Abstract:Sentiment analysis has been one of the research hotspots of natural language processing. The traditional sentiment analysis method can not obtain the global features well,and negative words, turning words and degree adverbs affect the sentence polarity judgment. According to the deep learning method, this paper proposes a short text sentiment analysis based on the CNN (Convolutional Neural Networks) and the attention mechanism of the Bi-GRU (Bidirectional Gated Recurrent Unit). In view of the CNN, the emotional integration is introduced to make full use of the emotional word self-information, and the global feature is obtained through the Bi-GRU.Further more the negative words, turning words and degree adverbs, which affect the polarity of the sentence, are introduced into the attention mechanism to retain important information. The experimental results show that the model can effectively improve the accuracy of sentiment classification compared with the existing related models.