FYI! -------------------------
COMPUTER SCIENCE COLLOQUIUM Kernel Methods for Word Sense Disambiguation and Abbreviation Expansion in the Medical Domain MAHESH JOSHI Computer Science Graduate Student Friday, June 23, 2006 10:00 a.m. HELLER HALL 306 ABSTRACT Word Sense Disambiguation (WSD) is the problem of automatically deciding the correct meaning of an ambiguous word based on the surrounding context in which it appears. The automatic expansion of abbreviations having multiple possible expansions can be viewed as a special form of WSD with the multiple expansions acting as "senses" of the ambiguous abbreviation. Both of these are significant problems even in a specialized domain of medical text such as abstracts of articles in scholarly medical journals and clinical notes taken by physicians. Most popular approaches to WSD involve supervised machine learning methods which require a set of manually annotated examples of disambiguated words or abbreviations to learn patterns that help in disambiguating future unseen instances. However, manual annotation imposes a limit on the amount of labeled data that can be made available to the supervised machine learning algorithms such as Support Vector Machines (SVMs). Kernel methods for SVMs provide an elegant framework to incorporate knowledge from unlabeled data into the SVM learners. This thesis explores the application of kernel methods to two datasets from the medical domain, one containing ambiguous words and the other containing ambiguous abbreviations. We have developed two classes of semantic kernels - Latent Semantic Analysis based Kernels and Word Association Kernels for SVMs, that are learned from unlabeled text containing the ambiguous words or abbreviations using unsupervised methods. We have found that our semantic kernels improve the accuracy of SVMs on the task of WSD in the medical domain. In particular, our second class of kernels (Word Association Kernels) perform better and the improvements are significant when the sense distribution for an ambiguous word is balanced. ------------------------ Yahoo! Groups Sponsor --------------------~--> Check out the new improvements in Yahoo! Groups email. http://us.click.yahoo.com/ulQhNB/fOaOAA/HwKMAA/x3XolB/TM --------------------------------------------------------------------~-> Yahoo! Groups Links <*> To visit your group on the web, go to: http://groups.yahoo.com/group/nlpatumd/ <*> To unsubscribe from this group, send an email to: [EMAIL PROTECTED] <*> Your use of Yahoo! Groups is subject to: http://docs.yahoo.com/info/terms/
