Note the following conventions for this ASCII version: logical disjunction: v list of synsems: L concatenation: + domain union: U string of phonemes P BALINESE ARGUMENT STRUCTURE AND VALENCE (draft version: comments welcome) Stephen Wechsler University of Texas, Austin wechsler@mail.utexas.edu I Wayan Arka University of Sydney 0. Introduction This paper is an HPSG treatment of argument structure (ARG-S) and valence in Balinese, an Austronesian language with about 3,000,000 speakers primarily on the Indonesian islands of Bali, western Lombok, and eastern Java. Balinese is a 'syntactically ergative' language in the sense that the unmarked form of a monotransitive verb has a Theme subject and Agent complement. We will present evidence from Balinese linking and binding, for a version of ARG-S where canonically terms precede obliques and items in each group are ordered by thematic prominence (Manning 1994), and relation-changing rules like passive or applicative systematically alter this order. We also argue that the account of Balinese applicatives can be generalized to a general theory of applicatives covering syntactically ergative as well as accusative languages. 1. 'Syntactic ergativity': In Balinese, the morphologically unmarked form of a monotransitive verb has a Theme subject and Agent complement. This form of the verb (the morphological zero) will be called the Objective Voice (OV): (1) Bawi-ne punika tumbas tiang (high register) pig-DEF that 0V.buy 1SG 'I bought the pig.' In section 2 we review evidence that Balinese has SVO order, so that _bawi-ne punika_ 'the pig' in (1) is the subject, and tiang 'I' is the object. When the verb appears in Agentive Voice (AV), marked with an initial homorganic nasal, the agent appears in subject position and the theme in object position: (2) Tiang numbas bawi-ne punika 1sg AV-buy pig-DEF that 'I bought the pig.' Languages with the Theme-Subject, Agent-Object linking illustrated in (1) are sometimes referred to as 'syntactically ergative', while languages with voice alternations dictating subject choice, as illustrated by the contrast between (1) and (2), are sometimes called 'mixed pivot languages.' Since OV is the normal form of the verb and is also morphologically unmarked, we use the term 'syntactically ergative' for Balinese, though nothing hinges on this terminology. 2. Subjecthood. Balinese has SVO order. The phenomena of raising, relativization, control, and quantifier float show that the position preceding the verb, where the sole argument of an intransitive appears, is the subject position (Artawa 1994). This hypothesis can be compared with the alternative view that the pre-verb position is the Topic position, and that the AV morpheme selects the Agent as Topic. In this section we review Artawa's (1994) arguments for subjecthood rather than topichood of this argument. Note first of all that this pre-V position is where the sole argument of an intransitive verb appears: (3) Ia pules. 3sg sleep '(s)he is sleeping.' This is prima facie evidence for taking this as the subject position. On the topic hypothesis, this would mean that every sentence has a topic. Raising, in which a semantic argument of a subordinate clause appears in a superordinate clause, is known to be restricted to subjects cross-linguistically. On the other hand there do not seem to be any languages where raising is restricted to Topics only. Balinese has certain raising adjectives which allow raising only of the purported subject argument. First consider an intransitive verb example: (4) a. Keweh ia pules. difficult 3sg sleep b. Ia keweh pules. 3sg difficult sleep 'It is difficult for him/her to sleep.' The subject _ia_ '(s)he' can be raised to the position to the left of the matrix predicate _keweh_ 'difficult.' Now consider transitive verbs, first an OV verb (5) and then an AV verb (6): (5) a. Aluh montor-e benahin tiang. easy car-DEF OV.fix 1sg b. Montor-e aluh benahin tiang. car-DEF easy OV.fix 1sg 'The car is easy for me to fix.' c. *Tiang aluh montor-e benahin. 