File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/88/c88-1021_metho.xml
Size: 28,456 bytes
Last Modified: 2025-10-06 14:12:05
<?xml version="1.0" standalone="yes"?> <Paper uid="C88-1021"> <Title>Anaphora Resolution: A Multi.Strategy Approach</Title> <Section position="3" start_page="96" end_page="96" type="metho"> <SectionTitle> 2. The Problem: Semantics and Pragmatics Dominate </SectionTitle> <Paragraph position="0"> Finding the appropriate anaphoric referent has been long recognized as a difficult problem, requiring lnuch ~emautic and pragmatic lmowledge. Consider, for instance, the tbllowing two sets of examples: John took the cake from the table and ate it.</Paragraph> <Paragraph position="1"> John took the cake from the table and washed it.</Paragraph> <Paragraph position="2"> Tile robot pushed the box towards the conveyor belt. But, it goojSd and dropped it on its way there.</Paragraph> <Paragraph position="3"> Semantic preference constraints (e.g., \[18, 11), if pmp:~rly coded, suffice to resolve the first example. The pt'eferre'l object of ingestation is an edible substance. It is a little more difficult to mechanize a process that excludes things such as cakes from being the object of washing. One cannot simply write a &quot;NOT(edibley restriction on the object case of the verb &quot;to wash&quot;. Alter ,all, vegetables and fruits are occasionally washed prior to eating them. Peltlaps a combination of typicality judgements with pragmatic knowledge exla'apolating the effects of attempting to drown a cake in sink full of water comes into play. Or, more abstract irffemntial constraints are appropriate, snch as requiring that the object of wash be unchanged by immersion in water. Interestingly, Subjects given only the &quot;...and washed it&quot; sentence report consistently that they didn't even consider the cake a reasonable referent for &quot;it&quot;. In the robot example, there ate four anaphoric referents, counting the possessive &quot;its&quot; and the locative &quot;there&quot;, referring to three different antecedents. Although subjects report little difficulty ascertaining the referent for each anaphor in a consistent manner, it appears that sophisticated semantic~ are required. Why is the referent for &quot;it&quot; in &quot;it goofed and dropped...&quot; the robot rather than the box or the conveyor belt? One could argue that the box cannot take action, but what allows a robot to goof and not a conveyor belt? Is it something as subtle as the degree to which the former can be anthropomolphized being greater than the degree to which the latter can be anthmpomorphized? The difficulty in anaphoric referent specification in narratives has been argued convincingly by many researchers including Chamiak in his work on children's story comprehension \[9\], where substantial pragmatic domain knowledge must be brought to bear, and by one of the authors \[4\], where knowledge of goals and personality traits is required to resolve difficult referents. Hence, the hypothesis that anaphor resolution in its fnll generality is at best a diffictflt problem, and at worst an almost intractable one, is well supported.</Paragraph> <Paragraph position="4"> Nevertheless, somewhat less ambitious endeavors can prove far more tractable, and yet be of major practical Import. Hayes \[13\] argued for the notion of limitod-domain anaphora in a natural language interface to an electronic mail system. Webber \[17\] demonstrated that intrasentential anaphora was more tractable than its intersentential counterpart, largely through the categorization of syntactic devices absent from larger textual or dialog segments.</Paragraph> <Paragraph position="5"> This paper explores an intemrediate position: addressing much larger classes of anaphors than those of Hayes \[13\] in a systematic mariner, but stopping short of full generality, which requires unbounded pragmatic knowledge and inference. We explore the central hypothesis that anaphora resolution may be best accomplished through fire combination of a set of strategies, rather than by a single monolithic method. The apparent complexities lie in the combination of these multiple strategies to produce syntactically, semantically arid pragmatically sound anaphoric resolutions. In the multiple examples analyzed, 3 unambiguous resolutions reported by human subjects correspond to situations where the applicable strategies concur on the referont of an anaphor, and disagreement on the con'eet referent by the human subjects corresponds to situations where the applicable strategies propose different candidate referents for file anaphor in questitm.