Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Identifying and Mitigating Social Bias Knowledge in Language Models

Ruizhe Chen, Yichen Li, Jianfei Yang, Joey Tianyi Zhou, Jian Wu, Zuozhu Liu

TL;DR

This work targets the risk that debiasing methods to enforce group parity can erode individual commonsense knowledge in language models. It introduces Fairness Stamp (FAST), a lightweight, localized debiasing module that identifies the decisive layer storing social bias knowledge and calibrates its outputs without destroying other knowledge, guided by a new benchmark, BiaScope, which evaluates both bias mitigation and knowledge retention via RS and PS metrics. The approach demonstrates superior bias reduction across multiple models and scales to larger architectures while maintaining language modeling and downstream performance, emphasizing the feasibility of fine-grained bias control. Overall, FAST and BiaScope offer a practical path toward fairer LLMs that retain essential factual and commonsense knowledge, with broad implications for deployment in real-world settings.

Abstract

Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.

Identifying and Mitigating Social Bias Knowledge in Language Models

TL;DR

This work targets the risk that debiasing methods to enforce group parity can erode individual commonsense knowledge in language models. It introduces Fairness Stamp (FAST), a lightweight, localized debiasing module that identifies the decisive layer storing social bias knowledge and calibrates its outputs without destroying other knowledge, guided by a new benchmark, BiaScope, which evaluates both bias mitigation and knowledge retention via RS and PS metrics. The approach demonstrates superior bias reduction across multiple models and scales to larger architectures while maintaining language modeling and downstream performance, emphasizing the feasibility of fine-grained bias control. Overall, FAST and BiaScope offer a practical path toward fairer LLMs that retain essential factual and commonsense knowledge, with broad implications for deployment in real-world settings.

Abstract

Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.
Paper Structure (71 sections, 10 equations, 7 figures, 25 tables)

This paper contains 71 sections, 10 equations, 7 figures, 25 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Preliminaries
  4. Social Bias Knowledge
  5. Task Formulation.
  6. Social Bias Knowledge Localization
  7. Contrastive Social Biases Localization.
  8. Determine the decisive layer.
  9. Bias Mitigation with Fairness Stamp
  10. Bias Mitigation.
  11. Knowledge Retention.
  12. BiaScope Benchmark
  13. Dataset Construction
  14. Create commonsense knowledge dataset.
  15. Create paraphrased social bias dataset.
  16. ...and 56 more sections

Figures (7)

  • Figure 1: (a) Expression towards different groups (e.g., mom/dad) does not necessarily constitute a bias. (b) Existing debiasing approaches indiscriminately neutralize different social groups, resulting in unreasonable predictions. (c) Our approach performs fine-grained calibration on biases, while retaining other knowledge.
  • Figure 2: An illustration of our framework: (a) We localize the bias knowledge that over-associates women with nurse than doctor in the language model. (b) We insert a fairness stamp to mitigate the bias knowledge at the localized layer.
  • Figure 3: (a) The average indirect effects of every layer in BERT. (b) Debiasing Performance on different layers in BERT. (c) Ablation on the Number of External Parameters. Experiments are conducted on BERT in terms of gender. SS is transformed by $\text{SS}=100-|\text{SS}-50|$ so that it is also higher is better.
  • Figure 4: An illustration depicting the construction process of the BiaScope benchmark.
  • Figure 5: Knowledge Locating results of GPT2 (left) and GPT2-XL (right).
  • ...and 2 more figures

Theorems & Definitions (1)

  • Definition 1