Cross-Domain Reinforcement Learning for Sentiment Analysis

By transferring knowledge from the source domain labeled data to the target domain, cross-domain sentiment analysis can predict sentiment polarity in the lacking of labeled data in the target domain. However, most existing cross-domain sentiment analysis methods establish the relationship of domains by extracting domain-invariant (pivot) features. These methods ignore domain-specific (non-pivot) features or introduce a lot of noisy features to make domain adaptability weak. Hence, we propose a cross-domain reinforcement learning framework for sentiment analysis. We extract pivot and non-pivot features separately to fully mine sentiment information. To avoid the Hughes phenomena caused by the feature redundancy, the proposed framework applies a multi-level policy to select appropriate features extracted from the data. And a sentiment predictor is applied to calculate delayed reward for policy improvement and predict the sentiment polarity. The decision-making capacity of reinforcement learning can effectively tackle the problem of noisy feature data and improve domain adaptability. Extensive experiments on the Amazon review datasets demonstrate that the proposed model for cross-domain sentiment analysis outperforms state-of-the-art methods.