Abstractive Dialogue Summarization With Enhanced Factuality Consistency:Research And Implemntation

The dialogue summarization aims to summarize the content of the dialogue and generate a brief text that covers all significant information in the dialogue.In recent years,dialogue summarization has attracted a lot of attention from researchers due to its applicability in various real-life scenarios such as meetings and healthcare.Many successful text summarization approaches have failed in handling dialogue summarization due to the unstructured context and multi-party first-person perspective features of the dialogues.In dialogue summarization tasks,the input dialogues are usually in colloquial style with omissions and co-reference information,while the output summaries are more formal and complete.Therefore,dialogue summarization models should be able to complete the omissions and co-reference information and generate appropriate summaries accordingly.However,current state-ofthe-art models focus more on the themes or structures of the summaries rather than the consistency between the summary and its input dialogue context,which may be affected by logical inconsistencies.This thesis proposes the following approaches based on analyzing related work:An utterance rewriting model is proposed to complete the omissions in the dialogue content and obtain complete utterances.Then,a coreference information data augmentation mechanism is used to replace referenced names with original names to enhance co-reference information.Finally,the BART model is used to generate summaries using the data augmented by sentence rewriting and co-reference information.A contrastive learning model based on AMR is proposed,which uses AMR tools to analyze the dialogues and performs a series of operations on entity-level,sentence-level,and dialogue-level text that are prone to logical inconsistencies,and then constructs them into negative samples for contrastive learning.The model is then trained using contrastive learning to improve the fact consistency of the summaries.The experimental results on the SAMSum and DialogSum datasets show that the model proposed significantly outperforms the baseline model in terms of quantitative evaluation and human evaluation.Finally,this thesis implements a dialogue summarization demo system to showcase the research results,which includes designing and implementing a front-end interactive interface,a back-end control module,and deployment.