Cross-Domain Sentiment Classification Based On Feature’s Sentiment Orientation

With the rapid increase of internet, there are more and more sentiment reviews appearing at the network. Both individual and organization became strongly relying on the information contained in these sentiment reviews to make their decisions. Sentiment classification which could classify review’s sentiment orientation is important for many applications, such as opinion mining, market analysis, opinion supervision, etc.However, the sentiment classification based on traditional machine learning makes at least two assumptions, first, it needs enough labeled data to train an accurate classifier, and second, the training data set and the testing data set should be drawn from the same distribution. All of these require us to label a large number of reviews in every domain, however, this is expensive. Cross-domain sentiment classification represents the most common way to overcome the problem of lacking in labeled data, has been attracting significant attention.This dissertation conducts research on the existing problems in cross-domain sentiment classification. The main work is as follows:(1) First a general overview of cross-domain sentiment classification is presented, including the background, significance and the research status.(2) In addition, a feature selection method for cross-domain sentiment classification is proposed, this method could reduce the high dimensional feature space so as to improve the efficiency while maintain the classification accuracy.(3) Then, a common subspace construction method (CSC) is proposed in our paper, to solve the problem of features which have different sentiments in different domains. Empirical studies demonstrated significant improvement of our method in classification accuracies.