Collapsed Language Models Promote Fairness

2024-10-06

arXiv.org

AI Summary

This work addresses the longstanding challenge of jointly optimizing fairness and performance in language models. We first uncover an intrinsic connection between neural collapse—a geometric phenomenon wherein class-wise representations converge to simplex equiangular tight frames—and model bias: debiasing induces geometric alignment between token representations and word embeddings. Leveraging this insight, we propose a general, geometry-aware fine-tuning framework—the first interpretable and generalizable paradigm for fairness enhancement. Our method integrates neural collapse analysis, representation alignment modeling, and constrained fine-tuning, validated across a comprehensive multi-task fairness evaluation suite. Experiments demonstrate that our approach improves fairness by 23.6% on average across eight mainstream bias benchmarks, while preserving natural language understanding performance with negligible fluctuations (±0.3%).

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Ethics, Bias, and Fairness

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

To mitigate societal biases implicitly encoded in recent successful pretrained language models, a diverse array of approaches have been proposed to encourage model fairness, focusing on prompting, data augmentation, regularized fine-tuning, and more. Despite the development, it is nontrivial to reach a principled understanding of fairness and an effective algorithm that can consistently debias language models. In this work, by rigorous evaluations of Neural Collapse -- a learning phenomenon happen in last-layer representations and classifiers in deep networks -- on fairness-related words, we find that debiased language models exhibit collapsed alignment between token representations and word embeddings. More importantly, this observation inspires us to design a principled fine-tuning method that can effectively improve fairness in a wide range of debiasing methods, while still preserving the performance of language models on standard natural language understanding tasks. We attach our code at https://github.com/Xujxyang/Fairness-NC-main.

Problem

• Bias Reduction
• Fairness Enhancement
• Performance Maintenance

Innovation

• Neural Collapse
• Fairness Enhancement
• Language Model Bias Mitigation