The capture and symbolization of data from the clinical problem list facilitates the creation of high-fidelity patient resumes for use in aggregate analysis and decision support. We report on the development of a UMLS-based semantic parser and present a preliminary evaluation of the parser in the recognition and validation of disease-related clinical problems. We randomly sampled 20% of the 26,858 unique non-dictionary clinical problems entered into OMR (Online Medical Record) between 1989 and August, 1997, and eliminated a series of qualified problem labels, e.g., history-of, to obtain a dataset of 4122 problem labels. Within this dataset, the authors identified 2810 labels (68.2%) as referring to a broad range of disease-related processes. The parser correctly recognized and validated 1398 of the 2810 disease-related labels (49.8 +/- 1.9%) and correctly excluded 1220 of 1312 non-disease-related labels (93.0 +/- 1.4%). 812 of the 1181 match failures (68.8%) were caused by terms either absent from UMLS or modifiers not accepted by the parser; 369 match failures (31.2%) were caused by labels having patterns not recognized by the parser. By enriching the UMLS lexicon with terms commonly found in provider-entered labels, it appears that performance of the parser can be significantly enhanced over a few subsequent iterations. This initial evaluation provides a foundation from which to make principled additions to the UMLS lexicon locally for use in symbolizing clinical data; further research is necessary to determine applicability to other health care settings.