Human-Aligned Rationale Learning for Explainable Hate Speech Moderation: A Behavioral Evaluation Framework for Reliable Decision Support

Automated hate speech moderation systems are increasingly deployed to support large-scale platform governance, yet their reliability and transparency remain critical concerns. While recent advances in explainable AI (XAI) provide tools to interpret model decisions, these methods are often applied post hoc and rarely evaluated in terms of their impact on moderation behavior under real-world conditions.In this work, we present Human-Aligned Rationale Learning (HARL), an explanation-aware training framework that integrates human-annotated rationales into the optimization of hate speech detection models. HARL combines standard classification loss with an attribution alignment objective, encouraging models to ground predictions in human-identified indicators of harmful content. Rather than proposing a new rationale-learning paradigm, the framework is designed to systematically study how explanation-guided supervision influences moderation performance, explanation quality, and deployment-relevant behavior.We evaluate HARL across multiple hate speech detection benchmarks, examining classification performance, explanation plausibility, and partial faithfulness using token-level agreement metrics and prediction-drop analysis. Beyond standard evaluation, we introduce a behavioral analysis framework that assesses moderation outcomes across identity-linked content and varying degrees of code-switching, providing insight into how models behave under socially and linguistically variable conditions.Our results show that explanation-guided supervision improves explanation grounding while maintaining competitive classification performance across both high-resource and low-resource settings. Furthermore, different attribution methods exhibit distinct trade-offs between plausibility, faithfulness, and computational efficiency, highlighting their suitability for different deployment scenarios such as real-time moderation and offline auditing.We position HARL as an applied moderation framework for explanation-aware training and behavioral evaluation, rather than as a new fairness intervention. The findings provide practical insights into how explanation-guided learning can support more transparent, reliable, and inspectable moderation systems in real-world deployment contexts.