In　recent　times,　numerous　models　tried　to　enhance　the　performance　of　Transformer　on　Chinese　NER　tasks.　The　model　can　be　enhanced　in　two　ways:　one　is　combining　it　with　lexicon　augmentation　techniques,　the　other　is　optimizing　the　Transformer　model　itself.　According　to　research,　fully　connected　self-attention　may　scatter　the　attention　distribution,　which　is　the　reason　for　worse　performance　of　the　original　Transformer　with　self-attention.　In　this　paper,　we　attempt　to　optimize　the　Transformer　model　especially　attention　layer.　Therefore,　a　novel　attention　mechanism,　Dilated　Shift　Window　Attention,　is　proposed　to　address　this　problem.　By　using　Window　Attention,　this　method　improves　the　model’s　capacity　to　deal　local　information,　meanwhile,　the　model　can　still　manage　long　text　and　long-distance　dependencies　owing　to　the　Window　Dilatation　mechanism.　Experiments　on　various　datasets　also　show　that　DSWA　replacing　fully　connected　self-attention　improves　the　model’s　performance　on　the　Chinese　NER　task.　Copyright　©　2023　by　KSI　Research　Inc.　and　Knowledge　Systems　Institute,　USA.