adjacent stops in certain phonologically, morphologically, syntactically, and stylistically defined circumstances. Data are given in (36) (see Hualde 1991: 127): bat paratu bat traban bat kurri guk piztu hitz tegi hitz keta haritz ki haritz mendi hotz tu irakts-i hauts-i itj-i

ba[p]aratu ba[t]raban ba[k]urri gu[p]iztu hi[çt]egi hi[çk]eta hari[çk]i

'put one' 'one stuck' 'run one' 'we light' 'dictionary' 'conversation' 'oak wood' 'oak mountain'

hari[çm]endi ho[çt]u iraka[st]en / t s - t ç /

'become cold-PERF' 'teach'

hau[st]en / t s - t ç / ijten /IHç/

'break' 'open'

148 The result is complete deletion when two stops are adjacent ((36)a), but a fricative results when an affricate originally preceded the stop ((36)b). Finally, when the imperfective suffix -tzen is added to an affricate-final verb form, the simplifications result in the deletion of the stop part of the first affricate and the fricative part of the second ((36)c). The description of this process is entirely straightforward, viz. in a sequence of two [stop] features, the first is deleted (see Hualde 1987, 1991). First, consider the operation of the process to stop-stop sequences. Since Place is dominated directly either by the feature [stop], it is predicted that in Stop Deletion no interpretable segment will remain behind when the [stop] feature is deleted: 15 (37)

C *

C

[stop]

I [stop]

I Place

I Place

The data in (36) show that what we derive in (37) is exactly right: no clean-up operations are necessary after the delinking of the [stop] feature. The operation of the rule when the first stop is an affricate is illustrated in (38): (38)

C [stopf\ [cont] I Place

C I [stop] I I Place

e.g. [tz t] — [z t] The segment which results from the operation of the process is a well-formed fricative. Notice that in (38) again it is not necessary to assume some convention in which Place material previously under the [stop] node is pruned or stray-erased. This is a direct consequence of the fact that [stop] is not linked to any Place features in underlying representation. The righthand segment may also be another affricate. This possibility is shown in (39), where [stop] is delinked from the first segment, as expected:

15

In a Manner-feature framework like thai of Rice (1992), in which obstruent stops are represented having no Manner features at all, this process is by definition inexpressible. Any attempt formalise the rule by referring to the absence of Manner features would also predict stops to deleted when no segment follows, for instance, or would have to complicate the rule unduly provide for such complications.

as to be to

149 C

C

(39)

[stop]

[stoj [cont]

[cont]

Place

Place

e.g. [tz tz] — [z t] As shown, a sequence of two affricates is also input to Fricativisation (see § 3.2.4 below), which deletes the rightmost of two [cont] features, as the examples in (36)c demonstrated. A final interesting aspect of the rule of Stop Deletion in Basque is that nasals and laterals also trigger the rule, as in /haritz mendi/-» [hariz mendi] 'oak mountain'. That nasals and laterals form a natural class together with obstruent stops is clear in a proposal for Manner primitives such as that put forward in chapter 5: both have a [stop] feature in their representation (although this appears on different levels for both). To conclude this subsection, an underlying representation of affricates like (4), in which [stop] and [cont] appear on independent tiers, and in which Place is attached to [cont] only, can account for Stop Deletion in Basque. A representation like that in (34)a, on the other hand, with an ordered contour of Manner features, is by definition incapable of accounting for the facts in a straightforward manner.

3 22 Glottal Stop Formation and Aspiration Hualde (1991: 130ff.) discusses Glottal Stop Formation and Aspiration (which in the Romance tradition refers to the replacement of a fricative / s / by [h]) in different languages. I will also discuss the formalisation of these processes in the model outlined here, since although neither occurs in Basque, the output of these processes may be identical to the output of other processes in Basque (see § 3.2.4 below). Hualde's account is based on Lass (1976), Steriade (1987), and others. As Lass (1976) points out, voiceless stops may be reduced to glottal stops in the course of the development of a language. This may be analysed as the loss of all supralaryngeal features, retaining the laryngeal features and the feature [-continuant]. Likewise, the reduction of fricatives to [h] may be described as the retention of laryngeal features and [+continuant]. A wrinkle in such an account, especially in feature geometries with a Supralaryngeal node, a Place node, and individual Place features, is that both Glottal Stop Formation and Aspiration may be described as delinking of the Supralaryngeal node, delinking of the Place node, or delinking of individual Place features. In all cases only

150 laryngeal features and a continuancy specification will remain behind, as the reader can easily check for herself. However, it is unfortunate that three formally distinct processes should all have the same phonetic effect. Either these operations should have different outcomes, or some principled way of deciding which node is delinked should be available. It should therefore be noted that in a model where Manner features dominate Place features directly, Glottal Stop Formation and Aspiration can only be formalised in one way. These are provided in (40) and (41), respectively: (40)

Glottal Stop Formation C I [stop] * place features

(41)

Aspiration C I [cont] + place features

What is the realisation of the segments that result form the operations in (40) and (41)? When the output of the process is a glottal stop [?] as in some dialects of English (see Lass 1976), we assume that the [stop] feature is realised 'laryngeally', viz. as [ +constricted glottis]. See also Anderson and Ewen (1987: 190f.), who follow Lass (1976), for the observation that the glottal stop and voiceless stops are phonetically similar. Although the state of the glottis during the production of both kinds of segments is quite different, both types of segments involve a complete blockage of the airstream. In Anderson and Ewen's notation, these segments form a natural class because they are the only segments that have { | C | } as the representation of the phonatory subgesture. However, as Anderson and Ewen are aware, these segments do not form a natural class in the sense that glottal stop and voiceless stops often occur together as targets or triggers of phonological rules. It is therefore useful also to consider the proposal of van der Hulst (1991), in whose representational system the features [stop] and [constricted glottis] are expressed by the same element. The location of this prime is responsible for its interpretation as either oral or laryngeal constriction. Hence, cases like Glottal Stop Formation form support for such 'crossgestural relations' (see van der Hulst 1995). Finally, Humbert (1994) offers an account of processes like Aspiration (also including nasalisation after nasal loss) in terms of structured components. In her proposal stops like / k / are composed of a

151 configuration which by itself represents a glottal stop and a Place feature. If the Place feature is lost, a glottal stop is automatically derived. To conclude, the presence of Supralaryngeal and Place nodes leads to nonrestrictive hypotheses with regard to the representation of phonological segments and processes. A number of alternative approaches have been briefly discussed. To the extent that our formalisation of linguistic processes can do without such nodes, they are redundant units in phonological theory. It has been shown above that the facts of Basque can be adequately accounted for without such units.

3.2.3 Sibilant Harmony Placing Place under the [cont] feature also makes the correct predictions with respect to Sibilant Harmony in Basque. This refers to a constraint which requires that all sibilants (affricates and fricatives) in a morpheme share the same point of articulation (see Hualde 1991: 132ff.). Across morpheme boundaries, harmony is quite freely violated. Examples are given in (42), where the sibilants are italicised (it is a coincidence that in these examples there are no intervening / t / s (there are other coronals like /I r n/; according to Hualde (1991: 132), whose data are based on Salaburu (1984), sibilant harmony is exceptionless): (42)

asote eltouiUse íama/í sasoin

zuzen zezen azazkal zimi/z

urtxintx

Regardless of which primitives are used to represent Place features, Sibilant Harmony can be understood as a process that spreads all Place features from a continuant which has the element I. A possible formalisation appears in (43): (43)

In a configuration [cont] I spread all Place features within the morpheme

Because [cont] dominates Place directly in affricates, the Place element(s) spread to this feature only, so that we explain rather than stipulate why they participate in the constraint, while stops do not. The way in which Sibilant Harmony is formalised in (43) accounts for the fact that a non-coronal continuant like / f / , which also occurs in Basque, may occur with a continuant of any other place of articulation.

152 3.2.4 Fricativisation Fricativisation transforms a fricative sibilant plus affricate sequence into a fricative plus stop cluster in certain morphological contexts. Data (from Hualde 1991: 134f.) in which the process operates are given in (44)b, while (44)a shows that it does not operate after other segments: perfective ipin-i ibil-i sar-tu eska-tu ikas-i ikus-i

imperfective ipintzen ibiltzen sartzen eskatzen ikasten (*ikastzen) ikusten (*ikustzen)

'put' Svalk' 'enter' 'ask' 'learn' 'see'

In this process, the righthand feature [cont] in a sequence of two is deleted under appropriate circumstances. The effects when the rightmost [cont] feature is part of an affricate are illustrated in (45): (45)

C I [cont] I

[cont] [stop]

Place

I Place

e.g. / s tz/ — [s t] What are the effects of this process for the [stop] part of the original affricate? What we derive in (45) is a [stop] segment unspecified for Place. In all examples provided by Hualde the output is a simple coronal stop [t], regardless of the original place of articulation of the affricate. We must therefore assume either that a Manner feature unspecified for Place is interpreted as Coronal (i.e. without further rules), or that a default rule of the type in (46) assigns this feature Coronal, or element I. (46)

If [stop] is left unspecified for Place, then insert I

As we argued above that Glottal Stop Formation consists of the delinking of Place features from [stop], which results in a configuration identical to that derived in (45), the default rule in (46) is appropriate for Basque: the operation of this rule does not result in a glottal stop (note that Humbert's 1994 proposal discussed above needs a similar fill-in rule). It will be interesting to see if languages exist in which a rule like Glottal Stop Formation applies next to rules like Fricativisation.

153

When two affricates are concatenated, the prediction is that the second one will lose its [cont] (and therefore also its Place) specification, because the two features [cont] are adjacent on their autosegmental tier. That this is in fact what happens has been shown above (see (44)-(45)). Finally, note that a model in which Place is 'doubly dependent' on [cont] and [stop] (see (35) above) results in an account which needs patching up. Compare the output of (45) to that of (47): (47)

C

C I [cont] I Place

[stop] [cont]

V Place

Here a [stop] segment is derived which is specified for the place of articulation of the original affricate. As far as the alveolar affricate is concerned, this is not a problem, of course, but in the case of the other two affricates we would appear to derive palatal stops. To arrive at the resulting simple coronal stop [t], we would then have to rely on, for instance, Structure Preservation (Kiparsky 1985) (see the next section). However, such an account is not available, since Basque dialects usually do have palatal stops as underlying segments (Hualde 1991: 10), and they are also created in another phonological process, namely Palatalisation (§ 3.2.6). It would therefore seem that Fricativisation provides crucial support for a representation of affricates in which Place is only dependent on [cont], rather than both Manner features.

3.2.5 Voice Assimilation In most Basque dialects there is a process which voices stops after nasals or laterals (Hualde 1991: 135ff.). In present-day Basque the process is morphologised, and affects only inflectional suffixes. Hualde gives the following examples for the perfective suffix /-tu/: (48)

perfective

/-tu/

afal-du lan-du ken-du

'have dinner' 'labour' 'take away'

neka-tu har-tu

'get tired' 'take'

cf.:

154 The process voices stops after the class of nasals and laterals. However, affricates are not voiced by this process, as can be seen from the following examples, provided by Hualde: (49)

imperfective afal-tzen [-ts-] lan-tzen [-ts-] ken-tzen [-ts-] neka-tzen har-tzen

'have dinner' 'labour' 'take away' 'get tired' 'take'

Why should voice assimilation fail to apply to affricates, that is, why should it apply to representations as in (50) and not to representations as in (51)? (50)

C

C

[+vce]

[stop]

[-vce]

I Place (51)

*

C

[+vce]

C [stop]

\ [cont] I Place

There is nothing in the structure of affricates that will forbid application of the rule. The reason for its non-application must therefore be found elsewhere in the language. Here I will follow Hualde (1991) by giving an account relying on Structure Preservation. For stops and fricatives there is an underlying voicing contrast, but Basque does not have voiced affricates in underlying representation. Apparently voiced affricates are not created either. We may express this as a segment structure condition (or a restriction on the combination of F-elements, along the lines of Archangeli and Pulleyblank 1993): (52)

*[stop, cont, +vce]

This condition forbids marking [+voice] on affricates, and blocks the application of the relevant rule.

155 3.2.6 Palatalisation The representation in (4) is also successful in accounting for the fact that Palatalisation, which turns the apico-alveolar coronals / t 1 n / into the palatals [c λ ji] after the vowel / i / (see examples in (53)a, from Hualde 1991: 61), does not apply to the apico-alveolar affricate / t s / (see Hualde 1991: 138ff.). Examples that show that Palatalisation does not apply to affricates are given in (53)b: [irudicu] [mendicik] [mutiXe] [ipijii] [itsela] [menditsue] [aritsue] [barâintsue] [bitse]

'imagine-PERF' 'from the mountain' 'the boy-ABS' 'put-PERF' 'huge' 'mountainous' 'stony' 'similar' 'foam'

It is clear that palatalisation affects coronals. However, only coronal stops are affected: that is, what the rule seems to require is an element I attached to a feature [stop]. I will not go into a discussion of the formal description of palatalisation here (relevant literature includes Clements 1976; Sagey 1986a; Lahiri and Evers 1991; Jacobs and van de Weijer 1992; see also § 2 in Ch. 6 above). For the sake of the argument, let us assume that palatalisation in Basque can be expressed as spreading of a feature Coronal from a vowel to a consonant. The target segment is crucial, as the rule in (54) shows: (54)

V

C I [stop]

I

I

The rule as formulated does not apply to affricates, because these do not have their element I under [stop]. Here the difference between (4) and the double-dependent representation in (35) is again crucial: there is nothing in (35) that would explain why Palatalisation should not apply to affricates. Note that in this case an alternative account relying on a segment structure condition (see (52) above) is not available, because affricates in the palatal region do exist in Basque.

156

3.3 Summary concerning Basque In this section I have shown that all processes concerning affricates discussed in Hualde (1991: Ch. 5) can be reanalysed using a simpler, more constrained, and more explanatory model of the affricate. The fact that the superficially quite intricate processes and constraints in Basque can be accommodated so simply is support for the representation of Basque affricates as in (4). The relation of direct dominance of the Place features by the Manner feature [cont] not only expresses the generalisation that fricatives and affricates appear at the same places of articulation in Basque, but was also shown to be relevant for the (non-) application of phonological rules.

4. Conclusion In this chapter I have drawn together two proposals regarding the structure of affricates and the relation between Manner and Place features. It has been proposed that the phonological arrangement of Manner features in affricates involves two unary features [stop] and [cont] (see Hualde 1987, 1988a,b, 1991; Lombardi 1990). It has also been proposed that Place features are directly dominated by the feature [± continuant] (Wetzels 1991). Uniting the two proposals led to the underlying representation of affricates that was proposed above. This representation was tested on the basis of a number of processes and constraints in Basque and a number of other languages.

Chapter 8 Prenasalised stops In this chapter I discuss in some more detail another class of segments that was referred to as complex in the previous chapters, namely that of prenasalised stops. This chapter is organised as follows: first I will introduce the representation of prenasalised stops proposed in chapter 5. From this representation a number of predictions will be derived, which will be shown to be correct (§ 1.1-3). Finally, a number of alternative proposals for the representation of prenasalised stops will be discussed (§ 2). These will be shown to be theoretically undesirable or empirically inadequate.

