Advanced Topics in HPSG

Transcription

1 Advanced Topics in HPSG Andreas Kathol Adam Przepiórkowski Jesse Tseng 1 Introduction This chapter presents a survey of some of the major topics that have received attention from an HPSG perspective since the publication of Pollard and Sag (1994). In terms of empirical coverage (of English and other languages) and analytical and formal depth, the analyses summarized here go well beyond the original theory as defined in Pollard and Sag (1987) and (1994), although these naturally remain an indispensable point of reference. 1 We will have to make a biased choice among the possible topics to cover here, and the presentation will of course be colored by our own point of view, but we hope that this chapter will give the reader a reasonable idea of current research efforts in HPSG, and directions for further exploration of the literature. In keeping with HPSG s emphasis on rich lexical descriptions, the first section ( 2) concentrates on the licensing of dependents by lexical heads. We begin with a discussion of the conceptual separation between argument structure and valence in current HPSG work. We examine how the the traditional distinction between arguments and adjuncts fits into this model, and then we turn to the highly influential idea of argument composition as a mechanism for dynamically determining argument structure. In 3, we concentrate on issues of linear order, beginning with lexicalist equivalents of configurational analyses and then considering more radical departures from the notion of phrase structure. The topics covered in 4 all have to do with syntactic abstractness. On the one hand, most work in HPSG avoids the use of empty categories in syntactic structure, preferring concrete, surface-based analyses. On the other hand, there is a current trend towards construction-based approaches, in which analyses are no longer driven only by detailed lexical information, but rely crucially on the definition of phrasal types, or constructions. One of the distinctive design features of HPSG is its integrated view of grammar. Information about syntax, semantics, morphology/phonology, and (potentially) all other components of the grammar represented in a single structure, with the possibility of complex interactions. In 5 we discuss a number of recent developments in the analysis of the syntax-semantics-pragmatics interface, in particular the treament of scope and illocutionary force, as well as information structure and the representation of speakers beliefs and intentions. The discussion of grammatical We would like to thank Bob Borsley, Miriam Butt, Ivan Sag, and especially Georgia Green for extensive comments on an earlier draft of this article. All remaining errors are ours. 1

2 interfaces continues in in 6, devoted to interactions between syntax and morphology. We conclude the chapter with a summary of recent developments in the formal logical foundations of HPSG ( 7). 2 Argument Structure One of the most significant conceptual changes distinguishing HPSG from Generalized Phrase Structure Grammar is the treatment of combinatorial properties. In GPSG, lexical items carry a numerical index that identifies the subcategorization frame in which they can occur, and there is a distinct immediate dominance rule for each subcategorization type, resulting in a large number of such rules for head-complement structures. In contrast, lexical descriptions in HPSG include a detailed characterization of their combinatorial potential encoded in a valence feature, and thus a much smaller set of highly general immediate dominance schemata is sufficient. In this way, HPSG has an affinity with Categorial Grammar, where the categories themselves are complex and encode combinatorial properties, allowing the assumption of a small number of general combination mechanisms. A number of linguistic problems have since been explored in HPSG and solutions have been developed that have significantly refined the original ideas and provided new insights into the nature of valence. 2.1 Valence and Argument Structure One significant development since the original presentation of the theory is the separation of the notions of valence and argument structure. In HPSG1 and HPSG2, valence was encoded in a single attribute, SUBCAT, containing a list of all syntactically selected dependents. Borsley (1987) pointed out, however, that this approach did not allow syntactic functions to be reliably distinguished. For example, the subject was originally defined as the single remaining element on SUBCAT, but this incorrectly identifies some prepositional complements and nominal specifiers as subjects. Borsley s proposals for treating syntactic functions as primitive notions, and splitting the SUBCAT list into three valence lists, SUBJ(ECT), SPECIFIER (SPR), and COMP(LEMENT)S, were adopted in HPSG3, and since then most authors assume these three lists as part of a complex VALENCE attribute. 2 The technical consequence of this move is that the head-complement, head-subject, and headspecifier schemata refer to the appropriate valence lists, rather than particular configurations of SUBCAT, and the SUBCAT Principle is replaced by the correspondingly more complex Valence Principle. An alternative default formulation of this principle is proposed by Sag (1997), 3 later incorporated into the default Generalized Head Feature Principle (Ginzburg and Sag, 2000). This approach offers a more economical notational representation (at the price of additional formal machinery for allowing default unification), but it can be argued that the essential content of the original Valence Principle that synsem objects are removed from the valence lists when they are syntactically realized is then encoded in a piecemeal fashion in the definitions of the individual ID schemata. 2

