A new paper on arXiv examines explanation fairness in hate speech detection, arguing that a model's high classification performance does not always mean its explanations are equally plausible for different demographic groups. The authors focus on how well model rationales align with human reasoning, rather than only looking at prediction accuracy.

Using the HateXplain dataset, the work studies how explanation-human alignment varies across demographic groups. To address the observed disparities, the paper proposes a rationale regularization approach, which appears designed to steer models toward explanations that are more consistent with human judgment across groups.

The abstract is truncated, so details of the method and experimental results are not fully available in the source. Still, the work highlights an important distinction: fairness in hate speech detection may require not just equitable predictions, but also equitable explanations.