Statistical Bistratal Dependency Parsing

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 561–569, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Statistical Bistratal Dependency Parsing Richard Johansson Department of Information Engineering and Computer Science University of Trento Trento, Italy johansson@disi.unitn.it Abstract We present an inexact search algorithm for the problem of predicting a two-layered dependency graph. The algorithm is based on a k-best version of the standard cubic- time search algorithm for projective de- pendency parsing, which is used as the backbone of a beam search procedure. This allows us to handle the complex non- local feature dependencies occurring in bistratal parsing if we model the interde- pendency between the two layers. We apply the algorithm to the syntactic– semantic dependency parsing task of the CoNLL-2008 Shared Task, and we obtain a competitive result equal to the highest published for a system that jointly learns syntactic and semantic structure. 1 Introduction Numerous linguistic theories assume a multistratal model of linguistic structure, such as a layer of surface syntax, deep syntax, and shallow seman- tics. Examples include Meaning–Text Theory (Mel’ˇcuk, 1988), Discontinuous Grammar (Buch- Kromann, 2006), Extensible Dependency Gram- mar (Debusmann et al., 2004), and the Functional Generative Description (Sgall et al., 1986) which forms the theoretical foundation of the Prague De- pendency Treebank (Hajiˇc, 1998). In the statistical NLP community, the most widely used grammatical resource is the Penn Treebank (Marcus et al., 1993). This is a purely syntactic resource, but we can also include this treebank in the category of multistratal resources since the PropBank (Palmer et al., 2005) and NomBank (Meyers et al., 2004) projects have an- notated shallow semantic structures on top of it. Dependency-converted versions of the Penn Tree- bank, PropBank and NomBank were used in the CoNLL-2008 Shared Task (Surdeanu et al., 2008), in which the task of the participants was to pro- duce a bistratal dependency structure consisting of surface syntax and shallow semantics. Producing a consistent multistratal structure is a conceptually and computationally complex task, and most previous methods have employed a purely pipeline-based decomposition of the task. This includes the majority of work on shallow se- mantic analysis (Gildea and Jurafsky, 2002, in- ter alia). Nevertheless, since it is obvious that syntax and semantics are highly interdependent, it has repeatedly been suggested that the problems of syntactic and semantic analysis should be carried out simultaneously rather than in a pipeline, and that modeling the interdependency between syn- tax and semantics would improve the quality of all the substructures. The purpose of the CoNLL-2008 Shared Task was to study the feasibility of a joint analysis of syntax and semantics, and while most partici- pating systems used a pipeline-based approach to the problem, there were a number of contribu- tions that attempted to take the interdependence between syntax and semantics into account. The top-performing system in the task (Johansson and Nugues, 2008) applied a very simple reranking scheme by means of a k-best syntactic output, similar to previous attempts (Gildea and Juraf- sky, …

Content truncated to 500 words