3 The decision to split syntactic valence into three lists makes it possible to express mismatches between the syntactic function of a constituent and the way that it is realized in the syntactic structure. This possibility has been exploited mainly in analyses where the synsem of the grammatical subject is encoded in the COMPS list. As a result, the subject is realized not by the head-subject schema, but by the head-complement schema. This has been proposed for verb-initial languages like Welsh (Borsley, 1989), and for finite clauses in German, where the subject appears in the Mittelfeld, just like the complements and adjuncts of the verb (Kiss, 1995). Another example of the same valence/function mismatch is the analysis of subject-auxiliary inversion in Sag and Wasow (1999), where a lexical rule empties the auxiliary s subject valence, which has the result of forcing the valence object corresponding to subject to appear as the first element of the COMPS list instead. This ensures that the subject will not be realized preverbally, but as the first complement following the auxiliary verb, which is the desired structure. It should be said that many analyses of this type are motivated primarily by word order considerations, and so a possible alternative approach would be to use surface linearization constraints, without actually modifying the basic syntactic structure via valence manipulation. After replacing SUBCAT by SUBJ, SPR, and COMPS, researchers soon realized that for the treatment of some phenomena (most notably Binding Theory), they still needed a single list encoding all of the arguments of a head. So the SUBCAT list was revived in the form of the ARG(UMENT)-ST(RUCTURE) list, with one crucial difference: while SUBCAT as a valence feature recorded the level of syntactic saturation for each higher phrase in the tree, ARG-ST was introduced as a static representation of the dependents of the lexical head. In its original conception, this information is only found in the representation of the lexical head (an object of type word). But a variety of recent work (for instance Przepiórkowski 2001) has argued that certain phenomena require that ARG-ST information also be visible on phrasal constituents projected from the head. In simple cases, the ARG-ST list is identified with the concatenation of SUBJ, SPR, and COMPS at the lexical level, i.e., before any valence requirements have been saturated. However, the lists in question do not always line up in this fashion and the possibility of mismatches gives rise to a number of analyses of otherwise puzzling phenomena. We will briefly discuss two of these here, pro-drop and argument realignments in Austronesian languages. The standard transformational approach to missing subjects in finite environments has been to posit a null pronoun (pro) that instantiates the syntactic subject position. In keeping with HPSG s general avoidance of unpronounced syntactic material, we can instead analyze the unexpressed subject as an ARG-ST element that does not have a corresponding valence expression. The following example from Italian (1a) and the corresponding lexical description of the verb mangia illustrate this idea: (1) a. Mangia un gelato. eat.3sg a icecream S/he is eating an icecream. b. ARG-ST NP3sg, NP SUBJ COMPS NP 3

4 Dependencies in which the subject participates, such as binding or agreement, can be accommodated straightforwardly if they are described as referring to the least oblique ARG-ST element, rather than the value of SUBJ. A more radical mismatch between valence and argument structure has been proposed by Manning and Sag (1998) and Manning and Sag (1999) for the realization of arguments in Western Austronesian languages such as Toba Batak. In this language clause-initial verbs form a VP with the immediately following argument NP. In the case of active voice (AV) morphology, this NP has the status of non-subject, as evidenced by the fact that a reflexive in that position has to be bound by a later ( higher ) NP. The example in (2) can be analyzed exactly like the corresponding English sentence (apart from the position of the subject NP). In particular, AGR-ST is the concatenation of SUBJ and COMPS: (2) a. Mang-ida diri-na si AV-saw self-his John saw himself i. PM John. John b.*mang-ida si John diri-na. AV-saw PM John self-his c. S SUBJ 1 1 NP SUBJ 1 COMPS 2 ARG-ST 1, 2 mang-ida 2 NP diri-na si John Compare this now with objective voice (OV) verbs. Again, using the distribution of reflexives as a diagnostic, we now have to assume that the VP-internal NP has the status of a subject. But this means that in the OV case, valence and argument structure are aligned in a way that is precisely opposite from the AV cases. (3) a.*di-ida diri-na si John. OV-saw self-his PM John b. Di-ida si John diri-na. OV-saw PM John self-his John saw himself i. c. 4

5 S SUBJ 2 2 NP SUBJ 2 COMPS 1 ARG-ST 1, 2 1 NP si John diri-na di-ida By separating information about valence (i.e., syntactic combinatorial potential) from argument structure (the lexically determined list of syntactic and semantic arguments) it becomes possible to provide a lexical treatment of a number of phenomena that would otherwise have to be handled in syntactic terms. In turn this keeps structural complexity (in terms of the inventory of genuine syntactic elements) to a minimum. The issue of structual complexity will also be of concern in the next subsection, and in Dependents and Lexical Amalgamation The following subsections deal with two issues in the area of argument structure that appear at first to be independent of each other but turn out to be closely linked in recent HPSG work. First, is there a fundamental distinction between complements and adjuncts and, second, what is the role of the syntactic head in licensing information about missing dependents? Complements and Adjuncts It is a common and generally unquestioned assumption in much of contemporary linguistics that there is a syntactic distinction between complements and adjuncts, and that these two classes of dependents occupy different tree-configurational positions (for example, sister of X 0 for complements vs. sister of X for adjuncts). This was also the position of early HPSG work. However, the evidence for this syntactically encoded complement/adjunct dichotomy has recently been re-examined within HPSG. For example, Hukari and Levine (1994, 1995) show that there are no clear differences between complement extraction and adjunct extraction, and Bouma et al. (2001) build on these observations and propose a unified theory of extraction based on the assumption that there is no structural distinction between complements and (at least a class of) adjuncts. Earlier, eliminating the configurational distinction was proposed in Miller (1992) (on the basis of French agreement facts, inter alia), van Noord and Bouma (1994) (on the basis of semantic ambiguities in Dutch clusters), and Manning et al. (1999) (on the basis of the behavior of Japanese causative constructions). This adjuncts-as-complements approach is further defended on the basis of case assignment facts in Finnish and other languages (Przepiórkowski 1999c, 1999a), and on the basis of diachronic considerations (Bender and Flickinger, 1999). The central idea of all these analyses is that (at least a class of) adjuncts must be added to the verb s subcategorization frame at the lexical level and are thus indistinguishable from 5

