Evaluating UMLS strings for natural language processing

The National Library of Medicine's Unified Medical Language System (UMLS) is a rich source of knowledge in the biomedical domain. The UMLS is used for research and development in a range of different applications, including natural language processing (NLP). In this paper we investigate the nature of the strings found in the UMLS Metathesaurus and evaluate them for their usefulness in NLP. We begin by identifying a number of properties that might allow us to predict the likelihood of a given string being found or not found in a corpus. We use a statistical model to test these predictors against our corpus, which is derived from the MEDLINE database. For one set of properties the model correctly predicted 77% of the strings that do not belong to the corpus, and 85% of the strings that do belong to the corpus. For another set of properties the model correctly predicted 96% of the strings that do not belong to the corpus and 29% of the strings that do belong to the corpus.