Combining Intra- and Multi-sentential Rhetorical Parsing for Document-level Discourse Analysis

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 486–496, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Combining Intra- and Multi-sentential Rhetorical Parsing for Document-level Discourse Analysis Shafiq Joty∗ sjoty@qf.org.qa Qatar Computing Research Institute Qatar Foundation Doha, Qatar Giuseppe Carenini, Raymond Ng, Yashar Mehdad {carenini, rng, mehdad}@cs.ubc.ca Department of Computer Science University of British Columbia Vancouver, Canada Abstract We propose a novel approach for develop- ing a two-stage document-level discourse parser. Our parser builds a discourse tree by applying an optimal parsing algorithm to probabilities inferred from two Con- ditional Random Fields: one for intra- sentential parsing and the other for multi- sentential parsing. We present two ap- proaches to combine these two stages of discourse parsing effectively. A set of empirical evaluations over two different datasets demonstrates that our discourse parser significantly outperforms the state- of-the-art, often by a wide margin. 1 Introduction Discourse of any kind is not formed by inde- pendent and isolated textual units, but by related and structured units. Discourse analysis seeks to uncover such structures underneath the surface of the text, and has been shown to be benefi- cial for text summarization (Louis et al., 2010; Marcu, 2000b), sentence compression (Sporleder and Lapata, 2005), text generation (Prasad et al., 2005), sentiment analysis (Somasundaran, 2010) and question answering (Verberne et al., 2007). Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), one of the most influential the- ories of discourse, represents texts by labeled hier- archical structures, called Discourse Trees (DTs), as exemplified by a sample DT in Figure 1. The leaves of a DT correspond to contiguous Elemen- tary Discourse Units (EDUs) (six in the exam- ple). Adjacent EDUs are connected by rhetori- cal relations (e.g., Elaboration, Contrast), form- ing larger discourse units (represented by internal ∗This work was conducted at the University of British Columbia, Vancouver, Canada. nodes), which in turn are also subject to this re- lation linking. Discourse units linked by a rhetori- cal relation are further distinguished based on their relative importance in the text: nucleus being the central part, whereas satellite being the peripheral one. Discourse analysis in RST involves two sub- tasks: discourse segmentation is the task of identi- fying the EDUs, and discourse parsing is the task of linking the discourse units into a labeled tree. While recent advances in automatic discourse segmentation and sentence-level discourse parsing have attained accuracies close to human perfor- mance (Fisher and Roark, 2007; Joty et al., 2012), discourse parsing at the document-level still poses significant challenges (Feng and Hirst, 2012) and the performance of the existing document-level parsers (Hernault et al., 2010; Subba and Di- Eugenio, 2009) is still considerably inferior com- pared to human gold-standard. This paper aims to reduce this performance gap and take discourse parsing one step further. To this end, we address three key limitations of existing parsers as follows. First, existing discourse parsers typically model the structure and the labels of a DT separately in a pipeline fashion, and also …

Content truncated to 500 words