6 complements in syntax. For example, in the analysis of Bouma et al. (2001), words are specified for the attribute DEPS(ENDENTS), in addition to the attributes ARG-ST and VALENCE discussed in the previous section. ARG-ST encodes the core argument structure, i.e., information about dependents that are more or less idiosyncratically required by the word. This information is eventually mapped into the word s VALENCE attributes, responsible for the syntactic realization of these dependents. However, in Bouma et al. s 2001 account there is an intermediate level between ARG-ST and VALENCE, namely DEPS, which encodes all dependents of the verb, both subcategorized (elements of ARG-ST) and non-subcategorized (adjuncts). In other words, DEPS extends ARG-ST to adjuncts, as schematically illustrated in (4). (4) Argument Structure Extension word... CAT... HEAD verb category DEPS ARG-ST 1 1 list(adjunct) The DEPS list is, in turn, mapped into the VALENCE attributes, according to the following schematic constraint. (5) Argument Realization category word... CAT SUBJ 1 VALENCE COMPS 2 list(gap) DEPS 1 2 According to this principle, all elements of ARG-ST, except gaps, must be present on VA- LENCE attributes. There are two things to note about (5). First, gaps (encoding information associated with extracted elements) are present on the DEPS list, but they are not mapped to VA- LENCE. This means that, according to this approach, there are no wh-traces (and, more generally, no empty elements) anywhere in the constituent tree. Second, the configurational distinction between complements and adjuncts is lost here: all elements of the extended argument structure DEPS are uniformly mapped to the VALENCE attributes, regardless of their complement/adjunct status. As we will in the next section, various grammatical processes are assumed to operate at the level of such an extended argument structure Extended Argument Structure Extraction Bouma et al. (2001) propose a theory of extraction that makes crucial use of the extended argument structure encoded in DEPS. They argue that extraction does not distinguish between various kinds of dependents and propose the following principle of SLASH amalgamation to account for this observation. 4 (6) SLASH Amalgamation: word SYNSEM LOC CAT DEPS SLASH 1,...,SLASH n BIND 0 SLASH ( 1... n) 0 6

7 This principle is responsible for collecting SLASH values from all dependents of a word, perhaps lexically binding some of them (this happens in case of words such as tough or easy, which are lexical SLASH-binders), and collecting all other elements of these SLASH sets into the word s own SLASH value. This SLASH value is then shared along the head projection of the word, in accordance with the principle of SLASH inheritance: 5 (7) SLASH Inheritance (schematic): hd-val-ph SLASH 1 HD-DTR SLASH 1 This approach differs from earlier HPSG approaches to extraction not only in that it treats dependent extraction and argument extraction uniformly, but it also establishes a different division of labor between parts of the grammar. In the analysis sketched above, the amalgamation of SLASH values takes place at the level of words, never at the level of phrases phrases only pass SLASH values to the head-filler phrase, where extracted elements are overtly realized. See Bouma et al for further details and examples. Similar lexical amalgamation is also assumed for the purposes of the lexical analysis of quantifier scoping in Manning et al. (1999) and Przepiórkowski (1997, 1998), and for the flow of pragmatic information in Wilcock (1999). One important aspect of the SLASH Amalgamation Principle (6) is that it does not distinguish between slashed arguments and slashed adjuncts: since, in principle, any DEPS element can be a gap, any DEPS element, whether an argument or an adjunct, may be extracted by the same mechanism, in accordance with the observations in Hukari and Levine (1994, 1995). Case Assignment Apart from extraction, another phenomenon that, contrary to common assumptions, does not seem to distinguish between complements and adjuncts is syntactic case assignment. For example, Maling (1993) argues at length that some adjuncts (adverbials of measure, duration and frequency) behave just like objects with respect to case assignment and, in particular, notes the following generalization about syntactic case assignment in Finnish: only one NP dependent of the verb receives the nominative, namely the one with the highest grammatical function; other dependents take the accusative. Thus, if no argument bears inherent case, the subject is in the nominative and other dependents are in the accusative (8), but if the subject bears an idiosyncratic case, it is the object that gets the nominative (9). Furthermore, if all arguments (if any) bear inherent case, and the next available grammatical function is that of an adjunct, then this adjunct takes the nominative (10) (11). (8) Liisa muisti matkan vuoden. Liisa.NOM remembered trip.acc year.acc Liisa remembered the trip for a year. (9) Lapsen täytyy lukea kirja kolmannen kerran. child.gen must read book.nom third time.acc The child must read the book for a 3rd time. 7

