File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/92/c92-1040_metho.xml
Size: 22,577 bytes
Last Modified: 2025-10-06 14:12:55
<?xml version="1.0" standalone="yes"?> <Paper uid="C92-1040"> <Title>CONCEPTUAL STRUCTURES AND CCG: LINKING THEORY AND INCORPORATED ARGUMENT ADJUNCTS</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 1 Introduction </SectionTitle> <Paragraph position="0"> Tile present paper represents the first attempt to integrate Mark Steedman's theory of Combinatorial Categorial Grammar (CCG) \[Ste90, Ste91\] with Ray Jackendoff's theory of Conceptual Semantics \[Jac90, ~lac91\]. The former is known for its successhd treatment of tong-distance dependencies, coordination, and, more recently, matters of discourse focus relating to intonation---none of which have been treated within Conceptual Semantics. The latter is known for its development of conceptual structures, which are mental representations intended to serve as the link between language and other areas of cognition, e.g. vision, action and inference--which CCG stops short of. Since CCG is a. lexically oriented theory of grammar, the two are entirely compatible, as well as complementary.</Paragraph> <Paragraph position="1"> The immediate motivation to attempt such an integration, and the focus of the present paper, is CCG's incomplete treatment of sameness of role (i.e. thematic) information. In CCG, semantic function-argument structures are compositionally produced through the course of a derivation. These structures identify, inter alia, which entities play the same roles in different events for expressions involving a wide range of coordinate constructs. For example, the semantic function-argument structure shown in (lb) is derived for the sentence in (la) via type-raising, composition, and co-ordination of the bracketed non-standard constituent, following the analysis of Dowty \[Dow88\]: 1 (la) Jack filled \[the urn with coffee\] and \[the thermos with milk\].</Paragraph> <Paragraph position="2"> (lb) (fill' urn' coffee' jack') & (fill' thermos' mill&quot; jack') Of course, such semantic function-argument structures are intended only for illustrative purposes; indeed, according to Steedman, semantic constants like urn' are &quot;mere placeholders for a real semantics, intended to do no more than illustrate this compositionality.&quot; Nevertheless, we may glean from these structures the requirement that urn' and thern*os' play the same semantic role, since they are both first arguments to fill', and likewise for coffee' and milk', since they are both second arguments. In the terminology of Conceptual Semantics, these requirements may be restated in terms of thematic roles as follows: urn' and lhermos' share the thematic role Goal in their respective events; like- null in this paper.</Paragraph> <Paragraph position="3"> ACT~ DE COLING-92., NANI'ES, 23-28 Aofrr 1992 2 4 6 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 wise, coffee' and milk' share tile thematic role Theme. 2 Now, while CCG can thus be said to identify thematic information across a wide range of expressions not easily analyzed in other theories, it does not do so across cases of verbal diathesis (i.e, argument structure alternations). For example, consider (2), together with two possible sets of interpretations that follow: (2a) Jack filled the urn {with decaf}.</Paragraph> <Paragraph position="4"> (2b) The urn filled {with decaf}.</Paragraph> <Paragraph position="5"> (2c) Coffee filled the urn {* with decal}.</Paragraph> <Paragraph position="6"> (3a) fill' urn' {decaf~ jack' (35) till' {decaf'} urn' (3c) tilt&quot; urn' coffee' (4a) tiU.' arn' {deck/'} jack' (45) fillb'{decaf'} urn,' (4c) fill~' urn' coffee' Here it would not do to derive the function-argument structures shown in (3), as they incorrectly equate semantic roles in some eases, t'br example, the roles of jack&quot; and coffee&quot; are incorrectly said to be the same for Jack filled the urn and Coffee filled the urn. This problem may be avoided by introducing distinct constants ill//' (with possibly varying arities), as shown in (4). Note, however, that this approach is incomplete, insofar as it fails to equate any semantic roles across the functions til~ ', at least in the absence of further conditions on these functions.