Simultaneous Clustering and Noise Detection for Theme-based Summarization

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 491–499, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Simultaneous Clustering and Noise Detection for Theme-based Sum- marization 1,2Xiaoyan Cai, 2Renxian Zhang, 2Dehong Gao, 2Wenjie Li 1College of Information Engineering, Northwest A&F University xiaoyanc@mail.nwpu.edu.cn 2Department of Computing, The Hong Kong Polytechnic University {csxcai,csrzhang,csdgao,cswjli}@comp.polyu.edu.hk Abstract Multi-document summarization aims to pro- duce a concise summary that contains salient information from a set of source documents. Since documents often cover a number of topical themes with each theme represented by a cluster of highly related sentences, sentence clustering plays a pivotal role in theme-based summarization. Moreover, noting that real- world datasets always contain noises which inevitably degrade the clustering performance, we incorporate noise detection with spectral clustering to generate ordinary sentence clus- ters and one noise sentence cluster. We are also interested in making the theme-based summaries biased towards a user’s query. The effectiveness of the proposed approaches is demonstrated by both the cluster quality analysis and the summarization evaluation conducted on the DUC generic and query- oriented summarization datasets. 1 Introduction The exponential growth in the volume of docu- ments available on the Internet brings the prob- lem of finding out whether a single document can meet a user’s complex information need. In order to solve this problem, multi-document summarization, which reduces the size of docu- ments while preserves their important semantic content is highly demanded. Most of the summa- rization work done till date follow the sentence extraction framework, which ranks sentences according to various pre-specified criteria and selects the most salient sentences from the origi- nal documents to form summaries. In addition to sentence salience, the other fun- damental issues that must be concerned in sum- marization are information redundancy and in- formation diversity (Radev et al., 2002). When the given documents are all supposed to be about the same topic, they are very likely to repeat some important information in different docu- ments or in different places of the same docu- ment. Therefore, effectively recognizing the sen- tences with the same or very similar content is necessary for reducing redundancy and covering more diverse informative content in a summary. This is normally achieved by clustering highly related sentences into topical themes. Summaries can then be produced, e.g., by extracting the rep- resentative sentence(s) from each theme cluster. Thus, good sentence clusters are the guarantee of good summaries in theme-based summarization. It is also important to stress that the noise sen- tences are clearly observed in the DUC datasets, i.e., the benchmark datasets for use by the sum- marization community (Wei et al., 2009). Take the DUC2005 d301i document set, which talks about ‘International Organized Crime’, as an ex- ample. The sentence like ‘This well-educated, well-spoken, cosmopolitan businessman is laughing all the way.’ absolutely goes too far off the point, and it is considered as a noise sentence in the context of our study. The existence of noises will inevitably degrade the clustering per- formance. Noise …

Content truncated to 500 words