<?xml version="1.0" standalone="yes"?>
<Paper uid="J90-4001">
  <Title>ANAPHORA RESOLUTION IN SLOT GRAMMAR</Title>
  <Section position="13" start_page="0" end_page="0" type="ackno">
    <SectionTitle>
NOTES
1. Earlier versions of the paper were presented to the SRI natural
</SectionTitle>
    <Paragraph position="0"> language group at Menlo Park, CA in June, 1990, and to the AT&amp;T Bell Laboratories natural language and speech generation group at Murray Hill, NJ in July, 1990. We are grateful to the participants of these two forums for their comments. We thank Mori Rimon for detailed and useful comments on an earlier version of the paper. We also very much appreciate the careful reading of the paper and the suggestions of three anonymous referees. We would particularly like to express our thanks to Fernando Pereira and Mary Dalrymple for extended discussion of this paper and the problems involved in VP anaphora resolution. Their own work in this area has provided us with considerable stimulation and insight.</Paragraph>
    <Paragraph position="1"> 2. See Reinhart (1976, 1981, 1984), and Chomsky (1981, 1986b) for alternative definitions of c-command, and discussions of the role of c-command in determining the possibilities for anaphora. See Lappin and McCord (1990) for comparisons between the pronominal anaphora filter in Slot Grammar and recent implementations of Chomsky's binding theory in GB-based parsers.</Paragraph>
    <Paragraph position="2"> 3. Shortly after we designed and implemented these three algorithms in Slot Grammar, Karen Jensen constructed three alternative procedures for anaphora resolution in the PEG grammar (see Jensen 1986 for a general description of PEG). Moreover, George Heidorn has implemented a version of our filter on pronominal anaphora in PEG.</Paragraph>
    <Paragraph position="3"> Jensen's procedures and Heidorn's implementation of our filter algorithm rely on and apply after a set of socond-pass operations that comprise a module referred to as PEGASUS. This module computes deep grammatical roles from the surface configurational structures constituting the PEG parse. (See Jensen and Heidorn 1990 for a brief description of PEGASUS and an outline of Jensen's anaphora resolution procedures.) By contrast, in Slot Grammar deep grammatical roles are obtained directly in the course of parsing, through the unification of complement &amp;quot;marker variables&amp;quot; with variables in the argument frames of their heads. While PEGASUS reconstructs deep grammatical role information (primarily) from surface configurational relations, the representation of these roles in Slot Grammar is lexically driven and is an integral part of the parsing process. Therefore, where Jensen's anaphora resolution procedures operate on the output of a second-pass module (they are, in effect, third-pass rules), our algorithms are formulated in terms of the head-complement structures provided directly by the Slot Grammar parser. See McCord (1984) and Lappin et al. (1989) for earlier systems that compute deep grammatical roles from PEG's surface parse structures.</Paragraph>
    <Paragraph position="4"> 4. The list of complement slots in the argument frame of a verb includes its subject. Therefore, SG represents argument structure in a manner analogous to that of LFG in that it makes no structural distinction between the subject as an external argument of a VP and the internal arguments of the verb, as does Government Binding theory.</Paragraph>
    <Paragraph position="5"> 5. The distinction between slot filler rules and ordering constraints parallels the difference between immediate dominance rules and linear precedence rules in GPSG. See Gazdar et al. (1985) for a characterization of ID and LP rules in GPSG. See McCord (1989b) for more discussion of the relation of Slot Grammar to other systems.</Paragraph>
    <Paragraph position="6"> 6. IP is an inflectional phrase, the category to which sentences correspond in current versions of X' theory. See Chomsky (1986b) for details of the IP analysis of sentences.</Paragraph>
    <Paragraph position="7"> 7. May's QR-based analysis of VP anaphora extends several of the ideas concerning the interaction of quantified NPs and VP anaphora originally proposed in Sag (1976). Webber (1978) adopts and modifies Sag's approach to VP anaphora within a computationally oriented framework. See Lappin (1984) for discussion of some of the difficulties that arise with Sag's original analysis. Lappin (in press) presents more detailed criticism of May's account, and of a variant of this analysis proposed in Haik (1987). This paper also deals with Larson's (1987, 1988) extension of May's account in ACD structures in adverbial phrases. Other treatments of VP anaphora are discussed, and motivation is given for the S-structure interpretation view adopted here. In the following, we limit ourselves to a brief presentation of two main arguments against the LF approach to VP anaphora resolution, and a summary of the S-structure alternative that we propose.</Paragraph>
    <Paragraph position="8"> 8. See Chomsky 1981 and 1986b for formulations of subjacency and arguments to the effect that it is an S-structure constraint. Haik presents an LF analysis of VP anaphora that classifies an empty antecedent-contained VP as a variable bound by a wh-(or empty) operator at S-structure. While Haik's account permits subjacency to constrain ACD structures, it requires that the VP variable be reanalyzed as an NP trace at LF in order to obtain a structure like 3. This is an ad hoc and otherwise unmotivated device. See Lappin (in press) for more detailed discussion of Haik's proposal.</Paragraph>
    <Paragraph position="9"> 9. See Bresnan (1975) and Chomsky (1977) for the classical discussion of subdeletion.</Paragraph>
    <Paragraph position="10"> 10. The idea that empty VPs are structured was initially proposed in Wasow (1972) and adopted in Williams (1977).</Paragraph>
