<?xml version="1.0" standalone="yes"?>
<Paper uid="J90-4001">
  <Title>ANAPHORA RESOLUTION IN SLOT GRAMMAR</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 VP ANAPHORA
3.1. THE RESOLUTION OF VP ANAPHORA AT
S-STRUCTURE
</SectionTitle>
    <Paragraph position="0"> Before presenting our algorithm for the interpretation of VP anaphora structures, we will provide motivation for the general view of VP anaphora that the algorithm implements. We characterize VP anaphora as a relation between the head V and selected arguments and adjuncts of a structured empty, or partially empty, elliptical VP, and the head A and corresponding adjuncts of an antecedent VP.</Paragraph>
    <Paragraph position="1"> This relation is identified on S-structure parse representations. The VP anaphora interpretation procedure copies the head A of the antecedent VP into the position of the head of the elliptical VP, and specifies which arguments and adjuncts of the antecedent A are inherited by the elliptical V.</Paragraph>
    <Paragraph position="2"> In this way, it provides an interpretation of the elliptical VP. It is important to recognize that this procedure operates on S-structure representations rather than on a more abstract level of LF.</Paragraph>
    <Paragraph position="3"> Let us briefly consider the case for an LF-based approach to VP anaphora resolution. The elliptical VP in the relative clause of the object NP in 1 is contained in the matrix VP, which is its antecedent.</Paragraph>
    <Paragraph position="4"> 1. Dulles suspected everyone who Angelton did.</Paragraph>
    <Paragraph position="5"> As May (1985) observes, if we copy the matrix VP into the position of the empty VP at S-structure, an interpretive regress results. The empty VP will reappear in the copied matrix VP. May proposes to solve this problem by applying the operation of quantifier raising (QR) to the object NP in  1. QR adjoins the quantified NP to the matrix sentence to derive the LF representation 2. 6 Computational Linguistics Volume 16, Number 4, December 1990 199 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar 2. \[W'\[NP, everyone who 1 Angelton did  \[vp\] \] \[ip Dulles \[vp suspected h\]\] \] The matrix VP of the IP in 2 is assigned to the empty VP of the adjoined NP to obtain 3, the desired interpretation of 1. 3. \[IP'\[NP, everyone who Angelton \[vp suspected h\]\] \[w Dulles \[vp suspected t l\] \] \] May concludes that antecedent-contained deletion (ACD) structures can only be interpreted by a VP copying rule that applies at LF.</Paragraph>
    <Paragraph position="6"> There are at least two serious difficulties with May's analysis of ACD structures. 7 First, as Haik (1987) points out, the wh-phrase in the relative clause of an ACD sentence such as 1 is constrained by subjacency.</Paragraph>
    <Paragraph position="7"> 4a. John read everything which Bill believes he did.</Paragraph>
    <Paragraph position="8"> b. *John read everything which Bill believes the claim that he did.</Paragraph>
    <Paragraph position="9"> On May's analysis, the VP in the relative clause in 1 and 4a-b is empty at S-structure, and the wh-phrase binds a trace only after VP copying has applied to the LF produced by the movement of the object NP. But it is generally agreed that subjacency is a condition that constrains operator-trace binding chains only at S-structure. 8 Given May's account, there is no trace at S-structure for the wh-operator to bind in 1 and 4a-b. Therefore, it is unclear how, on this analysis, subjacency can constrain wh-movement in ACD structures.</Paragraph>
    <Paragraph position="10"> May (in press) seeks to avoid this problem by suggesting that subjacency does, in fact, apply at LF. The examples in  5 indicate that this is not the case.</Paragraph>
    <Paragraph position="11"> 5a. At least one critic reviewed Mary's biography of each author.</Paragraph>
    <Paragraph position="12"> b. Who did at least one critic review \[Npa biography of t\] c. *Who did at least one critic review  &amp;quot;at least one critic.&amp;quot; On the narrow scope reading, a single critic reviewed all of Mary's biographies. When &amp;quot;each author&amp;quot; receives wide scope, there is at least one (possibly different) critic for each of Mary's biographies of an author. If we accept May's view that the scope of a quantified NP is, in part, defined in terms of the constituent to which it is adjoined by QR, the fact that &amp;quot;each author&amp;quot; can take wide scope relative to &amp;quot;at least one critic&amp;quot; indicates that QR can move the former NP out of the NP &amp;quot;Mary's biography of each author.&amp;quot; But 5b-c shows that the latter NP defines a syntactic island for wh-movement. It seems, then, that subjacency does constrain binding chains visible at S-structure, but not scope assignment.</Paragraph>
    <Paragraph position="13"> The fact that antecedent-contained VP anaphora exhibits subjacency effects strongly suggests that the elliptical VP in these cases is not necessarily empty at S-structure, but may contain a trace bound by a wh-phase (or other ope:rator).</Paragraph>
