基于神经网络的语境特征对must语义排歧影响研究

Influence of Different Contextual Features Upon WSD of English Modal Verb

英语情态动词的一词多义给自然语言处理带来了很大困难。情态动词语义对语境很敏感,发现影响情态动词语义的主要语境因素对情态动词特征选择、机器翻译等都十分重要。因此,采用神经网络技术对英语情态动词进行语义排歧,并确定不同语境特征对语义排歧结果的影响。基于一个100万字的语料库,以英语情态动词must为例,从语境中提取影响must语义的语义特征和句法特征,计算并确定这些特征向量值,建立可区分根情态与认识情态语义的BP神经网络,排歧正确率达到94%。在此基础上,通过实验研究确定不同语境特征对情态动词must语义排歧的影响程度等级。该研究结果为情态动词语义排歧及情态动词语义人工识别提供了重要依据。

Polysemy of English modals brings trouble for natural language processing.The senses of modal verbs are sensitive to the contexts in which they occur;therefore,to find out the main contextual factors influencing the senses of a modal verb has been an im⁃portant issue in feature selection and machine translation.Therefore,in this paper,English modal verb“must”is taken for instance,a word sense disambiguation(WSD)is conducted based on the technique of neural network.The influence of different contextual fea⁃tures upon the accuracies of WSD of“must”is investigated.Based on a corpus of 1 million words,some semantic features and syntac⁃tic features are extracted from the context and the feature vectors for WSD of“must”are prepared,and a BP neural network for disam⁃biguating root“must”from epistemic“must”is established with an accuracy of 94%.Then,the influence of different contextual fea⁃tures upon the WSD of“must”is examined and the gradient of the influences by different features is given.The results of the study pro⁃vide some important evidence for the WSD and human recognition of the sense of the modal verb.