This paper presents a new approach for combining different semantic disambiguation methods that are part of a Word Sense Disambiguation(WSD) system. The way these methods are combined greatly influences the overall system performance. The approach is based on generating training examples, for each sense of the word, based on the output of each disambiguation method. A set of rules is learned from the training examples and then applied to optimize the output of the WSD system. We tested this approach on disambiguating WordNet glosses. However the approach is applicable to any WSD system. Our approach yielded a 3% gain in performance when compared with more traditional approaches such as selecting the sense given by the best disambiguation method or summing up the contribution of each method.