Sentiment analysis by augmenting expectation maximisation with lexical knowledge

Sentiment analysis of documents aims to characterise the positive or negative sentiment expressed in documents. It has been formulated as a supervised classification problem, which requires large numbers of labelled documents. Semi-supervised sentiment classification using limited documents or words labelled with sentiment-polarities are approaches to reducing labelling cost for effective learning. Expectation Maximisation (EM) has been widely used in semi-supervised sentiment classification. A prominent problem with existing EM-based approaches is that the objective function of EM may not conform to the intended classification task and thus can result in poor classification performance. In this paper we propose to augment EM with the lexical knowledge of opinion words to mitigate this problem. Extensive experiments on diverse domains show that our lexical EM algorithm achieves significantly higher accuracy than existing standard EM-based semi-supervised learning approaches for sentiment classification, and also significantly outperforms alternative approaches using the lexical knowledge.