1. Prenasalised stops: three predictions In chapter 5 I proposed that a combination of a nasal and a stop specified for Place results in the underlying representation of prenasalised stops: (1)

C

/\ [stop] [stop] I I [nasal] Place The nasal part and the stop part of the prenasalised stop are both attached to the same root node. Since the [stop] features are ordered on their tier, the two parts are also ordered. The rightmost feature [stop] is attached to Place: there is only one Place specification in the segment (although this might itself be complex, e.g. palatal or labial-velar). In this chapter we will translate these general statements into three specific predictions, that is, first, prenasalised stops have a single root node and are therefore unit segments (§ 1.1), second, prenasalised stops are unlike affricates in that their parts are ordered (§ 1.2) and, third, Place is not specified on [nasal] underlyingly (§ 1.3).

1.1 Prenasalised stops as unit segments It has been argued by some authors that prenasalised stops do not exist at the underlying phonological level: although these researchers agree on the 'phonetic' unity of such segments (see, for instance, Herbert 1986: 10, 69), they argue that prenasalised stops are derived in the course of the derivation by a process of 'unification' (see especially Herbert 1977, 1986; Feinstein 1977; Poser 1979). In this section I

158 will examine this view in some detail and argue for the view that prenasalised stops do exist at the underlying level. One argument that is often advanced in favour of the view that prenasalised stops are derived instead of underlying is that a common line of reasoning put forward to support the idea that prenasalised stops are unit segments appears to be ill-founded. This is the argument that languages with prenasalised stops (or other 'complex' segments) do not allow branching onsets, that is, in these languages prenasalised stops are the only types of 'complex' onsets permitted. This is the case, for instance, in many Bantu languages. In many analyses these facts are presented as evidence that prenasalised stops are unit segments: if these segments are analysed in this way, all onsets of the language can be analysed as non-branching (this is referred to as the 'Canonical Pattern Fit' argument by Poser 1979: 97, who gives a number of examples of such analyses). Both Poser (1979) and Herbert (1986: 67ff.) point out that this reasoning is not valid: simply put, the possibility should be left open that prenasalised stops are the only exceptions to the generalisation that all onsets are non-branching.1 The fact that in a particular language usually just one consonant is allowed in onset position does not justify the conclusion that prenasalised stops also occupy a single consonant position. Rather, the evidence that prenasalised stops are unitary segments should be psychologically 'real' for the native speaker (see Herbert 1986: 76), which appears to mean that prenasalised stops must be regarded as unit segments when they function as such in the grammar. The objection raised by Poser and Herbert is a valid one. However, a number of points can be raised that lessen its impact. First, in a language which only permits prenasalised stops in onset position the question should also be raised why only these onset-like segments are permitted on the phonetic surface, and not others. Analysing prenasalised stops as sequences begs the question of why there should not be other, cross-linguistically more usual, sequences in the onset. Moreover, if such onsets are the 'only exceptions', they would be awkward phonologically, since they violate the sonority hierarchy (see Selkirk 1982, among many others). An analysis of such onsets as derived does not account for this, whereas an analysis of prenasalised stops as single (complex) segments may attribute a special status to such onsets. Besides, if the 'Canonical Pattern Fit' argument is not a valid argument in favour of analysing prenasalised stops as underlying unit segments, it is not an argument for a cluster analysis either. It seems reasonable to take the position that these segments are what they are phonetically (i.e. unit segments, as Herbert admits) unless there is phonological evidence to the contrary. For instance, the phonetic observation that prenasalised stops have approximately the same duration as single segments suggests the null hypothesis that prenasalised stops are single segments. Herbert (1986), Poser (1979) and Sagey ( 1986a,c) provide phonetic data to show that prenasalised stops have the approximate duration of segments that are unit segments under anyone's

1

This objection is not new: see also Martinet (1939: 101) and Hintze (1950: 17).

159

model, and their results are confirmed by Maddieson (1989). The burden of proof is therefore on those who claim that initial prenasalised stops are not unit segments, but sequences. This also seems to be the approach followed by Maddieson, who includes a large number of prenasalised segments in his cross-linguistic survey (Maddieson 1984). Evidence that prenasalised stops function in the grammar as unit segments comes from phonological processes and rules. This is what we turn to now. One argument, adduced by Sagey (1986a: 69ff.), comes from syllabification in Kinyarwanda, an eastern Bantu language spoken in Rwanda. Like many Bantu languages, Kinyarwanda only has open syllables, according to traditional descriptions; this is also the elicited syllabification pattern from native speakers. The only complex onsets allowed are labial-velars and prenasalised stops. This might not be sufficient evidence, for onsets consisting of nasal and stop (and labial and velar) might be the only exceptions. However, the way in which loanwords are incorporated into the language shows that the prohibition on branching onsets is a psychologically real one for the Kinyarwanda speaker, that is, prenasalised stops function as unit segments in the synchronic grammar. Consider the loanwords from German in (2) (see Sagey 1986a: 71): (2)

Gennari Alexander Republik Petroleum

Kinyarwanda aregisaanderi repuburika peeteroori

The borrowings receive epenthetic and final vowels to make them conform to the canonical syllable pattern. Notice that the nd in Alexander is not broken up, which suggests that the sequence is interpreted as a single segment, namely a prenasalised stop. A second source of phonological evidence uncovered by Sagey concerns compensatory lengthening. If prenasalised stops are single segments, the formation of a prenasalised stop from a nasal and a stop vacates a timing position. Under appropriate circumstances, this timing position will be available for spreading from neighbouring vowel positions, resulting in compensatory lengthening, i.e. in the creation of a long vowel. Evidence like this is presented in detail in Sagey ( 1986a,b). If such units can be formed in the course of the derivation, there is no reason to assume that they cannot exist underlyingly. Another source of phonological evidence that prenasalised stops function as units at a 'deep' phonological level are word games like the following, from the Bantu language Bakwiri (Hombert 1986; Gussenhoven 1991). The game consists of inverting the two syllables of disyllabic words:

160 (3)

mòkò lówá kóndì zàqgò

kòmò wáló ndikò qgózà

'plantain' 'excrement' 'rice' 'father'

The word for 'rice' in (3) is evidently syllabified kó.ndì, at the level of representation at which the game is played. An analysis in which prenasalised stops are onset units derives the output ndikò unproblematically, while analyses deriving prenasalised stops from an underlying heterosyllabic sequence have to assume an otherwise unmotivated resyllabification rule.2 Such language games usually occur before phonetic implementation takes place, i.e. at a lexical level (Vago 1985; Bagemihl 1988). Another piece of evidence comes from Saramaccan (Rountree 1972), a creole language spoken in the interior of Surinam. Many words are of West African, English, or Portuguese origin, and the language is one of the few tonal creole languages of the world. The dialect represented here is that spoken along the upper Surinam River. Rountree (1972: 22) gives the following consonant chart (labels and transcriptions retained; this also applies to the 'general nasal' /N/): labial

(4) nasals stops prenasalised voiced voiceless fricatives voiceless voiced

laminai

m

η

nj

mb b Ρ

nd d t

ndj dj tj

f ν

s ζ

lateral semivowels

apical

dorsal

ng g k

dorsolabial

gb kp

1 w

j

(general / N h/)

The 'N' represents nasalisation of the syllable and only occurs finally in the syllable. I suspect it is an example of the so-called anusvara (Trigo 1988); it alternates between a nasal stop, vowel nasalisation and a nasal glide [q], much as in Polish (Rowicka and van de Weijer 1992) or (European) Portuguese (Parkinson 1983).3 Note that

2 3

Bradshaw (1978) discusses a language game in Numbani which he argues is evidence that prenasalised stops are single segments. See also further discussion in section 3.2 in chapter 10.

161 Saramaccan has voiced, voiceless and prenasalised stops, which precludes an analysis of prenasalised stops as filling the voiced stops slots.4 The evidence that nasal plus voiced stop sequences are phonological units rather than sequences and that nasal plus consonants other than these are phonological sequences is particularly telling in this language. Rountree points out three reasons for regarding nasal plus voiced stop as a unit.5 First, voiced prenasalised stops occur word-initially, while other sequences of nasal plus consonant do not:6 (5)

'make' 'night' 'stutter' type of peanut

mbéi ndçti ngáku ndju

Second, native speakers indicate a syllable break before the prenasalised stops /mb nd ndj ng/ but after the 'general nasal' in /Np Nt Nk Nz/. These latter sequences are realised as [mp nt qk nz], respectively (see Rountree 1972: 26). Neither argument is conclusive evidence (since prenasalised stops might again be the 'only exceptions'). However, the third argument that Rountree adduces cannot be easily dismissed. This concerns an allophonic 'muffling rule' in Saramaccan, which appears to refer to the distribution of an ATR-distinction. The property is distinctive for the mid vowels and allophonic for the high vowels, as shown in the vowel inventory of this variety of Saramaccan, which is given in (6): (6)

i [i] e

u [V] o Ç

0

a The muffled counterparts of the mid vowels are 'lax' counterparts of the non-muffled, 'bright' ones. The distribution of the bright and muffled allophones is as follows: the bright allophones occur in non-nasal syllables and the muffled allophones in syllables that are closed by a nasal. This is shown in the following forms, where the dot under the / indicates the muffled allophone:

4

5

6

Many African languages have only two series of oral stops, e.g. voiceless stops and voiced prenasalised stops (e.g. Luvale (125) in Maddieson 1984). In such languages prenasalisation might be taken as a phonetic effect of underlying voicing, or at any rate as a redundant property. Such an analysis is not available for Saramaccan. Price and Price (1976: 40), in a discussion of play languages in the Saramaccan language area, also assume prenasalised stops in this language to be "single prenasalized consonants". As far as I can see, however, the play languages they discuss do not provide additional support for this view. Dorsolabials iltpl and /gif/ also occur initially, e.g. kpáta 'black and white Capuchin monkey' and g¡bamba 'meat', where these segments alternate with [k w g w ].

162 (7)

a. b.

/gulí/ /gbçliN/

[gulf] [gbçlî ~ gbçljtq]

'swallow' absolute indicator

In (7)a bright [u] and [i] occur in the open initial and final syllables, and in (7)b the final [i] is muffled (and variably nasalised), since the syllable is closed by a nasal. Now compare the behaviour of the nasal when followed by a voiced stop, before which there is no muffling ((8)a), and the nasal when followed by a voiceless stop, before which there is muffling ((8)b): (8)

a. b.

/vindç/ /viNtu/

[vindç] [vintu]

'throw' 'wind'

The data in (8)a show that [i] is not tautosyllabic with [n] that is followed by a voiced stop: the syllabification is therefore /vi.nde/. Hence, no muffling takes place. In (8)b it is shown that the vowel and the nasal are tautosyllabic when the nasal is followed by a voiceless stop: here the syllabification is /vin.tu/. An analysis of prenasalised voiced stops as single onset segments derives these facts without additional stipulations. Note that in Saramaccan, as in many other languages, only prenasalised voiced stops occur. In Maddieson (1984) there are no examples of prenasalised voiceless stops. This is explained in the representation we have proposed by the fact that prenasalised stops have a single root node; since we have assumed (without discussion) that laryngeal features are attached to the root node, it would seem that such segments cannot have opposite laryngeal specifications (the same is true of affricates and lateral affricates, for instance). If prenasalised voiceless stops exist, they must be represented differently, in particular, as two-root complex segments (see chapter 10). This would account for the fact that they have a voice contour. A theory which represents all prenasalised stops (voiced or voiceless) as double-root segments needs further stipulations to account for this asymmetry. On the basis of the above arguments, we conclude that at least in some languages prenasalised stops must be permitted in the underlying repertory of segments, so that phonological theory must provide adequate representations for them. Let us now turn to some other aspects of the representation of these segments.

1.2 Ordering Compare the representation of prenasalised stops in Sagey's (1986a) model ((9)a, details omitted) and that proposed here ((9)b):

163 (9)

a.

b.

C

/\

[ + nasal]

[ - nasal]

C / \

[stop] [stop] [nasal]

In Sagey's representation, the features [ +nasal] and [-nasal] appear on the same autosegmental tier, so that they are ordered. In our representation the same ordering effect is achieved as a result of the fact that the two [stop] features are ordered on their tier. The different ways in which the ordering effect is brought about makes certain phonological predictions. These will be focused on in this section. The available evidence suggests that the two parts of a prenasalised stop are ordered. In this respect the structure of prenasalised stops does not seem parallel to the structure of affricates. Evidence that in affricates the parts are not ordered was presented in chapter 7 above. If the structure of prenasalised stops were parallel to that of affricates, i.e. if the two parts of prenasalised stops were unordered, we might expect that such segments should be able to nasalise vowels following the prenasalised stop, as in (10): (10)

c

V

[nasal] Processes of this kind are not attested, as far as I know. In fact, in Gbeya (Samarin 1966), pure nasal consonants can be followed by nasal vowels, but prenasalised consonants must always be followed by oral vowels (see Anderson 1976a: 335f.). The same situation occurs in many other languages (see Piggott 1991). This is unexpected if the [nasal] part of prenasalised stops is unordered with respect to the [stop] part. In this section I will examine two phonological phenomena that bear on the ordering question: the first concerns Ngbaka morpheme structure conditions. These are discussed in the light of the possibility that prenasalised stops, too, might give rise to anti-edge effects. We will show that the Ngbaka effects have been wrongly analysed as OCP-related effects, and that an account based on the sonority of the segments involved is more insightful and empirically more adequate. The second case concerns segment simplification processes such as those of the Ganda Law (Schadeberg 1987). These will be discussed in the light of the claim that the ordering effect in prenasalised stops is due to the fact that these segments have two contiguous [stop] features, and not to the fact that they have a sequence of [ +nasal] followed by [-nasal]. Hence, both will be seen to provide support for the representation in (9)b, which, unlike Sagey's representation, allows one to retain the restrictive hypothesis that the feature [nasal] is unary.

164

U . I Ngbaka Ngbaka is a language of the Adamawa branch of the Niger-Congo family, spoken in the Central African Republic.7 The root cooccurrence restrictions in this language have attracted considerable attention since they were first reported by Thomas (1963) (see Wescott 1965; Chomsky and Halle 1968: 387; Clements 1982; Itô 1984; Churma 1984; Herbert 1986: 113-16; Mester 1986: 32-45; Sagey 1986: 260-66; Lombardi 1990: 376 (fn. 2); Selkirk 1991, among others). In this language, there are vowel cooccurrence restrictions (not discussed here), as well as restrictions on the occurrence of consonants with the same point of articulation. The latter are stated below (from Wescott 1965, who summarises Thomas 1963): (...) if a disyllabic word contains a voiceless consonant, it does not also contain the voiced counterpart of that consonant (that is, / p / excludes /b/, /s/ excludes /z/, etc.). Similarly, if a disyllabic word contains a voiced obstruent, it does not also contain the prenasalized counterpart of that obstruent (that is, / b / excludes / m b/, /z/ excludes /"z/, etc.); if such a word contains a prenasalized obstruent, it does not also contain the corresponding nasal (that is, / m b/ excludes /m/, /"z/ excludes /n/, etc.) (...). (Wescott 1965: 346) Mester (1986: 45) makes explicit that these restrictions are bidirectional: "both *b + mb and *mb + b are excluded", and points out that this is the reason why "it is crucial to regard the elements of the nasal contour as underlyingly unordered with respect to each other". He points out that the consonant cooccurrence restrictions have the form of a gradation system, as in (11), where adjacent elements exclude each other, whereas non-adjacent elements are compatible (we will see reason to question the empirical adequacy of this generalisation below, but let us accept it for the time being): (11)

ρ - b - mb - m

Identical segments are also permitted to combine. In addition to the combinations of identical segments, therefore, there are six possibilities that must be considered: (12)

7

*p ρ ρ *b b *mb

b mb m mb m - m

This discussion of the Ngbaka data also appears in van de Weijer (1994a).

