Translating under pressure: domain-aware LLMs for crisis communication

The study is about developing a domain-adaptive approach to improve multilingual crisis communication in situations where high-quality parallel data are limited. Starting from a small reference corpus, the proposed pipeline retrieves and filters relevant data from general-domain corpora to create an expanded crisis-specific dataset. This dataset is then used to fine-tune a small language model for emergency translation, followed by preference optimization to produce translations in CEFR A2-level English, making emergency information more accessible to diverse audiences.

The findings show that the proposed approach successfully improves the readability of crisis translations while maintaining strong translation adequacy, as demonstrated through both automatic and human evaluation. The results suggest that combining domain adaptation with simplified English provides an effective and practical solution for emergency communication, enabling simplified English to serve as a reliable lingua franca in crisis scenarios when comprehensive multilingual translation is not available.