Title:	Deep Learning and Natural Language Processing for Conflict Prediction
DNr:	NAISS 2025/22-697
Project Type:	NAISS Small Compute
Principal Investigator:	Julian Walterskirchen <julian.walterskirchen@gu.se>
Affiliation:	Göteborgs universitet
Duration:	2025-05-07 – 2026-06-01
Classification:	50604
Keywords:

Abstract

This project explores the potential of applying deep learning and natural language processing (NLP) techniques to enhance conflict early-warning models. Previous research in conflict forecasting highlights three primary areas needing improvement: effectively capturing spatiotemporal dependencies and nonlinear relationships in conflict variables, improving the interpretability of complex predictive algorithms without compromising their accuracy, and quantifying prediction uncertainty. To address these challenges, we employ deep learning and NLP methods, leveraging their proven success across various research domains. Recognizing that successful NLP applications in other fields may not directly translate to conflict research, we conduct a systematic evaluation of prominent NLP techniques specifically for conflict prediction tasks. Additionally, we investigate deep learning models tailored for regular time series forecasting tasks, with a particular emphasis on temporal fusion transformers. Our analysis compares features derived from a dedicated conflict dictionary, two sentiment dictionaries, a word-scaling approach, dynamic topic modeling, transformer-based models, and temporal fusion transformers on a standard conflict prediction scenario. Through this investigation, we aim to provide valuable insights into the effectiveness of NLP-driven and deep learning-based approaches in predicting conflicts.