Improving Combinatory Categorial Grammar Parse Reranking with Dependency Grammar Features

Proceedings of COLING 2012: Technical Papers, pages 1441–1458, COLING 2012, Mumbai, December 2012. Improving Combinatory Categorial Grammar Parse Reranking with Dependency Grammar Features Sunghwan Mac Kim1 Dominick N g2 Mark Johnson1 James R. Curran2 (1) Department of Computing, Macquarie University, Sydney, NSW, Australia 2109 (2) School of Information Technologies, University of Sydney, Sydney, NSW, Australia 2006 sunghwan.kim@students.mq.edu.au, dominick.ng@sydney.edu.au, mark.johnson@mq.edu.au, james.r.curran@sydney.edu.au ABSTRACT This paper presents a novel method of improving Combinatory Categorial Grammar (CCG) parsing using features generated from Dependency Grammar (DG) parses and combined using reranking. Different grammar formalisms have different strengths and different parsing models have consequently divergent views of the data. More specifically, dependency parsers are sensitive to linguistic generalisations that differ from the generalisations that the CCG parser is sensitive to, and which the reranker exploits to identify the parse most likely to be correct. We propose DG-derived reranking features, which are obtained by comparing dependencies from the CCG parser with DG dependencies, and demonstrate how they improve the performance of a CCG parser and reranker in a variety of settings. We record a final labeled F-score of 87.93% on section 23 of CCGbank, 0.5% and 0.35% improvements over the base parser (87.43%) and reranker (87.58%), respectively. KEYWORDS: Combinatory Categorial Grammar (CCG), Dependency Grammar (DG), Rerank- ing, Dependency Grammar-derived features, parsing, syntax. 1441 1 Introduction Reranking is the process of rescoring an n-best list with an external model, and it is an effective method for improving performance in NLP tasks. In parsing, rerankers are able to incorporate arbitrary global features from the entire parse tree that would be intractable in a base parser. More informative features can be considered in reranking as the entire parse tree is available, as opposed to the fragments considered in parsing. In this paper, we propose a simple method for improving the performance of the C&C Combi- natory Categorial Grammar (CCG) parser (Clark and Curran, 2007). We parse sentences using the C&C n-best parser and a 1-best dependency grammar (DG) parser, and generate DG-derived features by comparing the extracted dependencies from the C&C parser with the DG dependen- cies. We then incorporate the DG-derived features into the CCG reranker of Ng et al. (2010) to reorder the n-best CCG parses using the external parse information. We experiment with both the Maltparser (Nivre et al., 2007b) and the MSTparser (McDonald et al., 2005) as the DG parser. This is the first cross-formalism parser combination experiment for CCG parsing that we are aware of, combining the features and strengths of two different formalisms together. Previous work has shown that dependency parsers such as the Maltparser perform better on short-range dependencies (McDonald and Nivre, 2007), whereas the C&C parser deals with long-range dependencies more reliably (Clark et al., 2002; Rimell et al., 2009). Short-range dependency information has also been shown to improve parser accuracy (Chen et al., 2009). We show how our new DG-derived features substantially improve parser performance by 0.35% to 87.93%, and improve the accuracy of the C&C parser on both short and …

Content truncated to 500 words