Event Extraction Method Based On Syntactic Graph Representation Fusion

The task of event extraction plays an important role in various fields.By extracting objective facts from text data,it serves financial risk monitoring,vertical domain knowledge base construction,and judicial assistance in case judgment.The event extraction task mainly includes several sub-tasks of event event classification,trigger word recognition,and argument detection.In current event extraction methods,the problems of argument overlapping and error propagation among multiple tasks often affect the accuracy of event extraction methods.Among them,the cascading method can show superior performance,but this method regards text sentences as sequences and ignores the syntactic structure information contained in the sentences.This paper will model complex dependencies through graph networks by introducing syntactic dependency information in texts,and will focus on the syntactic structure differences between different event types to better assist multiple subtasks involved in event extraction.At the same time,this paper designs a syntactic fusion module to fuse text features and graph information composed of syntactic dependencies,and implements it through a cascaded extraction method,which can simultaneously utilize the semantic features of text and syntactic relationships.Most of the event extraction methods currently used in the industry are difficult to transfer to new data sets,and it is difficult to identify emerging event types.In order to achieve model transfer learning in few-sample scenarios and adaptability to new data,a generative-based model improvement is proposed,which can get rid of the limitation of data type on model parameters,and make full use of the context information in the text and the label information of the data to capture the correlation of different task stages.In summary,the core contributions of this paper are as follows:1.In this paper,a cascading event extraction framework based on dual-channel feature fusion is designed by combining syntactic dependencies and semantic information of texts.The entities involved in the syntactic dependencies in the text can often have overlapping relations with the arguments defined in the event pattern.Fully mining this information can model the dependencies between different entities.Therefore,this study models syntactic dependencies by introducing graph representations.In the face of two different types of features,the vector features based on text sequences and the graph representation information based on heterogeneous graph networks,this study introduces a feature fusion device based on two channels to fuse the information in different spaces.2.Contrastive learning-based methods that employ event-type discriminative tasks to enhance the structural information represented by graphs.The syntactic dependency of events often has a similar relationship with the event pattern defined in the event extraction task,and the syntactic graph structure constructed from the text of the same event type will also have a certain similarity.Therefore,this study focuses on the similarity relationship between event patterns and structures,and distinguishes the structural features of different events by designing an unsupervised training task.3.Transform the model based on the generative paradigm to enhance the migration ability of the existing model to new data.Most of the traditional event extraction models are implemented according to a labeling method.The model extracts elements under the event category in the dataset through preset classifier parameters,which is difficult to achieve when faced with new data types and fewsample transfer learning.Therefore,this study transforms the model based on the generative paradigm,reconstructing the input and task form of the model to realize transfer learning in few-sample scenarios.The event extraction method based on syntactic graph representation fusion proposed in this paper has achieved good results on multiple data sets,surpassing all current event extraction methods,and this method can compare with the baseline in the case of few samples of data.The model achieves good results.Combined with the specific experimental design,this paper analyzes the roles and contributions of different functional modules,and lays a foundation for the application of event extraction in different fields.