| Event extraction aims to present unstructured text containing event information in a structured form,which is an essential task in information extraction.And it has a wide range of applications in the automatic digest,automatic question and answer,information retrieval,and knowledge base construction.Event extraction consists of four subtasks trigger identification,event type classification,argument identification,and argument role classification.Currently,most event extraction methods rely heavily on manually labeled event triggers.However,the annotation process will be resource intensive.The incorrect annotation and ambiguous triggers will also impact subsequent event detection tasks.At the same time,pipelined event extraction methods often have disadvantages,such as error transmission and independence in each task.In view of the shortcomings of the existing methods,this paper proposes an event extraction method without triggers.The main research contents of this thesis are as follows:(1)This thesis proposes a type hierarchy enhanced event detection framework without triggers to avoid using labeled triggers.Firstly,the method constructs a type hierarchy concept module using an external knowledge graph Probase to enhance the semantic representation of event types.Secondly,the method divides the instances into lexical-level and context-level representations,which allows the model to use different levels of features.Experimental results show that our proposed approach achieves a good improvement and is significantly competitive with the mainstream trigger-based models.(2)This thesis proposes a joint event extraction method without triggers based on deep learning to avoid problems from pipelined extraction methods,such as error transmission.Firstly,the method construct global features based on Frame Net,the semantic network of frames,to capture dependency features across instances and subtasks.Then the event detection and argument extraction tasks are performed simultaneously;finally,use global features to make joint decisions on classification results to obtain event extraction results.Extensive experiments have validated the effectiveness of the proposed model.In summary,this thesis uses different external knowledge bases and employ suitable neural network models to alleviate the problem,such as over-reliance on annotated triggers and neglecting subtask relationships for event extraction.Finally,experimental results show that the proposed method performs well. |
