Feature Distributions in Domain Adaptation

The adaptation problem in sentiment classification is approached in this paper. Since the availability of labeled data required by sentiment classifiers is not always possible, given a set of labeled data from different domains and a small amount of labeled data of the target domain, it would be interesting to determine which subset of those domains has a feature distribution most similar to the target domain. In this paper, we propose a meaningful examination of the overlap between two feature sets in order to obtain the most similar distribution to the target domain. The results of the experimentation show how our method is oriented to select features that have a good correlation with the target domain.