1. INTRODUCTION
A crucial aspect of phonological patterns is the domain in which they apply. The massive body of research in Prosodic Phonology culminates in a prosodic hierarchy that encompasses a finite set of prosodic structures (Selkirk Reference Selkirk, Aronoff and Kean1980, Reference Selkirk1984; Hayes Reference Hayes, Kiparsky and Youmans1989; Nespor & Vogel Reference Nespor and Vogel2007). Those central to the present study are summarized in (1).
(1) Prosodic domains between the syllable and the phonological phrase Footnote 2
Each of the domains in (1) delimits the application of different phonological patterns. Whereas the syllable (σ) is, among others, the locus of internal phonotactic generalizations like the Sonority Sequencing Principle (e.g. Blevins Reference Blevins and Goldsmith1995), the foot (ω) forms the domain for rhythm generalizations (e.g. Hayes Reference Hayes1995) and other patterns which reference the combination of syllables. The prosodic word (ω) is different from the aforementioned domains in being mapped onto morphological structure, e.g. a stem and its affixes. A number of generalizations, including minimality constraints, phonotactic generalizations, and the application of phonological processes, impinge on this domain cross-linguistically (e.g. Peperkamp Reference Peperkamp1997, Hall & Kleinhenz Reference Hall and Kleinhenz1999, Dixon & Aikhenvald Reference Dixon, Aikhenvald, Dixon and Aikhenvald2002a). The phonological phrase (P) references syntactic phrases and provides the constituent for post-lexical processes. Although the domains in the prosodic hierarchy interact with morphosyntactic structure, the theory does not assume that prosodic and syntactic structures are always isomorphic. Quite to the contrary, mismatches between these modules of grammar justify the formulation of Prosodic Phonology as a component in its own right (Nespor & Vogel Reference Nespor and Vogel2007: Chapter 2).
A critical methodological assumption of the theory is that it only holds for word-level processes that apply generally across the lexicon and that these processes are distinct from processes at the word level that make direct reference to specific morphological elements (Nespor & Vogel Reference Nespor and Vogel2007: 18). The latter rules would take place in the Lexical Phonology component that precedes the Prosodic Phonology component, although the interaction of these two components might be even more complicated. We refer to this assumption as the ‘Generality Assumption’.Footnote 3
Prosodic Hierarchy theory makes a number of predictions. In this paper, we discuss two of them, labeled the ‘Clustering Prediction’ and the ‘Strict Succession Prediction’, respectively. The Clustering Prediction states that phonological domains cluster on the single universal set of domains enshrined in the prosodic hierarchy. In other words, it predicts that, within and across languages, phonological patterns will be sensitive to these domains and not more (Inkelas & Zec Reference Inkelas, Zec and Goldsmith1995: 547f.). Specifically, the theory predicts that word-related phonological patterns will cross-linguistically converge on one and only one prosodic word domain. This assumption forms the common denominator of most theories of prosodic word structure. Accordingly, more recent Optimality-theoretic treatments of prosodic structure implement the architecture of Prosodic Phonology as representational primitives of the generator (McCarthy & Prince Reference McCarthy and Prince1986, Prince & Smolensky Reference Prince and Smolensky1993, see also Selkirk Reference Selkirk, Beckman, Dickey and Urbanczyk1995). Lexical Phonology and its successors, although allowing greater variation with respect to morphophonological strata, nevertheless make a principled distinction between the lexical (word-related) and the post-lexical (phrasal) levels (Kiparsky Reference Kiparsky and Yang1982).
The Strict Succession Prediction claims that in the building of prosodic structure none of the levels of the hierarchy can be skipped, i.e. that each level Ln is followed by a level Ln−1 until the terminal level L0. For example, a prosodic word can only be composed of one or several feet, but not of one or several syllables (Nespor & Vogel Reference Nespor and Vogel2007: 7). An alternative formulation of the Strict Succession Prediction is in terms of Proper Headedness (Itô & Mester Reference Itô and Armin Mester1992). This principle ensures that every prosodic constituent must have a head, defined as an immediately dominated category of the immediate subordinate prosodic rank. At the word level, this entails that a prosodic word dominates at least one foot.
Whereas most work in Prosodic Phonology accepts these predictions as mostly confirmed, some contributions question their universal validity and empirical basis. As Inkelas & Zec (Reference Inkelas, Zec and Goldsmith1995: 548) put it, ‘[w]hile the predictions are clear, in practice the evidence is less so’. In agreement with a growing body of literature challenging many traditionally accepted phonological universals (e.g. Pierrehumbert Reference Pierrehumbert, Bod, Hay and Jannedy2003, Blevins Reference Blevins2004, Mielke Reference Mielke2004), we propose to submit the predictions of the Prosodic Hierarchy to empirical testing. In this paper, we report some of the results from a large-scale typological survey that allows such testing.
In its canonical form, the Prosodic Hierarchy is an absolute universal, an ingredient of our descriptive metalanguage that is deemed necessary in the analysis of any language, on a par with notions such as ‘distinctive feature’. Another way of putting this is to declare the elements of the hierarchy as atoms of Universal Grammar (e.g. Nespor & Vogel Reference Nespor and Vogel2007). Challenges to such declarations cannot come from individual data or typological surveys but only by demonstrating that the predictions of the hierarchy lead (i) to analyses that are self-contradictory (e.g. claiming and not claiming word status for the same string at the same time) or (ii) to analyses that violate simultaneously held assumptions (Bickel in press). An additional possibility for falsifying such declarations, which we accept in the following, has to do with overgeneration: a declaration fails if it generates more structure than what can be descriptively motivated in at least one language. Some theorists reject this mode of falsification, but even if one does reject it, it seems to us a fruitful enterprise to determine which elements of our descriptive metalanguage indeed are necessary in the analyses of any language, and which elements are not necessary in any language. The latter can be taken to constitute inventories of possible, but not necessary, structures, on a par with such notions as ‘±high tone’. An interpretation of the Prosodic Hierarchy as an inventory, from which languages may chose, has been advocated early on by Selkirk (Reference Selkirk, Aronoff and Kean1980). It seems to us to represent an increasingly widespread view of phonological theory and is the standard view in phonological typology.
In this paper, we attempt to demonstrate that the Generality Assumption either leads to a violation of the Strict Succession Prediction or overgenerates prosodic structure in Vietnamese (Austroasiatic), and that it leads to a violation of the Clustering Prediction in Limbu (Sino-Tibetan). Specifically, we will argue that applying the Generality Assumption in Vietnamese obstructs any evidence for distinguishing between phonological phrases and prosodic words. In the case of Limbu, we argue that more than one domain must be posited between the foot and the phonological phrase, violating the Clustering Prediction and resulting in an analysis under which the same string of morphemes is and is not a prosodic word.
To be sure, deviations from the predictions of the Prosodic Hierarchy have been noted before,Footnote 4 but we will argue that the proposed solutions capture only very specific kinds of deviations and may work well only if deviations comprise marginal phenomena. In Sections 2 and 3, we embark on in-depth analyses of Vietnamese and Limbu as representative examples illustrating that languages may lack or multiply the level of the word in their prosodic structure. Section 4 details an alternative perspective on prosodic domains as conceived of as language-particular, intrinsic and highly specific properties of individual rules or constraints, and lays out the merits of this perspective for cross-linguistic comparison. Section 5 concludes the paper and highlights the consequences of our proposal, namely that the prosodic word emerges through frequent reference of sound patterns to a given construction type in a given language.
2. CHALLENGE 1: VIETNAMESE, OR THE LACK OF PROSODIC WORD DOMAINS
In a number of languages, motivating a prosodic word domain under the Generality Assumption turns out to be a non-trivial matter. For example, in a recent study on the word in Dalabon, a polysynthetic Australian language of Arnhem Land, Evans, Fletcher & Ross (Reference Evans, Fletcher and Ross2008) apply a number of phonological tests for the definition of the prosodic word. Neither morphophonemic processes, such as vowel harmony, assimilation or dissimilation, nor phonotactic constraints, nor suprasegmental processes, such as stress assignment, provide conclusive evidence for word boundaries. In their discussion of the relation between the grammatical and the prosodic word, the authors thus have to rely on pause placement and intonation contours. Note that within Prosodic Phonology, pauses and intonation are usually taken to operate on the level of the intonational phrase and not the word.Footnote 5
Although structurally very different, Vietnamese, a Viet-Muong language of Vietnam, poses a comparable problem to the universality of the prosodic word, since on purely descriptive grounds ‘there is no significant unit in Vietnamese intermediate between the syllable and the phonological phrase’ (Thomas Reference Thomas1962: 521). When limiting the evidence to lexically-general processes, a monosyllabic word in Vietnamese is prosodically indistinct from other syllables and polysyllabic words behave prosodically like other polysyllabic strings, i.e. phrases. If we simply postulate all of the levels in (1) above for this language, we would incorrectly predict the application of word-level processes, i.e. we would overgenerate prosodic structures. If we try to capture the lack of word-level prosodic processes by the exclusion of the ω-node in the prosodic tree, we must conclude that both Clustering and Strict Succession/Proper Headedness are violated. Vietnamese thus constitutes a real challenge to the Prosodic Hierarchy and the Strict Layer Hypothesis.
Since prosodic words by definition map onto morphological structure, a discussion of the grammatical word in Vietnamese is mandatory before we examine candidates of word-related prosodic processes motivating a prosodic word domain.
2.1 Morphology and the structure of grammatical words in Vietnamese
Although the definition of the word in Vietnamese has stirred up some discussion (Thomas Reference Thomas1962, Thompson Reference Thompson1963, Noyer Reference Noyer1998), linguists working on the language usually agree on the point that on morphological grounds, one can isolate forms showing different degrees of morphological complexity and phonological length (Emeneau Reference Emeneau1951, Thompson Reference Thompson1965, Dinh-Hoa Reference Dinh-Hoa1997). The examples in (2) summarize the main options for the phonological shape of monomorphemic and polymorphemic grammatical words.Footnote 6
(2) Possible words in Vietnamese (Thompson Reference Thompson1963: 50f.; Nhàn Reference Nhàn1984: 181)
Monomorphemic words are most commonly monosyllabic, the few polysyllabic monomorphemic words are either place names, e.g. Sài-gòn, or loans, e.g. the French borrowing va-li ‘suitcase’. Polymorphemic words, on the other hand, are generally polysyllabic. The morpheme break-up of the deictic elements đ-ây ‘here’ and v-ầy ‘this way’ is controversial, but it has no repercussions on prosodic structureFootnote 7 Other morphologically complex forms are the result of the two productive morphological processes that are found in the language, viz. reduplication and compounding (see below).
The main rationale for considering the forms in (2) as grammatical words is the fact that they can occur as basic free forms and that they have a conventionalized coherence and meaning. However, as we will discuss in Section 2.1.3, their status as grammatical words is problematic with respect to other criteria for grammatical wordhood that have been proposed in the literature, namely the criteria of cohesiveness and fixed ordering (Dixon & Aikhenvald Reference Dixon, Aikhenvald, Dixon and Aikhenvald2002a: 19).Footnote 8
2.1.1 Reduplication
Reduplication is a morphological process in Vietnamese that expresses a number of meanings, such as distributive, iterative, attenuative, intensive and emphatic (see Thompson Reference Thompson1965: 139ff., Noyer Reference Noyer1998: 80f. for a summary). The various reduplicative patterns with prefixed and suffixed reduplicants are given in (3). In what follows, we underline the bases of the reduplicated forms.
(3) Patterns of reduplication in Vietnamese (Thompson Reference Thompson1965: 139f.)
In alliterative reduplication, only the onset of the base is copied to the prefixed or suffixed reduplicant. Riming reduplication is characterized by the fact that the rime of both base and reduplicant are identical. In some cases the tone of the reduplicant changes in the process. Comparing the two constituents in vocalic reduplications, we see that everything except for the vowel is repeated in the reduplicant. In tonal reduplication everything except for the tone is retained in the reduplicant. Finally, since both members of complete reduplications are entirely identical, we cannot decide whether the reduplicant is prefixed or suffixed in disyllabic forms. Apart from the segmental resemblances between the base and its reduplicant, some of these forms also exhibit a special tonal phonology, which we will discuss in some detail in Section 2.2.3.
