Shift-Reduce CCG Parsing with a Dependency Model

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 218–227, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Shift-Reduce CCG Parsing with a Dependency Model Wenduan Xu University of Cambridge Computer Laboratory wx217@cam.ac.uk Stephen Clark University of Cambridge Computer Laboratory sc609@cam.ac.uk Yue Zhang Singapore University of Technology and Design yue zhang@sutd.edu.sg Abstract This paper presents the first dependency model for a shift-reduce CCG parser. Mod- elling dependencies is desirable for a num- ber of reasons, including handling the “spurious” ambiguity of CCG; fitting well with the theory of CCG; and optimizing for structures which are evaluated at test time. We develop a novel training tech- nique using a dependency oracle, in which all derivations are hidden. A challenge arises from the fact that the oracle needs to keep track of exponentially many gold- standard derivations, which is solved by integrating a packed parse forest with the beam-search decoder. Standard CCGBank tests show the model achieves up to 1.05 labeled F-score improvements over three existing, competitive CCG parsing models. 1 Introduction Combinatory Categorial Grammar (CCG; Steed- man (2000)) is able to derive typed dependency structures (Hockenmaier, 2003; Clark and Curran, 2007), providing a useful approximation to the un- derlying predicate-argument relations of “who did what to whom”. To date, CCG remains the most competitive formalism for recovering “deep” de- pendencies arising from many linguistic phenom- ena such as raising, control, extraction and coordi- nation (Rimell et al., 2009; Nivre et al., 2010). To achieve its expressiveness, CCG exhibits so-called “spurious” ambiguity, permitting many non-standard surface derivations which ease the recovery of certain dependencies, especially those arising from type-raising and composition. But this raises the question of what is the most suit- able model for CCG: should we model the deriva- tions, the dependencies, or both? The choice for some existing parsers (Hockenmaier, 2003; Clark and Curran, 2007) is to model derivations directly, restricting the gold-standard to be the normal-form derivations (Eisner, 1996) from CCGBank (Hock- enmaier and Steedman, 2007). Modelling dependencies, as a proxy for the se- mantic interpretation, fits well with the theory of CCG, in which Steedman (2000) argues that the derivation is merely a “trace” of the underlying syntactic process, and that the structure which is built, and predicated over when applying con- straints on grammaticality, is the semantic inter- pretation. The early dependency model of Clark et al. (2002), in which model features were defined over only dependency structures, was partly moti- vated by these theoretical observations. More generally, dependency models are desir- able for a number of reasons. First, modelling dependencies provides an elegant solution to the spurious ambiguity problem (Clark and Curran, 2007). Second, obtaining training data for de- pendencies is likely to be easier than for syn- tactic derivations, especially for incomplete data (Schneider et al., 2013). Clark and Curran (2006) show how the dependency model from Clark and Curran (2007) extends naturally to the partial- training case, and also how to obtain dependency data cheaply from …

Content truncated to 500 words