Research On Event Extraction Method Based On Multi-Feature Fusion

Event extraction helps people quickly obtain structured event information from long articles,which is of great significance.The existing event extraction models based on graph neural networks usually use dependency syntax analysis to obtain structural dependency relationships between words in sentences.However,for some relatively lengthy sentences,the dependency syntax analysis results contain many directed edges with syntactic relationship labels,and using graph neural networks alone cannot fully learn the event knowledge contained in them.In addition,syntactic analysis results may bring unnecessary redundant information to the model,increase model training parameters,and reduce event extraction performance.In order to solve the above problems,this article improves the existing methods and proposes two new event extraction methods based on graph neural networks.The main research content is as follows:(1)A method for event extraction based on graph attention network and interactive attention mechanism was proposed.First,the bidirectional short-term memory network is used to fuse multiple feature information extracted from sentences to help the model understand the complex sentence structure contained in lengthy sentences.Secondly,the dependency parsing results are transformed into undirected graphs,and the shortest path algorithm is used to supplement the associated nodes,solving the problem of poor performance of graph attention networks in learning relationships between multi hop nodes.Finally,the interactive attention mechanism is used to fully integrate the contextual semantic features of the sentence and the structural dependency features output by the graph attention network,thereby improving the performance of event extraction.(2)A method for event extraction based on self built dependency relationships and graph convolutional networks is proposed.Firstly,the dependency parsing tree is pruned through self built rules to reduce the negative impact of redundant information on the model,and transformed into a self built dependency relationship graph after fusing named entity features.Secondly,an improved graph convolutional network is used to aggregate nodes in the graph and integrate the multi head attention mechanism,enabling it to further distinguish the importance of dependency relationships between nodes.Finally,dynamic fusion of contextual semantic features and dependency features is achieved through a gating mechanism,and event extraction is completed.This article uses the ACE2005 dataset as experimental data,compares existing event extraction baseline models,and further explores the effectiveness of the improved method through ablation experiments.The experimental results show that both methods have significantly improved F1 values,proving that the performance of the event extraction model proposed in this paper has achieved the expected effect and is reasonable to a certain extent.