Sentiment　classification　for　reviews　has　attracted　increasingly　more　attention　from　the　natural　language　processing　community.　By　embedding　prior　knowledge　into　learning　structures,　classifiers　often　achieve　a　better　performance　than　original　methods.　In　this　paper,　we　propose　a　sophisticated　algorithm　based　on　deep　learning　and　information　geometry　in　which　the　distribution　of　all　training　samples　in　the　space　is　treated　as　prior　knowledge　and　is　encoded　by　deep　belief　networks　(DBNs).　From　the　view　of　information　geometry,　we　construct　the　geodesic　distance　between　the　distributions　over　the　features　for　classification.　The　study　of　the　distributions　contributes　to　the　training　of　the　DBN,　since　the　distance　is　correlated　to　the　error　rate　in　the　classification.　Finally,　we　evaluate　our　proposal　using　empirical　data　sets　that　are　dedicated　for　sentiment　classification.　The　results　show　that　our　algorithm　results　in　a　significant　improvement　over　existing　methods.