Generating Discourse Structures for Written Text

Generating Discourse Structures for Written Texts Huong LeThanh, Geetha Abeysinghe, and Christian Huyck School of Computing Science, Middlesex University The Burroughs, London, NW4 4BT, United Kingdom {H.Le, G.Abeysinghe, C.Huyck}@mdx.ac.uk Abstract This paper presents a system for automati- cally generating discourse structures from written text. The system is divided into two levels: sentence-level and text-level. The sentence-level discourse parser uses syntactic information and cue phrases to segment sen- tences into elementary discourse units and to generate discourse structures of sentences. At the text-level, constraints about textual adja- cency and textual organization are integrated in a beam search in order to generate best dis- course structures. The experiments were done with documents from the RST Discourse Treebank. It shows promising results in a rea- sonable search space compared to the dis- course trees generated by human analysts. 1 Introduction Many recent studies in Natural Language Proc- essing have paid attention to Rhetorical Structure Theory (RST) (Mann and Thompson 1988; Hovy 1993; Marcu 2000; Forbes et al. 2003), a method of structured description of text. Although rhe- torical structure has been found to be useful in many fields of text processing (Rutledge et al. 2000; Torrance and Bouayad-Agha 2001), only a few algorithms for implementing discourse ana- lyzers have been proposed so far. Most research in this field concentrates on specific discourse phenomena (Schiffrin 1987; Litman and Hirsch- berg 1990). The amount of research available in discourse segmentation is considered small; in discourse parsing it is even smaller. The difficulties in developing a discourse parser are (i) recognizing discourse relations between text spans and (ii) deriving discourse structures from these relations. Marcu (2000)’s parser is based on cue phrases, and therefore faces problems when cue phrases are not present in the text. This system can apply to unrestricted texts, but faces combinatorial explosion. The dis- advantage of Marcu’s approach is that it pro- duces a great number of trees during its process, which is the essential redundancy in computa- tion. As the number of relations increases, the number of possible discourse trees increases ex- ponentially. Forbes et al. (2003) have a different approach of implementing a discourse parser for a Lexi- calized Tree Adjoining Grammar (LTAG). They simplify discourse analysis by developing a grammar that uses cue phrases as anchors to con- nect discourse trees. Despite the potential of this approach for discourse analysis, the case of no cue phrase present in the text has not been fully investigated in their research. Polanyi et al. (2004) propose a far more complicated discourse system than that of Forbes et al. (2003), which uses syntactic, semantic and lexical rules. Polanyi et al. have proved that their approach can provide promising results, especially in text summariza- tion. In this paper, different factors were investi- gated to achieve a better discourse parser, in- cluding syntactic information, constraints about textual adjacency and textual organization. With a given text and its syntactic information, the search space in which well-structured discourse trees of a text are produced is minimized. The …

Content truncated to 500 words