Multi-Turn Dialogue Model And Application Based On Graph Neural Network

Dialogue system has always been the focus of artificial intelligence research.The intelligent dialogue system is very important for the future research of human-computer interaction.The open-domain dialogue system also has been proved to be more important than task-based dialogue systems in many fields.At present,the end-to-end dialogue generation model is widely used in the field of open-domain dialogue,but the end-to-end model has some disadvantages.For example,the results generated by computer tend to be generalized,the emotional expression of human dialogue cannot be simulated,the dialogue with purpose cannot be realized,and the topic transfer in multiple rounds of dialogue is poor,and so on.For the end-to-end multi-round dialogue system,the lack of high-quality multiround dialogue corpus is the main reason for the above problems.This thesis tries to solve these problems by introducing external knowledge.Knowledge is generally constructed in the way of graph,and graph neural network is the best choice for studying and processing graph information.In this thesis,a Chinese GPT multi-round dialogue Model was first implemented according to the GPT Model,which was better than the baseline.Then,this thesis constructs a dialogue response Model with emotional features.For this purpose,this thesis proposed for the first time to use external knowledge to enrich sentence-level feature information and build the heterogeneous figure.On the basis of heterographs,this thesis uses the two-layer graph attention mechanism and implement the graph convolution model on heterographs.Then this thesis uses the graph attention mechanism and graph convolution to classify sentence emotion,which is better than others.This proves the importance of external knowledge for the characterization of sentence emotion features.Then,according to the existing Emotional conversation model,this thesis also extended the sentence features and improved the model to achieve a better Emotional conversation model.In order to make dialogue generation have a certain purpose,this thesis proposes a dialogue generation model with objective programming based on external knowledge graph.The model divides the dialogue task into three levels,builds the specific model at different levels,and then combines the model and carries out joint training on the model in reinforcement learning environment,so as to realize the goal planning dialogue generation model based on knowledge graph.This thesis compares the effect of the improved dialogue model with the existing model and proves the importance of the diversity of external knowledge for the diversity of dialogue.Finally,according to the content of the research and the model of this thesis,the open-domain dialogue system in the mental health assessment system is designed and implemented in accordance with the software engineering standards.The system is based on open human-computer voice interaction to realize the assessment of human mental health.