165 Obviously, restrictions have to take into account voicing, nasality and place of articulation. The Manner-Place dependency adopted in the model developed in this study, where Place elements are attached to the features [stop] or [cont], predicts correctly that these restrictions range over stops and nasals on the one hand (where Place is under [stop]), and over fricatives on the other (where Place is under [cont]). For voicing, I will assume that obstruent stops are underspecified for [-voice] and that nasal stops are underspecified for [+voice]. Voicing in the prenasalised stop is also predictable. The crucial segments can be represented as follows: (13)

C I [stop] I U

c [+voi] -"I [stop] I U

/P/

C /\ [stop] [stop] I I [nasal] U

M

/ mb/

[stop] [nasal]

U

M

The three combinations that are disallowed according to the generalisation sketched out above (*/p-b/, */b-mb/ and */mb-m/) are given in (14) below: (14) a. 7 p - b /

b. */b-mb/

c. */mb-m/

C I [stop] I U C r^[+voi] [stop] I U c /\ [stop] [stop] I I [nasal] U

C r-[+TOi] [stop] I U C /\ [stop] [stop] I I [nasal]

U

C I [stop] f \ [nasal] U

The combinations that are assumed to be well-formed are provided in (15) below:

166 (15)

a. / p - m b /

b. / p - m /

c. / b - m /

C

C

[stop]

[stop] [stop]

U

[nasal]

C

C

[stop]

[stop]

I U

[nasal]

C n+voi] [stop] U

U

U C [stop] [nasal]

U

I propose to formalise Mester's idea that two consonants in a root may not be 'too much alike' by the restriction that consonants which have adjacent U elements that are linked to [stop] must differ by at least two features, whether [nasal], [voice] or the prenasalisation complex ([stop] dominating [nasal]). That is, if they differ by only one feature, they are too much alike, and the sequence is ruled out. This analysis accounts for four of the six cases directly: it correctly rules out * / p - b / (which differ only in voicing specification) and * / m - m b / (where the prenasalised stop only has an additional [stop] feature), and correctly accounts for the well-formedness of / p - m b / (which differ in voicing and prenasalisation) and / b - m / (which differ in voice specification and nasality). This leaves the cases of / p - m / and / b - m b / , in the first of which the two root consonants only differ in one feature (and which is therefore predicted to be ill-formed), and in the second of which the two root consonants differ in two features (and which is therefore predicted to be wellformed). A careful search through the original Ngbaka source, Thomas (1963), revealed not a single instance of a root with / p - m / (with / p / and / m / in either order). Another source of the language, Sevy (1972), contains a dictionary of the language. This too did not contain a single instance of a root with / p - m / . In a third source, Derive (1975), one item was found with / p - m / : pàmè 'marsh swine' (Derive 1975: 153), which does not occur in either of the two dictionaries of the language that were available to me (Maes 1965 and Sevy 1972). As for roots with / b - m b / , a number of these are listed as 'exceptions' in Thomas (1963: 63), and a search through the available sources reveals that such roots are not uncommon. The Ngbaka constraint on minimal difference can be compared to other effects where adjacent segments must have some degree of 'minimal difference' to be acceptable. For instance, in English onsets /pi-/ and /pr-/ are allowed, while / p n - / is

167 not. This has been expressed by saying that the latter two segments are too similar (cf. the minimal sonority distance constraints of Selkirk 1982 and van der Hulst 1984). Since sonority may be computed by inspecting the whole of the segment, such cases are similar to the one of Ngbaka. More abstractly, the Ngbaka case has wrongly been analysed as exclusively due to OCP effects. Hence, the fact that both /mb-b/ and / b - m b / are ruled out is only apparently an 'anti-edge effect': the effect concerns the composition of the whole segment and does not refer to specific tiers within the segment. I conclude that the empirical side of the Ngbaka generalisations which have given rise to so much discussion might have to be reconsidered. Pending a more refined statement of the data, I offer the above analysis as an alternative for the 'gradation series' analysis offered in Mester (1986). Note that in the alternative analysis that I have proposed it is not crucial to assume that the parts of prenasalised stops are unordered. Hence, this potential piece of evidence for unordering the parts disappears under my analysis. Crucially, however, all segments involved in these restrictions have their Place specifications under a [stop] feature, which predicts the existence and scope of such conditions.

1.2.2 The Ganda Law and other rules In this section I examine a number of variants of Meinhofs Law (Schadeberg 1987), which also bears on the ordering of parts within prenasalised stops. I will show that, despite appearances to the contrary, these variants do not motivate a hypothesis that the parts of prenasalised stops are unordered. I will also compare the formalisation of the various rules under my representation of prenasalised stops with the formalisation that is needed when the representation of prenasalised stops in Sagey (1986a) is adopted. The latter representation is argued to be less insightful, whether the two parts ([ +nasal] and [-nasal]) are assumed to be ordered or not. Schadeberg argues that there are four variants of the Ganda Law: (16)

The Ganda variant : The Lamba variant : The UMbundu variant : The Kwanyama variant

prenasalised stops lose their stop part before nasals and before prenasalised stops prenasalised stops lose their stop part only before prenasalised stops (and not before nasals) prenasalised stops lose their stop part only before nasals (and not before prenasalised stops) prenasalised stops lose their nasal part after prenasalised stops (while ordinary nasals are unaffected)

Let us discuss these variants in turn (see Schadeberg 1987: 2f. for examples and a fuller discussion).

168 In the Ganda variant of Meinhofs Law prenasalised stops lose their stop part before nasals and before prenasalised stops. These two variants are characterised in (17)a and (b): (17)

a.

C

[stop] [stop] I I [nasal] Place b.

C I [stop] I [nasal] C / \

C

Afe [stop] [stop] I I [nasal] Place

[stop] [stop] I I [nasal] Place

The trigger of the rule can be formally identified as a [stop] feature which dominates [nasal]: it does not matter whether this [stop] feature is part of a prenasalised stop or makes up a nasal itself. The [stop] feature removes the target, which can be formally characterised as a [stop] feature (dominating Place), if this is part of a prenasalised stop (ordinary stops are not deleted before either nasals or prenasalised stops). Basically this is an OCP effect on contiguous [stop] features, but with a number of specific conditions.8 Consider now the formal equivalent to (17) in a framework such as that of Sagey (1986a): (18)

a.

C

C

[+nasal] b.

[-nasal] C

AH [+nasal]

[-nasal]

I [+nasal] C / \ [+nasal] [-nasal]

In Sagey's model the basic insight of the Ganda Law, viz. that an occlusion (or [stop] feature) is removed before another occlusion, cannot be expressed. On the contrary, the [-nasal] part of a prenasalised stop is lost before a [ +nasal] specification, which would make it a very unnatural type of rule. The only rationale that could be given is

8

It does not seem inconceivable that the rule was once more general in Bantu, and that different varieties developed these different conditions. See also Meeussen (1962: 29): "Reviewing all the available evidence, one gains the impression that Meinhofs rule dates back to Proto-Bantu, at least for the eastern half of its dialect domain".

169 that two [+nasal] features drive out a [-nasal] specification in between them to achieve a fully [ + nasal] stretch. Rules of this type are very rare, as far as I am aware. Now consider the question of whether the rule could be characterised in a helpful way if the parts of prenasalised stops were unordered: (19)

a.

C

C

I [+nasal]

[ + nasal] [-nasal] b.

C [ + nasal]

[+nasal] [ - nasal]

[ - nasal]

Only in (19)b can the deletion of the [-nasal] part be expressed as an OCP effect on features on the same tier: [-nasal] is removed before [-nasal]. However, I argue against the view that this part of the deletion process forms evidence in favour of a hypothesis which postulates that the two parts of prenasalised stops are unordered. First, the deletion of the [-nasal] part in (19)a must be expressed differently, and hence the two variants of the Ganda variant of Meinhofs Law cannot be expressed as a single rule under these representations. Second, ordinary non-nasals do not trigger the deletion of the non-nasal part, so that the OCP effect on [-nasal] is only partly productive. In the representations proposed in (17), on the other hand, trigger and target are adjacent, identical features, and the Ganda variant can be stated as a single (although complicated) phonological rule. In the Lamba variant of Meinhofs Rule prenasalised stops lose their stop part only before prenasalised stops. This is a more restricted version of the Ganda variant: only the (17)b variant is found. In a Sagey-like representation of prenasalised stops, this variant may be taken as an anti-edge effect, as was pointed out above (see (18)b), since ordinary nasals do not trigger the simplification. It would therefore seem that an oral occlusion is deleted before another oral occlusion, which are formally identified with the [-nasal] parts of the prenasalised stops. However, note that ordinary oral stops neither undergo nor trigger this effect, so that it is equally valid to assume that the rule of Lamba is (even) more restricted than that of Ganda, in the sense that the trigger has to be specified as a prenasalised stop. I conclude that the occurrence of this rule is not counterevidence to the proposed representation. 9

9

(i)

The Lamba variant has a parallel in Zande (see Tucker and Bryan 1966: 143, who also point out the parallel to the Ganda Law), where voiced nasal compounds often lose their explosive element when followed by another nasal compound: m(b)_ámbakada q(g)enzi

'to prepare-FREQ' 'Nile perch'

170 In the UMbundu variant of Meinhofs Rule prenasalised stops lose their stop part only before nasals (and not before prenasalised stops). Similarly, this is a more restricted variant of the Ganda Law: only the (17)a variant is found. This rule is fundamentally unexpressible as an OCP effect in Sagey's framework, since the [-nasal] part is deleted before a [ +nasal] segment. Although again an occlusion is deleted before an occlusion, this is not statable as such in Sagey's model. The representation proposed here, where the effect is due to the sequence of two [stop] specifications (whether these represent oral or nasal occlusions), does account for the UMbundu variant in a similar way as was proposed for Lamba: the rule is basically an OCP effect on [stop], with additional specifications on the trigger. In the Kwanyama variant of Meinhofs Law prenasalised stops lose their nasal part after prenasalised stops (while ordinary nasals are unaffected). Thus, this variant is different from the other three, since here the nasal part of a prenasalised stop is lost after prenasalised stops. It is therefore not surprising that, as Schadeberg notes, that the Kwanyama variant "is usually considered an independent rule" (Schadeberg 1987: 2 - my translation, JW). Its effect can be formalised as follows: (20)

C / \ [stop] [stop] I I [nasal] Place

C [stop] [stop] I I [nasal] Place

The rule is a mirror image of the Lamba variant. The trigger must be characterised as the [stop] feature that is part of a prenasalised stop. Consider the formalisation of the Kwanyama variant if the representation of prenasalised stops as proposed in Sagey (1986a) is adopted: (21)

C [+nasal]

C [-nasal]

[ +nasal]

[-nasal]

The rule cannot be expressed as the deletion of one occlusion before another. This might be taken as evidence that [+nasal] and [-nasal] are not ordered in prenasalised stops, and that the Kwanyama variant on the Ganda Law is an OCP effect on the feature [+nasal]. In that case, the rule could be formalised as follows: (22) [ + nasal]

[ + nasal] [-nasal]

[-nasal]

171

However, recall that ordinary nasals are neither affected by this deletion process nor do they trigger it. This fact would remain unexplained under this analysis. Hence, I reject (22) and propose to regard it as an instance of a local effect on [stop], with the conditions specified above. From the discussion of these four variants of Meinhofs Rule I conclude that they present no crucial evidence for a hypothesis that the oral and nasal part of prenasalised stops are unordered. Rather, all the effects examined above can be expressed as local effects on the occlusion feature [stop]. Since these variants can be stated as such, they present evidence for the representation of consonants argued for here, in particular of nasals (which crucially have a [stop] feature) and prenasalised stops (which are crucially composed of a feature matrix for nasals and one for stops).

1.3 Place attached to [stop] only Consider once more the proposed representation of prenasalised stops: (23)

C

/\

[stop] [stop] I I [nasal] Place The place of articulation of the nasal part of a prenasalised stop is fully predictable. Hence, it should be left out of the underlying representation. This is captured in the representation in (23). How does the nasal part receive its surface place of articulation? There are a number of possibilities: First, the Place specification in a segment might be automatically realised on all parts of the segment (i.e. the representation is interpreted as it stands in (23). This is basically what Sagey (1986a) assumes for all complex segments. Second, the [nasal] feature might receive a default coronal specification, after which Place assimilation and delinking takes place: (24)

C [stop] [nasal]

Coronal

Place

Another possibility is that the Place specification of the [stop] part might spread, segment-internally, to the [nasal] part:

172 (25)

C

/ \

[stop] [stop] r-^l [nasal] Place It should be noted of course that nasals typically (but not universally) are placeless, or that at least the coronal nasal assimilates very often to neighbouring stops. Such a process is much more common, as far as we know, than assimilation between stop and fricative. If the strategy formally described in (25) is adopted, this fact is reflected in the difference in representation between affricates and prenasalised stops: in the former, Place would spread between different Manner features (from [cont] to [stop]), whereas in the latter Place would spread between identical Manner features ([stop] in both cases). It seems a natural assumption that the former is less common than the latter. However, I have no crucial evidence for or against any of the three possibilities for the interpretation of Place on the nasal part of prenasalised stops.

2. Alternative representations of prenasalised stops In this section I discuss a number of other representations that have been proposed for prenasalised stops. In § 2.1 I discuss specific arguments against Sagey's contour representation of prenasalised stops, and in § 2.2 I discuss a representation of prenasalised stops as two-root complex segments advanced by Rosenthall (1992).

2.1 Prenasalised stops as contour segments Recall Sagey's (1986a) representation of prenasalised stops: (26)

C / \ [+nasal]

[-nasal]

The possibility of arranging feature values in this way may be convenient, since it also allows an easy expression of postnasalised stops and 'medionasalised' stops; the last have been reported for the South American language, Kaingáng (see Wiesemann 1972; Anderson 1976a). The medionasalised stops could be represented as in (27): (27)

C [-nasal] [ +nasal] [-nasal]

173 Such medionasalised stops do not occur underlyingly in any language. Hence, a theory of phonological segment representation should exclude segments such as that in (27) as possible phonemic segments. The model advanced here does not allow for their existence as such. However, I would like to offer some discussion on the process by which they arise. In this process an underlying nasal becomes medionasalised when adjacent at both sides to oral vowels. To describe such a process, it is unlikely that any current theory would represent oral vowels underlyingly with a [-nasal] feature. In current underspecification analyses [ - n a s a l ] may be inserted in the course of the derivation, while unary-feature proposals would argue that [nasal] is privative, so that [-nasal] can never play a role in the phonology. This entails that processes such as that in Kaingáng must be analysed as phonetic (when specifications like [-nasal] may play a role). The prediction is that such a process cannot become phonologised, and that medionasalised segments will never be underlying segments. Finally, note that after Sagey (1986a) there have not been many proposals to regard segments other than affricates or prenasalised stops as contour segments. The extremely powerful device of contour representation has not been extended to other segment types.

