The　optimal　structure　of　theme　set　can　be　automatically　learned　from　the　data　with　Hierarchical　Dirichlet　Process　(HDP)　topic　model.　However,　the　set　of　topics　can　not　meet　the　semantic　requirement.　And　in　some　theme　models　with　labels　it　is　difficult　to　set　the　parameters.　Therefore,　based　on　the　known　semantic　labels　and　the　certitude　degree　of　labels,　a　semi-supervised　labeled　HDP　topic　model　(SLHDP)　and　the　accuracy　evaluation　index　of　random　cluster　are　proposed　in　this　paper.　Higher　weight　is　given　by　the　known　semantic　labels.　Combined　with　the　property　of　the　finite　space　being　divided　infinitely　in　Dirichlet　process,　the　model　is　built　via　Chinese　restaurant　process.　The　experimental　results　on　several　Chinese　and　English　datasets　show　that　SLHDP　model　makes　the　topic　set　more　reasonable　in　the　text　classification　of　large　scale　datasets.　©　2017,　Science　Press.　All　right　reserved.