Australian Linguistic Society
Valence, Voice and Argument Structure in Austronesian
Michael Dukes, Department of Linguistics, University of Canterbury, Christchurch, firstname.lastname@example.org
Relation-changing operations such as passive and dative shift have in the past been treated in terms of changes in the valence properties of predicates derived by such operations (e.g. Bresnan 1982, Pollard & Sag 1987, 1994). A number of researchers have recently proposed analyses in which passive is treated as an operation on argument structure. While the details of these analyses vary considerably, an important source of evidence for several of them is the alleged distinction between passive, putatively an argument structure operation (as found for example in English), and the voice/focus systems observed in Austronesian languages, such as Tagalog, Toba Batak and Balinese, which are claimed to involve operations on the valence features associated with predicates but not changes in their argument structure.
In this paper, the differences between 'true' passive and Austronesian voice/focus are evaluated, and some recent proposals for treating passive as an operation on argument structure within Head-driven Phrase Structure Grammar (HPSG) are critically assessed. It is shown that Austronesian languages display a considerable spectrum of behaviours with respect to voice, suggesting that there can be no straightforward split between two types of voice operations (i.e. one applying to argument structure and the other to valence features). It is argued that Austronesian voice/focus and 'true' passive generally display a similar range of characteristics which are argued to be typical of valence operations. A uniform valence-based approach to voice operations is defended (which entails that binding constraints may be determined with respect to both valence and argument structure information).
2. The Role of Argument Structure
In a number of recent publications (Sag & Fodor 1994, Sag & Godard 1994, Alsina 1996 amongst others) a considerable body of evidence has been built up illustrating the utility of a notion of 'argument structure' (ARG-ST) in accounting for a range of grammatical phenomena. One important thread of work in this area has involved the examination of the role of ARG-ST in the treatment of binding phenomena, much of which has sought to justify the distinctness of ARG-ST from valence features (SUBJ, OBJ, COMPS, etc) and to support the claim that binding constraints are uniquely associated with ARG-ST and not at all with valence. Key evidence for this approach has been adduced from languages of the western branch of Austronesian (Western Malayo-Polynesian), in particular from Toba Batak and Balinese (Manning & Sag 1995, Manning 1995, 1996, Wechsler & Arka 1998), which often display a dissociation between surface grammatical functions and reflexive binding possibilities. While subjects may not be coargument-bound reflexives in English (e.g. *Himself hit John), it is suggested that analogous structures are grammatical in Toba Batak and Balinese. The explanation for this contrast, it is claimed, lies in the fact that the voice morphology widely observed in WMP languages rearranges the assignment of semantic roles to grammatical functions but does not affect the structure of the ARG-ST list. Thus binding possibilities in Toba Batak and Balinese are unaffected by the choice of subject in a particular clause, while binding possibilities in English are altered by the effect of passive because passive actually rearranges the ARG-ST information associated with the predicate (thus explaining the acceptability of John was shot by himself).
It is not my goal to dispute the general usefulness of the notion of ARG-ST in this paper. Indeed, I believe that such a notion (or something like it) is a universal grammatical attribute in human languages. However, what I do intend to dispute here is the treatment of passive as an ARG-ST operation, which has been defended in the papers cited above. I hope to show that such a view of passive is problematic on both empirical and theoretical grounds. The empirical problems with respect to the Austronesian data are discussed in section three, while some general theoretical issues are discussed in section four.
3. Binding and Voice in Austronesian
The General Proposal.The treatment of voice in Austronesian has been controversial for decades (Schachter 1976). The authors cited above (Manning & Sag 1995, Manning 1995, 1996, Wechsler & Arka 1998) propose to distinguish Austronesian voice from familiar passive by treating the former as a valence operation (which can rearrange the linking of semantic roles to valence features but does not alter the arrangement of ARG-ST information) and the latter as an ARG-ST operation (which embeds the ARG-ST information of the verbal stem and creates a new matrix ARG-ST list in which the original active subject is demoted to oblique. Schematic representations of the proposed contrast between these two phenomena are given in (1) (Toba Batak) and (2) (English). Note that the CONT feature represents the semantic roles associated with each predicate.
