There　is　a　problem　that　the　difficulty　in　text　classification　will　increase　when　the　number　of　classes　increases,　to　which　hierarchical　structure　is　a　viable　solution.　Well,　a　document＇s　hierarchical　structure　is　usually　maintained　only　by　hand,　which　require　substantial　manpower　to　find　the　correct　position　of　a　document　in　the　class　hierarchy　or　to　reconstruct　the　hierarchy.　Constructing　the　hierarchical　structure　automatically　by　clustering　the　training　samples　can　effectively　reduce　the　cost　of　manual　maintenance,　and　at　the　same　time,　it　can　avoid　the　conflict　between　the　prior　knowledge　and　the　statistical　properties　of　the　sample　set　caused　by　artificial　maintenance　of　the　hierarchy.