<?xml version="1.0" standalone="yes"?>
<Paper uid="P89-1032">
  <Title>A COMPUTATIONAL MECHANISM FOR PRONOMINAL REFERENCE</Title>
  <Section position="5" start_page="266" end_page="267" type="metho">
    <SectionTitle>
4. INTERACTION WITH SEMANTIC
INTERPRETATION
</SectionTitle>
    <Paragraph position="0"> Syntactic constraints will not always identify just one allowable referent for a pronoun. Consider (22): (22) The committee awarded the prize to itself.</Paragraph>
    <Paragraph position="1"> Syntactically, &amp;quot;itself&amp;quot; in this sentence can refer to either &amp;quot;the prize&amp;quot; or &amp;quot;the committee&amp;quot;. The additional use of semantic constraints is required to determine that the proper referent of the reflexive pronoun is &amp;quot;the committee&amp;quot;.</Paragraph>
    <Paragraph position="2"> Applying such constraints is the responsibility of the semantic interpretation component of our system.</Paragraph>
    <Paragraph position="3"> In the current implementation reported on here, semantic interpretation is applied after both parsing and the c-command tree-traversal have been performed. It is a two-stage process in which the first stage is concerned with &amp;quot;structural semantics&amp;quot;nthe semantic consequence of syntactic structurenand the second stage with &amp;quot;lexical semantics&amp;quot;~the specific meanings of individual words with respect to a given application domain. This architecture for semantic interpretation was adopted from the PHLIQA1 system (Bronnenberg, et al. (1980)) and has been used in ~'eating several difficult semantic phenomena (de Bruin and Scha (1988); Scha and Stallard (1988)).</Paragraph>
    <Paragraph position="4"> The structural semantics stage operates on the parse tree to produce an expression of a language called &amp;quot;EFL&amp;quot; (for English-oriented Formal Language). This language is a higher-order intensional logic which includes a single descriptive constant for each word in the lexicon, however many senses that word may have. (From this standpoint, therefore, EFL is actually an ambiguous logical language.) Expressions of EFL are produced from the parse tree by a system of semantic rules, paired one-for-one with the syntactic rules of the grammar, which compute the EFL translation of a tree node from the EFL translations of its daughter nodes. The single EFL of a word is stored in its entry in the lexicon.</Paragraph>
    <Paragraph position="5"> The lexical semantics stage operates on an expression of EFL to produce zero or more expressions of a language called &amp;quot;WML&amp;quot; (for World Model Language). WML is a higher-order intensional logic, with the same set of operations as EFL, but with un-ambiguous descriptive constants which correspond to the primitive concepts and relations of the particular application domain. WML expressions also have types, which are derived from the primitive disjoint categories of the application domain and which serve to delimit the set of meaningful WML expressions.</Paragraph>
    <Paragraph position="6"> A set of translation rules pair ambiguous constants of EFL with one or more unambiguous expressions of WML. Translation to WML is performed by producing all possible combinations formed from replacing the EFL constants with their translations, and filtering to remove combinations which are disallowed by WML's type system. In this way selectional restrictions are represented and enforced.</Paragraph>
    <Paragraph position="7"> The algorithms for producing EFL and WML are slightly modified in the case of anaphoric consituents: that is, reflexive pronouns, personal pronouns, and pro N-BARs~ When the structural semantics component encounters an anaphoric constituent in the course of translating a parse tree to EFL, it creates a new EFL constant &amp;quot;on the fly&amp;quot; to serve as the EFL translation of this constituent. It marks this constant specially and attaches to it the EFL translations of the syntactically possible antecedents of the constituent, along with semantic type information (such as for gender) constraining the antecedents which make sense for it. If the constituent is a personal pronoun or pro N-BAR (but not a reflexive pronoun), a special constant of WML is also attached, marked with the EFL translations of the impossible antecedents of the constituent. This special WML constant represents the possibility of extra-sentential resolution of the anaphor.</Paragraph>
    <Paragraph position="8"> The EFL to WML translation algorithm treats the anaphoric EFL constant specially, returning as its WML translations the translations of the &amp;quot;possible antecedents&amp;quot; that were attached in the EFL phase, together with the WML constant for extra-sentential reference (when this is appropriate). Expansion and filtering then proceed as described above.</Paragraph>