As noted above, the various processes of reduplication may apply successively and/or recursively on a base, yielding output forms of considerable length. Consider the following examples:
(4) Multiple reduplication in Vietnamese (Nhàn Reference Nhàn1984: 247–249)
(a) lục → lục-lọi → lục-lục-lọi-lọi ‘search’ ‘search inside out’ ‘search over and over’
(b) ngầm → tầm-ngầm → tẩm-ngẩm-tầm-ngầm ‘underground’ ‘very secret’ ‘extremely secretive’
In (4a), alliterative suffixing reduplication applies on the base lục in the first step of the derivation. The output is input to a rule of total reduplication applying to both members of the base lục-lọi, yielding the form lục-lục-lọi-lọi. In (4b), riming prefixing reduplication operates on the base ngầm in the first step of the derivation. The output of this process tầm-ngầm is then input to tonal prefixing reduplication which results in the form tẩm-ngẩm-tầm-ngầm. With respect to the degree of complexity which reduplications can exhibit, this morphological process behaves like the other productive morphological process in the language, namely compounding.
2.1.2 Compounding
Stems can be combined in a number of ways to derive compound words. Thompson (Reference Thompson1965) distinguishes two major types of compound structures in Vietnamese, syntactic compounds, which parallel phrases, and non-syntactic compounds, with a structure that is not paralleled by syntactic phrases, see (5).
(5) Types of compounds in Vietnamese (Thompson Reference Thompson1965: 126ff.)
Syntactic Non-syntactic
(a) bàn-ghế ‘furniture’ (b) sợ-hoảng ‘be terrified’
(bàn ‘table’+ghế ‘chair’) (sợ ‘be afraid’+hoảng
~ bàn ghế ‘tables and chairs’ ‘be panic stricken’)
(c) người ở ‘servant’ (d) học trò ‘schoolchild, pupil’
(người ‘person’ +ở ‘be located’) (học ‘to study’+trò
~ người ở ‘person residing’ ‘school-age child’)
Cross-cutting the distinction between syntactic and non-syntactic compounds, a further distinction can be made between coordinative (i.e. dvandva) formations, in which both elements are weighted equally, and adjunctive formation, which show internal complement–head structure (Nhàn Reference Nhàn1984: 270ff.). The generalizing compound in (5a) and the reinforcing compound in (5b) can be said to consist of two coordinated heads. The specializing compound in (5c) and the attributive compound in (5d) consist of a head and a modifier. In the former, the first element is the head and the second element is the modifier, in the latter, the second element is the head and the first element is the modifier. Some compound-like structures, such as quốc-gia ‘nation, country’ (quốc ‘nation’+gia ‘household, establishment’), are exceptional insofar as their components only appear in compound structures. Thompson (Reference Thompson1965: 120) calls such elements ‘pseudo-compounds’.
Compounds can recursively undergo further compounding, as shown in (6).
(6) Multiple compounding in Vietnamese (Thompson Reference Thompson1965: 136f.)
(a) ngôn-ngữ học ‘linguistics’ (ngôn ‘speech, word’+ngữ ‘language’+học ‘to study’)
(b) vô-tuyến điện-thoại ‘radio telephone’ (vô ‘without’+tuyến ‘wire’ +điện ‘electricity’+thoại ‘speech, conversation’)
In the derivation of the word in (6a), the two bound forms ngôn ‘speech, word’ and ngữ ‘language’ are combined to yield the compound ngôn-ngữ ‘language’, before this latter form is compounded with the stem học ‘to study’. The second derivation, in (6b), involves the combination of vô ‘without’ and tuyến ‘wire’ yielding the compound vô-tuyến ‘wireless’ and the combination of điện ‘electricity’ and thoại ‘speech, conversation’ yielding the compound điện-thoại ‘telephone’. These forms are again input to a rule of compounding which results in the complex form vô-tuyến điện-thoại ‘radio telephone’.
2.1.3 Grammatical wordhood
The outputs of reduplication and compounding often have conventionalized meanings that are no longer compositionally derived. From this point of view, the forms would seem to qualify as unitary grammatical words. But the forms fail on other conventional criteria of grammatical wordhood as terminal nodes in the syntax, such as non-interruptability and ordering constraints. In Vietnamese, polysyllabic strings with a conventional meaning can be interrupted by phrasal elements, regardless of whether the strings consist of one or more morphemes.
(7) Interruptability of Vietnamese words (Nhàn Reference Nhàn1984: 6; Noyer Reference Noyer1998: 82)
(a) cà-phê ‘coffee’ vs. cà với phê ‘coffee and the like’
(b) đo đỏ ‘reddish’ vs. đo không đỏ ‘not reddish’
(c) nhà cửa ‘house, home’ vs. Tôi xay nhà xay cửa ‘I build a house.’
In (7a), the monomorphemic French loan cà-phê ‘coffee’ is split by the conjunction với ‘with, and’, which is inserted between the two syllables of the string. The constituents of the reduplication đo đỏ ‘reddish’ in (7b) may also be separated by a phrasal element, in this case the element không ‘not’. The insertion of với and không could prima facie also be analyzed as endoclisis (Harris Reference Harris2002, Bickel et al. Reference Bickel, Banjade, Gaenszle, Lieven, Paudyal, Rai, Rai, Rai and Stoll2007). But such an analysis would not carry over to (7c). Here, the two constituents of the compound nhà cửa ‘house, home’ appear in interlocking order with the adjacent verb xay ‘build’. In such sentential constructions, a reduplicated twin of the adjacent word (xay) in the clause breaks the unity of the compound word. Analyzing xay as endoclitic would entail that most lexical words in Vietnamese can endocliticize, a consequence that challenges the observation that in all other known cases, endoclisis is an extremely limited process. As a result, the data in (7) cast serious doubt on the grammatical integrity of the forms.
In many languages, the constituents of grammatical words are subject to strict ordering constraints not found among phrasal elements. Vietnamese confronts us with variable orders like those presented in (8).
(8) Variable order in Vietnamese words (Nhàn Reference Nhàn1984: 6; Thompson Reference Thompson1965: 130)
(a) quần-áo vs. áo-quần ‘clothes’ (quần ‘trousers’+áo ‘tunic’)
(b) chọn lựa vs. lựa chọn ‘to select’ (chọn ‘choose’+lựa ‘choose’)
(c) bổi-rổi bối-rối vs. bối-rối bổi-rổi ‘be troubled’
(d) com-rom còm-ròm vs. còm-ròm com-rom ‘be emaciated’
The first two examples show two conjunctive compounds consisting of two nouns in (8a) and two verbs in (8b), in which the order of the constituents can be varied without a difference in meaning. In (8c) and (8d), we see reduplications with disyllabic reduplicative bases, bối-rối and còm-ròm, respectively. In the output form, the reduplicant of these bases can appear either before the base or after it, without any semantic differentiation. Although it is not clear to us how common such variable orderings are in Vietnamese, the fact that such forms are possible speaks against strict ordering constraints for grammatical words in the language.
There are languages where elements inside a grammatical word can freely permute (e.g. Chintang, Bickel et al. Reference Bickel, Banjade, Gaenszle, Lieven, Paudyal, Rai, Rai, Rai and Stoll2007). However, in Vietnamese, it is not even clear whether the domain of free permutation is a grammatical word: the only evidence that they are words is non-compositional semantics, but non-compositional semantics can also be a property of phrases, as witnessed by idioms. Since, as we observed before, the syllables of polysyllabic strings can mix with phrasal elements even if they have a conventional and non-compositional semantics, it is indeed likely that they are themselves terminal nodes in the syntax, and not subconstituents of terminal nodes. In line with this, Nhàn (Reference Nhàn1984) postulates the syllabeme as the minimal unit of grammatical analysis in Vietnamese. In the more recent analysis of Noyer (Reference Noyer1998), the distribution of such syllabemes is governed by general rules of syntax, which operate across various levels of grammatical structure. For practical purposes, we will use the notion ‘lexical unit’ to refer to syllabeme strings with conventional and non-compositional semantics.
To conclude, although descriptions of Vietnamese use notions of (grammatical) wordhood, there is no evidence that there are terminal nodes in the syntax longer than one syllable. This characterization forms the basis of the following discussion of the prosodic properties of monosyllabic and polysyllabic lexical units in Vietnamese.
2.2 Phonological domains in Vietnamese
In order to determine whether monosyllabic and polysyllabic words of the kind discussed above constitute prosodic words, we examine all candidates with respect to word-related phonological properties of the language and discuss their relation to morphosyntactic structure. It turns out that on phonological grounds monosyllabic lexical units are indistinguishable from other syllables and that polysyllabic lexical units are indistinguishable from other polysyllabic strings.
2.2.1 The syllable and the monosyllabic word
Traditionally, the study of Vietnamese phonology is primarily concerned with the syllable, which is the domain for phonotactic generalizations, the distribution of tone and the realization of stress. The table in (9) presents a schematized representation of syllable structure in Vietnamese.
(9) The syllable in Vietnamese
Each syllable obligatorily starts with an onset consonant, traditionally referred to as the initial. The initial can be any consonant except for the bilabial voiceless stop /p/. The rhyme obligatorily consists of a vocalic nucleus, which is filled by either a short, long or diphthongal vowel. The glide /w/ may optionally precede the nucleus. The coda position can optionally be filled by one of the eight segments /p t k m n ŋ y w/. The phonetic realization of the final velar consonants /k ŋ/ is determined by the quantity and quality of the preceding vowel. After the short, round vowels /u o ɔ/, the two consonants surface in their labio-velar allophones, i.e. [kp] and [ŋm], respectively. If the preceding vowel is short and front, they are realized as the palatals [c] and [ŋ], respectively (Pham Reference Pham2003: 138). This rule establishes the rhyme as a key domain of Vietnamese phonology.
The syllable is traditionally treated as the tone-bearing unit of the language.Footnote 9 The tonal distinctions in (10) are contrastive in Vietnamese.
(10) Vietnamese tones (Pham Reference Pham2003: 10)
The high series comprises the three tones ngang, sắc, and ngã.Footnote 10Huyền, nặng and hỏi can be grouped in a low series. Every syllable is realized with one of these six tones. Sắc and nặng tones have short allotones (called sắc2 and nặng2, respectively) whose distribution is restricted to syllables which are closed by a stop consonant. This phonotactic constraint on the distribution of allotones underlines the interdependence of phonotactics and the distribution of tones at the sub-syllabic level of the rhyme. Further evidence for this constituent comes from the phonetic realization of tone, such that the broken part of the ngã tone, for instance, appears in the middle of the rhyme and not in the middle of the vowel (Pham Reference Pham2003: 141).
According to Thompson (Reference Thompson1965: 106–107), each syllable in Vietnamese is accompanied by one stress. At higher levels of prosodic structure, several degrees of stress may be distinguished: weak, medium and heavy stress. The general pattern is that the degree of stress on the syllables within a pause group will be determined by the information load of the different elements, with heavy stress marking contrast, emphasis or the semantically most salient piece of information. In polysyllabic strings, we also encounter a strong preference for iambic stress asymmetries, which will be discussed in more detail in Section 2.2.2 below.
Many languages show a clear distinction between syllable structure and the canonical phonotactic structure of words. Mon, a distantly related language of the Mon-Khmer family, for instance, allows simple CV syllables with a short vowel like /pə/, i.e. monomoraic syllables. Due to the sesquisyllabic structure of the word, however, such a syllable can only occur as the first syllable of a disyllabic word, e.g. /pə-lac/ ‘tear down’. As soon as a simple CV syllable acquires word status, it surfaces with a lengthened vowel, e.g. /ʔa/→[ʔaː] ‘go’ (Bauer Reference Bauer1982, Jenny Reference Jenny2005). Whereas the former piece of evidence emphasizes the necessity to distinguish syllable- and word-related phonotactics in Mon, the latter piece of evidence goes to show that the bimoraic minimal word constraint is at work in the language (McCarthy & Prince Reference McCarthy and Prince1986). Note that establishing the bimoraic word already presupposes a number of levels of prosodic structure, such that the prosodic word dominates a foot domain which consists of two morae. Keeping this mode of reasoning in mind, consider the Vietnamese data in (11).
