White Paper: Human-AI Collaboration in Conflict Analysis: Text Classifier Development with Peacebuilders
Fine-tuned BERT models reduce cultural misclassification by involving local experts in annotation.
A new white paper from researchers in Kenya and Sudan details a collaborative approach to building AI text classifiers for conflict analysis. The team, including peacebuilders and data scientists, developed fine-tuned BERT-based models to detect online polarization and hate speech. By involving local practitioners in problem definition, annotation design, and iterative validation, the models achieved enhanced contextual alignment and significantly reduced misclassification driven by cultural nuance. The resulting open-source models—Kenya-polarization and Sudan-hate speech—are available on HuggingFace.
This participatory method not only improved technical performance but also increased practitioner ownership and trust in the AI tools. The study provides empirical evidence that involving domain experts in the AI development lifecycle can simultaneously improve technical robustness, contextual validity, and normative alignment. For tech professionals, this highlights a growing trend toward human-in-the-loop AI systems that prioritize cultural sensitivity and stakeholder engagement, especially in sensitive fields like conflict analysis and humanitarian work.
- Fine-tuned BERT classifiers were trained on collaboratively annotated datasets from Kenya and Sudan.
- Models showed enhanced contextual alignment and reduced misclassification due to cultural nuance.
- The open-source models (Kenya-polarization, Sudan-hate speech) are available on HuggingFace.
Why It Matters
Demonstrates that participatory AI development can improve both technical performance and real-world trust in sensitive domains.