融合背景知识和常识感知的对话生成OA北大核心CSTPCD

One of the key issues in background-based conversation is knowledge extraction.However,due to the insufficient information in some conversations,especially when there is less information in certain dialogues,choosing the appropriate knowledge becomes particularly challenging.Furthermore,the current generation methods lack the capability to dynamically se-lect background knowledge.To address these issues,this paper proposed the KIF model,which incorporated a knowledge en-hancement library and knowledge vectors and introduced a knowledge tracking module and a knowledge sentiment feedback module to solve the aforementioned problems.The model obtained weight vectors of external knowledge and background know-ledge through a dual matching matrix method and performs knowledge selection.For each decoding step,the model generated conversation based on historical dialogues and external knowledge.Finally,experiments on the Holl-E and WoW datasets show that the KIF model significantly outperforms previous models.