TASRIC: A Type‐Aware Semantic Retrieval Augmentation Framework With Iterative Correction for Bias Mitigation

Gender bias in Chinese natural language processing poses significant challenges due to its high degree of metaphorical expression and strong dependence on contextual cues. Existing approaches typically rely on fixed demonstrations or one-shot generation strategies, which may result in incomplete debiasing or loss of core semantics when handling metaphorical and long-tail biased expressions. To address these limitations, this paper proposes a type-aware semantic retrieval augmentation framework with iterative correction (TASRIC), a semi-parametric generative framework for mitigating bias. First, we construct a unified fine-grained recognizer based on adversarial training to enable robust fine-grained bias identification. Second, we introduce a type-aware semantic retrieval augmentation mechanism that leverages bias type information and dynamic semantic retrieval to impose contextual constraints, thereby alleviating the insufficient coverage of fixed demonstration prompts in complex metaphorical bias scenarios. Finally, an iterative correction strategy is designed to ensure general semantic preservation during the debiasing process. Experimental results demonstrate that, compared with a few-shot ChatGPT baseline, TASRIC improves the quality of debiased text generation by 3.7% in BLEU, an n-gram overlap metric, and 3.3% in METEOR, a metric that considers more flexible lexical matching.