Attention Based GRU Network for Domain Adaptation in Sentiment Classification

Abstract

Domain adaptation in sentiment classification, as a worthwhile research area, aims to use the available labeled data of a source domain to classify the unlabeled data in the target domain. This approach saves costs and efforts and also improves the efficiency of sentiment classification. This paper proposes the attention based GRU network for domain adaptation in sentiment classification (AGN). This algorithm learns the semantic meanings of the text based on GRU or Bi-GRU, and makes hierarchical mapping from words to sentences and then to the document. The Attention layer is added to capture the core emotional words and sentences. Then, the adversarial idea is brought in, that is, auxiliary domain classification task and sentiment classification task are performed to jointly learn feature representations. Review datasets of five domains on the Amazon online platform are used, and an experiment is performed with 20 pairs of data. The result indicates that the proposed algorithms are able to improve the accuracy of sentiment classification. Finally, the AGN algorithms are interpreted by visualizing the Attention layer.