1sg easy car-DEF fix (6) a. Enggal tiang ngadep umah. quick 1sg AV.sell house b. Tiang enggal ngadep umah. 1sg quick AV.sell house 'It is quick for me to sell a house.' c. *umah tiang enggal ngadep. house 1sg quick AV.sell It either voice only the purported subject can be raised: in a clause with OV it is the theme but not the Agent (5), while in AV it is the Agent but not the Theme (6). Quantifier float is also known to be sensitive to grammatical relations, a common restriction being that quantifiers can float from subjects but not complements. The subject 'the children' launches the quantifier _onya_ 'all' in the following example with an OV verb. (7) a. onya cerik-cerk-e tepukin beli. all child:PL-DEF OV.see elder.brother 'I saw all the children.' b. cerik-cerk-e onya tepukin beli. c. cerik-cerk-e tepukin beli onya. With an AV verb it is still the subject which can launch the quantifier: (8) a. onya cerik-cerk-e meli jaja. all child:PL-DEF AV.buy cake 'All the children bought cakes.' b. cerik-cerk-e onya meli jaja. c. cerik-cerk-e meli jaja onya. Note that (8c) cannot have the interpretation where the quantifier is construed with the object, that is, it cannot mean 'The children bought all the cakes'. Control is another classic subject test: in many languages only a subject can be controlled ('deleted under equi'). Once again, only the pre-verbal argument, whether the Theme of an OV verb or the Agent of an AV verb, can be a controllee: (9) a. tiang edot [PRO teka] 1sg want come 'I want to come.' b. tiang edot [PRO meriksa dokter] 1sg want AV.examine doctor 'I want to examine a doctor.' c. tiang edot [PRO periksa dokter] 1sg want OV.examine doctor 'I want to be examined by a doctor.' Finally, relativization also picks out the purported subjects in Balinese. According to Comrie's (1981) accessibility hierarchy, the subject is the most likely item to be subject to relativization (that is, if a language allows any argument to relativize, it allows subjects). In Balinese only the purported subjects can be relativized: (10) a. anak-e cenik ento [ane gugut cicing] person-DEF small that [REL OV.bite dog] 'the child whom the dog bit' b. *cicing [ane anak-e cenik ento gugut] dog [REL person-DEF small that OV.bite] 'the dog that bit the child' (11) a. I Warta [ane maca koran] Art Warta [REL AV.read newspaper] 'Warta who read the newspaper' b. *koran [ane I Warta maca] newspaper [REL Art Warta AV.read] 'the newspaper that Warta read' To summarize, raising, quantifier float, control, and relativization all support the claim that the NP appearing to the left of the verb is the subject, regardless of the verb's voice marking. 3. The complement Agent of an OV verb. While the Agent of an AV verb must appear as the subject-- (12) Ia ngedengin I Wayan potlot-e ento 3SG AV-show I Wayan pencil-DEF that '(s)he showed Wayan the pencil.' --the subject of a mono- or ditransitive OV verb is any term (NP) argument EXCEPT the Agent (13). The Agent of the OV verb instead appears as a complement in one of four forms. First, it can take the form of a 3sg clitic (12) on the verb (low register only [$$ fn. about register]): (13) a. Potlote ento edengin=a I Wayan. pencil-DEF that OV.show=3sg Art Wayan b. I Wayan edengin=a potlote ento. I Wayan OV.show=3sg pencil-DEF that '(s)he showed Wayan the pencil.' This clitic appears only in third person singular form. This clitic can be doubled by a PP headed by the preposition _teken_ (14): (14) Buku-ne jemak=a teken Wayan book-DEF OV.take=3sg PREP Wayan 'Wayan took the book.' Alternatively, the Agent can be realized as an indefinite NP (15). (15) Wayan gugut cicing /*cicinge ento Wayan OV.bite dog /dog that 'A/*the dog bit Wayan.' Finally, the Agent can be expressed as a post-V free pronoun (see (1) above). 4. Anaphoric binding. Binding relations in Balinese depend on the status of the arguments and not on grammatical relations, a situation familiar from many Austronesian languages. An Agent can bind a Theme, but not vice versa, regardless of how these roles map onto the grammatical relations. Hence with an OV verb, the complement Agent can bind into a subject: (16) Ragan idane cingakin ida. (high register) self OV.see him/her '(S)he saw himself/herself.' But the opposite binding relation, where the Theme subject binds an Agent object, is not possible: (17) a. *Wayan tepuk=a teken awakne Wayan OV.see=3sg PREP self 'Wayan was seen by himself.' b. Awakne tingalin=a self OV.see=3sg '(S)he saw herself/himself.' When we turn to AV, the subject Agent binds a complement Theme: (18) Ia ningalin awakne 3sg AV.see self '(S)he saw herself/himself.' This suggests that binding is defined at argument structure rather than in valence features. As Manning 1994 points out, the behavior of binding in syntactically ergative languages constitutes prima facie evidence for the