Research And Application Of Chinese Event Extraction Method Based On Span Regressio

Event extraction is an important task in natural language processing.It automatically extracts event information that users are interested in from large amounts of unstructured text and presents it to users in a structured format,providing technical support for automatic text summarization,event knowledge graph construction,intelligent question answering,etc.Event extraction includes two subtasks: event trigger word extraction and event argument extraction.However,traditional Chinese event extraction methods face problems of trigger word mismatch and inaccurate recognition in extracting event trigger words,inaccurate recognition of event arguments,and overlapping event argument roles in event argument extraction.To solve these problems,this paper proposes a Chinese event extraction method based on span regression.The specific research work is as follows:(1)In order to solve the problems of trigger word mismatch and inaccurate recognition,a trigger word extraction method based on span regression is proposed.This method considers that a specific length of character subsequence(span)in a sentence may constitute an event trigger word.The BERT pre-training language model of the bidirectional encoder based on Transformer is used to obtain the feature representation of the sentence,and then the candidate spans of the trigger word are generated.A classifier is used to filter out low-confidence candidate spans,and the boundaries of the candidate spans are adjusted by regression to accurately locate the trigger word.Finally,the adjusted candidate spans are classified to obtain the extraction results.Experimental results on the ACE2005 Chinese dataset show that the proposed method is superior to traditional models.(2)In order to solve the problems of inaccurate event argument recognition and overlapping argument roles,a method combining span regression and trigger word information is proposed for argument extraction.This method first accurately identifies event arguments by regression adjustment of candidate span boundaries,and adds argument length features when generating span feature representations to improve the recognition effectiveness of long event arguments.Then,by inserting typed markers to highlight event trigger words and event arguments in the text,the marked text is input into BERT to obtain a semantically enriched contextual sentence representation,and the feature representation of the trigger word start position and argument start position in the sentence representation is concatenated and input into a classifier for argument classification.Through experimental comparisons on public datasets,the effectiveness of the proposed method is verified.(3)A news event extraction system is designed and implemented.Based on the proposed method,an event extraction model is trained,and the event extraction model is deployed on a B/S architecture to realize a Web news event extraction system.The system supports automatic event extraction,event information visualization,and other functions,laying the foundation for the development of event extraction applications.