    <Paragraph position="9"> (22) is handled as follows. We will suppose the following &amp;quot;domain model&amp;quot; of WML constants and  The combinatorially possible WML translations are the following, where anomally with respect to the type system is marked with a ......</Paragraph>
    <Paragraph position="10">  The first interpretation is anomalous because the function &amp;quot;AWARD&amp;quot; is applied to an argument whose type is disjoint with the function's domain (in the third argument place). It is therefore discarded, leaving the second interpretation as the correct one.</Paragraph>
    <Paragraph position="11"> A different example; in which a pronoun could have an extra-sentential antecedent, is: (23) The committee awarded the prize to it.</Paragraph>
    <Paragraph position="12"> In this case, neither NP inside the sentence is syntactically allowable as an antecedent of &amp;quot;it&amp;quot;, and so only the extra-sentential possibility remains. The WML translation for (23) is:</Paragraph>
  </Section>
  <Section position="6" start_page="267" end_page="267" type="metho">
    <SectionTitle>
(AWARD (THE COMMITTEE) (THE PRIZES) iT001)
</SectionTitle>
    <Paragraph position="0"> where IT001 is a WML constant marked for disjoint reference:</Paragraph>
    <Paragraph position="2"> This information is necessary so that the module responsible for extra-sentential discourse can prevent external resolution of the pronoun to an internally (syntactically) forbidden antecadent--as could otherwise happen if &amp;quot;the committee&amp;quot; or &amp;quot;the prize&amp;quot; was mentioned in preceding discourse.</Paragraph>
    <Paragraph position="3"> Unless the anaphoric constituent is a reflexive pronoun, an extra-sentential alternative will always be present as a WML translation option, and survive type filtering (since it is given the most general possible type). When both intra- and extra-sentential alternatives survive type filtering, our current heuristic is to prefer the intra-sentential one.</Paragraph>
  </Section>
  <Section position="7" start_page="267" end_page="267" type="metho">
    <SectionTitle>
5. COMPARISON WITH RELATED WORK
</SectionTitle>
    <Paragraph position="0"> Hobbs (1978) has done the only previous work we know of to use traversal of a syntactic parse tree to determine pronominal reference and we compare our algorithm with his in this section. Hobbs proposes a syntactic tree-traversal algorithm for pronominal reference that is &amp;quot;part of a larger left-to-right interpretation process&amp;quot; (Hobbs (1978, p. 318)). When a pronoun is encountered, the algorithm moves up to the nearest S or NP node (our &amp;quot;minimal domain nodes&amp;quot;) that dominates the pronoun and searches to the left of the pronoun for any NP nodes that are dominated by an intervening $ or NP node to propose as antecedents.</Paragraph>
    <Paragraph position="1"> The algorithm then proceeds up to the next NP or S node and searches to the left of the pronoun for any NP nodes to propose as antecedents. At this level, search is also made to the right for NP nodes to propose as antecedents. This will handle cases of backwards pronominalization, as in (18). However, this portion of the search is bounded; it does not seek antecedents below any NP or S nodes encountered.</Paragraph>
    <Paragraph position="2"> The search for c-commanding antecedents and antecedents for backwards pronominalization continues in this fashion until the top S is reached. At this point, preceding utterances in the discourse are searched, going from most recent to least recent. Each tree is searched in a left-to-right, breadth-first manner for NPs to propose as antecedents.</Paragraph>
    <Paragraph position="3"> There are several differences between this a{gorithm and ours. The major one is that our algorithm is a single-pass, depth-first, exhaustive traversal whereas Hobbs' algorithm first walks down the tree, then up, and then back down and is not guaranteed to be exhaustive. Hobbs also imposes a &amp;quot;nearness&amp;quot; condition on the search for antecedents in the case of backwards pronominalization. However, as Hobbs points out, this restriction rules out the perfectly acceptable (24a) and (24b).</Paragraph>
    <Paragraph position="4"> (24) a. Mary sacked out in \[his\] apartment before \[Sam\] could kick her out.</Paragraph>
    <Paragraph position="5"> b. Girls who \[he\] has dated say that \[Sam\] is charming.</Paragraph>
    <Paragraph position="6"> These examples show that the question of what the correct nearness constraint, if any, is remains open.</Paragraph>
    <Paragraph position="7"> Finally, Hobbs' algorithm handles both intra-sentential and extra-sentential pronominal reference relations, while ours is only intended to handle intra-sentential cases.</Paragraph>
  </Section>
  <Section position="8" start_page="267" end_page="270" type="metho">
    <SectionTitle>
6. CURRENT STATUS AND FUTURE
RESEARCH
</SectionTitle>