    <Paragraph position="14"> Second, May's analysis does not extend to the subdeletion variety of ACD, where arguments and adjuncts of an empty verb are realized within the partially elliptical VP that it heads, as in the sentences in 6. 9 6a. John writes more books than Bill does articles.</Paragraph>
    <Paragraph position="15"> b. The university gives more money to the library for books than the city does to the orchestra for instruments. null c. The university gives more money to the library for periodicals than it does for books.</Paragraph>
    <Paragraph position="16"> d. John wrote more articles for the journal about politics than he did about linguistics.</Paragraph>
    <Paragraph position="17"> e. John showed everything to Mary which he did to Bill. f. Mary argues about politics with everyone who she does about linguistics.</Paragraph>
    <Paragraph position="18"> g. Mary arrived in London before Sam did in New York.</Paragraph>
    <Paragraph position="19"> h. John reviewed the play for The New York Times shortly after Bill did for The Washington Post.</Paragraph>
    <Paragraph position="20"> As these examples illustrate, subdeletion occurs in a variel:y of syntactic environments, including comparative NPs (6a-d), relative clauses (6e-f), and adverbial phrases (6g-h). Given that May's analysis treats VP anaphora as a global relation between an empty VP and an antecedent VP, :it does not cover subdeletion, where an anaphoric relation holds between the head and selected constituents of an elliptical VP and its antecedent. But the full and subdeletion varieties of ACD are closely related phenomena, and an analysis that provides a unified explanation of both types of V\]? anaphora is clearly preferable to an account that handles only one type of antecedent-contained anaphora. I:~. is possible to capture the properties of ACD structures that create problems for May's LF analysis if we assume that the empty VP of a VP anaphora environment is structured, and may contain arguments or adjuncts of the head. ldeg The arguments appearing in a partially empty elliptical VP can be lexically realized, or they may be traces (or their counterparts in Slot Grammar). If we apply this analysis to 1, we obtain 7a as its S-structure.</Paragraph>
    <Paragraph position="21"> 7a. Dulles suspected \[Nr\[N,everyone \[cpwhol Angelton did \[ve\[v\] \[Nph\]\]\]\]\] b. Dulles suspected \[NP\[N,everyone \[cpwhol Angelton \[vP\[vsuspected\] \[NPh\]\] \] \] \] In 7a, the object of the head of the elliptical VP is realized as a 'trace, and the VP is interpreted by copying only the head, &amp;quot;suspected,&amp;quot; of the antecedent VP into the position of the empty head to yield 7b. Therefore, the interpretive regress is avoided without QR. Moreover, subjacency violations can be identified at S-structure by computing the relation between the wh-phrase and the trace that it binds. Our proposal handles subdeletion in a natural and straightforward way. The only difference between the subdeletion structures in 6 and the ACD structure in 7a is that 200 Computational Linguistics Volume 16, Number 4, December 1990 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar arguments and adjuncts of the elliptical verbs in 6 are lexically realized, while the object of the elliptical verb in 7a is not. The interpretation procedure is the same for the two cases. The head of the antecedent VP is copied into the head of the elliptical VP. If any arguments or adjuncts are missing in the elliptical VP and corresponding arguments or adjuncts are realized in the antecedent VP, the latter are inherited by the head of the former.</Paragraph>
    <Paragraph position="22"> This approach can be extended to intersentential VP anaphora cases like 8.11 8. John arrived yesterday, and Mary did too.</Paragraph>
    <Paragraph position="23"> We simply treat the anaphoric relation in these cases as holding between the head and constituents of a structured empty VP, and the head and counterpart constituents of a full VP in another conjunct or sentence.</Paragraph>
    <Paragraph position="24"> Subdeletion is generally marginal with intersentential VP anaphora when arguments are left in the elliptical VP. 9a. ??John writes articles, and Bill does books.</Paragraph>
    <Paragraph position="25"> b. ?Mary spoke to Max, but Sam won't to Lucy.</Paragraph>
    <Paragraph position="26"> However, subdeletion with adjuncts in these structures is considerably better.</Paragraph>
    <Paragraph position="27"> 10a. John arrived today, and Bill did yesterday.</Paragraph>
    <Paragraph position="28"> b. Max spoke after Mary, and Sam will before Lucy.</Paragraph>
    <Paragraph position="29"> The fact that adjuncts can remain in partially empty VPs in these cases provides motivation for applying the structured (partially) empty VP analysis to intersentential VP anaphora. null</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3.2. AN ALGORITHM FOR VP ANAPHORA
INTERPRETATION
</SectionTitle>
