Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Newswire to Nexus: Using text-based actor embeddings and transformer networks to forecast conflict dynamics

Mihai Croicu, Simon Polichinel von der Maase

TL;DR

The paper tackles forecasting conflict dynamics at the actor level by marrying text-based actor embeddings with transformer-based forecasting, addressing the gap in predicting escalation and de-escalation. It builds a hybrid training corpus by aligning a large Factiva newswire collection with UCDP GED event data, retrains a domain-specific encoder (ConfliBERT), and then fine-tunes larger language backbones (DeBERTa and Mistral) to predict a four-state dynamic at the dyad-month level. The methodology introduces two data-augmentation strategies—low-context digests and high-context (RAG) digests—to provide rich, context-laden inputs for forecasting, and employs a Gaussian-process-based target measure to capture dynamics. Results show that encoder-based DeBERTa methods outperform a strong baseline on nowcasting and 1-month forecasts, with diminishing gains at longer horizons, while Mistral underperforms, highlighting data-corpus biases and the need for more integrated, end-to-end approaches for robust long-horizon predictions.

Abstract

This study advances the field of conflict forecasting by using text-based actor embeddings with transformer models to predict dynamic changes in violent conflict patterns at the actor level. More specifically, we combine newswire texts with structured conflict event data and leverage recent advances in Natural Language Processing (NLP) techniques to forecast escalations and de-escalations among conflicting actors, such as governments, militias, separatist movements, and terrorists. This new approach accurately and promptly captures the inherently volatile patterns of violent conflicts, which existing methods have not been able to achieve. To create this framework, we began by curating and annotating a vast international newswire corpus, leveraging hand-labeled event data from the Uppsala Conflict Data Program. By using this hybrid dataset, our models can incorporate the textual context of news sources along with the precision and detail of structured event data. This combination enables us to make both dynamic and granular predictions about conflict developments. We validate our approach through rigorous back-testing against historical events, demonstrating superior out-of-sample predictive power. We find that our approach is quite effective in identifying and predicting phases of conflict escalation and de-escalation, surpassing the capabilities of traditional models. By focusing on actor interactions, our explicit goal is to provide actionable insights to policymakers, humanitarian organizations, and peacekeeping operations in order to enable targeted and effective intervention strategies.

From Newswire to Nexus: Using text-based actor embeddings and transformer networks to forecast conflict dynamics

TL;DR

The paper tackles forecasting conflict dynamics at the actor level by marrying text-based actor embeddings with transformer-based forecasting, addressing the gap in predicting escalation and de-escalation. It builds a hybrid training corpus by aligning a large Factiva newswire collection with UCDP GED event data, retrains a domain-specific encoder (ConfliBERT), and then fine-tunes larger language backbones (DeBERTa and Mistral) to predict a four-state dynamic at the dyad-month level. The methodology introduces two data-augmentation strategies—low-context digests and high-context (RAG) digests—to provide rich, context-laden inputs for forecasting, and employs a Gaussian-process-based target measure to capture dynamics. Results show that encoder-based DeBERTa methods outperform a strong baseline on nowcasting and 1-month forecasts, with diminishing gains at longer horizons, while Mistral underperforms, highlighting data-corpus biases and the need for more integrated, end-to-end approaches for robust long-horizon predictions.

Abstract

This study advances the field of conflict forecasting by using text-based actor embeddings with transformer models to predict dynamic changes in violent conflict patterns at the actor level. More specifically, we combine newswire texts with structured conflict event data and leverage recent advances in Natural Language Processing (NLP) techniques to forecast escalations and de-escalations among conflicting actors, such as governments, militias, separatist movements, and terrorists. This new approach accurately and promptly captures the inherently volatile patterns of violent conflicts, which existing methods have not been able to achieve. To create this framework, we began by curating and annotating a vast international newswire corpus, leveraging hand-labeled event data from the Uppsala Conflict Data Program. By using this hybrid dataset, our models can incorporate the textual context of news sources along with the precision and detail of structured event data. This combination enables us to make both dynamic and granular predictions about conflict developments. We validate our approach through rigorous back-testing against historical events, demonstrating superior out-of-sample predictive power. We find that our approach is quite effective in identifying and predicting phases of conflict escalation and de-escalation, surpassing the capabilities of traditional models. By focusing on actor interactions, our explicit goal is to provide actionable insights to policymakers, humanitarian organizations, and peacekeeping operations in order to enable targeted and effective intervention strategies.
Paper Structure (18 sections, 4 equations, 5 figures)

This paper contains 18 sections, 4 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Literature review and theoretical considerations
  3. Existing approaches
  4. Gap in research: Actors and agency
  5. Where does the conflict data come from?
  6. Conflict patterns and the potential of LLMs for forecasting
  7. Conflict dynamics, escalation and de-escalation
  8. Methodology
  9. Extracting the corpus
  10. Matching with UCDP GED
  11. Assigning articles to actor-dyads and filtering out irrelevant items
  12. Generating contextual data
  13. The low-context approach
  14. The high-context, RAG approach
  15. Measuring escalation
  16. ...and 3 more sections

Figures (5)

  • Figure 1: The general architecture of the proposed text-to-forecast end-to-end system. Note that matching with UCDP articles is only needed at training time; the trained classifier can be used for dyad-months not containing any labelled data.
  • Figure 2: Multi-class micro-aggregated recall plotted against training epoch across the 40 experiments using the four different base BERT-based classifiers.
  • Figure 3: The fitted multi-level temporal Gaussian process smoother for three UCDP dyads in Yemen and Somalia (blue), including 90,95 and 99% credible intervals against the raw observed data. The first derivative of this trend is used as the target in our forecasting experiment.
  • Figure 4: Main metrics for the experiments over four time steps (now-casting, 1-month ahead, 3-months ahead and 6-months ahead). Multi-class prediction (4 classes, escalation, reduction, zero, plateau), micro-aggregation where probabilities are involved (ROC, AP). Confidence intervals were obtained from 1000 bootstraps of the predictions.
  • Figure 5: Per class predicted probabilities for two of the dyads containing the most text, forecast one month ahead. Those squares that are outlined in black and covered in hatches represent the actual class, as observed. For both predicted probabilities and displayed actuals, the newswire text used in the model to predict is one month into the past from the forecasting horizon; the calendar month is on the horizontal axis.