HPSG(3) binding theory, where binding is defined on ARG-S rather than phrase structure or grammatical relation. (19) HPSG Binding Theory Principle A. A locally a-commanded anaphor must be locally a-bound. Principle B. A personal pronoun must be locally a-free. Principle C. A non-pronoun must be a-free. definitions: a-command = the ordering relation on ARG-S x locally a-binds y = x a-commands y, x and y coindexed x a-binds y = x a-commands some z dominating y, x and y coindexed Ergative and accusative clauses are roughly alike with respect to binding conditions, but differ in the mapping between ARG-S and the VALENCE lists. For example, these are the approximate lexical signs for _tingalin_ 'see (OV)' and _ningalin_ 'see (AV)' as in (17b) and (18) respectively (see Manning and Sag 1995): |PHON | |SUBJ <[1]NP> | |COMPS <[2]NP> | |ARG-S <[1]NP-i, [2]NP-j> | |CONTENT |seeing-rel | | | |SEER i | | | |SEEN j | | |PHON | |SUBJ <[2]NP> | |COMPS <[1]NP> | |ARG-S <[1]NP-i, [2]NP-j> | |CONTENT |seeing-rel | | | |SEER i | | | |SEEN j | | The next section presents an account of the ARG-S to VALENCE list mapping. 5. Linking and Syntactic Ergativity This formalization follows Manning and Sag 1995 in spirit, but modifies it slightly to account for Balinese. Verb stems are sorted into intransitive and transitive: v-stem-> intrans-stem v trans-stem intrans-stem: [ARG-S & list(non-core-np)] (exactly 1 NP) trans-stem: [ARG-S ] (more than 1 NP) An intransitive stem has exactly one core NP, where a 'core NP' is roughly a direct argument NP. A transtive stem has more than one. Note that transitive stems subsume both mono- and di-transitives. For all verbs except verbs which have undergone pro-drop or express the agent as a clitic, the value of ARG-S is the domain union (indicated by 'U') of the values of SUBJ and COMPS (xUy preserves ordering on x and on y but freely intermingles them). The ARG-S is inherited from the verb stem: canon-verb: |SUBJ [1] | |COMPS [2] | |ARG-S [3][[1]U[2]] | |STEM |v-stem | | | |ARG-S [3] | | Verbs are sorted into agentive voice (syntactically accusative) and objective voice (syntactically ergative): canon-verb -> AV-verb v OV-verb In a syntactically accusative verb stem (as in English, or Balinese AV verbs), the value of ARG-S is canonically the append of the values of SUBJ and COMPS, in that order. Given the above specification for canonical verbs, we need only specify that the first argument is selected as the subject; all remaining arguments will map to the COMPS list, preserving order. AV-verb: |PHON n(P) | |SUBJ [1] | |ARG-S [1]+L | |STEM |PHON P | | where L is a list, and n(P) is obtained by replacing the initial consonant in P with a homorganic nasal, if P is consonant-initial; and n(P)=ng+P if P is vowel-initial. Linking for a syntactically ergative verb stem (as in Balinese OV) depends on transitivity: if intransitive, it is like AV-verb; if transitive, then any NP EXCEPT the a-subject maps to SUBJ (see (13) above). Hence the sort declaration is disjunctive: OV-verb: |PHON P | |SUBJ [1] | |ARG-S [1]+L | | v |STEM |intrans-stem | | | |PHON P | | |PHON P | |SUBJ [1] | |ARG-S + [L U [1]] | |STEM |trans-stem | | | |PHON P | | For completeness, we shall account for verbs with the agent clitic (see (13), section 3 above). Such a verb is non-canonical in the sense that its ARG-S list does not contain exactly the same elements as the domain union of the SUBJ and COMPS lists, but rather contains one more item. We know that this item remains on the ARG-S list since it can be a binder. |PHON P+a | |SUBJ [1] | |COMPS L (+) | |ARG-S [2][+[L U [1]]] | |STEM |trans-stem | | | |PHON P | | | |ARG-S [2] | | The a-subject (NP:ppro_i) acts as a pronominal for the purpose of binding theory. It can be coindexed with an optional teken-PP (see (14) above). The rest of ARG-S is a list which contains [1], which maps to the singleton Subject list, while everything else in the list (L) maps to COMPS. 