A Novel Capsule Based Hybrid Neural Network for Sentiment Classification

Sentiment classification of short text is a challenging task because of limited contextual information. We propose a capsule-based hybrid neural network model which can obtain the implicit semantic information effectively. Bidirectional gated recurrent unit (BGRU) is applied in this model to achieve the interdependent features with long distance. Moreover, the capsule network can extract richer textual information to improve expression ability. Compared with the attention-based model which combines self-attention mechanisms and convolutional neural networks (CNN), the capsule-based hybrid model has the advantage of less training time and simple network structure to achieve better performance. The performance is evaluated on two short text review datasets. Our capsule-based model outperforms other related models on movie review data and gets the highest accuracy of 0.8255. Meanwhile, it performs better than most of the systems in NLPCC2014 Task II and, especially achieves the best result on negative data.