Sentiment analysis through critic learning for optimizing convolutional neural networks with rules

Sentiment analysis is an important task in natural language processing. Previous studies have shown that integrating the knowledge rules into conventional classifiers can effectively improve the sentiment analysis accuracy. However, they suffer from two key deficiencies: (1) the given knowledge rules often contain mistakes or violations, which may hurt the performance if they cannot be adaptively utilized; (2) most of the studies leverage only the simple knowledge rules and sophisticated rules are ignored. In this paper, we propose a critic learning based convolutional neural network, which can address the two shortcomings. Our method is composed of three key parts, a feature-based predictor, a rule-based predictor and a critic learning network. The critic network can judge the importance of knowledge rules and adaptively use them. Moreover, a new filter initialization strategy is developed, which is able to take sophisticated rules into account. Extensive experiments are carried out, and the results show that the proposed method achieves better performance than state-of-the-art methods in sentiment analysis. (C) 2019 Published by Elsevier B.V.