Research On The Generation Method Of Dialogue Summary Based On Deep Learning

Dialogue summarization aims to condense a conversation into a shorter passage containing important information,which can help people quickly capture the highlights of a semi-structured and multi-participant dialogue without reviewing historical conversations.In recent years,with the popularity of mobile phones,e-mails and social network applications,people are increasingly sharing information in the form of dialogues.Especially due to the spread of COVID-19 worldwide,online multi-person chat or meeting has become an important part of people’s daily work.Therefore,how to use text summarization technology to quickly and accurately organize a large amount of dialogue data into short,natural and informative text has become increasingly important.Since the information flow in a conversation is exchanged between at least two interlocutors,the key information is often scattered in different utterances of different speakers,which leads to redundant or incorrect content in the dialogue summarization generated by the traditional text summarization model.Aiming at the problems that the traditional text summarization model doesn’t fully understand the context of conversation and is difficult to associate the speaker with its corresponding action when generating dialogue summarization,this thesis proposes a dialogue summarization method based on T-HDGN model.The method firstly uses the extracted action triples to explicitly model the conversation structure,considers utterance and action triple as two different types of data to construct heterogeneous dialogue graphs and models for both information through a Heterogeneous Dialogue Graph Network;Meanwhile,the model also adds speakers as heterogeneous nodes to facilitate information flow.Moreover,in the decoding stage,the topic word features are introduced to assist the summary generation.The experimental results on the SAMSum data set show that compared with the baseline model,the proposed model achieved higher ROUGE scores;The comparison results of example show that the model can correctly associate speakers with their corresponding actions and generate higher quality dialogue summarization.The traditional text summarization model only models the entire conversation from a single perspective,making it difficult for the model to capture comprehensive and nuanced conversation information,and any lack of granular information in the conversation encoder will cause a larger error cascade in the decoding process,thus affecting model generation effect.Aiming at the above problems,this thesis proposes a Transformer-based multi-view dialogue summarization method.This method makes full use of structured views extracted from conversations,encodes conversations from topic perspective and conversation stage perspective,and then uses a multi-view decoder to fuse encoded representations from different views;In addition,the model introduces pointer generation network and coverage mechanism to alleviate the OOV problem and repeated fragment problem in conversation summary.The experimental results show that this model segments conversations into blocks to model conversations from multiple perspectives,obtains informative context representation,and gets a higher quality summary.