The　traditional　end-to-end　task-oriented　dialogue　models　are　based　on　complex　recurrent　or　convolutional　neural　networks　that　include　an　encoder　and　a　decoder.　But　in　the　case　of　large　amounts　of　data,　there　are　many　types　of　questions.　It　performs　poorly　when　answering　multiple　types　of　questions,　memory　information　cannot　effectively　record　all　the　sentence　information　of　the　context.　In　view　of　the　above　this,　this　article　uses　a　modified　transformer　model　to　overcome　the　problems　mentioned　in　dialogue　tasks.　Transformer　is　a　model　constructed　using　attention　mechanisms,　which　completely　discards　the　method　of　RNN　(recurrent　neural　networks),　and　its　structure　includes　two　sub-parts　of　Encoder　and　decoder.　It　uses　residual　network,　batch　normalization,　and　self-attention　mechanism　to　build　the　model　structure,　uses　Positional　Encoding　to　capture　sentence　information,　which　can　speed　up　model　training　convergence　and　capture　Longer　sentence　information.　In　this　paper,　we　modified　the　activation　function　in　the　transformer　and　use　label　smoothing　to　optimize　the　training　to　make　the　model＇s　expressive　ability　better　than　previous.　©　2019　Published　under　licence　by　IOP　Publishing　Ltd.