Two-stage sentiment classification based on user-product interactive information

Document-level review sentiment classification aims to predict the sentiment category for given review documents written by users for products. Most of the existing methods focus on generating a good review document representation and classifying the review document directly. However, on the one hand, as document-level review sentiment classification usually includes many sentiment categories and the difference between these sentiment categories is not obvious, it may be difficult to obtain satisfying result by direct classification. On the other hand, this once classification process with review representation may fail to well interpret how the results are achieved. In addition, although some information such as user preference and product characteristics are incorporated when building models, the interactive information between user and product are usually ignored. In this paper, inspired by the deductive reasoning strategy of human doing multiple choice questions, we are motivated to propose a Two-Stage Sentiment Classification (TSSC) model to classify review documents in two stages: (1) Coarse classification stage, where model mainly adopts user-product interactive information to pre-judge the sentiment tendency of the review document without considering the review information; (2) Fine classification stage, where model uses text information of the review document for further classification based on the sentiment tendency obtained in coarse classification stage. Finally, the sentiment classification task is accomplished by combining both the results of coarse classification and fine classification. The experimental results demonstrate that our TSSC model significantly outperforms most of the related models (e.g., Trigram and NSC+UPA) on IMDB and Yelp datasets in terms of classification accuracy. When compared with the state-of-the-art HUAPA model, our TSSC model not only achieves slightly more accurate performance, but also has lower time complexity and stronger interpretability.