Inferring conflict fatalities from event narratives: a comparative evaluation of text-based classification models

Accurate identification of fatal conflict events is essential for timely monitoring in conflict-affected regions. While structured data are widely used, the predictive value of unstructured event narratives remains underexplored. This study examines whether textual narratives contain sufficient semantic information to distinguish fatal from non-fatal events, framing the task as a binary text classification problem and evaluating multiple modeling approaches within a unified framework. Results show that classical machine learning models provide reasonable baselines but struggle with detecting fatal cases. Deep learning models, particularly BiLSTM and CNN, improve performance by capturing contextual patterns, while Transformer-based models achieve the most robust and balanced results. BERT attains the highest overall and class-wise performance, with DistilBERT offering a more efficient alternative. These findings demonstrate that narrative text is a reliable indicator of conflict severity and highlight the value of contextual language models for conflict analysis.