Mapping Escalatory Patterns of Violent Conflict with Deep Learning and News Text

• In Event: Political Forecasting Methods: Challenges and Potential Solutions

Abstract

Despite advancements in early warning systems for forecasting violent conflict, predicting novel conflicts remains a formidable challenge. One obstacle is that most existing frameworks focus on forecasting conflict fatalities while relying on past conflict patterns as the dominant predictor. This inherently limits our ability to identify new onsets. Moreover, the reliance on slow-moving or inadequately measured structural features, such as socioeconomic factors, also falls short of providing us with the necessary capability to predict novel conflicts. To overcome this critical limitation, we propose a framework for predicting escalatory patterns at the actor level. Our approach bridges the methodological gap between structured data, such as socioeconomic factors, and unstructured data, such as text. We integrate data on armed actors from the UCDP and VIEWS with attention/transformer-based neural networks in an innovative architecture. These networks are applied to a comprehensive corpus of newswire text from Factiva, enabling us to capture and utilize contextual information. By leveraging synergies between structured and unstructured data, our framework aims to enhance the reliability and effectiveness of conflict forecasting, particularly in identifying novel conflict onsets. Thus, this research represents a considerable contribution to the field of conflict prediction, offering valuable insights for policymakers and researchers alike.

Authors

• Mihai Croicu, Uppsala University, Dept. of Peace and Conflict Research

• Simon Polichinel von der Maase, Peace Research Institute Oslo