Improving the Inference of Implicit Discourse Relations via Classifying Explicit Discourse Connectives

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 799–808, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Improving the Inference of Implicit Discourse Relations via Classifying Explicit Discourse Connectives Attapol T. Rutherford Department of Computer Science Brandeis University Waltham, MA 02453, USA tet@brandeis.edu Nianwen Xue Department of Computer Science Brandeis University Waltham, MA 02453, USA xuen@brandeis.edu Abstract Discourse relation classification is an im- portant component for automatic discourse parsing and natural language understanding. The performance bottleneck of a discourse parser comes from implicit discourse rela- tions, whose discourse connectives are not overtly present. Explicit discourse connec- tives can potentially be exploited to collect more training data to collect more data and boost the performance. However, using them indiscriminately has been shown to hurt the performance because not all discourse con- nectives can be dropped arbitrarily. Based on this insight, we investigate the interaction be- tween discourse connectives and the discourse relations and propose the criteria for selecting the discourse connectives that can be dropped independently of the context without chang- ing the interpretation of the discourse. Extra training data collected only by the freely omis- sible connectives improve the performance of the system without additional features. 1 Introduction The analysis of discourse-level structure has re- ceived increasing attention from the field in recent years (Feng and Hirst, 2012; Patterson and Kehler, 2013; Li et al., 2014). Discourse-level analysis is typically concerned with relations between clauses and sentences, linguistic units that go beyond sen- tence boundaries. There are a few conceptions of the discourse structure representation of a text such as a tree (Mann and Thompson, 1988), or a graph (Wolf et al., 2005). In the work we describe here, we adopt the view of the Penn Discourse Treebank (PDTB) (Prasad et al., 2008), which views a text as a series of local discourse relations, each of which consists of a discourse connective as a predicate tak- ing two arguments. Syntactically, these two argu- ments are typically realized as clauses or sentences. The discourse connective (underlined) can either be explicit, as in (1), or implicit, as in (2): (1) [The city’s Campaign Finance Board has re- fused to pay Mr Dinkins $95,142 in matching funds]Arg1 because [his campaign records are incomplete]Arg2. (2) [So much of the stuff poured into its Austin, Texas, offices that its mail rooms there simply stopped delivering it]Arg1. Implicit=so [Now, thousands of mailers, catalogs and sales pitches go straight into the trash]Arg2. Determining the sense of an explicit discourse re- lation such as (1) is straightforward since “because” is a strong indicator that the relation between the two arguments is CONTINGENCY.CAUSE. This task effectively amounts to disambiguating the sense of discourse connective, which can be done with high accuracy (Pitler et al., 2008). However, in the absence of an explicit discourse connective, inferring the sense of a discourse rela- tion has proved to a very challenging task (Park and Cardie, 2012; Rutherford and Xue, …

Content truncated to 500 words