Word sense disambiguation in clinical text

Abstract

Lexical ambiguity, the ambiguity arising from a string with multiple meanings, is pervasive in language of all domains. Word sense disambiguation (WSD) and word sense induction (WSI) are the tasks of resolving this ambiguity. Applications in the clinical and biomedical domain focus on the potential disambiguation has for information extraction. Most approaches to the problem are unsupervised or semi-supervised because of the high cost of obtaining enough annotated data for supervised learning. In this thesis we compare the application of a semi-supervised general domain state of the art WSI method to clinical text to the best known knowledge-based unsupervised methods in the clinical domain. We also explore making improvements to the general domain method, which is based on topic modeling, by adding features that incorporate syntax and information from knowledge bases, and investigate ways to mitigate the need for annotated data.