Predicting Discourse Structure using Distant Supervision from Sentiment

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2306–2316, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2306 Predicting Discourse Structure using Distant Supervision from Sentiment Patrick Huber and Giuseppe Carenini Department of Computer Science University of British Columbia Vancouver, BC, Canada, V6T 1Z4 {huberpat, carenini}@cs.ubc.ca Abstract Discourse parsing could not yet take full ad- vantage of the neural NLP revolution, mostly due to the lack of annotated datasets. We pro- pose a novel approach that uses distant super- vision on an auxiliary task (sentiment classi- fication), to generate abundant data for RST- style discourse structure prediction. Our ap- proach combines a neural variant of multiple- instance learning, using document-level su- pervision, with an optimal CKY-style tree generation algorithm. In a series of experi- ments, we train a discourse parser (for only structure prediction) on our automatically gen- erated dataset and compare it with parsers trained on human-annotated corpora (news do- main RST-DT and Instructional domain). Re- sults indicate that while our parser does not yet match the performance of a parser trained and tested on the same dataset (intra-domain), it does perform remarkably well on the much more difficult and arguably more useful task of inter-domain discourse structure prediction, where the parser is trained on one domain and tested/applied on another one. 1 Introduction Discourse parsing is a fundamental NLP task known to enhance key downstream tasks, such as sentiment analysis (Bhatia et al., 2015; Nejat et al., 2017; Hogenboom et al., 2015), text classification (Ji and Smith, 2017) and summarization (Gerani et al., 2014). In essence, a discourse parser should reveal the structure underlying coherent text as postulated by a discourse theory, of which the two most popu- lar are Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) and PDTB (Prasad et al., 2008). In this paper, we focus on RST-style pars- ing, but the proposed approach is theory agnostic and could be applied to PDTB as well. The RST discourse theory assumes a complete hierarchical discourse tree for a given document, where leaf nodes are clause-like sentence frag- ments, called elementary-discourse-units (EDUs), while internal tree nodes are labelled with dis- course relations. In addition, each node is given a nuclearity attribute, which encodes the importance of the node in its local context. In the past decade, traditional, probabilistic approaches, such as Support Vector Machines (SVM) (Hernault et al., 2010; Ji and Eisenstein, 2014) and Conditional Random Fields (CRF) (Joty et al., 2015; Feng and Hirst, 2014), have dominated the field. More recently, neural ap- proaches (Braud et al., 2016; Li et al., 2016; Braud et al., 2017; Yu et al., 2018; Liu and La- pata, 2017, 2018) have been explored, with lim- ited success (Morey et al., 2017; Ferracane et al., 2019). The main reason why recent advances in deep learning have not enhanced discourse pars- ing to the same extend as they have revolutionized many other areas of …

Content truncated to 500 words