Contrastive Learning-Based Cross-Domain Data Augmentation for Aspect-Based Sentiment Analysis

Cross-domain Aspect-Based Sentiment Analysis (ABSA) leverages unsupervised domain adaptation techniques to transfer knowledge from a source domain, which is rich in labeled data, to the target domain which lacks labeled data. Many recent studies have attempted to address this issue by generating a large amount of labeled target domain data, and the domain adaptive model \(\mathrm DA^2LM\) has achieved state-of-the-art results. However, training this model requires the use of target domain data, which should be annotated with pseudo labels. Therefore, it is important to generate high-quality pseudo labels effectively. Furthermore, when a substantial amount of labeled data for the target domain is obtained using the generative model, it becomes essential to train an effective model to predict the label of test data from the target domain. In this work, we propose a novel Cross-Domain Data Augmentation approach based on Contrastive Learning, named CLCDDA. This approach combines the Multiple Kernel Maximum Mean Discrepancy and enhanced contrastive learning to obtain high-quality target domain data with pseudo labels, which is used to improve the performance of the generative model. In addition, contrastive learning is introduced to improve the performance of the sequence labeling model. The experimental results on four benchmarks show that CLCDDA significantly outperforms previous approaches in both cross-domain End2End ABSA and AE tasks.