Research And Application Of Response Generation Model In Open-domain Dialogue System

With the advancement of large-scale parallel computing technology,dialogue system based on the neural network has developed rapidly in recent years.However,the following problems still exist in the current open-domain dialogue system: the dialogue model is easy to generate context-irrelevant,ungrammatical,and even self-contradictory responses.The model fails to complete the task of natural language understanding and generation.The research theme of this thesis is to build an open domain dialogue system that can generate high-quality responses.To this end,this thesis improves the existing model from the perspectives of language understanding and dialogue generation based on the pretrained language model.The main contents of this thesis are as follows:1)This thesis proposes a response-aware dialogue model which introduces the real response information.The model utilizes the response-aware method to extract hidden information in the real responses and generates responses closer to the real ones with the capability of the pretrained model.To solve the exposure bias problem caused by the introduction of real responses,this thesis uses scheduled sampling and responseprediction methods to bridge the gap between model input in the training and generation stages.Finally,the experimental results show that the model can perform better while combining scheduled sampling and response-prediction methods.In addition,considering the differences in model architecture between the pretrained model and the recurrent neural network,this thesis modifies the scheduled sampling method and makes it applicable to the pretrained model.This improves the computational efficiency of scheduled sampling.2)This thesis proposes a dialogue generation model based on natural language inference.This thesis utilizes natural language inference to filter irrelevant information from the complex dialogue context to correct the model input.To improve the accuracy of language inference,the model manages the dialogue context through natural language inference from the perspectives of dialogue contexts and real responses.The experimental results show that the filtered dialogue context can guide the model to produce better responses.Finally,this thesis also uses natural language inference to evaluate the consistency between the generated responses and the contexts,expanding the application of natural language inference in the dialogue system.3)This thesis uses the software engineering approach to analyze,design,and implement a dialogue system based on the improved dialogue model.To sum up,the work of this thesis focuses on improving the quality of generated responses in open-domain dialogue systems.The main idea is to improve the model's language understanding ability of real responses and dialogue contexts.Both automatic and human evaluation results illustrate the effectiveness of the proposed methods.