Domain adaptation with a shrinkable discrepancy strategy for cross-domain sentiment classification

Abstract

Cross-domain sentiment classification (CDSC) is used to predict the sentiment polarity of a text in an unlabeled target domain by analyzing the reviews in the labeled source domain. Domain adaptive approaches have become the preferred solution in recent years to the unsupervised domain migration problem. Among them, adversarial learning aligns the sample distribution of the two domains through domain confusion to transfer sentiment across domains. However, traditional adversarial learning often roughly measures domain discrepancy. Although scholars have attempted to adjust the decision boundary of different categories to eliminate the domain shift, such as maximum classifier discrepancy model, there are still two problems with this approach. First, it ignores the intra-domain structure, which causes the samples distributed on the decision boundary to be easily misclassified. Second, it only realizes coarse-grained sentiment migration and lacks a refined evaluation of the transferable information in the inter-domain, which causes a negative transfer. To solve these problems, we propose domain adaptation with a shrinkable discrepancy strategy (DA-SDS) for the task of CDSC. Specifically, we propose to shrink the category subspace in the intra-domain while building the decision boundary of classifiers, which reduces the misclassification by clustering samples to the category center. We also propose to measure the weighted domain discrepancy in the inter-domain, which mitigates the negative transfer through the refined assessment of domain discrepancy. Extensive evaluations showed that DA-SDS outperformed state-of-the-art methods on the Amazon Review dataset.