The Ressearch On Abstrctive Dialogue Summarization Based On Dialogue Structures

As an important technology to extract key information and reorganize and summarize in the field of natural language processing,text summarization has been paid much attention.Among them,dialogue summary is an emerging summary task,and dialogue summary also gives birth to many unique challenges due to the rich characteristics of dialogue,such as many people and many rounds,which leads to low logic and poor readability of dialogue summary generated by traditional methods;Although the early exploration of dialogue summary proposed to build dialogue structures for dialogue modeling and generate summaries,the construction process of these structures is strong manual,time-consuming and laborious;At the same time,due to the incomplete understanding of the dialogue summary task in the early stage,these works can not well design a dialogue structure that can comprehensively model the dialogue.Therefore,in view of these problems in the existing dialogue summary work,the main work of this paper includes the following points:1.A hierarchical structured representation model is proposed to represent various structural relationships naturally existing in dialogue text.The model can represent the dialogue structure of dialogue relationship,the dialogue structure of dialogue participant relationship,the dialogue structure of dialogue topic relationship and so on.And these dialogue structures can be unsupervised through the basic identifiers in the dialogue,which has high portability.2.An attention mechanism based on dialogue structure perception is proposed to integrate multiple dialogue structure information.In this mechanism,each discourse sentence is taken as the minimum coding semantic unit,and the above hierarchical dialogue structured representation model is transformed into the mask matrix in the self attention mechanism,which is integrated into the calculation process of attention,so as to perceive the structural information in the process of summary generation.The experimental test on SAMSUM dataset shows that the generated summary can give consideration to readability and logic,and has high quality3.Based on the above generative dialogue summary model based on the corresponding dialogue structure,an automatic summary system is built for the work order summary requirements of the "auto master challenge",and the generated work order summary results have reached the historical optimal results.The above work and further analysis show that the dialogue structure proposed in this study is simple,feasible and efficient,and the relevant models can generate dialogue summaries with both readability and logic.