Research On Abstractive Dialogue Summarization Technology To Enhance Dialogical Semantic Information

The widespread application of social networking platforms has produced a variety of types of dialogue data,including meeting,customer service,doctor-patient communication,email and daily chat.Dialogue summary generation technology aims to generate concise summary descriptions based on dialogue data,thereby improving the efficiency of users’ access to information and solving the information overload problem.With the widespread application of sequence-to-sequence model,abstractive text summarization technology has gradually become the mainstream technology of summary generation,and the research of abstractive dialogue summarization has become more and more extensive.For multi-participant dialogue summarization,the existing summary generation methods does not take into account the dialogue semantic information sufficient in the dialogue data,such as discourse structure information,fact information and coreference information,so it is easy to cause the problem of reference confusion and fact inconsistency in the generated summary.Therefore,this paper conducts an in-depth study on the abstractive dialogue summarization task,fusion of multiple dialogue semantic information,constructed a dialogue summarization generation model that integrates coreference information and discourse relations,and a dialogue summarization generation model that integrates fact information and discourse relations trained by deep reinforcement learning,and experiments are carried out to verify the role of dialogue semantic features in summary generation through the extraction of dialogue discourse relationship features,coreference information extraction and resolution,so as to improve the problems of reference confusion and fact inconsistency in the generated abstract.The main works are as follows:(1)An abstractive dialogue summarization model integrating coreference information and discourse relations is constructed.The coreference resolution model extracts the coreference information from the original dialogue data and generates the coreference weight matrix,uses the coreference attention layer to fuse the coreference information with the hidden vector,which effectively utilizes the coreference information in the dialogue.At the same time,the dialogue discourse parser is used to extract the dialog structure information from the source text and construct the dialog discourse relation graph.The sentence vector representation obtained by the bidirectional gated recurrent neural network is transformed into the discourse node in the dialogue discourse relation graph.The relations between the dialogue sentences are transformed into the relations node in the dialogue discourse relation graph,and the global node is added to obtain all the structural information of the dialogue,which effectively utilizes the interaction between the dialogue sentences.The experimental results on the dialogue datasets AMI and ICSI show that the proposed method can effectively improve the quality of the generated summary.(2)A structure graph-aware abstractive dialogue summary model based on deep reinforcement learning is constructed.In the summary generation model,this paper first uses Stanford University’s natural language analysis tool to extract the coreference information from the dialogue,replaces the first-person,second-person and third-person pronouns in the dialogue data with the speaker’s name or the words in the coreference cluster related to it,uses information extraction tool to extract the fact triple information from the modified data and generate a coreference resolution action graph.Then,uses the discourse parser to extract the discourse relations and generate the discourse relation graph.Finally,the graph attention network is used to calculate the two graph data respectively,the output of the graph is transmitted to the decoder for generating the summary,which effectively utilizes the structural information in the dialogue data.In terms of training,uses question answering-based summary fact consistency metric and improved question answering-based summary fact consistency metric as reward functions.First,the corresponding questions and answers are generated from the reference summary and the generated summary.Then use the reference summary and the generated summary to answer the questions.Calculate the similarity of the two sets of answers,and the similarity is averaged to obtain the reward of the model,then the Proximal Policy Optimization reinforcement learning algorithm is used to train the model.Experiments show that the summary generation model trained by question answering-based reward and the summary generation model trained by improved question answering-based reward perform better on the SAMSum dataset than other models that do not use reinforcement learning for training.