Cross-domain sentiment classification-feature divergence, polarity divergence or both?

Abstract

Sentiment classification, which aims to predict the polarity of users’ viewpoint hidden in reviews, is a domain-specific problem, it often fails to be tested in one domain as a classifier trained from another domain. It is hence significant and challenging for cross-domain sentiment classification due to the following two cases: (1) the same feature is used to express different sentiments in different domains (we call it polarity divergence), and (2) different features are used to express similar sentiments in different domains (we call it feature divergence). Existing efforts focus on the latter and consider little on the former. In this paper, we consider both cases in cross-domain sentiment classification and propose a novel algorithm by transferring the polarity of features (TPF). Since the polarity of features is informative for sentiment classification, our algorithm transfers the polarity of features from the source domain to the target domain with the independent features as the bridge. It is worth to note that the polarities of independent features are reset when they are involved in the former case. In addition, the resetting of independent features’ polarities in our algorithm can also be used as a preprocessing step in existing efforts. Empirical results show that our proposed method outperforms state-of-the-art methods in cross-domain sentiment classification.