Semi-supervised Learning for Sentiment Classification

With the growing need of identifying opinions and sentiments automatically from online text data, sentiment classification tasks have received considerable attention recently. One can treat sentiment classification as a text classification problem, however, it is very time-consuming and somewhat impractical to acquire enough labeled data to train a good sentiment classifier. This paper investigates a semisupervised learning method for sentiment classification which can take advantage of large amounts of unlabeled data. Specifically, we learn some structural information from both labeled and unlabeled data to form a good feature mapping for the sentiment classification tasks, and then bootstrap the learner to improve the classification performance. We present an empirical study on two different sentiment classification tasks which indicates the proposed method can make good used of unlabeled data and improve the classification performance.