Improving Multi-documents Summarization by Sentence Compression based on Expanded Constituent Parse Trees

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 691–701, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Improving Multi-documents Summarization by Sentence Compression based on Expanded Constituent Parse Trees Chen Li1, Yang Liu1, Fei Liu2, Lin Zhao3, Fuliang Weng3 1 Computer Science Department, The University of Texas at Dallas Richardson, TX 75080, USA 2 School of Computer Science, Carnegie Mellon University Pittsburgh, PA 15213, USA 3 Research and Technology Center, Robert Bosch LLC Palo Alto, California 94304, USA {chenli,yangl@hlt.utdallas.edu} {feiliu@cs.cmu.edu} {lin.zhao,fuliang.weng@us.bosch.com} Abstract In this paper, we focus on the problem of using sentence compression techniques to improve multi-document summariza- tion. We propose an innovative sentence compression method by considering every node in the constituent parse tree and de- ciding its status – remove or retain. In- teger liner programming with discrimina- tive training is used to solve the problem. Under this model, we incorporate various constraints to improve the linguistic qual- ity of the compressed sentences. Then we utilize a pipeline summarization frame- work where sentences are first compressed by our proposed compression model to ob- tain top-n candidates and then a sentence selection module is used to generate the final summary. Compared with state-of- the-art algorithms, our model has simi- lar ROUGE-2 scores but better linguistic quality on TAC data. 1 Introduction Automatic summarization can be broadly divided into two categories: extractive and abstractive summarization. Extractive summarization focuses on selecting salient sentences from the document collection and concatenating them to form a sum- mary; while abstractive summarization is gener- ally considered more difficult, involving sophisti- cated techniques for meaning representation, con- tent planning, surface realization, etc. There has been a surge of interest in recent years on generating compressed document summaries as a viable step towards abstractive summarization. These compressive summaries often contain more information than sentence-based extractive sum- maries since they can remove insignificant sen- tence constituents and make space for more salient information that is otherwise dropped due to the summary length constraint. Two general strate- gies have been used for compressive summariza- tion. One is a pipeline approach, where sentence- based extractive summarization is followed or pro- ceeded by sentence compression (Lin, 2003; Zajic et al., 2007; Vanderwende et al., 2007; Wang et al., 2013). Another line of work uses joint compres- sion and summarization. Such methods have been shown to achieve promising performance (Daum´e, 2006; Chali and Hasan, 2012; Almeida and Mar- tins, 2013; Qian and Liu, 2013), but they are typi- cally computationally expensive. In this study, we propose an innovative sen- tence compression model based on expanded con- stituent parse trees. Our model uses integer lin- ear programming (ILP) to search the entire space of compression, and is discriminatively trained. It is built based on the discriminative sentence compression model from (McDonald, 2006) and (Clarke and Lapata, 2008), but our method uses an expanded constituent parse tree rather than only the leaf nodes in previous work. Therefore we can extract rich features for every node …

Content truncated to 500 words