Topic Modeling with Sentiment Clues and Relaxed Labeling Schema

This paper proposes a method to extract sentiment topics from a text collection. The method utilizes sentiment clues and a relaxed labeling schema to extract sentiment topics. Experiments with a quantitative and a qualitative evaluations was done to confirm the performance of the method. The quantitative evaluation with a polarity classification marked the accuracy of 0.701 in tweets and 0.691 in newswire texts. These performances are comparable to support vector machine baselines. The qualitative evaluation of polarity topic extraction showed an overall accuracy of 0.729, and a higher accuracy of 0.889 for positive topic extraction. The result indicates the efficacy of our method in extracting sentiment topics.