A Data-Driven, Factorization Parser for CCG Dependency Structures

A Data-Driven, Factorization Parser for CCG Dependency Structures Yantao Du, Weiwei Sun∗and Xiaojun Wan Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {ws,duyantao,wanxiaojun}@pku.edu.cn Abstract This paper is concerned with building CCG-grounded, semantics-oriented deep dependency structures with a data-driven, factorization model. Three types of fac- torization together with different higher- order features are designed to capture different syntacto-semantic properties of functor-argument dependencies. Integrat- ing heterogeneous factorizations results in intractability in decoding. We pro- pose a principled method to obtain opti- mal graphs based on dual decomposition. Our parser obtains an unlabeled f-score of 93.23 on the CCGBank data, resulting in an error reduction of 6.5% over the best published result. which yields a signifi- cant improvement over the best published result in the literature. Our implementa- tion is available at http://www.icst. pku.edu.cn/lcwm/grass. 1 Introduction Combinatory Categorial Grammar (CCG; Steed- man, 2000) is a linguistically expressive gram- mar formalism which has a transparent yet el- egant interface between syntax and semantics. By assigning each lexical category a dependency interpretation, we can derive typed dependency structures from CCG derivations (Clark et al., 2002), providing a useful approximation to the underlying meaning representations. To date, CCG parsers are among the most competitive sys- tems for generating such deep bi-lexical depen- dencies that appropriately encode a wide range of local and non-local syntacto-semantic infor- mation (Clark and Curran, 2007a; Bender et al., 2011). Such semantic-oriented dependency struc- tures have been shown very helpful for NLP ap- ∗Email correspondence. plications e.g. Question Answering (Reddy et al., 2014). Traditionally, CCG graphs are generated as a by-product by grammar-guided parsers (Clark and Curran, 2007b; Fowler and Penn, 2010). The main challenge is that a deep-grammar-guided model usually can only produce limited coverage and corresponding parsing algorithms is of relatively high complexity. Robustness and efficiency, thus, are two major problems for handling practical tasks. To increase the applicability of such parsers, lexical or syntactic pruning has been shown nec- essary (Clark and Curran, 2004; Matsuzaki et al., 2007; Sagae et al., 2007; Zhang and Clark, 2011). In the past decade, the techniques for data- driven dependency parsing has made a great progress (McDonald et al., 2005a,b; Nivre et al., 2004; Torres Martins et al., 2009; Koo et al., 2010). The major advantage of the data-driven architecture is complementary to the grammar- driven one. On one hand, data-driven approaches make essential uses of machine learning from lin- guistic annotations and are flexible to produce analysis for arbitrary sentences. On the other hand, without hard constraints, parsing algorithms for spanning specific types of graphs, e.g. projec- tive (Eisner, 1996) and 1-endpoint-crossing trees (Pitler et al., 2013), can be of low complexity. This paper proposes a new data-driven depen- dency parser that efficiently produces globally op- timal CCG dependency graphs according to a dis- criminative, factorization model. The design of the factorization is motivated by three essential properties of the CCG dependencies. First, all ar- guments associated with the same predicate are highly …

Content truncated to 500 words