6. Theory of Argument Structure Following Manning 1994, we assume that ARG-S elements are ordered as follows: (20) (i) terms precede obliques. (ii) arguments within each group are ordered by the lexico-semantics of the verb For the present purposes 'terms' means NPs (in a case-marking language, NPs in oblique case might be classified as obliques). Regarding (ii), one might assume a 'thematic hierarchy' like ag < go < th < loc, or adopt an alternative such as Wechsler 1995 or Davis 1996. The particular system is not crucial and will not be discussed further here. Evidence for (20) comes from anaphoric binding in verb alternations like the following: (21) a. Mary showed Bill_i himself_i in the mirror. [ARG-S ] b. Mary showed Bill_i to himself_i in the mirror. [ARG-S ] In (a) the Goal a-commands the Theme while in (b) the Theme, being a term, a-commands the oblique Goal. The same pattern obtains in Balinese: (22) a Icang ngomong teken Made unduk awakne 1sg AV.talk to Made about self 'I talked to Made_i about himself_i b. *Icang ngomong unduk Made teken awakne 1sg AV.talk about Made to self 'I talked about Made_i to himself_i' AV.talk: [ARG-S ] (23) Icang ngedengang Wayan sig awakne 1sg AV.show Wayan to self 'I showed Wayan_i to himself_i.' We will see in the next section that applicativization, being an augmentation of ARG-S, also crucially affects its order. 7. Applicativization. Balinese has two applicative suffixes which signal that a term argument has been added to the argument structure. Each suffix is restricted to a range of semantic types for the added role: (24) -in: Locative or Source roles -ang: Instrumental, Benefactive, or Recipient roles Following is an example of the -in suffix. Sentence (a) uses the non-applied form, so the locative argument appears in a PP. Sentence (b) is the applied form, in objective voice, while (c) illustrates the applied form in Agentive Voice: (25) a. Cai teka ka kantor-e ene. 2sg come to office-DEF this 'You came to this office.' b. Kantor-e ene teka-in cai. office-DEF this OV.come-APPL 2sg c. Cai neka-in kantor-e ene. 2sg AV.come-APPL office-DEF this Note that the applied locative role is the subject in (b) and the object in (c). We analyze applicativization as an operation which adds a term argument in the second place in ARG-S, that is, immediately following the a-subject. base verb: [ARG-S < NP>+L] applied verb: [ARG-S < NP, NPa>+L] (where NPa is the applied argument) This is accomplished by means of a sort declaration for the applied verb. For example, the locative applicative would have roughly the following declaration: (26) |trans-stem | |PHON P+in | |ARG-S <[1], NP_i>+L | |CONTENT {[2], at'(j,i) v at'(k,i)} | |STEM |PHON P | | | |ARG-S <[1]>+L | | | |CONTENT [2]{|EVENT j |} | | | |(THEME k) | | For concreteness we take the CONTENT to have a set value, to accomodate the meaning added by the applied argument; other formulations are possible. The locative relation is designated by at'. The locative applicative can be interpreted as locating a particular argument (as in (25b,c)) or as locating the event as a whole (as in (26) below), so we disjunctively specify the second psoa in the list for the applicative verb's CONTENT (at'(j,i) v at'(k,i)). (We assume that CONTENT psoas can contain a Davidsonian event variable.) Notice that the 'output' of applicativization-- i.e. the top-level sort for (26)-- is a (transitive) stem, and not a verb. Applicativization is an operation on ARG-S and does not itself determine the mapping to Valence lists. Instead it creates the 'input' to the mapping as determined by the AV-verb and OV-verb sorts given above. Hence the applied NP becomes the subject of a monotransitive OV verb (as in (25b) but the object of an AV verb (as in (25c)). Turning to derived ditransitives, that is, monotransitive roots to which a third term argument has been added, our analysis makes two predictions: Prediction I: Applied OV ditransitives should be 'symmetrical' in the sense that either non-a-subject term can be the subject. This follows immediately since OV specifies that the SUBJ is any term in ARG-S except the a-subject. The correctness of this prediction is illustrated here: (27) a Toko-ne ento anggeh-in=a baas shop-DEF that buy.on.credit-APPL=3sg rice b. Baas anggeh-in=a tokone ento rice charge-APPL=3sg shop-DEF that '(S)he bought rice on credit in that shop.' Prediction II. Applied OV ditransitives are 'asymmetrical' with respect to binding: the applied NP can bind the other non-a-subject, but not vice versa. This also follows straightforwardly from the fact that the applied argument is the leftmost item, excepting the a-subject. Hence it a-commands all non-a-subjects. This is illustrated by the examples of binding by quantified noun phrases (QNPs): (28) a. *sabilang potrekan edeng-in=a jleman-ne every picture OV.show-APPL=3 person-POSS '(S)he showed every picture_i (to) its_i owner" b. Sabilang jlema