Automatically Determining a Proper Length for Multi-Document Summarization: A Bayesian Nonparametric Approach

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 736–746, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Automatically Determining a Proper Length for Multi-document Summarization: A Bayesian Nonparametric Approach Tengfei Ma and Hiroshi Nakagawa The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo {matf@r., nakagawa@}dl.itc.u-tokyo.ac.jp Abstract Document summarization is an important task in the area of natural language processing, which aims to extract the most important in- formation from a single document or a clus- ter of documents. In various summarization tasks, the summary length is manually de- fined. However, how to find the proper sum- mary length is quite a problem; and keeping all summaries restricted to the same length is not always a good choice. It is obvi- ously improper to generate summaries with the same length for two clusters of docu- ments which contain quite different quantity of information. In this paper, we propose a Bayesian nonparametric model for multi- document summarization in order to automat- ically determine the proper lengths of sum- maries. Assuming that an original document can be reconstructed from its summary, we describe the ”reconstruction” by a Bayesian framework which selects sentences to form a good summary. Experimental results on DUC2004 data sets and some expanded data demonstrate the good quality of our sum- maries and the rationality of the length deter- mination. 1 Introduction Text summarization is the process of generating a short version of a given text to indicate its main top- ics. As the number of documents on the web expo- nentially increases, text summarization has attracted increasing attention, because it can help people get the most important information within a short time. In most of the existing summarization systems, people need to first define a constant length to re- strict all the output summaries. However, in many cases it is improper to require all summaries are of the same length. Take the multi-document summa- rization as an example, generating the summaries of the same length for a 5-document cluster and a 50-document cluster is intuitively improper. More specifically, consider two different clusters of doc- uments: one cluster contains very similar articles which all focus on the same event at the same time; the other contains different steps of the event but each step has its own topics. The former cluster may need only one or two sentences to explain its infor- mation, while the latter needs to include more. Research on summary length dates back in the late 90s. Goldstein et al. (1999) studied the char- acteristics of a good summary (single-document summarization for news) and showed an empiri- cal distribution of summary length over document size. However, the length problem has been grad- ually ignored later, since researchers need to fix the length so as to estimate different summarization models conveniently. A typical instance is the Doc- ument Understanding Conferences (DUC)1, which provide authoritative evaluation for summarization systems. The DUC conferences collect news arit- cles as the input data and define …

Content truncated to 500 words