Unsupervised Opinion Targets Expansion and Modification Relation Identification for Microblog Sentiment Analysis

Abstract

Microblog brings challenges to existing researches on sentiment analysis. First, microblog short messages might contain fewer content features. Second, it’s difficult to know what users want to express without suitable contexts. On the other hand, people tend to express their opinions in microblog messages, which could be helpful to sentiment analysis. In this paper, we propose a sentiment analysis approach based on opinion target finding and modification relations identification in microblog. First, user comments on specific topics are collected from microblog and preprocessed to reduce noises. Then, opinion targets are expanded by discovering the most frequently co-occurring terms, named entities, and synonyms of the topic. Finally, according to modification relations among part-of-speech (POS) tags, we extract entities or aspects of the entities about which an opinion has been expressed and calculate the overall score of sentiment orientation. In our experiment on 1,000 reviews of 50 movies collected from Twitter, the proposed method can achieve an average accuracy of 84.4% and an average precision of 87.1%, which is better than content similarity with SVM and Naive Bayes. This validates the higher precision in sentiment orientation identification for the proposed approach.