User and Product Attention Model Based on Recommendation Algorithm for Sentiment Classification

Document-level sentiment classification aims to predict user's overall sentiment in a document about a product. Existing sentiment classification methods just utilized the document content without consideration of user and product information, which have crucial effects on the sentiment polarities. To address the issue, this paper proposes a Hierarchical Bidirectional Long Short Term Memory (H-BiLSTM) model, where the user and product information captured by recommendation algorithm, and introduce IUF (Inverse User Frequence) as optimization objective, which proves effective. In addition, our model can adapt to the situation when the training data is sparse. Experiments on IMDB and Yelp datasets demonstrate that results are close to the state-of-the-art methods.