File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/92/c92-1048_metho.xml
Size: 19,072 bytes
Last Modified: 2025-10-06 14:12:55
<?xml version="1.0" standalone="yes"?> <Paper uid="C92-1048"> <Title>VP Ellipsis and Contextual Interpretation</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Background: The Incremen- tal Interpretation System </SectionTitle> <Paragraph position="0"> A semantic representatiou in the Incremental Interpretation (henceforth II) System is called a &quot;Conditional Interpretation&quot;, which is defined as an assumption-sense pair, A:s, where A is a set of assumptions, and s is the sense. The sense can be thought of as the ordinary truth-conditional semantic representation. The assumption set consists of assumptions that have been introduced during the derivation, and must be discharged before the derivation is complete. The assumption set &quot;represents constraints ou how the sense may be further connected to its context.&quot; \[22\] The process of interpretation is defined by a set of structural rules and a set of discharge rules. Structural rules build the conditional interpretation of a phrase compositionally, from the conditional interpretation of its parts. Discharge rules remove assumptions. In principle all rules have an input and output discourse model, but only the discharge rules actually interact with the discourse model.</Paragraph> <Paragraph position="1"> The form of a structural rule is P ~ A:s ff P1 ~ Al:sl and ...and Pk ~ Ak:s:C/ The ~ denotes the interpretation relation between a node of a syntactic analysis tree (produced by the parser) and a node of a semantic derivation tree. P denotes a syntactic node, where its immediate constituents are denoted by variables P1 through Pk. The rule schema is to be understood as stating a constraint that P receives the interpretation A:s if it has constituents PI through Pk, and these constituents have the interpretations indicated.</Paragraph> <Paragraph position="2"> The form of a discharge rule is</Paragraph> <Paragraph position="4"> Here, A I ~- A - {P~}, where R is the discharged assumption. The discharge of R, together with the current state of the discourse model, determines some modifications to s, resulting in s ~.</Paragraph> <Paragraph position="5"> The assumption storage mechanism is based on Cooper storage \[3\], which was applied to quantifier phenomena. In the II system, this mechanism is applied to several additional phenomena. Below, I will ACT~ DE COLING-92, NANTES, 23-28 AoLrr 1992 3 0 4 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 describe the rules for pronominal anaphora and for quantifiers.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.1 Rules for Pronominal Anaphora </SectionTitle> <Paragraph position="0"> The treatment of pronominal anaphora in the Itsystem is similar to the approach in Discourse Representation Theory(\[13\], \[12\]): indefinite NP's introduce new elements in the discourse model. Pronouns and definite descriptions find their referent among elements in the discourse model.</Paragraph> <Paragraph position="1"> Four types of referential NP's are defined: pronouns, definite descriptions, indefinites, and names. They are represented as follows: bind(x,pronoun,number/gender): x bind(x,def,sor t): x bind(x,indef,sort): x bind(x,name,N): x In each case, the sense is represented by a parameter z, and a binding assumption expresses constraints on the way x will be replaced by an entity in the discourse model. This is achieved by discharging the bind assumption. The discharge rules are: A, bind(x,pronoun,number/gender): S =:, A: Six/el A, bind(x,def, sort):S =~ A: Six/el A, bind(x,indef,sort):S =*- A: Six/el A, bind(x,name,N): S :=~ A: Six/el In the ease of pronouns and definite descriptions, the element e must be a salient element in the input discourse model, satisfying the constraints expressed in the binding assumption. An indefinite assumption causes a new element e to be added to the output dis. course model. In each case, e is substituted for each occurrence of x in the sense S. At least for pronouns, there is a second possibility: instead of selecting e from the discourse model, some other, undischarged parameter can be selected. This allows a pronoun to be bound by a quantifier, as described below.