Research And Implementation Of Topic Coherence And Reply Diversity In Dialogue System

In the field of natural language processing(NLP),the desire for intelligent robots that can communicate with humans without barriers has greatly promoted the development of topics related to human-computer dialogue.In recent years,a large amount of manpower and material resources have been invested in related research and development.It is of great significance to promote the development of dialogue robots and accelerate the landing of related industries.This paper focuses on the research of the open-domain multi-round dialogue under the field of human-computer dialogue.Aiming at the two major problems that still exist in the current multi-round dialogue model,namely,low topic coherence and lack of diversity in responses,this paper proposes two innovative algorithms and build a WEB dialogue robot system.The specific work content is as follows:First,the GVDialog algorithm is proposed and implemented for the problem of low topic coherence.On the basis of the hierarchical dialogue model ReCoSa,a layer of variational autoencoder with random reconstruction context task as the training target is added,which can capture global information such as topics,dialogue background,etc.without manual labeling,and use these information to guide the model to generate topics-related replies.Both the comparative metric-based evaluation and human judgment with the baseline model prove that the algorithm can improve the coherence of responses.Second,to deal with the lack of diversity in model replies,this paper proposes the DisentFusion algorithm to achieve stylized generation,using different styles to improve the diversity of replies.This model proposes the separation of latent variables at three levels of style,theme,and content on the large-scale pre-training model Optimus,and designs a method of dilution and exclusion to better separate latent variables.In experiments against the baseline model,the algorithm exhibits a high degree of stylization and maintains good content consistency.Third,this paper implements a WEB dialogue robot system.The front end uses the Vue framework to build a chatbot interface,and the back end uses Flask to design a data interface.All the above two algorithms are implemented and deployed to the back end,and finally a complete dialogue system is realized.The system test shows that the system can realize the basic function of dialogue and meet the real-time requirement.