2.2 Prenasalised stops as two-root complex segments Rosenthall (1992) claims that prenasalised stops are two-root complex segments. This contrasts with Sagey's representation of prenasalised stops in which they have a single root node. The empirical evidence he adduces comes from compensatory lengthening and interaction with syllabification. Let us briefly review this evidence here. The first source of evidence concerns languages which have a process as a result of which a vowel becomes long when it is followed by a prenasalised stop. Clements (1986) accounts for this process in Luganda by proposing a rule of prenasalisation which delinks the nasal from its skeletal point and attaches the segment to the skeletal point of the obstruent. The skeletal point which was previously linked to the nasal is now linked to the vowel, which produces a long vowel. However, the argument of compensatory lengthening is by no means evidence that prenasalised stops have two root nodes. The change from stop to prenasalised stop may be expressed by spreading at different heights, and there is no evidence that this must be the root node. In (28)a I give Rosenthall's conception of prenasalised stop formation, and in (28)b I give mine:

174 (28)

a.

χ

χ

I V

* , - ' Ί C"' C I [stop]

χ

timing tier root tier

I [stop]

I [nasal] b.

χ

χ

χ

timing tier

I V

I C

I C

root tier

t — - 1 [stop] [stop] I [nasal] There is no reason to believe that compensatory lengthening is more to be expected under the first analysis than under the second. The only thing that needs to be assumed in the second analysis is that the bare consonantal root node that results from delinking is stray-erased. Such an assumption is probably necessary in the grammar anyway. Note, finally, that both (28)a and (28)b are possible phonological processes under both approaches to the representation of prenasalised stops, so that it is difficult to see why it should offer conclusive support for either version. The second argument that Rosenthali adduces comes from singular-plural alternations in Sinhalese. The facts are given in (29):10 SG kanda homba potta redds

PL kandu ho m bu potu redu

root kandw hombw potw redw

'hill' 'chin' 'core' 'cloth'

Nasal plus stop sequences alternate with prenasalised stops and geminates alternate with normal-length segments. This indicates that prenasalised stops behave like single segments in this language, which is an argument for regarding them as "psychologically real" in this language (see § 1.1 above). As far as I can see, Rosenthali assumes that the SG morpheme in this language must be derived from the plural by insertion of a root node within the stem: this would argue for a representation of prenasalised stops as having two root nodes. This analysis is not the only possible one, however. The plural can be derived from the singular form in various ways. On one assumption

10

See also Reynolds (1980: 7) for the observation that word-medially there is a contrast between clusters of nasal and stop and prenasalised stops in Sinhalese.

175 the / n d / sequences and the geminates in the singular are both single, ambisyllabic segments, prenasalised stops in the case of (29)a above and simple stops in the case of (29)b. n Plural formation can then be described in a simple and elegant way, namely as delinking the segment from the rhyme node of the first syllable: (30)

σ \

R

X/χ

O

σ /

I c Finally, recall that prenasalised voiced stops are the norm across the world's languages, while prenasalised voiceless stops are much rarer, although they exist (see e.g. Herbert 1986 for a number of examples). A representation in which prenasalised voiced stops have a single root node (and therefore a single laryngeal specification) accounts for this directly: prenasalised voiceless stops - if they exist - must be represented differently. If a two-root representation of such segments is adequate, it would explain their comparative rarity (in Ch. 10 we will see that two-root complex segments are on the whole much less common than complex segments of other types). I conclude that there is no crucial evidence to represent prenasalised stops with two root nodes. I note, however, that positive evidence for representing prenasalised stops with one root node is also lacking. This would come, for instance, from templatic systems in which prenasalised stops are mapped as single segments to a template. Such evidence is available for affricates (see § 3.1.1 in Ch. 7), but not for prenasalised segments. It may be a coincidence that the set of languages which have templatic systems does not intersect the set of languages which have prenasalised stops.

3. Conclusion I conclude that the representation of prenasalised stops provided by the theory advanced in this study does not face grave empirical or theoretical problems. This representation has aspects in common with the contour representation of Sagey (1986a), but achieves the ordering of parts without the use of ordered values of a binary feature. Only a single Place specification is present underlyingly, as demanded by the Mismatch condition. Such a representation is compatible with the empirical

11

Sec van der Hulst and Smith (1982) and Borowski, Ito and Mester (1982) for the idea that length and ambisyllabicity are related notions. See also fn. 1 in the next chapter for an application of this idea to /s/ plus stop clusters.

176 facts examined here, and in certain cases (e.g. Ngbaka, Ganda) is more useful for a formal account. In the next chapter I turn to a variety of complex segments that have also attracted a great deal of attention, namely / s / plus stop clusters.

Chapter 9

/ s / plus stop clusters

The peculiar status of / s / plus stop clusters, especially in onset position, has frequently been noted in the literature. Various linguists have tried to account for the special behaviour of / s / plus stop clusters by treating them as single, complex segments. Cygan (1970), for instance, discusses the English clusters /sp st sk/ in initial position. He notes that these are the only underlying clusters in English which violate the sonority hierarchy and that some linguists have attempted to treat them as complex units for this reason. Other linguists prefer to speak simply of a third kind of stop here, for instance as a 'stigmatized' stop (Hill 1966) (see Herbert 1986: 56). Kuryiowicz (1966: 199) notes their "compound character" in alliterative verse (see below), and Fudge (1969: 278) treats them as "single phonemic elements". The representation of / s / plus stop clusters proposed here is that of a combination of a continuant unmarked for Place and a stop that is specified for Place features under a single root node. This representation is given in (1):

(1)

C

/\

[cont] [stop] I Place In this chapter we discuss the representation in (1) in detail. Previous discussion of the peculiar status of / s / plus stop clusters appears in, among others, Kuryiowicz (1966), Kohler (1967), Fudge (1969), Fujimura and Lovins (1978), Ewen (1982), Selkirk (1982), Booij (1984), van der Hulst (1984) and Broselow (1991). First, we will turn to some issues related to syllabification (§ 1). An account will be presented in which / s / plus stop clusters are single segments. A number of alternative ways of accounting for this behaviour will also be discussed. Second, we turn to the behaviour of / s / plus stop clusters in phonological rules and constraints (§ 2).

1. Syllabification It is well known that / s / plus stop clusters violate the generalisation that segments tend to be arranged in order of decreasing sonority when approaching syllable margins. In this section I will discuss two accounts of this violation: the first is an account in which / s / plus stop clusters are regarded as complex segments, and the second is one in which / s / is outside the syllable. If the representation of / s / plus stop clusters as in (1) is adopted, the violation of the sonority sequencing generalisation disappears. The fricative part of the complex segment, which is unspecified for Place in underlying representation (it is predictably

178 coronal) precedes the stop part, which is specified for Place. The ordering between the two elements within the segment is not determined by sonority considerations, but by the generalisation that heads are realised second in complex segments (see § 3 in Ch. 5).1 An additional advantage of looking upon / s / plus stop clusters as filling a single segment slot is that the onset in English (and other languages) can be limited to just two positions, of which the first can only be filled by obstruents, including / s / plus stop clusters (see Fudge 1969: 273). The latter argument parallels the one that prenasalised stops occupy a single position in the syllable structure of languages that normally allow only single-consonant onsets; hence, it is not proof that / s / plus stop clusters are single segments, but at least it is a strong suggestion. An alternative way of accounting for the appearance of / s / plus stop clusters in onset positions is by reference to the special status of coronals (see many of the contributions to Paradis and Prunet 1991, among others). In such an approach it may be proposed that coronals are not syllabified in underlying representation, and do not become incorporated in syllabic structure until a later stage in the syllabification (this approach is argued for by Steriade 1982, 1988; Levin 1985, among others). Such a solution begs the question of why it should be only / s / that appears in pre-onset position, and not other coronals, such as /t η 1/. Second, there is a potential empirical consequence. Recall that languages such as Korean do not allow coronal consonants before front vowels (see § 2.3 in Ch. 2 above). If a syllable like /si-/ is represented as in (2) below, the morpheme structure condition cannot adequately express the domain of the constraint.

(2)

/\ O

R I i

The domain of the constraint cannot be the syllable, since / s / is outside the syllable. Rather, the constraint should refer to elements both within and outside the metrical structure. Hence, if coronals are not part of the syllable structure underlyingly, such morpheme structure conditions cannot be expressed in a straightforward way.

1

Notice that there is no sonority violation in final position: in English task, for instance, there is no reason to assume a special (monosegmental) status of the segments /s/ and /le/ (see Ewen 1982: 51). However, sonority violations may also arise word-finally, as in tax, for instance, where the higher-sonority /s/ occurs after ¡YJ. A final extrametrical element (sometimes referred to as an 'appendix', see Fudge 1969; Fujimora and Lovins 1978, among others) might have to be incorporated to account for these cases. If such a device is only necessary in word-final position (because initial /s/ plus stop clusters are complex segments), phonological theory is arguably more constrained than when such a device is needed both word-initially and finally, especially in the light of the fact that for stress only final extrametricality seems to be required (Hayes 1995). However, it is not clear if all cases in which initial extrametricality has been postulated can be accounted for in this way: see the case from Italian discussed below.

179 Moreover, such an approach faces potential difficulties if in such cases the same constraints hold word-medially. Of course, the idea that coronals are underlyingly extrasyllabic also has certain merits. I note that it is not necessarily incompatible with the idea that / s / plus stop clusters are single segments. To show that / s / extrametricality may be independently necessary, let us consider the il~lo allomorphy in Italian. In this language the masculine definite article can have two forms: il or lo. The first of these occurs before nouns and adjectives that begin with a single consonant (including / s / , as in il senso 'the sense'), or with a specific consonant cluster (see examples in (3)a), and the second of these occurs before nouns and adjectives that begin with a vowel or glide or a set of other specific consonant clusters (see (3)b). Before a vowel lo is truncated to /'. All the examples in (3) have been taken from Davis (1990: 44). il il il il il il il il

burro capitale clima drago globo grado traffico premio

'the 'the 'the 'the 'the 'the 'the 'the

butter' capital' climate' dragon' globe' grade' traffic' prize'

b.

l'est l'osso lo spirito lo studente lo scampo lo slanco lo smalto lo sfarzo

'the 'the 'the 'the 'the 'the 'the 'the

east' bone' spirit' student' rescue' outburst' pavement' pomp'

It is easy to see that all clusters with / s / behave in the same way with respect to the selection of the article, regardless of whether the second part of the cluster is a stop, a fricative or a sonorant. Davis (1990) proposes that / s / in all these cases is outside the initial syllable. If there is an initial syllabified element, il is selected ((3)a); otherwise lo appears ((3)b). It should be noted, however, that lo also appears before another set of consonants, namely those that could be called complex·. (4)

lo lo lo lo

zio zero scialle gnomo

'the 'the 'the 'the

uncle' zero' shawl' gnome'

The first two of these examples begin with affricates ([ts dz]), and the latter two begin with the palatals ([î ji]). To account for the fact that lo also occurs before these segments, Davis assumes (following Chierchia 1986, among others) that they are underlyingly geminates, and that the first part of the geminate is also outside the syllable. Initial geminates, however, seem extremely rare and have therefore been treated with some suspicion in the literature (see Hayes 1989). The assumption that the affricates and palatals are geminates is not necessary if it is assumed that lo not only appears before a vowel or an extrametrical consonant (such as in the examples

180 with an / s / followed by a nasal or fricative in (3)b above), but also before complex segments, such as in the examples with / s / plus stop clusters in (3)b and the examples in (4). That is, il appears when the onset is properly filled, and when there is no branching in the segmental material: 2 (5)

select il in the following structure: NP[0

I χ

skeleton

I C

root node

I [. . .]

segmental material

In all other cases, that is, when there is no onset (or, in other words, when this is not filled), when there is an extrasyllabic root node (which is not linked to any syllabic structure), or when segmental branching occurs, lo is selected. I conclude that syllabification of / s / plus stop clusters as single segments is, minimally, a viable alternative to other approaches which seek to account for the special status of such clusters. Let us now turn to an examination of the behaviour of / s / plus stop clusters in phonological rules and constraints. These will show that in other cases the single-segment approach to / s / plus stop clusters is demonstrably more explanatory than an extrasyllabicity approach.

2. Phonological rules and constraints In this section I discuss the behaviour of / s / plus stop clusters in phonological rules and constraints. 3 First, in § 2.1, I turn to their behaviour in some reduplication systems. Then (in § 2.2) I turn to the phenomenon that these clusters have geminatelike properties (Broselow 1991). Finally, there is a variety of other evidence that is relevant (§ 2.3). The evidence presented below does not discriminate between a representation of / s / plus stop clusters as proposed in (1), and a representation of

2

Note that the association line between onset and skeleton must be interpreted non-exhaustively: before two properly syllabified consonants il is also selected (this violates Hayes' 1986 Linking Constraint). Between skeleton and root node, however, the association lines must be interpreted exhaustively for the constraint to work. There may be more Linking Constraint violations of this kind, for instance in cases where both short and long vowels trigger a phonological process such as palatalisation or umlaut. This may entail that the scope of the Linking Constraint should be limited to segment-internal association lines or that it should be limited to refer to targets of phonological processes, such as the well-known Tigrinya inalterability effects (see Schein and Steriade 1986). Discussion of such matters falls outside the scope of the present work.

3

Part of the material below also appears in a slightly different form in van de Weijer (1993a).

181

these as two-root complex segments, with the continuant part and the stop part divided over separate segments linked to a single timing slot. Since the representation in (1) fits more naturally in the framework developed here, I will adopt it. A quest for evidence that could discriminate between the two rival representations of / s / plus stop clusters will be left for further research.

