Research On Event Detection Method Of Graph Neural Networks Based On Dependency Awareness

In recent years,with the explosive increase in the amount of information on the Internet,researches on the extraction information have received widespread attention.Event detection has become an important subtask in information extraction,and it has also become the focus of NLP research,and many achievements have emerged.As an effective method in NLP,dependency parsing can parse out the dependencies between words in a text,and Graph Neural Network(GNN,Graph Neural Network)is a mature deep learning technologies.Many scholars combine dependency parsing with GNN to improve the effect of event detection.However,most of the existing methods for modeling dependencies using GNN only consider the structural information of dependencies and ignore the label information of dependencies.To address of problems existing in trigger recognition and event classification,the two subtasks of event detection,this paper designs two kinds of models.The main work of this paper is as follows:(1)Trigger recognition that integrates multi-hop relation labels and dependency syntactic structure information.Due to the existing methods that are difficult to fully capture the correlation between words in terms of dependencies,this paper have designed a new multi-hop search algorithm based on the semantic and syntactic features of the dependency syntax tree.Aims to capture multi-hop long-distance associations between trigger words and their associated words.In addition,the multi-level structure information of the dependency syntax tree is modeled by the MOGANED model,and the multi-hop relationship label information of the dependency relationship and the multi-level structure information of the dependency syntax tree are also considered,so as to fully improve the syntactic feature information of trigger words and improve recognition effectively.(2)Hierarchical Modular Event Classification Based on Dependency Awareness.In the past event classification,fine-grained event types are regarded as independent and scattered tasks,ignoring that each fine-grained event type has its own upper-level event type information.Therefore,this paper designs an event hierarchical modular network,and it uses the global attention mechanism to obtain the upper-layer coarse-grained event type information of the trigger word.Then,by using the graph convolutional neural network to dynamically perceive the dependencies between words,the fine-grained event type information of the event trigger word is obtained.Finally,the fine-grained event type is concatenated with the coarse-grained upper-level event type information through weighted average,which aims to reduce the function solution space in event classification,improve the efficiency of event classification,and further improve the performance of event detection.Experiments show that in trigger recognition,the new multi-hop algorithm proposed in this paper improves effect by nearly 2% on the F1 value,which proves the effectiveness of the algorithm.In the event classification task,the hierarchical modular network proposed in this paper achieves 79.3% on the F1 value,proving that the method is also effective.