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Rules for Quantifiers </SectionTitle> <Paragraph position="0"> The treatment of qunntifiers in the II system essentially duplicates that of Cooper\[3\]. A quantified NP is represented by storing a quantifier assumption, together with a parameter representing the sense. At some later stage in the derivation, the quantifier assumption is discharged, determining the scope of the quantifier. There are two general rules for quantifiers, governing the introduction and discharge of quantifier assumptions. A quantified NP is represented as: bind(x,q,n): x where x is a parameter, q is the quantifier, and n is the common noun. For example, &quot;every jet&quot; is represented null bind(x,every,jet): x Simplifying slightly, the discharge of quantifier assumptions cml be represented as follows: bind(x,q,s): Pt =*&quot; (q s x) p As an example, bind(x,everyjet): fly(x) =~ (every jet x) fly(x) As mentioned above, when a pronoun assumption is discharged, its parameter is replaced either by an entity in the discourse model, or by some, yet undischarged parameter. A pronoun becomes &quot;bound&quot; by a quantifier if the quantifier parameter replaces the pronoun parameter in this way.</Paragraph> </Section> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 The Account of VP Ellipsis </SectionTitle> <Paragraph position="0"> 1 now describe a semantic account of VP ellipsis in terms of some simple extensions to the II system. The approach parallels the above approach to pronominal anaphora. I will define a rule to add VP-meanings in the discourse model, and a rule for recovering those VP-meanings to resolve an elliptical VP. Thus full VP's are analogous to indefinite NP's, in that they both typically introduce semantic objects into the discourse model, and elliptical VP's are analogous to pronouns, in that their interpretation requires the selection of an appropriate object from the discourse model. The discourse model will have two sets: SE, the set of salient entities, and SP, tile set of salient predicates.</Paragraph> <Paragraph position="1"> To add VP-meoafings to the discourse model, I allow all lexical verbs to introduce all assumption which adds the VP-meaning to the discourse model. I call this binding assumption type &quot;pred&quot;. It is discharged as follows: A, bind(pred):S :=~ A: S where DMou, (SP) = DMi. (SP) U {A:S) That is, tile discharge results in the semantic representation of the VP (i.e.,the ~.ssumption-sense pair A:S) being added to the SP set of the output discourse model.</Paragraph> <Paragraph position="2"> l add the requirement that all arguments except the subject must be filled hefore the assumption is discharged. That is, the discharge of this a~sumption is permitted only if tile sense is of the form ACRES DE COLING-92, NANTES, 23-28 AOI~T 1992 3 0 5 PROC. O1; COL1NG-92, NANTES, AUO. 23-28, 1992 P(SUB~I, at ..... an) where SUBJ represents an unfilled subject argument position, with the remaining arguments al through a.</Paragraph> <Paragraph position="3"> filled.</Paragraph> <Paragraph position="4"> The assumption for recovering a VP-meaning is introduced by a lexical auxiliary verb; this assumption is termed &quot;epred', for elliptical predicate.</Paragraph> <Paragraph position="5"> The discharge rule is: bind(epred): AUX ::~ A:S where A : S is some element of the SP set in DMIn. That is, upon discharge of the epred assumption, an auxiliary verb is replaced by some VP-meaning in the input discourse model.</Paragraph> <Paragraph position="6"> The crucial point in these rules is that the antecedent VP is represented as an assumption-sense pair, since it is the assumptions that represent dependencies on context. For example, the representation of the VP &quot;help him&quot; might be bind(x,pronoun,male): help(SUBJ,x).</Paragraph> <Paragraph position="7"> This expresses the constraint that the object position must be filled by some entity in the discourse model according to constraints of pronominal reference. Two copies of this VP, as antecedent and target in VP ellipsis, could allow the pronoun to refer to different entities, depending on the state of the current discourse model.</Paragraph> <Paragraph position="9"/> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 An Example ~,e(e~,~, </SectionTitle> <Paragraph position="0"> I describe the derivation of example (2), which is repeated here. bh~exkit~)~b / (2) Every boy~ in Bill's class hoped Mary would / ask hhnl out, but a boyj in John's class / actually knew that she would. \[ask himj out\] The derivation is displayed in Figure 14 , in the form of a simplified derivation tree. The derivation tree is defined as follows: each node contains a conditional interpretation, a current discourse model, and a derivation rule R, such that the node is consistent with the b~(Hklji~bo~:z application of 1~ to the node's daughters. For brevity, the discourse models are not displayed, and only certain rules are labeled. A~'s~,s The antecedent VP &quot;ask him out&quot; is represented as bind(pred), bind(y, pronoun,male): ask-out(SUBJ,x).