</Paragraph> <Paragraph position="7"> To handle these cases as well, the present paper demonstrates how to adapt the solution developed in Conceptual Semantics to fit tile CCG paradigm. This approach may be seen as one method of specifying, in a principled fashion, the further conditions on constants like filli' necessary to give a complete account of thematic role identities. 3 It should not be viewed, however, as a variant of purely syntactic approaches to verbal diathesis, such as the Unaccusative tlypothesis \[Bur86\] in GB, which posit movement between an underlying and a surface structure and traces to recover thematic roles) 2 This restatetaent is actually a cor~iderable atrengtheldngb as CCG is not committed to anything stronger thmt the individual thematic role view (cf. \[Dow91D; thai is, it requi~s no nmre than jack' play the &quot;filler&quot; role, urn' and thermos' play the &quot;filled&quot; role, etc.</Paragraph> <Paragraph position="8"> 3Another viable approach is of course to use ntemffng postulates. A detailed discussion of these alternatives is beyond the scope of tiffs paper (though cf. the discussioa in \[dacPO\]). 4Agaln, cf. also \[J~cgo 3 for independent aa-guments in favor of the Concepttml Sematltics approach.</Paragraph> <Paragraph position="9"> The essence of the present approach is to redefine the Linking Theory component of Conceptual Semantics in teals of CCG categories, so that derivations yield conceptual structures representing the desired thematic information; in this way no changes are required on the CCG side. While this redefinition is largely straightforward, an interesting problem arises ill the case of Conceptual Semantics' Incorporated Argument Adjuncts. In examining these, the paper shows that they cannot be treated as adjuncts in the CCG sense without introducing new machinery, nor without compromising the independence of the two theories. For this reason, the paper instead adopts the more traditionM approach of treating them as oblique arguments.</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Preliminaries </SectionTitle> <Paragraph position="0"> This section reviews the details of CCG mid Conceptual Semmltics needed to understand their integration.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.1 CCC=, </SectionTitle> <Paragraph position="0"> Exan~ple (2) sullices to review the necessary details of CCG. A CCG deriw~tion starts with lexical lookup, which identifies the functional type and semantics of cach constituent. For example, the category of the verb fill needed for .lack filled the urn is as shown in (5): (5) ~ill := ($\NP)/NP : fill_a' In this notation, a category consists of a syntactic category paired via an infix colon with a semantic function. Syntactic categories have arguments appearing to the right of slashes, results to the left. The direction of the slash indicates tile direction of tlle argument. Thus the syntactic category (S\NP)/NP defines a fimction that takes all NP to tile right and returns a function from an NP on the left to an S. Categories may combine via forward or backward functional application, indicated as ACTF~S DE COLING-92, Nhrcre.s, 23-28 AO~'r 1992 2 4 7 PROC. OF COL1NG-92, NANTEs, AUG. 23-28, 1992 > and < in Figure 1. Categories may also combine by other means such as composition, often yielding multiple derivations of the same string. For present purposes this is of no significance, as all of the derivations of will produce the stone compositional meaning. Derivations for the rest of the examples in (2) are quite similar, differing only in the lexically specified category for fill.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Conceptual Semantics </SectionTitle> <Paragraph position="0"> Example (2) again suffices to review the necessary details of Conceptual Semantics. The version of Conceptual Semantics presented below is that of \[Jac90\] prior to the introduction of Linking Theory, plus a few modifications. Let us begin with the representation of an urn. Jackendoff represents an urn as the conceptual structure shown in (6): (~) \[Thing URN\] This represents an entity of ontological type Thing that meets the featural description URN. 5 To distinguish different urns, I will follow Zwarts and Verkuyl \[ZV91\] in requiring all conceptual structures to have an index, as shown in (7a):</Paragraph> <Paragraph position="2"> Note that under the Zwarts and Verkuyl formalization, (7a) is roughly equivalent to the more familiar (7b).