Improving LSA-based Summarization with Anaphora Resolution

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 1–8, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Improving LSA-based Summarization with Anaphora Resolution Josef Steinberger University of West Bohemia Univerzitni 22, Pilsen 30614, Czech Republic jstein@kiv.zcu.cz Mijail A. Kabadjov University of Essex Wivenhoe Park, Colchester CO4 3SQ, United Kingdom malexa@essex.ac.uk Massimo Poesio University of Essex Wivenhoe Park, Colchester CO4 3SQ, United Kingdom poesio@essex.ac.uk Olivia Sanchez-Graillet University of Essex Wivenhoe Park, Colchester CO4 3SQ, United Kingdom osanch@essex.ac.uk Abstract We propose an approach to summarization exploiting both lexical information and the output of an automatic anaphoric re- solver, and using Singular Value Decom- position (SVD) to identify the main terms. We demonstrate that adding anaphoric information results in significant perfor- mance improvements over a previously developed system, in which only lexical terms are used as the input to SVD. How- ever, we also show that how anaphoric in- formation is used is crucial: whereas using this information to add new terms does re- sult in improved performance, simple sub- stitution makes the performance worse. 1 Introduction Many approaches to summarization can be very broadly characterized as TERM-BASED: they at- tempt to identify the main ‘topics,’ which gen- erally are TERMS, and then to extract from the document the most important information about these terms (Hovy and Lin, 1997). These ap- proaches can be divided again very broadly in ‘lex- ical’ approaches, among which we would include LSA-based approaches, and ‘coreference-based’ ap- proaches . Lexical approaches to term-based sum- marization use lexical relations to identify cen- tral terms (Barzilay and Elhadad, 1997; Gong and Liu, 2002); coreference- (or anaphora-) based ap- proaches (Baldwin and Morton, 1998; Boguraev and Kennedy, 1999; Azzam et al., 1999; Bergler et al., 2003; Stuckardt, 2003) identify these terms by run- ning a coreference- or anaphoric resolver over the text.1 We are not aware, however, of any attempt to use both lexical and anaphoric information to iden- tify the main terms. In addition, to our knowledge no authors have convincingly demonstrated that feed- ing anaphoric information to a summarizer signif- icantly improves the performance of a summarizer using a standard evaluation procedure (a reference corpus and baseline, and widely accepted evaluation measures). In this paper we compare two sentence extraction- based summarizers. Both use Latent Semantic Analysis (LSA) (Landauer, 1997) to identify the main terms of a text for summarization; however, the first system (Steinberger and Jezek, 2004), dis- cussed in Section 2, only uses lexical information to identify the main topics, whereas the second sys- tem exploits both lexical and anaphoric information. This second system uses an existing anaphora reso- lution system to resolve anaphoric expressions, GUI- TAR (Poesio and Kabadjov, 2004); but, crucially, two different ways of using this information for summarization were tested. (Section 3.) Both sum- marizers were tested over the CAST corpus (Orasan et al., 2003), as discussed in Section 4, and sig- 1The terms ’anaphora resolution’ and ’coreference resolu- tion’ have been variously defined …

Content truncated to 500 words