A Study Of The Influence Of Linguistic Features On The Senses Of Modal Verb Must By Naive Bayesian Model

Word Sense Disambiguation (WSD) is a hot topic of the field of Natural Language Processing (NLP), while the study of modality, which has a long history, is a very important branch of linguistic field. Therefore, the semantic study of modal verbs by the methods of NLP is a cross-disciplinary study and of great significance. This thesis adopts the WSD method to conduct a semantic investigation into the modal verb must.This thesis takes a corpus-based qualitative and quantitative research method. Based on Coates'(1983) research results of modal auxiliaries, this thesis intends to use a supervised WSD method, namely Na?ve Bayes (NB) method, to further explore the following two questions. First, to what degree do the different linguistic features influence the senses of the modal verb must? Second, what is the possibly best feature combination for the disambiguation of the modal verb must?The creativity of this thesis can be displayed by the following three points. First, it applies Na?ve Bayes method to disambiguate the modal verb must. This makes a contribution to freeing the WSD studies from the situation that they are mainly limited to the disambiguation of nouns, verbs or adjectives whose senses are easy to be defined, so that the study of WSD steps into the level of modal verbs. Second, the indirect use of Mutual Information (MI) enables Na?ve Bayes method to take MI as an input vector, so that this important linguistic parameter can play a role in the WSD process. Third, the use of Conditional Mutual Information (CMI) in the process of selecting linguistic features makes the research results possess more scientificity.The experiment results show that different linguistic features have different degrees of influence on the senses of must. The relative degrees of influence of the linguistic features can be ranked in a descending order as follows: the perfective aspect of the verb after mustï¹¥the stativeness of the verb after must, the agentiveness of the verb after mustï¹¥MI of must and the verb after itï¹¥the person of the subjectï¹¥the passive voice of the verb after must, the negation of must, the animateness of the subject. Of these features, the stativeness and the agentiveness of the verbs have similar degree of influence on the senses of must. The passive voice of the verb after must, the negation of must and the animateness of the subject all have little influence on the senses of must. When the five features, MI of must and the verb after it, the perfective aspect of the verb after must, the stativeness of the verb after must, the agentiveness of the verb after must and the person of the subject, are combined to disambiguate must by Na?ve Bayes method, the disambiguation effect is the best. In this case, the disambiguation accuracy reaches 91% when 100 sentences are tested. In addition, the experiments in this thesis prove the ideas of such NB researchers as Harry Zhang, Ludmila I. Kuncheva that the conditional independence assumption is not the'sufficient and necessary'condition for Na?ve Bayes to reach its disambiguation optimality.