Word Sense Disambiguation Of English Modal Verb Will By Support Vector Machines

Word sense disambiguation is the task to identify the intended meaning of an ambiguous word in a certain context. Due to its wide application in machine translation, information retrieval, speech recognition, text categorization, it has been one of the hot and tough issues in natural language processing. Although techniques of word sense disambiguation have advanced greatly, the research objects have mainly centered on the common nouns and verbs. Modality is concerned with the speaker's opinion or attitude toward the proposition, which is realized mainly by modal verbs. Therefore, accurate identification of the meanings of modal verbs becomes very important for understanding and grasping speaker's opinion or attitude. The three types of semantic indeterminacy of modal verbs: gradience, ambiguity and merger, make it difficult to assign the correct meaning to the modal verbs. Therefore, building an effective and precise word sense disambiguation model of English modal verbs is of great importance.The research, based on one million and two hundred thousand words corpus, extracts eight linguistic features for identifying the sense of will. The thesis adopts Support Vector Machines for the word sense disambiguation of the English modal verb will. The experiment results show that the sense disambiguation accuracy by Support Vector Machines reaches 98.33%. The experimental result verifies the effectiveness of the Support Vector Machines for the sense disambiguation of the English modal verb will. Meanwhile, the effectiveness of the extracted eight linguistic features is also testified. For the purpose of verifying the superiority of the Support Vector Machines for the sense disambiguation, the thesis adopts BP neural network, RBF neural network and Probabilistic neural network to establish the sense disambiguation models of the English modal verb will respectively. By contrasting and analyzing the sense disambiguation results of each model, the thesis finds that word sense disambiguation model by Support Vector Machines has better generalization performance and produces more reliable results in practical application than artificial neural networks. Word sense disambiguation model by Probabilistic neural network outperforms RBF neural network. And RBF neural network is superior to BP neural network. In addition, the thesis investigates the causes for the wrong disambiguation of the sense disambiguation models.Further study is carried out to investigate contributions of different linguistic features to the word sense disambiguation of English modal verb will. By deleting semantic features or syntactic features one time from the original training data and testing data, and then train and test the new model. The experimental results show that semantic features contribute more to the word sense disambiguation of English modal verb will than syntactic features. Wherein, Mutual Information between modal verb will and the verb after it ranks first; namely, Mutual Information between modal verb will and the verb after it weighs more to judge the meaning of will. By F-score algorithm, a simple and effective method, the top four linguistic features influencing the sense disambiguation of will are obtained. They are Mutual Information between will of root meaning and the verb after it, Mutual Information between will of epistemic meaning and verb after it, third person and first person. The reliability of the results is verified by experiments.Successful building of word sense disambiguation model of English modal verb will not only contributes a lot to the realization of automatic sense-tagging of corpus, reducing researchers'heavy workload but also can improve the quality of machine-translation. The verified eight effective linguistic features provide an objective evidence to judge the meaning of modal verb will.