    <Paragraph position="0"> In this section, we conclude by discussing some of the strengths and weaknesses of the current implementation and areas for future research. The shortcomings fall into two general categories: limitations of the implementation proper and limitations of the theory of pronominal reference that was implemented. null There are two general sorts of limitations to the mechanism described here: those that may be overcome by adding additional filtering devices to the basic tree-walking engine and those that may require a change in that basic engine. We begin with limitations of the first sort.</Paragraph>
    <Paragraph position="1"> Currently, the algorithm does not do any checking on the potential antecedents of a pronoun or bound anaphora to see if they agree in person and number, s For bound anaphors, this is straightforward: a bound anaphor and its antecedent must agree in person and number. For personal pronouns, on the other hand, eCuwently, NPs are not specified for gender in our system, so this cannot be checked.</Paragraph>
    <Paragraph position="2">  the situation is more complicated. In the singular, first ('T', &amp;quot;me&amp;quot;), second (&amp;quot;you&amp;quot;), and third (&amp;quot;he&amp;quot;, &amp;quot;him&amp;quot;, &amp;quot;she&amp;quot;, &amp;quot;her&amp;quot;, &amp;quot;it&amp;quot;) personal pronouns require agreement in both person and number. In the plural, however, the number requirement is dropped because of &amp;quot;split antecedents&amp;quot; cases, in which more than one NP forms part of the antecedent of a pronoun, as in: (25) \[John\] told \[Bill\] that \[they\] should leave.</Paragraph>
    <Paragraph position="3"> where &amp;quot;John&amp;quot; and &amp;quot;Bill&amp;quot;, together, antecede &amp;quot;they&amp;quot;. Third person plural pronouns still require that each antecedent of a split antecedent itself be third person.</Paragraph>
    <Paragraph position="4"> First person (&amp;quot;we&amp;quot;, &amp;quot;us&amp;quot;) and second person (&amp;quot;you&amp;quot;) pronouns also allow split antecedents, but with looser person agreement requirements: (26)a. \[I\] told \[John\] that \[we\] should go.</Paragraph>
    <Paragraph position="5"> b. \[I\] told \[you\] that \[we\] should go.</Paragraph>
    <Paragraph position="6"> c. \[Bill\] told \[you\] that \[you\] should go.</Paragraph>
    <Paragraph position="7"> d. I told \[you\] that \[you\] should go.</Paragraph>
    <Paragraph position="8"> e. ~John told __Bill that w._.ee should go.</Paragraph>
    <Paragraph position="9"> f. John told Bil.._ll that you should go.</Paragraph>
    <Paragraph position="10"> Note that a first person plural pronoun allows split antecedents only if at least one of them is itself first person; contrast (26a) and (26b) with (26e). Similarly, a second person plural pronoun allows split antecedents only if at least one of them is also second personmcontrast (26c) with (26f)--but not if one is first person; contrast (26c) with (26d).</Paragraph>
    <Paragraph position="11"> While the constraints on singular and third person plural pronouns could be implemented as a local agreement check (e.g. as a pre-condition for being added to a pronoun's :possible-antecedents slot), the person agreement constraint on first and second person plural pronouns would require a separate postprocess, since it is not a local constraint on individual split antecedents, but a global constraint on the set of them. Currently, since our algorithm imposes no agreement checks, it allows both the good cases of split antecedents as well as the impossible ones. We need to add the check to our algorithm and extend the semantics to also deal with split antecedents.</Paragraph>
    <Paragraph position="12"> The algorithm also does not check for &amp;quot;crossover&amp;quot; cases. Roughly speaking, these are examples similar to backwards pronominalization cases such as (18) (repeated here as (27a)), in which the potential antecedent is a quantifier or a trace of a moved WH element. In such cases, overlapping reference is impossible. Contrast (27a) with (27b) and (27c).</Paragraph>
    <Paragraph position="13"> (27) a. \[His\] mother loves \[John\].</Paragraph>
    <Paragraph position="14"> b. ~His mother loves everyone.</Paragraph>
    <Paragraph position="15"> c. Who does hi._? mother love twho? These particular cases can be handled by adding a check to clause \[I.C\] to prohibit quantified NPs and WH-traces from participating in backwards pronominalization. However, the more general problem of how elements dislocated by WH movement or by topicalization interact with the algorithm given here is a topic that requires further work beyond this simple measure.</Paragraph>
    <Paragraph position="16"> More seriously, there is also a well-known case of pronominal reference within NPs that is not handled by the algorithm. A constraint from the syntactic theory of reference implemented by our algorithm is that if the antecedent-anaphor relation holds between two positions, disjoint reference also holds between them; see examples (10) and (11), and (13) and (14).</Paragraph>
