Multi-view and multi-task training of RST discourse parsers

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1903–1913, Osaka, Japan, December 11-17 2016. Multi-view and multi-task training of RST discourse parsers Chlo´e Braud CoAStaL Dep. of Computer Science University of Copenhagen braud@di.ku.dk Barbara Plank Computational Linguistics CLCG University of Groningen b.plank@rug.nl Anders Søgaard CoAStaL Dep. of Computer Science University of Copenhagen soegaard@di.ku.dk Abstract We experiment with different ways of training LSTM networks to predict RST discourse trees. The main challenge for RST discourse parsing is the limited amounts of training data. We combat this by regularizing our models using task supervision from related tasks as well as alternative views on discourse structures. We show that a simple LSTM sequential discourse parser takes advantage of this multi-view and multi-task framework with 12-15% error reductions over our baseline (depending on the metric) and results that rival more complex state-of-the-art parsers. 1 Introduction Documents are not just an arbitrary collection of text spans, but rather an ordered list of structures forming a discourse. Discourse structures describe the organization of documents in terms of discourse or rhetorical relations. For instance, the discourse relation CONDITION holds between the two discourse units (marked with square brackets) in example (1a) and a relation MANNER-MEANS holds between the segments in example (1b).1 (1) a. [The gain on the sale couldn’t be estimated] [until the “tax treatment has been determined.”] b. [On Friday, Datuk Daim added spice to an otherwise unremarkable address on Malaysia’s pro- posed budget for 1990] [by ordering the Kuala Lumpur Stock Exchange “to take appropriate action immediately” to cut its links with the Stock Exchange of Singapore.] Different theories of discourse structure exist. For instance, Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) analyzes texts as constituency trees covering entire documents. This theory has led to the RST Discourse Treebank (RST-DT) (Carlson et al., 2001) for English and the development of text- level discourse parsers (Hernault et al., 2010; Joty et al., 2012; Feng and Hirst, 2014; Ji and Eisenstein, 2014b). Such parsers have proven to be useful for several downstream applications (Taboada and Mann, 2006; Daum´e III and Marcu, 2009; Thione et al., 2004; Sporleder and Lapata, 2005; Louis et al., 2010; Bhatia et al., 2015; Burstein et al., 2003; Higgins et al., 2004). Another corpus has been annotated for discourse phenomena in English, the Penn Discourse Treebank (Prasad et al., 2008) (PDTB). In contrast to RST-DT, PDTB does not encode discourse as tree structures and documents are not fully covered. In this study we focus on the RST-DT, but among other things, we consider the question of whether the information in PDTB can be used to improve RST discourse parsers. Discourse parsing is known to be a hard task (Stede, 2011). It involves several complex and inter- acting factors, touching upon all layers of linguistic analysis, from syntax, semantics up to pragmatics. Consequently, also annotation is complex and time consuming, and hence available annotated corpora are sparse and limited in size. The aim of this paper …

Content truncated to 500 words