</Paragraph> <Paragraph position="3"> In addition to the ontological type Thing, an entity may be of type Place, Path, Event, State, Manner or Property. The Place in t, he urn, for example, would be represented as in (8a): (8&) \[Place IN(\[Thins URN\]/)\]p</Paragraph> <Paragraph position="5"> Here we have a conceptual function IN: Thing ~ Place mapping the urn j to the locatiou inside the urn p. Example (8b) is again an approximate notational variant.</Paragraph> <Paragraph position="6"> Moving on to the stative reading of example (2c), Coffee filled the urn, we introduce the conceptual function BE: Thing x Place -, State (note that as in this example, ontological categories and indices will often ~Sma\]l caps will be used to indicate feature~ that are atomic in Conceptual Structure, serving oldy ~ links to other areas of cognition. be suppressed for typographical convenience):</Paragraph> <Paragraph position="8"> Extending \[Jac90\], I have included the conceptual atom FILL in (9a). As is the case of other categories, this atom serves as a pointer to semantic information not captured by the decomposition. Thus the state s is to be understood as one characterized by the atom FILL and by the feature BE(I,p). Note that the variant in (9b) is reminiscent of the neo-Davidsonian approach adopted by Parsons \[Parg0\].</Paragraph> <Paragraph position="9"> To get the inchoative reading of (2c), we need only add the conceptual function INCH: State -~ Event shown in (10):</Paragraph> <Paragraph position="11"> The conceptual structure for example (2b), The urn filled {with deca~, would differ minimally from (10) by having \[DECAF\] as the Theme instead of \[COFFEE\], or by having the Theme left implicit.</Paragraph> <Paragraph position="12"> We are now in a position to construct the conceptual structure for example (2a), Jack filled the urn {with decal}, by adding the External Instigator func- null Here the representation of the inchoative event serving as the second argument of CS has an implicit Theme k, which the with-PP would specify if present. Note also that the entity \[JACK\]/ serves as both Actor and External Instigator, and likewise \[URN\]/ serves as both Patiant mid Goal, by virtue of coindexation. And again, the variant in (1 lb) indicates the similarity of this approach to the neo-Davidsonian one.</Paragraph> <Paragraph position="13"> At this point we may observe that representations in (9) - (11) capture the similarities and differences in semantic roles observed in tile arguments of the verb ACTES DE COLING-92, NANTES, 23-28 Ao(rr 1992 2 4 8 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 fill in (2). This follows straightforwardly from the inclusion of representations (9) mid (10) within (11), together with the semantic coindexation.</Paragraph> <Paragraph position="14"> Next we turn to a brief description of how these representations are constructed in \[Jac90\]. Two representative lexical entries, that of the stative Jill of (2c) and causative-inchoative fill of (2a), are shown below:</Paragraph> <Paragraph position="16"> in (12), the verb fill subcategorizes an object NP indexed j, as well as an external argument indexed i by convention. Sinfilarly, (13) subcategorizes an object NP and a wi~h-PP. Arguments to the verb are integrated into the above conceptual structure using the Argument Fusion Rule, which links the coindexed constituents in the obvious way, as long as they are semantically compatible.</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 Linking Theory </SectionTitle> <Paragraph position="0"> This section details how the Linking Theory component of Conceptual Semantics can be redefined in terms of CCG categories, so that derivations yield conceptual structures like (9) - (11). Before introducing Linking Theory, however, we shall first examine how tim version of Conceptual Semantics presented in the last section can he adapted to fit the CCG paradigm.</Paragraph> <Paragraph position="1"> As was suggested in ,Section 1, the present approach may be seen as specifying constraints on the constants \]Jill ' so that the desired thematic role identities are captured. This may be done by simply redefining lexical entries like (12) and (13) as follows:</Paragraph> <Paragraph position="3"> IIere the subcategorization frmnes have been replaced by the appropriate CCG categories, and the conceptual structures have been made into the appropriate functions corresponding to the filli ' constants. Because this information is supplied lexically, no changes need be made on the CCG side. Thus conceptual structures for sentences like those in (la) and (2) may be easily derived with the addition of just a few more lexical items like those below: 6 Given such lexical items, the constants appearing in (lb) and (4) may be replaced yielding fimctions like the first one appearing in Figure 2, which is equivalent mudulo an appropriate definition of \]?