    <Paragraph position="0"> We define the predicate P is contained in Q recursively as follows. A phrase P is immediately contained in a head Q iff (i) P is an argument of Q, or (ii) P is an adjunct ofQ. P is contained in Q iff (i) P is immediately contained in Q, or (ii) P is immediately contained in a head R, and (the phrase with head) R is contained in Q. The following is a schematic description of our algorithm for VP anaphora resolution. null</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
VP ANAPHORA ALGORITHM
</SectionTitle>
    <Paragraph position="0"> A. Identify an elliptical verb-antecedent verb pair (V,A) as follows.</Paragraph>
    <Paragraph position="1"> 1. An elliptical verb V is identified by the presence of an auxiliary verb or the infinitival complementizer &amp;quot;to,&amp;quot; where the auxiliary verb or the complementizer does not have a realized verb complement. 2. A candidate A for an antecedent of V is a verb that is not elliptical and not an auxiliary verb with a realized complement.</Paragraph>
    <Paragraph position="2"> 3. Check that A and V stand in at least one of the  following relations: a. V is contained in the clausal complement of a subordinate conjunction SC, and the SC-phrase is either (i) an adjunct of A, or (ii) an adjunct of a noun N and N heads an NP argument of A, or N heads the NP argument of an adjunct of A.</Paragraph>
    <Paragraph position="3"> b. V is contained in a relative clause that modifies a head noun N, N is contained in A, and, if a verb A' is contained in A and N is contained in A', then A' is an infinitival complement of A or of a verb contained in A.</Paragraph>
    <Paragraph position="4"> c. V is contained in the right conjunct of a sentential conjunction S, and A is contained in the left conjunct of S.</Paragraph>
    <Paragraph position="5"> B. Generate a new tree in which A is substituted for V as the head of the elliptical verb phrase VP' that V heads, and A is assigned the agreement features required by the head of VP'. (We will refer to this new occurrence of A as A').</Paragraph>
    <Paragraph position="6">  C. Consider in sequence each argument slot Sloti in the argument frame of A.</Paragraph>
    <Paragraph position="7"> 1. If Slot i is filled by a phrase C, then a. If there is a phrase C' in VP' that is of the appropriate type for filling Slot i, then fill Slot i in the argument frame of A' with the marker variable of C'. Else, b. Fill Slot i in A' with the marker variable of C, and list C as a new argument of A'.</Paragraph>
    <Paragraph position="8"> 2. If Slot i is empty in the frame of A, it remains empty in the frame of A'.</Paragraph>
    <Paragraph position="9"> D. For each adjunct Adj of A, if there is no adjunct of the same type as Adj in VP', then list Adj as a new adjunct of A'.</Paragraph>
    <Paragraph position="10"> Part A of the algorithm specifies the procedures for identifying pairs whose first element is the head of an antecedent VP and whose second element is the head of an elliptical VP. Elliptical VPs are identified by the presence  of a bare auxiliary verb or the bare complementizer &amp;quot;to.&amp;quot; In fact, for reasons of convenience, we take bare auxiliaries and bare complementizers as standing for the head of an elliptical VP, and so the algorithm treats them as surrogate VP heads. A.3 defines the structural relations that hold between the head of an elliptical VP and the head of a possible antecedent VP. 12 Part B describes the operation of generating a new interpreted VP anaphora tree in which the head of the antecedent VP is substituted for the head of the elliptical VP, and the features of the new head of the interpreted elliptical VP are adjusted in accordance with the requirements of this VP.</Paragraph>
    <Paragraph position="11"> Part C characterizes a procedure for filtering the arguments of the antecedent verb to determine which of them are inherited by the head of the interpreted elliptical VP. Similarly, Part D describes the filtering process that gives the adjuncts of the antecedent verb that are inherited by the interpreted elliptical verb. The combination of the new verb heading the elliptical VP and the lists of arguments and adjuncts it inherits from the antecedent verb provide the interpretation of the elliptical VP.</Paragraph>
    <Paragraph position="12"> To illustrate the Prolog implementation of this algorithm Computational Linguistics Volume 16, Number 4, December 1990 201 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar on the basis of the network representation, we will give the clauses pertinent to A.3.b, which identifies the case in which the elliptical verb V is contained in a relative clause. The top-level predicate for testing that A is an antecedent of an elliptical verb V (used for implementing A.3) is anaph(A,V). The clauses for this predicate relevant to</Paragraph>
    <Paragraph position="14"> (Here A =/B means &amp;quot;A is not equal to B&amp;quot; in IBM Prolog.) The relation relcont(V,N) holds if V is contained in a relative clause adjunct of noun N. The predicate pharg (P,Q), which says that P is an argument of Q, was defined in Section 2 in terms of the network representation. The relation phadjunet(P,Q) says that P is an adjunct of Q, and is also straightforwardly defined in terms of the network.</Paragraph>