2.1 Reduplication In this section I discuss a number of reduplication processes in which / s / plus stop clusters behave differently from other clusters. I show that the assumption that / s / plus stop clusters are single segments can account for such cases quite naturally, while additional stipulations are necessary in other approaches. A strong case for regarding / s / plus stop clusters as single segments comes from Gothic reduplication. In Gothic there is a class of verbs which form their preterite by means of (partial) reduplication (Wright 1910: 146ff.; Braune 1912: 91; Mossé 1942: 126ff). For example: a. b.

c.

haita h w opa fraisa grëtan slëpan (ga-)stalda skaida

Ί am called' Ί boast' Ί try' 'to weep' 'to sleep' Ί acquire' Ί sever'

haihait hwaihwop faifrais gaigröt saislëp staistald skaiskaij)

Ί Ί Ί Ί Ί Ί Ί

was called' boasted' tried' wept' slept' acquired' severed'

In the forms in (6)c the / s / plus stop clusters behave exactly like single consonant onsets. Ordinary clusters, also with / s / , are simplified ((6)b). Note that the complex segment / h w / is also reduplicated as a single consonant ((6)a). Thus, the template for reduplication is simply a CV syllable, with the melodic content of the vowel prespecified as [ai]. An approach to / s / plus stop clusters in which / s / is extrasyllabic cannot account for the Gothic data. To derive the forms in (6)c correctly, it must be assumed that / s / is incorporated into syllable structure before reduplication takes place. This would leave the fact that, for instance, / l / is lost in the form saislép Ί slept' completely unaccounted for. Consider now Sanskrit reduplication (Kiparsky 1979; Steriade 1982, 1988: 119ff.), which forms an interesting counterpart to Gothic. Sanskrit perfective reduplication presents a problem for the principle of Template Satisfaction of McCarthy and Prince (1986, 1990), because while ordinary complex onsets simplify to their first member ((7)a), in / s / plus stop clusters the / s / does not map onto the préfixai template ((7)b) (irrelevant further changes omitted, see Steriade 1988: 120):

182 (7) a.

b.

root prath dru tsar k$ad mna: snih smi sru svar skand stambh stu sphut sprd h

perfect (full grade) 'to spread' pa-práth-a du-druv 'to run' ta-tsar 'to approach stealthily' ka-kjad-a 'to divide' ma-mná:-u 'to note' 'to be sticky' si-çnih 'to smile' si-çmi 'to flow' su-jru 'to sound' sa-svár-a 'to leap' ka-skánd-a 'to prop' ta-stámbh-a 'to praise' tu-çtu 'to burst' pu-$phut h 'to contend' pa-spfd

These data may be accounted for by making the initial /s/ extrametrical, as Steriade proposes, so that it does not map onto the template. Note that it must be stipulated that / s / is only extrametrical in case a sonority violation would occur. This is necessary to account for the last few forms of (7)a, which start with / s / plus nasal, liquid or glide. Under the analysis of / s / plus stop clusters proposed here, there is an alternative approach to Sanskrit perfective reduplication, which does not make use of extrasyllabicity. If / s / plus stop clusters are segmental units, these facts in Sanskrit can only be analysed as the deletion of the part of the segment that is unmarked for Place. A stop at the original place of articulation is predicted to remain: C

(8) [coi

[stop] I Place The deletion of the [cont] part of the /s/ plus stop cluster may seem ad hoc. However, parallel cases of reduplication in which other classes of complex segments lose their non-head part exist. That is, in affricates the stop part may be deleted in reduplication, and in prenasalised stops the nasal part may be subject to loss, and so on. In all cases loss of complexity results, which may be regarded as the underlying cause of the segmental change. Let us briefly consider some more cases of simplification of segmental structure. Consider first the case of Nisgha, a native American language spoken in the Nass River Valley and coastal area in northwestern British Columbia, described by Shaw

183 (1987, 1991). In this language final affricates / t s / and lateral affricates /t4-/ lose their stop part when mapping onto the (templatic) prefix (stress omitted): (9)

pats q'uts ts'atf' q'ati'

pis-pats q'as-q'uts ts'if-ts'at*' q'al-q'atf'

'to lift, carry something' 'to cut something' 'music record-SG/to have rippled surface-PL' 'to be slightly crooked'

To explain this fact, Shaw (1991) represents affricates as [ - c o n t ] underlyingly, and leaves fricatives unspecified for continuancy. She proposes that during reduplication [ - c o n t ] is delinked, so that the segment that remains is interpreted as a fricative. There are two problems with this solution. First, final oral stops do map to the template, so that the analysis must specifically separate stops from affricates. Second, it seems questionable that [ - c o n t ] is the 'marked' value for continuancy, in view of the preponderance of stops over fricatives in language inventories. Facts from language acquisition, in which fricatives are consistently replaced by stops (at least in certain positions of the syllable; see Fikkert 1994) point towards the same conclusion. In the framework of Manner-Place dependencies developed here, what happens to Nisgha affricates under reduplication is that they lose the part unmarked for Place. This is shown in (10)a, where it can be seen that a fricative at the original place of articulation is the predicted result. In (10)b it is shown how lateral affricates are affected (for the representation of lateral affricates, see § 3 in Ch. 5). Here the process is the same and results in a lateral fricative. We must assume that the original sonorancy value of the whole segment, that of an obstruent, is retained. See § 2.1 in Ch. 5 for two possible ways in which this might be achieved: (10)

a.

root

b.

[stop] [cont]

ι

Place

root [stop] [cont]

r \

[stop] Place

Thus, both in Nisgha and in Sanskrit complex segments are liable to simplification under reduplication. In both cases simplification takes the same form: the part of the segment that is unmarked for place is deleted. The analysis here in some ways resembles the analysis of the Nisgha facts presented by Shaw (1987), who discusses a suggestion by McCarthy and Prince (1986: 17) to regard the least sonorous member of a branching structure as the head. Shaw notes that such an analysis, while it works for Nisgha, does not generalise to the Sanskrit case. In Nisgha, the fricative part of the affricate, which is the most sonorous part of the segment, is retained, while in Sanskrit it is deleted. She therefore proposes (Shaw 1987: 301) that "the righthand branch of a contour segment functions as head", pointing out that in morphology too structures have been proposed to be right-headed

184

('The Right-Hand Head Rule', Williams 1981).4 This definition of headedness leads to the same results as the one adopted here, in which the part of the segment that is specified for Place in underlying representation is regarded as the head (see § 3 in Ch. 5). Both these analyses generalise over the Nisgha and Sanskrit cases, because the (lefthand) stop parts of affricates are deleted in Nisgha while the (lefthand) fricative parts of / s / plus stop clusters are deleted in Sanskrit. Of course, it is not obligatory that affricates or / s / plus stop clusters lose their 'weak' parts under reduplication. The facts provided above indicate that if reduplication simplifies affricates, these may only lose the stop part, while / s / plus stop clusters may only lose the continuant part. Hence, it is possible to propose a 'complex segment simplification' parameter: (11)

Complex segment reduplication: Delete non-head Default: no

Under the default setting complex segments reduplicate as complete segments. If the parameter is set to 'yes', however, complex segments are split up. The parameter setting then determines that it is the non-head part, i.e. the part not specified for Place underlyingly, that is lost. Recall that in Sanskrit perfective reduplication clusters like /pr-/ and /ks-/ are simplified to / p - / and /k-/ (see (7) above). In various phonological frameworks such as Dependency Phonology and Government and Charm Phonology (Kaye, Lowenstamm and Vergnaud 1985, among others) the internal structure of clusters also involves a difference between heads (such as obstruent stops) and dependents (such as sonorants). Hence, the strategy of deleting non-heads is also applicable to clusters. Finally, Shaw (1987) mentions simplification of diphthongs in Ewe reduplication as a case in which branching structure is simplified. Consider the data in (12) (from Sagey 1986a: 86): fia bia avo-sia fie kplo gbla jira

4

fa-fia-a ba-bia-m avo-sa-sia fe-fle-e kpo-kplo gba-gbla-m jia-jira-la

'burn vs. burnt' 'ask vs. asking' 'cloth-dry vs. cloth-drying' 'buy vs. bought' 'lead vs. leading' 'exert onself vs. exerting onself 'rave vs. a raver'

For some languages (e.g. Turkish) it has been observed that there is an alternation of headedness from the morphology (or even higher) downwards: whereas compounds are right-headed (since the main constituent is on the right), words are left-headed (since the stressed syllable is on the left), syllables are right-headed (since the nucleus is on the right), and onsets are left-headed (since the first member 'governs' the second, as in Kaye, Lowenstamm and Vergnaud 1985). Next in line is the segment, which by this rationale would be right-headed. However, note that the left-headedness of words can only be demonstrated for languages with initial stress, such as Turkish.

185 In the data in (12)a the righthand member of the diphthong is preserved under reduplication. This provides support for the definition of heads that Shaw proposes, i.e. that in which the righthand part of a complex segment is the head. However, the set of data in (12)b cannot be explained in this way, since here the lefthand part of a cluster is retained. Hence, the notion of head of a cluster must be defined differently from that of head of a segment (cf. above, and note 4), or these data must be accounted for in yet another way (for instance, by appealing to the 'core syllable', as Shaw proposes, following McCarthy and Prince). 5 Finally, consider one of the extremely rare cases (known to me) in which prenasalised stops are split up in reduplication. On the basis of the above, it is predicted that the nasal part would be subject to loss in reduplication. A reduplication process with this effect appears to be operative in Ninzam, a language of Nigeria. Faraclas and Williamson (1984) report (on the basis of unpublished work by Hörner 1980) that in reduplication "prenasality is sometimes lost in word-initial syllables when the consonant is a voiceless labial" (Faraclas and Williamson 1984: 3), and provide the first example in (13), from the Amar Tita dialect. 6 (13)

i-nfinfitu à-npùnpùzà

finfítú pùnpùzà

'hair of head (PL vs. SG)' 'thigh (PL vs. SG)'

Here prenasalisation is lost from the initial prenasalised labial fricative. A larger body of data would clearly be welcome, but tentatively it supports the generalisation that in segments which are complex for Manner, the part of the segment not specified for Place may be subject to context-free deletion.

22 Partial geminates In many languages that do not allow branching onsets in the native vocabulary loanwords with initial / s / plus stop clusters are treated differently from loanwords with other onset clusters. Either they are incorporated unchanged, or, as Broselow (1991) notes, prothesis applies instead of epenthesis. Thus, these clusters behave like 'partial geminates': it has been noted that epenthesis cannot sever the two parts of long or geminate consonants (see Schein and Steriade 1986; Hayes 1986). Consider the following data from Modern Standard Hindi (Singh 1985). In (14)a below 'ordinary' clusters are broken up, in (14)b it is shown that / s / plus stop clusters receive a prothetic vowel instead (the last case of (14)b is the only example of an / s / plus fricative cluster that Singh gives; unfortunately, there are no examples of / s / plus nasal clusters):

5 6

In chapter 1 (sc. 1.3.2) some facts of Nupe reduplication were presented, in which vowels with more than one Place element simpliGed by deleting one of the Place elements. The second example in (13) was provided to me by Prof. Gerhardt of the Seminar für Afrikanische Sprachen und Kulturen of the University of Hamburg. This help is gratefully acknowledged.

186 please pfizer frock fruit flirt cloth slipper sleeping-bag

/pikz/ /pafazsr/ /firDk/

/firut/ /fllAlt/ /kib0/

/siliper/ /silipiq basg/

b.

spelling spoon speech stool station school screw sphere

/ispeliq/ /ispun/ /ispitj/ /istul/ /isfejan/ /iskul/ /iskru/ /isfiar/

Broselow (1991) reports similar data for Bengali, Central Pahari, Sinhalese, Egyptian Arabic, Amharic, Wolof, and Telugu. Let us consider the data from Sinhalese in some more detail (Sannasgala 1976). In this language, which is spoken on Sri Lanka, clusters in loanwords from Dutch were broken up along roughly the same pattern as those in Modern Standard Hindi above. Loanwords with ordinary clusters are broken up by epenthesis ((15)a), and / s / plus stop clusters remained either unchanged or received a prothetic vowel ((15)b):7 Dutch plan procuratie prop troef trap klamp kraan vrouw spatie stai stoep strijkijzer schroef schop schoppen

Sinhalese päläna perakaläsi-ya poroppa-ya turumpa-va tarappu-va kalampa karäma-ya porova ispäsu-va stäla-ya/istäla-ya stöppa-va/istöppa-va stirikka-ya/istrikka-ya skuruppu-va/iskuruppu-va sköppa-ya/isköppa-ya sköppa

'plan' 'procuration' 'cork' 'trump' 'flight of stairs' 'clamp' 'tap* 'queen (cards)' 'space' 'stable' 'footpath' 'iron' 'screw' 'spade' 'spades'

Note that the Dutch /sx/ clusters in the last few items are replaced by / s / plus stop clusters.8 Unfortunately, again, there are no examples of loanwords (neither from Dutch, nor from Portuguese or Malay) that would show how / s / plus nasal clusters are broken up.

7 8

Compare also Sinhalese iskoolee 'school', isfeesama 'station', ispiritaalee 'hospital' (from Reynolds 1980). In fact, I have no information about the pronunciation of sch- in school, schop(pen), etc. in the beginning of the seventeenth century, when contact between Dutch and Sinhalese existed.

187 Data such as those from Hindi and Sinhalese form, in my view, a precise indication of the phonological nature of / s / plus stop clusters: they are monopositional, but complex segments. Both in the case of ordinary clusters and / s / plus stop clusters the onset branches, albeit at different prosodie levels: (16)

a.

branching within the onset: Onset / \ χ

b.

χ

skeleton

branching within the segment: Onset I χ

skeleton

C

root

/\

In languages which do not allow any kind of branching in the onset simplification of both branching structures must take place. However, in the case of complex segments ((16)b) the usuál strategy to achieve this, viz. epenthesis, is not allowed, because this would violate the line crossing constraint (Goldsmith 1976), as shown in (17): (17)

*

C

V

,[cont] A / ' •' Place

[stop] I Place

To achieve the desired simplification, prothesis takes place, after which resyllabification can dissolve the complex segment into two separate syllables (possibly dividing the complex segment into two separate segments). This is shown in (18):

188 (18)

a. insertion of a rhyme and projection of a vowel: χ I R

χ I O

i V

I

c

[cont]

\ [stop] I Place

b. resyllabification: σ

σ

I R I X

O I X

ν

ι c

/

As a result of (18)b, the complex segment is not split up, but becomes ambisyllabic. It is therefore phonetically interpreted as a true cluster. I conclude that the behaviour of /s/ plus stop clusters as partial geminates, noted by Broselow (1991), argues in favour of the representation of these clusters as single segments. Note, finally, that a representation of /s/ as initially extrametrical cannot easily account for facts such as those discussed above. In such an approach there are a number of possibilities in a language which has a prohibition on clusters. First, if / s / is underlyingly outside the syllable and is allowed to remain extrasyllabic, there is no cluster that needs to be broken up, so that no epenthesis (or prothesis) is expected. Second, if / s / does become incorporated in the first syllable of the word later in the derivation, there is no reason to assume that the clusters that are created in this way should be treated differently from other clusters. Finally, if / s / is allowed to be incorporated as a separate syllable (as in Itô's 1989 Persistent Syllabification approach), it would be expected that epenthesis or prothesis would occur in accordance with the direction of syllabification in the language concerned; in any case / s / plus stop clusters would, again, not be expected to be treated differently from other clusters. Hence, the extrametricality approach does not make the prediction that / s / plus stop clusters should be treated differently from other clusters in simplification. Either no epenthesis is expected or it is predicted that the two types of clusters should be

189

treated on a par. Additional stipulations are necessary to account for the different treatment of /s/ plus stop clusters and other clusters in languages such as Hindi where prothesis takes place.

2.3 Other evidence In this section we briefly discuss some other kinds of considerations with respect to / s / plus stop clusters, which also suggest that these clusters are best analysed as single segments. These concern alliteration, language games, vowel-glide alternations and the acquisition of / s / plus stop clusters. Kuryiowicz (1966: 195) notes that in the inherited Germanic verse a consonant followed by a vowel may alliterate with itself followed by another consonants. Thus words in initial ρ alliterate with words in initial pr-, pl-\ h alliterates with hr-, hl-, hn-, etc. There is an important exception to the above rule. The clusters sk-, st-, sp- are not allowed to alliterate with each other or with words beginning with s not followed by a consonant; sk can alliterate only with sk, st only with st, sp only with sp9 Consider the following lines (taken from Kuryiowicz 1966): (19)

Beowulf Heliand

212 288 173

on síefn síÍ3on; síréamas wundon scearp scyldwÍ3a 3&scád witan spràca bilôsit, thoh hê spâhan hugi

These lines show that / s / plus stop clusters only alliterate with themselves in this kind of verse, and is an indication that they may be analysed as monopositional. Kuryiowicz notes that in other verse, however, the possibilities for alliteration are less restricted. This indicates that /s/ plus stop clusters need not be analysed as single segments. We return to this typical variability below. Hammond (1990) analyses a name game in English, in which the onsets of names are replaced by single segments [b] or onsets starting with [b].10 On the basis of his data, he distinguishes three 'dialects' of forming the game names:

9 10

See Hogg (1992: 45) for the same observation and the suggestion that "the groups /sp, st, sk/ filled a single slot in the onset structure, as in [present-day English]" (Hogg 1992: 45). Thanks to Carlos Gussenhoven for pointing this case out to me.

