减轻大语言模型性别偏见的正则化方法  

Regularization Methods to Mitigate Gender Bias in Large Language Models

作　　者：吴逢宇 叶阿勇[1] WU Fengyu;YE Ayong(College of Computer and Cyber Security,Fujian Normal University,Fuzhou,China,350117)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福州350117

出　　处：《福建电脑》2025年第2期18-23,共6页Journal of Fujian Computer

基　　金：移动网络中面向位置信息生命周期的隐私保护关键技术(No.61972096)资助。

摘　　要：大语言模型凭借其强大的语言理解能力在简历筛选任务上展现出了优秀的性能,然而其存在的性别偏见可能会影响筛选结果的公平性。为了解决这一问题并保证模型性能,本文提出了一种基于正则化的模型训练方法,引入损失惩罚来引导模型降低偏见,并结合模型微调对简历筛选任务进行优化。在真实的简历和岗位描述数据集上进行的实验表明,采用该方法训练的模型能够有效减少性别偏见,并在简历筛选中保持良好的性能。Large language models have been shown to perform excellently in resume screening tasks due to their strong language comprehension abilities.However,the presence of gender bias in these models may affect the fairness of screening outcomes.To address this issue and maintain model performance,a regularization-based training method was proposed in this paper.A loss penalty is introduced to guide the model in reducing bias,while model fine-tuning is employed to optimize the resume screening task.Experiments conducted on real-world resume and job description datasets demonstrate that the models trained with this approach can effectively reduce gender bias while maintaining strong performance in resume screening.

关 键 词：大语言模型 性别偏见 正则化策略 简历筛选 信息检索 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]