A Novel Discriminative Framework for Sentence-Level Discourse Analysis

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 904–915, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics A Novel Discriminative Framework for Sentence-Level Discourse Analysis Shafiq Joty and Giuseppe Carenini and Raymond T. Ng {rjoty, carenini, rng}@cs.ubc.ca Department of Computer Science University of British Columbia Vancouver, BC, V6T 1Z4, Canada Abstract We propose a complete probabilistic discrim- inative framework for performing sentence- level discourse analysis. Our framework com- prises a discourse segmenter, based on a bi- nary classifier, and a discourse parser, which applies an optimal CKY-like parsing algo- rithm to probabilities inferred from a Dynamic Conditional Random Field. We show on two corpora that our approach outperforms the state-of-the-art, often by a wide margin. 1 Introduction Automatic discourse analysis has been shown to be critical in several fundamental Natural Lan- guage Processing (NLP) tasks including text gener- ation (Prasad et al., 2005), summarization (Marcu, 2000b), sentence compression (Sporleder and Lap- ata, 2005) and question answering (Verberne et al., 2007). Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), one of the most influential theories of discourse, posits a tree representation of a discourse, known as a Discourse Tree (DT), as exemplified by the sample DT shown in Figure 1. The leaves of a DT correspond to contiguous atomic text spans, also called Elementary Discourse Units (EDUs) (three in the example). The adjacent EDUs are connected by a rhetorical relation (e.g., ELAB- ORATION), and the resulting larger text spans are recursively also subject to this relation linking. A span linked by a rhetorical relation can be either a NUCLEUS or a SATELLITE depending on how central the message is to the author. Discourse anal- ysis in RST involves two subtasks: (i) breaking the text into EDUs (known as discourse segmentation) and (ii) linking the EDUs into a labeled hierarchical tree structure (known as discourse parsing). Figure 1: Discourse structure of a sentence in RST-DT. Previous studies on discourse analysis have been quite successful in identifying what machine learn- ing approaches and what features are more useful for automatic discourse segmentation and parsing (Sori- cut and Marcu, 2003; Subba and Eugenio, 2009; du- Verle and Prendinger, 2009). However, all the pro- posed solutions suffer from at least one of the fol- lowing two key limitations: first, they make strong independence assumptions on the structure and the labels of the resulting DT, and typically model the construction of the DT and the labeling of the rela- tions separately; second, they apply a greedy, sub- optimal algorithm to build the structure of the DT. In this paper, we propose a new sentence-level discourse parser that addresses both limitations. The crucial component is a probabilistic discriminative parsing model, expressed as a Dynamic Conditional Random Field (DCRF) (Sutton et al., 2007). By representing the structure and the relation of each discourse tree constituent jointly and by explicitly capturing the sequential and hierarchical dependen- cies between constituents of a discourse tree, our DCRF …

Content truncated to 500 words