    <Paragraph position="15"> Let us consider several examples of the VP algorithm's results. The system produces the following output. For each ESG analysis of the input sentence, the parse tree is displayed, and then all pairs (antecedent verb, elliptical verb) found by the algorithm are displayed. Then, for each such pair, the following three things are displayed: (1) the new arguments inherited by the interpreted elliptical verb from its antecedent; (2) the new adjuncts inherited by the interpreted elliptical verb from its antecedent; and (3) the interpreted VP anaphora tree, as a modification of the original parse tree.</Paragraph>
    <Paragraph position="16">  1 \] in Figure 2 shows the output for May's ACD example 1. In the parse tree, variable X5 in the object slot of the complement frame for the auxiliary &amp;quot;did&amp;quot; unifies with the  phrase marker of the head of the relative &amp;quot;everyone&amp;quot; (and that of the wh-phrase &amp;quot;who&amp;quot;). Moreover, this variable is marked as a trace in the internal representation of the phrase structure from which the tree is projected. (Such a trace is marked in the feature structure of the verb--although it is not shown in the following abbreviated display.) Hence, the parse tree corresponds to the S-structure given in 7a. The VP anaphora algorithm identifies &amp;quot;suspected&amp;quot; as the antecedent of the elliptical verb (represented by the auxiliary), and substitutes it for the auxiliary in the interpreted VP anaphora tree. No arguments or adjuncts are inherited from the antecedent verb. The interpretecl VP anaphora tree in 11 is the SG counterpart of the interpreted S-structure 7b.</Paragraph>
    <Paragraph position="17"> A similar ACD case involving an elliptical VP that follows a bare occurrence of the complementizer &amp;quot;to&amp;quot; is given in 12 in Figure 3. Here, the interpreted verb &amp;quot;write&amp;quot; inherits the object argument &amp;quot;notes,&amp;quot; while its indirect object argument is a trace bound by the wh-phrase corresponding to the head of the relative clause.</Paragraph>
    <Paragraph position="18"> In the following examples, we give the output in an abbreviated way in order to save space. The uninterpreted tree is not shown, and the analysis resulting from the ii. Dulles suspected everyone who Angelton did.</Paragraph>
    <Paragraph position="19"> Syntactic analysis time = 87 msec.</Paragraph>
    <Paragraph position="21"/>
    <Paragraph position="23"> algorithm is shown in an abbreviated linear form, consisting simply of head words and their arguments.</Paragraph>
    <Paragraph position="24"> In 13, both &amp;quot;promise&amp;quot; and its infinitival complement &amp;quot;read&amp;quot; satisfy the condition given in A.3.b on the antecedent of an elliptical verb in a relative clause. Therefore, the algorithm correctly generates two possible interpreted VP anaphora trees. Tree 1 gives the reading on which &amp;quot;promise&amp;quot; is taken as the antecedent of the head of the empty VP in the relative clause, and the infinitival clause headed by &amp;quot;read&amp;quot; is inherited as a new argument of the interpreted verb. Tree 2 specifies the interpretation where &amp;quot;read&amp;quot; is substituted for the head of the empty verb, and the trace of the relative operator &amp;quot;which&amp;quot; is its (noninherited) argument. The algorithm correctly excludes &amp;quot;said&amp;quot; as a possible antecedent of the empty verb in the relative clause in 14.</Paragraph>
    <Paragraph position="25"> This is because it has a tensed rather than an infinitival complement that contains the verb that contains the noun modified by the relative clause. Therefore, the algorithm produces only two possible interpretations for 14, which are represented by the two interpreted VP anaphora trees that it generates.</Paragraph>
    <Paragraph position="26"> 13. John promised to read everything which Mary did.</Paragraph>
    <Paragraph position="28"/>
    <Paragraph position="30"> 15 is a subdeletion case in which all of the arguments of the elliptical verb are filled locally within the elliptical VP.</Paragraph>
    <Paragraph position="31"> Hence, the algorithm substitutes the antecedent verb &amp;quot;write&amp;quot; for the auxiliary, and fills the direct and indirect argument slots in its frames with &amp;quot;notes&amp;quot; and &amp;quot;to Bill,&amp;quot; respectively. It should be pointed out that the algorithm corrects ESG's parse of &amp;quot;to Bill&amp;quot; as a PP adjunct of &amp;quot;notes&amp;quot; in the original tree. This misparse is due to the fact that &amp;quot;do&amp;quot; does not allow an indirect object in its frame. The algorithm recognizes this adjunct as a possible filler for the indirect object slot in the frame of &amp;quot;write,&amp;quot; and uses it to fill the slot when &amp;quot;write&amp;quot; is substituted for &amp;quot;do&amp;quot; in the interpreted tree.</Paragraph>
    <Paragraph position="32"> 15. Max writes more letters to Sam than Mary does notes to Bill.</Paragraph>
