Towards Understanding Task-agnostic Debiasing Through the Lenses of Intrinsic Bias and Forgetfulness

TL;DR

This paper tackles the challenge of task-agnostic debiasing in pretrained language models (PLMs) and the risk of relearning social biases during downstream fine-tuning. It establishes an empirical bias lower bound, showing that the bias level of a debiased model approximates a lower bound for bias in downstream tasks, and introduces ProSocialTuning, a framework that combines Causal Mediation Analysis (CMA), Counterfactual Data Augmentation (CDA), and PAC-Bayes-based generalization estimation to regularize only successfully debiased attention heads during fine-tuning. The approach formalizes a generalization objective using $L_{train}$ and $L_{PAC}$ and a head-level regularization term $\mathcal{E}_{tuning}$ to balance fairness and accuracy, achieving improved extrinsic bias scores across BERT-base and RoBERTa-base on NLI-bias, BiasBios, and STS-B, while maintaining language modeling performance to a reasonable extent. The work provides a practical, data-efficient path toward mitigating bias in downstream systems, highlights the forgetting mechanism as a key factor in debiasing effectiveness, and outlines directions for scaling the framework to larger models and broader bias domains.

Abstract

While task-agnostic debiasing provides notable generalizability and reduced reliance on downstream data, its impact on language modeling ability and the risk of relearning social biases from downstream task-specific data remain as the two most significant challenges when debiasing Pretrained Language Models (PLMs). The impact on language modeling ability can be alleviated given a high-quality and long-contextualized debiasing corpus, but there remains a deficiency in understanding the specifics of relearning biases. We empirically ascertain that the effectiveness of task-agnostic debiasing hinges on the quantitative bias level of both the task-specific data used for downstream applications and the debiased model. We empirically show that the lower bound of the bias level of the downstream fine-tuned model can be approximated by the bias level of the debiased model, in most practical cases. To gain more in-depth understanding about how the parameters of PLMs change during fine-tuning due to the forgetting issue of PLMs, we propose a novel framework which can Propagate Socially-fair Debiasing to Downstream Fine-tuning, ProSocialTuning. Our proposed framework can push the fine-tuned model to approach the bias lower bound during downstream fine-tuning, indicating that the ineffectiveness of debiasing can be alleviated by overcoming the forgetting issue through regularizing successfully debiased attention heads based on the PLMs' bias levels from stages of pretraining and debiasing.

Figures (3)

• Figure 1: StereoSet Scores of BERT Models When Bias Level and Training Dataset Size Vary. The (StereoSet) intrinsic bias scores of the pretrained, debiased, and fine-tuned models are assessed concerning different bias levels and training dataset sizes present in specific datasets for downstream tasks. The fine-tuned model is based on the debiased one and fine-tuning indicates fine-tuning of the pretrained model with task-specific data. Models are considered to be less biased when closer to 50.
• Figure 2: Visualization of CMA Effects of Attention Heads. From left to right, these figures show the effect of CMA on attention heads in the pretrained BERT-base model, debiased BERT-base model, and fine-tuned BERT-base model on benchmarks of NLI-bias, STS-B, and BiasBios respectively. The default random seed is 1. The fine-tuned model is based on the debiased model.
• Figure 3: Language Modeling Scores. These figures present the language modeling scores of the pretrained, debiased, and fine-tuned models with respect to different bias levels and dataset sizes in downstream tasks.