190

dialects Steve Scott Bruce Claire Gwen Kyoko Beula

(20)

1 b b b b b b b

2 b b b b b by by

3 b b br bl bw by by

Note that speakers of dialect 3 treat the /st/ and /sk/ of Steve and Scott differently from the onsets in the other names: they are replaced by single-segment onsets. Again this suggests a monopositional status of these onset clusters.11 Third, consider French vowel-glide alternations (Clements 1990b). An underlying vowel becomes a glide if it fits into an onset, i.e. if there is only one consonant preceding it ((21)a). If there are two consonants, the vowel remains and a diphthong results ((21)b). If an / s / plus stop cluster precedes, the result is variable: either the vowel remains or may become a glide ((21)c): (21)

a. b. c.

Ij0 (*li0) plie (*plje) skie ~ skje

lieu plier skier

'place' 'to fold' 'to ski'

These data suggest again that / s / plus stop clusters may be interpreted as single segments. However, the last few cases are a clear indication of the Variability' of the single segment status of / s / plus stop clusters: not all verse and name game dialects treat them the same, while the French rule for glide formation is variable. Of course, / s / plus stop clusters also have the duration of non-suspect clusters. This leads Ewen (1982: 49) to suspect that /s/ plus stop clusters are truly clusters, but that there is (at least) a stronger internal cohesion between the segments in /s/ plus stop clusters than between the segments in other clusters: "the elements of [/s/ plus stop clusters] are, at least, more intimately linked than those of normal consonant clusters, whether or not this means that they should be interpreted as single segments". The formal framework developed in this study is not equipped to express the difference in degree of cohesion that Ewen notes. As a final point, however, note that there are (at least) two possible well-formed phonological representations of a phonetic [st] sequence:

11

The question of whether /s/ plus stop clusters act as units in speech errors, such as slips of the tongue, cannot be answered on the basis of the available data. We would expect that these clusters are more frequently transposed as a unit, and that the two parts are less easily severed than ordinary clusters. However, from the data that are known it is clear that if these tendencies exist, they would not be absolutes, as examples like that's a sticky point -»... spicky point, split pea soup -» plit spea soup (from Fromkin 1971: 31f.) show (while affricates always transpose as a unit, and are never severed, as observed in sc. 2.1.3 in Ch. 7).

191

one is the structure discussed in this section, but another has two separate root nodes and timing slots: (22)

a. C / \ [stop] [cont] I ι

b.

C I [stop]

C I [cont]

I ι

I ι

The fact that both are well-formed representations (although the second one violates the sonority hierarchy, when it occurs in the onset of the syllable) indicates that it might not be inconceivable that different languages or varieties may analyse phonetic [st] sequences differently, perhaps as a result of different patterns in the language or variety concerned. Fikkert (1994) notes that children acquire either / s / plus stop clusters or obstruent plus liquid clusters as their first clusters (Fikkert 1994: 109).12 She proposes to analyse the initial / s / in any / s / plus consonant cluster as extrasyllabic, and rejects the hypothesis that / s / plus stop clusters are unit segments. To account for the two different courses of development, the child would then either first learn extrasyllabicity (resulting in / s / plus consonant clusters), or would first learn to make a branching onset (resulting in obstruent plus liquid clusters). I would propose that /s/ plus stop clusters are also unit segments in child language. On this view, the two 'learning paths' could be stated as two different loci where branching is learned. This would be either the prosodie level (resulting in branching onsets of the obstruent plus liquid type) or the segmental level (resulting in branching segments of the /s/ plus stop cluster type). This accounts for the observation that there is a difference in acquisition between /s/ plus stop clusters on the one hand, and / s / plus fricative or nasal clusters on the other. It makes it possible to restate the two learning paths as the setting of a parameter whose options are naturally related.13

3. Ordering In chapter 7 we noted some cases in which affricates behaved as if their parts were phonologically unordered. If / s / plus stop clusters are a kind of 'reversed affricates' (since their Place specification is under the [stop] Manner feature), the same kind of

12 13

Thus, contrary to what Fikkert (1994: 117) maintains, /s/ plus stop clusters do not have the same status as /s/ plus fricative clusters. This hypothesis might entail that in a language like English, which has both affricates and /s/ plus stop clusters, complex segments are learned roughly around the same time. This hypothesis cannot be checked for Dutch.

192 effects might be expected for these segments. However, no such effects are known to me. The only type of ordering effect in which /s/ plus stop clusters take part are mispronunciations, i.e. their parts are especially prone to be pronounced in the wrong order: e.g. Dutch [weps] for wesp Svasp', and [geps] for gesp 'buckle' (Ewen 1980: 252). Such mispronunciations can be attributed to a process of coda simplification, if coronals like [s] occur outside the coda, or in a separate metrical constituent known as the 'appendix' (see e.g. Fujimura and Lovins 1978; Charette 1984; Kager 1989; see also note 1 above). Under such a conception the change from [wesp] to [weps] can be formalised as in (23): (23)

σ

ε

s7 ρ

Since the stop part and the continuant part of the / s / plus stop cluster are not ordered underlyingly, it is not surprising that metathesis processes like this affect / s / plus stop clusters rather than other clusters.

4. Conclusion On the basis of the behaviour of / s / plus stop clusters in syllabification and a variety of phonological phenomena, we can conclude that at least in some cases there is evidence that they behave as single segments. A representation of / s / (or coronal segments in general) as extrametrical cannot easily account for all of these cases. However, there is a certain degree of variability, and there is no reason to dismiss the extrametricality approach completely. The variable behaviour was related to the fact that a phonetic signal [st], for instance, may be phonologically analysed as either a complex segment or as a true cluster.

Chapter 10 Two-root complex segments Consider the following representations of simple and complex consonants and vowels: (1)

C

ι

C

/\

V

[stop]

[stop] [cont]

I

I υ

V

ι

/\ A

U

I ι

simple consonant

complex consonant

simple vowel

complex vowel

A crucial difference between vowels on the one hand and consonants on the other is that consonants have at least one Manner feature, and that vowels lack such features (except that they may have [nasal]; see § 3.2 below). Note that there is a certain redundancy between major class and Manner here: the V-status of root nodes that dominate no Manner features can be predicted, or, alternatively, if a root node is marked as vocalic, it may not dominate Manner features. I will not attempt to answer the question which of these two pieces of redundant information should be omitted, or, in other words, what follows from what. In this chapter I turn to a category of complex segments which are different from the type of representations in (1). These are complex segments consisting of two root nodes dominated by a single timing position, i.e. involving the type of complexity referred to as 'two-rootedness' in chapter 3. An example of a two-root complex segment given there was a consonant with secondary articulation. Consider the representation of / k w / in (2): (2)

χ

timing tier

/\ C

V

I [stop]

I U

root node tier

I A Λ7 In this chapter I will examine such structures in further detail, with special attention for the representation of root nodes. In (2) these are labelled as C and V. This notation was adopted in § 1.1 in chapter 1, and below it will be discussed in more detail. Note that no claim is made about the relation between the number of root nodes and phonetic length.

194 In current models of phonological representation the assumption is made that the unity of phonological segments should be expressed by a node which dominates all features that make up a segment. This node is usually referred to as the root node. The expression of length, as in long vowels or geminate consonants, is more controversial. There are two main approaches. The first assigns single timing slots to short vowels and consonants, and two timing positions (or 'x-slots') to long vowels and consonants (see, for instance, Clements and Keyser 1983). The Dutch stem zee 'sea', which has a long vowel, is represented in (3) by way of illustration (structure between the syllable node and the timing tier omitted): (3) timing tier root node tier segmental content The second approach to length expresses timing in terms of moms and is syllablebased: a syllable which either has a long vowel or is closed has two moras and a syllable with a short vowel has one mora (see e.g. Hayes 1989). The representation of geminate consonants in moraic theory is less straightforward, and will not concern us here. Two possible moraic representations of Dutch zee 'sea' are given in (4): (4)

a.

σ

b.

σ

Λ

syllable tier mora tier

C

V

C

V

root node tier

I ζ

I e

ζ

e

segmental content

In these representations, the onset consonant may be linked to either the first mora or to the syllable node; these are the variants (4)a and (4)b, the latter of which is adopted by Hayes (1989). In the first approach to length it is possible that the association between x-slots and root nodes is not one-to-one. That is, not only is it possible to associate two timing units to a single root node (as in a long vowel or consonant), but it is also possible - at least in principle - to associate two root nodes (whether consonantal or vocalic) to a single timing slot, which results in phonological complexity. The consonant with secondary articulation in (2) above is an example. I will investigate these complex segments by focusing on the major class specifications that are marked on these root nodes. 'C' is used to refer to a root node marked

195

as [+consonantal], and 'V' to refer to a root node marked as [-consonantal]. This follows McCarthy (1988) in marking major class specifications not on the timing tier, but on the root nodes themselves. I will assume that these roots consist of the major class specifications: the features are therefore not associated to the root nodes. Below I address the question of whether [± consonantal] is a binary feature or whether two unary features (referred to as [cons] and [voc]) exist (see § 3.1); for the sake of presentation I use 'C' and 'V' as abbreviations for these specifications. Since there are two types of root nodes, three types of two-root complex segments are predicted to exist, provided in (5) below: (5)

a.

χ

b.

/\ C

C

χ

c.

/\ C

V

χ

/\ V

V

The representation in (5)b is especially interesting since the two root nodes dominated by the timing slot are phonologically unordered if the root nodes consist of the major class features [cons] and [voc], whereas they are ordered if the major class feature is a separate feature that is linked to the root node, or if the major class dimension is expressed by a binary feature [± consonantal]. In the two other cases ((5)a and (5)c), the root nodes are ordered under either hypothesis. Since the timing tier is part of the representation of two-root complex segments, it is expected that processes affecting length can play a role in their (diachronic) formation. Specifically, the hypothesis is that two-root complex segments can be formed diachronically by 'compressing' two segments under separate x-slots into one, retaining all their specifications. I will refer to such processes as 'compression':1 (6)

Compression: X

X

X

X

or root

root

root

root

We will see that compression processes play a role in the formation of at least some of the two-root complex segments to be considered. A large part of the discussion of the (5)a type of representations will be devoted to clicks. Other segments that are candidates for this representation are labialcoronals and postnasalised stops. These will be more briefly dealt with.

1

In (6) the root node is delinked from the timing slot. Delinking of the root node from the timing slot does not always have to take place, however. For instance, in the formation of a consonant with secondary articulation the vocalic root node is not delinked, perhaps because of the prosodie demand that the nucleus may not be vacated. This will be illustrated below (sc. 3.1).

196

1. Two consonants under one x-slot I repeat (5)a as (7): (7)

χ

/\ C

C

Single segments such as affricates are expected to occur in initial and final position in the syllable. The same prediction is made for x-slots such as those in (7). On the phonetic surface this amounts to the impression that identical consonant clusters appear in initial and final position. Observations like these have been made in a number of cases, so that in these cases there is reason to assign such clusters some kind of single-unit status. There are - pre-theoretically - no constraints on what kinds of C (or V) segments may be attached to a single x-slot (as was pointed out in § 2.1.1 in Ch. 7 above). In the following subsections I will consider the following types of segments as candidates for being x-slots that dominate two consonantal roots, namely clicks (§ 1.1), labial-coronals (§ 1.2) and postnasalised stops (§ 1.3).

1.1 Clicks

Clicks, which almost exclusively2 occur in a range of southern African Khoisan languages and a number of neighbouring Bantu languages where they are loan segments, are usually analysed in terms of an influx and an efflux (see Chomsky and Halle 1968: 319ff., who follow Beach 1938 in their terminology, based on studies of clicks in Hottentot). Under the influx of the click are subsumed the features that are relevant for the primary constriction; all other click features are subsumed under the heading efflux. The latter comprise such features as glottalisation, aspiration and nasalisation, and also combinations of these such as "weak unvoiced velar plosive" and "velar glottalic affricative" (see Chomsky and Halle for discussion). Examples of influx types are the Hottentot dental affricate, the denti-alveolar implosive, the lateral affricate and the alveolar implosive. It may also be a bilabial. These primary articulation types can each be combined with a type of efflux (for general descriptions see Jakobson 1968; Ladefoged 1971; Ladefoged and Traill 1984; Sagey 1986a, among others).

2

Clicks have also been reported for at least one Australian language, namely Damin, the secret jargon of initiation used by speakers of the Australian language Lardil, which "has a number of consonantal segments which are totally lacking elsewhere in Australia: among these are nasalised clicks (bilabial/m!/, dental/nh!/, apico-alveolar/n!/, apico-dorsal/n!/) (...)." (Hale 1973: 443). Ponelis (1992) maintains that clicks were also borrowed into northwestern dialects of Afrikaans. Data are unavailable, however.

