Simultaneous Ranking and Clustering of Sentences: A Reinforcement Approach to Multi-Document Summarization

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 134–142, Beijing, August 2010 Simultaneous Ranking and Clustering of Sentences: A Reinforcement Approach to Multi-Document Summarization 1Xiaoyan Cai, 1Wenjie Li, 1You Ouyang, 2Hong Yan 1Department of Computing, The Hong Kong Polytechnic University {csxcai,cswjli,csyouyang}@comp.polyu.edu.hk 2Department of Logistics and Maritime Studies, The Hong Kong Polytechnic University lgthyan@polyu.edu.hk Abstract Multi-document summarization aims to produce a concise summary that contains salient information from a set of source documents. In this field, sentence ranking has hitherto been the issue of most concern. Since documents often cover a number of topic themes with each theme represented by a cluster of highly related sentences, sentence clustering was recently explored in the literature in order to provide more informative summaries. Existing cluster- based ranking approaches applied clustering and ranking in isolation. As a result, the ranking performance will be inevitably influenced by the clustering result. In this paper, we propose a reinforcement approach that tightly integrates ranking and clustering by mutually and simultaneously updating each other so that the performance of both can be improved. Experimental results on the DUC datasets demonstrate its effectiveness and robustness. 1 Introduction Automatic multi-document summarization has drawn increasing attention in the past with the rapid growth of the Internet and information explosion. It aims to condense the original text into its essential content and to assist in filtering and selection of necessary information. So far extractive summarization that directly extracts sentences from documents to compose summaries is still the mainstream in this field. Under this framework, sentence ranking is the issue of most concern. Though traditional feature-based ranking approaches and graph-based approaches employed quite different techniques to rank sentences, they have at least one point in common, i.e., all of them focused on sentences only, but ignored the information beyond the sentence level (referring to Figure 1(a)). Actually, in a given document set, there usually exist a number of themes (or topics) with each theme represented by a cluster of highly related sentences (Harabagiu and Lacatusu, 2005; Hardy et al., 2002). These theme clusters are of different size and especially different importance to assist users in understanding the content in the whole document set. The cluster level information is supposed to have foreseeable influence on sentence ranking. Figure 1. Ranking vs. Clustering In order to enhance the performance of summarization, recently cluster-based ranking approaches were explored in the literature (Wan and Yang, 2006; Sun et al, 2007; Wang et al, 2008a,b; Qazvinian and Radev, 2008). Normally these approaches applied a clustering algorithm to obtain the theme clusters first and then ranked the sentences within each cluster or by exploring the interaction between sentences and obtained clusters (referring to Figure 1(b)). In other words, clustering and ranking are regarded as two independent processes in these approaches although the cluster-level information has been incorporated into the sentence ranking process. As a result, Ranking Ranking Clustering Ranking Clustering (a) (b) (c) 134 the ranking performance is influenced by the clustering result. inevitably To …

Content truncated to 500 words