-reduction to tile one appearing; beh)w it. Such a defnition must mirror that of Argument Fnsion, insofar as it must append features specified by tile argument to those specified by the head. A schematic version appears in (18): (18) Argument Fusion as ~-reduction Schema: (Ax.\[... \[XFEATS\] .... \] \[YFEATS\]v ) Tlirning now to the introduction of Linking Theory, we may observe that there is nothing in theory as presented to this point which would eliminate hypotheticM verbs such as delli~ below \[Car88\]) which would have (19) meaning Jack filled the urn.</Paragraph> <Paragraph position="4"> (19) * The urn dellifed Jack.</Paragraph> <Paragraph position="5"> To capture such generalizations, Jackendoffproposes to eliminate rigid coindexation between syntactic and semantic structures, opting to introduce Linking Theory to handle this task instead. Lexieal entries are therefore modified to indicate only which conceptual constituents must be specified, and not which syntactic constituents must specify them. The selected conceptual arguments are annotated with an A, or A-marked. In present terms, this means changing entries like (15) to ones like (20): ~' fill, dellif</Paragraph> <Paragraph position="7"> Categories like the one in (15) thus become derived inrThe w/~h-PP is unselected for exp~ititory reasons only.</Paragraph> <Paragraph position="8"> stead of lexically specified, with Linking Theory specifying constraints on such derivations to permit Aji as the only possible argument ordering. The central idea behind such constraints is as follows: Given (independently motivated) syntactic and semantic hierarchies, do not allow inconsistent orderings. This is stated more formally in (21): (21) Linking Principle: A semantic function headed by Aa:l...xn in a CCG category must not have Xi -~sem ~j and ~i ~'~yn x/, or vice-versa, where -~sern and &quot;~sy, encode the semantic and syntactic hierarchies, respectively.</Paragraph> <Paragraph position="9"> Note that if Actor &quot;~sem Patient and Subject -%yn Direct Object, then the ordering Aij (with indices as before) required for dellif is indeed ruled out by the Linking Principle. s As developed so far, tile status of the Linking Principle in the present framework is that of a filter on representations. The Linking Principle may be made more constructive by eliminating syntactic specifications from lexical entries, following (say) Rappaport and Levin \[RL88\] or Pinker \[Pin89\], deriving them instead via Linking Rules which obey the Linking Principle. :\]ackendoff does not rule out this possibility, but chooses to develop instead an approach in which both syntactic and semantic subcategorization is retained.</Paragraph> <Paragraph position="10"> I shall part company with Jackendoff on this issue, as I find his arguments in favor of retaining subcategorization unconvincing. These arguments are twofold. First, verbs appear to idiosyncratically specify prepositions. Such verbs may be accommodated within the present framework by simply providing fully specified categories like (15). Second, and more interestingly, some Incorporated Argument Adjuncts arc syntactically obligatory. This argument presupposes, of course, a Observe that the syntactic ordering cannot be defined from the syntactic type if subjects are to precede complements.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> CCG With-Theme Adjunct, </SectionTitle> <Paragraph position="0"> the correctness of the Incorporated Argument Adjunct analysis, to which we now turn.</Paragraph> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Incorporated Argument Ad- </SectionTitle> <Paragraph position="0"> juncts Jackendoff observes that with-PPs may specify an optional Theme argument across a wide range of verbs. This observation leads him to hypothesize that such with-Themes should not be treated as subcategorized arguments, bnt rather as adjuncts. Such au analysis is partieulary appealing in cases involving an incorporated Theme, such as butter, as in Jack butfeted ~he bread with that yacky stuff. 