</Paragraph> </Section> <Section position="4" start_page="96" end_page="98" type="metho"> <SectionTitle> 3. Multiple Resolution Strategies </SectionTitle> <Paragraph position="0"> In this section we propose a general framework for anaphor resolution ba~d on the integration of multiple knowledge sources: sentential syntax, case-frame semantics, dialog structure, and general world knowledge. The underlying theoretical tenet is: Anaphor resolution is not a monolithic autonomous process; it requires access and integration of all the knowledge sources necessary for dialog and text lnterprel'ation. These linguistic knowledge sources are brought to bear as constraints or preferences encoded as multiple resolntion strategies.</Paragraph> <Paragraph position="1"> Each source of knowledge usethl in resolving intersentential anaphnra is presented below, along with corresponding examples, and a statement of the anaphoric resolution strategy.</Paragraph> <Section position="1" start_page="96" end_page="96" type="sub_section"> <SectionTitle> 3.1. Local A naphor Constraints </SectionTitle> <Paragraph position="0"> Certain anaphot~ carry with them constraints (number, gender, case, etc.) which must be satisfied by the candidate referents. For instant'e, gender uniquely specifies the anaphor in: John at~d Mary went shopping. He bought a steak.</Paragraph> <Paragraph position="1"> \[he=John\] Tile strategy here is trivial: Eliminate from consideration all candidate referents that violate the local constraints of the anaphor its question. A variant of this strategy has been implemented in RUS and in XCALIBUR.</Paragraph> </Section> <Section position="2" start_page="96" end_page="96" type="sub_section"> <SectionTitle> 3.2. Case-role Semantic Constraints </SectionTitle> <Paragraph position="0"> Here the ease-role semantics impose constraints on what can fill them. If they are filled by an anaphor (which specifies few if any semantic features), the case role constraints must be also satisfied by the referent of tile anaphor, thus eliminating from consideration all candidate anaphor referents that violate constraints on the case role occupied by the anaphor. Consider our previous example, where the semantic constraints on the object case of &quot;to eat&quot; and &quot;to wash&quot; impose restrictions on the possible case fillem and prove sufficient to select a unique referent.</Paragraph> <Paragraph position="1"> John took the cake from the table and ate it. lit:cake\] John took the cake from the table and washed it. lit=table\] The slrategy here is also fairly simple: Eliminm'e from consideration all candidate referents that violate any case-constraint imposed on the anaphor its question~ Prefer those candidates that accord with typical ease fillers, in the absence of hard constraints.</Paragraph> <Paragraph position="2"> XCALIBUR implements this strategy directly though use of its case-frame grammar. With the I-rule mechanism, it was possible to implement an ad-hoc variant of this strategy in RUS as well.</Paragraph> </Section> <Section position="3" start_page="96" end_page="96" type="sub_section"> <SectionTitle> 3.3. Preconttition/Postcondition Constraints </SectionTitle> <Paragraph position="0"> Using real-world knowledge and pragmatics, it is possible to say that a candidate antecedent cannot be the referent of an anaphor because some action occurring betwee n the referent and the anaphor invalidates the assumption that they denote one and the same object or event.</Paragraph> <Paragraph position="1"> John gave Tom an apple. He ate the apple. \[he=Tom\] Here, &quot;he&quot; refers to Tom, as Jolm no longer has the apple. The postcondition on give is that the actor no longer have the object being given, which ,:onfiicts with the precondition on eat that the actor have the item being eaten, if the actor is assumed to be John.</Paragraph> <Paragraph position="2"> The strategy is simple, but requires a fairly large amount of knowledge to be useful for a broad range of cases: Eliminate from consideration all candidate referents associated with actions whose postconditions violate the preconditions of the action containing the anaphor.</Paragraph> </Section> <Section position="4" start_page="96" end_page="96" type="sub_section"> <SectionTitle> 3.4. Case.role Persistence Preference </SectionTitle> <Paragraph position="0"> We observe a pervasive form of &quot;linguistic inertia&quot; that manifests as a preference to assign the referent of an anaphor to the linguistic entity in the discourse context that filled the corresponding semantic case role in an earlier utterance. This is a generalized form of ease-role parallelism, which has proven crucial in ellipsis resolution \[8, 7, 5\], although in anaphora resolution it is demoted from the status of a categorical constraint to that of a preference.