Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations

Abstract

We describe the design, development, and API for two discourse parsers for Rhetorical Structure Theory. The two parsers use the same underlying framework, but one uses features that rely on dependency syntax, produced by a fast shift-reduce parser, whereas the other uses a richer feature space, including both constituent-and dependency-syntax and coreference information, produced by the Stanford CoreNLP toolkit. Both parsers obtain state-of-the-art performance, and use a very simple API consisting of, minimally, two lines of Scala code. We accompany this code with a visualization library that runs the two parsers in parallel, and displays the two generated discourse trees side by side, which provides an intuitive way of comparing the two parsers. IntroductionThis paper describes the design and development of two practical parsers for Rhetorical Structure Theory (RST) discourse (Mann and Thompson, 1988). This work contributes to the already vast body of research on RST parsing (see, inter alia, Soricut and Marcu, 2003; Feng and Hirst, 2012; Joty et al., 2013, Joty et al., 2014) with the following:1. We propose two parsers that use constituentbased and dependency-based syntax, respectively. The underlying framework, other than the syntax-based features, is identical between the parsers, which permits a rigorous analysis of the impact of constituent and dependency syntax to RST parsing. We describe the parsers in Section 2 and empirically compare the impact of the two syntactic representations in Section 3. Our analysis indicates that both parsers achieve state-of-the-art performance. The parser based on dependency syntax performs marginally worse (by 0.1 F 1 points) but runs approximately 2.5 times faster than the parser based on constituent syntax. On average, the faster parser processes one document from the RST corpus in 2.3 seconds. Both parsers have been released as open-sourceScala code with a very simple API; consisting of, minimally, two lines of code. We discuss this API in Section 4.3. We also introduce a visualization tool that runs the two parsers in parallel, and displays the two generated discourse structures side by side. This allows users to directly compare the runtimes and outputs of the two parsers. This visualization tool will be the centerpiece of the proposed demo session. We summarize this tool in Section 5. The Two ParsersThe proposed parsing approach follows the architecture introduced by Hernault et al. (2010), andFeng and Hirst (2012). The parser first segments the text into elementary discourse units (EDUs) using an i.i.d. classifier that identifies which tokens end an EDU. Then the parser iteratively constructs the discourse tree (consisting of binary relations between discourse units) using a greedy bottom-up approach that interleaves two classifiers: the first de-1 tects which two adjacent discourse units are most likely to be connected given the current sequence of units; and the second labels the corresponding relation. The resulting discourse unit produced by the new rel...