Methodology for creating UMLS content views appropriate for biomedical natural language processing

Given the growth in UMLS Metathesaurus content and the consequent growth in language complexity, it is not surprising that NLP applications that depend on the UMLS are experiencing increased difficulty in maintaining adequate levels of performance. This phenomenon underscores the need for UMLS content views which can support NLP processing of both the biomedical literature and clinical text. We report on experiments designed to provide guidance as to whether to adopt a conservative vs. an aggressive approach to the construction of UMLS content views. We tested three conservative views and two new aggressive views against two NLP applications and found that the conservative views consistently performed better for the literature application, but the most aggressive view performed best for the clinical application.