Transformer based Chinese Sentiment Classification

This paper deals with the task of Chinese sentiment classification. We propose the MITE (Multi-Inputs Transformer Encoder) model, draw on the transformer encoding thought, mining the emotional information of Chinese contents. MITE introduce self-attention to find the emotional dependence between words, which we think is important for analyzing text sentiment categories. Experiments prove that our method improve the correctness of sentiment classification, which proves the emotional tendency is influenced by the sentimental polarity of the words in the sentence.