Semi-supervised Chinese contextual polarity classification with automatic feature selection

Common approaches to the tasks of sentiment analysis start with a list of words with prior polarities, which are context-free. However, a word can exhibit different polarities in different contexts, which are termed as contextual polarities. In this paper, viewing polarities of words as properties of word senses, we treat the Chinese contextual polarity classification as word sense disambiguation (WSD), and manually labeled a Chinese dataset for training and testing. Due to the insufficiency of labeled data, semi-supervised methods are adopted; to find the effective features for the contextual polarity classification, two automatic feature selection algorithms are proposed. We combine the semi-supervised methods with automatic feature selection algorithms in order to utilize the strengths of both. The experimental results show that the semi-supervised methods, automatic feature selection, and the combination of both help to improve the Chinese contextual polarity classification above a supervised baseline model.