ICT:A System Combination for Chinese Semantic Dependency Parsing

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 514–518, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics ICT:A System Combination for Chinese Semantic Dependency Parsing Hao Xiong and Qun Liu Key Lab. of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {xionghao, liuqun}@ict.ac.cn Abstract The goal of semantic dependency parsing is to build dependency structure and label seman- tic relation between a head and its modifier. To attain this goal, we concentrate on obtain- ing better dependency structure to predict bet- ter semantic relations, and propose a method to combine the results of three state-of-the-art dependency parsers. Unfortunately, we made a mistake when we generate the final output that results in a lower score of 56.31% in term of Labeled Attachment Score (LAS), reported by organizers. After giving golden testing set, we fix the bug and rerun the evaluation script, this time we obtain the score of 62.8% which is consistent with the results on developing set. We will report detailed experimental results with correct program as a comparison stan- dard for further research. 1 Introduction In this year’s Semantic Evaluation Task, the organiz- ers hold a task for Chinese Semantic Dependency Parsing. The semantic dependency parsing (SDP) is a kind of dependency parsing. It builds a depen- dency structure for a sentence and labels the seman- tic relation between a head and its modifier. The semantic relations are different from syntactic rela- tions. They are position independent, e.g., the pa- tient can be before or behind a predicate. On the other hand, their grains are finer than syntactic re- lations, e.g., the syntactic subject can be agent or experiencer. Readers can refer to (Wanxiang Che, 2012) for detailed introduction. Figure 1: The pipeline of our system, where we com- bine the results of three dependency parsers and use max- entropy classifier to predict the semantic relations. Different from most methods proposed CoNLL-2008 1 and 2009 2, in which some re-in searchers build a joint model to simultaneously gen- erate dependency structure and its syntactic relations (Surdeanu et al., 2008; Hajiˇc et al., 2009), here, we first employ several parsers to generate depen- dency structure and then propose a method to com- bine their outputs. After that, we label relation be- tween each head and its modifier via the traversal of this refined parse tree. The reason why we use a pipeline model while not a joint model is that the number of semantic relations annotated by or- ganizers is more than 120 types, while in the for- mer task is only 21 types. Compared to the former task, the large number of types will obviously drop the performance of classifier. On the other hand, the performance of syntactic dependency parsing is ap- proaching to perfect, intuitively, that better depen- dency structure does help to semantic parsing, thus we can concentrate on improving the accuracy of de- pendency structure construction. The overall framework of our system is …

Content truncated to 500 words