Dependency-based Syntactic–Semantic Analysis with PropBank and NomBank

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 183–187 Manchester, August 2008 Dependency-based Syntactic–Semantic Analysis with PropBank and NomBank Richard Johansson and Pierre Nugues Lund University, Sweden {richard, pierre}@cs.lth.se Abstract This paper presents our contribution in the closed track of the 2008 CoNLL Shared Task (Surdeanu et al., 2008). To tackle the problem of joint syntactic–semantic anal- ysis, the system relies on a syntactic and a semantic subcomponent. The syntactic model is a bottom-up projective parser us- ing pseudo-projective transformations, and the semantic model uses global inference mechanisms on top of a pipeline of clas- sifiers. The complete syntactic–semantic output is selected from a candidate pool generated by the subsystems. The system achieved the top score in the closed challenge: a labeled syntactic accu- racy of 89.32%, a labeled semantic F1 of 81.65, and a labeled macro F1 of 85.49. 1 Introduction: Syntactic–Semantic Analysis Intuitively, semantic interpretation should help syntactic disambiguation, and joint syntactic– semantic analysis has a long tradition in linguis- tic theory. This motivates a statistical modeling of the problem of finding a syntactic tree ˆysyn and a semantic graph ˆysem for a sentence x as maximiz- ing a function F that scores the joint syntactic– semantic structure: ⟨ˆysyn, ˆysem⟩ = arg max ysyn,ysem F(x, ysyn, ysem) The dependencies in the feature representation used to compute F determine the tractability of the search procedure needed to perform the maximiza- tion. To be able to use complex syntactic features c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. such as paths when predicting semantic structures, exact search is clearly intractable. This is true even with simpler feature representations – the problem is a special case of multi-headed dependency anal- ysis, which is NP-hard even if the number of heads is bounded (Chickering et al., 1994). This means that we must resort to a simplifica- tion such as an incremental method or a reranking approach. We chose the latter option and thus cre- ated syntactic and semantic submodels. The joint syntactic–semantic prediction is selected from a small list of candidates generated by the respective subsystems. 2 Syntactic Submodel We model the process of syntactic parsing of a sentence x as finding the parse tree ˆysyn = arg maxy F(x, y) that maximizes a scoring func- tion F. The learning problem consists of fitting this function so that the cost of the predictions is as low as possible according to a cost function ρ. In this work, we consider linear scoring functions of the following form: F(x, y) = w · Ψ(x, y) where Ψ(x, y) is a numeric feature representation of the pair (x, y) and w a vector of feature weights. We defined the syntactic cost ρ as the sum of link costs, where the link cost was 0 for a correct de- pendency link with a correct label, 0.5 for a correct link with an incorrect label, and 1 for an incorrect link. A widely …

Content truncated to 500 words