As　a　simple　classification　method　VSM　has　been　widely　applied　in　text　information　processing　field.　There　are　some　problems　for　traditional　VSM　to　select　a　refined　vector　model　representation,　which　can　make　a　good　tradeoff　between　complexity　and　performance,　especially　for　incremental　text　mining.　To　solve　these　problems,　in　this　paper,　several　improvements,　such　as　VSM　based　on　improved　TF,　TFIDF　and　BM25,　are　discussed.　And　then　maximum　mutual　information　feature　selection　is　introduced　to　achieve　a　low　dimension　VSM　with　less　complexity,　and　at　the　same　time　keep　an　acceptable　precision.　The　experimental　results　of　spam　filtering　and　short　messages　classification　shows　that　the　algorithm　can　achieve　higher　precision　than　existing　algorithms　under　same　conditions.　©　2008　American　Institute　of　Physics.