(1) Toba Batak a. Active Voice b. Objective Voice Mang-ida si Ria si Torus Di-ida si Torus si Ria AV-see PM PM OV-see PM PM 'Torus saw Ria' 'Ria was seen by Torus' SUBJ  SUBJ  COMPS  COMPS  ARG-ST <,> ARG-ST <,> CONT [SEER,SEEN] CONT [SEER,SEEN] (2)English a. Active Voice b. Passive Voice Torus saw Ria Ria was seen (by Torus) SUBJ  SUBJ  COMPS  COMPS (PP) ARG-ST <,> ARG-ST <,> CONT [SEER,SEEN] CONT [SEER,SEEN] STEM [ARG-ST <,>] [CONT ]
The principle differences between the analyses of English passive and Toba Batak can be summarized as follows. In the English passive, the active subject is 'demoted' to optional oblique at ARG-ST, so that despite the change from an agentive subject to a patient subject, the first element on the matrix ARG-ST list is still represented as the subject. The grammatical information associated with the active predicate is embedded under the STEM feature in the passive but actually seems to play no further role in any grammatical phenomena in English (though it is appealed to in the analysis of binding phenomena in other languages (e.g. Russian, Manning & Sag 1995)). Note that the semantic information embedded in STEM is shared with that found in the matrix; thus both CONT features have the value denoted by the numeral .
In the case of the Toba Batak data, there is no embedding of grammatical information in either the active or objective voice. The ARG-ST list is the same for both predicates, but the value of SUBJ differs. The most agentive argument is subject in the active voice while the less agentive argument is subject in the objective voice. Note that while the agent is demoted to oblique in the (English) passive, the (Toba Batak) objective voice analysis treats the agent as an NP complement. This distinction is supported by the fact that the agentive NP in the objective voice is obligatory and is positionally fixed with respect to the verb (Schachter 1984). This distinction between complement and oblique is taken to be an important diagnostic for identifying whether a voice phenomenon in some language involves an ARG-ST operation or a valence operation.
A related diagnostic is that involving binding possibilities. In Toba Batak, the agent may bind a reflexive nonagent (but not vice versa) in both the active and the objective structures (Mangida dirina si Torus, Diida si Torus dirina, 'Torus saw himself'), while in English the subject may bind the nonsubject in both cases but not vice versa. These binding possibilities are accounted for solely by appeal to the order of elements on the ARG-ST.
Empirical Problems. The analysis of Toba Batak sketched above (as proposed by Manning & Sag 1995) is very plausible and corresponds closely with the general analysis of voice phenomena I wish to defend here. However, while Toba Batak itself provides an excellent exemplar for the claim that 'true' passive is formally distinct from WA objective voice, I believe that the data from a number of other WA languages is far more equivocal. I will restrict myself primarily to discussion of the data from Balinese, which has also been claimed to involve WA 'objective voice' (Manning & Sag 1995, Wechsler & Arka 1998)), and Malagasy.
While nonsubject agents are clearly complements in Toba Batak, they are plausibly treated as obliques in both Balinese (contra Artawa & Blake 1997, Wechsler & Arka 1998) and Malagasy. Wechsler & Arka claim that such agents are complements in Balinese because they are obligatory NPs. Furthermore, they form some sort of tight constituent with the verb so that nothing may intervene between verb and agent. However, the postverbal agent in Balinese displays a number of properties that are not consistent with complementhood, as noted in Clynes 1995. Firstly, postverbal agents, unlike postverbal patients, can only be rather 'short' in size; that is, they are either free pronouns, bare indefinite nouns or a third person clitic pronoun (-(n)a). Clynes (p.c.) suggests that such nominals are not even NPs but are in fact incorporated nouns and thus not true complements (though Wechsler & Arka 1998 do provide one example ((34)b.) of a true phrasal indefinite NP that is problematic for Clynes' claim). Strikingly similar data in Indonesian have also been analysed in terms of agent incorporation (Sukarno 1996). No such restrictions are observed with non-agent complements of active verbs, illustrating the fact that unlike Toba Batak, Balinese displays considerable asymmetry between the nonsubject agent and nonsubject patient.
Secondly, when the agentive argument of a objective voice verb is third person and marked by the clitic pronoun -(n)a, it can be doubled by a full (definite) phrasal argument. But this full phrasal argument is actually a PP headed by the preposition teken 'with / by' (or ajak in Singaraja Balinese) and cannot be an NP (Clynes 1995). The fact that this argument is a PP suggests that it is an oblique rather than a term, since Wechsler & Arka claim that terms are always NPs in Balinese. Even more damaging for the claim of complementhood is the fact that agents demoted by a 'true' passive rule in Balinese mark their agents with the same preposition. To avoid this impasse, Wechsler & Arka take the position that sentences containing teken are always passive. But this in turn forces them to claim that any sentence containing the third person clitic -(n)a is actually ambiguous between 'true' passive and objective voice and is only disambiguated by the addition of a teken PP (which, in their account, forces the passive analysis). Such a distinction is subtle and supports the contention defended here that the difference between 'passive' and 'objective voice' structures is graded and rather language-specific. Even if one accepts the ambiguity claim and the associated hypothesis that the objective voice clitic is being reanalyzed as a passive morpheme (Wechsler & Arka 1998, section 5.3.1), the fact that one and the same morpheme can synchronically indicate passive or objective voice suggests that there is unlikely to be any categorical distinction between the two.
