A Deep Learning Model Enhanced with Emotion Semantics for Microblog Sentiment Analysis

Word embedding based on neural language model can automatically learn effective word representation from massive unlabeled text dataset, and has made essential progress in many natural language processing tasks. Emoticons in microblog are important emotion signals for microblog sentiment analysis. There have been a lot of research works exploiting emoticons to improve sentiment classification performance for microblog effectively. Commonly used emoticons are adopted to construct an emotion space as feature representation matrix RE from their word embedding. On the basis of vector based semantic composition, the projection to emotion space is performed as matrix-vector multiplication between RE and other embedding. Then, the results are forward to MCNN to learn a sentiment classifier for microblog. This new model is named as EMCNN, short for Emotion-semantic enhanced MCNN, which seamlessly integrates emotion space projection based on emoticon into deep learning model MCNN to enhance its ability of capturing emotion semantic. On the datasets of NLPCC microblog sentiment analysis task, EMCNN achieves the best performance in several sentiment classification experiments and surpass the state-of-the-art results on all the performance metrics. Comparing to MCNN, EMCNN not only improve the classification performance, but also reduce the training time, i.e. 36.15% for subject classification and 33.82% for 7-class sentiment classification. © 2017, Science Press. All right reserved.