Stacking of Dependency and Phrase Structure Parsers

Proceedings of COLING 2012: Technical Papers, pages 849–866, COLING 2012, Mumbai, December 2012. Stacking of Dependency and Phrase Structure Parsers Richárd Farkas and Bernd Bohnet Institute for Natural Language Processing University of Stuttgart {farkas,bohnetbd}@ims.uni-stuttgart.de ABSTRACT We investigate the stacking of dependency and phrase structure parsers, i.e. we define features from the output of a phrase structure parser for a dependency parser and vice versa. Our features are based on the original form of the external parses and we also compare this approach to converting phrase structures to dependencies then applying standard stacking on the converted output. The proposed method provides high accuracy gains for both phrase structure and dependency parsing. With the features derived from the phrase structures, we achieved a gain of 0.89 percentage points over a state-of-the-art parser and reach 93.95 UAS, which is the highest reported accuracy score on dependency parsing of the Penn Treebank. The phrase structure parser obtains 91.72 F-score with the features derived from the dependency trees, and this is also competitive with the best reported PARSEVAL scores for the Penn Treebank. KEYWORDS: parsing, stacking, dependency parsing, phrase structure parsing. 849 1 Introduction Both phrase structure and dependency parsers have developed considerably in the last decade, cf. (Nivre et al., 2004; McDonald et al., 2005b; Charniak and Johnson, 2005; Huang, 2008). The development has taken rather different directions as phrase structure parsers and dependency parsers employ different techniques to parse sentences. Phrase structure parsers usually apply probabilistic context free grammars and focus on the relationships among phrases. Dependency parsers use edge factored models for parsing that primarily model the interaction between the a head word and a dependent word. Second order graph-based parsers consider in addition for the decision on a dependency edge the interaction with siblings and grandchildren. Phrase structure and dependency parsers have both shown to be efficient and to provide accurate parsing results. Different parsing approaches have different strengths on distinct linguistic constructions, cf. (Nivre and McDonald, 2008). In this paper, we exploit the difference of the representations of dependency and phrase structure parses and the divergence in the parsing techniques. We use features derived from the 1-best automatic phrase structure parse to augment the feature set of the dependency parser and vice versa. This way of combining parsers proved to be effective and provides a substantial accuracy gain for both parsing approaches. Our ultimate objective is to advance one parser by a second one. This motivation is different from previous approaches for combining phrase-structure and dependency parsers (cf. (Klein and Manning, 2003; Carreras et al., 2008; Rush et al., 2010)) which aimed to achieve a joint optimum of the two approaches. Our proposal has three advantages over previous work. (i) Our approach is simple to implement by defining new feature templates, yet very effective. (ii) It does not require a joint representation of different parse structures. (iii) The participating dependency and phrase-structure parsers can be easily replaced, and each can be developed and optimized independently. In our experiments both parsers …

Content truncated to 500 words