AN INTEGRATED MODEL FOR ANAPHORA RESOLUTION Ruslan Mitkov Institute of Mathematics Acad. G. Bonchev str. bl.8, 1113 Sofia, Bulgaria ABSTRACT "distributed architecture", but their ideas a) do not seem to capture enough The paper discusses a new knowledge- discourse and heuristical knowledge and based and sublanguage-oriented model b) do not concentrate on and investigate a for anaphora resolution, which integrates concrete domain, and thus risk being too syntactic, semantic, discourse, domain general. We have tried nevertheless k) and heuristical knowledge for the incorporate some of their ideas into our sublanguage of computer science. Special proposal. attention is paid to a new approach for tracking the center throughout a discourse segment, which plays an imtx~rtant role in THE ANAPttORA RESOLUTION proposing the most likely antecedent to MODEL the anaphor in case of ambiguity. Our anaphora resolution model integrates modules containing different types of INTRODUCTION knowledge - syntactic, semantic, domain, discourse and heuristical knowledge. All Anaphora resolution is a complicated the modules share a common problem in computational linguistics. representation of the cunent discourse. Considerable research has been done by computational linguists ([Carbonell & The syntactic module, for example, Brown 88], IDahl & Ball 90], knows that the anaphor and antecedent [Frederking & Gchrke 87], [Hayes 81], must agree in number, gender and [Hobbs 78], [lngria & Stallard 89], person. It checks if the c-command [Preug et al. 9411, [Rich & LuperFoy 88[, constraints hold and establishes disjoint [Robert 89]), but no complete theory has reference. In cases of syntactic emerged which offers a resolution parallelism, it prefers the noun phrase procedure with success guaranteed. All with the same syntactic role as the approaches developed - even if we restrict anaphor, as the most probable antecedent. our attention to pronominal anaphora, It knows when cataphora is possible and which we will do throughout this paper - can indicate syntactically topicalized noun from purely syntactic ones to highly phrases, which are more likely to be semantic and pragmatic ones, only antecedents than non-topicalized ones. provide a partial treatment of the problem. The semantic module checks for semantic Given the complexity of the problem, we consistency between the anaphor and the think that to secure a comparatively possible antecedent. It filters out successful handling of anaphora semantically incompatible candidates resolution one should adhere to the following the cun-ent verb semantics or following principles: l) restriction to a the animacy of the candidate. In cases of domain (sublanguage) rather than focus semantic parallelism, it prefers the noun on a particular natural language as a phrase, having the same semantic role as whole; 2) maximal use of linguistic the anaphor, as a most likely antecedent. information integrating it into a uniform Finally, it generates a set of possible architecture by means of partial theories. antecedents whenever necessary. Some more recent treatments of anaphora ([Carbonell & Brown 88], [Preug et al. The domain knowlcdge module is 941, [Rich & LuperFoy 8811) do express practically a knowlcdge basc of the the idea of "multi-level approach", or concepts of the domain considered and 1170 thc discourse knowledge module knows implementcd; its development, however, how to track the center throughout the is envisaged for later stages of the project. current discourse segment. The syntactic and semantic modules The heuristical knowledge module can usually filter the possible candidates and solnetimes bc helpful in assigning the do not propose an antecedent (with the antecedent. It has a set of useful rules exception of syntactic and semantic (e.g. the antecedent is to be located parallelism). Usually the proposal for an preferably in thc current sentence or in the antecedent comes from the domain, previous one) and can forestall certain heuristical, and discourse modules. The impractical search procedures. latter plays an important role in tracking the center and proposes it in many cases The use of colnmon sense and world as the most probable candidate for an knowledge is in general commendable, antecedent. but it requires a huge knowledge base and set of inference rules. The present version Figurc 1 illustrates the general structure of of the model does not have this mcxtule our anaphom resolution model. IIIiURISTICAI, 1)OMAIN I)|SCOURSI ~, KNOWI ,I ';lX;t i KNOWI ,El X]! i KNOW I ,l ilX]l'~ l)omain lleuristics l)omain Concept Rating Rules '1'racking Center Ktlowledgc 1~ase Recency [ Rl:ilqlilil iNTIA], l ihJ ANAPIIORA ANTI iCI il)ENT ANAPIIOR-~--~.I ixptaisstoN P,I {SOl ,VI {R SYNTACTIC KNOW] ,t il )(}l i Number Agrccmenl St,',MANTI( Gender Agfeelncnl KNOW1,1 '~I)GE PCI'SOll A~l'Celllellt Semm~tic Consistency l)isjoim Reference Case Roles (~-(~ommaud Constraints Semantic Parallelism Cataphora Animacy Syntactic Paralldislll Set Generalion Syntactic Topicalization Figure 1: Anaphora resolution model THE NEED FOR DISCOURSE criteria are therefore needed. CRITERIA As an illustration, considerthe following Although the syntactic and semantic text. criteria for the selection of an antecedent are already very strong, they are not Chapter 3 discusses these additional or always sufficient to discriminate alnong a auxiliary storage devices, wlfieh mc set of possible candidates. Moreover, similar to our own domestic tape they serve more as filters to eliminate cassettes and record discs. Figure 2 unsuitable candidates than as protx)sers of illustrates lheir connection to the main the most likely candidate. Additional cenlral memory. 1171 anaphora resolution. Usually a center is In this discourse segment neither the the most likely candidate for syntactic, nor the semantic constraints can pronominalization. eliminate the ambiguity between "storage devices", "tape cassettes" or "record There are different views in literature discs" as antecedents for "their", and thus regarding the preferred candidate for a cannot turn up a plausible antecedent from center (focus). Sidner's algorithm among these candidates. A human reader ([Sidner 811), which is based on thematic would be in a better position since he roles, prefers the theme o[ the previous would be able to identify the central sentence as the focus of the current concept, which is a primary candidate for sentence. This view, in general, is pronominalization. Correct identification advocated also in ([Allen87]). PUNDIT, of the antecedent is possible on the basis in its current implementation, considers of the pronominal reference hypothesis: in the entire previous utterance to be the every sentence which contains one or potential focus ([Dahl&Ball 901). Finally, more pronouns must have one of its in the centering literature ([Brennan et al. pronouns refer to the center 1 of the 87]), the subject is generally considered previous sentence. Therefore, whenever to be preferred. We have found, we have to find a referent of a pronoun however, that there are many additional which is alone in the sentence, we have to interrelated factors which influence upon look for the centered clement in the the location of the center. previous sentence. Wc studied the "behaviour" of center in Following this hypothesis, and various computer science texts (30 recognizing "storage devices" as the different sources totally exceeding 1000 center, an anaphora resolution model pages) and the empirical observations would not have problems in picking up enabled us to develop efficient the center of the previous sentence sublanguage-dependent heuristics for ("storage devices") as antecedent for tracking the center in the sublanguage of "their". computer science. We summarize the most important conclusions as follows: We see now that the main problem which arises is the tracking of the center 1) Consider the primary candidates throughout the discourse segment. for center from the priority list: Certain ideas and algorithms for tracking subject, object, verb phrase. focus or center (e.g. [Brennan et al.87]) have been proposed, provided that one 2) Prefer the NP, representing a knows the focus or center of the first domain concept to the NPs, which sentence in the segment. However, they are not domain concepts. do not try to identify this center. Our approach determines the most probable 3) If the verb is a member of the center of the first sentence, and then Verb set = {discuss, present, tracks it all the way through the segment, illustrate, summarize, examine, correcting the proposed algorithm at each describe, define, show, check, step. develop, review, report, outline, consider, investigate, explore, assess, analyze, synthesize, study, TRACKING THE CENTER IN "['HE survey, deal, cover}, then consider SUBLANGUAGE OF COMPUTER the object as a most probable SCIENCE center. Identifying center can be very helpful in 4) If a verbal adjective is a member of" the Adj set = {defined, called, so-called}, then consider the NP 1Though "center" is ml uncrancc specific notion, they refer to as the probable center we refer to "sentence center", because in many cases the centers of the uttermmes a senlence may of the subsequent clause/current consist of, coincide. In a complex sentence, sentence. however, we distinguish also "clause centers" 1172 5) If the subject is "chapter", proposer module which identifies the "section", "table", or a personal most likely center. We must point out that pronoun - 'T', "we", "you",