Research On Dialogue Generation With Multi-Information Fusion

Dialogue system is a human-computer interaction technology that enables communication between machines and humans through natural language processing and machine learning.It can be applied in various fields such as customer service,intelligent voice assistants,and smart homes,making people’s lives more convenient.However,traditional dialogue systems offen only allow for simple and mechanical question-and-answer task due to a lack of knowledge,personality,and emotion.By adding knowledge,dialogue systems can become more intelligent and accurate,while adding personality can make them become more personalized and friendly.Additionally,incorporating emotions can make dialogue systems more human-like.To make the dialogue system more intelligent,personalized and human-like,this thesis conducts research on dialogue generation from three perspectives:knowledge,personality,and emotion.The main contributions are listed as follows:Firstly,in order to make dialogue systems more intelligent and accurate,this thesis proposes a dialogue generation method that integrates knowledge and soft label from a knowledge perspective.During the training process,the method improves the accuracy of knowledge selection by using the reference response as posterior,and integrates selected knowledge into the dialogue to enhance the model’s understanding of the dialogue.At the same time,it converts hard labels into soft labels to mitigate model’s over-confidence.The experimental results shows that by adding knowledge,this method can provide the model with more information,while adopting soft label can effectively alleviate the model’s overconfidence,thus improving the diversity and robustness of the model.Secondly,in order to make dialogue systems more personalized and friendly,this thesis proposes a dialogue generation method that integrates demographic from the perspective of personalization.This method uses dialogue context and demographic of characters to select relevant knowledge,and uses graph neural networks to incorporate the selected knowledge into the dialogue.Then,through contrastive learning,the method enhances the personality of generated responses by contrasting with knowledge that differs from the demographic.The experimental results show that demographic play a crucial role in knowledge selection and dialogue generation.Finally,in order to make dialogue systems more human-like,this thesis proposes a method of incorporating emotional information into dialogue generation based on the role of emotions in dialogue,building upon the integration of demographic.The method selects knowledge from three perspectives:dialogue context,demographic,and emotions,and uses graph attention mechanism to learn knowledge expression in specific context.Then,by considering the context,demographic,and emotions,the method incorporates contextual knowledge into dialogue generation to enhance the personalized and emotional expression ability of generated responses.The experimental results show that emotional information not only helps the model select relevant knowledge but also promotes the integration of knowledge.