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

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Gender bias in Chinese NLP is challenging due to metaphorical language and context dependence.
• Current debiasing methods struggle with metaphorical and long-tail biased expressions, risking semantic loss.
• Existing approaches often use fixed demonstrations or one-shot generation, limiting effectiveness.

Purpose of the Study:

• To propose a novel semi-parametric generative framework, TASRIC, for mitigating gender bias in Chinese NLP.
• To improve the handling of complex metaphorical and long-tail biased expressions.
• To ensure semantic preservation during the debiasing process.

Main Methods:

• Developed a unified fine-grained bias recognizer using adversarial training for robust identification.
• Introduced a type-aware semantic retrieval augmentation mechanism with dynamic retrieval for contextual constraints.
• Implemented an iterative correction strategy to maintain core semantics during debiasing.

Main Results:

• TASRIC significantly improves debiased text generation quality compared to a few-shot ChatGPT baseline.
• Achieved a 3.7% increase in BLEU score and a 3.3% increase in METEOR score.
• Effectively addresses limitations of fixed prompts in complex metaphorical bias scenarios.

Conclusions:

• The proposed TASRIC framework offers a more effective approach to mitigating gender bias in Chinese NLP.
• TASRIC balances bias reduction with semantic preservation, outperforming existing methods.
• This research contributes to fairer and more accurate natural language understanding and generation.