Transductive data augmentation framework for cross-domain ASTE using large language models

Purpose This study aims to address the dual challenges of source-domain annotation subjectivity and target-domain pseudo-label noise in cross-domain Aspect Sentiment Triplet Extraction (ASTE). The purpose is to enhance model reliability and robustness under domain shift, thereby supporting fine-grained sentiment understanding and trustworthy decision-making for intelligent agents in web information systems (WIS). Design/methodology/approach Adopting a “text-label-text” paradigm, transductive data augmentation framework for cross – domain ASTE using large language models (TDA-LLM) systematically orchestrates large language models (LLMs) as controllable augmentation agents across three stages: source-domain annotation refinement to reduce label ambiguity, target-domain pseudo-data generation to bridge domain gaps and prediction post-processing with structural validation. A bidirectional consistency-based filtering mechanism prunes unreliable samples, followed by transductive inference for dependency-aware validation. Findings Extensive experiments on six cross-domain benchmarks show that TDA-LLM outperforms strong state-of-the-art baselines by an average of 1.78% in F1 score. The framework effectively mitigates annotation inconsistency and pseudo-label noise, improving structural coherence and prediction reliability in cross-domain sentiment triplet extraction. Originality/value The originality of TDA-LLM lies in its systematic integration of LLMs as controllable augmentation agents throughout the entire data pipeline – from source-domain annotation refinement and target-domain pseudo-label generation to transductive inference with structural validation. It introduces a novel bidirectional consistency-based filtering mechanism to ensure semantic and syntactic fidelity. The value resides in its demonstrated ability to robustly mitigate cross-domain supervision noise, thereby advancing the reliability and trustworthiness of fine-grained sentiment understanding systems for intelligent agents in WISs.