The data involving nonactive agents in Malagasy passives is broadly similar to that observed in Balinese, with at least one important exception. Nonactive agents in Malagasy, even more so than in Balinese, form a very tight constituent with the verb and are marked as genitives in exactly the same way that genitive possessors combine with a head noun (Keenan 1991). However, unlike agents in Balinese and Toba Batak, those in Malagasy are not obligatory. It is certainly the case that they appear more often than not (Manorohanta & Keenan 1998 claim that around 60% of nonactive clauses in Malagasy contain an agent phrase), but they must generally be treated as optional. Are these nonactive agents demoted by a true ARG-ST passive or is there a valence-affecting process of objective voice in Malagasy? Their optionality suggests they are obliques, as do their morphological parallels with possessors, which are also clearly obliques. However, markedness and binding data suggest that they are somewhat more like complements (according to the criteria outlined above). As in Balinese, morphologically unmarked predicates are 'patient-prominent' (i.e. take a patient subject) and may take a genitive agent even though there is no overt process of passivization applying to them. Active predicates by contrast are morphologically complex (see (3)).
(3) Resy 'be defeated' a. Resi-ko izy defeated-1.Sg.Gen 3.Sg.Nom 'He was defeated by me' b. N-an-dresy azy aho Past-Act-defeated 3.Sg.Acc 1.Sg.Nom 'I defeated him'
Thus there is no sense in which the patient is 'promoted' to subject by any morphological marker. Furthermore, since active structures are morphologically derived from 'passive' ones, it would be necessary to assume under an embedding approach, that the active predicate actually embeds the underlying passive ARG-ST. The empirical consequences of this are unclear and the approach seems somewhat counterintuitive.
Binding facts show that the genitive agent can naturally bind an accusative reflexive (as can a nominative agent in an active clause). Furthermore, a genitive agent can also bind a nominative reflexive, though with some degradation, apparently due to the fact that the nominative reflexive must be grammatically definite, as in (4) (examples from Keenan 1991).
(4) a. Amono-an-'ny ray aman-dreny rehetra tena ny zanaky kill-circ-the parents all self the children 'All parents kill themselves for their children' b. ?No-vono-in-dRabe ny tena-ny Past-kill-pass-Rabe the self-his 'Rabe killed himself' (Lit: 'Rabe killed his body')
The mixed properties of the Malagasy voice phenomena again suggest that there is no easy split between 'WA voice' and 'true' passive. Similar mixed data can be cited from other WA languages (see e.g. the discussion of Chamorro below, where 'WA voice' appears even more passive-like). In short, for every WA language that behaves like Toba Batak, it is possible to present data from another whose properties with respect to voice appear considerably closer to a standard passive type process, suggesting that these phenomena range across a continuum rather than falling into two discrete categories.
4. Theoretical Problems
Alongside the empirical problems sketched above, there are a number of theoretical issues that are raised by the 'passive as embedding' analysis. Space restrictions preclude much discussion of these here. Some of these problems arise directly in the analysis of Chamorro grammatical relations and are discussed in more detail in Dukes 1998. In particular, it is odd that while passive is claimed to involve embedding (like causativization and restructuring), only passive requires that the CONTENT value of the embedded predicate is identical to the CONTENT value of the matrix. This is an unexplained coincidence under the embedding analysis but falls out as a straightforward consequence of the view of passive advocated here, in which it is treated in a (more or less) traditional fashion as a process that simply reorders elements on the valence lists, while preserving ARG-ST information (i.e. 'underlying grammatical relations'). The fact that passive and dative shift are invariably monoclausal operations (unlike causativization) also calls into question the embedding analysis of relation-changing operations sketched earlier, as does morphosyntactic evidence from Chamorro wh-extraction, which clearly supports the claim that objects derived by dative shift are not underlying objects in the same way that objects derived by causativization are (i.e. via creation of a new matrix ARG-ST list, see Dukes 1998 for details). In short, the embedding analysis of relation-changing operations seems to create considerably more structure than can actually be motivated in accounting for the data.