(11) Monosyllabic and polysyllabic words in Vietnamese (Nhàn Reference Nhàn1984)
The first column in (11) lists the possible phonotactic shells of monosyllabic words in Vietnamese. The minimal word /di/ ‘go’ in (11a), which can also constitute a minimum pause group, exhibits a simple CV structure and contains a short vowel as its nucleus. The other forms in the first column of (11) add up to demonstrate that the phonotactic shape of monosyllabic words is coextensive with the generally available syllable templates of the language, i.e. CV, CwV, CVC and CwVC. On the basis of this observation, there is no need to posit a bimoraic minimal word. Additionally, the data so far suggest that there is no difference between syllable-related and word-related phonotactics. However, it could be that such differences turn up in larger structures, e.g. the disyllabic word. In order to test whether this expectation is borne out, the second and third columns of (11) give disyllabic reduplicative words with prefixed and suffixed reduplicants, respectively. In each case, the shape of the reduplicants adheres to the syllable template given in (9), i.e. the phonotactic generalizations about the syllable remain valid for each and every syllable in the language, irrespective of their position within larger structures. Note that since the forms given in (11) present the possible inputs for (possibly recursive) processes of compounding and reduplication, differences between syllable-related and word-related phonotactics cannot be expected in larger words either.
To conclude this section, the phonotactics of the language does not motivate a distinction between syllable-related and word-related patterns, i.e. all syllables exhibit the same structure, irrespective of their grammatical status as affixes, stems, or words. Phonotactically, monosyllabic words are thus indistinguishable from other syllables. Whether suprasegmental features, such as stress and tone, are sensitive to a prosodic word domain will be explored in the next sections.
2.2.2 Stress and the polysyllabic word
Although the analysis of stress in Vietnamese still offers rich research opportunities, the descriptive consensus at this point is that (i) weak and strong stresses alternate in sequences of up to three syllables and (ii) heavy stress is usually placed on the last syllable of the sequence (Thomas Reference Thomas1962: 521). In other words, the basic rhythmic pattern of Vietnamese is one of iambic (or anapestic) phrasing, (syllables with main prominence are marked by ˈ), see (12).
(12) Phonological phrasing in Vietnamese (Thompson Reference Thompson1965: 107)
(ˈNói) (phải ˈcó người) (nói ˈđi) (nói ˈlại) (chớ ˈbắt) (người ta ˈnói) (một mình ˈhoài)! speak ought exist person speak go speak come prohibit constrain someone speak alone continually‘For a conversation, [you] ought to have people talking back and forth, not make somebody talk alone all the time!’
The utterance in (12) consists of seven stress units that vary in phonological length from monosyllabic to trisyllabic. The first unit consists of a monosyllabic lexical unit only, whereas the second one is made up of three syllables. Although final stress is the default, the placement of heavy stress within a stress unit is ultimately governed by the semantic saliency of the elements within the structure. This principle accounts for the non-final stress in the second stress unit of the utterance, in which the head có ‘exist’ receives heavy stress in penultimate position of the phrase. As we will see below, such cases of non-final stress placement are also evidenced in polysyllabic lexical units.
In Section 2.2.1, we already noted that each syllable is realized with one stress. Accordingly, the monosyllabic units in (11) will be realized with heavy stress if they constitute a stress unit or even an utterance when uttered in isolation. The crucial question is how stress assignment operates in the polysyllabic strings discussed above. If there is a separate layer of word-level stress in Vietnamese, stress assignment in such polysyllabic strings should differ from the general principles of phonological phrasing illustrated above. Consider (13), listing polysyllabic strings of different morphosyntactic types and exhibiting varying degrees of complexity.
(13) Stress in Vietnamese (Thompson Reference Thompson1965: 126ff.; Nhàn Reference Nhàn1984: 101)
(a) va-ˈli ‘suitcase’
(b) nói ˈnói ‘keep talking and talking’
(c) mơ ˈmơ màng ˈmàng ‘deep in the state of dreaming’
(d) người ˈta ‘somebody’
(e) một ˈmình ‘alone’
(f) Liên-ˈhiệp quốc ‘United Nations’
(g) hoa ˈhồng ‘rose’ (lexicalized, ‘single-word’ reading)
(h) hoa ˈhồng ‘pink flower’ (‘phrasal’ reading)
(i) Tôi không ˈbiết. ‘I don't know.’
Stress placement in lexical units is exemplified in (13a–g), with monomorphemic, reduplicative and compound strings. Stress assignment in monomorphemic units such as (13a) does not differ from that in reduplication (13b–c) or in compounding, as in (13d–g). In each case, the final syllable constitutes the locus for main prominence. Note that the tetrasyllabic reduplication in (13c) is parsed as two stress units in iambic phrasing.Footnote 11 Since the stress assignment in these structures is identical to that in the phrases (13h) and (13i), we must conclude that stress units in Vietnamese are construed on a number of adjacent syllables irrespective of their internal morphosyntactic composition. Polysyllabic strings with monomorphemic, reduplicative, or compounding structure are thus all subject to exactly the same principles of stress assignment as bona fide phrases. Example (13f) illustrates the only deviation from the default final stress. Here, the pseudo-compound has its origin in structural borrowing from Chinese. In such Sino-Vietnamese compounds, stress is assigned to the left branch of the construction, in which stress is realized on the final syllable.
The examples in (13) all involve strings uttered in isolation. Comparing the two isolated realizations of the morphologically complex forms người ta ‘somebody’ and một mình ‘alone’ in (13d) and (13e) with their counterparts in actual phrasing in (12), we notice that within larger phrases, the last element is regularly stressed, whereas the compounds do not take heavy stress: người ta ˈnói and một mình ˈhoài, respectively. This observation provides additional evidence for the analysis of a stress unit in which stress is assigned irrespective of morphosyntactic constituency.Footnote 12
The phonetic evidence relating to stress is often taken with skepticism, especially when it is based on qualitative auditory judgments of individual analysts. In the case just discussed, one objection would be that although the stress placement in lexical units and phrases is both final, there might be differences in the degree of stress which entails that the phonetic realization of stress at the phrasal level should be more salient. Minimal pairs like the one in (13g–h) constitute perfect test cases to clarify the issue: in terms of stress placement the examples are identical, but with respect to the phonetics of stress we would expect differences if we assume that stress assignment differs between the word and the phrase level. This question has been the focus of a number of recent production and perception experiments by Ingram & Nguyen (Reference Ingram, Nguyen, Warren and Watson2006), Nguyen & Ingram (Reference Nguyen, Ingram, Warren and Watson2006) and Nguyen & Ingram (Reference Nguyen and Ingram2007). With respect to the phonetic cues of syllable duration, F0 range, intensity, spectral slope and formant frequencies, no significant difference is detectable between compound-phrase minimal pairs of the type hoa hồng ‘rose’ vs. hoa hồng ‘pink flower’ under a normal speech production condition of a picture naming task. Under a maximal contrast condition, where native speakers are encouraged to bring out the difference in meaning between the two forms, the disambiguation strategy that speakers chose most often consists in inserting a phrasal juncture between the two constituents of the phrase. This prosodic juncture has a number of phonetic effects such as a pause after the first syllable and concomitant lengthening, widening of the pitch range and lowering of intensity in the pre-pausal syllable. Crucially, however, the chosen disambiguation strategy does not rely on word-level stress assignments.
The acoustic evidence from the production experiments does not support the idea of a separate layer of word stress in Vietnamese. This finding is further backed up by the results of a perception experiment by Nguyen & Ingram (Reference Nguyen and Ingram2007). Native speakers showed a response bias towards compound forms when asked to distinguish compounds from phrasal structures on the basis of the stimuli obtained from the normal speech production condition of a picture naming task. When confronted with noun phrases lacking a phrasal juncture between the two constituents, listeners performed at chance level when asked to identify the phrasal structures. Performance was slightly better with stimuli from the maximal contrast condition where the two syllables of the noun phrase were separated by a juncture. This shows that listeners do not rely on any phonetic cues for distinguishing compounds and phrases beyond phrasal juncture cues, which were most pronounced under the maximal contrast condition. In line with their findings from the production experiments, the authors conclude that there is no word-level phonological mechanism for distinguishing disyllabic lexical units from phrases. However, there is evidence for phrase-level phonology, which may be recruited to disambiguate the structures when necessary.
To summarize, Vietnamese provides ample evidence for a genuine stress domain that is preferably disyllabic and maximally trisyllabic. Within this domain, stress is realized on the final syllable in the default case. Crucially, this domain is computed irrespective of the morphosyntactic status of its constituent syllables, i.e. stress phonology does not distinguish between a word-level and a phrasal-level of prosodic structure. Metrically, polysyllabic words are thus indistinguishable from other combinations of syllables. Since the most complex structures which are referenced by the rules for iambic rhythm are phrasal, stress may most adequately be attributed to the prosodic domain of the Phonological Phrase.
2.2.3 The tonal phonology of reduplication
In Section 2.2.1, we discussed the syllable as the domain of tone and the relevance of the rhyme in the phonetic realization of tone with respect to timing. In running speech, the realization of tones is remarkably stable, i.e. there are no tonal sandhi rules which operate on phrasal or sentential levels (see Pham Reference Pham2003: 3). However, there are several patterns of reduplication in Vietnamese which result in interesting tone patterns. Recall that the different reduplication patterns illustrated in (3) exhibit different correspondences between the base and the reduplicant, ranging from complete identity to partial segmental identity with respect to initial or rhyme or partial identity in tone. The following data illustrate alliterative suffixing reduplications with bases covering all eight phonetic tones.
(14) Tone in Vietnamese reduplication (Pham Reference Pham, Jensen and Herk2000: 228)
(a) mau ‘fast’ → mau mắn ‘very fast’ (ngang – sắc)
(b) láu ‘clever’ → láu lỉnh ‘very clever’ (sắc – hỏi)
(c) vất ‘laborious’ → vất vả ‘very hard’ (sắc2 – hỏi)
(d) mỡ ‘greasy’ → mỡ màng ‘very greasy’ (ngã – huyền)
(e) tàn ‘worn out’ → tàn tạ ‘very worn out’ (huyền – nặng)
(f) lạnh ‘cold’ → lạnh lẽo ‘very cold’ (nặng – ngã)
(g) ngặt ‘severe’ → ngặt nghèo ‘very hard’ (nặng2 – huyền)
(h) đỏ ‘red’ → đỏ đan ‘very red’ (hỏi – ngang)
The examples in (14a–d) show reduplications of bases with high tones, namely ngang, sắc, sắc2 and ngã, respectively. The bases in the reduplications in (14e–h) carry low tones, namely huyền, nặng, nặng2 and hỏi, respectively. The tone of the suffixed reduplicant is never identical to the one of the base. However, in the first example, the tone of the base and the reduplicant belong to the same tone series, such that both ngang and sắc are high tones. In (14e) and (14g) we find parallel cases with low tones: in the former example we find the combination huyền – nặng, in the latter the combination nặng2 – huyền. Based on such observations, one may generalize that within reduplications the tone of base and reduplicant harmonize with respect to tone series (or tone ‘register’).
The formalization of tone series harmony faces a number of problems, however. Recall that ngã is traditionally grouped with the high tones. But in examples (14d) and (14f) this tone patterns with low tones, yielding the combinations ngã – huyền and nặng – ngã, respectively. A similar complication arises with hỏi, traditionally grouped with the low tones, which patterns with high tones in the combinations sắc – hỏi, in (14b), sắc2 – hỏi, in (14c), and hỏi – ngang, in (14h), respectively. In order to rescue the tone harmony analysis, an additional mechanism must be stipulated that reverses the series membership of the ngã and hỏi tones for the sake of the tonal phonology of reduplication, e.g. a rule of Concave Tone Reversal (Nhàn Reference Nhàn1984: 78) or a Flip-Flop-Rule (Pham Reference Pham, Féry, Green and van de Vijver2001). Another option lies in the re-categorization of the tone series, such that ngã becomes a ‘low’ tone and hỏi becomes a ‘high’ tone. This approach, however, produces a mismatch between the phonetics of tone, which is the basis of the traditional classification, and the phonological patterning of tone, which is evidenced in tone harmony. Although the mechanisms listed above can account for the deviant behavior of the two tones ngã and hỏi, they offer no explanation for other deviations from tone harmony. For instance, in the form cứng-cựng ‘very hard’ the high tone sắc is combined with the low tone nặng. In this case of tone register dissimilation (Thompson Reference Thompson1965: 155–157), tone register harmony is violated, but neither ngã nor hỏi is involved.
