A Proposition-Based Abstractive Summariser

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 567–578, Osaka, Japan, December 11-17 2016. A Proposition-Based Abstractive Summariser Yimai Fang, Haoyue Zhu, Ewa Muszy´nska, Alexander Kuhnle and Simone Teufel University of Cambridge Computer Laboratory 15 JJ Thomson Avenue Cambridge CB3 0FD, United Kingdom {yf261,hz315,emm68,aok25,sht25}@cam.ac.uk Abstract Abstractive summarisation is not yet common amongst today’s deployed and research systems. Most existing systems either extract sentences or compress individual sentences. In this paper, we present a summariser that works by a different paradigm. It is a further development of an existing summariser that has an incremental, proposition-based content selection process but lacks a natural language (NL) generator for the final output. Using an NL generator, we can now produce the summary text to directly reflect the selected propositions. Our evaluation compares textual quality of our system to the earlier preliminary output method, and also uses ROUGE to compare to various summarisers that use the traditional method of sentence extraction, followed by compression. Our results suggest that cutting out the middle-man of sentence extraction can lead to better abstractive summaries. 1 Introduction Abstraction, rather than extraction, is generally seen as the more desirable method of performing auto- matic summarisation. It implies generating an entirely new text from the information contained in the input text. Amongst its advantages is the promise of better information packaging: conveying more in- formation using less output text. But end-to-end abstractive summarisation is still uncommon in current systems, mostly due to the problems with NLP text analysis and knowledge representation that would plague any “deeper” method. Current research into abstractive summarisation therefore mainly concerns compressing individual sentences or merging sentences of similar content. Sentence compression using machine learning and/or constraint-solving methods is an active area of research. Common methods use syntactic, lexical and discourse-based features to determine which words should be dropped or para- phrased (Knight and Marcu, 2000; McDonald, 2006; Clarke and Lapata, 2008; Cohn and Lapata, 2007, 2008; Yoshikawa et al., 2012). More recently, neural language model has also been applied to the prob- lem (Rush et al., 2015). For multi-document summarisation, sentence fusion (Barzilay and McKeown, 2005) allows the synthesis of common information across documents in one sentence, using grammatical dependencies as a substitute for a semantic representation. These methods have the ability to produce high-quality output, but their sentence generation is a local step in the summarisation pipeline, i.e. iso- lated from the content selection, which is essentially global. The only exceptions we know are Martins and Smith’s (2009) system, and Nishikawa’s (2014) system for Japanese text, both of which optimise sentence selection jointly with sentence compression. A different paradigm for abstractive summarisation would be to avoid the “middle man” of sentence extraction altogether, and to operate over propositions instead. Such a system would first transform the text into sub-sentential semantic units, then perform summarisation operations over these (possibly performing inference and creating new semantic units), and finally use NL generation techniques to verbalise the semantic units that make up …

Content truncated to 500 words