OSU_CHGCG at SemEval-2016 Task 9 : Chinese Semantic Dependency Parsing with Generalized Categorial Grammar

Proceedings of SemEval-2016, pages 1218–1224, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics OSU CHGCG at SemEval-2016 Task 9: Chinese Semantic Dependency Parsing with Generalized Categorial Grammar Manjuan Duan Lifeng Jin William Schuler Department of Linguistics The Ohio State University {duan,jin,schuler}@ling.osu.edu Abstract This paper introduces our Chinese semantic dependency parsing system for Task 9 of Se- mEval 2016. Our system has two components: a parser trained using the Berkeley Grammar Trainer on the Penn Chinese Treebank rean- notated in a Generalized Categorial Grammar, and a multinomial logistic regression classi- fier. We first parse the data with the auto- matic parser to obtain predicate-argument de- pendencies and then we use the classifier to predict the semantic dependency labels for the predicate-argument dependency relations ex- tracted. Although our parser is not trained directly on the task training data, our system yields the best performance for the non-local dependency recovery for the news data and comparable overall results. 1 Introduction Semantic dependency parsing is an important lan- guage processing task which is useful in information extraction and question answering. In this paper, we introduce a Chinese semantic dependency parsing system which is built upon a categorial analysis of the Chinese language. Categorial grammar annotations are attractive be- cause they have a transparent syntactic-semantic interface and provide a natural account of long- distance dependencies in a language. For this sys- tem, we adopt a Generalized Categorial Grammar framework, GCG, (Bach, 1981), for our language analysis. GCG annotations, compared with other Categorial Grammars, have a larger set of language- specific rules and a smaller set of lexical categories, which on the one hand retains the desirable fea- tures of a categorial grammar, such as straightfor- ward compositionality of its syntactic derivations and elegant analysis of filler-gap phenomenon, and on the other hand, mitigates the sparse data prob- lem faced by any heavily lexicalized annotations. Parsers trained with English GCG annotations have been shown to have state-of-the-art parsing perfor- mance and better long-distance dependency recov- ery (Rimell et al., 2009; Nguyen et al., 2012). Parsers trained with Chinese GCG annotations have been shown to achieve better parsing accuracy than the parser trained with Chinese Combinatory Cate- gorial Grammar, CCG, (Steedman, 2000; Steedman, 2012) annotations (Tse and Curran, 2010; Tse and Curran, 2012) for those trees which both grammar annotations assign the same tree structures (Duan and Schuler, 2015). The current experiment is our first experiment with dependency relations generated from the Chi- nese GCG annotations. We evaluate them against the manually annotated semantic dependencies in the current SemEval task. Since the purpose of the system is to verify the semantic dependencies gen- erated by the Chinese GCG parser are reasonable, we adopt a minimalist machine learning scheme for this system to accomplish the evaluation. We first train the Berkeley parser (Petrov and Klein, 2007) with GCG annotations converted from the currently 1218 reannotated protion of around 71% of the Chinese Treebank (Xue et al., 2005) sentences. With this parser, we parse the sentences …

Content truncated to 500 words