Research On The Theory And Method Of Dialogue Response Diversity And Persona-consistency Based On Neural Network

In recent years,with the rapid improvement of computer computing power and the emergence of large-scale available dataset,artificial intelligence(AI)has been greatly developed.Dialogue system,as one of the main research directions of natural language processing,an important branch of artificial intelligence,has gradually attracted the attention of academia and industry and has become a current research hotspot.At present,although some progress has been made in the research on dialogue systems and the corresponding technical results have been applied to real-world scenarios such as customer service dialogues,legal consultations,and virtual assistants,they still face challenges from data,algorithms,and optimization.This article focuses on the research on the persona-consistency and diversity of responses in chit-chat conversational system.Although quite natural for human beings to communicate based on their own personality characteristics in daily life,it is quite challenging for neural dialogue systems to do the same.This is because the general dialogue systems are difficult to generate diverse responses while maintaining consistent persona information.Existing methods basically only focus on one of them,but for daily dialogue systems,the lack of role consistency in response will reduce the credibility of the system,and the lack of diversity will affect the fun and sustainability of humanmachine interaction(HMI).Therefore,this study considers improving the personaconsistency and diversity of responses at the same time.The research content specifically includes the following aspects:(1)Exploring the multi-stage response generation technology to achieve an organic combination of response diversity and role consistency research: It is difficult to achieve the unity of diversity and consistency performance by relying solely on a single end-toend model.Starting from the multi-stage response generation model,this paper proposes a two-stage response generation framework,which plays a role in linking the response diversity and persona-consistency.Under this framework,the response generation model we finally proposed successfully achieves simultaneous improvements.The diversity of responses and the consistency of persona information.(2)Exploring the methods of improving response diversity : First,through a comparative analysis of existing diversity-promoting methods and combining the specific needs of this research,we choose the Conditional Variational Autoencoder(CVAE)-based method as the basic method of response diversity research in this subject.Then,for the lack of persona-consistency in the latest model of CVAE-based method,we propose a diverse response generation model guided by conversational attribute,which uses the persona information used in the current dialogue as a prior knowledge to participate in the encoding process of latent variables to enhance the model's perception of persona information in different responses.This is helpful for the hidden variables to model the diversity response under the given persona setting.(3)Exploring the role consistency improvement method: In order to further improve the persona-consistency of diverse responses,we propose new persona-consistency checking and rewriting modules under the Two-stage response generation framework,that is,using pre-training model BERT detects the persona-consistency of the generated responses and marks inconsistent words,and then combines the pointer generation network to rewrite the inconsistencies through secondary decoding to obtain the final persona-consistent and diverse responses.We use a combination of objective and subjective evaluation to evaluate the performance of the proposed model.Experiments show that compared with the baseline model,our proposed model not only inherits its advantages in diversity,but also significantly improves the persona-consistency,which proves the effectiveness of our proposed model.