With　the　rapid　development　of　information　technology,　short　texts　arising　from　socialized　human　interaction　are　gradually　predominant　in　network　information　streams.　Accelerating　demands　are　requiring　the　industry　to　provide　more　effective　classification　of　the　brief　texts.　However,　faced　with　short　text　documents,　each　of　which　contains　only　a　few　words,　traditional　document　classification　models　run　into　difficulty.　Aggressive　documents　expansion　works　remarkably　well　for　many　cases　but　suffers　from　the　assumption　of　independent,　identically　distributed　observations.　We　formalize　a　view　of　classification　using　Bayesian　decision　theory,　treat　each　short　text　as　observations　from　a　probabilistic　model,　called　a　statistical　language　model,　and　encode　classification　preferences　with　a　loss　function　defined　by　the　language　models　and　the　external　reference　document.　According　to　Vapnik＇s　methods　of　Structural　risk　minimization　(SRM),　the　optimal　classification　action　is　the　one　that　minimizes　the　structural　risk,　which　provides　a　result　that　allows　one　to　trade　off　errors　on　the　training　sample　against　improved　generalization　performance.　We　conduct　experiments　by　using　several　corpora　of　microblog-like　data,　and　analyze　the　experimental　results.　With　respect　to　established　baselines,　results　of　these　experiments　show　that　applying　our　proposed　document　expansion　method　produces　better　chance　to　achieve　the　improved　classification　performance.