    <Paragraph position="34"> In 16, the indirect object slot of the interpreted verb &amp;quot;write&amp;quot; is filled locally by &amp;quot;to Bill,&amp;quot; but &amp;quot;letters&amp;quot; is inherited from the antecedent. The algorithm also corrects the misparse of &amp;quot;to Bill&amp;quot; as a PP zdjunct of the elliptical verb in the original tree by a strategy similar to the one used to correct the parse of the PP in 15 16. Max writes more letters to Sam than Mary does to</Paragraph>
    <Paragraph position="36"> 18 and 19 show the operation of argument filtering in an ACID passive case. 13 As in 13 and 14, the algorithm corrects the misparse, in the original tree, of the second &amp;quot;by&amp;quot; phrase as a PP adjunct. It raises it to the status of the agent (deep subject) argument of the head of the new verb in the interpreted tree.</Paragraph>
    <Paragraph position="37"> 18. John was interviewed by Bill before Mary could have been by Max.</Paragraph>
    <Paragraph position="39"> 20 and 21 illustrate the adjunct filtering procedure of the algorithm in an intersentential case of VP anaphora. The adverbial &amp;quot;yesterday&amp;quot; is inherited in 20, but not in 21.</Paragraph>
    <Paragraph position="40"> 20. John arrived yesterday, and Mary did too.</Paragraph>
    <Paragraph position="42"> 22 and 23 exhibit the effects of adjunct filtering of a more complex variety in an ACD structure. Both the adverb &amp;quot;briefly&amp;quot; and the PP adjunct &amp;quot;with Bill&amp;quot; are inherited by the interpreted verb &amp;quot;walked&amp;quot; in 22, but only &amp;quot;briefly&amp;quot; is inherited in 23.</Paragraph>
    <Paragraph position="43"> 22. John walked briefly with Bill before Mary did.</Paragraph>
    <Paragraph position="45"/>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 A SYNTACTIC FILTER ON PRONOMINAL
ANAPHORA
</SectionTitle>
    <Paragraph position="0"> The filter consists of six conditions for NP-pronoun non-coreference within a sentence. To state these conditions, we use the following terminology. The agreement features of an NP are its number, person, and gender features. We will say that a phrase P is in the argument domain of a phrase N iff P and N are both arguments of the same head. We will say that P is in the adjunct domain of N iff N is an argument of a head H, P is the object of a preposition PREP, and PREP is an adjunct of H. P is the NP domain of N iff N is the determiner of a noun Q and (i) P is an argument of Q, or (ii) P is the object of a preposition PREP and PREP is an adjunct of Q.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 FILTER ON PRONOMINAL ANAPHORA
</SectionTitle>
      <Paragraph position="0"> A pronoun P is noncoreferential with a (nonreflexive or nonreciprocal) noun phrase N if any of the following conditions hold.</Paragraph>
      <Paragraph position="1"> I. P and N have incompatible agreement features. II. P is in the argument domain of N.</Paragraph>
      <Paragraph position="2"> III. P is in the adjunct domain of N.</Paragraph>
      <Paragraph position="3"> IV. P is an argument of a head H, N is not a pronoun, and N is contained in H.</Paragraph>
      <Paragraph position="4"> V. P is in the NP domain of N.</Paragraph>
      <Paragraph position="5"> VI. P is a determiner of a noun Q, and N is contained in Q. Condition I rules out coreference between a pronoun and an NP with incompatible agreement features. It will identify the co-indexed expressions in 24a-c as noncoreferential. null 24a. *He i said that they i came.</Paragraph>
      <Paragraph position="6"> b. *The woman i said that he i is funny.</Paragraph>
      <Paragraph position="7"> c. *I i believe that she i is competent.</Paragraph>
      <Paragraph position="8"> The filter treats (&amp;quot;he,&amp;quot; &amp;quot;they&amp;quot;) as a noncoreferring pair, which entails only that the intended denotation of &amp;quot;he&amp;quot; cannot be taken as identical to that of &amp;quot;they.&amp;quot; The referent of &amp;quot;he&amp;quot; can, of course, be a part of the referent of &amp;quot;they,&amp;quot; and, in appropriate contexts, a discourse interpretation system, like the LODUS system of Bernth (1988, 1989), should be able to recognize this possibility.</Paragraph>
      <Paragraph position="9"> Condition II covers cases in which a pronoun and an NP are arguments of the same head, and so it rules out coreference between the coindexed expressions in 25a-d.  25a. *Mary i likes her i.</Paragraph>
      <Paragraph position="10"> b. *She i likes her i.</Paragraph>
      <Paragraph position="11"> c. *John i seems to want to see him i.</Paragraph>
      <Paragraph position="12"> d. *This is the girl i Mary said she i saw.</Paragraph>