8 (10) Kekkoseen luotettiin yksi kerta. Kekkonen.ILL trust.passp one time.nom Kekkonen was trusted once. (11) Kekkoseen luotettiin yhden kerran Kekkonen.ILL trust.passp one time.acc Kekkonen was trusted for one year once. yksi one vuosi. year.nom Maling (1993) concludes that syntactic case is assigned according to the grammatical hierarchy and that (at least some) adjuncts belong in this hierarchy. On the basis of these facts, as well as other case assignment facts in Korean, Russian, and especially Polish, Przepiórkowski (1999a) provides an HPSG account of syntactic case assignment taking extended argument structure (i.e., DEPS, assuming Bouma et al. s 2001 feature architecture) as the locus of syntactic case assignment. (See 6.3 below, and Przepiórkowski 1999a, ch. 10 for details.) 2.3 Argument Composition Moving subcategorization information into lexical descriptions is at first blush a simple redistribution of labor between the syntax and the lexicon. But it turns out that this move affords a much wider perspective on the kinds of relationships that are lexically encoded. In particular, the lexicalization of valence makes it possible to express second-order dependencies i.e., for a word to refer to the valence of its valence elements. The HPSG analysis of controlled complements can be seen as an application of this basic idea, in that the subject requirement of the selected VP is identified with the subject requirement of a predicate selecting that VP: 6 (12) SUBCAT 1, V SUBCAT 1 synsem More generally, since structure-sharing tags in HPSG can be variables over any kind of structure, they can range over the entire list of valence elements of the selected predicator. The valence list of that predicator consists of the verbal complement followed by (using the list-append notation ) the list of dependents of that same complement. This is illustrated in (13), where 1 is used as a variable over lists. (13) SUBCAT 1 V SUBCAT 1 list(synsem) As a result, the arguments of the higher predicator are composed from those of the selected (typically verbal) complement. Another way of thinking about such cases is in terms of the higher predicator attracting the valence requirements of the lower one. Many phenomena for which separate operations of clause union have been assumed in other syntactic frameworks can thus be treated in terms of a rather straightforward head-driven extension of HPSG s original valence mechanism. 7 Among the original applications of argument composition is Hinrichs and Nakazawa s analysis of the German verb cluster, the clause-final sequence of verbal forms (Hinrichs and Nakazawa 8

9 1989, 1994). Starting with Bech (1955), two modes of verbal complementation have been assumed for German. The first (known as the incoherent construction) is very similar to English VP-complement constructions, as for instance in (14): (14) a. Sandy tries VP to read the book. b. daß Otto versucht VP das Buch zu lesen. that Otto tries the book to read A plausible analysis of (14) is that lesen combines with its NP complement (das Buch) and the resulting phrase serves as the VP complement to versucht. However, it is highly debatable whether the same should be assumed for the relation between gelesen and its notional object das Buch in constructions such as (15), where the main verb co-occurs with the tense auxiliaries haben and werden. (15) a. daß Peter das Buch gelesen haben wird. that Peter the book read-psp have-inf will-fin that Peter will have read the book. Hinrichs and Nakazawa propose that in coherent constructions of this kind, the valence requirements of the main verb (here, lesen) are inherited by the governing tense auxiliaries (haben and wird), so that the satisfaction of the main verb s valence requirements are now mediated by the highest governing head element (here, wird). Suggestive evidence for such an analysis comes from the fact that the object of the main verb is subject to the same range of order variation as if the main verb itself had been the sole predicator in the clause. Thus, in (16a) the pronominal object es occurs before the subject Peter, which is precisely parallel to the simple case in (16b): (16) a. daß es Peter gelesen haben wird. that it Peter read-psp have-inf will-fin that Peter will have read it. b. daß es Peter las. that it Peter read that Peter read it. Transformational analyses have usually assumed that such cases are the result of a scrambling transformation that dislocates the object (es) from the phrase that it forms with the main verb. 8 Dislocation constructions are generally treated as filler-gap dependencies in HPSG, because they can typically hold across finite clause boundaries. Since cases like (16) are restricted to a single clause, an analysis in terms of dislocation is inappropriate. Instead, order variation of this kind has been analyzed in terms of permissive linear precedence conditions within a local syntactic domain (typically, a local phrase structure tree). If both subject and object end up as arguments of the highest predicator wird via argument composition, the scrambled order in (16a) can be explained in terms of order variations among daughters within the same local tree, just as in (16b). 9

10 Further evidence against the main verb and its notional object forming a constituent comes the fact that in relative clauses, the two do not form a frontable relative phrase ( VP pied piping ), as seen in (17a). This is in contrast to cases such as (17b), where the governing verb versuchen does combine with a frontable VP dependent: (17) a.*ein Buch das gelesen Peter haben wird a book that read Peter have will b. ein Buch das zu lesen Peter versuchte a book that to read Peter tried a book which Peter tried to read As has been pointed out by Kathol (2000, ), linking the valence requirements of verbal material by means of argument composition does not, in fact, determine the phrase structural relations among the participating verbs. Thus, for typical head-final cases as in (15), there have been proposals that assume no subconstituents among the verbal elements at all (Baker 1994, 1999; Bouma and van Noord 1998b; Bouma and van Noord 1998a), a constituent with rightbranching structure (Kiss 1994; 1995), or a constituent with left-branching structure (Hinrichs and Nakazawa 1989; Kathol 2000) illustrated in (18): 9 (18) 2 3 Vinf... SBCT 1 Vfin... SBCT 1 Vfin... SBCT 1 3 Vinf Vinf... SBCT 1 NPnom,NPacc... SBCT 1 2 lesen können Empirical evidence in favor of such structures is presented by Hinrichs and Nakazawa (1989), who point out, among other things, that the order variation known as Oberfeldumstellung (or aux-flip ) receives an elegant account in terms of reordering of constituents under their leftbranching analysis: (19) Vfin Vfin... SBCT 1 3 wird 2... SBCT 1 3 Vinf... SBCT 1 wird Vinf Vinf... SBCT 1 NPnom,NPacc... SBCT 1 2 lesen können 10

