Agentic Multilingual NLP for Conflict Forecasting from Open-Source Text Streams

Poli Nemkova, Mark V. Albert

Keywords: Natural Language Processing, Multilingual Models, Retrieval-Augmented Models, Agentic LLMs, Uncertainty and Calibration, AI for Social Good

Abstract: We present an agentic multilingual NLP framework for forecasting conflict and unrest from open-source text streams. Unlike static classifiers, our system iteratively refines predictions through an agentic reasoning loop, retrieving additional evidence and adapting prompts when uncertainty is high. It combines multilingual embeddings with region-aware modules and a hybrid retrieval–forecasting model, enabling robustness across diverse languages and geopolitical contexts. Preliminary experiments on news streams from seven countries reveal substantial regional variation in baseline performance (minority F1 = 0.07–0.55), motivating the need for adaptive approaches that improve calibration, fairness, and transparency in high-stakes forecasting.

Submission Number: 429