Discourse Chunking and its Application to Sentence Compression

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 257–264, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Discourse Chunking and its Application to Sentence Compression Caroline Sporleder and Mirella Lapata School of Informatics University of Edinburgh 2 Buccleuch Place Edinburgh EH8 9LW, UK {csporled,mlap}@inf.ed.ac.uk Abstract In this paper we consider the problem of analysing sentence-level discourse struc- ture. We introduce discourse chunking (i.e., the identification of intra-sentential nucleus and satellite spans) as an al- ternative to full-scale discourse parsing. Our experiments show that the proposed modelling approach yields results com- parable to state-of-the-art while exploit- ing knowledge-lean features and small amounts of discourse annotations. We also demonstrate how discourse chunking can be successfully applied to a sentence com- pression task. 1 Introduction The computational treatment of discourse phenom- ena has recently attracted much attention, partly due to their increasing importance for potential appli- cations. In summarisation, for example, the extrac- tion of sentences to include in a summary crucially depends on their rhetorical status (Marcu, 2000; Teufel and Moens, 2002); one might want to extract contrastive or explanatory statements while omit- ting sentences that contain background information. In information extraction, discourse-level knowl- edge can be used to identify co-referring events (Humphreys et al., 1997) and to determine their tem- poral order. Discourse processing could further en- hance question answering systems by interpreting the user’s question either in isolation or in the con- text of preceding questions (Chai and Jing, 2004). Discourse analysis is often viewed as a parsing task. Rhetorical Structure Theory (RST, Mann and Thomson, 1988), one of the most influential frame- works in discourse processing, represents texts by trees whose leaves correspond to elementary dis- course units (edus) and whose nodes specify how these and larger units (e.g., multi-sentence seg- ments) are linked to each other by rhetorical rela- tions (e.g., Contrast, Elaboration). Discourse units are further characterised in terms of their text im- portance: nuclei denote central segments, whereas satellites denote peripheral ones. Recent advances in discourse modelling have greatly benefited from the availability of resources annotated with discourse-level information such as the RST Discourse Treebank (RST-DT, Carlson et al., 2002). Even though discourse parsing at the document-level still poses a significant challenge to data-driven methods, sentence-level discourse mod- els (e.g., Soricut and Marcu, 2003) trained on the RST-DT have attained accuracies comparable to hu- man performance. The availability of discourse an- notations is partly responsible for the success of these models. Another important reason is the devel- opment of robust syntactic parsers (e.g., Charniak, 2000) that can be used to provide critical structural and lexical information to the discourse parser. Un- fortunately, discourse annotated corpora are largely absent for languages other than English. Further- more, reliance on syntactic parsing renders dis- course parsing practically impossible for languages for which state-of-the-art parsers are unavailable. In this paper we propose discourse chunking as an alternative to discourse parsing. Analogous to sen- tence chunking, discourse chunking is an interme- diate …

Content truncated to 500 words