Shallow Discourse Parsing with Conditional Random Fields

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1071–1079, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Shallow Discourse Parsing with Conditional Random Fields Sucheta Ghosh Richard Johansson Giuseppe Riccardi Sara Tonelli Department of Information Engineering and Computer Science, University of Trento FBK-IRST {ghosh,johansson,riccardi}@disi.unitn.it satonelli@fbk.eu Abstract Parsing discourse is a challenging natural language processing task. In this paper we take a data driven approach to iden- tify arguments of explicit discourse con- nectives. In contrast to previous work we do not make any assumptions on the span of arguments and consider parsing as a token-level sequence labeling task. We design the argument segmentation task as a cascade of decisions based on con- ditional random fields (CRFs). We train the CRFs on lexical, syntactic and seman- tic features extracted from the Penn Dis- course Treebank and evaluate feature com- binations on the commonly used test split. We show that the best combination of fea- tures includes syntactic and semantic fea- tures. The comparative error analysis in- vestigates the performance variability over connective types and argument positions. 1 Introduction Automatic discourse processing is considered one of the most challenging NLP tasks due to its de- pendency on lexical and syntactic features and on the inter-sentential relations. While automatic dis- course processing of structured documents or free text is still in its infancy, a number of applications of this technology in practical NLP systems have been proposed. For instance, Somasundaran et al. (2009) describe the use of discourse structure for opinion analysis. Other applications include con- versational analysis and dialog systems (Tonelli et al., 2010). In this work we divide the whole task of discourse parsing into two sub-tasks: connec- tive classification and argument segmentation and classification. Several successful attempts have already been made in the direction of automatic classification of connectives, while token-level argument segmentation has not been explored. Therefore in this paper we will focus on the seg- mentation and labeling of discourse arguments (Arg1 and Arg2) with full spans, as defined in the annotation protocol of the Penn Discourse Treebank (PDTB) (Prasad et al., 2008). We present a methodology that, given explicit discourse connectives, automatically extracts dis- course arguments by identifying Arg1 and Arg2 including the corresponding text spans. We call this approach shallow following Prasad et al. (2010) as opposed to tree-like representations of discourse, as in Rhetorical Structure Theory (Mann and Thompson, 1988). Indeed, we provide a flat chunk classification of discourse relations, building a non-hierarchical representation of the relations in a text. The discourse parser is designed as a cascade of argument-specific CRFs trained on different sets of lexical, syntactic and semantic features. The evaluation is made in terms of exact and partial match of arguments. The partial match condition may be useful in the case of noisy input or for ap- plications that do not require exact alignment. The paper is structured as follows: in Section 2 we present related work to discourse parsing. In Section 3 we detail argument …

Content truncated to 500 words