Abstract

With the growing use of Natural Language Processing (NLP) techniques for information extraction and concept indexing in the biomedical domain, a method that quickly and efficiently assigns the correct sense of an ambiguous biomedical term in a given context is needed concurrently. The current status of word sense disambiguation (WSD) in the biomedical domain is that handcrafted rules are used based on contextual material. The disadvantages of this approach are (i) generat ing WSD rules manually is a time-consuming and tedious task, (ii) maintenance of rule sets becomes increasingly difficult over time, and (iii) handcrafted rules are often incomplete and perform poorly in new domains comprised of specialized vocabularies and different genres of text. This paper presents a two-phase unsupervised method to build a WSD classifier for an ambiguous biomedical term W. The first phase automatically creates a sense-tagged corpus for W, and the second phase derives a classifier for W using the derived sense-tagged corpus as a training set. A formative experiment was performed, which demonstrated that classifiers trained on the derived sense-tagged corpora achieved an overall accuracy of about 97%, with greater than 90% accuracy for each individual ambiguous term.

Author Keywords: natural language processing; word sense disambiguation; corpus-based machine learning; MedLEE; UMLS; MEDLINE.