Weakly supervised techniques for domain-independent sentiment classification

An important sub-task of sentiment analysis is polarity classification, in which text is classified as being positive or negative. Supervised machine learning techniques can perform this task very effectively. However, they require a large corpus of training data, and a number of studies have demonstrated that the good performance of supervised models is dependent on a good match between the training and testing data with respect to the domain, topic and time-period. Weakly-supervised techniques use a large collection of unlabelled text to determine sentiment, and so their performance may be less dependent on the domain, topic and time-period represented by the testing data. This paper presents experiments that investigate the effectiveness of word similarity techniques when performing weakly-supervised sentiment classification. It also considers the extent to which the performance of each method is independent from the domain, topic and time-period of the testing data. The results indicate that the word similarity techniques are suitable for applications that require sentiment classification across several domains.