Research On Dialogue Text Summarization Method Based On Knowledge Graph

The summarization of the dialogue text aims to extract,summarize,and conclude the main content of a conversation.In recent years,with the increasing number of online meetings,online consultations,and other scenarios,the summarization of the dialogue text has gradually become a hot research topic.There are significant differences in structure and language habits between dialogue texts and traditional paragraph texts.In the current research,there are still many unresolved issues,such as personal reference errors,dialogue interruptions,topic shifts,factual errors,etc.This thesis focuses on the summarization of the dialogue text as the topic,introduces knowledge graph technology,and explores how to solve the problems of personal reference errors and factual errors in the summarization of the dialogue text.Starting from the structure of dialogue text,this thesis focuses how to build a corresponding summarization generative model according to the structure of dialogue text and how to introduce knowledge graph technology into the model.Firstly,this thesis constructs a dialogue structure diagram based on the dialogue text,and integrates the knowledge nodes queried from the knowledge graph into the dialogue structure diagram.The purpose is to use the graph structure to describe the relationship between the speaker and the discourse in the dialogue text,and to associate the relationship between the discourses with the knowledge nodes.Secondly,this thesis constructs a node encoder and a transformer based heterogeneous graphics encoder to encode the dialogue structure diagram,and generates summarization through a decoder.This thesis uses the Rouge evaluation method to test model performance on the SAMSum dataset.The experimental results showed that the knowledge graph based dialogue text summarization model constructed in this thesis achieved a Rouge-1 score of 41.08%,a Rouge-2 score of 17.02%,and a Rouge-L score of 38.04%.Compared with the Transformer model of using end-to-end sequences,the Rouge-1 score increased by 4.46%,the Rouge-2 score increased by 5.84%,and the Rouge-L score increased by 4.98%.Through ablation experiments and summarization quality analysis,it was verified that the addition of knowledge graph helps to improve the problems of factual error,and the addition of speaker nodes helps to improve the problems personal reference error.To further improve the performance of the model and improve the quality of summarization generation,this thesis constructs a knowledge inference module,which utilizes knowledge inference technology to mine deeper knowledge information.The experimental results showed that after adding knowledge reasoning technology,the Rouge-1 score of the model increased by0.08%,Rouge-2 score increased by 0.05%,and Rouge-L score increased by 0.07%.By analyzing the quality of the generated summarization,it can be proven that the summarization generated by the model with the addition of knowledge reasoning technology have higher quality and fewer factual errors.And it verifies the effectiveness of knowledge reasoning technology in improving the quality of summarization.