Evaluating a Focus-Based Approach to Anaphora Resolution

Evaluating a Focus-Based Approach to Anaphora Resolution* Saliha Azzam, Kevin Humphreys and Robert Gaizauskas {s. azzam, k. humphreys, r. gaizauskas}©dcs, shef. ac. uk Department of Computer Science, University of Sheffield Regent Court, Portobello Road Sheffield S1 4DP UK Abstract We present an approach to anaphora resolution based on a focusing algorithm, and implemented within an existing MUC (Message Understand- ing Conference) Information Extraction system, allowing quantitative evaluation against a sub- stantial corpus of annotated real-world texts. Extensions to the basic focusing mechanism can be easily tested, resulting in refinements to the mechanism and resolution rules. Results show that the focusing algorithm is highly sensitive to the quality of syntactic-semantic analyses, when compared to a simpler heuristic-based ap- proach. 1 Introduction Anaphora resolution is still present as a signi- ficant linguistic problem, both theoretically and practically, and interest has recently been re- newed with the introduction of a quantitative evaluation regime as part of the Message Under- standing Conference (MUC) evaluations of In- formation Extraction (IE) systems (Grishman and Sundheim, 1996). This has made it pos- sible to evaluate different (implementable) the- oretical approaches against sizable corpora of real-world texts, rather than the small collec- tions of artificial examples typically discussed in the literature. This paper describes an evaluation of a focus- based approach to pronoun resolution (not ana- phora in general), based on an extension of Sidner's algorithm (Sidner, 1981) proposed in (Azzam, 1996), with further refinements from development on real-world texts. The approach * This work was carried out in the context of the EU AVENTINUS project (Thumair, 1996), which aims to develop a multilingual IE system for drug enforcement, and including a language-independent coreference mech- anism (Azzam et al., 1998). is implemented within the general coreference mechanism provided by the LaSIE (Large Scale Information Extraction) system (Gaizauskas et al., 1995) and (Humphreys et al., 1998), Shef- field University's entry in the MUC-6 and 7 evaluations. 2 Focus in Anaphora Resolution The term focus, along with its many relations such as theme, topic, center, etc., reflects an in- tuitive notion that utterances in discourse are usually 'about' something. This notion has been put to use in accounts of numerous linguistic phenomena, but it has rarely been given a firm enough definition to allow its use to be evalu- ated. For anaphora resolution, however, stem- ming from Sidner's work, focus has been given an algorithmic definition and a set of rules for its application. Sidner's approach is based on the claim that anaphora generally refer to the cur- rent discourse focus, and so modelling changes in focus through a discourse will allow the iden- tification of antecedents. The algorithm makes use of several focus re- gisters to represent the current state of a dis- course: CF, the current focus; AFL, the altern- ate focus list, containing other candidate foci; and FS, the focus stack. A parallel structure to the CF, AF the actor focus, is also set to deal with agentive pronouns. The algorithm updates these registers after each sentence, …

Content truncated to 500 words