</Paragraph> <Paragraph position="1"> 4Note: it is a~umed that the auxiliary verb contributea tense and polarity. This contribution is ignored in the derivation for the ~e of ~implicity.</Paragraph> <Paragraph position="3"> AVrES o13 COLING-92, NANTES, 23-28 ^olYr 1992 3 0 6 Paoc. OF COLING-92, NANTES, AUO. 23-28, 1992 The discharge of the pred assumption results in bind(y, pronotm,maln): ask-out(SUBJ,x) being added to the discourse model. Later, the binding assumption for the pronoun is discharged, allowing it to be bound by the quantifier every boy. 5 In the interpretation of the elliptical VP, the auxiliary &quot;would&quot; is represented bind(epred):would The discharge of the epred assumption results in the selection of a VP-meaning from the current discourse model: in this case, blnd(y, pronoun,mule):ask-out (SUB3,x) is selected. Later, the binding assumption for the pronoun is discharged, allowing the pronoun to be bound to '% boy&quot;.</Paragraph> <Paragraph position="4"> This example receives a straightforward derivation in the system I am proposing. In other accounts, it appears that examples of this sort cannot be acconlmodated. It is clear, for example, that the example violates the alphabetic variance condition imposed in the Sag/Williams approach. This condition requires that the antecedent mid target VP's be identical up to renaming of bound variables. In the example, the antecedent VP contains a free variable (him/) which becomes him./ in the target, violating the alphabetic variance condition. Partee and Bach\[21\] adopt essentially the same alphabetic variance condition, as does Klein\[15\], so that their accounts also rule out the example. Lappin's\[17\] acemmt explicitly rejects the alphabetic variance condition, replacing it with the following condition: For any two occurrences c~ and fl of an open VP intension C/xl, tr can serve as the antecedent of/~ iff both occurrences of xl can be naturally interpreted as having the same intended range of possible values.</Paragraph> <Paragraph position="5"> The example clearly violates this condition: since him/ ranges over members of Bill's class, and him`/ picks out a member of John's class, they do not have &quot;the same intended range of possible values&quot;. Next, I consider the higher order matching approach of Dalrymple, Shieber and Pereira\[4\] (DSP). In this approach, a matching operation is performed to solve an equation, in which a second order variable represents the elliptical VP. To set up the equation, it is SThe quantifier &quot;every boy in Bill's class&quot; is represented in the derivation as &quot;every boy&quot;, for the sake of brevity. Similarly for &quot;a boy in John's class&quot;. necessary to determine the antecedent clanse and the &quot;parallel elements&quot;, and DSP provide no method Ibr making this determination. &quot;l~\]pieally, with VP ellipsis, there are two adjacent clauses, in which tile second clause contains an elliptical VP. Then the first clause is the antecedent clause and the two subjects are the parallel elements. Applying this to the current example, we have &quot;Mary would ask hiual out&quot; as the antecedent clause, and &quot;Mary&quot; mad &quot;she&quot; as parallel elements. The equation to solve is</Paragraph> <Paragraph position="7"> In this case, the desired solution, Az.ask-out(x,himj) is not a possible solution to this equation, according to tile matching operation used by DSP. This is the most straightforward method of determining parallelism to set up the equation, and it does not permit the derivation of the desired reading, tlowever, it may be that all extended notion of parallelism might solve the problem. While this has not been investigated by DSP, such an approach has been advocated in another recent account, that proposed by Priist et a1(\[23\],\[24\]). It appears that this account can accommodate the example, based on Priist et al's requirement that if a pronoun p is bound to Q in the antecedent, the corresponding pronoun p' must be bound to a &quot;structurally parallel&quot; Q~ in the target, where this is intended as matching syntactic and semantic structure. However, example (3) indicates that the two quantifiers need not be in struc- null turally parallel positions. Indeed, example (5) shows that there is no requirement for a corresponding quantifier at all.