Alignment-Based Pseudo-Label Generation With Collaborative Filtering Mechanism for Enhanced Cross-Domain Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) in areas such as online shopping and restaurants can effectively facilitate specific service improvements. However, ABSA performance heavily relies on high-quality labeled data, posing a major challenge in data-scarce domains. To solve the data scarcity problem, existing cross-domain ABSA research employs semi-unsupervised learning using pseudo-labels. Unfortunately, low-quality pseudo-labels may mislead model learning, resulting in suboptimal outcomes. Two key questions are raised in this paper, how to improve the quality of pseudo-label generation and how to accurately filter invalid pseudo-labels. First, although traditional adversarial training aims to improve model generalization, its instability and convergence difficulties hinder its effectiveness in managing complex inter-domain differences. To address this challenge, we introduce a scene-style pre-alignment strategy. This method effectively reduces style feature differences between domains and solves adversarial training limitations in terms of convergence and stability, thus improving the quality of pseudo-tag generation. Secondly, we propose a pseudo-label collaborative filtering mechanism, which focuses on filtering labels and texts that would mislead the model by comparing the scene feature similarity between the pseudo-label-based generated texts and target domain texts. Our approach achieves significant performance gains on several public datasets, demonstrating that the combination of aligned pre-training and adversarial training, as well as collaborative filtering mechanisms, indeed improves the model’s domain adaptation.