    <Paragraph position="17"> However, there is one position in English where this generalization is known not to hold: the possessive position of an NP. A bound anaphor is possible here, but a pronoun in the same position is not subject to disjoint reference; see (28): (28) a. \[The men\] read \[each other's\] books.</Paragraph>
    <Paragraph position="18"> b. \[The men\] read \[their\] books.</Paragraph>
    <Paragraph position="19"> (28a) is correctly handled by the algorithm as already outlined; pass-down-c-commanding-nodes treats the nodes internal to the current minimal domain as internal nodes for the possessive in a Noun Phrase, so the NP &amp;quot;the men&amp;quot; will be added to the :possible-antecedents slot of a bound anaphor in this position. However, the same characteristics of the algorithm will also result in the NP &amp;quot;the men&amp;quot; being assigned to the :impossible-antecedents slot of &amp;quot;their&amp;quot; in (28b). One possible remedy for this situation is to add a clause to update-node that checks for possessive pronouns separately from other pronouns and that allows NPs both internal and external to the current minimal domain to be possible antecedents.</Paragraph>
    <Paragraph position="20"> However, the more far-reaching modifications proposed in the discussion below of the theory of pronominal reference would obviate this change.</Paragraph>
    <Paragraph position="21"> There are several areas where our implementation points out problems with the structural theory of pronominal reference. The first of these is the definition of c-command itself. 7 Under Reinhart's (1976) original definition, a node A c-commands node B iff the branching node most immediately dominating A also dominates B and A does not dominate B. The difference between the two definitions can be seen in Figure 2-1; in addition to the c-command statements given there, Reinhart's definition adds the following: E c-commands B, C, F, D, and G F c-commands D and G G c-commands C and F These statements are true under Reinhart's definition of c-command, because no branching category intervenes between the c-commanding and c-commanded nodes, but not under that used in the implemented algorithm, since there is no sisterhood among the nodes. We have found this modified definition to be easier to implement; moreover, various researchers (e.g. Aoun and Sportiche (1983)i have pointed out  The implementation has also brought to light asymmetries in the strictness of c-command used to determine the antecedents of a bound anaphor and that used to determine the non-antecedents of a pronoun. In particular, none of the conjuncts of a conjoined NP can be the antecedent of a reflexive: (29) &amp;quot;John and Mary like himself.</Paragraph>
    <Paragraph position="22"> However, all of the conjuncts of a conjoined NP are impossible antecedents for any pronoun for which the entire conjoined NP is an impossible antecedent. In (30) John and Mary like him.</Paragraph>
    <Paragraph position="23"> &amp;quot;John&amp;quot; cannot be an antecedent of &amp;quot;him&amp;quot;, despite the fact that &amp;quot;John&amp;quot; does not c-command &amp;quot;him&amp;quot;. Contrast this with (19) where a non-c-commanding possessive can be the antecedent of a pronoun. This is handled correctly in the implementation. Whenever our algorithm adds a conjoined NP to the :impossible-antecedents slots of a pronoun or a non-pronominal NP, it adds all the conjuncts of that NP, as well. While this works, there is clearly something that is being missed here. Presumably, it should follow by definition that no individual conjunct of a conjoined NP can be a possible antecedent of a Noun Phrase with which the entire conjoined NP is disjoint in reference, s A more serious problem with the theory of pronominal reference elaborated in Chomsky (1980) and (1981), and which our algorithm implements, is the crucial assumption that referentially dependent Noun Phrases can be exhaustively partitioned into bound anaphors vs. personal pronouns and that, therefore, they will be in complementary distribution.</Paragraph>
    <Paragraph position="24"> However, examples such as (28), as well as (31) (pointed out by Kuno (1987)) and (32) indicate that the notion of exhaustive partitioning of bound anaphors against personal pronouns is incorrect in the general .case, even though it may be the typical state of affairs.</Paragraph>
    <Paragraph position="25"> (31) a. \[John\] put the blanket under \[himself\].</Paragraph>
    <Paragraph position="26"> b. \[John\] put the blanket under \[him\].</Paragraph>
    <Paragraph position="27"> (32) a. Ill buy myself a beer.</Paragraph>
    <Paragraph position="28"> b. rll buy me a beer.</Paragraph>
