Controlling Gender Bias in Retrieval

Women remain underrepresented pretty much everywhere power hangs out — about 27% of MPs worldwide (Aug 2025) and ~27.5% of managerial roles (2022). Not exactly a rounding error. See IPU Parline and UN Women.

tl;dr: Ever searched “developer” and gotten wall-to-wall men, or “nurse” and seen mostly women? That’s data bias doing cosplay as relevance. This paper, with Prof. Mohammad Aliannejadi’s group at the University of Amsterdam, converts Backpack language models (J. Hewitt et al., 2023) — which split each token into multiple, interpretable “sense” vectors — into an encoder, so the senses that carry gender signal can be identified and turned down at inference, with no retraining required. On MS MARCO and a gender-bias IR benchmark, it reduces RaB/ARaB skew (N. Rekabsaz et al., 2020) with only a small cost to NDCG/MRR, and the resulting ranker still edges out a similarly sized GPT-2 baseline.

Read the paper on arXiv.

Recommended citation: Amirabbas Afzali*, Amirreza Velae*, Iman Ahmadi, and Mohammad Aliannejadi. "Controlling Gender Bias in Retrieval." In Proceedings of ECIR 2026. (*Equal contribution)
Download Paper