Finding document topics for improving topic segmentation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 480–487, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Finding document topics for improving topic segmentation Olivier Ferret CEA LIST, LIC2M 18 route du Panorama, BP6 Fontenay aux Roses, F-92265 France ferreto@zoe.cea.fr Abstract Topic segmentation and identification are of- ten tackled as separate problems whereas they are both part of topic analysis. In this article, we study how topic identification can help to improve a topic segmenter based on word reiteration. We first present an unsu- pervised method for discovering the topics of a text. Then, we detail how these topics are used by segmentation for finding topical similarities between text segments. Finally, we show through the results of an evaluation done both for French and English the inter- est of the method we propose. 1 Introduction In this article, we address the problem of linear topic segmentation, which consists in segmenting doc- uments into topically homogeneous segments that does not overlap each other. This part of the Dis- course Analysis field has received a constant interest since the initial work in this domain such as (Hearst, 1994). One criterion for classifying topic segmen- tation systems is the kind of knowledge they de- pend on. Most of them only rely on surface features of documents: word reiteration in (Hearst, 1994; Choi, 2000; Utiyama and Isahara, 2001; Galley et al., 2003) or discourse cues in (Passonneau and Lit- man, 1997; Galley et al., 2003). As such systems do not require external knowledge, they are not sensi- tive to domains but they are limited by the type of documents they can be applied to: lexical reiteration is reliable only if concepts are not too frequently ex- pressed by several means (synonyms, etc.) and dis- course cues are often rare and corpus-specific. To overcome these difficulties, some systems make use of domain-independent knowledge about lexical cohesion: a lexical network built from a dic- tionary in (Kozima, 1993); a thesaurus in (Mor- ris and Hirst, 1991); a large set of lexical co- occurrences collected from a corpus in (Choi et al., 2001). To a certain extent, these lexical networks enable topic segmenters to exploit a sort of concept reiteration. However, their lack of any explicit topi- cal structure makes this kind of knowledge difficult to use when lexical ambiguity is high. The most simple solution to this problem is to ex- ploit knowledge about the topics that may occur in documents. Such topic models are generally built from a large set of example documents as in (Yam- ron et al., 1998), (Blei and Moreno, 2001) or in one component of (Beeferman et al., 1999). These sta- tistical topic models enable segmenters to improve their precision but they also restrict their scope. Hybrid systems that combine the approaches we have presented were also developed and illus- trated the interest of such a combination: (Job- bins and Evett, 1998) combined word recurrence, co-occurrences and a thesaurus; (Beeferman et al., 1999) relied …

Content truncated to 500 words