Research On Controlled Dialogue Generation Of Medical Dialogue System

Healthcare is a topic of constant concern to people,and we hope that we can get timely feedback when encountering health problems.Therefore,the application of the dialogue system to the medical field reflects its value,and the medical dialogue system can be used as a medium to provide users with medical knowledge inquiries.The dialogue system can be roughly composed of several parts from the system structure: dialogue understanding,dialogue management and dialogue generation.Among them,the dialogue generation module is a relatively important part,because the generated dialogue directly reflects the advantages and disadvantages of the dialogue system.At the same time,in the medical dialogue generation,it is a very big challenge to convert complex medical knowledge into smooth and natural dialogue.The current dialogue generation is mainly realized by the language model of deep learning.If we want to change the theme,emotion and key information carried in the generated sentences,we need to consider a controlled dialogue generation method.The control methods in the dialogue generation process can be divided into two categories,one is conditional dialogue generation,which adds additional conditions in the input stage of the language model;the other is constraint dialogue generation,which adjusts the original output to achieve control over the results.This paper aims to explore the method of generating controlled text in the context of medical dialogue generation,and integrate medical knowledge into the generated dialogue.The work done in the full text is as follows:(1)Conduct a comprehensive exploration of controlled text generation technology,including conditional text generation and constraint text generation.Because there are many and complex controlled text generation methods,few people have divided and classified them in the past.This paper sorts out the two types of controlled text generation methods,paving the way for the subsequent application of controlled text generation methods to medical dialogue generation.(2)In the conditional text generation method,the CONV-PE model is proposed to be applied to the open dialogue model,which carries the knowledge information generated by the medical entity as a controlled dialogue.The model is divided into two modules: dialogue understanding and dialogue generation.In the dialogue generation stage,the Prompt strategy is mainly used to add medical entities to change the input conditions of the model,so as to provide context information to the language model.After experimental verification,compared with the baseline model,the average improvement in various indicators is 10.8%.(3)In the conditional text generation method,the CONVE-SF model is proposed to realize the task-based dialogue model of slot filling to generate medical dialogues.The model is divided into three parts: dialogue understanding,dialogue management and dialogue generation,which are constructed in the dialogue generation stage Prompt mode applicable to Dialogue Act is generated.Through experimental verification,the model has an average improvement of 6.4% in various indicators of dialogue generation compared with the baseline model.(4)On the constrained text generation method,auxiliary training is proposed,and the generative models of the above two systems are replaced by constrained dialogue generation to compare the performance of the two control types.Auxiliary training is to add a trainable auxiliary model to encode constraint information,add its output to the original output of the model,so as to change the original probability distribution,and realize the controlled generation of constraints,which are carried out in the generation phase of the above two models respectively application.After comparison,the generation method using constrained text is slightly lower than conditional text generation in BLEU,but higher than conditional text generation in ROUGE metric.