Text Sentiment Classification Based on Layered Attention Network

Abstract

The emerging attention based methods are widely used in sentiment classification, achieving the accuracy improvement of sediment classification tasks. However, these methods usually work improperly in the task of film review classification, in which positive and negative comments are often mixed and interpreting the comments from different perspectives may be diametrically opposite sentiments. In this paper, we propose a new attention based neural network architecture based on HAN model where context layer is added. Compared with the HAN, the addition of the context-aspect layer can remove the impact of unimportant sentences and improve the accuracy of sentiment classification. The experiment results on IMDB dataset show that the proposed model outperforms other existing methods, achieving an accuracy improvement of 3.11% as compared to the state-of-the-art method. The experiment results also show that our model has the better accuracy and the lower iteration time, as compared to the baseline model.