Research On Dialogue Summarization Technology Based On Pre-Trained Language Models

In recent years,dialogue summarization has become one of the research hotspots in the field of natural language processing.Its goal is to extract the most important information from semi-structured,multi-party dialogues and generate concise summaries so that users can quickly understand the topics and contents of the dialogue.However,the complex linguistic properties of dialogue pose challenges for summarization tasks,such as the scarcity of data resources and important information always being scattered among multiple participants.To address these challenges,this thesis proposes a dialogue tasummarization model that combines prompt learning with dynamic templates,aiming to explore low-resource dialogue summarization tasks.Furthermore,this thesis constructs a customer service dialogue summarization dataset to alleviate the problem of resource scarcity in dialogue summarization specific fields and proposes a dialogue summarization model that introduces dialogue state tracking for this dataset.Finally,this thesis designs and implements a dialogue summarization system.The main work of this thesis includes the following:Firstly,this thesis proposes a dialogue summarization model called Dynamic PET,which combines prompt learning with dynamic templates.Based on the idea of prompt learning,the model constructs dynamic templates using key information in the dialogue dataset,fully mining the potential knowledge in pre-trained language models to guide the model in generating high-quality dialogue summaries.Experimental results show that Dynamic PET significantly outperforms baseline models on the QMSum and Media Sum datasets under low-resource conditions.Secondly,this thesis constructs a customer service dialogue summarization dataset called CSDSumm and proposes a dialogue summarization model called DST-DS that combines dialogue state tracking technology.Specifically,the CSDSumm dataset is constructed through a person-in-the-loop method,and the characteristics of the dataset are analyzed in detail.In addition,for the CSDSumm dataset,this thesis proposes the DST-DS model,which can obtain the state information of the dialogue text and integrate it into the model to improve the model’s ability to generate dialogue summaries.Experimental results show that the DST-DS model significantly improves performance on the CSDSumm and TODSum datasets under low-resource conditions,demonstrating the effectiveness of this approach.Finally,this thesis designs and implements a dialogue summarization system based on the two proposed dialogue summarization models.During the system implementation process,functional requirements analysis and non-functional requirements analysis were conducted to determine the overall system architecture and technical architecture design.After continuous development and testing,a fully functional dialogue summarization system has been successfully implemented,which can generate accurate and concise dialogue summaries based on the model selected by the user.Test results show that the system has good usability,reliability,and scalability,and can effectively handle dialogue summarization requests,providing users with high-quality dialogue summarization services.