Integrating Document Structure into a Multi-Document Summarizer

International Conference RANLP 2009 - Borovets, Bulgaria, pages 45–49 Integrating Document Structure into a Multi-Document Summarizer Aur´elien Bossard and Thierry Poibeau Laboratoire d’Informatique de Paris-Nord CNRS and Universit´e Paris 13 99, avenue Jean-Baptiste Cl´ement — F-93430 Villetaneuse {firstname.lastname}@lipn.univ-paris13.fr Abstract In this paper, we present a novel approach for automatic summarization. CBSEAS, the sys- tem implementing this approach, integrates a method to detect redundancy at its very core, in order to produce more expressive summaries than previous approaches. The evaluation of our system during TAC 2008 —the Text Analysis Conference— revealed that, even if our system performed well on blogs, it had some failings on news stories. A post-mortem analysis of the weaknesses of our original system showed the im- portance of text structure for automatic summa- rization, even in the case of short texts like news stories. We describe some ongoing work dealing with these issues and show that first experiments provide a significant improvement of the results. Keywords Multi-document Summarization; Text structure; Evaluation; Text Analysis Conference. 1 Introduction During the past decade, automatic summarization, supported by evaluation campaigns and a large re- search community, has shown fast and deep improve- ments. Indeed, the research in this domain is guided by strong industrial needs: fast processing despite ever increasing amount of data. We have developed a system called CBSEAS that integrates a new method to detect redundancy at its very core, in order to produce more expressive sum- maries than previous approaches. We have evaluated our system by participating in two tasks of TAC 2008 (the Text Analysis Conference): • Opinion Task (Summarizing opinions found in blogs); • Update Task (News stories summarization and detecting updates). We obtained very competitive results during TAC 2008 on the “Opinion Task”. However, our system did not rank as well on the “Update Task”. A post-mortem analysis of the weaknesses of our original system re- vealed the importance of text structure for automatic summarization, even in the case of short texts like news stories. Therefore, we will only focus on the “Update task” in this paper. We present our approach for automatic summarization and the first results of our current work dealing with the detection of document structure along with its integration for the production of summaries. The reader who wants to get information on the sys- tem we have developed for the Opinion task —for which we obtained among the best results— may refer to the system description in the TAC 2008 proceed- ings, see [1]. The rest of this paper is structured as follows: we first give a quick overview of the state of the art. We then describe our system, focusing on the most im- portant novel features implemented and on the results obtained for the TAC 2008 “Update” task. Lastly, we show that news stories structure is meaningful and we detail some preliminary techniques that improve the results. 2 Related Works Interest in creating automatic summaries began as soon as in the 1950s with the work by Luhn at IBM …

Content truncated to 500 words