Toward Unified Moderation of Cyberbullying Across Social Media and Video Games

The growing integration of social media and video games in adolescents' daily lives has heightened the risk of cyberbullying, contributing to severe mental health challenges such as depression and suicidal ideation. Traditional moderation techniques, including manual review and rule-based filters, often fail to capture contextual nuances. Existing machine learning models also face limitations due to their reliance on large labeled datasets and lack of cross-platform adaptability. Our experiments further reveal that models trained on social media content perform poorly on video game data and vice versa, leading to inefficient moderation. To address these challenges, we propose a novel approach that leverages Large Language Models (LLMs) with Chain-of-Thought (CoT) reasoning for cyberbullying detection across both domains. This method eliminates the need for extensive training data while significantly improving accuracy. It achieves 90.3% and 91.1% accuracy on social media and video game datasets, respectively, offering a scalable and resource-efficient solution for cyberbullying moderation.