The text representation in text classification is usually a sequence of terms. As the number of terms becomes very high, it is greatly time-consuming to perform existed text categorization tasks. In this paper we presented a novel text representation model for text classification which greatly reduced the required resources. This model represents text with several features. Each feature corresponds to a theme that emerged from a set of related articles. We also introduce an efficient way to build the model. The proposed model has been applied to naive bayes classifier and experiments on Reuters-21578 corpus have shown that the efficiency is greatly improved without sacrificing classification accuracy even when the dimension of the input space is significantly reduced.