    <Paragraph position="11"> 11. See Lappin (1984) and the references cited there for discussions of intersentential VP anaphora.</Paragraph>
    <Paragraph position="12"> 12. The VP anaphora algorithm identifies an elided VP by the presence of a bare auxiliary or infinitival complementizer. Therefore, it will not deal with elided VPs that are not introduced by auxiliaries or the complementizer &amp;quot;to,&amp;quot; as in (i)a-b.</Paragraph>
    <Paragraph position="13"> (i)a. John wrote more papers than Mary.</Paragraph>
    <Paragraph position="14"> b. Bill arrived before Lucy.</Paragraph>
    <Paragraph position="15"> Complex syntactic and semantic factors must be invoked to distin-Computational Linguistics Volume 16, Number 4, December 1990 211 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar 13.</Paragraph>
    <Paragraph position="16"> 14.</Paragraph>
    <Paragraph position="17"> guish &amp;quot;bare&amp;quot; VP ellipsis cases of this sort from structurally similar sentences that do not contain elided VPs, such as (ii)a-b.</Paragraph>
    <Paragraph position="18"> (ii)a. John gave more papers than books to Mary.</Paragraph>
    <Paragraph position="19"> b. Bill arrived before the beginning of the concert.</Paragraph>
    <Paragraph position="20"> Extending the algorithm to cover &amp;quot;bare&amp;quot; VP ellipsis is clearly a nontrivial task, which is beyond the scope of our current work. We hope to take up this matter in future research.</Paragraph>
    <Paragraph position="21"> Unlike the examples where the elliptical verb is identified by the auxiliary &amp;quot;do,&amp;quot; the elliptical verb in 19 is represented by the auxiliary &amp;quot;be&amp;quot; rather than the antecedent verb in the list of elliptical verb-new argument pairs. This is due to the fact that with auxiliaries other than &amp;quot;do&amp;quot; the algorithm copies the elliptical verb immediately after the auxiliary in the interpreted tree, while in the case of &amp;quot;do&amp;quot; the antecedent is substituted for the auxiliary. The filtering component of the algorithm applies after the antecedent has been inserted into the new tree, and so it identifies the antecedent with the elliptical verb for purposes of listing new arguments and adjuncts when the antecedent replaces the auxiliary, but not otherwise.</Paragraph>
    <Paragraph position="22"> Unfortunately, Condition III also incorrectly blocks coreference in cases like (i)a-b, (discussed in, for example, Reinhart (1981)) when &amp;quot;near him&amp;quot; is taken as an adverb modifying the head verb &amp;quot;saw.&amp;quot; (i)a. Dani saw a snake near him i.</Paragraph>
    <Paragraph position="23"> b. Near him~, Dan i saw a snake.</Paragraph>
    <Paragraph position="24"> The problem with the cases of this kind is that the possibilities for pronominal coreference are notoriously susceptible to lexical variation, as indicated by (ii)-(iii).</Paragraph>
    <Paragraph position="25"> (ii)a. John~ took a book with him~.</Paragraph>
    <Paragraph position="26"> b. *John i took a walk with him~.</Paragraph>
    <Paragraph position="27"> (iii)a. Maryi heard music near herl.</Paragraph>
    <Paragraph position="28"> b. ??Mary~ played music near her~.</Paragraph>
    <Paragraph position="29"> A variety of syntactic and lexical semantic factors seem to be involved in determining the possibility of coreference in these cases. However, when no complement intervenes between the subject of a verb and the pronominal object of its PP adjunct, coreference is always excluded. In light of this fact and the lexically governed complexity of the coreference patterns in structures like (i)-(iii), we h~tve decided to retain Condition III in its present general form. Clearly, it would be desirable to refine it to allow for the distinctions illustrated in the examples given here.</Paragraph>
    <Paragraph position="30"> 15. Dalrymple, Shieber, and Pereira (1990) obtain both scope readings fi~r 47 by permitting interaction between storage and an equational procedure for ellipsis resolution. This procedure involves solving an equation in which the interpretation of the source clause appears on one side and a predication containing a higher order property wxiable corresponding to the ellided VP is on the other. The representation of the source clause in the equation can contain either a released or an unreleased stored NP interpretation. Lappin (1984) employs an analogous free interaction between NP storage and interpretation to derive both the globally wide and locally (conjunct) wide readings of &amp;quot;many windows&amp;quot; in (i). Sentences of this kind were originally discussed in Hirshbuhler (1982).</Paragraph>
    <Paragraph position="31"> (i) A Canadian flag was hanging in front of many windows, and an American flag was too.</Paragraph>
    <Paragraph position="32"> 16. It will be necessary to constrain copying at all levels by the filter on pronominal anaphora, as the possibilities for pronominal coreference in ellided VPs are restricted by the conditions that the filter implements. Thus, even if only the marker variable of a pronoun is inherited by an interpreted verb, it will be identified as the (denotation) marker variable of a pronoun, and subject to the pronominal coreference filter within the VP that the interpreted verb heads.</Paragraph>
    <Paragraph position="33"> Similarly, copying of reciprocal NPs must be restricted by the anaphor binding algorithm at every level of copying. The situation with respect to reflexives is less clear, given the possibility of strict as well as sloppy readings for reflexives in the interpretation of anaphoric VPs.</Paragraph>
    <Paragraph position="34"> 17. The variable of the wh-phrase (which unifies with that of the head noun of the relative clause) fills the object slot in the frame of the elliptical (auxiliary) verb in the original tree. It is correctly reanalyzed as filling the indirect object slot of the substituted verb.</Paragraph>
  </Section>
class="xml-element"></Paper>