Multidimensional Event Analysis Based On Generative Summaries

In recent years,news breaking events,such as the New Coronavirus pandemic and the Russia-Ukraine conflict,have received attention from the academic community because of their significant impact on human society and the environment.In the era of mobile Internet,the contradiction between fragmented reading needs and complicated online information has become increasingly prominent,making the distillation of key information crucial.In order to deeply understand breaking events,it is more appropriate to use multidimensional event analysis methods to analyze news events by comparing multiple event analysis methods.Multidimensional information,including time,place,and affected area,helps readers to grasp the development of the event in a comprehensive manner and thus avoid falling into a single perspective.In addition,multidimensional information distillation helps identify key nodes and influencing factors,providing targeted suggestions for decision makers and thus improving the effectiveness of response measures.Through an in-depth discussion of automatic summarization technology,this thesis systematically analyzes the background,importance,and current status of domestic and international research on automatic text summarization and multidimensional event analysis technology,and designs an improved summarization algorithm to address the above issues.In addition,based on the above algorithm process,a new set of information extraction algorithm process,i.e.,multidimensional event analysis algorithm,is proposed,and a multidimensional event analysis system facing news breaking events is constructed based on this.The main research contents of this thesis are shown as follows:1.In this thesis,we discuss the challenges of existing summarization methods in depth,and design a GC-GPT model algorithm for news text summarization tasks based on the GPT-2 model by incorporating GSG training objectives,an improved copy mechanism,and the Spacemax function to address these problems and research objectives.In addition,the algorithm performance is evaluated using ROUGE-1,ROUGE-2,and ROUGE-L metrics,which have about 0.3,0.1,and 0.3 improvement over the original GPT-2 model,respectively,confirming the practical effect of the optimization method.2.In order to improve the accuracy of dimension extraction,this thesis improves the SIFRank method and proposes the BF-SIFRank dimension generation algorithm,whose key point is to use BERT-Flow instead of SIFRank’s representation network ELMo,thus enhancing text understanding and key phrase generation.After that,a multidimensional event analysis sequence is successfully constructed using the summary generation model and the multidimensional event analysis algorithm proposed in this thesis.The designed algorithm flow consists of three modules: text pre-processing,summary and dimension generation,and multidimensional event analysis sequence generation,and it is verified that the accuracy and recall of dimension extraction are improved.3.In this study,a multidimensional emergent event analysis system is designed.The system provides users with a simple interface for using multidimensional analysis of breaking events and also gives a wide application potential for the algorithmic process proposed in this thesis.