AI-driven language technologies are increasingly used in hiring, but they may encode and reproduce harmful social stereotypes. Prior work often studies bias mitigation methods in isolation and outside realistic application settings. We examine the combined effects of data-level and model-level debiasing in a hiring-related context, using Norwegian-language academic bios and a proxy STEM/non-STEM classification task. Specifically, we study masking sensitive information, GenWriter-based rewrites (CITATION), and adversarial debiasing (CITATION). We evaluate these interventions using downstream task performance, group fairness metrics, intrinsic bias tests based on WEAT (CITATION), and measures of gender leakage from hidden representations. We find that combining masking, GenWriter rewrites, and adversarial debiasing substantially reduces gender leakage while maintaining or improving downstream performance. However, effects on fairness gaps and intrinsic bias are mixed, underscoring the need for downstream, context-sensitive evaluation of bias mitigation methods in hiring-related NLP.

Bias in Natural Language Processing (NLP) applications has become a critical issue, with many methods developed to measure and mitigate bias in word embeddings and language models. However, most approaches focus on single categories such as gender or ethnicity, neglecting the intersectionality of biases, particularly in non-English languages. This paper addresses these gaps by studying both single-category and intersectional biases in Italian word embeddings and language models. We extend existing bias metrics to Italian, introducing GG-FISE, a novel method for detecting intersectional bias while accounting for grammatical gender. We also adapt the CrowS-Pairs dataset and bias metric to Italian. Through a series of experiments using WEAT, SEAT, and LPBS tests, we identify significant biases along gender and ethnic lines, with particular attention to biases against Romanian and South Asian populations. Our results highlight the need for culturally adapted methods to detect and address biases in multilingual and intersectional contexts.