In this paper, we propose a weight-based feature extraction approach to reduce the number of features for text classification. The number of extracted features is equal to the number of document classes and the feature values are obtained according to the distributions of words over class partitions. Each word of the original word set contributes a weight to each extracted feature and a transformation matrix is formed. By using the transformation matrix, the original document set is converted to a new set with a smaller number of features. The proposed approach has two advantages. Trial-and-error for determining the appropriate number of extracted features can be avoided. Computation demand is small and the method runs fast. Experimental results obtained from real-world data sets have shown that our method can perform better than other methods.