Coattention-Based Recurrent Neural Networks for Sentiment Analysis of Chinese Texts

Sentiment analysis aims to predict user’s sentiment polarities of a given text. In this study, we focus on the sentiment classification task on Chinese texts, which are highly relevant in many online customer services for opinion monitoring. Recently, Recurrent Neural Networks (RNNs) perform very well on solving the classification problem of sentences. Compared with other languages, Chinese text has richer syntactic and semantic information, which leads to form an intricate relationship between words and phrase. In this paper, we propose a Coattention-based RNN for analyzing the sentiment polarities of Chinese short texts, in which the bidirectional RNN with the input word embedding is applied to learn representations of context and target, and coattention mechanism could obtain more effective sentiment feature. In the last, results on two public datasets demonstrate the superiority of our proposed methods over the state-of-the-art methods.