Adversarial Generative Model for Cross-Domain Aspect-Based Sentiment Analysis

The cross-domain aspect-based sentiment analysis task seeks to utilize source domain data with sentiment labels to perform sentiment analysis on target domain data lacking labels. Existing methods rely on the resemblance of the source and target domains, leading to poor results when the distributions differ significantly. To do this, we provide AG-CDSA, an adversarial training-based generative technique that takes labeled statements from the source domain and uses them to produce target domain statements with fine-grained labels. Specifically, to predict fine-grained pseudo-labels, we begin by training it on labeled data from the source domain and then applying it to unlabeled data from the target domain. These pseudo-labels may be noisy, based on which we generate natural language sentences for data augmentation to train for more accurate models. In addition, we introduce perturbations through adversarial training to strengthen the model's generality and robustness. Experimental findings from four reference datasets illustrate the superior performance of the AG-CDSA technique compared to cutting-edge methods.