In order to evaluate the relevance of the distribution of tone within reduplication for the motivation of a disyllabic word domain, the forms discussed above must be situated within the broader picture of Vietnamese reduplication. In Section 2.1.1, the major structural distinctions between the various types of reduplication have already been summarized. Complete reduplication, in which the entire segmental and suprasegmental content of the base is copied in the reduplicant, is a common and fully productive process in the language. Since the reduplicant is an identical copy of the base in these formations, the fact that the two syllables of the reduplication agree in their tone series is not due to an explicit rule of tone harmony, but rather a side effect of the process itself.Footnote 13 In a second productive process of reduplication, the reduplicant differs from the base with respect to tone and coda consonants. For instance, when the form biệt ‘disappear’ constitutes the base, its prefixed reduplicant takes the even tone of the same tone series. In this case, the base carries the tone nặng of the low series and, accordingly, the reduplicant will be assigned the even low tone, i.e. huyền. Concomitant with the tone alternation, the coda of the reduplicant is subject to a consonant alternation in which the final stop is replaced by homorganic nasal. The output of the reduplication process is therefore biền-biệt. It is evident that in this second reduplication process less specifications of the base are copied into the reduplicant. At the segmental level, everything except for the manner feature of the coda consonant is copied. The change from a stop to a nasal consonant is inherently associated with the reduplication process. At the suprasegmental level, the tone series is copied, but all other tone features are provided by a requirement of the specific process, which demands an even tone.Footnote 14 Apart from these two productive processes, there are innumerable unproductive patterns differing with respect to how much features of the base are copied and how much features are assigned by the specific rules underlying the derivation. The shape of the reduplicant in an alliterative reduplication, for instance, cannot be predicted by general phonological rules. Although the onset consonant will be copied from the base, the rhyme can take one of a number of lexically pre-specified shapes.Footnote 15 A further obstacle in this context stems from the fact that a given form can be the input for several reduplication processes. For example, given the base cứng ‘be hard’, at least the three reduplicated forms cứng-cáp, cứng-cát and cứng-cỏi are attested. Finally, note that tone harmony does not in general apply to monomorphemic and compound lexical units: in such units, tones of the two series can combine without limit, e.g. bàn-ghế ‘furniture’ (huyền-sắc).
This brief survey of the nature of Vietnamese reduplication illustrates how diverse the processes and patterns found in this morphological domain are. The process and its accompanying phonological concomitants cannot be captured in a unified way. Quite to the contrary, a number of sub-processes must be distinguished, each of them exhibiting its own degree of productivity and phonological predictability. In the extreme case, the idiosyncratic phonological relationship between a base and its reduplicant is more efficiently handled by a lexicon entry than by the formulation of a phonological rule. As a result, no general rule of tone harmony can be said to characterize Vietnamese reduplication in its entirety. In line with this, a more adequate description must distinguish the different reduplicative processes and must specify in which of these a rule of tone register harmony is at work. We will elaborate on this point in 4, where we situate these cases in a broader theoretical context.
Since the tone patterns under review operate on the output of the morphological process of reduplication, the domain of these patterns could potentially provide evidence for the prosodic word domain (see DiCanio Reference DiCanio2005 for register harmony within the prosodic word in Mon and Khmer). However, as just noted above, the tone patterns are subject to lexical specification and only hold for specific subtypes of one morphological process (reduplication). The Generality Assumption that underlies the Prosodic Hierarchy does not admit such patterns as evidence for prosodic structure.
2.3 Discussion
In the previous sections, we surveyed the morphological and phonological properties of ‘words’ in Vietnamese. The grammatical word as terminal node in the syntax has a dubious status since the individual syllables of monomorphemic, reduplicative or compounding lexical strings allow variable ordering and can mix with phrasal elements. This suggests that Vietnamese syllables, or ‘syllabemes’, are direct targets of syntactic operations, not shielded off from the syntax by a more complex grammatical word node (cf. already Emeneau Reference Emeneau1951: 2–4, 44–45).
In phonology, the structure of monosyllabic lexical units is identical to that of other syllables,Footnote 16 whereas the metrical constituency of polysyllabic lexical units is indistinguishable from that of other polysyllabic strings. The tonal phonology of reduplication reveals a lexically specified morphophonological rule of tone harmony that does not count as evidence for a prosodic domain by standard criteria. On the basis of this evidence, we must conclude that there is no lexically general phonological rule which exclusively references the prosodic word in Vietnamese.
2.3.1 A prosodic hierarchy for Vietnamese
The representation of prosodic structure with appeal to the prosodic hierarchy introduced in (1) is far from trivial for Vietnamese. A stubborn application of all prosodic layers defined a priori yields the following prosodic tree:
(15) A prosodic tree for Vietnamese
This representation adheres to the architecture for prosodic structure expressed in the Prosodic Hierarchy and the Strict Layer Hypothesis. Each level between the syllable and the phonological phrase is employed in such a way that neither Clustering nor Strict Succession, nor Proper Bracketing is violated. However, the prosodic tree in (15) suffers from severe discrepancies with the actual empirical findings.
With respect to the first level of prosodic structure, the construction of syllable nodes is uncontroversial and follows the general phonotactic and suprasegmental generalizations concerning this domain as exemplified in Section 2.2.1 above. The evidence for the foot nodes is less straightforward. In the context of the minimum pause group đi /di/ ‘go’, we noted that the bimoraic minimal word constraint is not at work in Vietnamese. Accordingly, there is no evidence for a bimoraic foot dominated by the word node to provide a domain for the application of rules to fulfill minimal word requirements, such as compensatory lengthening. In (15), the foot node in some cases merely captures the fact that an element is bimoraic, e.g. tôi ‘I’. In the other cases, however, it would inadequately predict bimoraicity where it is not manifested or it would predict the application of foot-based processes, such as compensatory lengthening, where in fact none apply. This problem of overgeneration continues at the level of the prosodic word. As soon as this level of structure is present in the representation, it predicts the application of word-related phonological processes. The survey in Section 2.2 demonstrates that there is no positive evidence for such a domain. Moreover, experimental findings (Ingram & Nguyen Reference Ingram, Nguyen, Warren and Watson2006, Nguyen & Ingram Reference Nguyen, Ingram, Warren and Watson2006, Reference Nguyen and Ingram2007) suggest that this domain lacks psycholinguistic reality. The last node of the prosodic tree discussed here, the phonological phrase, seems to be as uncontroversial as the syllable node. The insights from stress phonology motivate this domain as the locus for the iambic rhythm of the language.
Overgeneration is thus the major problem of strict adherence to the Prosodic Hierarchy and the Strict Layer Hypothesis in the context of Vietnamese. Most of the predictions made by the prosodic representation are not manifested by phonological patterns evidenced in the language. Paring the tree down to those elements for which there is positive evidence in Vietnamese leaves us with the structure in (16), which violates the Strict Succession Prediction.
(16) A revised prosodic tree for Vietnamese
2.3.2 Skipping levels of prosodic structure
The discussion so far suggests that the syllable is the direct daughter of the phonological phrase in the prosodic representation of Vietnamese and that the building of prosodic structure in this language must consistently skip the level of the word (and the foot), i.e. Strict Succession must be violated in every prosodic tree.
In the literature on prosodic domains, a number of prosodic structures which violate Strict Succession have been proposed and defended, and it has been shown that only such structures can correctly predict the application of the domain-specific rules involved. For instance, a number of prosodic structures violating Strict Succession can be motivated for procliticization in Dutch (in (17a)) and German (dialectal, in (17b) and Middle Frankish in (17c)).
(17) Violations of Strict Succession (Kabak & Schiering Reference Kabak and Schiering2006: 82)
The examples above show that proclitics can constitute single segments, syllables or feet which are adjoined to the prosodic word at the level of the phonological phrase. In each case, the left branch of the tree violates Strict Succession in skipping the level of the word, (17c), the level of the word and the level of the foot, (17b), and the level of the word, the level of the foot, and the level of the syllable, (17a). The prosodic trees nevertheless correctly predict the facts encountered in the various languages. For instance, the structure in (17c) correctly predicts the application of foot-based processes such as trochaic rhythm and syllabification within function word contractions but rules out the application of processes applying across word boundaries within a phonological phrase in such constructions (see Kabak & Schiering Reference Kabak and Schiering2006 for details). Note that it is still true that a node Ln dominates one Ln–1 constituent.
Data like these, where only one element is dominated by Ln–1, motivate Itô & Mester's Reference Itô and Armin Mester1992 version of Weak Layering. In order to account for word structures such as ((tere)ϕbi)ω ‘television’ in Japanese, in which the final syllable is included in the prosodic word without constituting a prosodic foot, the authors present a restatement of Prosodic Layering which critically rests on the requirement (Proper Headedness) that every prosodic constituent must have a head, defined as an immediately dominated category of the immediate subordinate prosodic rank. Accordingly, the prosodic trees in (17) are acceptable because they still respect Proper Headedness, such that every phonological phrase contains at least one prosodic word.
But the Vietnamese facts challenge even this weakened interpretation of Layering since the phonological phrase would not contain any single prosodic word and thus violate Proper Headedness. In this respect, the structures which have been discussed in the literature differ significantly from the constellation which we would have to assume for Vietnamese if we do not want to postulate overgenerating structures.
Would Vietnamese fit theoretical expectations better if the prosodic hierarchy were modeled in terms of violable constraints, as in Optimality Theory? In Selkirk's (Reference Selkirk, Beckman, Dickey and Urbanczyk1995) approach, for example, the Strict Layer Hypothesis is decomposed into four constraints on prosodic domination: Layeredness (no Ci dominates a Cj, j>i), Headedness (any Ci must dominate a Ci–1 (except if Ci=σ)), Exhaustivity (no Ci immediately dominates a constituent Cj, j<i–1), and Nonrecursitivity (no Ci dominates Cj, j=i). Whereas the first two constraints are presented as holding universally, the last two are conceived of as violable. As already said, Vietnamese seriously questions this assumption with respect to (Proper) Headedness, since its prosodic structures contain P-nodes which do not dominate ω. But if we accept that Headedness is also a violable constraint, we have to ask which constraint overrides what is required by Strict Succession. Selkirk (Reference Selkirk, Beckman, Dickey and Urbanczyk1995: 453) alludes to Prince & Smolensky's (Reference Prince and Smolensky1993: 25) *Struc constraint, which ensures that structure is constructed minimally, essentially saying ‘less is best’. The tableau in (18) shows that the empirically adequate tree is chosen as the optimal candidate if *Struc is ranked higher than Headedness.
(18) *Struc >>Headedness
Candidate (18a), in which every level of the prosodic hierarchy is projected, yields two violations of *Struc, i.e. the unmotivated foot and word nodes. In this constraint ranking, it is discarded even though it does not violate Headedness. Candidate (18b) embodies a violation of *Struc at the word level but also a violation of Headedness at the foot level. Finally, although the prosodic structure in (18c) disobeys Headedness at the foot and the word levels, it wins as the optimal candidate since it respects *Struc completely.
This brief analysis illustrates the fact that recent approaches may well account for the Vietnamese data, if one accepts the basic assumptions of Optimality Theory. However, by allowing Headedness to be violable, the ‘essence of the Strict Layer Hypothesis’ (Selkirk Reference Selkirk, Beckman, Dickey and Urbanczyk1995: 443) has to be given up. As a consequence, the Prosodic Hierarchy loses its predictive power in stating what a prosodic structure can look like, what is universal about it, and which prosodic systems are possible in the languages of the world.
2.3.3 Summary
Under the Generality Assumption, two distinct levels of prosodic structure can be established for Vietnamese, namely the syllable as the locus for generalizations related to phonotactics, stress and tone, and the phonological phrase, which provides the basis for the iambic (or anapestic) rhythm of the language. Neither the prosodic foot nor the prosodic word can be motivated for the language. If we formalize these findings in a structure that encompasses all levels of the Prosodic Hierarchy, we face the problem of overgeneration since several levels are represented for the individual language although they are not actually manifested. In order to avoid this inadequacy, we opt for a structure that dispenses with the unmotivated levels of the foot and the word.Footnote 17 In the context of the Strict Layer Hypothesis, this solution forces us to postulate a prosodic tree which skips the levels of the foot and the prosodic word, and thus violates Strict Succession as well as (Proper) Headedness.
The case study from Vietnamese suggests that the parsing of the prosodic word domain as well as Strict Succession or (Proper) Headedness are essentially violable in individual languages. If this observation also translates to other levels of the Prosodic Hierarchy and other components of the Strict Layer Hypothesis, the overall predictive power of the theory is seriously challenged.
