Text-level Discourse Dependency Parsing

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 25–35, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Text-level Discourse Dependency Parsing Sujian Li1 Liang Wang1 Ziqiang Cao1 Wenjie Li2 1 Key Laboratory of Computational Linguistics, Peking University, MOE, China 2 Department of Computing, The Hong Kong Polytechnic University, HongKong {lisujian,intfloat,ziqiangyeah}@pku.edu.cn cswjli@comp.polyu.edu.hk Abstract Previous researches on Text-level discourse parsing mainly made use of constituency structure to parse the whole document into one discourse tree. In this paper, we present the limitations of constituency based dis- course parsing and first propose to use de- pendency structure to directly represent the relations between elementary discourse units (EDUs). The state-of-the-art depend- ency parsing techniques, the Eisner algo- rithm and maximum spanning tree (MST) algorithm, are adopted to parse an optimal discourse dependency tree based on the arc- factored model and the large-margin learn- ing techniques. Experiments show that our discourse dependency parsers achieve a competitive performance on text-level dis- course parsing. 1 Introduction It is widely agreed that no units of the text can be understood in isolation, but in relation to their context. Researches in discourse parsing aim to acquire such relations in text, which is funda- mental to many natural language processing ap- plications such as question answering, automatic summarization and so on. One important issue behind discourse parsing is the representation of discourse structure. Rhe- torical Structure Theory (RST) (Mann and Thompson, 1988), one of the most influential discourse theories, posits a hierarchical genera- tive tree representation, as illustrated in Figure 1. The leaves of a tree correspond to contiguous text spans called Elementary Discourse Units (EDUs)1. The adjacent EDUs are combined into 1 EDU segmentation is a relatively trivial step in discourse parsing. Since our work focus here is not EDU segmenta- tion but discourse parsing. We assume EDUs are already known. the larger text spans by rhetorical relations (e.g., Contrast and Elaboration) and the larger text spans continue to be combined until the whole text constitutes a parse tree. The text spans linked by rhetorical relations are annotated as either nucleus or satellite depending on how sali- ent they are for interpretation. It is attractive and challenging to parse the whole text into one tree. Since such a hierarchical discourse tree is analogous to a constituency based syntactic tree except that the constituents in the discourse trees are text spans, previous researches have explored different constituency based syntactic parsing techniques (eg. CKY and chart parsing) and var- ious features (eg. length, position et al.) for dis- course parsing (Soricut and Marcu, 2003; Joty et al., 2012; Reitter, 2003; LeThanh et al., 2004; Baldridge and Lascarides, 2005; Subba and Di Eugenio, 2009; Sagae, 2009; Hernault et al., 2010b; Feng and Hirst, 2012). However, the ex- isting approaches suffer from at least one of the following three problems. First, it is difficult to design a set of production rules as in syntactic parsing, since there are no determinate genera- tive rules for the …

Content truncated to 500 words