Rethinking the Information Inside Documents for Sentiment Classification

Document-level sentiment classification aims to predict the sentiment rating of a particular document. Most of the methods take word embeddings as the inputs of a neural network. However, most existing methods fail to account for the fact that a specific word usually contains a certain amount of redundant information, and generally different words contain different amounts of redundancy. Such ambiguous word representation will cause models to misunderstand a text, and thus to incorrectly predict the sentiment, meanwhile simply treating words with different amounts of redundancy equally is not appropriate. Moreover, these methods take the user ratings as the training target, which leads to the fact that the information selected by the model is usually limited to the rating itself and there is no obvious sentiment tendency. Accordingly, we propose a Rethinking mechanism (R-TM) to rethink the information inside documents. More specifically, R-TM filters out redundancy contained in different words from different levels, and selects information from positive and negative two different perspectives. Our experimental results demonstrate that the proposed mechanisms can achieve consistent improvements compared to state-of-the-art methods.