RobustDebias: Debiasing Language Models using Distributionally Robust Optimization

TL;DR

This work addresses bias amplification in pretrained language models by debiasing during MLM fine-tuning instead of pretraining. It introduces RobustDebias, a DRO-based method that combines a latent-subgroup autoencoder with a weighted reconstruction objective to debias multiple demographics simultaneously without relying on predefined word lists. Empirical results across StereoSet, SEAT, and CrowS-Pairs show RobustDebias achieving strong bias mitigation while maintaining competitive language modeling performance across BERT-family and related models, outperforming several baseline DRO variants and existing debiasing methods. The approach demonstrates robustness across dataset splits and model architectures, offering a scalable, demography-spanning debiasing solution with practical impact for real-world deployments. Future work points to stronger bias metrics, prompt-based evaluations, multiclass gender debiasing, and extension to other modalities.

Abstract

Pretrained language models have been shown to exhibit biases and social stereotypes. Prior work on debiasing these models has largely focused on modifying embedding spaces during pretraining, which is not scalable for large models. Fine-tuning pretrained models on task-specific datasets can both degrade model performance and amplify biases present in the fine-tuning data. We address bias amplification during fine-tuning rather than costly pretraining, focusing on BERT models due to their widespread use in language understanding tasks. While Empirical Risk Minimization effectively optimizes downstream performance, it often amplifies social biases during fine-tuning. To counter this, we propose \textit{RobustDebias}, a novel mechanism which adapts Distributionally Robust Optimization (DRO) to debias language models during fine-tuning. Our approach debiases models across multiple demographics during MLM fine-tuning and generalizes to any dataset or task. Extensive experiments on various language models show significant bias mitigation with minimal performance impact.