SESS: A Self-Supervised and Syntax-Based Method for Sentiment Classification

This paper presents a method for sentiment classification, called SESS (SElfSupervised and Syntax-Based method). SESS includes three phases. Firstly, some documents are initially classified based on a sentiment dictionary, and then the sentiments of phrases and documents are iteratively revised. This phase provides some accurately labeled data for the second phase. Secondly, a machine learning model is trained with the labeled data. Thirdly, the acquired model applies on the whole data set to get the final classification result. Moreover, to improve the quality of labeled data, the affect of compound and complex sentences on clause sentiment is examined. For three types of compound and complex sentences, i.e., coordination, concession or condition sentence, the clause sentiment is revised accordingly. Experiments show that, as an unsupervised method, SESS achieves comparative performance to state-of-the-art supervised methods on the same data.