Long　tail　effect　and　excessive　out-of-vocabulary(OOV)　words　in　social　media　texts　result　in　severe　feature　sparsity　and　reduce　classification　accuracy.　To　solve　the　problem,　a　social　media　text　classification　method　based　on　character-word　feature　self-attention　learning　is　proposed.　Global　features　are　constructed　at　the　character　level　to　learn　attention　weight　distribution,　and　the　existing　multi-head　attention　mechanism　is　improved　to　reduce　parameter　scale　and　computational　complexity.　To　further　analyze　character-word　feature　fusion,　OOV　sensitivity　is　proposed　to　measure　the　impact　of　OOV　words　on　different　types　of　features.　Experiments　on　several　social　media　text　classification　tasks　indicate　that　the　effectiveness　and　classification　accuracy　of　the　proposed　method　are　obviously　improved　in　terms　of　fusing　word　features　and　character　features.　Moreover,　the　quantitative　results　of　OOV　vocabulary　sensitivity　index　verify　the　feasiblity　and　effectiveness　of　the　proposed　method.　©　2020,　Science　Press.　All　right　reserved.