A Hybrid Sentiment Analysis Method of Transformer and Capsule Network for Hotel Reviews

Abstract

As a significant branch of natural language processing, deep learning-based sentiment analysis has dominated in text sentiment analysis instead of lexicon graph and machine learning methods. To further improve the quality of sentiment analysis, we propose a hybrid sentiment analysis method of transformer and capsule network for hotel reviews. The proposed approach takes advantages of both self-attention mechanism in transformer and detailed representation in capsule network to capture bidirectional semantic features well. Compared with the traditional RNN, CNN and pure transformer, the hybrid sentiment analysis method of transformer and capsule network performs 13.83%, 8.97%, and 8.02% higher accuracy for an open-source dataset of hotel reviews respectively. The comprehensive experiments results demonstrate that our proposed method achieves higher quality of sentiment analysis than latest methods.