Representation Learning for Improved Generalization of Adversarial Domain Adaptation with Text Classification

Domain Adaptation techniques remain limited in accuracy and robustness due to data sparsity. In this paper, we present a new approach called Domain Adversarial network with Representation Learning (DARL), to improve domain adaptation by introducing an encoding layer as part of DARL model learning. We integrate a Stacked Denoising Autoencoder and Adversarial learning for the domain adaptation process. The advantage of the proposed method is that it can extract descriptive features under noisy conditions while still learning task discriminative features. The encoding under noisy reconstruction ensures both higher accuracy and increased robustness in the learning process. We evaluate DARL using Amazon review data set and the results showed superior accuracy and robustness compared to Domain Adversarial Neural Networks (DANN).