Studies On Automatic Recognition Of Functional Words Usages And Application On Dependency Parsing

Modern Chinese words are mainly divided into two categories of content words and functional words, and functional words include adverbs, prepositions, conjunctions, modal words, location words, and auxiliary. Although the functional words can not act as syntactic components, its usages are more complex and diverse. The same functional word may have different parts of speech in different contexts, even if they have the same part of speech, their meanings and usages may different. Therefore it is necessary to analysis and research on various usages of functional words, and it not only offers facilities for the understanding of the text, but also makes for the in-depth study of the modern Chinese.In this paper, it builds the "Trinity" knowledge base of modern Chinese functional words. On this basis the modern Chinese functional words usages automatic recognition, take adverbs for instance, are researched with the rule-based and the statistics-based approach, which uses conditional random field model, maximum entropy model and support vector machine model for research and analysis. The experimental results show that the statistics-based approach is in general better than the rule-based method, and the support vector machine model has the best result in three statistical models. However, from the analysis of single usage recognition, the rule-based approach is better on some usage. Therefore, according to the advantages of both rule-based and statistics-based approach, this paper proposes the idea of combining rules and statistics. The experimental results show that this method achieved good results in the automatic recognition research on adverbs usages.On the basis of functional words usages automatic recognition, this paper analyzes its applications on Chinese dependency parsing. On Chinese dependency syntax analysis, it uses HIT-IR-CDT Treebank and language technology platform LTP of Harbin Institute of Technology, which contains24dependencies. It founds that the recognition effect of coordination relations is poor after the detail analysis about the LTP. This paper sums up the labels of coordination relations, and then recognizes the parallel structure phrases of the sentence with the conjunction usages. Finally, the recognition results of the dependency parser can be processed with the parallel structure phrases to improve Chinese dependency parsing. The experimental results show that the recognition effect of coordination relations is obviously improved after using the parallel structure information.