Most classification studies are done by using all the objects data. It is expected to classify objects by using some subsets data in the total data. A rough set based reduct is a minimal subset of features, which has almost the same discernible power as the entire conditional features. Here, we propose multiple reducts with confidence, which are followed by the k-nearest neighbor to classify documents to improve the classification accuracy. To select better multiple reducts for the classification, we develop a greedy algorithm for the multiple reducts, which is based on the selection of useful attributes for the documents classification. These proposed methods are verified to be effective in the classification on benchmark datasets from the Reuters 21578 data set.