Global-Local Feature Fusion Mechanism for Sentiment Classification

Short text is a common form of information and opinions in social life. The sentiment analysis of short-text is helpful to obtain the user's attitude towards items, events and services. However, because of sparse data problem of the shorttext, sentiment analysis for short-text is still in face of great challenges. To address this problem, this paper proposes a novel sentiment classification model, Cap_ONCNNA, which includes word embedding module, sentence representation module and capsule classification. Specially, in the sentence representation module, we proposed a new text representation mechanism ONCNNA. This method can generate global structure and local feature semantic representation by attention-based ON-LSTM and CNN respectively and then fuse them to enhance the sentence representation. In classification module, we are able to get more hidden information by introducing capsule network. The experiments on MR, SST1, and SST2 datasets prove our model is superior to other models. Compared with the best model of the three datasets, the accuracy of our model is improved by 0.2% (MR), 1.84% (SST1), 0.03% (SST2) respectively.