Emergent Event Detection Based On Graph Neural Networks

At present,various emergent public opinion events are increasingly frequent in social networks.The mass dissemination of event-related texts is easy to cause the acceleration of public opinion.Timely detection of the emergent events is helpful for public opinion analysis to avoid the deterioration of public opinion environment in a short time.Existing event detection methods require sufficient labeled samples of events,which makes it difficult to effectively detect the emergent events that lack labeled information.Therefore,intelligent emergent event detection has become an urgent problem to be solved.In order to solve the above problems,the research makes full use of the topology information in public opinion networks and adopts few-shot learning to propose a emergent event detection method.Considering the poor robustness of random labeled texts in few-shot scenarios,the research uses prompt learning algorithm for text attribute learning.The method converts unstructured text information into interpretable structured attribute representation,so as to introduce few-shot attribute representation based on expert knowledge to replace random labeled texts and improve the robustness of the model in the task of emergent event detection.The research proposes a self-supervised text attribute learning method based on prompt learning for text attribute learning tasks.The method uses prompt learning algorithm to obtain external prior knowledge from the pre-trained language models.We construct self-supervised tasks based on topological structure information in graph data and train the model without labeled texts.By learning the attribute representation of the text,the method extracts the discriminant information concerned in the current event detection task for each text.The experimental results on three public datasets show that this method can effectively learn the text attributes.In the downstream text classification task,the accuracy of this method is improved by 0.6%～3.3% compared with baseline methods.Aiming at the few-shot learning tasks in graph data,The research proposes a graph few-shot classification algorithm based on attribute matching.The method uses a meta-learning paradigm to classify new categories with minimal expert knowledge by using prior information from historical categories.We capture and encode attribute distribution differences among different meta-tasks in meta-learning,and construct attribute matching vectors for each meta-task,so as to enhance the information transfer efficiency of meta-learning model among different meta-tasks.We improve the model’s ability of few-shot node classification,which provides a model basis for the emergent event detection tasks.The experimental results on five public datasets show that the proposed algorithm can effectively learn high-quality node representations,and achieves an improved accuracy of 1.8%～6.4%compared with the baseline methods in few-shot node classification tasks.In order to further verify the effectiveness of the emergent event detection method based on graph neural networks in real scenarios,the research collected and constructed an event detection dataset for 7 public opinion sub-events related to the postgraduate entrance examination and the COVID-19 epidemic on Sina Weibo.We verify the proposed method by real dataset experiments in the task of emergent event detection.The experimental results show that this method can effectively learn the interpretable attribute representation of event text and detect the emergent events efficiently and accurately under the few-shot scenario.