Research Of Chinese Frame Identification Technology Based On Semantic Feature Of Lexical Units

In recent year, semantic analysis has been paid attention to in natural language processing. With SemEva12007Task19proposed, more and more researchers have focused on. Frame disambiguation is a sub-task of the Frame Semantic Structure Extraction. As intermediate links, frame disambiguation occupies a fundamental and important position. The most commonly used method is to manually set the model and feature templates, studies show that this method cannot make full use of the semantic features of the frame evoked by the target words. Therefore, this paper studies the Chinese frame disambiguation based on semantic features of lexical units. The two most important steps of frame disambiguation are model selection and feature selection. For model selection, Chinese frame disambiguation is considered as a single point classification problem, so study the effect of SVM model and maximum entropy model to Chinese frame disambiguation. For feature selection, current method is manual feature selection. However, this method does not effectively use the semantic feature of each target word. In addition, many experiments indicate that their feature templates are different when the target words achieve best results. Hence, this paper proposes Chinese frame disambiguation method based on semantic feature of lexical units. The main research contents are as follow:(1) The model of Chinese frame disambiguation based on semantic feature of lexical units. Introduce the SVM model and maximum entropy model, and analyze the effects of these two models in the lexical level and syntactic level to performance of the Chinese frame disambiguation. The results show that the maximum entropy model is more suitable for the Chinese framework disambiguation(2) The selection feature of Chinese frame disambiguation based on semantic feature of lexical units. Use the maximum entropy model, adopt the traditional manual feature selection methods and automatic feature selection method proposed by the thesis for Chinese frame disambiguation. The results show that the automatic feature selection method is significantly better than manual feature selection method in time complexity and space complexity, also simplifies the feature template.(3) Using most-frequent frame method as baseline, and compare with related literature. In addition, verifying the effectiveness of automatic feature selection by analyzing the results based on word level and syntactic level.The paper has the aid of Chinese FrameNet corpus, utilize a maximum entropy model and feature templates selected by automatic feature selection algorithm for each ambiguous target word with5-fold cross validation, and the average precision could achieve84.46%.