Since　the　introduction　of　the　GPT　model　as　the　mainstream　model　for　dialog　generation,　a　hot　issue　is　how　to　enable　the　model　to　extend　the　prediction　length　and　generate　longer　dialog　texts　while　the　training　context　length　remains　constant.　Rotary　incorporates　absolute　position　information　in　the　form　of　relative　position　encoding,　an　approach　that　has　been　shown　to　be　advantageous　for　extrapolating　on　transformer　models.　However,　in　the　actual　extrapolation　process,　when　the　prediction　length　is　too　long,　its　perplexity　climbs　sharply.　In　order　to　solve　the　above　problems,　this　paper　proposes　a　method　of　scaling　frequency　basis.　In　contrast　to　the　previous　manual　test　for　selecting　the　optimal　parameter　I＇,　this　method　allows　the　model　to　dynamically　train　the　parameters　based　on　the　input　sequence　length　to　obtain　the　optimal　I＇.　Using　this　method,　we　demonstrate　that　the　GPT　model　can　effectively　extrapolate　to　contexts　that　are　longer　than　its　original　pre-Training　length.　In　addition,　in　order　to　verify　the　effectiveness　of　this　method　in　practical　applications,　this　paper　applies　it　to　the　field　of　medical　dialog　generation,　which　is　a　more　complex　scenario,　and　the　experimental　results　demonstrate　that　the　extrapolation　length　is　several　times　the　length　of　the　training　context,　which　greatly　enhances　the　extrapolation　performance　compared　to　the　original　absolute　position　coding.　©　2023　ACM.