Learning Contextually Informed Representations for Linear-Time Discourse Parsing

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1289–1298 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Learning Contextually Informed Representations for Linear-Time Discourse Parsing Yang Liu and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB Yang.Liu2@ed.ac.uk,mlap@inf.ed.ac.uk Abstract Recent advances in RST discourse parsing have focused on two modeling paradigm- s: (a) high order parsers which jointly predict the tree structure of the discourse and the relations it encodes; or (b) linear- time parsers which are efficient but mostly based on local features. In this work, we propose a linear-time parser with a novel way of representing discourse constituents based on neural networks which takes into account global contextual information and is able to capture long-distance dependen- cies. Experimental results show that our parser obtains state-of-the art performance on benchmark datasets, while being effi- cient (with time complexity linear in the number of sentences in the document) and requiring minimal feature engineering. 1 Introduction The computational treatment of discourse phe- nomena has recently attracted much attention, due to their increasing importance for potential ap- plications. Knowing how text units can be com- posed into a coherent document and how they re- late to each other e.g., whether they express con- trast, cause, or elaboration, can usefully aid down- stream tasks such summarization (Yoshida et al., 2014), question answering (Chai and Jin, 2004), and sentiment analysis (Somasundaran, 2010). Rhetorical Structure Theory (RST, Mann and Thompson 1988), one of the most influential frameworks in discourse processing, represents texts by trees whose leaves correspond to Elemen- tary Discourse Units (EDUs) and whose nodes specify how these and larger units (e.g., multi- sentence segments) are linked to each other by rhetorical relations. Discourse units are further The projections are in the neighborhood of 50 cents a share to 75 cent, compared with a restated $1.65 a share a year earlier, when profit was $107.8 million on sales of $435.5 million. nucleus satellite [Cray Research Inc. said]e1 [it sold one of its newest and largest computer systems, the Cray Y-MP/832, to the United Kingdom Meteorological Office.]e2 [The system is the first]e3 [to be sold through the joint marketing agreement between Cray and Control Data Corp.]e4 [Only a few months ago, the 124-year-old securities firm seemed to be on the verge of a meltdown,]e1 [racked by internal squabbles and defections.]e2 [Its relationship with parent General Electric Co. had been frayed since a big Kidder insider-trading scandal two years ago.]e3 e1 e2 consequence e1:2 e3 list e1:3 Figure 1: Example text (bottom) composed of two sentences (three EDUs) and its RST discourse tree representation (top). characterized in terms of their importance in tex- t: nuclei denote central segments, whereas satel- lites denote peripheral ones. Figure 1 shows an example of a discourse tree representing two sen- tences with three EDUs (e1,e2, and e3). EDUs e1 and e2 are connected with a mononuclear relation (i.e., Consequence), where e1 is …

Content truncated to 500 words