Hierarchical adversarial learning with culture-aware attention and LLM knowledge distillation for cross-cultural sentiment analysis

Abstract

Cultural frame bias significantly degrades sentiment analysis performance across different media systems, often causing accuracy drops of up to 30% during cross-cultural transfer. Existing pre-trained models struggle to disentangle culture-specific editorial framing from true sentiment semantics. To address this, we propose CA-HLAD (Culture-Aware Hierarchical Adversarial Learning with LLM-Guided Knowledge Distillation). This framework explicitly mitigates cultural bias through three innovations: a culture-aware attention mechanism that dynamically adjusts feature representations based on learned cultural prototypes; a hierarchical adversarial structure enforcing domain invariance simultaneously at word, sentence, and document levels; and LLM-guided knowledge distillation leveraging GPT-4 to generate soft labels and counterfactual explanations for a lightweight student model. We constructed a comprehensive benchmark dataset of 15,000 English news articles from ten diverse international outlets (2020–2024). Extensive experiments demonstrate that CA-HLAD achieves a cross-cultural transfer accuracy of 87.4%, outperforming the state-of-the-art adversarial method by 6.9% and BERT-base by 14.7%. Notably, our model reduces the same-culture to cross-culture performance gap from 13.7% to a mere 2.7%, while decreasing the Maximum Mean Discrepancy by 47%. Ablation studies confirm the significant contributions of each architectural component. Qualitative analyses further reveal that CA-HLAD successfully isolates universal sentiment indicators from culture-specific entities. This research provides a robust paradigm for cross-cultural sentiment analysis, offering valuable tools for international media bias detection and global public opinion monitoring.