</Paragraph> <Paragraph position="1"> Mary gave an apple to Susan. John also gave her an orange. \[her=Susan\] Mary gave an apple to Susan. She also gave John an orange. \[she=Mary\] The first anaphor relers to Susan, whereas the second anaphor refers to Mary. Clearly it is not a matter of primacy or recency, as the sentence structures are identical. Rather it is a case of structural parallelism. And, the semantic structnre dominates over the syntactic one. For instance, in the first example, &quot;Susan&quot; is the object of the &quot;to&quot; prepositional phrase, whereas the corelerent anaphor is in the indirect ol~iect position: two different syntactic roles that map into the same semantic case, recipient. In the second example above, both syntactic mad semantic structures coincide, and therefore the preference is stronger. Note, moreover, that the subject or direct-object form of the pronoun (&quot;she&quot; vs &quot;her&quot;) is not the primary source of discriminant knowledge. For instance, in the example below, one has only the anaphor &quot;it&quot;, but the same referent discrimination occurs by semantic case-role parallelism: The robot gave the dog a bone. John also gave it some water, lit=dog\] The robot gave the dog a bone. It also gave John some water, lit=robot\] To provide more ammunition in support of semantic case role persistence, consider the following final example, with three possible referents to the anapher &quot;him&quot;. It is clear that &quot;Peter&quot; is the preferred referent, once again due to the persistence of the underlying semantic recipient case.</Paragraph> <Paragraph position="2"> John carried the box of papers from Bill to Peter.</Paragraph> <Paragraph position="3"> He also sent him Mary's books. \[he=John, him=Peter\] The semantic preference strategy can be stated as follows: Search first for acceptable referents in the antecedent phrase (or phrases) that occur in the same semantic case role as the attaphor, lf a match satisfying all constraints is found, look no further; else search the other case roles.</Paragraph> <Paragraph position="4"> To our knowledge, this preference strategy has been neither proposed nor implemented prior to our work on the Universal Parser (reported below), yet it counts for a large number of anaphor resolutions in our sample set.</Paragraph> </Section> <Section position="5" start_page="96" end_page="97" type="sub_section"> <SectionTitle> 3.5. Semantic Alignment Preference </SectionTitle> <Paragraph position="0"> A form of pragmatic &quot;Occam's razor&quot; exists in not postulating extra roles for the same objects in different sentences in the discourse. This preference is a more gcner~d and looser form of case role inertia, discussed above, in that the we have inmtia of the underlying action.</Paragraph> <Paragraph position="1"> For instance, in the example below, this preference manifests as preferring all departures to be from the park, and all arrivals m be at the club: Mary drove from the park to the club. Peter went there too.</Paragraph> <Paragraph position="2"> \[there=chth\] Mary drove from the park to the chub. Peter left there too.</Paragraph> <Paragraph position="3"> \[there=park\] The locative anaphor &quot;there&quot; refers to &quot;the club&quot; in the first example above, but refers to &quot;the park&quot; in the second example, yet both sentences share the identical syntactic structure and the same basic underlying semantic case structure. However, discourse cohesion prefers to make the sentences coreferential (pragmatically parallel) with respect to the same underlying action (leaving the pink and going to the club). Therefore, the former aligns with the second (destination) part, whereas the latter aligns with the first (source) part. The strategy here is a bit more difficult to state, and certainly has not been implemented in any system to date: If the clause in which the anaphor is embedded aligns with a previous clause (&quot;aligns&quot; means that it can represent the same underlying action, perhaps with d~fferent instantiated case fillers), or with part of a previous clause, search first for referents of the anaphor in that clause. If there are no allowable re#rents in the semantically aligned clause, expand the search to other antecedent clauses; else halt the search.