Research And Application Of Meeting Text Generation Method Based On Style Transfer

Meeting is an important medium for people to exchange information in daily life.How to get an accurate meeting record report and how to extract the basic information and central content of the meeting from the vast amount of meeting content efficiently and quickly,is significant for people to improve office efficiency and reduce work burden.In recent years,with the advancement of technology,people begin to use online channels for meetings to facilitate communication.However,meeting records stored in audio format are not easy to search and locate.Transcriptions obtained after automatic speech recognition may contain some colloquial information,which is not conducive to readers’ understanding.Therefore,this thesis focuses on the research and application of the meeting text generation method based on style transfer,which aims to convert spoken transcriptions into written reports in Chinese and further generate the meeting minutes with written reports.The specific research of this thesis is as follows:(1)For the task of converting spoken transcriptions into written reports in Chinese,this thesis improves the existing BART model based on hierarchical decomposition encoding.The model adopts an end-to-end approach and extrapolates the maximum length(512 tokens)that the model can originally handle,allowing the model to handle texts of longer length and reduce data loss.We construct a Cantonese meeting-style transfer dataset using publicly available meeting audio files and records of the Legislative Council of Hong Kong,and the improved model generates better text quality on the dataset than unimproved model.(2)For the task of meeting minutes text generation,this thesis proposes an extractive-generative method for meeting minutes generation based on structural information.Considering that the paragraph structural information of a text has guiding significance for text understanding,the thesis combines extractive and generative methods.It introduces structural information into the model training process.The thesis constructs a meeting minutes dataset with structural information using the meeting records of the United Nations General Assembly.The final results on several evaluation metrics show that the introduction of structural information of meeting texts can effectively improve the quality of meeting minutes generation.