A New Approach to Automatic Document Summarization

A New Approach to Automatic Document Summarization Xiaofeng Wu National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences Beijing, China xfwu@nlpr.ia.ac.cn Chengqing Zong National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences Beijing, China cqzong@nlpr.ia.ac.cn Abstract In this paper we propose a new approach based on Sequence Segmentation Models (SSM) to the extractive document summa- rization, in which summarizing is regarded as a segment labeling problem. Comparing with the previous work, the difference of our approach is that the employed features are obtained not only from the sentence level, but also from the segment level. In our approach, the semi-Markov CRF model is employed for segment labeling. The pre- liminary experiments have shown that the approach does outperform all other tradi- tional supervised and unsupervised ap- proaches to document summarization. 1 Introduction Document summarization has been a rapidly evolving subfield of Information Retrieval (IR) since (Luhn, 1958). A summary can be loosely defined as a text that is produced from one or more texts and conveys important information of the original text(s). Usually it is no longer than half of the original text(s) or, significantly less (Radev et al., 2002). Recently, many evaluation competitions (like the Document Understanding Conference DUC “http://duc.nist.gov”, in the style of NIST’s TREC), provided some sets of training corpus. It is obvious that, in the age of information explosion, document summarization will be greatly helpful to the internet users; besides, the techniques it uses can also find their applications in speech tech- niques and multimedia document retrieval, etc. The approach to summarizing can be catego- rized in many ways. Some of them are: 1) indica- tive, informative and evaluative, according to func- tionality; 2) single-document and multi-document, according to the amount of input documents; 3) generic and query-oriented, according to applica- tions. Yet the taxonomy currently widely em- ployed is to categorize summarization into abstrac- tive and extractive. According to (Radev et al., 2002), all methods that are not explicitly extractive are categorized as abstractive. These approaches include ontological information, information fusion, and compression. Abstract-based summarization never goes beyond conceptual stage, though ever since the dawn of summarization it has been argued as an alternative for its extract-based counterpart. On the other hand, extractive summarization is still attracting a lot of researchers (Yeh et al., 2005) (Daum´e III and Marcu, 2006) and many practical systems, say, MEAD “http://www.summarization.com/mead/”, have been produced. Using supervised or unsuper- vised machine learning algorithms to extract sen- tences is currently the mainstream of the extractive summarization. However, all pervious methods focus on obtaining features from the sentence gra- nularity. In this paper we focus on generating summariza- tion by using a supervised extractive approach in which the features are obtained from a larger gra- nularity, namely segment. The remainder of the paper is organized as follows: Section 2 introduces the related work concerning the extract-based summarization. Section 3 describes our motiva- tions. Our experiments and results are given in Section 4, and Section 5 draws …

Content truncated to 500 words