      <Paragraph position="13">  It is important to note that the conditions of the filter apply to pronouns and NPs regardless of whether they are lexically realized in argument position (25a-b), or bind the argument slots which they fill in their heads at a distance through control (25c) and unbounded dependency relations (25d). This is due to the fact that the variable that fills an argument slot is unified with the phrase marker of the head of the phrase to which it corresponds. Therefore, it is not necessary to incorporate empty categories such as traces and PRO into the parse output, and compute appropriate binding chains for these categories in order for the algorithm to handle noncoreference in cases involving control and wh-movement. Mechanisms of this kind are required for implementations of Chomsky's binding theory in Government Binding-based parsers, such as those described in Correa (1988) and Ingria and Stallard (1989).</Paragraph>
      <Paragraph position="14"> Computational Linguistics Volume 16, Number 4, December 1990 205 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar Condition III rules out coreference between an argument of a verb V and the object of a prepositional adjunct of V, as in 26a-bJ 4 26a. *Mary i arrived with her i.</Paragraph>
      <Paragraph position="15"> b. *Who i did John say wants to sit near himi? Condition IV prevents coreference between a pronoun that is an argument of a head H, and a nonpronominal NP contained in H, as in 27a-c.</Paragraph>
      <Paragraph position="16"> 27a. *Who i did she i say Mary i kissed? b. *This is the man i he i said Max i wrote about.</Paragraph>
      <Paragraph position="17"> c. *He i likes Johni's mother.</Paragraph>
      <Paragraph position="18"> The filter does permit coreference in 28a-b. &amp;quot;His&amp;quot; in 28a is not an argument of &amp;quot;likes,&amp;quot; and so (&amp;quot;his,&amp;quot; &amp;quot;John&amp;quot;) do not satisfy Condition IV (or any other condition of the filter). An ordering constraint attached to Condition IV requires that a possessive adjunct of a noun contained in a head H follow a pronominal argument of H for this condition to apply to the pair.</Paragraph>
      <Paragraph position="19"> 28a. His i mother likes John i.</Paragraph>
      <Paragraph position="20"> b. Johni's mother likes him i.</Paragraph>
      <Paragraph position="21"> Finally, V and VI in effect apply conditions II and III, respectively, to NP internal cases. They prevent coreference in 29a-c, while allowing it in 29d.</Paragraph>
      <Paragraph position="22"> 29a. *His i portrait of John i is interesting.</Paragraph>
      <Paragraph position="23"> b. *Johni's portrait of him i is interesting.</Paragraph>
      <Paragraph position="24"> c. *His i description of the portrait by John i is interesting. null d. Johni's description of the portrait by him i is interesting. null The filter on pronominal anaphora restricts the search space that a discourse system of anaphora resolution must consider. Bernth has integrated the filter into LODUS (Bernth 1988, 1989), which resolves pronominal anaphora and NP denotation through semantic and pragmatic rules of inference. The anaphora resolution component of LOI)US applies only to the pronoun-NP pairs that the syntactic filter has not identified as noncoreferential. An example of the filter algorithm's output is given in 30 in Figure 4. The list of noncoreferential pronoun-NP pairs appears after the parse tree. Ulrike Schwall has successfully implemented the algorithm in German Slot Grammar 30. This is the girl who she wanted Mary to talk to.</Paragraph>
      <Paragraph position="26"> 31. paul gab Peter das Buch, um ihn zu beeindrucken. (Paul gave Peter the book to impress him.) subj Paul(X2) .-- to geb(Xl,X2,X3,X4,u)</Paragraph>
      <Paragraph position="28"/>
    </Section>
  </Section>
  <Section position="8" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 AN ANAPHOR-BINDING ALGORITHM
</SectionTitle>
    <Paragraph position="0"> We take the set of anaphors to include reflexive pronouns and the reciprocal NP &amp;quot;each other.&amp;quot; The notion higher argument slot used in the formulation of the algorithm is defined by the hierarchy of argument slots given in 32.</Paragraph>
    <Paragraph position="1"> 32. subj &gt; agent &gt; obj &gt; (iobj I pobj) subj is the surface subject slot, agent is the deep subject slot of a verb heading a passive VP, obj is the direct object slot, iobj is the indirect object slot, and pobj is the object of a PP complement of a verb, as in &amp;quot;put NP on NP.&amp;quot; We assume the definitions of argument domain, adjunct domain, and NP domain given in Section 4.</Paragraph>
  </Section>
  <Section position="9" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5.1 ANAPHOR-BINDING ALGORITHM
</SectionTitle>
    <Paragraph position="0"> A noun phrase N is a possible antecedent binder for an anaphor A iff N and A do not have incompatible agreement features, and one of the following five conditions holds.</Paragraph>
    <Paragraph position="1"> I. A is in the argument domain of N, and N fills a higher argument slot than A.</Paragraph>
    <Paragraph position="2"> II. A is in the adjunct domain of N.</Paragraph>
    <Paragraph position="3"> III. A is in the NP domain of N.</Paragraph>
    <Paragraph position="4"> IV. N is an argument of a verb V, there is an NP Q in the argument domain or the adjunct domain of N such that Q has no noun determiner, and A is (i) an argument of Q, or (ii) A is an argument of a preposition PREP and PREP is an adjunct of Q.</Paragraph>
