Intelligent Word Sense Disambiguation Of English Modal Verbs

Word sense disambiguation (WSD) is the ability to identify the meaning of words in context in a computational manner. Most English modal verbs are characteristic of polysemous (according to the statistic result, each modal verb has averagely 6.7 senses) and sense indeterminacy. Therefore, the WSD of English modal verbs is an important but tough problem in both semantic studies and the studies of natural language processing. The traditional semantic studies of the modal verbs have been focused on their sense classification and the description of their syntactic features and functions. The WSD of English modal verbs is mainly limited to the analysis of sense categorization and the description of the syntactic and contextual features.Intelligent WSD has been an important research direction in the natural language processing both at home and abroad. The studies of WSD have mainly focused on the representation of semantic information, the improvement of the technique and the perfection of the algorithms for the WSD of the notional verbs and nouns. Maybe it is because of the complexity of the senses of the modal verbs that the study of the WSD of English modal verbs has not been found, let alone the study of the influence of different semantic and syntactic features on the effects of WSD.Modal verbs play an important role in interpersonal communication and man- machine communication. The intelligent WSD of the modal verbs would be significant for man-machine communication, machine translation, information retrieval and text classification, as well as for the semantic study of the modal verbs. It would also be significant for providing correct modality information for the emotional robots to establish the anthropomorphic emotional models.This dissertation studies the WSD of English modal verbs based on large English corpora. On the basis of the theory of connectionism cognitive psychology, the statistic learning theory, the theories of the Bayes probability and fuzzy logic, it comprehensively studied the WSD and word sense inference of some primary English modal verbs, such as may, must, can and will by the approaches of BP neural network, support vector machine, na?ve Bayesian model, adaptive neuro-fuzzy inference system and fuzzy c-means cluster, and established the models for disambiguating epistemic modalities from root modalities and the models for fuzzy sense inference and fuzzy sense classification of English modal verbs.1) Based on the theory of the connectionism cognitive psychology:"psychological activity is like brain", this dissertation tries to reveal the nature of the natural language through simulating the working mechanism of human brains by neural network. Two BP neural network models for the WSD of may and must are established, respectively, and both of them reach an ideal ratio of correct disambiguation. The influences of different semantic and syntactic features on the effects of the WSD of the two modal verbs are analyzed and ranked through experiments. Besides, a comparison is made to analyze the similarity and difference in features influencing the senses of the two modal verbs. It is found that under the environment of BP neural network, voice plays an important role in the WSD of both modal verbs. The mutual information of modal verb and main verb has a greater influence on the effect of WSD than the mutual information of subject and modal verb, and semantic features have a greater influence on the effect of WSD than the syntactic features.2) A support vector machine for the WSD of must is established based on the semantic and syntactic features, and its correct disambiguation rate reaches 96%. A comparison of the results of the WSD is made between the BP neural network and the support vector machine. It is found that the SVM is more stable and effective as a sense classifier and is also better in the ability of generalization than the BP neural network. However, since the BP neural network is sensitive to the change in linguistic features, it is more suitable for investigating the influence of individual linguistic features on the effect of the WSD of the modal verbs.3) A syntactic feature based na?ve Bayes classifier for the WSD of must is established and its correct disambiguation rate reaches 94%. The results of the study indicate that although this approach has some limitation in the options of linguistic features, as long as the proper linguistic features are extracted, it can achieve an idea result of WSD since it is based on the working principle of probabilities which is different from those of the BP neural network and SVM. 4) An adaptive neuro-fuzzy inference system for must (strong obligation→weak obligation) is established based on the semantic features, syntactic features and the membership degrees from the feature matrix. Moreover, another two adaptive neuro-fuzzy inference systems for can (ability→possibility) and can (permission→possibility) are established, respectively, based on the semantic features, syntactic features and the membership degrees given by personal subjective judgment. The rules for inferring the senses of must (strong obligation→weak obligation), can (ability→possibility) and can (permission→possibility) are obtained. The 3 systems all reach the ideal accuracies of inference with the average testing errors of: 0.18107, 0.24341, and 0.25452, respectively.5) A model of fuzzy c-means clustering for will is established, and the rate of correct clustering reaches 90%. The model clusters 4 senses of will: willingness, intention, predictability and prediction, and clearly shows the membership degrees of each will's belonging to the 4 senses, i.e., the amount of the 4 senses each will holds. The results of the clustering provide valuable reference for the semantic study of will.6) Based on the results of the above studies and through the comparison and mutual verification of them, the primary and secondary factors of influencing the word senses and the results of the WSD of English modal verbs are summarized. Generally, semantic features bring greater influence on the results of WSD of the modal verbs than syntactic features, and among the semantic features, the mutual information of modal verb and main verb have greater influence on the effects of WSD than the mutual information of subject and modal verb. These results provide important evidence for the feature selection for the WSD of English modal verbs.The innovations of this study are as follows:(1) The intelligent WSD of English modal verbs is realized in this study, which is an extension from the WSD of nouns and notional verbs. The categoris of modality of the modal verbs are determined.(2) The regularities of influence by different semantic and syntactic features upon the senses of English modal verbs are revealed and the features are ranked according to the degrees of influence. These provide valuable references for the feature selection in the design of the models for intelligent WSD of English modal verbs, for the sense recognition of English modal verbs by human intuition and for the semantic study of English modal verbs.(3) The fuzzy sense clustering of English modal verb reveals the membership degrees of different senses held by each modal verb and the distributions of the senses, which provides valuable references for the studies of semantic categorization, semantic analysis and semantic evolution of English modal verbs.(4) The models and systems established in this study are with generality, and they can be used for the WSD of other English modal verbs after some modifications. In addition, these models and systems can be used in the automatic natural language processing systems, such as in the sense recognition of modal verbs for the emotional robot system, machine translation, information retrieval and text classification.(5) The results of this study provide theoretical and practical references for the studies of intelligent WSD of other semantically complex words.