A Predictive Model for Anticipated Hate and Speech Violence in Social Media: Large Language Model Approach

This study explores the application of Large Language Models (LLMs), specifically a BERT-based architecture, coupled with predictive analytics for proactive hate speech and violence identification and forecasting in social media comment moderation. The research addresses the critical need to shift from reactive to proactive content moderation strategies in order to enhance digital safety and foster inclusivity. Using a dataset of 44,000 public comments from Facebook that are unfiltered and generalized, the methodology includes data collection through the Facebook Graph API, preprocessing phases, and model training using a balanced dataset to improve detection accuracy. The study highlights the importance of ethical deployment in artificial intelligence, noting the role of predictive analytics in identifying patterns and signals that indicate the existence of harmful content before its widespread circulation. From this, the results show that a balanced dataset helped the model to achieve strong performance metrics: accuracy of 82.63%, precision of 82.5%, recall of 82.88%, and F1 score of 82.69%. Results like this have shown the promise that advanced AI technologies hold when integrated into content moderation to successfully handle and pre-empt online hate speech and violence. This research helps to contribute to the larger discussion on ethical AI, responsible digital citizenship, and safer online communities through the promotion of proactive moderation systems.