A paper by Nari Yoo, professor at the University of Michigan School of Social Work and a member of MAP’s board, and MAP Executive Director Youngah Kim, ‘Agentic LLM for Information Access among Refugees: A Field Report from a Resource-Constrained Organization’, appears in the 5th NLP for Positive Impact workshop (NLP4PI), held with EMNLP 2026 in Budapest, Hungary, from October 24 to 29, 2026.
A government notice translated into a refugee’s language is still hard to act on when the Korean original is dense administrative prose. The paper reports what happened when MAP put a large language model in front of the translation step of Refugee News, which it has run since 2018. Scoped with the staff who answer refugees’ questions, the model rewrites the Korean source in plain language, explains Korea-specific terms, checks whether figures and deadlines are out of date, and fixes organisation names to a glossary before the text is translated into seven languages.
It covered most of the catalogue and shortened the sentences readers meet, and it also failed in ways a translation-only system never did. The stages allowed to search the web made corrections no translation could make, and they also wrote every serious error found, including policy updates that were simply wrong.