Research And Implementation Of Event Extraction Method For Low Resource Scenarios

An event refers to a significant change in the state of something that occurs,develops,and ends within a certain time and space range,which can be a natural phenomenon,social activity,personal behavior,etc.An event consists of event elements,such as the participants,the time and location of the event.Accurately identifying event related information from text has wide applications in various fields such as intelligence analysis,public opinion monitoring,and security monitoring.The current development of event extraction is limited by low resource factors such as insufficient annotated data.Therefore,this article explores event extraction methods at sentence-level and discourse level in low resource scenarios.For sentence-level event extraction,although current methods based on remote supervision have to some extent alleviated the low resource problem,such methods can generate data noise and affect subsequent model training.In the joint extraction model of trigger words and arguments,most current methods only consider the pipeline method,which first identifies the trigger words and then identifies the arguments.However,this method can lead to irreversible parameter error replay problems.To address the above issues,this thesis proposes an end-to-end event extraction algorithm based on the BERT pre training model.On the one hand,the proposed algorithm trains trigger word vectors with strong semantic representation capabilities based on the BERT model;On the other hand,the proposed algorithm is based on entity recognition,treating event argument extraction as a multi classification problem,improving the accuracy of event argument extraction.Experiments on public datasets have shown that the proposed end-to-end model improves the accuracy of sentence-level event extraction in low resource scenarios.For document-level event extraction,the existing document-level event extraction method based on active learning alleviates the problem of low resources,but this method requires a large number of people to participate in,and cannot fully extract event information automatically.To address the above issues,this thesis proposes a document-level event extraction algorithm based on text enhancement.The algorithm is based on the generative adversarial networks for text enhancement,and focuses on the semantic information between sentences based on the attention mechanism.Experiments on publicly available benchmark datasets show that the proposed method improves the accuracy of document-level event extraction.For the event extraction system,the event extraction system designed and implemented in this thesis includes two granularity functions of event extraction,data transmission,and content management.After testing,the event extraction system implemented in this thesis has achieved the expected goals,and the system functions meet the basic requirements of the design.