Research On Event Knowledge Extraction Of COVID-19 Based On Epidemic Notification Corpus

Epidemic notification is an important channel for governments at all levels to release information on the occurrence and development of infectious diseases.In this New Coronavirus pneumonia prevention and control,the behavior track of the case is the key content of the epidemic notification,and has received wide attention and discussion from the whole society.Effective extraction and integration of these behavior trajectory information can provide valuable data resources for epidemic traceability and monitoring,which is a prerequisite for intelligent and accurate epidemic prevention and control.On the basis of in-depth analysis of the insufficiency of knowledge extraction in emergency cases,this thesis studies the knowledge extraction of COVID-19 in low resource environment,and develops a prototype system according to the case behavior track information in the epidemic notification of Beijing municipal health and Health Committee.The details are as follows:(1)Completed the COVID-19 event knowledge extraction task design for epidemic traceability.In this study,the demand for COVID-19 traceability was analyzed,and a COVID-19 traceability model based on PROV was constructed to effectively model the behavior trajectory of cases.On this basis,a set of model-task mapping rules were designed to realize the task definition of event knowledge extraction driven by traceability model,and the annotated data set of epidemic event knowledge extraction covering six types of events was constructed through Text crawling,corpus preprocessing,and remote supervised corpus annotation.(2)AT-CPREE,a joint model of epidemic event extraction based on deep confrontation network,is designed and implemented.To solve the problem of insufficient labeled samples in low resource situation,a lightweight event extraction joint extraction model based on Bi LSTM+CRF+ sigmoid +Free AT was proposed by combining adversative learning and deep learning.In view of the fact that the event argument complexity is high but the event structure is relatively simple,this model adopts Bi LSTM+CRF+ sigmoid,a relatively simple structure that emphasizes argument recognition and light role extraction,to achieve fast convergence of the model on smallscale training sets by reducing parameters.Furthermore,an adversarial training mechanism based on Free AT is introduced to alleviate the over-fitting problem of lightweight model learning on small training sets.Experimental results show that the F value of the proposed event extraction method reaches 72.23% under low resource conditions,which is 4.78%-13.9% higher than that of the control model,and is significantly better than the existing mainstream event extraction pipeline model and joint model.(3)KC-CPRERE,an event causality extraction model based on the double constraints of common sense domain,is designed and implemented.In view of the problems of insufficient labeled samples and sparse dominant connectives in the extraction of epidemic event relationship under the condition of low resources,this study integrates knowledge coding,constraint learning and deep neural network,and develops a knowledge enhanced document level event causality extraction model.The model uses MLP-based knowledge coding and introduces external domain knowledge to enhance the representation of event causality.In addition,this study designed a constraint learning mechanism based on common sense and domain consistency to accelerate the convergence of the model on small training sets.Experimental results show that the average F value of the proposed method in the validation data set reaches85.45%,which is 5.4% higher than that of the traditional relational extraction model and superior to the traditional relational extraction model.It can effectively mine implicit relationships and improve the accuracy of knowledge extraction.