Leveraging Large Language Models for Cyberbullying Detection

Abstract

This work explores the application of Large Language Models (LLMs), such as LLaMA and Mistral, in advancing cyberbullying detection. Unlike traditional rule-based approaches, LLMs leverage contextual understanding and emotional cues to identify nuanced forms of online abuse, including sarcasm, coded language, and informal dialects. The study employs fine-tuned models adapted to annotated datasets, utilizing multitask learning for both hate classification and intent detection. Key challenges addressed include cultural bias, privacy risks, and the unpredictable nature of online language. Results indicate that LLMs significantly enhance detection accuracy and adaptability. Future directions involve integrating multimodal data (e.g., text with images or audio), leveraging federated learning for privacy preservation, and aligning detection systems with ethical AI frameworks to ensure fairness and user safety. This approach highlights the potential of LLMs to enable proactive and responsible cyberbullying mitigation on digital platforms