</Paragraph> <Paragraph position="4"> 3.6. Syntactic paralielism preference Although semantic and pragmatic parallelism (case-persistence, and alignment, respectively, in the discussion above) appear to dominate over syntactic parallelism, the latter plays an important role if two clauses are directly contrasted (e.g., in a coordinate structure, or by means of explicit discourse cohesion markers \[14\]). Consider the following examples: The girl scout leader paired Mary with Susan, but she had paired her with Nancy last time. \[she=leader, her=Mary\] The girl scout leader paired Mary with Susan, but she had paired Nancy with her last time. \[she=leader, her=Susan\] There is no reason to prefer different referents for the pronoun &quot;her&quot; in each sentence above, other than retaining as much as possible the surface syntactic order from the first coordinate clause in the second clause. The strategy here is summarized as follows: In coordinated clauses, adjacent sentences or explicitly contrasted sentences, prefer the anaphoric referent that preserves the surface syntactic role from the first clause.</Paragraph> </Section> <Section position="6" start_page="97" end_page="98" type="sub_section"> <SectionTitle> 3.7. Syntactic Topicalization Preference </SectionTitle> <Paragraph position="0"> Topicalized structures are searched first for possible anaphoric referents. Consider, for instance, the following pseudo-cleft constmctions: It was Mary who told Jane to go to New York. Why did site do it? \[she=Mary\] It was Jane who went to New York at Mary's bidding. Why did she do it? \[she=Jane\] It was Mary who told Peter to go to New York. Why did he do it? \[he=Peter\] It was Peter who went to New York at Mary's bidding. Why did he do it? \[he=Peter\] In the first set of examples, describing essentially the same tcaderlying action, the topicalized person becomes the referent of the anaphor &quot;she:&quot; &quot;Mary&quot; in the first sentence, &quot;Jane&quot; in the second. And, the action associated with that person become the r~ferent of &quot;it.&quot; However, to stress that topicalization is a preferential rather than categorical strategy, consider the second set of examples above. The exact same semantic and syntactic structures yield &quot;Peter&quot; both times as the referent of &quot;lie&quot;, because localized constraints so dictate, regardless of who is topiealized. Thus, it is important to distinguish constraints from preferences in anaphora resolution. The topicalization strategy may be stated as follows: Search first a syntactically topiealized part of the candidate antecedent clause (or clauses) for the referent of the anaphor. If an acceptable referent is found, search no further; else search the rest of the clause(s).</Paragraph> <Paragraph position="1"> This strategy surprisingly enough has not been exploited h~ any system to our knowledge, although it is easy to establish syntactic topiealization (indicated by linguistic devices such as fronting, and cleft constructions). In contrast, the much more complex phenomenon of pragmatic topiealization by dialog focus or actor focus (discussed below) was suggested by Sidner \[lS\] We also believe that dialog l:'oeus can yield a useful preference for anapnmac reference selection, but lacking a computationally-adequate theory for dialog-level focus trackh~g (Sidner's is a partial theory), we could not yet implement such a strategy.</Paragraph> </Section> <Section position="7" start_page="98" end_page="98" type="sub_section"> <SectionTitle> 3.8. Intersentential Recency Preference </SectionTitle> <Paragraph position="0"> Thus far we have focused on the problem of selecting the best anaphoric referent among several candidates, all from a single previous sentence (or coordinated clause). When prior context contains many sentences, the question naturally ari~s of how far back to search for the anaphoric referent, and how to prioritize that search.</Paragraph> <Paragraph position="1"> At the paragraph (or dialog) level level, we advocate searching sentences in reverse chronological ordm, applying all the constraints and preferences to select among possible candidates within each sentence. If there are no satisfactory candidates in the previous sentence, then the one before that is considered, and so on. Although we are investigating more sophisticated tectmiques, these await a more comprehensive (non-linear) theory of discourse structure - and one that is precise enough to permit implementation.</Paragraph> </Section> </Section> <Section position="5" start_page="98" end_page="98" type="metho"> <SectionTitle> 4. Integrating the Strategies </SectionTitle> <Paragraph position="0"> In order to apply a diverse set of strategies, such as those presented in this paper, one needs to make a distinction between constraints (which cannot be violated), and preferences (which discriminate among candidates satisfying all constraints). The latter may be ranked in a partial order (as the goals trees in \[4\]), or may be offered a voting scheme where the stronger preferences get more votes, and where conflicting preferences of equal voting power indicate true ambiguity.