Cross-Domain Sentiment Classification via Deep Reinforcement Learning

Abstract

Cross-domain sentiment classification aims to transferring the labeled knowledge of the source domain to the target domain for sentiment classification. In the case of lacking labeled target domain data, it reduces manual marking cost of labeling data. Most existing works perform transfer learning by extracting pivot and non-pivot features between domains. These methods are susceptible to noise data and ignore important sentiment information. To solve the above problem, we propose the cross-domain sentiment classification via deep reinforcement learning. We propose a deep reinforcement learning framework, which formulates feature selection policy to solve the noise data problem and pay attention to important data features. The framework is integrated with a policy network and a classification network. The policy network is applied to the feature selection. The classification network makes predictions based on feature selection and calculate the reward of the policy network. Experiment results on the Amazon review datasets demonstrate that the proposed method considerably outperforms other state-of-the-art methods.