Epidemic Event Extraction from News Media using Large Language Models

Abstract

Extracting structured epidemic events from unstructured news is vital for public health surveillance, yet remains difficult in low-resource languages. We present a practical pipeline that leverages large language models (LLMs) to extract epidemic events from Khmer-language news, and introduce a majority-voting consensus framework to evaluate the quality of extraction without requiring expert annotations. Evaluating six LLMs across 100 Khmer health articles, we find that Claude attains the best overall performance (F1 = 0.63), while DeepSeek offers the strongest cost-effective open-source alternative (F1 = 0.60). Few-shot prompting does not consistently outperform zero-shot, suggesting that sufficient prior knowledge is available for this task. Preventive and symptomatic events are most prevalent in the corpus, while argument identification and role classification remain the main bottlenecks across models. Our contributions are: (1) the first systematic assessment of LLMs for Khmer epidemic event extraction; (2) a consensus-based evaluation method suited to low-resource settings; and (3) practical guidance on cost–performance trade-offs for real-world surveillance. These findings provide actionable baselines for deploying automated epidemic intelligence in low-resource environments.