GenderCARE: A Comprehensive Framework for Assessing and Reducing Gender Bias in Large Language Models

TL;DR

GenderCARE presents a comprehensive framework to quantify and mitigate gender bias in large language models. It introduces CGEB, a six-dimension criteria for fair benchmarks, and GenderPair, a robust pair-based bias assessment that inclusively covers TGNB identities. Debiasing combines counterfactual data augmentation with LoRA-based fine-tuning, achieving averages above 35% and peaks over 90% bias reduction across 17 LLMs while preserving mainstream task performance within 2%. The work demonstrates that a structured, inclusive approach can yield substantial fairness improvements with practical applicability across diverse model architectures and prompt structures.

Abstract

Large language models (LLMs) have exhibited remarkable capabilities in natural language generation, but they have also been observed to magnify societal biases, particularly those related to gender. In response to this issue, several benchmarks have been proposed to assess gender bias in LLMs. However, these benchmarks often lack practical flexibility or inadvertently introduce biases. To address these shortcomings, we introduce GenderCARE, a comprehensive framework that encompasses innovative Criteria, bias Assessment, Reduction techniques, and Evaluation metrics for quantifying and mitigating gender bias in LLMs. To begin, we establish pioneering criteria for gender equality benchmarks, spanning dimensions such as inclusivity, diversity, explainability, objectivity, robustness, and realisticity. Guided by these criteria, we construct GenderPair, a novel pair-based benchmark designed to assess gender bias in LLMs comprehensively. Our benchmark provides standardized and realistic evaluations, including previously overlooked gender groups such as transgender and non-binary individuals. Furthermore, we develop effective debiasing techniques that incorporate counterfactual data augmentation and specialized fine-tuning strategies to reduce gender bias in LLMs without compromising their overall performance. Extensive experiments demonstrate a significant reduction in various gender bias benchmarks, with reductions peaking at over 90% and averaging above 35% across 17 different LLMs. Importantly, these reductions come with minimal variability in mainstream language tasks, remaining below 2%. By offering a realistic assessment and tailored reduction of gender biases, we hope that our GenderCARE can represent a significant step towards achieving fairness and equity in LLMs. More details are available at https://github.com/kstanghere/GenderCARE-ccs24.

Figures (3)

• Figure 1: Illustration of the limitations of template-based benchmarks (left) and phrase-based benchmarks (right).
• Figure 2: The GenderCARE framework for comprehensive gender bias assessment and reduction in LLMs. It consists of four key components: (I) Criteria for gender equality benchmarks; (II) Assessment of gender bias in LLMs using the proposed GenderPair benchmark aligned with the criteria; (III) Reduction of gender bias via counterfactual data augmentation and fine-tuning strategies; (IV) Evaluation metrics at both lexical and semantic levels for bias quantification.
• Figure 3: Assessment of the Alpaca and Vicuna 7B and 13B models using GenderPair with three different prompt structures (Sec. \ref{['sec5.4.1']}). The results for each metric are mean values across three gender groups.