197 Chomsky and Halle (1968: 322) support the use of the term 'primary constriction' by the fact that any of these constrictions may be combined with any of the efflux types in the Korana dialect of Hottentot. However, this argument is vacuous for languages in which any of the effluxes may be combined with any of the influxes such as Korana. The fact that different dialects of Hottentot differ as to which effluxes are present may be an argument to regard the influxes as primary. A reason to classify the efflux as primary might be that there are more effluxes than influxes. I return to the question of which articulation in clicks is primary below. Let us now turn to the phonological representation of clicks, focusing on their duration and complex character. First I will briefly discuss the representation of clicks advanced in Sagey (1986a). Then I will make a counterproposal on the basis of observations made by Traill (1981, 1992), who proposes that clicks are not unit segments but clusters. A two-root representation of clicks will then be advanced as a compromise between the two positions: this makes it possible to retain the traditional insight that these complex phonetic events are unit segments, and resolves Traill's objections to Sagey's representation of clicks. Sagey (1986a: 126ff., 257ff.) assumes that clicks have a single timing slot and also a single root node. Her representation of an alveolar click is provided in (8) (see Sagey 1986a: 258): (8)

Root

[-coi Place

/

Dorsal The pointer from the root node to the dorsal node is only necessary in a language which has both (egressive) coronal-velars and clicks, such as !Xü (Sagey 1986a: 257f.). Sagey argues that the pointer is independently needed, i.e. also for complex segments not involving suction. The pointer is also used to distinguish between Nupe /p w / and /k w /, for instance (see Sagey 1986a: 216):

198

(9)

Root

Root

Dorsal

Dorsal [ + round]

[ + round] [+back]

[ + back]

It is impossible to carry over Sagey's representation of clicks into the framework developed here, since segments with a coronal and a dorsal node are interpreted as palatals or palato-alveolars (see Ch. 6). Headedness (which is the obvious translation of the pointer mechanism employed by Sagey) of either the coronal or the dorsal node is interpreted as a distinction in place of articulation, not as a difference in airstream mechanism. Still, Sagey's idea of abandoning the special suction feature is very attractive, as it represents the logical endpoint of a line of research starting with Halle and Stevens' (1971) proposal to reduce the number of laryngeal and airstream features. Below, I will show that a special suction feature is not necessary under the approach advocated here either. I propose that clicks are represented as two-root complex segments, with two consonantal root nodes under a single timing slot. The reason to diverge from Sagey's representation with a single root node is not only a theory-internal one. It has been argued that clicks are not single-root segments, namely by Traill (1981, 1992), who proposes that they are clusters. Traill (1981) notes, first, that a cluster analysis of clicks removes the crosslinguistic 'oddity' of the Khoisan languages, whose segment inventories are enormous when clicks are considered as unit segments, but can be cut down to ordinary size when the click parts are considered to be separate segments. Second, he notes that a sequence of clicks removes the need for ad hoc features like [delayed release of secondary closure] (Chomsky and Halle 1968) or 'implosive vs. infricative' (Snyman 1976). Third, Traill (1981: 135) claims that a sequence analysis permits a proper description of phonetic identities between clicks and non-click segments. He points out the fact that in languages like Korana [x], [kx?] and [?] occur both as click effluxes and as independent consonants. Moreover, the presence of an efflux implies the presence of the independent consonant, which remains unexplained under the unit segment analysis. Traill (1992) adds: "in those Khoisan languages undergoing changes to the click system, changes affecting the effluxes extend to the independent consonants. Thus in Nama, both the efflux [kx'] and the independent consonant [kx'] have been replaced with ['] suggesting that a single change has taken place involving only

199

one element. And !Xü lacks both uvular effluxes and independent uvular stops, whereas !Χόο has both." Finally, phonetic evidence also suggests that clicks are different from ordinary complex segments such as affricates: "the mean duration of basic clicks [in !Χόό] is 175 ms., only 3 ms longer than basic pulmonic stops (...). They are slightly shorter when the effluxes q, x, follow. The duration of click + q is almost identical to that of the basic click (only 4% longer), but the duration of click + χ and click + ' is one and a half times longer. These differences do not follow from the assumption that all click + efflux complexes are associated with a single timing slot. In fact, it is known that the duration of a cluster of two consonants is between one and a half and twice the duration of a single consonant and the duration of a unitary complex consonant (i.e. one associated with a single timing slot) is comparable to the duration of a single consonant (Maddieson and Ladefoged 1989). The mean duration of the contour segment /ts/ in !X6ö is 158 ms. which suggests that this phonetically complex segment is controlled by a single timing slot and is therefore unlike the superficially similar click + x." Thus, a single-segment representation of clicks has a number of disadvantages. However, I will not follow Traill and adopt a sequence analysis of clicks. As Traill himself notes, this would not do justice to the traditional intuition that these are single segments. A second argument to regard clicks as single segments is that they were borrowed as whole segments in neighbouring languages (see note ? above). Finally, a cluster analysis would merely transfer the oddity of Khoisan languages to another sphere: instead of languages with enormous consonant inventories they would be languages with anomalous syllable types. What I propose is a compromise representation of clicks in which these are tworoot complex segments. That is, the influx and the efflux are represented on separate root nodes, both of which are consonants. In this conception, the alveolar click, for example, can be represented as follows: (10)

χ

/\ c c I I [stop] [stop] I I I A

Either part of the click may be affricated, i.e. involve multiple stricture (see Maddieson 1984: 217). The representation in (10) solves a problem that Traill notes with regard to the laryngeal specifications of different parts of the click: both parts may be separately specified for laryngeal features. However, it does not go so far as to claim that clicks consist of entirely different segments, which would be on separate x-slots.

200 In such a conception of clicks, special suction features may not be necessary, just as in Sagey (1986a). Similar to Sagey's stipulation that a corono-dorsal with a pointer to dorsal is interpreted as a click, it has to be stipulated here that suction may take place in a segment with two C-roots, one of which dominates a dorsal stop. The issue of headedness in segments such as (10) can also be resolved in a principled way. Recall from chapter 4 that the Mismatch condition dictates that the dependent part of a complex segment must be simple, i.e. have a single Place element at most. This means that the influx part of the click must be phonologically primary, since here palatals are allowed, as well as finer distinctions in the coronal area. This in turn implies that the efflux is secondary (just as has been traditionally assumed; see above), and predicts that it must be simple in terms of Place. This is correct: the efflux can only be located at the velum (and is therefore predictable specified as A), although it may be accompanied by glottal or supraglottal frication. The same segment with headedness reversed may then be interpreted as the corono-velars of !Xü (recall that the presence of I and A under one root node represents palatals or palato-alveolars). If these elements are under separate root nodes, the relation between both Place elements is understandably looser. The representation of clicks as in (10) leads us to consider the question of whether clicks can have secondary articulation types like labialisation. Clicks with secondary labialisation would have the structure in (11): (11)

χ / N C C V I υ

Such a structure is incompatible with the hypothesis that complex phonological structure is binary branching (see § 2.1 in Ch. 3; see also van de Weijer 1992a: 268). The prediction that clicks cannot have secondary labialisation is borne out for the languages surveyed in Maddieson (1984). It is also true for the / / A n i language studied by Voßen (1986). However, velarised clicks are reported in Maddieson (1984). It appears to be the case, however, that this velarisation refers to a fricated release of the velaric closure (see Maddieson 1984: 419). Hence, it seems that on the basis of the available sources it does not seem necessary to allow for underlying secondary articulation on clicks. Another point that bears on this issue is that clicks can be either prenasalised or fully nasal. Probably, prenasalisation in a click would have to be represented as prenasalisation of one of the parts of the click, as in (12):

201

χ

(12) C / \ [stop] [stop] [nasal]

C α

ß

Full nasalisation would imply that either both parts of the clicks are specified for [nasal] or the existence of a separate [nasal] specification, perhaps attached to the xslot. Let us turn to the evidence that clicks are specified for two (sets of) Place elements. In !Xóò (Traill 1985), the dental clicks behave as both coronal and dorsal with respect to phonological processes and restrictions. Sagey (1986a: 127ff.) shows that the dental clicks front and raise the vowel / a / in specific environments. The language also has a constraint which requires back vowels after dorsal consonants, including clicks. Finally, Sagey (1986a: 130) mentions that the first person singular pronoun [ή-] and the verbal formative [-ή-] assimilate to the coronal articulation of the click. These facts show that clicks are simultaneously coronal and dorsal. This forces us to assume that I and A above must be unordered, although they are dominated by separate (ordered) root nodes. The historical evolution of clicks is uncertain. 3 It is usually assumed that there is hardly ever a historical relation between clicks and non-clicks, although it has also been suggested that there may be a relation between the bilabial click and a labialised bilabial (note that both are two-root complex segments in the framework developed here). There are also clear cases in the Khoe languages of clicks shifting to non-clicks synchronically. The complete alveolar click series ([!], etc.) is replaced by velars, and the complete palatal series ([φ], etc.) is replaced by palatals. Traill (p.c.) posits that similar shifts from non-click to click or vice versa might have taken place in the past, but that the data necessary to allow such a reconstruction are not available. Finally, Ladefoged (1968: 9) describes three ways of articulating labial-velar stops in West-African languages, two of which involve a (partial) velane airstream mechanism. Consistent with the approach above would be that clicks were formed from sequences of segments by the process referred to above as compression: (13)

χ

χ

C

C

X

/\ c

c

However, as was pointed out, the data do not allow a certain reconstruction.

3

I thank Tony Traill (personal communication) for discussion on the points made in this paragraph.

202 In this section I have dealt with the representation of clicks, and proposed that they may be represented as two-root complex segments. Such a representation preserves the traditional idea that clicks are unit segments, and resolves the objections Traill (1981, 1992) raises to such an analysis.4 In the following subsections, a number of other categories of segments will be dealt with that are also candidates for representations with two consonantal root nodes under one timing slot.

L2 Labial-coronals Sagey (1986a) provides evidence that Margi (Hoffmann 1963; Ladefoged 1968) has a range of labial-coronal segments, which she represents as single-root complex segments. These include segments like /pT/, /bct/, /bz"/ and / n i n / (see Sagey 1986a: 176). In all such labial-coronal segments the two parts agree in voicing, but not necessarily in continuancy. The evidence that these segments are single segments is compelling. Margi has a reduplication processes which treats the labial-coronals differently from other clusters, which are derived by a synchronic syncope rule. Three reduplications involving singleconsonant clusters and genuine clusters are given in (14)a below. Examples illustrating the reduplication of labial-coronals are given in (14)b (from Sagey 1986a: 179): (14)

a.

b.

tapara s(u)kuda tjkwar(i) mjia bda ptî»

'to 'to 'to 'to 'to 'to

vomit' push' touch' rebuke' chew' wash'

tatapara ssskuds tj9tjkwar(i) mjiamjia bdabds ptjiptj9

'to vomit many times' 'to push bit by bit' 5 'to touch (many things)' 'to rebuke very much' 'chewed' 'clean, washed'

Maddieson (1983) argues that the labial-coronals of Bura, a close relative of Margi, are consonant clusters rather than complex segments. He argues that these segments are phonetic sequences of labial followed by coronal, on the basis of the fact that they have a longer duration than single consonants. Sagey (1986a: 180) notes that if his observation is also correct for Margi, then it would remove phonetic motivation for the representation of labial-coronals on single x-slots. However, Sagey points out that it is still necessary to represent them on a single x-slot in view of the phonological evidence consisting of the reduplication patterns discussed above. Again it is possible, and in my view appropriate, to propose a compromise representation with a single

4 5

Traill (1985) treats the clicks of !X6ó, a Bushman language in the same way; unfortunately this work is unavailable to me. Note that the /s/ plus stop cluster in this form must be analysed as two separate segments; only the /s/ reduplicates. The optional epenthetic vowel in this cluster bears out the variable status of this cluster.

203 timing slot and two root nodes. Segments such as / b z y , which differ in continuancy and Place, cannot be represented in any other way in the framework developed so far, and hence must be represented as two-root complex segments. The result is that a segment like / b z 7 is represented as in (15): (15)

χ c

c

I I [stop] [cont] I υ

I ι

A two-root representation accounts for the reduplication facts discussed by Sagey as evidence for the single-segment status of the Margi labial-coronals. Note that it must be stipulated that the two consonantal roots share the same laryngeal node. This is not unexpected, however, and does not constitute evidence for a single-root representation: obstruent clusters in many other languages also have to agree in voicing, and this has never been taken as an argument for single-segment status. Incidentally, the representation in (15) also sheds light on Sagey's observation that the Margi labial-coronals "were derived historically from consonant clusters created by syncope" (Sagey 1986a: 176f.). Hence, their historical derivation seems to be one in which compression played a role. In relation to this, I would like to speculate on a possible intermediate stage between a language with full vowels in a CV-structure and a language with double-root segments like the ones above. A possible development contains a stage in which there is an empty vowel between the two consonants: (16)

C I α

V

C I β J

V I 7

C I θ

—

I a

V

C

V I

y

I θ

0 The empty vowel position in between the consonants might continue to exist in underlying representations, possibly as an intermediate stage in which the two consonants are joined under a single timing slot. Evidence for the empty vowel stage comes from Leti (van der Hulst and van Engelenhoven 1994). This language has 'crazy onsets' (originating historically from the loss of a vowel between contiguous onsets), which are analysed by van der Hulst and van Engelenhoven (1994) as consisting of C followed by empty nucleus followed by another C, as in (16). It is important to note that the complex segments of Bura do not behave the same as the complex onsets of Leti, which reveals their different status. In Leti the

204 parts of the onset clusters can be split up under reduplication (unlike the case in Bura), as van der Hulst and van Engelenhoven show: (17)

tmuela tmuel mela ptiali -»· ptia tali

'dark' 'easy-going'

From the facts that complex segments in Leti can be broken up, I conclude that these onsets have a different status from the complex segments of Bura and Margi.

1.3 Postnasalised slops Anderson (1974a: 301) reports that postnasalised stops occur in South American and in some Malayo-Polynesian and New Caledonian languages. For a description of the occurrence of these segments see also Herbert (1986: 74), who refers to Capell (1967) on Aranda (which is also the only example of a language with postnasalised stops in Maddieson 1984). Finally, a detailed study of postnasalised stops appears in Poser (1979). I would like to suggest that postnasalised stops are also two-root complex segments. One reason is that postnasalised stops are also created historically by processes such as compression. Consider the case of Némi, a language spoken in New Caledonia, which has pre- as well as postnasalised stops (Rivière 1975: 347). The evidence that in this and neighbouring languages postnasalised stops are unit segments comes largely from the Canonical Pattern Fit argument (see § 1.1 in Ch. 8). This is the evidence that Haudricourt (1961: xi) provides: "The monophonematic character of this consonant resorts in the fact that there is never a cluster of consonants at the beginning of the word in these languages". On the origin of these segments Haudricourt notes: "From a diachronic perspective the origin of this type of consonant must be sought in old nasal infixes which are no longer productive, and in contractions such as the word for 'mother', *tina > fe". Nasal-initial words also receive a prothetic oral element when they are used in a locative sense: ηa 'house' > 0a 'in the house', maac 'reef, group of rocks on the waterside' > pmaac 'on the waterside' (my translations, JW). I have not been able to find further information on the sources or behaviour of these segments. If they are truly single segments (recall that the Canonical Pattern Fit argument is not sufficient to prove this), a representation like that of clicks or Bura labial-coronals may be appropriate. Thus, the structure proposed for [pm] would be as in (18):

205 (18)

χ / \ C C I I [stop] [stop] I [nasal]

In these segments there is only one Place node, since postnasalised segments have a single place of articulation. It is not immediately clear, however, if the Place specification should be attached to the left or the rightmost branch. I leave this issue unresolved.