3. CHALLENGE 2: LIMBU, OR MULTIPLE PROSODIC WORD DOMAINS
Vietnamese and similar languages challenge the theory of the Prosodic Hierarchy by evidencing fewer domains than expected. Another true challenge, we claim, comes from Limbu, an Eastern Kiranti (Sino-Tibetan) language of Nepal (Weidert & Subba Reference Weidert and Subba1985; Michailovsky Reference Michailovsky1986, Reference Michailovsky2002; van Driem Reference van Driem1987, Reference van Driem and Bradley1997; Ebert Reference Ebert1994; Bickel Reference Bickel1998, Reference Bickel, Yadava and Glover1999, Reference Bickel, Bradley, LaPolla, Michailovsky and Thurgood2003; Hildebrandt Reference Hildebrandt2007). For this language, at least two distinct prosodic word domains must be recognized when standard tests for motivating prosodic domains are applied, arranged in such a way that the multitude cannot be reduced by positing additional domains, recursive structures or prosodic subcategorization. We first delimit the grammatical word, and then turn to prosodic domains.
3.1 Morphological domains in Limbu
The grammatical word in Limbu is characterized by the morphological properties of different morphemes, such as stems, affixes and enclitics as well as the morphological processes that govern their combination. Affixes, which are attached to the stem by prefixation, suffixation or circumfixation, are realized in various positions within nominal and verbal word forms. Stems themselves can be combined to form complex compound verbs. Limbu also has a class of enclitics that syntactically have phrasal distribution and scope, but nevertheless are integrated into a phonological domain at the word level.
3.1.1 Affixation
In both nominal and verbal morphology, the grammatical word may be defined as consisting of grammatical elements that must occur together following a strictly ordered template. Nouns inflect for possessor agreement (in person and number), number, definiteness and case. Definiteness is marked only in the singular (by -ʔin, which in turn assimilates to -ʔil before the ergative/instrumental in -le) and only in the absolutive, genitive, and ergative/instrumental case.Footnote 18 Since singular number is zero-marked, the two categories appear in a single position: poss–stem–def/number–case.
Nominal inflection is obligatory, except for non-relational nouns where possessor agreement is optional, and for inanimate nouns where number marking is optional. All affixes realizing these categories exclusively subcategorize for nominal stems, which indicates that we are dealing with morphological elements and not independent phrasal elements. The only exception is the ergative/instrumental case, which is also found on finite, nominalized verb forms. However, when the ergative/instrumental appears on such forms (as well when following derivational suffixes in -a), definiteness can no longer be marked, and the sequence definite-ergative (-ʔin-le >-ʔille) is reanalyzed as the post-consonantal allomorph of the plain ergative -lle which appears only and always after vowels. Derivational morphology of nominal stems shows regular affixation, but also circumfixation. One case of circumfixation is the formation of free from bound color roots such as mak ‘black’. In this derivation, the discontinuous affix ku-root-la is circumfixed to the color morph, e.g. ku-mak-la ‘black’.
Verb morphology makes use of the same affixation types, but it involves many more affixes and categories. The Limbu verb agrees with the intransitive subject, and among transitives, with both the subject and the primary object (i.e. the sole object of monotransitives and the most goal-like argument of ditransitives). Agreement categories are person and number. Verbs also inflect for tense, aspect and mood. The marking of all of these categories is distributed across four prefix and thirteen suffix positions. Evidence that that entire string of prefixes, a stem (or several) and suffixes form a single grammatical word comes from the following: (i) all inflectional categories are obligatory in finite verbs; (ii) all affixes involved are exclusively subcategorized for verb stems and do not appear on other stems or word forms; (iii) some categories are simultaneously expressed by a prefix and one or more suffixes, e.g. negation; (iv) no phrasal element can ever intrude, including clitics. That all inflectional elements are contained within the same grammatical word, and that none of them behaves like a clitic is furthermore shown by the fact that agreement affixes (prefixes and suffixes in verbs, prefixes in nouns) regularly co-occur with agreement-triggering NPs in argument position, and that agreement affixes cannot be gapped under identity.
3.1.2 Complex stems
Both verbal and nominal stems can form compounds, but they differ in their morphological structure (and as we will see, in their phonological structure as well). The following are nominal compounds:
(19) Nominal compounds in Limbu (van Driem 1987: 27, 54)Footnote 19
All these are comprised of two nominal stems and are presented in their bare forms in the first column. The second column illustrates these compounds in affixation. In (19a), one possessive prefix is attached at the left edge of the compound, i.e. in this case, the compound behaves like a simple stem. In (19b), on the other hand, prefixation of the possessive marker applies twice, both to the first and the second member of the compound. For the sake of prefixation, the compound in (19b) thus constitutes two morphological stem domains. With respect to suffixation, example (19c) shows that even such compounds that form two stem domains in prefixation behave like single stems when it comes to plural marking.
Verbal compounds differ from nominal compounds with regard to prefixation possibilities. Consider the data in (20).
(20) Verbal compounds in Limbu (van Driem 1987)
The first column presents complex stems in infinitival (citation) form. The second column shows affixation to these stems. While in nominal compounds, agreement prefixes are found on one or both stems, in verbal compounds prefixes only ever occur on the rightmost stem.Footnote 20
3.2 Prosodic words in Limbu
Sino-Tibetan linguistics has traditionally focused on the syllable as the key domain of both phonological and morphological regularities in the languages of the family (e.g. Matisoff Reference Matisoff, Traugott and Heine1991a, Reference Matisoffb), with occasional extensions to the foot (e.g. Bickel Reference Bickel, Bradley, LaPolla, Michailovsky and Thurgood2003). However, a number of studies have revealed the relevance of additional prosodic domains in the application of phonological processes. For instance, the tone domain in many Bodish languages includes roots and various bound morphemes and can, accordingly, best be analyzed with appeal to the prosodic word (Mazaudon Reference Mazaudon1973, Denwood Reference Denwood1999, Hildebrandt Reference Hildebrandt2003, Noonan Reference Noonan, Thurgood and LaPolla2003). In what follows, we present an analysis of the prosodic word domains in Limbu. We discuss how multiple prosodic word domains are referenced by phonological generalizations and processes.Footnote 21 As a byproduct it turns out that the syllable and the bimoraic trochaic foot have salient status in the prosodization of words. In order to distinguish the various prosodic words from higher levels of prosodic structure, we also discuss processes that apply at the level of the phonological phrase. The data discussed in the following make our second challenge to the Prosodic Hierarchy obvious enough: phonological processes do not always cluster on one prosodic word domain intermediate between the foot and the phonological phrase.
3.2.1 The major prosodic word
In Limbu, a number of phonological patterns converge on a prosodic domain referencing the stem, its prefixes, its suffixes, and enclitics. Since this domain includes a larger portion of the grammatical word than the domain to be discussed in Section 3.2.2, we refer to it as the major prosodic word. This domain can be motivated with appeal to at least two processes: the suprasegmental rules of word-level stress placement and the segmental rule of coronal to labial assimilation.
The first piece of evidence for the major prosodic word comes from stress phonology. Since stress in Limbu ‘is not very pronounced and … non-distinctive’ (van Driem Reference van Driem1987: 15), grammatical descriptions have so far only rudimentarily treated the distribution and placement of primary stress. In the majority of cases, stress placement is predicable in a straightforward manner from the morphological composition of the word. To substantiate the claim that the stress domain includes the stem, all its affixes and enclitics, we present a more thorough analysis of stress in Limbu, based on a detailed phonetic study.Footnote 22
(21) Stress in Limbu
(a) /ku-laːp/ [ˈkulaːp] 3poss-wing‘its wing’
(b) /peːg-i/ [ˈpeːgi] go-1pl.s‘We go.’
(c) /a-oŋ-eː/ [ʔaˈʔoŋˌŋeː] 1poss-brother.in.law-voc‘My brother in law!’
(d) /ku-taŋ=mο/ [kuˈtaŋmο] 3poss-horn=ctr‘its horn, on the contrary’
(e) /mο-thaŋ-e=aŋ/ [mοˈthaŋˌjaŋ] 3ns-come.up-pst=and‘they come up and …’
With respect to the distribution of stress, the forms in the examples above all take only one primary stress. This generalization holds true for combinations of prefix and stem, as in (21a); stem and suffix, (21b); prefix, stem and suffix, (21c); prefix, stem and enclitic, (21d); and finally prefix, stem, suffix and enclitic (21e). On the basis of this evidence we can postulate a prosodic word domain defined by the presence of one and only one primary stress and referencing a morphosyntactic structure which contains the stem, its prefixes, its suffixes and enclitics, e.g. (mο-ˈthaŋ-ˌe-aŋ)ω.
The data in (21) also allow us to analyze rules of stress placement within the stress-defined word. (21b–e) show the default pattern, in which primary stress is realized on the stem-initial syllable. Depending on the morphological structure, this entails that stress will be word-initial if there is no prefix present, as in (21b), and on the second syllable if the word is prefixed. A deviation from this default stress placement is illustrated in (21a). Stem-initial stress would result in an illicit iambic foot here and stress shifts to the prefix.
The stress facts exemplified above also suggest an additional level of prosodic organization, namely the foot. In this context, the examples (21c) and (21e) illustrate that within the prosodic word several bimoraic stress domains are constructed in which primary and secondary stresses are assigned. In (21c), the bimoraic stem syllable carries primary stress and the following bimoraic syllable is parsed as a foot for the sake of secondary stress, i.e. (ʔa(ˈʔoŋ)ϕ(ˌŋeː)ϕ)ω. The same principle of stress assignment is also evidenced in (21e), such that the string will be prosodized as (mο(ˈth aŋ)ϕ(ˌjaŋ)ϕ)ω. Note that monomoraic syllables within the word remain unfooted. Accordingly, no secondary stress is assigned in (21d) where the final syllable does not constitute a bimoraic foot, i.e. (ku(ˈtaŋ ϕ)mο)ω.
The second piece of evidence which motivates the major prosodic word domain comes from a productive segmental rule of regressive coronal to labial assimilation:
(22) Regressive coronal to labial assimilation (van Driem 1987: 17)
The rule states that the coronal phonemes /t, n/ regressively assimilate for place of articulation to the bilabial phonemes /m, p/, respectively. This rule is sensitive to the prosodic word domain as defined above and its application includes stems, prefixes, suffixes and enclitics, as illustrated by the following data.
(23) Coronal to labial assimilation in Limbu (van Driem Reference van Driem1987: 17, 136, 230)
(a) /ɔːmɔt-maʔ/ [ʔɔːmɔpmaʔ] look.at-inf‘to look at’
(b) /mο-n-mο t-paŋ/ [mο mmο ppaŋ]ns.a-neg-tell-1sg>3.pst‘I did not tell him’
(c) /si-aŋ-mο n-pa/ [sjaŋmο mba] die-1sg.s.pst-cond-ipfv‘I might die’
(d) /myaŋluŋ=phοlle hο n=phοlle/ [mjaŋluŋbhοlle hο mbhοlle] Myaŋluŋ=sub what=sub‘What does Myaŋluŋ mean?’
In (23a), the rule of regressive coronal to labial assimilation applies across the morpheme boundary of the stem ɔːmɔt ‘look at’ and the nominalizer suffix -maʔ. In the second example, the application of the rule is demonstrated for a morphologically complex word consisting of a stem, two prefixes and a suffix. The coronal of the second prefix n- assimilates to the following labial of the stem mοt and the stem-final coronal assimilates to the labial of the suffix -paŋ. (23c) shows that the rule also applies across the morpheme boundary of two adjacent suffixes. When the conditional suffix -mοn is followed by the imperfective marker -pa, its final coronal assimilates to the labial of the following suffix. Note that in this example the labial of the second suffix also undergoes voicing assimilation from /p/ to [b] (discussed in Section 3.2.3). (23d), finally, shows that clitics are integrated into the prosodic word domain for coronal to labial assimilation. The element -phοlle is a subordinator attaching to the right margin of a clause. As a result of cliticization, the initial labial of the clitic triggers regressive assimilation of the final coronal in the host word hοn ‘what’. Additionally, the initial labial is subject to voicing assimilation from /ph/ to [bh]. In terms of prosodic status, the phrasal subordinator behaves exactly like the imperfective suffix -pa in example (23c). We can sum up the discussion of coronal to labial assimilation by concluding that this morphophonological process applies across morpheme boundaries in the major prosodic word domain which references the stem, its affixes and enclitics.
