Exploring Text Links for Coherent Multi-Document Summarization

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 213–223, Osaka, Japan, December 11-17 2016. Exploring Text Links for Coherent Multi-Document Summarization Xun Wang, Masaaki Nishino, Tsutomu Hirao, Katsuhito Sudoh and Masaaki Nagata NTT Communication Science Labortories Kyoto, 619-0237, Japan {wang.xun,nishino.masaaki,hirao.tsutomu sudoh.katsuhito,nagata.masaaki@lab.ntt.co.jp} Abstract Summarization aims to represent source documents by a shortened passage. Existing methods focus on the extraction of key information, but often neglect coherence. Hence the generated summaries suffer from a lack of readability. To address this problem, we have developed a graph-based method by exploring the links between text to produce coherent summaries. Our approach involves finding a sequence of sentences that best represent the key information in a coherent way. In contrast to the previous methods that focus only on salience, the proposed method addresses both coherence and informativeness based on textual linkages. We conduct experiments on the DUC2004 summarization task data set. A performance comparison reveals that the summaries generated by the proposed system achieve comparable results in terms of the ROUGE metric, and show improvements in readability by human evaluation. 1 Introduction Automatic summarization is extremely useful in this age of information overload. It provides readers with easier access to information without the labour of reading the source text. According to the number of documents dealt with, summarization falls into two categories: single document summarization and multi-document summarization. While they both aim to represent the source text using a shorten passage, the latter deals with a set of documents sharing the same topic. Based on the method adopted, existing approaches to summarization can be divided into two kinds: abstraction based or extraction based. The difference lies in the sentences they use to generate summaries: the former selects sentences (clauses, or other text units, hereafter we refer to all of them as sentences.) from source documents and the latter generates new sentences. Most existing summarization systems are extraction-based because abstraction- based methods require the use of natural language generation technology, which is still a growing field. This paper, without exception, also employs extraction-based methods and we focus on multi-documents summarization. Currently the extraction-based methods face some major challenges. One is informativeness, which means we need to maintain the important information of source documents in summaries. This is the focus of almost all research on summarization. Another challenge is presentation, namely that the ex- tracted text should be well presented, i.e., it should contain little redundancy and be coherent so as to be readily understandable. Previous work has addressed the problem of redundancy, and some successful solutions like Maximum Marginal Relevance (MMR) (Carbonell and Goldstein, 1998) have been pro- posed and widely adopted (e.g., (Li and Li, 2013)), but very few try to deal with coherence. Therefore the generated summaries generally suffer as regards readability and are very difficult to use for practical applications. In the report of the TAC 2011 summarization task (Owczarzak and Dang, 2011), it is stated that “in general, automatic summaries are better than baselines1, except Readability.” …

Content truncated to 500 words