Predicting organized collective violence through hostile discourse in social media

While substantial evidence indicates that patterns of organized collective violence can be influenced through social and political discourse, existing approaches to estimating these effects have often focused on specific actors, geographic regions, and languages, potentially obscuring the more general patterns that underlie efforts to promote collective violence across human societies. To address this, we leverage the capabilities of multilingual language models (MLLMs), combined with a language-agnostic spatio-temporal approach to sampling social media discourse, to train a neural network classifier that learns to detect systematic linguistic precursors to collective violence at global scales. Our findings show that this approach outperforms traditional machine learning models, generating more accurate predictions by leveraging the abilities of MLLMs to generalize across diverse cultural and linguistic contexts. We further demonstrate, using zero-shot natural language inference, that our model has learned to detect semantically meaningful emotional dimensions in social media texts which are systematically linked to the model’s ability to distinguish between the discursive precursors to violent conflict and the discursive consequences of violent conflict. By making all code, data, and trained models publicly available, we aim to provide a foundation for future development of conflict early-warning systems that can operate effectively at global scales.