Research On Event Graph Construction For Social Events Analysis

Social events are public events that broke out at a certain moment,aroused widespread public opinion,and may have a major impact on public safety and social stability.Research on social events can not only provide more high-quality services at the user level,but also play a vital role in the monitoring of macro public opinion at the national level.With the development of news media and social media,a large amount of data will be generated on these platforms when a social event occurs.The structured processing of these event data is the basis of the analysis task of social event.Building an event graph is one of the methods to structure these event data.The existing event graph mainly focuses on atomic events composed of a set of related actions and entity concepts while social events are more complex than atomic events and generally consist of multiple atomic events.In this article,such events are called public opinion events.The above facts lead to the problem of the granularity mismatch in the construction of event graph for social events.This work combines the characteristics of social events and related technologies of existing event graph to design a two-layer graph schema that includes atomic event layer and public opinion event layer.This work also provides a three-step framework and corresponding solutions for constructing an event graph based on the designed two-layer graph schema:(1)Construction of Atomic Event Graph.This stage mainly completes the task of structuring a document describing a public opinion event into an atomic event graph.This work first uses the pre-trained language model and the conditional random field model to complete the extraction of atomic events,including the identification of trigger words and the identification of arguments and their corresponding roles.This work also uses the pretrained language model to build a multi-classification model to complete the recognition of atomic event relationships,and finally get an atomic event graph.(2)Atomic Event Graph Fusion.This stage mainly completes the task of fusing atomic event graphs generated from different documents describing the same social hot public opinion event.In this work,the graph neural network is used to align the nodes in the atomic event graph,and then follow-up problems such as deduplication,merging and conflict resolution are solved through rules to get the final atomic event graph fusion result.(3)Public Opinion Event Relevance Analysis.This stage mainly completes the task of mining the relevance analysis between public opinion events on the atomic event graph.This work uses external knowledge and graph convolutional network to design model,uses relative relevance to complete the training process,and obtains the result of relevance weight between public opinion events.