Research And Application Of Meeting Minutes Generation Method Based On Structure Modeling

Meeting is an important way for people to obtain information in their daily life and work.It is time-consuming and labor-intensive to write the meeting minutes manually because of the increasing demand for meetings.It is necessary to use the automatic generation technology of meeting minutes to quickly extract important information from massive meeting data.The meeting structure information is more universality.If we can know the structural distribution of the various topics of the meeting,it can help to understand the key information of the meeting content.The structural information is critical to the formation of meeting minutes.Therefore,this thesis studies the generation method of meeting minutes based on structure modeling,and has the following research contents:(1)Paragraph segmentation method based on meeting content is proposed.We improve the Text Tiling algorithm and propose an unsupervised paragraph segmentation method using sentence-level information for meeting content structure.Further,we introduce the Bi LSTM to realize a supervised meeting paragraph segmentation model based on sentence embedding.And we build the meeting paragraph dataset by using the text records of the United Nations General Assembly.(2)Extractive meeting minutes generation method based on structural information is proposed.We take the extractive text summarization task as a text matching task,and introduce the paragraphs information of meeting data to improve the model’s ability to understand the overall semantics of the text and the learning ability of structural information features.Further,we build the meeting summary dataset that can be used for extractive text summarization task.Finally we perfrom a series of experiments for our method.The experimental results prove that the introduction of meeting structure information is beneficial to the improvement of the quality of meeting minutes results,it is necessary to study the text summarization task combined with the text structure modeling task.