Projective Methods for Mitigating Gender Bias in Pre-trained Language Models

TL;DR

This work addresses gender bias in pre-trained language models by testing whether projective debiasing of internal BERT representations can reduce both intrinsic NSP bias and downstream NLI bias without updating model parameters. It contributes an enhanced StereoSet with two new metrics, Strength $S$ and Distance $D$, and a gender-swapped augmentation to better quantify intrinsic bias. The authors implement multiple projection-based interventions—including an attention-based one—at various interior layers, revealing that inner-layer debiasing can yield strong intrinsic bias reductions but may trade off downstream task performance, and that intrinsic-bias improvements do not necessarily predict downstream fairness, as shown by a low Spearman correlation between $S$ and $beta$. The findings advocate task-specific hyper-parameter search and information-weighting strategies, and suggest that debiasing methods should be evaluated with downstream tasks in mind rather than relying solely on intrinsic benchmarks; the approach is readily adaptable to other architectures and downstream bias settings, with practical implications for rapid, non-destructive debiasing in real-world systems.

Abstract

Mitigation of gender bias in NLP has a long history tied to debiasing static word embeddings. More recently, attention has shifted to debiasing pre-trained language models. We study to what extent the simplest projective debiasing methods, developed for word embeddings, can help when applied to BERT's internal representations. Projective methods are fast to implement, use a small number of saved parameters, and make no updates to the existing model parameters. We evaluate the efficacy of the methods in reducing both intrinsic bias, as measured by BERT's next sentence prediction task, and in mitigating observed bias in a downstream setting when fine-tuned. To this end, we also provide a critical analysis of a popular gender-bias assessment test for quantifying intrinsic bias, resulting in an enhanced test set and new bias measures. We find that projective methods can be effective at both intrinsic bias and downstream bias mitigation, but that the two outcomes are not necessarily correlated. This finding serves as a warning that intrinsic bias test sets, based either on language modeling tasks or next sentence prediction, should not be the only benchmark in developing a debiased language model.

Figures (1)

• Figure 1: Abstracted representation of BERT.