This paper proposes a federated learning driven intelligent translation framework to solve the problems of corpus dispersion, non-uniform terminology and insufficient cultural adaptation in cross-regional earthquake education text translation. Based on Transformer, the framework integrates region Adapter, seismic term alignment and cultural adaptive feedback mechanism to realize multi-region collaborative training without uploading the original corpus. The experiment is based on 30600 bilingual sentence pairs for earthquake education. The results show that the BLEU of FedCA-NMT reaches 42.7, the COMET reaches 0.842, the term accuracy is 95.1%, and the cultural adaptability score is 87.4, which are better than those of the comparison models. Research shows that this method can improve the translation accuracy and action guidance effect of campus evacuation, community risk avoidance and disaster prevention for foreign residents.
