Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists

TL;DR

This paper introduces an entropy-based attention regularization method that reduces bias and overfitting in NLP models without relying on predefined lists of identity terms, improving fairness and performance.

Contribution

The authors propose a novel, knowledge-free regularization technique that discourages overfitting to specific terms by penalizing low-entropy attention, enhancing bias mitigation in NLP models.

Findings

Matches or exceeds state-of-the-art performance on hate speech classification.

Effectively reveals overfitting terms contributing to bias.

Improves fairness metrics across multiple benchmark datasets.

Abstract

Natural Language Processing (NLP) models risk overfitting to specific terms in the training data, thereby reducing their performance, fairness, and generalizability. E.g., neural hate speech detection models are strongly influenced by identity terms like gay, or women, resulting in false positives, severe unintended bias, and lower performance. Most mitigation techniques use lists of identity terms or samples from the target domain during training. However, this approach requires a-priori knowledge and introduces further bias if important terms are neglected. Instead, we propose a knowledge-free Entropy-based Attention Regularization (EAR) to discourage overfitting to training-specific terms. An additional objective function penalizes tokens with low self-attention entropy. We fine-tune BERT via EAR: the resulting model matches or exceeds state-of-the-art performance for hate speech classification and bias metrics on three benchmark corpora in English and Italian. EAR also reveals overfitting terms, i.e., terms most likely to induce bias, to help identify their effect on the model, task, and predictions.