11 In 3, we will return to evidence presented by Kathol (2000) that a purely phrase structure-based view fails to cover the full range of order variation among verb cluster elements seen in German and Dutch. Argument composition analyses in effect establish extended valence relations between a governing verb and the phrasal argument of a more deeply embedded verb. This property has been the basis for novel proposals for the treatment of passives suggested by Kathol (1994) and Pollard (1994). Instead of simply copying the valence requirements of the embedded verb, a passive auxiliary can be thought of as actively manipulating the set of valence elements that they inherit from the governed dependents. As a result, passives on the clausal level can be analyzed as a form of object-to-subject raising. Compared to more standard manipulation of the verb s valence in terms of lexical rules, such an approach has the advantage of only assuming one participle; no distinction between morphologically identical passive and past participles is needed. 10 Other areas of German grammar for which argument composition analyses have been proposed include derivational morphology (Gerdemann 1994) and the problem of preposed complements of nouns (De Kuthy and Meurers 1998). In addition, Abeillé and Godard (1994) have argued that tense auxiliaries in French should be analyzed as inheriting the arguments of their main verbs via argument composition, albeit with a flat constituent structure. Abeillé et al. (1998) show how this idea can be extended to certain causative constructions with faire. Another language for which argument composition has yielded insightful analysis is Korean, both for auxiliaries (Chung, 1993) and control verb constructions (Chung 1998a, cf. also Bratt 1996). Finally, Grover (1995) proposes an analysis of English tough-constructions by means of argument composition, as an alternative to the more standard approach that treats missing objects inside the VP complements of tough-adjectives as the result of an extraction. 3 Phrase Structure and Linear Order 3.1 Configurationality A theme running through much of the HPSG literature is the lexicalization of relationships that have been treated in tree-configurational terms in other theories. HPSG s binding theory is a prime example of how certain asymmetries among co-arguments can be reinterpreted in terms of obliqueness on valence/argument structure. As a result, there is no longer a need for expressing such asymmetries using structural notions such as c-command. Similarly, variation in phrase order of the kind seen in Japanese or German has typically been seen in terms of liberal linear precedence constraints over flat clausal tree structures rather than the result of manipulating highly articulated phrase structures via scrambling movements (see for instance Uszkoreit 1987 and Pollard 1996 for German and Chung 1998a for Korean). 11 HPSG analyses of this kind are thus similar to recent LFG proposals for describing nonconfigurational languages in terms of flat clause structures (cf. Austin and Bresnan 1996). For instance, free order among nominative and accusative dependents in a verb-final language can be described in terms of the linear precedence constraint in (20a), which requires NPs to precede verbal elements, 11

12 without specifying any order among the NPs. As a result, both constituent orders in (20b) and (20c) are licensed. (20) a. NP V b. S c. S NPnom NPacc V NPacc NPnom V An issue closely related to order variation among phrasal dependents is that of the placement of verbal heads in the Germanic languages (and elsewhere). Given a flat structure analysis for the phrasal constituents of the clause, the different positions of the finite verb in verb-initial and verb-final clauses then reduce to clause-initial vs. clause-final placement of that verb (typically mediated by a binary-valued feature such as INV, familiar from GPSG/HPSG analyses of English subject-auxiliary inversion), cf. Pollard (1996): (21) a. S V+INV NP NP liest Otto das Buch b. S (daß) S NP NP V INV Otto das Buch liest Analyses of this kind diverge starkly from the standard transformational approach in terms of movement of the finite verb from its clause-final base position to a clause-initial position (Comp) via head movement. The underlying intuition that verb placement is dependent on constituent structure is in fact also shared by various HPSG-based proposals that offer a number of different ways in which verb movement may be implemented in HPSG, cf. Kiss and Wesche (1991), Kiss (1995), Netter (1992), Frank (1994). The representation given in (22) illustrates how to capture the dependency between the finite verb (liest) and its putative base position (occupied by an empty category) in terms of the additional nonlocal feature DSL (for double slash ): 12 12

13 (22) SUBCAT V DSL {} V SUBCAT SUBCAT V DSL {V} SUBCAT V DSL {V} liest NP Otto NP SUBCAT NP V DSL {V} SUBCAT NP,NP V DSL {V} das Buch t Thus, much like SLASH is used to thread information about phrasal constituents from the gap site to the filler, DSL does the same for finite verbs occurring in verb-first or verb-second constructions in German. Accounts of verb placement in terms of nonlocal dependencies of this kind are discussed by Kathol (1998) and Kathol (2000), who points out that none of the putative evidence for a dislocation-based analysis in fact holds up under closer scrutiny. 13 In addition, Kathol notes a number of technical and conceptual problems involving the locality of the dependency and the existence of dislocated heads. One area in which verb dislocation approaches appear to provide better analyses than those based on ordering variation within local trees is the interaction between finite verbs and complementizers. German and most other Germanic languages exhibit a characteristic complementarity of distribution between initial finite verbs and complementizers in root and subordinate clauses, respectively. 14 If verbs and complementizers are not subconstituents of the same local tree, it is not clear how they can be made to interact positionally. In contrast, verb movement analyses are able to express a direct functional analogy between those two categories, which can account for the distributional facts. However, like their transformational counterparts, such analyses fail to generalize to phrasal clause-initial categories that is, wh-phrases in subordinate interrogative and relative clauses which share the basic distributional and functional properties of complementizers (cf. Kathol and Pollard 1995 and Kathol 2000, for extensive discussion of this point). In fact, one of the major motivating factors behind the linearization-based approach to Germanic clausal syntax pursued in Kathol (2000) is precisely to express this basic parallelism in a comprehensive account of the linear underpinnings of Germanic clause structure. As we will see in the next section, the required extensions of the phrase structure substrate of the HPSG linguistic theory affords a fairly flexible and elegant approach to problems of discontinuous constituency within HPSG. 3.2 Nonconcatenative Approaches to Linear Order In much of contemporary syntactic theory, the correlation between hierarchical organization and linear order in terms of a left-to-right concatenation of the leaves of the syntactic tree ( terminal 13

