Grounded Plain-Language Narratives for Humanitarian Dashboards: An Empirical Evaluation on UNHCR-Derived Refugee Trends and OCHA FTS Funding Flows

Abstract

Public humanitarian dashboards are used to communicate displacement and funding evidence without turning people into abstractions or making unverified claims. This paper reports a completed empirical evaluation of an evidence-locked narrative generation workflow for two NGO communication tasks: plain-language annotation of refugee trend charts and transparent explanation of OCHA Financial Tracking Service funding flows. The study used frozen, reproducible snapshots derived from UNHCR Refugee Data Finder and OWID/World Bank refugee indicators for 2012–2021, together with OCHA FTS 2021 country, sector, and donor funding snapshots for six response plans. Four generation conditions were evaluated: a deterministic numeric template, a generic unguided LLM-style prompt, a grounded LLM prompt, and a grounded prompt with humanitarian-safety constraints. The experiments generated and scored 23 refugee-trend annotations and 42 funding explanations, producing all reported tables and figures from the included code. The grounded+safety condition achieved 100.0% factual accuracy, 100.0% specificity on refugee trends, 100.0% transparency on funding explanations, and 100.0% caveat coverage in both tasks. The generic prompt produced factual accuracy of 60.87% for refugee trends and 36.31% for funding explanations because rounded claims, missing denominators, and absent gap logic broke source consistency. Results show that dashboard storytelling can combine readable language, empathy, and accountability when narrative generation is locked to computed facts and audited before publication.