Related Experiment Videos
Evaluating UMLS strings for natural language processing.
A T McCray1, O Bodenreider, J D Malley
1National Library of Medicine, National Institutes of Health, USA.
Proceedings. AMIA Symposium
|February 5, 2002
Summary
This study evaluates strings from the Unified Medical Language System (UMLS) for natural language processing (NLP) utility. Findings show specific string properties can predict their presence in biomedical corpora like MEDLINE.
Area of Science:
- Biomedical Informatics
- Natural Language Processing
- Computational Linguistics
Background:
- The Unified Medical Language System (UMLS) is a key resource for biomedical knowledge.
- UMLS is utilized in various research and development applications, notably in Natural Language Processing (NLP).
Purpose of the Study:
- To investigate the characteristics of strings within the UMLS Metathesaurus.
- To assess the utility of these strings for NLP tasks.
- To develop predictive models for string presence in biomedical corpora.
Main Methods:
- Identification of string properties predictive of corpus presence.
- Application of a statistical model to test these predictors.
- Corpus derived from the MEDLINE database.
Main Results:
- One set of properties achieved 77% accuracy in predicting strings absent from the corpus and 85% accuracy for strings present.
- A second set of properties demonstrated 96% accuracy for predicting absent strings and 29% accuracy for present strings.
Conclusions:
- String properties can effectively predict their relevance and presence within biomedical text corpora.
- This predictive capability is valuable for optimizing NLP applications using the UMLS.