OV.edengina potrekanne every person show-APPL=3 picture-POSS '(S)he showed every personi his/her_i picture' The agent is expressed by the 3rd person clitic. The goal is the applied role, so it a-commands the theme. Hence a QNP goal can bind a theme (as in (b)) but not vice versa, as shown in (a). It is generally assumed that goals outrank themes, so if both are terms then the above facts might seem to follow independently of our assumption that applicativization adds the second role. From that point of view, stronger evidence comes from the behavior of locatives, which are generally assumed to fall below Theme on the thematic hierarchy. An applied locative can bind into a theme, but not vice versa: (29) a Sabilang toko anggeh-in=a baas-ne ane paling luunga every shop OV.charge-APPL=3sg rice-POSS best 'In every shop_i (s)he bought its_i best rice on credit.' b. * Sabilang sate jang-in-a bodang-ne every satay OV.put-APPL basket-POSS '(S)he put every satay_i in its_i basket' Being an applied role, the locative falls into the second position in ARG-S, above the theme. Interestingly, applicativization can lead to an ARG-S ordering which departs slightly from the basic 'root' ordering given in (20), assuming that Themes are ordered above Locatives on the thematic hierarchy or whatever other means is used to determine basic ARG-S ordering. 8. Passive. Balinese has a passive construction, probably borrowed from Javanese, which occurs mainly in the high register. The verb bears the passive ka- prefix, and the a-subject appears as an optional PP headed by teken ('by-phrase'): (30) Padi-ne ka-ambil (teken Made). rice-DEF PASS-take (PREP Made) 'The rice has been taken (by Made).' The passive cannot cooccur with AV marking. We analyze the Balinese passive as designating that the a-subject is a teken-PP. pass-stem: |PHON ka+P | |ARG-S + L | |STEM |PHON P | | | |ARG-S +L | | Since AV marking maps the a-subject to the subject, and subjects must be NPs, we explain why the passive is incompatible with Agentive voice. Two further predictions follow: Prediction I: Either non-a-subject term argument of a passive ditransitive can be the subject, since the Subject of an OV-verb is any term except the a-subject. Sentence (a) is an AV ditransitive; (b) and (c) are passives in which the recipient and the theme, respectively, surface as the subject. (31) a Ida sampun numbas-ang okan idane motor anyar 3 already AV.buy-APPL child 3-DEF motor-bike new '(s)he bought his child a new motor bike' b. Okan idane sampun ka-tumbas-ang motor anyar (antuk ida) child 3-DEF already PASS-buy-APPL motor-bike (by 3) c. Motor anyar sampun ka-tumbas-ang okan idane (antuk ida) Prediction II: Since the passive teken-PP is the highest element in the ARG-S list, it can bind into the subject: (32) Ragan idane tan ka-runguang antuk ida. self 3 NEG PASS-care by 3 '(s)he does not take care of him/herself.' Recall from (14) above that a teken-PP can also double the clitic in a construction with a (non-passive) OV verb. This OV teken-PP cannot bind into the subject: (33) ?*Awakne tusing runguang=a teken ia. self NEG OV.care=3sg by 3sg '(s)he does not take care of himself/herself' This contrast follows because OV and AV affect the mapping from ARG-S to Valence, but do not alter the canonical ordering of ARG-S described in (20) above. According to that ordering, terms a-command obliques, so the teken-PP falls at the end of the ARG-S list. But passive leads to a violation of that basic, canonical ordering, since in a passive ARG-S the teken-PP remains at the top of the list, where it a-commands terms. 9. Conclusion. This study of Balinese argument structure and valence supports these conclusions: (i) Binding is determined at ARG-S. (ii) Ergativity involves the mapping between ARG-S and the valence lists. (iii) Applicativization operates on the ARG-S list by adding an argument just following the a-subject. References Artawa, Ketut 1994. Ergativity and Balinese Syntax. PhD thesis, La Trobe University, Bundoora, Australia. Davis, Tony 1996. Linking and the Hierarchical Lexicon". PhD thesis, Stanford. Manning, Chris 1994. Ergativity: Argument Structure and Grammatical Relations. PhD thesis, Stanford. Manning, Chris and Ivan Sag 1995. 'Dissociations Between Argument Structure and Grammatical Relations.' Draft, http://www.lcl.cmu.edu/manning/papers/tuebingen.ps. Wechsler, Stephen 1995. The Semantic Basis of Argument Structure. CSLI Publications.