Compositional polarity classification approach for product reviews

In this paper, we examine the effectiveness of compositional polarity classification technique which uses semi-supervised classifier with the help of a domain-independent unsupervised classifier for sentiment classification problem. For compositional polarity classifiers, we create a pseudo-labeled training set by using an unsupervised classifier that relies on a lexical resource and train a base SVM classifier over the training set, and then investigate four semi-supervised learning methods (self-training, Transductive SVM, spectral graph transduction and semi-supervised learning based on a Deterministic Annealing approach) on four Chinese datasets which span two different domains: digital products and hotel. Through comparative experiments, we conclude that compositional classification technique is effective and helpful to improve the accuracy of sentiment classification without using labeled data.