Peking: Building Semantic Dependency Graphs with a Hybrid Parser

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 927–931, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics Peking: Building Semantic Dependency Graphs with a Hybrid Parser Yantao Du, Fan Zhang, Xun Zhang, Weiwei Sun∗and Xiaojun Wan Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {duyantao,zhangxunah,ws,wanxiaojun}@pku.edu.cn zhangf717@gmail.com Abstract This paper is a description of our system for SemEval-2015 Task 18: Broad-Coverage Semantic Dependency Parsing. We imple- ment a hybrid parser which benefits from both transition-based and graph-based parsing ap- proaches. In particular, the tree approximation method is explored to take advantage of well- studied tree parsing techniques. Evaluation on multilingual data sets demonstrates that con- siderably good semantic analysis can be au- tomatically built by applying state-of-the-art data-driven parsing techniques. 1 Introduction Dependency grammar is a long-standing tradition that determines syntacto-semantic structures on the basis of word-to-word connections. It names a family of approaches to linguistic analysis that all share a commitment to typed relations between or- dered pairs of words. Partially due to the power- ful expressiveness of bi-lexical dependency struc- tures, the corresponding parsing problem has been widely studied especially in the last decade. The majority of these studies, however, only focus on tree-structured representations. Beyond tree-shaped structures, SemEval-2014 Task 8 (Oepen et al., 2014) seeked to stimulate the dependency parsing community to move towards more general graph processing. Quite a number of teams all over the world participated in this shared task, which sug- gests a growing community interest in parsing into graph-shaped dependency representations. ∗Email correspondence. SemEval-2015 Task 18 is a subsequent task of SemEval-2014 Task 8. Following several well- established syntactic theories, this task proposes us- ing graphs to represent semantics and provides high- quality annotations for three typologically different languages. We have developed a system, dubbed DZSW14 (Du et al., 2014) for the task last year. The system employed a hybrid architecture which benefits from both transition-based and graph-based parsing approaches. Evaluation on multiple English data sets provided by SemEval-2014 indicated that DZSW14 is able to obtain high-quality parsing re- sults. Following the key idea to employ hetero- geneous models to enhance hybrid parsing, we ex- tend DZSW14 by developing more tree approxima- tion models, namely the weighted tree approxima- tion models. Evaluation on multilingual data sets provided by this year’s task confirms the effective- ness of the techniques we have studied. In this paper, we first give an introduction of the architecture of the baseline system DZSW14. Then we demonstrate the weighted tree approximation models. Finally we show the experiment results on SemEval-2015 Task 18. The tree approximation sys- tem can be downloaded at http://www.icst. pku.edu.cn/lcwm/grass. 2 Baseline System: DZSW14 Our system is based on the system we constructed for SemEval-2014 Task 8. In this section we present a brief overview of its architecture. Refer to (Du et al., 2014) for more information. Inspired by the research on discriminative de- pendency tree parsing, DZSW14 employed a hy- brid parsing …

Content truncated to 500 words