Mistake-driven learning with thesaurus for text categorization

This paper extends the mistake-driven learner WINNOW to better utilize thesauri for text categorization. In our method not only words but also semantic categories given by the thesaurus are used as features in a classi- fer. New fltering and disambiguation methods are used as pre-processing to solve the problems caused by the use of the thesaurus. In order to verify our methods, we test a large body of tagged Japanese newspaper articles created by RWCP1 . Experimental results show that WINNOW with thesauri attains high accuracy and that the proposed fltering and disambiguation methods also contribute to the improved accuracy.