    <Paragraph position="5"> V. A is a determiner of a noun Q, and (i) Q is in the argument domain of N and N fills a higher argument slot than Q, or (ii) Q is in the adjunct domain of N. Conditions I and II cover anaphoric binding in cases like 33-34, respectively.</Paragraph>
    <Paragraph position="6"> 33a. They i wanted to see themselvesi.</Paragraph>
    <Paragraph position="7"> b. Mary knows the people i who John introduced to each other i.</Paragraph>
    <Paragraph position="8"> 34a. He i worked by himself i.</Paragraph>
    <Paragraph position="9"> b. Which friends i plan to travel with each otheri? Condition III handles binding of an anaphor inside an NP by the determiner of the NP, as in 35, and IV deals with NP internal anaphors which are bound from outside of the NP, as in 36.</Paragraph>
    <Paragraph position="10"> 35. John liked Billi's portrait of himself i.</Paragraph>
    <Paragraph position="11"> 36. They i told stories about themselves i.</Paragraph>
    <Paragraph position="12"> Condition V applies to cases in which a reciprocal determiner is bound by an argument in the same clause as the NP containing the reciprocal. 37 is an example of this binding relation, and 38 illustrates the combined effect of IV and V.</Paragraph>
    <Paragraph position="13"> 37. \[John and Mary\]i like each otheri's portraits. 38. \[John and Mary\]i like each otheri's portraits of themselves i.</Paragraph>
    <Paragraph position="14"> An example of the anaphor binding algorithm's output is presented in 39 in Figure 5. Notice that the sentence in this example is ambiguous concerning antecedents for &amp;quot;himself,&amp;quot; and the algorithm correctly identifies both &amp;quot;who&amp;quot; and &amp;quot;John&amp;quot; as possible binders of the reflexive. When a discourse interpretation system makes use of this algorithm, it must, of course, constrain the interpetation of anaphors by requiring that exactly one binding pair be selected from the list of pairs that the algorithm provides for any given anaphor, relative to the clause in which it appears.</Paragraph>
    <Paragraph position="15"> Ulrike Schwall has implemented the algorithm in GSG, and 40 in Figure 5 illustrates its output for a German sentence.</Paragraph>
  </Section>
  <Section position="10" start_page="0" end_page="0" type="metho">
    <SectionTitle>
6 AN INTEGRATED SYSTEM FOR ANAPHORA
RESOLUTION
</SectionTitle>
    <Paragraph position="0"> Any two of the algorithms described in Sections 4-6 can operate in conjunction with each other. Examples of the results provided by such combinations are given in 41-44 (see Figure 6 for 41). In 45, both the filter and anaphor binding algorithms have been integrated into the VP anaphora algorithm, and operate on the interpreted VP anaphora tree it generates.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.1 VP ANAPHORA ALGORITHM WITH
</SectionTitle>
      <Paragraph position="0"/>
    </Section>
  </Section>
  <Section position="11" start_page="0" end_page="0" type="metho">
    <SectionTitle>
PRONOMINAL ANAPHORA FILTER
</SectionTitle>
    <Paragraph position="0"> 42. John talked to him, and Bill did too.</Paragraph>
    <Paragraph position="2"> Noncoreferential pronoun-NP pairs: th.ey. 1 - portrait. 5, they. 1 - John.9, they. 1 - coord(and, John, Mary). 10, they. 1 - Mary. 11 Antecedent NP-anaphor pairs: they. 1 - (each.other).3, (each.other).3 - themselves.7, coord(and, John, Mary). 10 - (each.other).3 Our integrated system for anaphora resolution is syntactically based, and it must be supplemented by additional semantic procedures to yield fully adequate interpretations of e!iliptical VP structures. This can be seen quite clearly by considering the interpreted VP anaphora tree of 42. Here &amp;quot;him,,&amp;quot; the indirect object of &amp;quot;talk,&amp;quot; is inherited by the interpreted verb in the second conjunct, and its marker  variable X10 is unified with the marker of the indirect object slot in the argument frame of this verb. Therefore, the interpreted VP anaphora tree correctly represents the fact that the second conjunct in 42 must be understood as asserting that Bill spoke to the same person as John did. The list of noncoreferential pronoun-NP pairs specifies that &amp;quot;him&amp;quot; is distinct in reference from both &amp;quot;John&amp;quot; and &amp;quot;Bill.&amp;quot; However, in its present form, the VP anaphora algorithm unifies the marker variables of all inherited arguments with the appropriate slots in the frame of an interpreted verb. This will yield incorrect results for a sentence like 46, where &amp;quot;a book&amp;quot; is the inherited argument.</Paragraph>
    <Paragraph position="3"> 46. John read a book, and Mary did too.</Paragraph>
    <Paragraph position="4"> On at least one possible reading of the sentence, John and Mary read distinct books. To complete the interpretation of elliptical VPs, it will be necessary to add procedures for substituting new marker variables for the occurrence of inherited arguments and adjuncts in the interpreted VP, when these expressions need not be taken as having the same denotations that they receive as arguments and adjuncts of the antecedent verb.</Paragraph>