Coronal to labial assimilation does not cross word boundaries, as is evidenced by a separate but related allophonic alternation applying to the coronal plosive /t/. Of relevance here are two variants of /t/: an unreleased, glottalized variant [tʔ], which occurs in word-medial coda position, and a variant comprised of a glottal-plus-plosive cluster with simultaneous lateralization [ʔtl], occurring in word-final position (van Driem Reference van Driem1987: 3). Only the word-medial, glottalized variant is subject to the coronal to labial assimilation rule, while the word-final, lateralized variant does not undergo this assimilation:
(24) Allophonic alternation of /t/ in Limbu (van Driem Reference van Driem1987: 377, 485, 279)
(a) [ɔːmɔtʔ-n-i-ŋ] look.at-1>2-pl.p-1sg.a‘I look at you (plural).’
(b) [ɔːmɔʔtl] look.at‘look at’
(c) [ɔːmɔp-maʔ] look.at-inf‘to look at’
(d) nirοŋ-sumlοŋ [cwaʔtlmaːʔtl] lɔk piːt-maʔ poːŋtwice-thrice water be.finished only suck-infexig‘If it [beer mixture] has sat for many days, it will ripen…’
(24a) shows the word-medial allophone of /t/, while (24b) shows the word-final allophone. Word-medial /t/ undergoes coronal to labial assimilation in (24c). Crucially, example (24d) demonstrates that across word boundaries (within a clause), the lateralized glottal-plosive variant occurs, and assimilation does not apply.
3.2.2 The minor prosodic word
Although the motivation of the major prosodic word in Limbu is unproblematic with appeal to the standard methodology of analyzing prosodic structure, a complication arises when other salient word-related phonological processes of the language are also considered. These phonological patterns converge on a prosodic domain which is smaller than the major prosodic word in that it systematically excludes the prefix. We refer to this domain as the minor prosodic word. This domain can be motivated with appeal to two generalizations: the [l] ~ [r] alternation and the process of [ʔ]-insertion.
One regularity referencing the minor word is the alternation between the two allophones [l] and [r] of the phoneme /l/.Footnote 23 The distribution of these can be stated with exclusive reference to syllable structure, i.e. with reference to the syllable as a prosodic domain, only for two contexts. First, in loans, /l/ is realized as [l] in syllable-final position, e.g. beːl ‘plant name’. Second, in native words, the phoneme surfaces as [r] when it is the second member in initial clusters, for instance in the second syllable of cɔk.krɔk.ma ‘uvula’. Apart from this, the distribution of the allophones is governed by word structure. Syllable-initial /l/ always surfaces as [l] in word-initial position, while in word-medial position it is realized as either [r] or [l], depending on the structure of the preceding syllable. The realization is [r] if the preceding syllable is open or ends in a glottal stop (which, accordingly, could be analyzed as a vocalic feature: Michailovsky Reference Michailovsky1986). The realization remains [l] after closed syllables, i.e. after consonants, see (25).
(25) Distribution of syllable-initial, word-medial /l/ in Limbu (van Driem Reference van Driem1987: 4ff.)
/l/→[r]/{V, ʔ}__
The domain to which this distributional rule applies is smaller than the major prosodic word because it always excludes prefixes. Consider the following data, which illustrate the application of the rule in morphologically simple and complex words. The forms on the left show /l/-onsets after open syllables, the forms on the right after closed syllables.
(26) The [l] ~ [r] alternation in Limbu (van Driem Reference van Driem1987: 4ff.)
(a) nοrοt vs. lupli
‘heart’ ‘earthquake’
(b) pha-re siŋ vs. mik-le raŋ
bamboo-gen wood eye-gen color
‘the wood of bamboo’ ‘the color of the eyes’
(c) kο-lɔʔ vs. mο-l-lο-baŋ (<mο-n-lο-baŋ)
2-say neg-neg-know-1sg>3.pst
‘you say’ ‘I didn't know [it].’
In line with (25), /l/-onsets are realized as [r] in post-vocalic position and as [l] in post-consonantal position (26a–b). However, as shown by the data in (26c), when /l/ is located at a prefix–stem boundary, it does not exhibit the predicted alternation and surfaces as [l] both after open and after closed syllables. This suggests that prefixes are outside the domain in which the /l/-alternation applies.Footnote 24 This can be accounted for by positing a prosodic word structure like kο-(lɔʔ)ω. In this structure, where stem-initial /l/ now appears at the left edge of the prosodic word, it always surfaces as [l].
While at the beginning, the domain of the /l/-alternation is delimited by the stem boundary, at the end, it includes everything that is part of the same grammatical word (suffixes, other stems in compounding) and even clitics. The following data show the predicted alternations with enclitics and complex stems.
(27) The [l] ~ [r] alternation with enclitics and complex stems
(a) peːg-i=roː vs. peːg-aŋ=loː
go-pl=ass go-1sg.pst=ass
‘Come on, let's go!’ ‘I'm on my way!’
(b) kɔŋ lοːs-u=rəcə vs. maːŋgha kο-n-nis-u-n=ləcə
this know-3p=depr far 2-neg-see-3p-neg=depr
‘He appears to know.’ ‘You seem to be myopic.’
(c) yaʔ+raːk-maʔ vs. laːk-maʔ
paddy+trample-inf trample-inf
‘to perform the rice dance’ ‘to trample, stamp, walk underfoot, kick’
In all cases, /l/-onsets surface as [l] after consonants and as /r/ after vowels and the glottal stop within the minor prosodic word.
If, as we claim, the domain includes everything following the initial stem, this predicts that prefix–stem boundaries inside this domain would show the /l/-alternation, unlike boundaries between prefixes and the initial stem. Domain-internal prefix–stem boundaries are regularly found as a result of verb compounding, because, as noted in Section 3.1.2 above, only the rightmost stem can ever host prefixal morphology in such compounds. The prediction is borne out by the following data.
(28) yaʔ+gο-raːkt-u paddy+2-trample-3p ‘You performed the rice dance.’
In sharp contrast to the non-cohering status of the prefix with simple stems (see (26c)), the prefix which is attached to the rightmost syllable of the first part of a complex stem is prosodically integrated into the prosodic word domain. The initial segment of laːk-maʔ, accordingly appears in word-medial position in such a construction, i.e. (yaʔ+gο-raːkt-u)ω, and therefore surfaces as [r].Footnote 25
The same domain as the one referenced by the /l/-alternation is also relevant for the distribution of non-contrastive glottal stops in Limbu.Footnote 26 One variant involves glottal stop insertion.Footnote 27 This is found before vowel-initial syllables at the left edge of initial stems.
(29) Glottal stop insertion in Limbu (van Driem 1987: 15)
(a) /iŋghɔŋ/ [ʔiŋghɔŋ] ‘message’
(b) /ku-iŋghɔŋ/ [kuʔiŋghɔŋ] 3poss-message ‘his news’
(c) /a-iːr-ο/ [ʔaʔiːrο] 1-wander-pst ‘We (plural, inclusive) wandered.’
(29a) shows a monomorphemic word iŋghɔŋ ‘message’ which begins with a vowel underlyingly. In the surface form, the empty onset position is filled by the prothetic glottal stop. In (29b) the same word forms the base for the prefixation of the possessive prefix ku-. The glottal stop is still inserted to provide an onset for the vowel-initial stem. This shows that the prefix is not included in the domain referenced by the word-based rule of glottal stop insertion, see also (29c). Unlike prefixes, suffixes and enclitics are integrated into the word domain referenced by glottal stop insertion. Therefore, if a vowel-initial suffix or clitic follows a stem or another suffix, no glottal stop is inserted. In these positions, hiatus is not resolved by glottal insertion,Footnote 28 but by diphthongization (or, trivially, by resyllabification if a vowel-initial suffix follows a consonant or by glide formation if the suffix follows a vowel that has a glide alternant). The following data show this with minimal pairs of /u–e/, /a–e/, /ο–i/, and /a–i/ sequences; the forms on left involve prefix–stem, the forms on the right stem–suffix and host–clitic boundaries, respectively.
(30) Glottal stop insertion/diphthongization across different morpheme boundaries
(a) /ku-eːk/ [kuʔeːk] vs. /a-mphu-eː/ [ʔamphueː]
3poss-back 1poss-brother-voc
‘its/his/her back’ ‘Brother!’
(b) /a-eːk/ [ʔaʔeːk] vs. /yuma-eː/ [yumaeː]
1poss-back grandmother-voc
‘my back’ ‘Grandma!’
(c) /kο-im/ [kοʔim] vs. /naks-ο=iː/ [naks
iː]2-sleep go.crazy-pst=q
‘You sleep.’ ‘Has he gone crazy?’
(d) /a-iːr-ο/ [ʔaʔiːrο] vs. /nu-ba=iː/[nubaiː]
1-wander-pst be.alright-nom=q
‘We wandered.’ ‘Is this good?’
iː]The only difference between the forms with prefixes and those with suffixes is that prefixes are outside of the prosodic domain in which glottal insertion applies, while suffixes are inside this domain. Thus, the domain is the same as the one observed above for the /l/-alternation, i.e. prefix–(stem–suffix=clitic)ω.
To sum up, /l/-alternation and glottal prothesis evidence a prosodic word including stems and suffixes (and enclitics) but crucially excluding prefixes. This is in conflict with other phonological processes of Limbu referencing the major prosodic word domain that includes stems, prefixes, suffixes and enclitics.
3.2.3 The phonological phrase
In order to rule out that the major prosodic word is really a phrasal domain and not a word domain, we turn to higher levels of prosodic structure, i.e. the phonological phrase. There is at least one phonological process that applies across word boundaries within this domain. (31) illustrates that word-initial plosives optionally assimilate to the preceding word with respect to the feature voice (which is not distinctive in Limbu).
(31) Voicing assimilation in Limbu (van Driem 1987: 18f.)
A: anige hοn joːkmaʔ? cf. /coːkmaʔ/ ‘What shall we do?’
B: peːkmaʔ boːŋ. cf. /poːŋ/ ‘It's time [for you] to go.’
In the first turn of this conversational exchange, the word coːkmaʔ ‘to do, make, build’ is realized following a word ending in a voiced segment /n/. According to the phrasal rule of voice assimilation, the word-initial unaspirated voiceless lamino-postalveolar affricate /c/ [ʨ] is replaced by its voiced counterpart [ʥ] (written as 〈j〉) in this context. The auxiliary word poːŋ in the second sentence of (31) assimilates to the preceding word in the same fashion. Note that the glottal stop does not block voice assimilation in this example, cf. its vocalic status noted in (25).
Nominal compounds constitute two word domains for the sake of the /l/-alternation. For instance, in compounds such as haʔ+luŋ (fire+stone) ‘fireplace’ and makhi+lam (blood+road) ‘artery’ the initial segment of the second compound member always surfaces as [l], although it appears after a glottal stop in the first compound and intervocalically in the second example. This suggests that /l/ in these examples does not appear in word-medial position, but in word-initial position, so that both stems appear to have prosodic word status, i.e. (haʔ)ω+(luŋ)ω and (makhi)ω+(lam)ω. But the two prosodic words are conjoined into a single phonological phrase. Let us reconsider the nominal compound cum+deːŋ ‘buddy’, which is composed of the two elements cum ‘friend’ and teːŋ ‘comrade’. Here, the initial plosive of the second member surfaces in its voiced counterpart in the compound. In an analysis that treats the two stems as two separate prosodic words which are joined in a phonological phrase, this fact is predicted by the application of the phrase-level rule of voicing assimilation, i.e. [(cum)ω+(deːŋ)ω]P. In summary, nominal compounds fit into the higher domain structures in constituting two prosodic words, which are in turn subsumed under a single phrase node.
Our knowledge of Limbu phonology is still too limited to propose an analysis of intonation and other phonological processes which might allow the postulation of further prosodic domains for the language, such as the intonational phrase and the utterance. However, our point is that by concentrating on the phonological processes which operate between the levels of the syllable and the phonological phrase we find highly specific and overlapping domains whose structure is not predicted by the Prosodic Hierarchy.
3.3 Discussion
In the previous section, we discussed the word-level prosodic constituency in Limbu, while making reference to other domains of prosodic structure, namely the syllable, the foot and the phonological phrase. At first sight, the domains which need to be distinguished in Limbu adhere to the prosodic layers predicted by the Prosodic Hierarchy. However, the crucial deviation lies in the fact that the word-related processes themselves do not converge on a single domain, but constitute two domains which reference different portions of the grammatical word. The first one, which references the entire grammatical word+enclitics, is evidenced by stress assignment and coronal to labial assimilation. The second one, which references the initial stem, subsequent stems, suffixes and enclitics, but which excludes the prefix, is instantiated by the [l]~[r] alternation and the [ʔ]-insertion.