</Paragraph> <Paragraph position="1"> Our resolution method works by applying the constraints first to reduce the number of candidate referents for the anaphor in question. Then, the preferences are applied to each of the remaining candidates. If more than one preference applies, and each Suggests different candidate referents for the anaphor in question, 'all of which have passed the constraint tests, then we consider the anaphor to have a truly ambiguous referent. Thus, when faced with conflicting knowledge sources of equal strength, we simply reduce the space of possible anaphoric referents to those that are accepted by constraints and indicated as preferred by one or more preferences. Earlier hand simulations of a slightly different method 4 on 70 examples (including those presented earlier in this paper) yielded 49 unique resolutions, 17 conflicting possibilities, and 4 anomalous cases. Human judgements correlate very well in terms of identifying the same referent as that suggested by the system in the 49 unique cases. 5 Moreover, the majority of the 17 multiple-referent cases were judged ambiguous by our subjects (the rest required complex world knowledge to establish a unique referent). Therefore, we believe that one can indeed achieve human-like performance with the multi-strategy method of determining referents to anaphors using different sources of linguistic knowledge in a semi-modular fashion.</Paragraph> </Section> <Section position="6" start_page="98" end_page="99" type="metho"> <SectionTitle> 5. A Practical Implementation </SectionTitle> <Paragraph position="0"> We have developed an anaphor resolver using Local Constraints, form of lexical-functional grammar\[3\] unifying symactie and semantic knowledge sources to produce a complete parse of each sentence. The anaphor resolver operates post facto on the set of instantiated semantic case frames and syntactic trees, attempting to resolve anaphors in the parse of the newest sentence using earlier parses (semantic and syntactic) as context to mine for candidate referents. We expect the resolver to become an integral part of our multi-lingual machine translation effort.</Paragraph> <Paragraph position="1"> Candidate ~'cferents are derived by extracting the noun phrases from the most-receipt previous sentences that the resolvcr has processed. The number of sentences examined may be changed, allowiug the future addition of discourse phenomena to further restrict the sentenees which are examined for candidate referents.</Paragraph> <Paragraph position="2"> The pmfere,Lces use a voting method to detennine which candidate referent is most preferred. Each preference strategy is given an individual weight, and may vote with less than its fuU weight for less-preferred candidates, such as case role persistence in a referent several sentences removed from the anaphor.</Paragraph> <Paragraph position="3"> In addition to ailing out candidates, the case-role and local anaphor constraints may also cast votes tot those allowable candidates which are most clo,~ely matched to the anaphor or con'espond to typical fillers. In elf,',ct, fllese strategies indicate a preference in the absence of hard conso'aints. For ex~unple, ti~e gender constraint would prefer a candidate reference of female gender over one of indeterminate gender when resolving an auaphor of female gender, while at the same time eliminating all candidates of male gender.</Paragraph> <Paragraph position="4"> After applying the preferences, the most preferred candidate referent is unified with the reference to restrict the range of possible values as mu~:h as possible. For example, if she is determined to refer to doctor, all future anaphorie references to the doctor will be required to have female or unknown gender. However, if multiple candidates have received nearly the same number of votes, the anaphor is coasidered to be anthiguous.</Paragraph> <Paragraph position="5"> Ttle anaphor resolver i~ able to resolve partially-specified definite noun phrases with an antecedent noun phrase. To do so, along with the other lo(-al constraints, the head nouns ,'uld the remaining slots in the noun ph~a.~c are checked for agreement with the reference. The head noun ol the candidate must be the samc as, or an instance of, the head noun o(&quot; the reference. For the remaining slots, it suffices for corresponding slots to be uniliable with each other or missing from either the d~,,finite noun phrase or the candidate referent. Unlike anaphors, which must have a suitable referent, it is not considered an error if there are no referents which pass all constraints. We believe that the ability to resolve definite noun phrases with basically the saute approach as anaphors is an indication of the generality of our strategies anal their implementation exploiting semantic and syntactic constraint ut~ification methods.