Recursive Deep Models for Discourse Parsing

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2061–2069, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Recursive Deep Models for Discourse Parsing Jiwei Li1, Rumeng Li2 and Eduard Hovy3 1Computer Science Department, Stanford University, Stanford, CA 94305, USA 2School of EECS, Peking University, Beijing 100871, P.R. China 3Language Technology Institute, Carnegie Mellon University, Pittsburgh, PA 15213, USA jiweil@stanford.edu alicerumeng@foxmail.com ehovy@andrew.cmu.edu Abstract Text-level discourse parsing remains a challenge: most approaches employ fea- tures that fail to capture the intentional, se- mantic, and syntactic aspects that govern discourse coherence. In this paper, we pro- pose a recursive model for discourse pars- ing that jointly models distributed repre- sentations for clauses, sentences, and en- tire discourses. The learned representa- tions can to some extent learn the seman- tic and intentional import of words and larger discourse units automatically,. The proposed framework obtains comparable performance regarding standard discours- ing parsing evaluations when compared against current state-of-art systems. 1 Introduction In a coherent text, units (clauses, sentences, and larger multi-clause groupings) are tightly con- nected semantically, syntactically, and logically. Mann and Thompson (1988) define a text to be coherent when it is possible to describe clearly the role that each discourse unit (at any level of grouping) plays with respect to the whole. In a coherent text, no unit is completely isolated. Dis- course parsing tries to identify how the units are connected with each other and thereby uncover the hierarchical structure of the text, from which mul- tiple NLP tasks can benefit, including text sum- marization (Louis et al., 2010), sentence compres- sion (Sporleder and Lapata, 2005) or question- answering (Verberne et al., 2007). Despite recent progress in automatic discourse segmentation and sentence-level parsing (e.g., (Fisher and Roark, 2007; Joty et al., 2012; Sori- cut and Marcu, 2003), document-level discourse parsing remains a significant challenge. Recent attempts (e.g., (Hernault et al., 2010b; Feng and Hirst, 2012; Joty et al., 2013)) are still consid- erably inferior when compared to human gold- standard discourse analysis. The challenge stems from the fact that compared with sentence-level dependency parsing, the set of relations between discourse units is less straightforward to define. Because there are no clause-level ‘parts of dis- course’ analogous to word-level parts of speech, there is no discourse-level grammar analogous to sentence-level grammar. To understand how dis- course units are connected, one has to understand the communicative function of each unit, and the role it plays within the context that encapsulates it, taken recursively all the way up for the entire text. Manually developed features relating to words and other syntax-related cues, used in most of the re- cent prevailing approaches (e.g., (Feng and Hirst, 2012; Hernault et al., 2010b)), are insufficient for capturing such nested intentionality. Recently, deep learning architectures have been applied to various natural language processing tasks (for details see Section 2) and have shown the advantages to capture the relevant semantic and syntactic aspects of units in context. As word distributions are composed …

Content truncated to 500 words