9 We shall see, however, that this analysis cannot be adapted into the present framework without adding substantial new machinery, 1deg nor without compromising the independence of the two theories. In contrast, the traditional oblique argument analysis will be seen to surmount these difficulties in a natural way.</Paragraph> <Paragraph position="1"> Jaekendoff's informal version of the With-Theme Adjunct Rule is repeated below: (22) With-Theme Adjunct Rule: In a sentence containing with NP in the VP, if the Theme position is not indexed in the verb's lexical entry, then the object of with cazl be interpreted as Theme.</Paragraph> <Paragraph position="2"> With (22) in mind, one might try to redefine (17) as follows:</Paragraph> <Paragraph position="4"> Here with is defined as a function from an NP to a VPnit shtmld be noted that Jackendoif does not adequately address the issue of why the class of apread~ appropriate for the verb butter ia larger than the class apppopriate for the ,mue noun.</Paragraph> <Paragraph position="5"> ldegAs was done in m~ ear|ier vexalon of this paper.</Paragraph> <Paragraph position="6"> modifier, where the constant WithTheme stands in for the function that fuses the Theme with the specified NP. A sample syntactic derivation using (23) is shown in Figure 3.</Paragraph> <Paragraph position="7"> There are two problems with adequately specifying the function withTh~me. First, one might question its introduction on theoretical grounds, as it marks a substantial departure from the simple rule (18), /3reduction as Argument Fusion, compromising the independence of the two theories. Second, there is an empirical problem of avoiding examples like (2c), * CoSec filled the urn wi~h decaf . In (22), Jackendoff stipulates that the Thenm position be unindexed in the verb's lexical entry. This argmneut indexing information is no longer available, however, at the point in the derivation in which the withThcm e constant is to perform its magic, since the function (f x) is already saturated.</Paragraph> <Paragraph position="8"> Again, while adequate fixes might be possible, any such approach would seem quite ad hoc.</Paragraph> <Paragraph position="9"> Instead of treating these wilh-PPs as adjuncts, we may rehrterpret Jackendoff's (22) as a Linking Rule for oblique with-Theme PP arguments. This rule would then bc just one of those necessary to derive the cate~ gory in (15) from tim lexical entry in (24) below; other rules would map Actors to Subject NPs, Patients to Direct Object NPs, etc. Note that in this entry the subcategorization of the Theme argmnent is indicated to be optional by tim curly braces: fill V (24) Under this formulation, both of the problen~q mentioned above disappear: first, the Theme's specifica~ tion again becomas like that of any other argument, and second, the ungrammatieality of * Coffee filled the urn with decal again becomes a straightforward consequence of the independently motivated (Neo) 0Criterion. null At this point we may return to Jaekeadoff's argument in favor of retaining syntactic subcategoriza~ tion. After having chosen to treat oblique argumeuts as Incorporated Arguumnt Adjuncts, Jackendoff then observes that they are not always optional. Rathcr than retreat, however, he suggests that these are cases of syntactic subcategorization not matching semantic subcategorization. For example, consider (25): (25) Jack rid the room {* 0 / of insects}.</Paragraph> <Paragraph position="10"> The verb *~d is like empty in taking an of-Theme Pl', AcrEs lJ~ COLlNG-92, NANTES, 23-28 AO~r 1992 2 $1 PROC. or: COLING-92, Na~rr~.s, AUG. 23-28, 1992 semantically the inverse of the with-Theme PP. Unlike empty, however, the PP is obligatory for r/d. This leads Jackdendoff to posit a lexieal entry like (26), in which the Theme is not A-marked but the PP is obligatory. Such lexical entries are then used to argue in favor of retaining syntactic subeategorizatiou. This rather unusual move does not seem to be necessary, however. ~ Consider the representation adopted in the present framework, appearing in (27). This representation adequately captures r/d's idiosyncratic selection facts by simply requiring the A-marking of the Theme, forcing the appearance of the with-PP. Of course, to the extent that the existence of lexical entries like (26) is called into question, the argument following from their existence becomes likewise suspect.</Paragraph> </Section> class="xml-element"></Paper>