Cross-Domain Sentiment Classification with Attention-Assisted GAN

In the problem setting of cross-domain sentiment classification, two different domains are introduced, and we refer to them as source and target domains respectively. In the source domain, sentiment labels to instances are available, while labels to instances in the target domain are not available. This problem is critical and practical, as in the real world, data in some domains (source) are abundant, while those in other domains (target) may become scarce. In this paper, we propose a cross-domain sentiment classification framework based on Generative Adversarial Networks (GANs) with the assistance of the attention mechanism, which aims to leverage the information available from the source domain to the target domain. Existing state-of-the-art methods mainly use multi-task learning to minimize the distance between the source and the target instances in a latent feature space. However, the projections may suffer as the deep model always tries to overfit the cross-domain adaptation task. We introduce a framework with multiple tasks, including adversarial example generation, cycle reconstruction, and cross-domain classification. Empirical evaluation and analysis on real-world datasets are being performed to validate the effectiveness of our proposed algorithm compared to state-of-the-art techniques.