TLC is a supervised training (S) system that uses a Bayesianstatistical model and features of a word's context to identifyword sense. We describe the classifier's operation and how itcan be configured to use only topical context cues, only localcues, or a combination of both. Our results on Senseval'sfinal run are presented along with a comparison to theperformance of the best S system and the average for S systems.We discuss ways to improve TLC by enriching its featureset and by substituting other decision procedures for the Bayesianmodel. Future development of supervised training classifiers willdepend on the availability of tagged training data. TLC canassist in the hand-tagging effort by helping human taggers locateinfrequent senses of polysemous words.