14 yield ) is taken for granted. However, an interesting consequence of the sign-based approach is that the very ingredient that gave HPSG its name ( phrase structure grammar ) turns out to be a nonessential part of the formalism. While simple concatenation is one mode of computing the phonology of a sign from the phonology of its constituent parts, other relations are perfectly compatible with the sign-based approach. There is now a significant literature that explores such alternatives. Concatenative approaches lends themselves rather straightforwardly to the description of the relation between constituency and order in a language like English. However, it is far less clear whether this also holds of languages such as German. For instance, Reape (1993, 1994, 1996) observes that in German nonfinite complementation constructions of the kind illustrated in (23a), the verb zu lesen occurs separated from its notional object dieses Buch unlike in the English counterpart (23b). (23) a. daß dieses Buch niemand zu lesen versuchte. that this book.acc no one.nom to read tried that no one tried to read this book. b. that no one tried to read this book. The argument composition approach sketched above in 2 attributes this discontinuity to the formation of a complex predicate (zu lesen versuchte). Reape instead proposes analyzing the German and English constructions in terms of the same basic constituent types (in particular VPs), yet realized in a discontinuous fashion in German. This is illustrated in (24), where each sign now is augmented with a list-valued feature representing that sign s (WORD) ORDER DO- MAIN. Linear order is determined by mapping the phonology of the domain elements onto the phonology of the phrase, rather as the terminal yield of the constituent structure. This is indicated below in (24) by arrows linking the phonology of individual domain elements to the phonology of the entire constituent. While in standard phrase structure grammar, the region of (potential) order variation is the local tree, order domains expand that region to include elements that are not immediate constituents of the sign in question. For instance, in (24), the NP dieses Buch as a complement of zu lesen is not an immediate constituent of the clause; nevertheless it occurs together with the verbal head versuchte within the clause s order domain. As a result, both the clausal and the higher VP node have order domains that contain more elements than immediate syntactic daughters. 14

15 (24) S PHON dieses Buch niemand zu lesen versuchte dieses Buch niemand zu lesen versuchte DOM,,, NP NP VP NP dieses Buch zu lesen versuchte niemand DOM,, NP V V V V V versuchte VP dieses Buch zu lesen DOM, NP V NP dieses Buch V zu lesen Reape s proposal bears a strong resemblance to previous approaches to discontinuous constituents, in particular Pullum and Zwicky s notion of liberation (Pullum 1982, Zwicky 1986) (for related ideas in Categorial Grammar, see Bach 1981 and especially Dowty 1996). Thus, the VP in (24) can be thought of as being liberated in the sense that its immediate constituents may intermingle with elements from outside the VP. Unlike Pullum and Zwicky s proposals, HPSG order domains provide a level of syntactic representation from which the range of possible intermingling effects can be represented directly. Thus, while the VP dieses Buch zu lesen gives rise to two list elements in the clausal domain, the NP dieses Buch contributes only one element. Since domain elements cannot themselves be broken apart, it is predicted that discontinuities are allowed only in the former case, but not in the latter. Finally, if order domains take the place of local trees as the range of potential order flexibility, it is natural to interpret linear precedence constraints as well-formedness conditions over order domains rather than as order constraints on daughter nodes in trees Linearization-Based vs. Valence-Based Approaches The initial appeal of structures such as (24) is that they allow an analysis of German that, despite differences in linear order, is remarkably similar to the constituency commonly proposed for the equivalent English sentences in nontransformational approaches. Therefore, it appears that argument composition and order domains constitute two alternative ways of allowing for embedded verbs and their objects to occur discontinuously in a middle distance dependency construction. There are, however, empirical reasons for preferring one approach over the other. As discussed in detail in Kathol (1998), Reape s DOMAIN analysis is 15

16 ultimately unsatisfactory in that it fails to link the argument structure of more deeply embedded predicates to that of the governing verb which is precisely the main intuition behind the argument composition approach. Evidence that such linkage is in fact necessary comes from a phenomenon known as remote (or long) passive (cf. Höhle 1978). In (25), the NP der Wagen is the direct object of the embedded verb zu reparieren, yet its nominative case marks it as the subject of the passivized predicate wurde versucht. (25)?Der Wagen wurde zu reparieren versucht. the car-nom was to repair tried Someone tried to repair the car. If all predicates of the versuchen-class invariably embed VPs, as suggested by Reape, the direct object of the embedded verb (den Wagen in (25)) would never be visible to the valence change that accompanies the passivization of versuchen. 15 Thus, Reape s approach wrongly predicts that such constructions should not exist. In contrast, the argument composition approach can easily accommodate such cases because the syntactic arguments of the most embedded verbal predicate become the syntactic dependents of the governing predicates. Even though facts such as these cast doubt on the appropriateness of order domains in description of the particular phenomena that they were originally developed for, there nevertheless appear to be other discontinuous constituency phenomena for which order domains represent an elegant descriptive tool. For instance, Kathol (1998) points out that argument composition of the kind proposed by Hinrichs and Nakazawa fails to correctly account for certain orderings within Dutch verb clusters. In Dutch we typically find head-first ordering between the governing verb and the governed subcomplex. For example, in (26a), moet as the highest governing verb precedes hebben gelezen. Combinations of tense auxiliaries and their dependent verbs can generally occur in either order; when they occur in head-final order, as in (26b), the preferred occurrence of the governing verb moet (in standard Dutch) turns out to be between gelezen and hebben. This kind of ordering cannot be described assuming only argument composition and binary branching verbal complexes of the kind initially proposed by Hinrichs and Nakazawa. (26) a. dat Jan dit boek moet 1 hebben 2 gelezen 3. that Jan this book must-fin have-inf read-inf that Jan must have read the book. b. dat Jan dit boek gelezen 3 moet 1 hebben 2. that Jan this book read-psp must-fin have-inf Kathol (2000) shows how facts such as these can be accounted for if argument composition is combined with order domains that permit the discontinuous linearization of governed subcomplexes such as gelezen hebben in (26). Such an analysis goes a long way toward a uniform account of the ordering possibilities in a number of varieties of German and Dutch by factoring out dialect-independent constituency and dialect-dependent linearization constraints. 16

