Discriminative Representation Learning for Cross-Domain Sentiment Classification

Cross-domain sentiment classification aims to solve the lack of labeled data in the target domain by using the knowledge of the source domain. Most existing approaches mainly focus on learning transferable feature representations for knowledge transfer across domains. Few of them pay attention to the feature discriminability, which contributes to distinguish different sentiment polarity and improves the classification accuracy. In this work, we propose discriminative representation learning, which extracts transferable and discriminative features. Specifically, we use spectral clustering to reduce the negative effect of low prediction accuracy on the target domain. Centroid alignment enforces samples of the same polarity with smaller distance in the feature space and enlarges the difference between samples of different polarities. Then intra-class compactness benefits true centroid by reducing samples distributed at the edges of the clusters. Experiments on the multiple public datasets demonstrate that discriminative representation learning outperforms state-of-the-art methods.