    <Paragraph position="29"> We can keep the insight of the structural theory of pronominal reference (i.e. that structural relations play a role in delimiting reference possibilities), while still incorporating these facts, if we give up the restriction that bound anaphors and personal pronouns are always in complementary distribution. One possible approach to this problem is to use feature decomposition to characterize bound anaphors and pronouns: the feature +short-distance indicates whether a pronominal can be used as a short-distance anaphor while the feature :l:long-distance indicates whether it eThanks to Leland George for this insight, as well as for discussion of short and long distance enephors.</Paragraph>
    <Paragraph position="30"> can be used as a long-distance anaphor. 9. While, in the normal case, personal pronouns in English are specified to be long-distance anaphors that cannot be used as short-distance anaphors, (i.e. as I-short-distance +long-distance\]) this system would allow the feature governing a pronominal's use as a short-distance anaphor to be left free (i.e. as ?short-distance) in certain syntactic contexts in English, such as the possessive position of a Noun Phrase, the object of certain prepositions, and the indirect object position of verbs. 1deg Such a view of the syntax of personal pronouns could be implemented in a unification grammar fairly straightforwardly.</Paragraph>
    <Paragraph position="31"> While such a treatment of personal pronouns as short-distance anaphors does not handle all the counter-examples to the syntactic theory of pronominal reference raised by researchers such as Kuno, it does begin to address them seriously.</Paragraph>
    <Paragraph position="32"> Clearly, it is more in accord with the facts than a theory that postulates an exhaustive partitioning of bound anaphors vs. personal pronouns, and so constitutes, in our opinion, a promising start towards handling the full range of pronoun reference facts in a reasonable manner.</Paragraph>
    <Paragraph position="33"> Finally, we consider alternate ways of combining our pronominal reference mechanism with parsing and semantic interpretation. One possibility is a fully incremental architecture in which c-command constraints, semantic interpretations, and external reference resolution are computed simultaneously with the parse. Such an architecture might seem particularly attractive for processing large sets of alternatives, such as are encountered when processing spoken input. The intra-sentential reference phenomena described in this paper pose a problem for such an incremental approach, however. The possiblities for internal resolution for an anaphor cannot all be known locally to the anaphor, but must be obtained from elsewhere in the sentence. In many cases antecedents will lie to the left of the anaphor in the sentence, and thus will have been seen by a left-to-right parser by the time the anaphor is reached. But consider a case of backward pronominalization, as in (18), repeated here as (33): (33) His mother loves John.</Paragraph>
    <Paragraph position="34"> A wholly incremental mechanism, parsing the NP &amp;quot;his mother&amp;quot; first, would have to conclude that the referent of &amp;quot;his&amp;quot; was extra-sentential, since no intra-sentential referent was seen to the left. And if no extra-sentential referent could be found, the NP would have to be rejected. To be successful, such an incremental mechanism would have to be modified to include a kind of &amp;quot;lazy evaluation&amp;quot; which could rule out certain ~'hle is akin to the feature system :l:anaphodC/ :l:pronomlnal of Chomsky (1981) 'degThis suggestion was originally made by Lust, et aL (1989) who support it on the basis of language acquisition data.</Paragraph>
    <Paragraph position="35">  referents for an anaphor but never rule an anaphor empty of referents until utterance processing had been completed.</Paragraph>
    <Paragraph position="36"> Another alternative would be to separate intra-sentential anaphor resolution from semantic interpretation, performing it instead in conjunction with extra-sentential discourse processing. A possible problem for this approach can be seen in sentences where the anaphor is combined with another ambiguous element, so that proliferation of semantic interpretations occur, as in: (34) John's car is better than Bill's.</Paragraph>
    <Paragraph position="37"> where the pro N-BAR, left completely unspecified during semantic interpretation, is free to generate all sorts of combinations with the possessive, including those in which the possession is appropriate to various &amp;quot;relational&amp;quot; interpretations of the pro N-BAR (de Bruin and Scha (1988)).</Paragraph>
    <Paragraph position="38"> In future work, we plan to combine parsing and semantic interpretation into a single unification grammar incorporating semantic information in additional features. Part of that work will be to look for the optimal method of combining it with the pronominal reference mechanism presented here.</Paragraph>
  </Section>
class="xml-element"></Paper>