Research On Automatic Summarization Algorithm For Meeting Speech Transcribed Text

Meeting is an efficient and common way of multi-person discussion while daily office work.This would have been a labor-intensive task,but with the development of natural language processing technology,meeting summaries are now possible to be generated automatically.Automated meeting summarization has high practical value,and research institutions,including enterprises,have a strong interest in it.Different from the of traditional genres such as news,the conference corpus is more colloquial and more loosely related in context.It also contains information such as multiple speakers and implicit sub-topics of the conference.These differences make the meeting summary more difficult,the general summary method can not get a good effect,it is necessary to build a targeted model.Aiming at the problems above,the main work of this paper is as follows:1.We proposed a meeting summarization model based on heterograph of meeting elements.In order to deal with the problem of serious colloquialism and loose context of the meeting and make better use of the unique elements of the meeting such as speakers and subtopics,we designed a heterograph of meeting elements,containing speaker node,subtopic node,word node and sentence node,at the same time,we leverage relational graph convolutional network to encode meeting context.In the end,we extract summarizations of the meeting.2.A meeting summary model based on multi-granularity supervising information is proposed on the basis of heterograph of meeting elements.Through global summarization and speakers’summarization supervising tasks,models can be learned from different focuses,and problems of over-smoothing and information bottleneck of graph neural network can be alleviated,so as to improve the representation learning ability of graphneural network and improve the performance of the model.3.We conducted model experiment and evaluation on AMI dataset,obtained ROUGE-1、ROUGE-2、ROUGE-L score of 0.8、0.69 and 0.78,and verified that the model could reach a competitive level.In addition,through ablation experiment,we can verify the rationality of the construction method of hetero-graphs in this work and the effectiveness of the multi-granularity supervising information strategy.