17 3.2.2 Further Applications Another area in which the adoption of order domains has arguably led to significant progress is in the syntax of left-peripheral elements in German. As was pointed out above, the striking interplay between finite verbs and complementizers (and wh-phrases, for that matter) which forms the basis of transformational verb movement accounts has been captured only insufficiently in purely phrase structure-based approaches. However, if order domains are combined with the concept of topological fields from traditional German grammar, these facts can be described straightforwardly in purely nonderivational terms (cf. Kathol 2000). The basic idea is to allow elements with different grammatical roles within the clause verbal head, phrasal complements, filler phrase, complementizer, etc. all to occur within the clause s order domain and assign each of them to a topological field such as Vorfeld (vf.) (roughly equivalent to Spec,CP), linke Satzklammer (l. S.) (roughly equivalent to Comp), or Mittelfeld (mf.), etc., determined either lexically or by the combination schema. With the further constraint that the leftmost topological fields (Vorfeld, linke Satzklammer) can be instantiated by at most one element, the distributional complementarity of complementizers and finite verbs follows as a natural consequence. Thus, in (27) the finite verb cannot be associated with the same field as the complementizer and must instead occur clause-finally (rechte Satzklammer (r. S.)). (27) DOM l. S. mf. mf. r. S. PHON daß, PHON Lisa, PHON die Blume, PHON sieht COMPL NPEM NOM NPEM ACC VEM FIN In verb-first constructions such as (28) by contrast, there is no complementizer blocking the l. S. position, hence the finite verb can (and in fact, must) occur there. (28) DOM l. S. PHON sieht VEM FIN mf. mf., PHON Lisa, PHON die Blume NPEM NOM NPEM ACC Typical verb-second declarative clauses involve the instantiation of Vorfeld by a non-wh-phrase and linke Satzklammer by a finite verb, as shown in (29): (29) DOM vf. PHON die Blume NPEM ACC l. s. mf., PHON sieht, PHON Lisa VEM FIN NPEM NOM Kathol (1999, 2000) further describes how clausal domains of this kind can be utilized in a constructional approach (see 4.2 below) to German sentence types with various kinds of illocutionary force potential. While much of the work employing order domains has concentrated on German (see also Richter 1997 and Müller 1999, ch. 11, Müller 2000), there have been numerous adaptations of linearization-based ideas for a variety of other languages, including Breton (Borsley and Kathol 2000), Danish (Hentze 1996 and Jensen and Skadhauge 2001), Dutch (Campbell-Kibler 17

18 2002), English (Kathol and Levine 1992), Fox (Crysmann 1999b), French (Bonami et al. 1999), Japanese (Calcagno 1993 and Yatabe 1996, 2001), Ojibwe (Kathol and Rhodes 1999), European Portuguese (Crysmann 2000b), Serbo-Croatian (Penn 1999a, 1999b), and Warlpiri (Donohue and Sag 1999). One of the ongoing issues in the literature on nonconcatenative approaches to syntax is the precise informational content of the elements of order domains. In Reape s original formulation, order domains contain HPSG signs. But this allows for the formulation of many linear precedence constraints for which there is little or no empirical evidence. As a result, there have been proposals (cf. Kathol 1995) to limit the informational content of domain elements, i.e. the features appropriate for order domain elements. This can be seen as closely related to other proposals that utilize the architecture of features to express linguistically contentful constraints ( geometric prediction ). For instance, the idea that dependents are represented on valence lists as objects of type synsem rather than sign makes predictions about which properties of dependents can be selected by heads (e.g. syntactic category and semantic type, but not phonology). In the case of linearization, the equivalent issue is which aspects of linguistic information appear never to be relevant for linear precedence relations; these features should be rendered inaccessible by means of the feature geometry. For instance, it appears that linear precedence constraints are not sensitive to internal phrase structure, i.e. the number and kind of immediate constituents, as encoded in the DAUGHTERS value. The DOMAIN model should therefore be restricted in certain ways, but it can also be extended in other ways. For the analysis of phenomena involving floating affixes, it has been proposed that domain elements can represent objects smaller than words. 16 This makes it possible to use linearization constraints to handle discontinuous realization of words in the same way as discontinuous phrases. 4 Syntactic Abstractness and Reductionism In this section we survey some developments in HPSG that seem to be primarily methodological issues, but on closer inspection also have empirical ramifications. These have to do with the reality of phonologically empty syntactic constituents and the division of labor between the lexicon and the combinatorial apparatus in expressing syntactic generalizations. The overriding concern in both is the question of how abstract we should assume syntactic representations to be. 4.1 The (Non-)Reality of Syntactic Traces With the introduction of the structure preserving constraint on transformations in the seventies, the notion of a trace as the residue of movement operations became a core ingredient of transformational theories of grammar. The presence of inaudible copies of dislocated elements within the syntactic representation has been of crucial importance for the formulation of many principles in transformational theories, including binding, scope of quantificational expressions, distribution of case-marked elements, and constraints on extraction. The definition of a trace in HPSG is quite straightforward. One can simply see it as a phrasal element of some category (usually nominal or prepositional) that is phonologically empty and 18

