Research On Event Prediction Based On Graph Neural Network

Event prediction refers to the prediction of the trend level of events or the occurrence of an event that affects the stability of a country or a region.It can provide auxiliary support for decision makers,and can be widely applied in the domain of strategic situation analysis and public opinion analysis.To solve the problem of inadequate feature extraction and lack of interpretability when predicting major event trend based on the public news data,starting from the perspective of event cognition,the idea of using knowledge graph to organize conceptual knowledge is adopted in this paper,and the core elements of events are firstly organized in the form of graph after extracting events from text data.The trend level prediction for major events and the subsequent event prediction for emergencies are then studied by using the technique of natural language processing and graph neural network comprehensively.The main research contents are listed as follows:(1)To solve the problem of lack of semantic understanding in traditional prediction methods based on frequency statistics,a trend prediction method of major events based on Graph Convolutional Network(GCN)is proposed.Firstly,the topic detection and tracking technology based on the aging theory is used to form the topics,and then the related topics are manually screened according to the actual problems and merged into thematic data.Secondly,the event extraction technology based on pattern matching is used to extract structured event data from news text.Thirdly,the semantic association graph of events is constructed according to the overlap of elements between events.Finally,GCN is used to aggregate node neighborhood information to train the prediction model and output the prediction results.The accuracy of the test set reaches 76.92% when predicting nuclear behavior trend level of North Korea,which is better than the method based on word frequency characteristics,and demonstrates the effectiveness and feasibility of the proposed method.(2)As the prediction of emergency focuses on prediction of its development trend and ignores the prediction of subsequent related events,a subsequent event prediction method combining Gated Graph Neural Network(GGNN)and attention mechanism is proposed.GGNN is applied to perform event feature extraction based on the characteristic that events between each other are orderly correlated and this number is uncertain.Considering the difference of contribution of context events to candidate events,the attention mechanism is introduced to match the candidate event.The emergency prediction dataset,including earthquake,fire,traffic accident,terrorist attack and food poisoning,is built in this paper based on an open emergency corpus.which is used to predict the subsequent related events of the emergency.Long Short Term Memory(LSTM),GCN and GGNN are used to extract event feature,respectively,and their prediction accuracies of the built emergency prediction dataset are 62.77%,64.08% and 68.76%,respectively.Finally,the work of this paper is summarized and prospected.