Research On Key Algorithms In Knowledge Base-oriented And Personalized-oriented Dialog System

The introduction of the Turing Test in 1950 caused a wave of research on the dialog system in academia.With the advent of information age in the 21 st century and the vigorous development of the Internet,people's daily activities have become more closely linked to the Internet.Therefore,it is easier to collect dialogue data from the Internet,which provides a good data foundation for the research of dialog systems.In recent years,because of the rise of deep learning and the improvement of computer hardware performance,the dialog system no longer only relies on rules matching and retrieval,but has gradually transformed into generative dialog systems.This thesis focuses on the open domain generative dialog system,which is different from the dialog system in a specific domain.Its goal is not to complete the tasks specified by users,but to attract and retain users,so that users are interested in continuing to talk.At present,the open-domain generative dialog systems mainly use sequence-to-sequence models,but there are some problems in the standard sequence-to-sequence model:(1)Due to the large number of security responses in the corpus,the model tends to generate short and tedious general responses;(2)Because the model does not incorporate external knowledge information,the generated response contains a small amount of information,which is not attractive;(3)Due to the lack of character personality in the model,the responses generated are usually inconsistent for input statements with the same or similar semantics,which reduces the user's trust in the model.In view of the above problems,the main work of this thesis is as follows:(1)An open domain response generation model KGDlg combined with knowledge base is proposed.The knowledge graph information is fused with the user input through the fusion network,and the fusion result is input into the decoder,so that the decoder can use the knowledge graph information related to the user input to generate a response.In addition,the model uses user responses(retrieval responses during the test phase)to modify the decoder's choice of knowledge graphs so that the decoder can utilize the correct knowledge graph information.(2)A personalization-oriented open domain response generation model,PersonalDlg,is proposed.In the model training phase,the ground truth is used as a posterior distribution to help judge the importance of each personalized information to generate a response.This distribution can give greater weight to the personalized information actually used in the ground truth,which helps guide the decoder to generate a response containing the correct personalized information.Since there is no ground truth during the test phase,a prior distribution is used to approximate the posterior distribution,so that even if there is no ground truth,the model can select the appropriate personalized information to generat the response.In general,this thesis mainly studies the knowledge base-oriented and personalizedoriented open domain response generation models and trains the two models on real datasets.Through comparative experiments,it is proved that the KGDlg model can generate information-rich responses,and the PersonalDlg model can generate personalized responses consistent with personalized information.