| Human-machine dialogue is an important research content in the field of natural language processing.Document based dialogue response generation technology is a direction of human-machine dialogue research.According to the application scenario,this technology can be divided into many types.Among them,knowledge response oriented human-machine dialogue pays attention to the scene where the user and the dialogue system communicate with some knowledge.The system obtains knowledge from documents and generates reasonable replies to inform users,which has high practical value.Therefore,this paper is aiming to study the document based dialogue response generation technology for knowledge response.At present,the relative researches are still in the primary stage,which is mainly reflected in two aspects:the responses generated by the system are not able to correctly reflect the document knowledge and the lack of reasonable automatic evaluation indicators to evaluate the ability of the response generated by the system to reflect the document knowledge.In order to solve the above problems,this study is carried out on the basis of extensive research.The specific contents are as follows:This paper proposes a new dataset for multi document driven knowledge oriented dialogue,which is called DocKDConv,and puts forward a new evaluation metric called knowledge exploit score(KES).Compared with other similar datasets,DocKDConv includes many fields and a large-scale document database.The dialogue system is required to obtain relevant documents by retrieval method,select document knowledge and generate reply,which makes it more similiar to the human to human dialogue scene.The KES score can well reflect the accuracy of the generated reply to the document knowledge by calculating the reflection of the generated reply to the marked knowledge.This paper proposes a response generation model based on topic retention and transfer and reply prototype method.In the knowledge selection stage,the model improves the accuracy of selecting documents by modeling the retention or transfer behavior of topics related to different turns of dialogue;In the reply generation stage,the model decomposes the target response into response prototype and target knowledge,obtains the response prototype by generation method,and obtains the target knowledge by retrieval method,which improves the accuracy of knowledge selection in the response.Based on the above data sets and models,we realized a multi domain and multi document response generation system.The users can interact with the system through dialogue to know the relevant information in the fields of film,music and tourism. |
