Real-Time Profiling and Mitigation of Irony and Stereotype Spreaders on Twitter Using a Domain-Adaptive NLP Pipeline

Real-time detection and mitigation of irony and stereotype spreaders on Twit-ter is vital for content moderation. We propose a domain-adaptive NLP pipeline combining transformer models, graph neural networks, and reinforcement learning to profile and prioritize such users. Using domain-adaptive embeddings and user interaction analysis, our approach excels across domains like politics, entertainment , and sports. Evaluated on the TwiBot-22 dataset, augmented with additional irony and stereotype annotations from over 200 million tweets, it achieves a 5–8% F1-score increase (reaching 85–88%) and 4–7% accuracy boost (reaching 79–82%) over state-of-the-art baselines, offering a scalable solution for Twitter’s content ecosystem.