19 contributes its own local information to the set of nonlocal SLASH information: (30) PHON LOCAL 1 SYNSEM NONLOCAL SLASH { 1 } However, the reliance on such phonologically empty syntactic elements is generally considered to go against the spirit of HPSG as a surface-oriented theory. This holds for all kinds of empty categories, not only traces (e.g. wh-trace and NP-trace), but also pro, PRO, and the many empty operators and empty functional heads that are assumed in other frameworks. The discussion in this section focuses on wh-trace, because most of the other empty categories have never been proposed in standard HPSG analyses. For example, PRO is not needed in infinitival constructions, because the unrealized subject is identifiable as an unsaturated valence element, and NP-trace is not needed in the HPSG treatment of the passive alternation, which involves related but distinct verbal lexical entries. It should be said that some authors do in fact take advantage of the fact that HPSG can technically accommodate empty categories. As discussed in 3.1 above, a number of proposals for German clause structure assume a head movement analysis with clause-final verbal traces. And the account of relative clauses in Pollard and Sag 1994 relies crucially on syntactically complex but phonologically empty relativizing operators. For both of these cases, however, subsequent research has shown that alternative analyses are available that do not involve empty categories (recall 3.2 and see the next section). The main issue that remains to be considered is therefore the elimination of wh-trace. And in fact, the treatment of extraction in terms of traces in the syntactic structure proposed in HPSG2 was supplanted right away in HPSG3 by a traceless approach involving several lexical rules, and later by the unified head-driven constraint-based analysis sketched in Extraction is encoded as a mismatch between the list of potential syntactic dependents DEPS and the elements on the valence lists, which correspond to canonically realized dependents. An extracted element is instead identified as a gap, a non-canonical subtype of synsem, and its LOCAL value is added to the SLASH set. 17 SLASH information propagates by head-driven inheritance and eventually licenses the appearance of a filler that discharges the long-distance dependency. The syntactic evidence typically offered in support of wh-traces can be equally well accounted for by referring to the ARG-ST list, whose membership remains unchanged even if arguments are extracted. For instance, fillers in English topicalization constructions can be reflexives with an antecedent in the following clause (31a) notwithstanding the fact that the reflexive is presumably in a configurationally higher position than its antecedent (Pollard and Sag 1994:265). Similiarly, an extracted subject as in (31b) can still serve as antecedent for a reflexive object of its original verb. (31) a. (John and Mary are stingy with their children.) But themselves i, they i pamper. b. Which man i do you think perjured himself i? 19

20 In transformational analyses, these in situ effects are analyzed by assuming the presence of a trace at the extraction site, but this is unnecessary in HPSG, because the relevant reflexive binding constraints apply to the ARG-ST list of the verb. Many aspects of extraction phenomena are open to both trace-based and traceless analyses in HPSG, but there are empirical motivations for preferring one technical approach to the other. As has been argued by Sag and Fodor (1994), the evidence for the existence of traces proposed in the literature is often extremely weak. At the same time there are phenomena that can be explained more straightfowardly if no traces are assumed in the syntactic structure. As an example of arguments of the first kind, consider wanna-contraction, one of the most celebrated pieces of evidence in favor of traces. The basic idea is that wh-traces disallow the phonological contraction of want and to. The relative clause in (32a) is ambiguous between a subject or object control reading for the understood subject of succeed. In contrast, the variant in (32b) is only said to permit the subject control reading, supposedly because of the impossibility of contraction across a wh-trace. (32) a. This is the man I want to succeed. b. This is the man I wanna succeed. However, as has been pointed out by Pullum (1997), there are numerous technical and conceptual problems with this explanation. For instance, whether contraction is possible appears to be highly lexically specific: gonna, hafta, but *intenna (intend to), *lufta (love to), *meanna (meant to). This suggests that contraction cannot be a general process. Instead, a fully lexical, traceless analysis of the above contrast is available if wanna is thought of as syntactically underived subject-control verb that does not license an object. Pullum is able to explain all of the phenomena previously discussed in the literature, in addition to data distinguishing his proposal from others that have been advanced. Turning to positive evidence against traces, a strong argument in favor of their abolition comes from data involving extractions from coordination, first discussed by Sag (2000) (see also Bouma et al. 2001). The well-known Coordinate Structure Constraint requires that each conjunct be affected equally in extractions from conjoined phrases; in particular, extraction must apply in an across-the-board fashion. This straightforwardly explains the ungrammaticality of (33): (33) *Who i did you see i and Kim? However, as Sag points out, examples such as the following are also ungrammatical, even though here, the extraction affects each conjunct in a parallel fashion: (34) a.*who i did you see i and a picture of i? b.*which student i did you find a picture of a teacher of i and i? c.*who did you compare i and i? The pertinent generalization is that no conjunct can consist of an extraction site with no other material. This Conjunct Constraint has to be stipulated in addition to the across-the-board 20