Fostering Multilinguality in the UMLS: A Computational Approach to Terminology Expansion for Multiple Languages

We here report on efforts to computationally support the maintenance and extension of multilingual biomedical terminology resources. Our main idea is to treat term acquisition as a classification problem guided by term alignment in parallel multilingual corpora, using termhood information coming from of a named entity recognition system as a novel feature. We report on experiments for Spanish, French, German and Dutch parts of a multilingual UMLS-derived biomedical terminology. These efforts yielded 19k, 18k, 23k and 12k new terms and synonyms, respectively, from which about half relate to concepts without a previously available term label for these non-English languages. Based on expert assessment of a novel German terminology sample, 80% of the newly acquired terms were judged as reasonable additions to the terminology.