Classification and clustering are frequently-used methods in data excavation technology. This paper introduces the idea of text clustering into the categorization algorithm study. The authors also attempt to use the text categorization pattern of self'-initiated learning to design a clustering-based text categorization algorithm, in the purpose of reducing the dimension of training set and raising the efficiency of categorization implement. A series of experiments prove that this algorithm can greatly raise the efficiency while slightly reducing the accuracy of categorization, and then balance the contradiction between them.