Zero-Shot Cross-Domain Aspect-Based Sentiment Analysis via Domain-Contextualized Chain-of-Thought Reasoning

Cross-domain aspect-based sentiment analysis (ABSA) aims to learn domain-specific knowledge from a source domain to perform various ABSA tasks in a target domain. Recent works mainly focus on leveraging domain adaptation techniques to transfer the domain-agnostic features from the labeled source domain to the unlabeled target domain, yet manually collecting target domain data remains impractical owing to the facts like privacy concerns in banking or insurance. To alleviate this issue, we propose ZeroABSA, a unified zeroshot framework for cross-domain ABSA that effectively eliminates dependency on target domain annotations. Specifically, ZeroABSA consists of two novel components, namely, (1) a LLM-driven augmentation module synthesizing domain-adaptive target data through iteratively evaluating the metrics (e..g, vocabulary richness, semantic coherence, and sentiment/domain consistency) of augmented exemplars for refinement; (2) a domain-contextualized chain-of-thought (COT) strategy trains models on augmented data while explicitly modeling domain-invariant reasoning to bridge the well-known cross-domain gap. Extensive evaluations across four diverse domains demonstrate that ZeroABSA surpasses the state-ofthe-arts, which advances the practicality of cross-domain ABSA in real-world scenarios where target domain data is unavailable.