3.3.1 A prosodic hierarchy for Limbu
If we follow the Prosodic Hierarchy blindly, we would have to postulate the following prosodic tree for Limbu.
(32) A prosodic tree for Limbu
This representation makes a number of correct predictions concerning low level constituency in Limbu. The syllable captures the syllabification of the given string, which includes the resyllabification of the stem-final consonant to fill the onset position of the last syllable. The second and third syllables constitute bimoraic feet, which qualifies them as the locus of stress placement at the higher level of the word. Here, the stem syllable is assigned primary stress and the final syllable is assigned secondary stress. The initial, monomoraic syllable is adjoined to the prosodic word without being parsed as a foot. Although this violates Strict Succession, it is still in line with the principle of Proper Headedness. The word node in this example provides the domain for stress placement and for the application of the coronal to labial assimilation. Finally, the inclusion of the prosodic word in the phonological phrase captures the possibility that the final voiced element of the word will trigger voicing of the initial segment of the following prosodic word within the phonological phrase.
However, the prosodic representation in (32) suffers from some inconsistencies. These relate to the fact that only one word domain can be recruited from the universal repertoire of domain types provided by the Prosodic Hierarchy. With appeal to the prosodic tree given above it is impossible to formalize the fact the [l]~[r] alternation and the [ʔ]-insertion reference a domain which is smaller than that delimited by the word node in the tree, i.e. the prosodic tree undergenerates. An obvious solution would be to posit a second word domain intermediate the foot and the major prosodic word to optimize the descriptive adequacy of the representation.
(33) A revised prosodic tree for Limbu
In (33), the minor word node provides the domain for [l]~[r] alternation and the [ʔ]-insertion, whereas the major word node provides the domain for stress assignment and coronal and labial assimilation. Although the revised prosodic tree correctly predicts the application of all relevant prosodic patterns, it violates the Clustering Prediction, which specifies that word-related processes should converge on one and only one prosodic word domain. Second, since the major word node immediately dominates a node of the same prosodic category, i.e. the minor word, the prosodic tree also violates Strict Succession or Proper Headedness.
It is noteworthy that a picture similar to that of Limbu is offered by other Eastern Kiranti languages and is likely to be a diachronically robust pattern of this branch of the family (perhaps even beyond that). In Belhare, for example, there is an intervocalic voicing rule applying within stem–suffix strings but not across prefix–stem sequences; and a rule of deleting velar stops at the end of a domain including prefixes along with the stems and suffixes (Bickel Reference Bickel1996, Bickel & Hildebrandt Reference Bickel and Hildebrandt2005). In Chintang, a rule of glottal insertion (similar to the one in Limbu) and the host definition of endoclitics identifies a domain excluding prefixes; rules of stress assignment and the host definition of enclitics identifies a domain that includes prefixes (along with stems and suffixes); a third domain is referenced by intervocalic voicing which optionally applies within sequences of prefixes, or within sequences of stems and suffixes, but not across prefix–stem boundaries (Bickel et al. Reference Bickel, Banjade, Gaenszle, Lieven, Paudyal, Rai, Rai, Rai and Stoll2007).
3.3.2 Multiplying prosodic word domains
Limbu is not the only language which deviates from the predictions of the Prosodic Hierarchy because its prosodic system distinguishes multiple non-isomorphic domains for word-related processes. Better-known examples come from Bantu (Hyman Reference Hyman1987, Reference Hyman1998, Reference Hyman2008; Odden Reference Odden1987; Reference Odden and Katamba1995, Reference Odden1996; Myers Reference Myers, Hyman and Kisseberth1998; Downing Reference Downing2001; Bickmore Reference Bickmore2007) or Athabaskan (McDonough Reference McDonough, Fernald and Platero2000). What makes the challenge from Limbu extraordinary is the fact that even the additional mechanisms of prosodization which have been proposed in the literature cannot resolve the problems raised by the data.
One answer to deviations from the Prosodic Hierarchy at the level of the word simply lists them by specifying which morpheme strings (e.g. stem–suffix vs. stem=clitic vs. stem–suffix=clitic strings) support which prosodization. This is obviously not a theoretically appealing solution to begin with, but, worse, any solution along these lines fails to account for the fact that in Limbu, one and the same morpheme string can simultaneously support two non-isomorphic prosodic structures. In a string like ku-laːp (3poss-wing) ‘its wing’, which surfaces as [ˈkulaːp], both major and minor words are present simultaneously. The minor word domain parses the string as ˈku(laːp)ω for the sake of the [l]~[r] alternation, which accounts for the fact that the stem-initial /l/ surfaces with its word-initial allophone [l]. The major word domain parses the same string as (ˈkulaːp)ω for the sake of stress placement and this accounts for initial stress. Crucially, unlike languages like Dutch where two classes of suffixes can be distinguished which differ in their prosodic status, the prefix in Limbu is at the same time integrated into a word domain and excluded from another. As a result, one would have to simultaneously claim that ku- is part of the prosodic word and that it is not part of the prosodic word.
To make this point even clearer, the Limbu situation is crucially different from well-known cases of non-cohering prefixes as, for instance, in Bantu. In Kimatuumbi, for example, the object prefixes and the inflectional stem constitute the ‘superstem’ domain. The next higher level domain of the ‘simple verb’ encompasses the superstem plus the preprefixed subject and tense markers (Odden Reference Odden1996: 71). In such a system, there are distinct types of prefixes, differentiated with respect to which domain they attach to. But it is not the case that a given prefix is at the same time within and outside a domain. It is not the case in Kimatuumbi that the object prefix is part of the inflectional stem for the sake of some process and at the same time outside this domain for another process.Footnote 29
Yet this is exactly what one finds in Limbu.
A second solution to multiple word domains offered in the literature is to posit an additional prosodic domain. An argument along those lines would re-categorize the major word domain in the prosodic tree (33) as a Clitic Group node (C), yielding the bracketing (ˈku(laːp)ω)C for the form discussed above. In Nespor & Vogel's (Reference Nespor and Vogel2007) original proposal, the C domain references clitic elements which can be identified by morphosyntactic criteria and binds them to their phonological host. The crucial evidence for the motivation of such an intermediate domain comes from regular phonological processes which specifically apply in host–clitic combinations, such as enclitic stress in Latin. The Limbu word domains cannot be resolved with appeal to such a domain because the prefix in kulaːp is not a clitic in either phonological or morphological terms: it is not a clitic in phonological terms because all phonological processes that link it to the stem also apply between stems and suffixes (cf. the coronal to labial assimilation and the stress placement rules), i.e. it does not trigger a special host–clitic phonology. The prefix is not a clitic in morphological terms, either: as noted in Section 3.1, Limbu possessive agreement prefixes, like all other nominal affixes, exclusively subcategorize for nominal stems and are therefore not phrase-level affixes; they are regular elements of noun inflection; they cannot be gapped under identity; they systematically co-occur with agreement-triggering argument NPs; and, as agreement markers, one would conventionally analyze their scope as being over the head, not the phrase. Moreover, those elements in Limbu, which are true phrase-level clitics, are regularly included in both word domains under discussion, as demonstrated in Sections 3.2.1 and 3.2.2. Analyzing (despite all evidence) the prefixes as cliticsFootnote 30 would therefore necessitate two distinct Clitic Group nodes, one node that includes enclitics and prefixes analyzed as proclitics and one node that includes enclitics, but no proclitics. As a result, the Clustering Prediction would again be violated, only now on the Clitic Group instead of the word level.Footnote 31
A similar problem arises if one tries to address the Limbu challenge by equating the major word domain with the phonological phrase, e.g. (ˈku(laːp)ω)P. This would simply shift the problem of multiple domains from the word to the phrase level because there already is clear evidence from intervocalic voicing (reviewed in Section 3.2.3) for a phonological phrase that encompasses longer sequences than any of the two word domains.
A third formal mechanism which has been proposed in the literature to account for multiple domains is recursion. One such analysis is Peperkamp's (Reference Peperkamp and Kleinhenz1996) analysis of stress assignment in Neapolitan Italian. Applied to Limbu, one would then posit a stacked structure like (ˈku(laːp)ω)ω, in which both word domains are of the same rank. In such an approach the smaller domain would first be built, and in a second step, the larger domain would be constructed by the recursive application of the domain-defining processes. But, in contrast to the Italian stress data, the evidence for two domains in kulaːp does not come from recursive application of one and the same phonological process, but from very distinct processes: [l]~[r] alternation and glottal stop insertion for the smaller and stress assignment and coronal to labial assimilation for the larger domain. Put differently, a recursive structure entails that the stacked domains are one and the same prosodic domain. This would then predict that they have the same phonological properties. For instance, applied to Limbu, it would predict that the same principles of stress assignment will recursively apply at all levels. In the stacked structure (ku(laːp)ω)ω, stress should accordingly be assigned to the smaller domain first and then again within the larger domain. The actual form ˈkulaːp proves this prediction patently false.Footnote 32
A fourth possible solution lies in assigning different domains to different phonological tiers, such as the tone and quantity domains in Luganda (Hyman, Katamba & Walusimbi Reference Hyman, Katamba and Walusimbi1987). However, this solution cannot be carried over to Limbu because the different domains are associated with phonological patterns from different phonological tiers: for example, the segmental process of coronal to labial assimilation and the suprasegmental process of stress assignment converge on the major word domain. In Limbu, the different prosodic word domains are not exclusively confined to one specific phonological tier but cross-cut several components of the phonology.
A fifth solution is to assume a much more fine-grained theory of prosodic structure (Downing Reference Downing, Hall and Kleinhenz1999, Reference Downing2001). Even under this approach, however, there remain crucial problems for the analysis of Limbu. Following the proposed distinction between the prosodic stem and the prosodic word has no consequences for the analysis of Limbu since all the phonological patterns relevant for motivating prosodic domains referencing morphological structure target larger portions of the grammatical word and not exclusively the stem. It might be objected that the prosodic stem might show language-particular mappings to morphological structures extending the stem, just like the prosodic word might include or exclude prefixes across language-particular grammars. From that point of view, one probably could state mapping rules for Limbu which incorporate the stem–suffix–enclitic sequence into a prosodic stem domain. As we saw above with the Clitic Group approach, the crucial problem in this proposal is the fact that the lexical and post-lexical level are mixed in non-trivial ways. Already at the level of the prosodic stem, post-lexical clitics are integrated and the derivation would need to step back to include lexical prefixes at a later stage. Note that this problem seems not to arise with the Bantu languages for which the prosodic stem domain has been motivated. In languages like Cilungu (Bickmore Reference Bickmore2007), there is evidence for a principled prosodic distinction between the stem and the macrostem integrating lexical affixes at different levels. Clitics, however, are prosodized in a higher prosodic domain, presumably the Clitic Group.Footnote 33
Finally, one might opt for a framework in which the Prosodic Hierarchy is replaced by morpheme-bound prosodic subcategorization frames (Inkelas Reference Inkelas1989). The varying status of the prefix across the different word domains in Limbu would be an equal challenge for analyses in such terms. One and the same Limbu prefix is phonologically both cohering and non-cohering; the choice is not lexical but depends on the rule one looks at. The prefix would therefore have to come with two simultaneous subcategorization frames, specified for the relevant phonological processes: one frame selecting prosodic words would have to be specified for the [l]~[r] alternation, while another frame selecting some smaller prosodic constituents (e.g. bimoraic feet) would have to be specified for stress assignment. Although we are sympathetic to construction-based co-phonologies (Itô & Mester Reference Itô and Mester1995, Orgun Reference Orgun1996, Antilla Reference Antilla2002, Booij Reference Booij, Dressler, Kastovsky, Pfeiffer and Rainer2005, Inkelas & Zoll Reference Inkelas and Zoll2005), this problem also extends to this more recent theory of the morphology–phonology interface. This model assigns a specific co-phonology to individual morphological constructions, such as, for example, an affixation construction. In the Limbu case, the prefixing construction /ku/-[laːp], would have to be associated with one co-phonology which parses the prefix as part of the prosodic word and at the same time a second one which parses it as outside the prosodic word.
3.3.3 Summary
Having presented an in-depth analysis of the phonological and morphological properties of words in Limbu and situating the evidence in the broader picture of recent achievements in the theory of prosodic structure, we are now able to precisely formulate the language-particular challenges to the theory of the Prosodic Hierarchy.
