FairFlow: Mitigating Dataset Biases through Undecided Learning for Natural Language Understanding

2025-03-22

Conference on Empirical Methods in Natural Language Processing

AI Summary

Language models are vulnerable to dataset biases—such as shortcut learning and spurious correlations—leading to degraded cross-domain generalization and compromised fairness. To address this, we propose *Undecided Learning*, a novel paradigm that actively rejects high-confidence yet potentially biased predictions via an uncertainty-driven prediction suppression mechanism. Methodologically, we design a bias-aware multi-view generative and contrastive learning framework, integrating both data- and model-level perturbations to jointly model and suppress both known and unknown biases in a unified manner. Empirically, our approach significantly outperforms existing debiasing methods on cross-domain transfer and challenging sample benchmarks, while preserving source-domain performance. Notably, it achieves the first demonstrated robust mitigation of *unknown* biases—without requiring prior knowledge or bias annotations—thereby offering a principled pathway toward improved model generalization and fairness.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Bias, Fairness & Privacy
• Natural Language Processing: Safety and Robustness

Application Category

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

Abstract

Language models are prone to dataset biases, known as shortcuts and spurious correlations in data, which often result in performance drop on new data. We present a new debiasing framework called FairFlow that mitigates dataset biases by learning to be undecided in its predictions for data samples or representations associated with known or unknown biases. The framework introduces two key components: a suite of data and model perturbation operations that generate different biased views of input samples, and a contrastive objective that learns debiased and robust representations from the resulting biased views of samples. Experiments show that FairFlow outperforms existing debiasing methods, particularly against out-of-domain and hard test samples without compromising the in-domain performance.

Problem

• Mitigating dataset biases in language models
• Learning undecided predictions for biased data
• Improving performance on out-of-domain test samples

Innovation

• Undecided learning mitigates dataset biases
• Data and model perturbation generate biased views
• Contrastive objective learns debiased representations