LSTM sentiment polarity analysis based on LDA clustering

Abstract

Sentiment polarity analysis is a major problem in the field of sentiment analysis, especially emotional words will have different emotional tendencies under different scenarios. This paper aims to solve the problem of emotional polarity confusion caused by polysemy of sentiment polarity in different domains, and proposes an LSTM (Long-Short Term Memory) network sentiment polarity analysis method based on LDA (Latent Dirichlet distribution) clustering. The method firstly employs LDA topic model clustering on datasets. Then, the LSTM algorithm is used to train the emotion base learners for clusters. At last, all base learners are integrated with weighting by using the topic probability distribution. Experiments are made on three datasets: (1) Unlocked mobile phone reviews on Amazon; (2) Electronic products reviews on Amazon; and (3) Movies reviews on IMDB. Our experimental result shows the classification accuracy of this method is obviously better than that of using only LSTM method.