Cross-domain sentiment classification with word embeddings and canonical correlation analysis

A common approach for automatic sentiment classification is using classifiers trained on labeled text data (reviews, blog posts etc.) to predict the sentiment polarity of new data. Because people express sentiment differently in different domains, this approach requires annotated corpora for each domain. However, annotating data for every domain of interest is laborious and impractical. In this paper, we address the domain adaptation problem for sentiment classification. We explore the effect of generic methods for feature learning and feature subspace mapping, namely word embeddings and canonical correlation analysis (CCA), on cross-domain sentiment classifiers. We show that by using only such rather generic methods, it is possible to get results very competitive with those of sophisticated methods specially developed for the considered problem. An advantage of using word embeddings and CCA is their availability out-of-the-box, which is important for the applicability of the proposed method. Experiments on a widely used benchmark dataset shows that both word embeddings and CCA contribute to accuracy improvement and their combination provides the best results.