Meaning Matters: Senses of Words are More Informative than Words for Cross-domain Sentiment Analysis

Getting labeled data in each domain is always an expensive and a time consuming task. Hence, cross-domain sentiment analysis has emerged as a demanding research area where a labeled source domain facilitates classifier for an unlabeled target domain. However, cross-domain sentiment analysis is still a challenging task because of the differences across domains. A word which is used with positive polarity in the source domain may bear negative polarity in the target domain or vice versa. In addition, a word which is used very scarcely in the source domain may have high impact in the target domain for sentiment classification. Due to these differences across domains cross-domain sentiment analysis suffers from negative transfer. In this paper, we propose that senses of words in place of words help to overcome the differences across domains, which in turn leads to improvement in cross-domain sentiment classification accuracy. Results show that senses of words provide a better sentiment classifier in the unlabeled target domain in comparison to words for 12 pairs of source and target domains.