Research On Key Technologies Of Event Extraction And Event Reasoning Based On Graph Neural Network

With the rapid development of the Internet,lots of valuable information has emerged in the Internet field.Therefore,how to extract valuable information from massive data and implement the application becomes increasing crucial.As knowledge mapping becomes more and more sophisticated,event graph with event information as the core has been paid attention to by researchers.Compared with knowledge graph,event graph can better reflect the time sequence information of events and better fit people’s understanding of events.In the research and application field of event graph,event extraction as the core of event graph automatic construction process has been paid much attention by researchers.Event reasoning,as the main application of event graph,can be used to reason and answer tasks with the help of valuable event attribute and timing information in event graph.At present,in the field of event extraction,a single sentence-level event extraction algorithm has achieved high accuracy.However,there are a large number of nested events in the text,many events need to be extracted across sentences,and the sentence-level event extraction algorithm can no longer meet the requirements,so the text level event extraction algorithm emerged.However,the same word in text often contains different semantic information in different sentences.The traditional graph neural network cannot effectively represent the one-to-many association relationship of a single sentence corresponding to multiple words.Therefore,the current text level event extraction algorithm cannot fully learn the information transfer between words.In addition,with the increase of the number of layers,the graph neural network will cause some problems such as over-smoothing of node features.Therefore,in view of the problems existing in the existing document-level extraction models,a novel algorithm based on hypergraph convolutional network is proposed in this thesis.For the first time,we introduce and improve the hypergraph convolutional network in the context-level event extraction task,so that it can effectively learn the one-to-many correlation features between sentences and words in the text and information transmission,and at the same time reduce the oversmoothing problem by controlling the number of layers in the network.In addition,we introduced and improved the Bi Affine Graph Parser to learn the syntax information more efficiently.Finally,we conducted experiments on the biomedical data sets MLEE and GE,and the experimental results show that the algorithm is superior to other benchmark models in overall extraction effect.In the field of event reasoning,graph neural network is used to extract the time series features of event chain mostly by its powerful learning ability of graph structure data.However,the traditional graph neural network is unable to effectively learn the information transmission among parameters in the whole event chain,resulting in the failure to learn the comprehensive event context features.In addition,in the existing event inference model,the attention mechanisms only focus on the correlation between historical events and candidate events,and ignores the transmission of information between candidate events and historical events.Therefore,in view of the problems existing in the existing script event reasoning models,an algorithm of multi-granularity features enhanced representation learning for script event reasoning was proposed in this thesis.To be able to learn the multi-grained characteristics of events simultaneously,we introduce LSTM to learn the feature of event parameter sequence,and introduce PNA to extract the node feature of event graph to learn the chain-level features of events and introduce the ALN to extract the contextual semantic features of events.To take advantage of the above three features,we have designed a feature fusion approach with Transformer as a framework.Meanwhile,we adopted the threshold calculation method to select excellent features for feature fusion,so as to reduce noise features.In addition,we first introduce the Cross-Attention mechanism in the event inference task to enhance to learn the information transmission from candidate events to historical events.Finally,we experiments on datasets NYT,and experimental results show that our model is superior to other baseline models in scripting event reasoning tasks.In this thesis,algorithm innovation is carried out for documents-level event extraction task and script event reasoning task respectively,and a corresponding algorithm model is proposed.Experiments show that the above two algorithms have certain theoretical and application value.