On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning

TL;DR

Problem: Bias in fine-tuned LMs can persist across tasks and domains. Approach: Upstream Bias Mitigation (UBM) debiases the encoder during upstream fine-tuning and transfers it to downstream tasks. Methods: two algorithms—explanation regularization and adversarial debiasing—are evaluated across hate speech, toxicity, occupation, and coreference tasks, including cross-domain transfer and multi-bias scenarios. Findings: UBM effects generally transfer to downstream models, reducing bias while preserving or improving task performance, though some domain/task cases show limited gains. Implications: UBM offers a more efficient, accessible pathway to debiasing LM fine-tuning without task-specific data or annotations.

Abstract

Fine-tuned language models have been shown to exhibit biases against protected groups in a host of modeling tasks such as text classification and coreference resolution. Previous works focus on detecting these biases, reducing bias in data representations, and using auxiliary training objectives to mitigate bias during fine-tuning. Although these techniques achieve bias reduction for the task and domain at hand, the effects of bias mitigation may not directly transfer to new tasks, requiring additional data collection and customized annotation of sensitive attributes, and re-evaluation of appropriate fairness metrics. We explore the feasibility and benefits of upstream bias mitigation (UBM) for reducing bias on downstream tasks, by first applying bias mitigation to an upstream model through fine-tuning and subsequently using it for downstream fine-tuning. We find, in extensive experiments across hate speech detection, toxicity detection, occupation prediction, and coreference resolution tasks over various bias factors, that the effects of UBM are indeed transferable to new downstream tasks or domains via fine-tuning, creating less biased downstream models than directly fine-tuning on the downstream task or transferring from a vanilla upstream model. Though challenges remain, we show that UBM promises more efficient and accessible bias mitigation in LM fine-tuning.

Figures (3)

• Figure 1: Comparison between the focus of our study (d) and previous works (a,b,c). We study the viability of obtaining an upstream model that could reduce bias in a number of downstream classifiers when fine-tuned.
• Figure 2: Experiment setups to study Upstream Bias Mitigation (UBM) for Downstream Fine-Tuning. We consider the settings with the same or different upstream and downstream domains and tasks, while addressing one or more bias factors (e.g., both dialect bias and gender bias). The framework consists of two stages: (1) an upstream (source) model $f_s=h_s \circ g_s$ is trained with bias mitigation algorithms and (2) the encoder $g_s$ is transferred to the downstream (target) model $f_t$ for fine-tuning.
• Figure 3: Gradient of importance attribution on group identifiers $\phi(w, {\mathbf{x}})$ over time (in solid lines) and the corresponding values of $\phi(w,{\mathbf{x}})$ (in dash lines) during downstream fine-tuning. The cross-marks show the gradient measured in the upstream model (before re-initialization of the classifier layer). UBM$_{reg}$ not only reduces importance attributed to group identifiers, but also the gradient norm of the importance.