Using Context Inference to Improve Sentence Ordering for Multi-document Summarization

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1055–1061, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Using Context Inference to Improve Sentence Ordering for Multi-document Summarization Peifeng Li Guangxi Deng Qiaoming Zhu School of Computer Science and Technology, Soochow University 1 Shizi St, Suzhou, China, 215006 {pfli,gxdeng,qmzhu}@suda.edu.cn Abstract In this paper, we propose a novel context inference-based approach for sentences ordering in multi-document summarization application. Our method first detects whether or not two summarization sentences should be adjacent according to the similarity between one summarization sentence and the context of the other one, and then it computes the reliability of the adjacent summarization sentences based on the similarity and their relative position. To be specific, the first sentence will be selected according to features of sentence, and the second sentence will be selected if and only if it has the maximum reliability with previous sentence. Evaluation result shows that our method outperforms the state-of-the-art ones on DUC2004 corpus. 1 Introduction Multi-document summarization is an automatic procedure aimed at extraction of information from multiple texts written about the same topic. Whereas, sentence ordering, the last task of multi-document summarization, will finally affect the quality of summarization. Furthermore, the task of sentence ordering in multi-document summarization is harder than that in single document summarization because multiple documents are created by different authors who have different writing styles and backgrounds. Therefore, no natural order of texts can be extracted as the basis of sentence ordering judgment. How to conduct an efficient and effective method for sentences ordering is a difficult but important task for both multi- document summarization and other text processing job, e.g. Question Answering. Currently, a variety of studies have been reported on sentence ordering. Some methods adopted chronological information (McKeown et al., 1999; Lin et al., 2001; Barzilay et al., 2002; Okazaki et al., 2004) while others learned the natural order of sentences from source documents or large corpus (Lapata, 2003; Barzilay and Lee, 2004; Nie et al., 2006; Ji and Nie, 2008). However, chronological information cannot be easily extracted from those non-news documents and constructing a large corpus also is not so easy. Furthermore, those results of all above methods are far from satisfactory. Therefore, how to achieve coherent summarization still is an issue for us. This paper proposes a novel method to infer the order of summarization sentences for multi- document summarization according to their context. We first judge whether or not two summarization sentences should be adjacent based on the similarity between one summarization sentence and the context of the other one, and then compute the reliability based on the similarity and relative position of the adjacent summarization. To be specific, the first sentence will be selected based on its features, and the second sentence will be selected if and only if it has the maximum adjacency reliability with previous sentence. The experimental results on DUC2004 corpus show that our method outperforms state-of-the-art ones. The reminder of this paper is …

Content truncated to 500 words