TARGE: large language model-powered explainable hate speech detection.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• User-generated content on social media poses challenges for detecting hate speech at scale.
• Traditional manual moderation is impractical; automated solutions are needed.
• Existing deep learning models lack transparency, limiting their utility in nuanced content moderation.

Purpose of the Study:

• To develop transparent hate speech detection systems using advanced language models.
• To leverage large language models (LLMs) for explainable content moderation.

Main Methods:

• Utilized Mistral-7B, a state-of-the-art language model.
• Developed a framework where LLMs generate explicit rationales for hate speech detection.
• Integrated LLM-generated rationales into specialized classifiers for explainable moderation.

Main Results:

• Incorporating LLM-generated explanations significantly enhanced hate speech detection interpretability.
• The methodology improved the accuracy of identifying inflammatory and discriminatory speech.
• The system provides clear analytical rationales for each moderation decision.

Conclusions:

• LLM-generated rationales enhance transparency and accuracy in automated hate speech detection.
• This approach addresses the critical demand for explainable AI in content moderation.
• The framework offers a practical solution for nuanced and scalable online content moderation.