Fine tuning method affects gender and stereotype bias in HR embeddings
Fine-Tuning Strategy, Not Just Data: A Statistically Rigorous Audit of Gender and Stereotypical Bias in Embedding Models for HR Recommendation Systems
Information RetrievalComputation and LanguageComputers and SocietyMachine Learning
Summary
Job recommendation systems use mathematical models to match candidates with jobs based on descriptions and resumes. These models can unintentionally include biases related to gender and stereotypes. The authors checked several popular models with careful statistics and found that the way the model is fine-tuned can increase or not affect these biases. This means companies building hiring tools can choose fine-tuning strategies to reduce biased recommendations. Some aspects remain uncertain due to limited data or software restrictions.
What this means in practice
- •For hr technology teams: Adjust fine-tuning techniques of embedding models to reduce gender and stereotype biases in job recommendation systems.
- •For natural language processing engineers: Use robust statistical tests to more reliably evaluate social biases when developing or adapting multilingual embedding models.
Authors
Ekta Yadav
Abstract
Embedding models underpin modern job-candidate matching: resumes, job descriptions, and queries are encoded as dense vectors, and recommendations come from nearest-neighbor retrieval over that space. Any social bias in the embeddings therefore feeds directly into HR recommendations. We audit six multilingual sentence-embedding models (LaBSE- and E5-family, both public and custom fine-tuned) using Bias-in-Bios (occupational gender bias) and StereoSet (stereotypes across profession, race, gender, and religion). We treat statistical validity as a first-class concern: we identify a non-independence error in a naive significance test, where pooling non-independent pairwise similarities inflated a trivial 0.0046 mean difference to $p \approx 0$, and replace it with a profession-level paired Wilcoxon signed-rank test, individual-resampling bootstrap confidence intervals, and permutation tests for StereoSet. Under the corrected methodology, knowledge-distillation (KD) fine-tuning of E5-Base significantly increases measured bias on Bias-in-Bios ($Δ= -0.177$, $p < 10^{-8}$; more negative means more bias), although this measure is sensitive to global embedding geometry. The KD result is corroborated by pro-stereotype patterns in the StereoSet profession category (uncorrected $p < 0.05$). Binary-classification fine-tuning of the same base model shows no statistically distinguishable change ($p = 0.218$). Fine-tuning strategy is thus an actionable design lever for practitioners building HR recommenders. We also report which benchmark categories are underpowered, and note that two domain-adapted models could not be included in significance testing due to a software-access constraint, which limits our conclusions.