Research And Implementation Of Network Group Event Extraction Technology Based On Q&A Task

Network group events are the events which spread through the Internet and are attended by many netizens,these events are generally topical or controversial.Event extraction is one of the important tasks in the field of information extraction,which aims to transform unstructured event description text into structured event description.The study of event extraction technology in Chinese Network Group Event field has great value to content security,emergency management,public opinion analysis and other fields.Considering serious error propagation which caused by long pipeline of Event Extraction model,this paper proposes a pre-training model based method named Chinese Event Extraction by Machine Reading Comprehension(CEEMRC),which simplifies event extraction task to a cascade of only two question answering models.We construct the question for event trigger word extraction,event type determination and attribute extraction.It predict the start and end positions of the answer.to complete required extraction based on the pre-training model RoBERTa.In this paper,DuEE Chinese event data set is used for the experiment,The F1 value of trigger extraction and element extraction are better than similar methods.Considering the vague of network group events field’s definition and the lack of annotation corpus,this paper gives the framework of network group events’ descriptions based on the summary of Chinese news and Twitter.It gives the type of events,triggers and event’s arguments.Under the framework’s guide,we have completed annotation of event data with iteration rules and data enhancement.Considering the small size of annotated corpus and the high cost of annotating network group event data,this paper uses semi-supervised learning method based on UDA(Unsupervised Data Enhancement)to train the model.The model smoothness is enhanced by minimizing the consistency loss between the unlabeled data and the enhanced data.In addition,a text classifier based on FastText is added to improve the efficiency of model prediction.The network group event labeled data and unlabeled data are used in the experiments.Experimental results show that compared with the training method which totally based on annotated data,this method can improve the model F1 value by more than 3%.Finally,based on the above model method,this paper designed and implemented the system of network group event extraction,and completed the realization of data upload,data annotation,text event extraction,model training and other functions.Functional testing’s results show that this system can meet the requirements of event extraction in the field of network group event.