Aspect-Level Sentiment Analysis Based on a Generalized Probabilistic Topic and Syntax Model

A number of topic models have been proposed for sentiment analysis in recent years, which rely on extensions of the basic LDA model. In this paper, we apply a generalized topic and syntax model called Part-of-Speech LDA (POSLDA) to sentiment analysis, and propose several feature selection methods that separate entities from the modifiers that describe the entities. Along with a Maximum Entropy classifier, we can use the selected features to conduct sentiment analysis at both document and aspect levels. The advantage of using POSLDA is that we can automatically separate semantic and syntactic classes, and easily extend it to aspect level sentiment analysis by mapping topics to aspects. However, words in the noun-related classes, which are also treated as semantic classes, should be removed as much as possible to reduce their impact on sentiment analysis. To evaluate the effectiveness of our solutions, we conducted experiments on two collections of review documents and obtained the accuracy results competitive to the previous work on sentiment analysis.