Understanding hate speech: the HateInsights dataset and model interpretability.

Area of Science:

• Natural Language Processing
• Computational Social Science
• Artificial Intelligence Ethics

Background:

• Online hate speech remains a significant challenge on social media platforms.
• Existing hate speech detection models lack sufficient interpretability and explainability.
• There is a need for comprehensive datasets that support explainable AI in this domain.

Purpose of the Study:

• Introduce the HateInsights dataset, a novel benchmark for hate speech research.
• Evaluate the explainability of state-of-the-art hate speech detection models.
• Investigate the impact of human-annotated rationales on model performance and explainability.

Main Methods:

• Developed the HateInsights dataset with dual annotations: 3-class classification (hate speech, offensive language, normal discourse) and supporting rationales.
• Utilized state-of-the-art models for hate speech classification and explainability evaluation.
• Analyzed model performance on both classification and explainability metrics (plausibility, faithfulness).

Main Results:

• Advanced hate speech detection models show limitations in explainability metrics like plausibility and faithfulness.
• Models trained with human-annotated rationales demonstrate improved explainability.
• The HateInsights dataset provides a valuable resource for advancing research in explainable hate speech detection.

Conclusions:

• Explainability remains a critical challenge in hate speech detection, even for high-performing models.
• Human-annotated rationales offer a promising avenue for enhancing model transparency and fairness.
• The release of the HateInsights dataset and codebase aims to foster collaborative research and drive progress in ethical AI for content moderation.