Extractive Summarization Using Multi-Task Learning with Document Classification

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2101–2110 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Extractive Summarization Using Multi-Task Learning with Document Classification Masaru Isonuma,Toru Fujino, Junichiro Mori, Yutaka Matsuo and Ichiro Sakata The University of Tokyo, Japan {isonuma, jmori, isakata}@ipr-ctr.t.u-tokyo.ac.jp, tr.fujino@scslab.k.u-tokyo.ac.jp, matsuo@weblab.t.u-tokyo.ac.jp Abstract The need for automatic document sum- marization that can be used for practi- cal applications is increasing rapidly. In this paper, we propose a general frame- work for summarization that extracts sen- tences from a document using externally related information. Our work is aimed at single document summarization using small amounts of reference summaries. In particular, we address document sum- marization in the framework of multi- task learning using curriculum learning for sentence extraction and document classi- fication. The proposed framework en- ables us to obtain better feature representa- tions to extract sentences from documents. We evaluate our proposed summarization method on two datasets: financial report and news corpus. Experimental results demonstrate that our summarizers achieve performance that is comparable to state- of-the-art systems. 1 Introduction With rapid increase in the volume of textual data that are available both online and offline, the need for automatic document summarization that can be implement in practical scenarios is in- creasing (Li et al., 2016; Chopra et al., 2016; Takase et al., 2016). Among the several sum- marization systems, extractive summarization ap- proaches (Erkan and Radev, 2004; McDonald, 2007; Wong et al., 2008) are widely used. These techniques identify and subsequently concatenate relevant sentences automatically from a docu- ment to create its summary while preserving its original information content. Such approaches are popular and widely used for practical appli- cations because they are computationally cost- effective and less complex. Extractive summariza- tion approaches based on neural network-based approaches (K˚ageb¨ack et al., 2014; Cao et al., 2015; Yin and Pei, 2015; Cao et al., 2016) have ad- vanced rapidly. Recently, an attentional encoder- decoder for extractive single-document summa- rization was proposed and its application to the news corpus was demonstrated (Cheng and Lap- ata, 2016; Nallapati et al., 2017). The neural network-based approaches rely heavily on large amounts of reference sum- maries for training neural models, and conse- quently, for tuning a large number of parameters. The reference summaries are manually or semi- automatically created in advance. Some existing studies employ parallel corpora as artificial ref- erence summaries (Woodsend and Lapata, 2010; Cheng and Lapata, 2016). However, preparing such large volumes of reference summaries man- ually is sometimes costly. Particularly, it is infea- sible for humans to create hundreds of thousands of reference summaries in cases where summa- rization requires domain-specific or expert knowl- edge. Such cases include financial reports, finan- cial and economic news (Filippova et al., 2009), and scientific articles (Parveen et al., 2016). A fundamental requirement in extractive sum- marization is the identification of salient sentences from a document, i.e., sentences that represent key subjects mentioned in the document. Such sub- jects are often described in the form …

Content truncated to 500 words