Semi-Supervised Sentiment Classification Based on Sentiment Feature Clustering

Sentiment classification for text is an important aspect of opinion mining. This paper proposes a semi-supervised sentiment classification method based on sentiment feature clustering. The method only requires a small number of labeled training data instances. Firstly, the method extracts common text features and sentiment features. Common text features can be used to train the first sentiment classifier. Then the spectral clustering-based algorithm is employed to map sentiment features into extended features. The extended features and common text features are combined together to form the second sentiment classifier. The two classifiers select instances from the unlabeled dataset into the training dataset to train the final sentiment classifier. Experimental results show that the proposed method can reach higher sentiment classification accuracy than both the self-learning SVM-based method and the co-training SVM-based method.