HateXScore: A Metric Suite for Evaluating Reasoning Quality in Hate Speech Explanations

TL;DR

HateXScore introduces a four-component metric suite to evaluate the reasoning quality of hate-speech explanations, addressing gaps in traditional label-centric metrics. By combining Hate-Type Check, Quotation Faithfulness, Target-Group Identification, and a Consistency Check, the framework probes explicit conclusions, causal grounding, policy-aware group targeting, and internal coherence. Evaluations across six multilingual datasets and seven LLMs show HateXScore uncovers annotation inconsistencies and interpretability failures invisible to accuracy or macro-F1, with strong alignment to human judgments. The approach provides a practical, diagnostic tool for transparent moderation and highlights when annotation labels warrant human review due to justification quality. It also demonstrates how explanation-focused evaluation can guide improvements in hate-speech detection systems and annotation practices.

Abstract

Hateful speech detection is a key component of content moderation, yet current evaluation frameworks rarely assess why a text is deemed hateful. We introduce \textsf{HateXScore}, a four-component metric suite designed to evaluate the reasoning quality of model explanations. It assesses (i) conclusion explicitness, (ii) faithfulness and causal grounding of quoted spans, (iii) protected group identification (policy-configurable), and (iv) logical consistency among these elements. Evaluated on six diverse hate speech datasets, \textsf{HateXScore} is intended as a diagnostic complement to reveal interpretability failures and annotation inconsistencies that are invisible to standard metrics like Accuracy or F1. Moreover, human evaluation shows strong agreement with \textsf{HateXScore}, validating it as a practical tool for trustworthy and transparent moderation. \textcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}

Figures (2)

• Figure 1: An example of a hate speech detection scored using HateXScore.
• Figure 2: Sensitivity of HateXScore to the Quotation Faithfulness threshold $\tau$ across six datasets. Each subplot shows the average HateXScore for all models as $\tau$ varies from 0.1 to 0.9. Higher curves indicate stronger consistency between explanation and decision.