</Paragraph> <Paragraph position="6"> The curt'cut test suite consists of ten examples, totalling 3l ,sentences (:outaining 27 anaphors and three definite noun phrases with prior reii~rents. 6 The anaphor resolver correctly resolves &quot;all but four of the anapho~s, ,'mr determines the correct referent for all of the definite noun ptu'ases; In two of tbe four problematic cases, the anaphor is an it referring to an action only indirectly mentioned, which is beyond the scope of the resolver. Tile remaining two anaphors are in the example John carried the box from Bill to Peter. He also sent him Macy's books.</Paragraph> <Paragraph position="7"> Here, him remains an~biguous, and he also remains ambiguous between John and Bill (with the current voting scheme, John is preferred over Bill).</Paragraph> <Paragraph position="8"> The follo,~eing rtm of the anaphor rcsolver (edited to save space) illustrates several of the strategies. Each candidate referent is tagged with a number indicating how many votes it has received so far. The intersentential recency preference is applied at the same time that the candidates a~e collected and tagged because of its computational efficiency; l ttus, the initial list of candidates already includes the votes from intersentential recency. The ease-role persistence preference is applied between pre/postcenditlon constraints and local constraints, because removal of eliminated candidates (in this implementation) also removes tim information on which previous sentence a candidate originates from. Then, case-role constraints are applied, and if multiple candidates remain, the rem~fining preferences (currently only syntactic topicalizatio10 are applied.</Paragraph> <Paragraph position="9"> ; sentence 6: The doctor gave John a glass of llnterclausal anaphora in coordinate constructions behaves much like a constrained version of intersentential anaphora, where syntactic parallelism (between the coordinated clauses) plays a more dominant role.</Paragraph> <Paragraph position="10"> 2No claim;;, however, are made for file relative frequency or utility of resolving intersentential vs intrasentential anaphors in processing narrative or expository texts.</Paragraph> <Paragraph position="11"> 3Although many of our anaphora instances come from actual user utterances in our experience with domain-oriented human-computer interfaces, we expect that the strategies developed here am of more general applicability. For clarity of exposition in this paper, we have selected exmnples not from our human-computer dialogs, but from everyday events.</Paragraph> <Paragraph position="12"> '*Using preferenccs to determine whict* candidates are tested against the constrain~ s 5Olten, more than one slrategy suggested the same referent, increasing or.r confidence. Language is redundant, and it may prove useful to exp'loit that redundancy.</Paragraph> <Paragraph position="13"> 6The sentences in our corpus used to test the implementation are: John gave Mary two aspirin. She took them from him.</Paragraph> <Paragraph position="14"> Mary had a h,mdache. John gave her two aspirin tablets. She took them.</Paragraph> <Paragraph position="15"> The doctor gme John a glass of water. John drank it. He gave him an aspirin. ~Ie took it with another glass of water.</Paragraph> <Paragraph position="16"> Mary gave art apple to Susan. John also gave her an orange.</Paragraph> <Paragraph position="17"> Mary gave arl apple to Susan. She also gave John an orange.</Paragraph> <Paragraph position="18"> John took the cake from the table. He ate it.</Paragraph> <Paragraph position="19"> Jotm took the cake from the table. He washed it.</Paragraph> <Paragraph position="20"> John took the cake from the table \[ambig\]. He washed it.</Paragraph> <Paragraph position="21"> John carried the box from Bill to Peter. He also sent him Mary's books.</Paragraph> <Paragraph position="22"> It was Mary who told Jane to go to New York. Why did she do it? It was Jane who went to New York at Mary's bidding. Why did she do it? Jotm gave Peler an apple. He ate it.</Paragraph> <Paragraph position="23"> Jack (age 10) went up the hill. John (age 32) went up the hill.</Paragraph> <Paragraph position="24"> The boy fell down.</Paragraph> <Paragraph position="25"> Jack went up the hill. The boy fell down.</Paragraph> </Section> class="xml-element"></Paper>