Accepting the standard analytic tools for motivating a prosodic domain forces us to distinguish two distinct levels of prosodic word structure for Limbu. The major word provides a domain for stress assignment and the segmental process of coronal to labial assimilation. The prosodic domain references a stem, its prefixes, suffixes, and enclitics. The minor word is the locus of the [l]~[r] alternation and initial glottal stop insertion. The morphological string that this domain targets consists of the stem, its suffixes and enclitics, to the exclusion of the prefix. Furthermore, phonotactic restrictions and resyllabification evidence the syllable domain, secondary stress assignment within the word motivates the bimoraic foot, and a rule of voicing assimilation across words suggests the relevance of the phonological phrase. If we attempt to represent these findings in a structure that only employs the levels of the Prosodic Hierarchy, we face the problem of undergeneration. In such a prosodic tree, the fact that the domain of some word-related processes includes the prefix and the domain of some other processes excludes the prefix cannot be captured. To avoid this inadequacy, we opt for the postulation of two distinct prosodic word domains, the major and the minor word. In the context of the Strict Layer Hypothesis, this solution forces us to accept violations of Clustering and Strict Succession.
In this context, we point to Itô & Mester's Reference Itô, Mester and Parker2009 version of the prosodic hierarchy, which allows prosodic adjunction to multiply levels of prosodic structure. In their framework, a minimal word can in various steps be enlarged by prosodic adjunction to constitute a larger word domain, up to the maximal word domain. In face of the Limbu data, this move seems to be justified, since it would allow us to refer to the minor word for the sake of some processes and to the major word for the sake of others. However, this approach has itself serious drawbacks, the most severe being that it is too powerful and imposes too little restrictions on prosodic systems. This objection has to be emphasized given our claim in Section 2 that the standard version of the Prosodic Hierarchy already overgenerates.Footnote 34 Ultimately, revisions along these lines again prune the predictive power of the theory.
The case study from Limbu suggests that there is no reason to assume that phonological processes should universally cluster on one prosodic word domain intermediate between the foot and the phonological phrase. It also demonstrates that deviations from the theoretical predictions of the Prosodic Hierarchy may lie well outside the range of phenomena which are accountable by the additional prosodization mechanisms proposed in the literature.
4. THE PROSODIC WORD AS AN EMERGENT CATEGORY
Up to this point, our discussion has been pessimistic towards the standard descriptive framework for the study of word-related prosodic patterns and the architecture of prosodic structure as enshrined in the Prosodic Hierarchy. In cross-linguistic comparison, this approach forces us, on the one hand, to make languages fit the theory by systematically excluding variation, for instance, in the form of non-general processes referencing morphological structure; on the other hand, the study of cross-linguistic variation forces us to revise the theoretical apparatus to such an extent that it loses its predictive power.
As a reaction to this interim conclusion, we would like to advocate a different view on prosodic domains. We will argue that constellations such as those evidenced in Vietnamese and Limbu can only be handled if we accept that prosodic domains are language-particular, intrinsic and highly specific properties of individual phonological rules or constraints. Our heuristics for the cross-linguistic study of prosodic domains allows us to inductively investigate how much prosodic systems can vary and what the limits of this variation are. Instead of positing absolute universals a priori, we seek statistical universals which help us understand how a probable prosodic system might look like (see Newmeyer Reference Newmeyer2005, Bickel Reference Bickel2007).
4.1 Decomposing the ‘prosodic word’
The evidence from our two cases suggests that the prosodic word should not be conceptualized as a holistic entity with a fixed set of properties. Even within one language, the prosodic word may be defined by different sets of processes or rules, cf. the minor and major prosodic word in Limbu with their distinct properties. The desire to distinguish ‘different kinds’ of the prosodic word has already been expressed by Hyman (Reference Hyman2008: 335–336) who proposes the following list:
(34) Different kinds of PW
(a) The demarcative word: a property marks the beginning or end of the word
(b) The culminative word: a feature occurs only once per word
(c) The harmonic word: a feature is realized throughout the word
(d) The metrical word: a word consists of hierarchically arrayed moras or syllables
(e) The minimal word: a word must consist of a minimum of moras or syllables
(f) The maximal word: a word can consist of a maximum of moras or syllables
(g) The phonotactic word: a word permits only certain output segments/sequences
(h) The morphophonotactic word: a word permits only certain input segments/sequences
Facing the fact that even such finer-grained word notions differ across languages, we would go one step further in positing only language-particular prosodic words. Consider the following prosodic domains for some word-related processes in Vietnamese and Limbu.
(35)
(a) The Vietnamese Tonal Dissimilation Word(BASE cứng–RED cựng)
Replace the high tone of the base with a low tone in the reduplicant.
(b) The Vietnamese Tonal Harmony Word(BASE–RED)
Be faithful to the tone series of the base in alliterative reduplication; note that ngã and hỏi change their series membership.
(c) The Vietnamese Sino-Vietnamese Word(STEM–STEM)
Stress the left constituent of the compound.
(d) The Limbu Velar Assimilation Word(PREFIX–STEM)
/n/ assimilates to the following velar place feature and surfaces as [ŋ]
(e) The Limbu Glottal Stop Insertion Word(PREFIX)(STEM–SUFFIX–ENCLITIC)
Insert an initial glottal stop to avoid a vowel-initial syllable.
(f) The Limbu Primary Stress Word(PREFIX–STEM–SUFFIX–ENCLITIC)
Stress the stem-initial syllable; if an iamb would arise in prefix–stem combinations, stress the initial syllable.
For each language-particular domain, we specify the morphological structure relevant for a given process and also the details of its application.Footnote 35 As a matter of fact, the domain specifications vary with respect to their scope. In (35a) specific instances of reduplication in Vietnamese are singled out, whereas (35b) refers to an entire subtype of reduplication. In some cases, word-related processes are only active in a certain area of the lexicon, as in (35c), an information also included in the domain specification. Note that the three domains for Vietnamese do not count as evidence for a domain in the Prosodic Hierarchy, since they do not apply generally within the lexicon. In contrast, we consider generality across the lexicon a matter of degree and leave the empirical test of whether this property has an effect on the organization of prosodic structure for future research. In any case, the Limbu domains (35d–f) differ from the Vietnamese ones in that their domain specifications generalize over morphological constructions, such as e.g. prefixation or the entire grammatical word.Footnote 36
Equipped with lists like (35), we are now in the position to explore empirically, e.g. using standard statistical tools of data-mining such as multidimensional scaling, what if any patterns there are in the distribution of prosodic domains across languages. In Bickel, Hildebrandt & Schiering Reference Schiering2009, we test whether there are probabilistic clusters of similarly-sized domains relative to the kind of sound pattern involved, e.g. tonal vs. segmental processes. In sharp contrast to the predictions of the Prosodic Hierarchy, we find a cross-linguistic trend for languages to multiply prosodic word domains. However, we also find statistical support for the following probabilistic universal:
(36) Stress-defined domains tend to be significantly larger than other domains.
In comparison to the huge body of absolute universals stated in the Prosodic Hierarchy, this universal seems to have rather small scope and to be rather local. However, it is empirically founded and, at least in our sample of 70 languages, stands the test for genealogical and areal bias.
4.2 The emergent prosodic word
The findings just reported suggest that like in many other areas of grammar, the absence of absolute universals does not mean that the distribution of structure is random. This is also evident from the observation that certain domain specifications are referenced more frequently by phonological processes than others. Consider the frequency of reference to some of the prosodic domains described for Limbu (Hildebrandt Reference Hildebrandt2007).
(37) Prosodic words in Limbu
The Limbu Primary Stress Word(PREFIX–STEM–SUFFIX–ENCLITIC)
The Limbu Labial Assimilation Word(PREFIX–STEM–SUFFIX–ENCLITIC)
The Limbu [l]~[r] Alternation Word(STEM–SUFFIX–ENCLITIC)
The Limbu Glottal Stop Insertion Word(STEM–SUFFIX–ENCLITIC)
The Limbu *ω(ŋ Word(STEM–SUFFIX–ENCLITIC)
The Limbu Velar Assimilation Word(STEM–SUFFIX)
The Limbu Velar Assimilation Word(PREFIX–STEM)
The Limbu Glottal Stop Insertion Word(PREFIX)
We note that some domains, such as the prefix domain, or the smaller prefix–stem and stem–suffix domains, are referenced only by one process each. In contrast, several processes converge on the prefix–stem–suffix–enclitic domain and even more cluster on the stem–suffix–enclitic domain. It is not a coincidence that these domains figure most prominently in our discussion of the Prosodic Hierarchy in Limbu. In this line of thought, one could attribute the concept of the prosodic word simply to the domain that is most frequently referenced by phonological patterns in a given language. The prosodic word can thus be redefined as a language-particular category which emerges through frequent reference within the prosodic system. In Vietnamese, on the other hand, no domain seems to be frequent enough to postulate a prosodic word domain in the traditional sense.
The concept of the emergent prosodic word again offers rich possibilities for cross-linguistic comparison. We observe that languages tend to show bipolar distributions in the frequency of reference to prosodic domains. From that point of view Limbu, in which two domains are more frequent than the other attested ones, mirrors a common property of prosodic systems. But we can also ask whether the evidence for the prosodic word is equally strong across languages. We already used the related measure of ‘word dominance’, i.e. the frequency of reference to the most frequently referenced domain, to test whether languages formerly classified as ‘stress-timed’ indeed show more evidence for prosodic words than languages of other rhythm classes as hypothesized by Auer Reference Auer1993 (Schiering Reference Schiering2009). It turns out that there is a trend for stress-timed languages in Indo-European and Sino-Tibetan to have more patterns reference their most frequent domains than syllable- or mora-timed languages. This finding, among others, suggests that prosodic word domains are stable in time, such that languages tend to retain the prosodic profile they inherited from the proto-language (see also Bickel et al. Reference Bickel, Hildebrandt, Schiering, Grijzenhout and Kabak2009). Presumably, frequent domains have a gravitating effect within the system, attracting phonological patterns which evolve in the course of sound change. However, at this point we can only speculate about the wider relevance and the psycholinguistic reality of the emergent prosodic word, but we consider the concept interesting enough to lead future research in the area of prosodic structure.
5. Conclusion
In this paper, we discussed the prosodic structure of two languages which challenge the theoretical assumptions of the Prosodic Hierarchy. Vietnamese is a language in which a prosodic word domain cannot be motivated on the basis of purely phonological rules. The phonology of the language thus exhibits less structure than would be predicted by Prosodic Phonology. For the architecture of prosodic structure this entails violations of Clustering and a consistent violation of Strict Succession. Limbu is a language in which multiple word domains can be motivated which overlap in non-trivial ways by referencing different portions of the grammatical word. The phonology of this language thus exhibits more structure than would be predicted by Prosodic Phonology. For the architecture of prosodic structure this entails violations of Clustering. The problems this evidence poses can neither be solved by the standard theory nor by more recent versions which allow more flexibility by weakening the Strict Layer Hypothesis or by providing more fine-grained inventories of prosodic domains or prosodic subcategorization.
The challenge from Vietnamese can be responded to by rejecting the idea that structures posited as universal must be descriptively necessary in any language. This can mean either of two things. (i) It can mean that Vietnamese speakers have prosodic words of the standard kind in their minds, but the words do not manifest themselves in any known phonological effect. (ii) We reinterpret Prosodic Hierarchy theory as constituting a universal inventory of choices rather than as an absolute universal; then, Vietnamese can simply be said to lack prosodic words.
The challenge from Limbu is stronger: whether one interprets the Prosodic Hierarchy as an absolute universal or as an inventory of choices, the theory leads to systematic violations of the Clustering Prediction, or, if one seeks to rescue these predictions, to self-contradictory analyses (‘the same prefix is and is not part of a prosodic word at the same time’). A radical alternative approach is typological in spirit: prosodic domains are conceived of as language-particular, intrinsic and highly specific properties of individual phonological rules or constraints, and all we need is tools for comparing domains across languages and for measuring their similarities and differences in precise ways. We sketched how we explored this in companion papers (Bickel et al. Reference Bickel, Hildebrandt, Schiering, Grijzenhout and Kabak2009, Schiering Reference Schiering2009). Note that under such a conception, prosodic domains are by no means superfluous or that the phonological properties of morphological constructions can be accounted for by a direct reference mechanism. The point is rather that a more reliable methodology should start from the individual processes at work in the language and should construct prosodic structure on that basis rather than starting with a limited number of domain types defined a priori. This leads us to conclude that the prosodic word is a language-particular category which emerges through frequent reference of phonological patterns to a given morphological construction type. We hope that this reassessment inspires future research on prosodic structures.
