Label Distribution Representation Learning in Document-Level Sentiment Analysis

Document-Level Sentiment Analysis (DSA) aims at predicting the user rating for a product based on the user text comments. One major challenge of DSA is the inconsistency between user comments and their ratings. So far, various methods have been proposed to address this challenge, and significant progress has been made. However, existing works ignore that user behaviors may be inconsistent even for the same product. For example, a user may leave similar reviews for the same product at different times, but his/her ratings may be completely different. This will make user labels in training data no longer reliable. To address this issue, we propose a method called Label Distribution Representation Learning (LDRL) for DSA. LDRL assumes the presence of noise in ground truth data and proposes a probability distribution scheme to denote data labels. During the training process, we assign different distribution standard deviations to the users with different reliability levels, thereby reducing the impact of noise. We conduct comprehensive experiments in three real world datasets to evaluate the proposed LDRL. The experimental results show that LDRL outperforms all the baselines in terms of Accuracy and RMSE.