Semi-supervised Sentiment Classification with Self-training on Feature Subspaces

Abstract

In sentiment classification, labeled data is often limited while unlabeled data is ample. This motivates semi-supervised learning for sentiment classification to improve the performance by exploring the knowledge in unlabeled data. In this paper, we analyze the possibility and the difficulty of semi-supervised sentiment classification and indicate that noisy features may be the main reason for badly influencing the performance. To overcome this problem, we propose a novel self-training approach where multiple feature subspace-based classifiers are utilized to explore a set of good features for better classification decision and to select the informative samples for automatically labeling. Evaluation over multiple data sets shows the effectiveness of our self-training approach for semi-supervised sentiment classification.