Abstractive Multi-Document Summarization via Phrase Selection and Merging

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1587–1597, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Abstractive Multi-Document Summarization via Phrase Selection and Merging∗ Lidong Bing§ Piji Li♮ Yi Liao♮ Wai Lam♮ Weiwei Guo† Rebecca J. Passonneau‡ §Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA USA ♮Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong †Yahoo Labs, Sunnyvale, CA, USA ‡Center for Computational Learning Systems, Columbia University, New York, NY, USA §lbing@cs.cmu.edu, ♮{pjli, yliao, wlam}@se.cuhk.edu.hk †wguo@yahoo-inc.com, ‡becky@ccls.columbia.edu Abstract We propose an abstraction-based multi- document summarization framework that can construct new sentences by exploring more fine-grained syntactic units than sen- tences, namely, noun/verb phrases. Dif- ferent from existing abstraction-based ap- proaches, our method first constructs a pool of concepts and facts represented by phrases from the input documents. Then new sentences are generated by selecting and merging informative phrases to max- imize the salience of phrases and mean- while satisfy the sentence construction constraints. We employ integer linear op- timization for conducting phrase selection and merging simultaneously in order to achieve the global optimal solution for a summary. Experimental results on the benchmark data set TAC 2011 show that our framework outperforms the state-of- the-art models under automated pyramid evaluation metric, and achieves reasonably well results on manual linguistic quality evaluation. 1 Introduction Existing multi-document summarization (MDS) methods fall in three categories: extraction-based, compression-based and abstraction-based. Most ∗ The work described in this paper is substan- tially supported by grants from the Research and De- velopment Grant of Huawei Technologies Co. Ltd (YB2013090068/TH138232) and the Research Grant Coun- cil of the Hong Kong Special Administrative Region, China (Project Codes: 413510 and 14203414). The work was done when Weiwei Guo was in Columbia Uni- versity summarization systems adopt the extraction- based approach which selects some original sen- tences from the source documents to create a short summary (Erkan and Radev, 2004; Wan et al., 2007). However, the restriction that the whole sen- tence should be selected potentially yields some overlapping information in the summary. To this end, some researchers apply compression on the selected sentences by deleting words or phrases (Knight and Marcu, 2000; Lin, 2003; Zajic et al., 2006; Harabagiu and Lacatusu, 2010; Li et al., 2015), which is the compression-based method. Yet, these compressive summarization models cannot merge facts from different source sentences, because all the words in a summary sentence are solely from one source sentence. In fact, previous investigations show that human-written summaries are more abstractive, which can be regarded as a result of sentence ag- gregation and fusion (Cheung and Penn, 2013; Jing and McKeown, 2000). Some works, albeit less popular, have studied abstraction-based ap- proach that can construct a sentence whose frag- ments come from different source sentences. One important work developed by Barzilay and McK- eown (2005) employed sentence fusion, followed by (Filippova and Strube, 2008; Filippova, 2010). These works first conduct clustering …

Content truncated to 500 words