Double LDA: A Sentiment Analysis Model Based on Topic Model

The sentiment mining is a fast growing topic of both academic research and commercial applications, especially with the widespread of short-text applications on the Web. A fundamental problem that confronts sentiment mining is the automatics and correctness of mined sentiment. This paper proposes an DLDA (Double Latent Dirichlet Allocation) model to analyze sentiment for short-texts based on topic model. Central to DLDA is to add sentiment to topic model and consider sentiment as equal to topic, but independent of topic. DLDA is actually two methods DLDA I and its improvement DLDA II. Compared to the single topic-word LDA, the double LDA I, i.e., DLDA I designs another sentiment-word LDA. Both LDAs are independent of each other, but they combine to influence the selected words in short-texts. DLDA II is an improvement of DLDA I. It employs entropy formula to assign weights of words in the Gibbs sampling based on the ideas that words with stronger sentiment orientation should be assigned with higher weights. Experiments show that compared with other traditional topic methods, both DLDA I and II can achieve higher accuracy with less manual needs.