Zhijun Wu: Chinese Semantic Dependency Parsing with Third-Order Features

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 430–434, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Zhijun Wu: Chinese Semantic Dependency Parsing with Third-Order Features Zhijun Wu Xuan Wang Xinxin Li Computer Application Research Center School of Computer Science and Technology Harbin Institute of Technology Shenzhen Graduate School Shenzhen 518055, P.R.China mattwu305@gmail.com wangxuan@insun.hit.edu.cn lixin2@gmail.com Abstract This paper presents our system participated on SemEval-2012 task: Chinese Semantic De- pendency Parsing. Our system extends the second-order MST model by adding two third-order features. The two third-order fea- tures are grand-sibling and tri-sibling. In the decoding phase, we keep the k best results for each span. After using the selected third-order features, our system presently achieves LAS of 61.58% ignoring punctuation tokens which is 0.15% higher than the result of purely second-order model on the test dataset. 1 Introduction Recently, semantic role labeling (SRL) has been a hot research topic. CoNLL shared tasks for joint parsing for syntactic and semantic dependencies both in the year 2008 and 2009, cf. (Surdeanu et al., 2008; Hajič et al., 2009; Bohnet, 2009). Same shared tasks in SemEval-2007 (Sameer S., 2007). The SRL is traditionally implemented as two sub- tasks, argument identification and classification. However, there are some problems for the seman- tic representation method used by the semantic role labeling. For example, the SRL only considers the predicate-argument relations and ignores the rela- tions between a noun and its modifier, the meaning of semantic roles is related with special predicates. In order to overcome those problems, semantic dependency parsing (SDP) is introduced. Semantic dependencies express semantic links between pre- dicates and arguments and represent relations be- tween entities and events in text. The SDP is a kind of dependency parsing, and its task is to build a dependency structure for an input sentence and to label the semantic relation between a word and its head. However, semantic relations are different from syntactic relations, such as position indepen- dent. Table 1 shows the position independent of semantic relations for the sentence XiaoMing hit XiaoBai with a book today. Today, XiaoMing hit XiaoBai with a book. XiaoBai was hit by XiaoMing today with a book. With a book, XiaoMing hit XiaoBai today. XiaoMing hit XiaoBai with a book today. Table 1: An example position dependency Although semantic relations are different from syntactic relations, yet they are identical in the de- pendency tree. That means the methods used in syntactic dependency parsing can also be applied in SDP. Two main approaches to syntactic dependency paring are Maximum Spanning Tree (MST) based dependency parsing and Transition based depen- dency parsing (Eisner, 1996; Nivre et al., 2004; McDonald and Pereira, 2006). The main idea of 430 MSTParser is to take dependency parsing as a problem of searching a maximum spanning tree (MST) in a directed graph (Dependency Tree). We see MSTParser a better chance to improve the parsing speed and MSTParser provides the state- of-the-art performance for both projective and non- projective tree banks. For the …

Content truncated to 500 words