Research On Multi-Document Summarization Based On Multi-Granularity Fusion And Knowledge Enhancement

With the rise of the Internet,an overwhelming amount of repetitive information has emerged.This has made it difficult for users to quickly and accurately extract the key information they need when browsing through massive amounts of content on their smart devices.To address this issue and help users obtain key information from each article while integrating common information from multiple documents in a short amount of time,the development of multidocument summarization technology is crucial.Although many existing neural network-based models consider the contextual semantic information of the text for multi-document summarization,they often fail to capture deeper and more granular semantic information as well as ensure factual consistency in the generated summaries.Additionally,existing generative summarization models lack constraints and controls on the generated text,which can be problematic when users have specific needs for the generated summaries.To overcome these challenges,this paper focuses on developing a multi-document summarization technology that addresses the aforementioned issues.The main research objectives of this paper are:(1)This paper presents a neural network model that aims to address the issue of inconsistency in generated text facts in multi-document summarization.The proposed model leverages entity information enhancement and multi-granularity fusion techniques to enrich the semantic representation of source texts.In the pre-training stage,external knowledge graphs and multi-document texts are used to fuse entity information,enabling the model to capture deeper semantic information.Existing neural network-based models often fail to consider deeper semantic information,which motivated the proposed approach to my implicit expressions at three granularities: words,entities,and sentences.The results of experiments demonstrate that this approach effectively improves the performance of the multi-document summarization model by enhancing the fluency,information content,and fidelity of generated text.Moreover,incorporating external knowledge graphs into the model contributes to improving its factual consistency.Overall,the proposed model represents a significant step towards solving the issue of inconsistency in multi-document summarization and highlights the potential of external knowledge incorporation in deep learning-based models.(2)This paper proposes a conditional text generative model for generative multi-document summarization to address the lack of constraints and controls on generated text.The proposed model utilizes knowledge embedding and entity control to guide the summary output of the conditional text generator,enabling the controllable generation of generative text at the entity granularity.In the decoding stage,entities are added as control signals to guide word generation probability,which optimizes the previous method of solving the problem of factual consistency of the generated text.To address the issue of knowledge noise resulting from the introduction of external knowledge triples that can cause a shift in sentence meaning,this paper proposes a solution based on soft position coding and visible matrices in K-BERT.The approach controls the relationship between injected knowledge triples from different nodes,ensuring they remain invisible to each other,which effectively solves the knowledge noise problem.Experiments have verified the effective improvement and feasibility of the control signal-guided generation method proposed in this paper for the performance of the generative summarization model.