In　text　classification,　dimension　reduction　on　the　original　features　space　is　very　necessary　to　improve　accuracy　and　efficiency.　Feature　selection　is　a　kind　of　simple　and　effective　methods　in　dimension　reduction.　In　this　paper,　we　designed　two　feature　selection　methods　based　on　the　feature＇s　category　discriminability　(CD)　and　the　inflence　of　features　co-occurrence　to　classifcation.　The　experiment　show　that　our　methods　proposed　is　much　better　than　traditional　methods　and　the　classification　results　have　respectively　improved　by　5%　and　6.9%　at　most.　©2010　IEEE.