Using unlabeled data to handle domain-transfer problem of semantic detection

Due to highly domain-specific nature, supervised sentiment classifiers typically require a large number of new labeled training data when transferred to another domain. This is so-called domaintransfer problem. In this work, we attempt to tackle this problem by combining old-domain labeled examples with new-domain unlabeled ones. The basic idea is to use old-domain-trained classifier to label some informative unlabeled examples in new domain, and train the base classifier again. The experimental results demonstrate that proposed method dramatically boosts the accuracy of the base sentiment classifier on new domain.