</Paragraph> <Paragraph position="8"> (3) Almost every boyl in the class hope Mary will ask hlmi out, but I know there are a few boys i who hope that she won't. \[ask him`/ out\] (4) Every boyi in Mrs. Smith's class hoped she would pass himl. In John's`/ case, 1 think she will. \[pass him`/\] Examples (1) - (4) illustrate the flexibility required in interpreting pronouns within the antecedent and target VP's. I have shown how the proposed approach permits this flexibility. None of the alternative accounts discussed can accommodate these examples.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 5 Constraints on Selecting an </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Antecedent </SectionTitle> <Paragraph position="0"> I have argued that tile current approach has significant advantages over other approaches to problem (2), concerning tile level of representation at which VP ellipsis ACRES DE COL\]NG-92, NANTf.S, 23-28 AOl~r 1992 3 0 7 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 is resolved. In addition, this approach suggests some poesible ways of v~ldreasing problem (1), concerning the selection among alternative potential antecedents.</Paragraph> <Paragraph position="1"> Since the approach parallels the treatment of pronominal anaphora, storing semantic representations of both VP's and NP's in the discourse model, a natural hypothesis is that similar constraints govern the selection of an antecedent in both the pronominal and the VP ellipsis cases.</Paragraph> <Paragraph position="2"> The problem of selecting among alternative VP antecedents has been virtually ignored in the literature. 6 The corresponding problem in pronominal anaphora has received a significant amount of attention. The Centering model (\[6\],\[7\],\[1\]) of pronominal anaphora is a leading example, applying a variety of constraints dealing with such factors as recency, salience, and attention. In addition, it is generally agreed that there are syntactic configurational constraints governing pronominal anaphora of the sort described in the &quot;Binding Theory&quot; of GB\[2\].</Paragraph> <Paragraph position="3"> For each of these types of constraints, there are interesting parallels with the case of VP ellipsis.</Paragraph> <Paragraph position="4"> Syntactlc/configuratlonah It appears that VP ellipsis obeys the &quot;precede and command&quot; constraint, as pointed out by Jackendo~8\], ruling out examples such 8.8 (5) * Charlie will, if his mother-in-law doesn't leave town.</Paragraph> <Paragraph position="5"> Recency: Just as in the pronominal case, the vast majority of cases involves an antecedent in the current or immediately preceding utterance. In a survey of VP ellipsis in the Brown Corpus\[10\], I found this to be true about 95% of the time.</Paragraph> <Paragraph position="6"> Salience: VP's in less salient positions seem to be less available as antecedents for VP ellipsis. For exampie, Halliday and Hasan\[9\] give the following example: (6) A: The policeman paid no attention to the girl who was driving the car.</Paragraph> <Paragraph position="7"> (7) *B: Was she really? Presumably the unavailability of the VP &quot;driving the car&quot; is related to the fact that it appears in a restrictive relative clause and is thus not particularly salient. Attentionah There is evidence that a &quot;center shift&quot;, i.e., shifting attention from one entity to another, might be correlated with the availability of VP antecedents. This is suggested by experimental work of Malt\[19\], who describes experiments that show, in her terms, that &quot;changing the focus&quot; in an intervening eA note by Klein ~d Stmlnton-EUiJ \[16l, points out the im- portance of problem (l), sentence tends to make a previous VP antecedent inaccessible. Thus the first example (taken from Malt's experiment) was understood more readily than the second: null (8) a. &quot;I liked the Monet exhibit,&quot; Heather remarked. null b. &quot;It was really very nice&quot;.</Paragraph> <Paragraph position="8"> c. &quot;I did too,&quot; Richard replied.</Paragraph> <Paragraph position="9"> (9) a. &quot;I liked the Monet exhibit,&quot; Heather remarked. null b. &quot;Renoir is my favorite, though.&quot; c. * &quot;I did too,&quot; Richard replied.</Paragraph> <Paragraph position="10"> There is reason to believe, then, that constraints relating to factors such as recency, syntactic configuration, salience, and attention, might apply to VP ellipsis and pronominal nnaphora in a similar way. A simplified version of these constraints is implemented in the pronoun case of the Incremental Interpretation System, and it would a simple matter to allow the same constraints to apply to VP antecedents.</Paragraph> </Section> </Section> class="xml-element"></Paper>