Research On Personalized Dialogue Generation Method Based On Deep Learning

Social robots aim to meet different user needs in different dialogue scenarios by communicating with users,and its core is the dialogue generation model.According to different scenarios,the model can be divided into chatting dialogue model,question-answer dialogue model and task-orient dialogue model in vertical field.This thesis focus on the problem of personalized dialogue generation in chat dialogue model.Specifically,an excellent chatting dialogue model should be anthropomorphic,which are characterized by a fluent conversational response with description of personality-related attributes.Huge social networks and users with abundant information of personality make personalized expression a potential behavior in communication.Therefore,how to highlight personal attributes in a communication with users to make the conversation more personalized and anthropomorphic is an important topic in the study of chatting dialogue model.This thesis first describes some dialogue models and training strategies related to the topic,and analyzes the inherent defects of some current work of personalized dialogue generation.Then,based on the end-to-end dialogue model,this thesis solves problem from the perspectives of structured personality information and unstructured personality information respectively,and improves the overall performance of the personalized dialogue generation model.The main work and innovation of this thesis is as follows:(1)For unstructured personality information,a multi-task personalized dialogue generation method based on meta-learning is proposed.In the test phase,a few of personalityrelated dialogue history can be used to quickly fit the model to the target personality,and a dual decoder structure is proposed to alleviate the problem of data sparsity.The proposed method can effectively solve the problems that personal attributes are not prominent and cannot be expressed explicitly in communication,and the validity of the model is verified on public datasets.(2)For structured personality information,this thesis summarizes the pre-existing personality information of the users in social network based on the existing methods and combined with the mainstream social platforms,and analyses the difference between the public training set and the personality information in the actual scenario,as well as its impact on the performance of the model.A personalized conversation generation model with sparse tags is proposed to solve the performance degradation caused by the loss of user’s personality information in the real scenario.At the same time,a binary role information fusion module is designed to more effectively combine the personality information of both sides of the dialogue to generate more fluent and personalized responses.