Summary
Some computer programs that spot hateful messages explain their decisions differently depending on who the messages are about, which is unfair. The authors looked at this problem using a special dataset with human explanations and made a way to measure explanation fairness. They found that a particular method called adversarial training helped their models give better and fairer explanations for all groups. However, making the predictions fair did not always fix explanation fairness, so these are two separate challenges. They also created a new approach to reduce these disparities, helping make AI decisions more understandable and fair.
What this means in practice
- •For content moderation teams: Enhance hate speech detection systems to provide more consistent and fair explanations across diverse user groups, improving transparency and trust.
- •For ai ethics compliance teams: Develop methods to measure and reduce explanation disparities in AI models, supporting regulatory and ethical standards for fairness in automated decisions.
Tested on one dataset.
Abstract
We study explanation fairness in hate speech detection. High classification performance does not always mean that a model produces equally plausible explanations for different demographic groups. Using the HateXplain dataset, we introduce Plausibility-EFG, a metric that measures the largest difference in explanation plausibility across demographic groups by comparing model-generated token attributions with human-annotated rationales. We evaluate BERT, RoBERTa, and ADV-RoBERTa, an adversarially trained model, using Integrated Gradients and SHAP. Our results show that ADV-RoBERTa improves classification performance and alignment between explanations and human rationales while reducing explanation disparities across demographic groups. These improvements are observed with both attribution methods and remain consistent across training, validation, and test splits. We also show that reducing prediction-level bias alone does not eliminate disparities in explanations, suggesting that predictive fairness and explanation fairness are related but distinct objectives. Finally, our fairness-performance analysis shows that stronger adversarial regularization can further reduce Plausibility-EFG, although it may cause a small decrease in macro-F1. Overall, this work presents explanation fairness as an important objective in hate speech moderation and introduces Fair-REF, a framework designed to reduce disparities in model explanations across demographic groups.