When dealing with large amounts of textual information, we may require an automatic system to organize them into known taxonomies which are arranged in a hierarchy. This learning task is called hierarchical classification. In such case, usually there are huge numbers of terms. We need apply certain techniques to remove irrelevant and redundant features for saving computation time without losing too much classification accuracy. In this article, we will first propose a new feature selection method called MD. After that a new hierarchical classification method based on MD is proposed and compared with existing methods on a real dataset.