Research on methods for progressive cross-domain aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) aims to extract sentiment polarities associated with different aspects or specific targets, and are widely used for public opinion monitoring, business intelligence, and decision-making support. Traditional cross-domain ABSA methods typically focus on generating labeled data for the target domain and using them to train cross-domain models. However, these approaches often encounter challenges such as significant semantic deviations between the generated and original target domain texts, limited sentence structure diversity, and low-quality pseudo-labels. These limitations hinder the model's ability to effectively learn target-domain features, resulting in suboptimal transfer performance. To address these challenges, this paper proposes a framework for progressive cross-domain ASBA, which consists of four sequential stages: (1) generating high-quality pseudo-labels for the target domain, (2) reconstructing target-domain texts aligned with the pseudo-labels, (3) refining the reconstructed texts to enhance diversity and quality, andłinebreak (4) performing final cross-domain adaptation through joint training with source-domain data. Model parameters are progressively transferred across stages, facilitating progressive domain adaptation while generating high-quality labeled target-domain data. Extensive experiments conducted on four public benchmark datasets demonstrate the effectiveness of the proposed method. For the cross-domain triplet extraction task, the proposed method outperforms existing cross-domain approaches by an average of 1.28% in F1 score, and surpasses non-transfer models by an average of 9.57%, highlighting its superior cross-domain generalization capabilities.