Mutual Information-Based Word Embedding for Unsupervised Cross-Domain Sentiment Classification

Unsupervised cross-domain sentiment classification is a challenging task, which trains a sentiment classifier using labeled source domain data and could be applied to unlabeled target domain. This problem can be solved by training cross-domain word embedding. This paper proposes a method of mutual information-based word embedding. Firstly, source domain word embedding is trained based on source domain data. Then the pivots with strong sentimental polarities are selected based on mutual information (MI). According to the MI value of pivots, the transfer coefficient from the source domain to the target domain is calculated. Finally, the transfer coefficient is used as a constraint of the objective of the skip-gram model to train the word embedding of the target domain. Experiments show that the proposed method is superior to the cross-domain word embedding based on word frequency.