Scholarlink

<p>Dense neural retrievers have improved retrieval effectiveness but can also amplify social biases in ranked results. This paper investigates gender bias in retrieval systems and introduces a fairness-aware training approach that regularizes standard ranking losses with bias and fairness terms. The formulation applies penalty or reward signals at the document level within pairwise objectives, enabling a tunable trade-off between effectiveness and fairness. We evaluate the approach on MS MARCO-derived benchmarks using two encoders (BERT-mini and ELECTRA-small) and two query sets (gender-neutral and socially sensitive). Across ARaB, LIWC, and NFaiRR, our best configurations substantially reduce gender bias while preserving MRR@10 within small to moderate deltas, and in some cases improving effectiveness. We also compare against fairness-aware baselines such as adversarial and neutrality-regularized rankers and find competitive or superior bias reduction under comparable effectiveness. The findings are empirical and scoped to binary gender bias in English on the evaluated datasets and models, without claims of broader generality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Regularization Framework for Gender Bias Mitigation in Dense Neural Rankers

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Shirin Seyedsalehi,
  • Morteza Zihayat,
  • Ebrahim Bagheri

摘要

Dense neural retrievers have improved retrieval effectiveness but can also amplify social biases in ranked results. This paper investigates gender bias in retrieval systems and introduces a fairness-aware training approach that regularizes standard ranking losses with bias and fairness terms. The formulation applies penalty or reward signals at the document level within pairwise objectives, enabling a tunable trade-off between effectiveness and fairness. We evaluate the approach on MS MARCO-derived benchmarks using two encoders (BERT-mini and ELECTRA-small) and two query sets (gender-neutral and socially sensitive). Across ARaB, LIWC, and NFaiRR, our best configurations substantially reduce gender bias while preserving MRR@10 within small to moderate deltas, and in some cases improving effectiveness. We also compare against fairness-aware baselines such as adversarial and neutrality-regularized rankers and find competitive or superior bias reduction under comparable effectiveness. The findings are empirical and scoped to binary gender bias in English on the evaluated datasets and models, without claims of broader generality.