The Research On Personalized Response Generation Method Of Open Domain Dialogue System

As a very important research subfield in the field of artificial intelligence,dialogue systems are also a challenging and important task in the field of human-computer interaction.The open-domain dialogue system communicates with users in the form of chitchat with no limited purpose,topic or domain,and can provide users with a more natural and humanized human-computer interaction experience.It has become a research topic that has been widely concerned by many researchers in recent years.At present,the open-domain dialogue generation system based on the Sequenceto-Sequence model has made great progress in the task of dialogue generation,but such models still suffer from the tendency to generate trivial and generic responses and the lack of consistent character personalities in dialogue content.To address such problems,the personalized dialogue system has been discovered by researchers and proved to be essential for realizing a natural and authentic dialogue systems.The goal of the personalized dialogue system is to allow the chatbot to use the pre-defined persona information to generate high-quality responses that not only conform to the personality of the person,but also remain consistent with the context.However,there are still some shortcomings in the performance of existing personalized dialogue generation models in in the performance of complex dialogue scenarios.The existing work mainly suffers from the following two shortcomings: 1)the accuracy of personality selection is low,and the personality characteristics in the generated responses do not match the current dialogue context;2)Insufficient utilization of personalized information makes it difficult to generate personalized information-rich responses.Based on the above problems,this article takes the improvement of the accuracy of personality selection and making full use of personality as the starting point to study the personalized response generation method of the open domain dialogue system.The main research contents include:(1)We propose a response generation method called Personalized Response Generation Model based on Context Enhancement.Firstly,the semantic features of the context are enhanced by extracting key information related to the conversation topic and personality in the context to improve the accuracy of persona selection.Secondly,a persona attention mechanism is designed in the decoding stage to promote the generation of persona information in responses.In addition,this method adds an additional cross-entropy loss to the model for persona selection by labeling personalized text labels to further select the correct persona.(2)We propose a response generation method called Persona Enhanced Dual Alternating Learning Network.This method includes two dialogue generation subnetworks with different learning goals.One of the sub-networks learns to select a proper persona as well as ensure the contextual relevance of the predicted response,while the other sub-network learns to enhance the utilization of persona information when generating the response by weakening the disturbance of specific content in the conversation context.The two sub-networks are built upon a common encoder-decoder backbone,and allow the model to learn persona selection and persona integration separately through multi-task alternate training,so as to generate more appropriate and richer personality in responses.In this paper,we conduct experimental validation on the public data set PersonaChat for personalized conversation tasks.The experimental results show that the proposed approach can effectively improve the diversity and personalization of generated responses compared with the existing conversation generation models.