CVAE-Attention: CVAE based Semi-Supervised Sentiment Classification using Attention

Abstract

Text sentiment classification is an important domain in NLP, and the related technical research has been mature. The sentiment classification of text with the "but" contrastive marker is a challenging problem. In this paper, a semi-supervised framework based on conditional variational autoencoder using attention, called CVAE-Attention, is proposed for sentiment classification. In the CVAE-Attention framework, the attention mechanism is introduced to cope with the contrastive structure. The latent semantic information of the clause after "but" (but-clause) is extracted through the attention model, and is incorporated into the generative model to enlarge the effect of the but-clause. Experiments show that the proposed method is effective compared with other state-of-the-art semi-supervised methods.