A Parameter-Efficient Multi-Objective Approach to Mitigate Stereotypical Bias in Language Models

Yifan Wang, Vera Demberg2024 · DOI

Abstract

Pre-trained language models have shown impressive abilities of understanding and generating natural languages. However, they typically inherit undesired human-like bias and stereotypes from training data, which raises concerns about putting these models into use in real-world scenarios. Although prior research has proposed to reduce bias using different fairness objectives, they usually fail to capture different representations of bias and, therefore, struggle with fully debiasing models. In this work, we introduce a multi-objective probability alignment approach to overcome current challenges by incorporating multiple debiasing losses to locate and penalize bias in different forms. Compared to existing methods, our proposed method can more effectively and comprehensively reduce stereotypical bias, and maintains the language ability of pre-trained models at the same time. Besides, we adopt prefix-tuning to optimize fairness objectives, and results show that it can achieve better bias removal than full fine-tuning while requiring much fewer computational resources.

Introduction

Large Language Models (LLMs) often mirror and amplify societal biases found in their training data, particularly regarding gender, race, and profession. Mitigating these biases presents two primary challenges. First, bias is multifaceted, manifesting both locally (e.g., token-level associations like "he" with "doctor") and globally (e.g., overall negative sentiment toward specific demographics). Most existing methods target only one form, leaving others untouched. Second, traditional debiasing via full fine-tuning is computationally expensive and often incurs an "alignment tax," degrading the model's general language modeling capabilities.

Methods

The proposed framework utilizes a Multi-Objective Probability Alignment approach combined with prefix-tuning. Instead of updating all model parameters, prefix-tuning adds a small set of trainable "virtual tokens" (a prefix) to the input, steering the frozen pre-trained model toward fairer outputs. This "surgical" approach preserves the model's core knowledge while filtering biased pathways.

Overview of the multi-objective debiasing framework using prefix-tuning and probability alignment.

The optimization process incorporates four distinct loss functions:

  • Language Modeling Loss ($L_{LM}$): Maintains fluency using counterfactual data (gender-swapped sentences).
  • Neutralization Loss ($L_{neu}$): Ensures gender-neutral words (e.g., "nurse", "engineer") have equal probability regardless of the gendered context.
  • Token-level Equalizing Loss ($L_{eq\_tok}$): Equalizes the probability of target pairs (e.g., he/she) in neutral contexts.
  • Sequence-level Equalizing Loss ($L_{eq\_seq}$): Captures global bias by equalizing probabilities across entire sentences.

Training data was derived from the News-Commentary V15 dataset, augmented via Counterfactual Data Augmentation (CDA).

Results

The method was evaluated on GPT-2 and Llama-2 models using benchmarks such as StereoSet and CrowS-Pairs. Results demonstrate that the proposed approach significantly outperforms baselines in bias reduction while maintaining language modeling performance.

Comparison of debiasing performance on StereoSet (Intrasentence and Intersentence) across various models and categories.

As shown in Figure 2, the proposed method ("Ours") achieved an ICAT score of 81.010 on StereoSet Intrasentence, a substantial improvement over the vanilla GPT-2 (68.740). The ICAT score balances bias reduction (Stereotype Score, SS) and language preservation (Language Modeling Score, LMS). Furthermore, the method is highly efficient, requiring training for only 12.36% of the parameters compared to full fine-tuning.

In open-ended generation tasks, the "regard difference" (measuring social perception gaps) dropped to 0.052, the lowest among tested methods. Human evaluation further validated these findings, with the proposed model receiving a fairness rating of 4.50/5.0, compared to just 2.83/5.0 for the vanilla model.

Conclusion

This work demonstrates that a multi-objective approach using parameter-efficient prefix-tuning can effectively mitigate complex stereotypical biases in LLMs. By targeting both local and global bias representations while keeping the base model frozen, the framework achieves superior debiasing with minimal computational overhead. Future research will explore extending this method to non-binary gender identities, race, and religion, as well as investigating other parameter-efficient techniques like LoRA.

Summary updated

Abstract

Pre-trained language models have shown impressive abilities of understanding and generating natural languages.However, they typically inherit undesired human-like bias and stereotypes from training data, which raises concerns about putting these models into use in real-world scenarios.Although prior research has proposed to reduce bias using different fairness objectives, they usually fail to capture different representations of bias and, therefore, struggle with fully debiasing models.In this work, we introduce a multi-objective probability alignment approach to overcome current challenges by incorporating multiple debiasing losses to locate and penalize bias in different forms.Compared to existing methods, our proposed method can more effectively and comprehensively reduce stereotypical bias, and maintains the language ability of pre-trained models at the same time.Besides, we adopt prefix-tuning to optimize fairness objectives, and results show that it can achieve better bias removal than full fine-tuning while requiring much fewer computational resources.
BibTeX
@misc{wang2024parameter,
  author = {Yifan Wang and Vera Demberg},
  title = {{A Parameter-Efficient Multi-Objective Approach to Mitigate Stereotypical Bias in Language Models}},
  year = {2024},
  doi = {10.18653/v1/2024.gebnlp-1.1},
}
Figure