    <Paragraph position="5"> A related problem concerns scope assignment in sentences like 47.</Paragraph>
    <Paragraph position="6"> 47. Mary spoke to everyone after Max did.</Paragraph>
    <Paragraph position="7"> Dalrymple, Shieber, and Pereira (1990) point out that 47 is ambiguous between a narrow scope reading on which Mary spoke to everyone after Max spoke to everyone, and a wide scope reading according to which everyone is such that Mary spoke to him/her after Max spoke to him/her. At this point, the VP anaphora algorithm generates only the former reading, as &amp;quot;everyone&amp;quot; is inherited as an argument by the interpreted head of the ellided VP.</Paragraph>
    <Paragraph position="8"> We could capture the wide scope reading by modifying our S-structure copying analysis of VP anaphora to allow copying to apply to more abstract semantic representations. This approach involves adopting an interpolated copying theory of VP ellipses on which copying is permitted not only at the level of S-structure, but also after the antecedent clause has been assigned a partial or full semantic interpretation. In the case of 47, copying could apply after &amp;quot;everyone&amp;quot; has been assigned scope through the operation of NP storage and a semantic variable appears in its place) 5 The result of such copying would be an interpretation on which &amp;quot;everyone&amp;quot; would have wide scope by virtue of the fact that it binds variables in both the antecedent and interpreted VPs. The interpolated copying analysis could be implemented within Slot Grammar by permitting either expressions or simply their marker variables to be inherited. The former case corresponds to S-structure copying of a constituent, the latter to copying at a level of representation to which interpretation has already applied. If &amp;quot;everyone&amp;quot; is inherited in 47, it is, in effect, copied, and the narrow scope reading of the sentence results. When only its marker variable is inherited, the semantic variable within its scope is copied, which yields the wide scope interpretation. 16 43 is particularly interesting. The sentence is a variant of an example that May (in press) claims provides evidence for this QR treatment of ACD structures. He maintains that only after QR has been applied to &amp;quot;everyone who he did&amp;quot; and the matrix VP &amp;quot;sent John to t&amp;quot; copied into the empty VP in the relative clause, can Principle C of Chomsky's binding theory rule out coreference between &amp;quot;he&amp;quot; and &amp;quot;John.&amp;quot; In fact, the application of our filter to the interpreted VP anaphora tree provides the correct results for this case. This is due to the fact that the VP algorithm identifies &amp;quot;John&amp;quot; as the inherited object of the verb that it Computational Linguistics Volume 16, Number 4, December 1990 209 Shalom Lappin and Michael McCord Anaphora Resolution in Slot Grammar substitutes for the elliptical verb in the new tree. Condition II of the filter algorithm is then satisfied by (&amp;quot;he,&amp;quot; &amp;quot;John&amp;quot;). z7 This example provides strong support for our treatment of VP anaphora.</Paragraph>
    <Paragraph position="9"> In the interpreted VP anaphora tree of 44, the substituted verb &amp;quot;writes&amp;quot; inherits &amp;quot;herself&amp;quot; as a new argument, and so we capture the &amp;quot;sloppy&amp;quot; reading of this sentence, on which each occurrence of the reflexive is bound by the subject of the clause in which it occurs. To obtain the &amp;quot;strict&amp;quot; interpretation, according to which &amp;quot;herself&amp;quot; is bound only by &amp;quot;the girl,&amp;quot; it will be necessary to allow copying of the marker variable associated with &amp;quot;herself,&amp;quot; in the manner required for the wide scope reading of &amp;quot;everyone&amp;quot; in 47 (see footnote 16).</Paragraph>
    <Paragraph position="10"> The fully integrated algorithm provides the desired results for the sentence in 45. &amp;quot;They&amp;quot; is identified as noncoreferential with any of the NPs contained in the head of which it is the subject. &amp;quot;They&amp;quot; binds &amp;quot;each other&amp;quot; in the matrix clause, and &amp;quot;John and Mary&amp;quot; binds &amp;quot;each other&amp;quot; as the determiner of the inherited argument &amp;quot;portrait&amp;quot; in the adverbial phrase. &amp;quot;Each other&amp;quot; binds &amp;quot;themselves,&amp;quot; and so, by transitivity of binding, &amp;quot;they&amp;quot; binds &amp;quot;themselves&amp;quot; in its occurrence in the object NP headed by &amp;quot;portrait&amp;quot; in the matrix clause, and &amp;quot;John and Mary&amp;quot; binds &amp;quot;themselves&amp;quot; in its occurrence in this NP as the inherited object of the substituted verb in the adverbial phrase.</Paragraph>
  </Section>
class="xml-element"></Paper>