2. Two vowels under one x-slot: short diphthongs Consider the representation in (19): (19)

χ V

V

The interpretation of (5)b, repeated here as (19), is straightforward. If single vowels are represented with one root node and one x-slot, two root nodes under a single xslot form a short diphthong. It is interesting to note that these segments in certain cases have developed in much the same way as the clicks, labial-coronals and postnasalised stops, namely by compression. In this section we will present some evidence for the existence of short diphthongs. Short diphthongs are rare in synchronic systems and may result from rules like compression. See, for example, Hayes (1990: 39): "incontrovertible cases of short diphthongs are found in Chevak Yupik Eskimo (...), Icelandic (...) and Faroese (...). However, none of these is the result of (spontaneous) diphthongisation (...); rather they involve an originally long diphthong, compressed in a closed syllable by the loss of a V position" (emphasis added, JW). Another language for which the existence of short diphthongs is well established is Old English. Very early in the prehistoric Old English period the front vowels /ae e i/ (both short and long) were diphthongised when immediately followed by a velar or velarised consonant or consonant group (see Gussenhoven and van de Weijer 1991: 315ff.; Hogg 1992: 84ff. for a description and analysis of the environment). The traditional name for the process is Breaking (Grimm 1822). Where the broken vowel is short, the resultant diphthong is equivalent in length to a short vowel, and hence the new contrast in Old English between long and short diphthongs first arises at the time of breaking. Hogg describes the scenario of the changes as follows: "between the

206 front vowels and the [velar or velarised] consonant groups there developed a transitional glide, which in the first instance would be a non-low, nonsyllabic back vowel, that is, / u / or / o / . At a later stage, certainly by the time of the earliest written texts, this glide developed into a non-syllabic back vowel of the same height as the preceding front vowel". T h e process can be analysed as 'pre-articulation' in the sense of Wetzels and Sluyters (1995). Wetzels and Sluyters document and analyse a complex of consonantvowel interactions in Maxakali, a native language of Brazil. They argue that all consonants in this language develop 'pre-glides', which then spread into the nucleus. T h e timing tier of lexical items is not affected. The process seems an exact parallel to that in Old English, with the proviso that in Old English only a limited set of consonants develop the pre-glide. The development of the pre-glide in Old English is shown in (20): (20)

χ V

A

C

T h e subsequent move into the nucleus is shown in (21): (21)

χ V

χ "v I A

c I A

T h e result of the process is a short diphthong as in (19) above. 6 Another language where short diphthongs arise is Finnish (Keyser and Kiparsky 1984; see also Steriade 1982: 24-27). In Finnish they are the result of morphological concatenation, rather than the result of phonological changes. Finnish does not allow long vowels or diphthongs in unstressed syllables. As a result of morphological concatenation, however, sequences of vocalic nuclei do arise. Keyser and Kiparsky (1984) show that such nuclei are adjoined under single timing slots, and hence become short diphthongs. One argument comes from the rule of f-deletion. In Finnish the ending of the partitive singular is -ta after consonants, and -a after vowels:

6

Note, incidentally, that this way of describing the process of Breaking does not rely on a formal distinction between a Place node and separate features, as Gussenhoven and van de Weijer (1990) argue the process does: the root node functions as the organising node that is necessary to distinguish between single-feature assimilation and class node assimilation.

207 (22)

a.

sammal-ta tul-ta vet-tä lun-ta vieras-ta

'm0SS-PART-SG' 'fire-PART-SG' 'water-PART-SG' 'snow-PART-SG' 'guest-PART-SG'

b.

lokero-a paperi-a talo-a

'pigeonhole-PART-SG' 'paper-PART-SG' 'house-PART-SG'

The ending -ta also appears after long vowels and after stressed short vowels: (23)

a.

vapaa-ta

'free-PART-SG'

b.

mi-tä

'what'

Keyser and Kiparsky formulate the rule of /-deletion as follows (where the syllable must also be unstressed): (24)

t -» 0 / C

V] s

V

The underlying structure of 'house-PART-SG' is talo + i + ta. It is clear that the combination [oi] does not behave like a long diphthong, since t is deleted (taloia, glided to [taloja]). Keyser and Kiparsky argue that the two segments [o] and [i] are joined under a single timing slot (see Keyser and Kiparsky 1984: 22): (25)

/\ V

V

I [ ]

I [ ]

ν / \

[ ] [

The slight difference between this formulation and the formalism adopted here is that the V-units of (25) correspond to timing slots (represented here as x) and the segmental material in brackets correspond to vocalic root nodes. A second piece of evidence for the structure of short diphthongs in Finnish comes from consonant gradation. The genitive plural suffix in Finnish is -den instead of -ten after a sonorant and before a vowel (Keyser and Kiparsky 1984: 15-16). Consonant gradation also degeminates geminates. However, after / l a k o + i / the inessive suffix -ssa is not degeminated. This indicates that the combination of [o] and [i] is short. A third consideration is that if the diphthongs involved occur in a trisyllabic word, they may receive secondary stress, and are therefore lengthened. In that case, rules like degemination do apply. To conclude, we have considered some evidence for the existence of short diphthongs in synchronic systems. A sample diphthong like [io] may be represented as in (26) below:

208 (26)

χ / \ V

V

I I

Α Α

υ

3. A vowel and a consonant under one x-slot I repeat (5)c as (27): (27)

χ

/\ C

V

There are two possible ways of interpreting such a structure: either the consonantal or the vocalic root is the head. I discuss these possibilities in turn.

3.1 Consonants with secondary articulation If in a two-root complex segment such as (27) the consonantal root is head, the interpretation of such structures as consonants with secondary articulation is straightforward. The Mismatch condition discussed in chapter 4 dictates that the secondary articulation be of a simplex type, i.e. either labialisation, velarisation (including pharyngealisation and uvularisation) or palatalisation. These correspond to the primitives U, A and I, respectively. The representation of consonants with secondary articulation as having two root nodes is not new. In Pulleyblank (1984b) it is proposed that the consonant and the glide part of a consonant with secondary articulation each has its own root node linked to the same timing slot (see Pulleyblank 1989: 390). In Sagey (1986a), however, this kind of representation is not allowed since in her theory branching may take place only at the feature level (see chapter 3 above for discussion). For this reason, Sagey proposes that segments like / k w / and / k j / are represented as in (28) (see Sagey 1986a: 216f.; Pulleyblank 1989: 389):

209 (28)

Root

[+round]

Root

[+back]

[-back]

/k7

AV

The use of the pointer to indicate which of the two articulations is major has been widely criticised (see for instance Padgett 1991). Furthermore, there is nothing that specifies the [-consonantal] articulation of the 'minor' features [+round] and [-back] features in (28), as Pulleyblank (1989: 389f.) points out. Both Pulleyblank and Padgett argue on these grounds that double-root segments are the best alternative to Sagey's representation. In the present theory, it seems the only possible way of analysing consonants with secondary articulation (see also Jacobs and van de Weijer 1992: 133). An important issue here is that of the ordering of the two parts of the complex segment. Recall that C and V in (27) are abbreviations for major class specifications, for example in terms of the feature [± consonantal]. If these are two specifications of one binary feature, the parts of a segment with secondary articulation are predicted to be phonologically ordered ((29)a below). If they are two separate, single-valued features, they are predicted to be unordered, as in (29)b: (29)

a.

χ

/\ [-cons] [+cons]

b. [cons] [voc]

There is overwhelming evidence that the representation in (29)b is correct, i.e. that the major class difference between consonants and vowels is best expressed by two unary autosegments such as [cons] and [voc]. First, secondary palatalisation may occur both before or after front vowels and /}/. In varieties of Breton (Jackson 1967), for instance, "palatal [k'] and [g'] occur at le Bourg Blanc as allophones of the / k / and / g / phonemes after [i] (not before it), even though they are followed by back vowels" (Jackson 1967: 54). The same is true of Carib (Hoff 1968): "with the exception of the glottal stop ? and the voiced glottal fricative h all the consonantal phonemes have two allophones, viz. a palatalized allophone and a non-palatalized one. Generally speaking (...) the occurrence of these allophones is determined by the preceding vowel: if this is a short or long i or an i-

210 diphthong, then the palatalized allophone occurs; in all other cases the non-palatalized allophone is found" (Hoff 1968: 31). In the case of velars, palatalization results in a shift in primary place of articulation, but in all other cases secondary articulation occurs. The following examples show this (Hoff 1968: 39): (30)

piip'o piit'o siim'o

'skin' 'flatus' 'liana'

okoip'o kuitJa aim'a

species of tree 'spindle' 'to smoke'

It is interesting to note, as Hoff does, that "palatalization does not take place when consonants or consonant sequences are not only preceded but also followed by / / / or //.-/ (...)". The reason why palatalisation does not take place before / i ( : ) / may be related to the fact that in Carib syllable-initial /ji-/ does not occur (see Hoff 1968: 64). This is the kind of sequence that palatalisation would create. A second instance comes from Lenakel. Secondary labialisation in this language (Lynch 1978) rounds the low vowel [a] to [D] before as well as after labialised consonants. Lenakel has labialised /p w m w / as well as plain / p m / . The following examples (from Lynch 1978: 11) illustrate the rounding rule: (31)

/apwa/ /epwaik/ /namwi-n/ /nimwataa-n/

[óbwt>] [eb w ójk h ] [ni>mwin] [ni m w odán]

locative marker 'to fold' 'his footprints' 'his back'

Ordinary labials do not have the same effect: (32)

[amgám9k h ] [íamáj]

'dirty' 'it is flowing'

A third consideration comes from Hausa, where / a : / frequently becomes [ai] before palatalised velars and [au] before labialised velars, according to Dunstan (1969: 78), although both velars presumably have a phonetic offglide. The Hausa case also shows that the consonantal and the vocalic slot are phonologically unordered. Consider next the treatment of loanwords from Yoruba in the distantly related Kwa language Nupe (Hyman 1970: 66). The latter allows palatalisation and labialisation, but the complex vowels / e o / are not allowed in this language. These vowels are nativised in the following way (tone omitted):

211 Yoruba [keke] [egbe] [tore] [kobo] [oko] [kpo] [eje] [owo]

Nupe [kWa] — [egbya] — [twarya] - * [kwabwa] w — [ek a] [kpwa] - * [egya] [egwa] —

'bicycle' place name 'to give a gift' 'penny' 'spear' 'to be cheap' 'blood' 'hand'

In the present model the change from Yoruba is purely one of organisation: no segmental elements are inserted or deleted. The most sonorous of the two elements in the complex vowels / e 0/ is retained, and the other element attaches to the preceding consonant, on an inserted vocalic root node (I/U here means either I or U): (34)

A

V

V I/U

A

In mora theory it is also possible to represent this process, if intermediate nodes like C- and V-place are assumed. Similarly, only in this way can mora theory represent consonants with secondary articulation in synchronic systems. In chapter 3 I argued that such nodes are not necessary. In Luganda (Katamba 1974), the formation of consonants with secondary articulation actually shows that a timing slot is vacated ((35)a below). Postvocalically, the underlying / i / turns into a [j]: since it does not vacate a timing slot the following vowel does not lengthen ((35)b). The data have been taken from Steriade (1984): /N-kuale/ /mu-luan-i/ /ku-li-a/ /mi-oio/ /a-ial-a/ /a-iogel-a/ /e-iak-a/ /ki-eia/ /lu-ieio/ /mu-som-ie/

[qkwaile] [mulwaini] [kulja:] [mjoijo] [ajala] [ajogela] [ejaka] [ce:ja] [lweijo] [musomje]

'partridge' 'fighter' 'to eat' 'souls' 'he spreads out' 'he talks' 'it burns' 'drought' 'brooms' 'you have read'

212 These data indicate that a timing slot is vacated as the secondary articulation is created. Most straightforwardly this can be expressed as the adjunction of the vocalic root node to the timing slot of the consonant, which immediately frees a timing slot. Finally, consider the typical historical development of secondary articulation in general (for a description of secondary palatalisation, see Bhat 1978). Usually a consonant acquires some characteristics of a neighbouring vowel. This can be formally expressed as the spreading of a node that previously dominated the vowel. In the model developed here, this is the root node: (36)

χ

χ

C

V

Note that delinking of the vowel slot would lead to full compression of the two segments into one (cf. (6) above). I assume that this does not normally happen because it would cause the nucleus to become completely empty: (37)

*

O I χ

Nu I χ

Γ—• c

ν

32 Vowels with consonantal properties We now turn to the second possible way of interpreting (27), namely when the vocalic root in the two-root complex segment is the head. Again, the Mismatch condition demands an 'unpolluted' consonantal place specification, i.e. a single A, I or U element. Furthermore, it would appear possible that vowels may also have a nasal or liquid second element in the consonantal position: that is, in such cases the C-node must be of a sonorant type. First, let us investigate those cases where vowels have a consonantal constriction. As observed above, the Place of this constriction can be one of three types, Coronal, Labial or Dorsal. These three possibilities are represented schematically below (Manner features on the consonantal root nodes suppressed): (38)

a.

χ / \ V

b. C I ι

χ / \ V

c. C I υ

X / \ V

C

213 With regard to the representation in (38)a, a number of interpretations come to mind. First, retroflexed vowels (e.g. in Tarascan (747) or Mandarin Chinese (500); see Bhat 1974a; Maddieson 1984) could be described as vowels with an element I on a separate, consonantal root. The American English / r / , which may serve as a nucleus (as in bird, [b3~d]), could also fit this heading. The controversy of whether this nuclear element is a consonant or a vowel can then be resolved: it is a bit of both. Another language with vowels with secondary articulation may be Badaga (see Ladefoged and Maddieson 1990: 116; Clements 1990b, which both follow Emenau 1939). This language has a five-vowel system which occurs in two quantities, short and long. Furthermore, three 'categories of resonance' (Clements 1990b: 24) are found for each vowel quality and in both quantities, viz. normal, half-retroflexed, and fullyretroflexed. Hence, in addition to vowels with an extra consonantal I-element, the most 'retracted' of these vowels might be represented as having A in the same position ((38)c). This is suggested by Emenau's description that "in the vowels with fully-retroflexed resonance the whole tongue is strongly retracted, the edges are curved upward towards the hard palate well behind the alveolar ridge (...)" (see Ladefoged and Maddieson 1990: 116). Other 'pharyngealised' vowels may also be candidates for a representation like (38)c (such as those of Neo-Aramaic, see Clements 1990b: 23). Finally, in Ladefoged and Maddieson (1990: 109ff.), the pharyngealised vowels of Even, a Tungus language of North-Central Siberia, are discussed in some detail. Ladefoged and Maddieson observe that the X-ray tracings in Novikova (1960) suggest that the shape of the vocal tract for these pharyngealised vowels is similar to the ATR vowels in Akan, with the difference that in Even there is "a greater degree of retraction" (Ladefoged and Maddieson 1990: 111). This description suggests that the phonological element involved in both categories of vowels may be the same but that in pharyngealised vowels it may be more preponderant or, as is proposed here, more independent than in ATR vowels. Finally, the element U on the consonantal root of a two-root vowel (38)b could serve to distinguish the well-known Swedish in-rounded from out-rounded vowels. The latter have been described as diphthongs with a consonantal, fricative element (see Clements 1990b: 15: "The vowels [a: u:] are described as ending with a period of consonantal turbulence, producing a segment similar, it seems, to the voiced bilabial approximant [ß] found in other languages. Another description of [»: u:], then, would treat them as closing diphthongs [yß uß], containing short vowel nuclei followed by the consonantal element [ß]". Clements' description and analysis match the spirit of the two-root representation perfectly. A phonetic analysis like that of McAllister, Lubker and Carlson (1974) confirms such a proposal (see also Ladefoged and Maddieson 1990: lOOf.). Only in the last of these cases was it possible to determine the Manner characteristics of the accompanying consonantal root node with some degree of certainty: these vowels were described as fricatives. Let us now turn to some vowels which more

214 clearly have consonantal Manner characteristics. Among these are the nasal vowels of Polish. Polish is the last of the Slavic languages to retain nasal vowels in its phonology. These vowels are spelled q and ç. Their pronunciation varies among dialects, but in the standard variety there are two allophones. The first is an oral vowel followed by a nasal glide [fx|, and the second is an oral vowel followed by a nasal consonant. The former allophone occurs when the nasal vowel appears before fricatives or pause, and the latter occurs before stops (including affricates). Examples (taken from Rowicka and van de Weijer 1992) appear in (39): w[oû( s] kr[oúj 3]y