Research On Emergency Relation Recognition Based On Deep Learning

Emergencies refer to incidents that occur suddenly and produce serious hazards.They are sudden,uncontrollable and unpredictable.Due to their impact on human society,relevant departments are often required to make emergency response strategies in a timely manner.The recognition of emergency relations can reflect the logical connection between emergencies,which is an inevitable requirement for deep understanding of emergency texts.Efficient,timely and accurate recognition of emergencies and their relations is of great significance for mastering the ecological cycle of emergencies and related upper-level applications.Aiming at the problems of low coverage,poor fast parallel ability and insufficient representation ability in the current research on event relationship recognition,this paper proposes an emergency relationship recognition method based on deep learning by using natural language processing and deep learning related technologies and combining with the characteristics of emergencies.The method constructs three deep learning models.They are BERT-BiGRU-CRF's emergency recognition model,double-layer CNN-BiGRU-CRF's event causal relationship recognition model and Transformer-CRF's event following relationship recognition model respectively,and are verified by comparative experiments in CEC corpus.Provide decision-making basis for emergency response agencies to effectively respond to all kinds of emergencies.This paper is oriented to the field of emergencies,and has carried out a series of key theoretical and technical researches on the task of event identification,event causality and follow-up identification.The main research contents are as follows:(1)Event recognition refers to identifying events in textual information based on event trigger word arguments.Aiming at the problem of emergency event recognition,a based on BERT-BiGRU-CRF model emergency event recognition method is proposed.This method first uses sample data to fine-tune the BERT model,and then fuses the feature vector matrix generated by the BERT model with the four internal structural features of the emergency;then the fused multi-features are input to the BiGRU network for deep feature extraction;finally,The output uses a conditional random field model(CRF)to label trigger words based on the maximum probability.Analysis of the experimental results on the expanded experimental set shows that the F value of the method reaches 94.96%,which proves the effectiveness of the method.(2)Causality reflects the logical relationship between antecedents and consequences,causes and effects between emergencies,which is extremely common in emergent text messages.Aiming at the problems of current causality extraction of events,such as weak causality boundary recognition ability and insufficient semantic representation,an event causality extraction method based on a two-layer CNN-BiGRU-CRF model is proposed.This method uses a step-by-step processing idea to convert the causality identification task into two sequence labeling tasks,which are respectively completed by two layers of CNN-BiGRU-CRF models.In each layer model,first,use the CEC corpus to fine-tune the BERT model to form a text representation model of the emergency domain to obtain the feature vector matrix;then,pass it to the CNN and BiGRU models for local and global deep feature extraction,respectively;Finally,the residual idea is used to input the highly distinguished features of weighted fusion into the CRF model to decode the sequence labeling task.By comparing the experimental results on the CEC experimental corpus with other advanced working methods,the event-causality extraction F value is 91.81%,which verifies the effectiveness of the proposed method.(3)The occurrence of an event is strongly related to another event,and an event usually follows the appearance of another event.In order to deal with the sudden and rapid characteristics of emergencies and identify the follow-up relationship of an emergency,a method for identifying the follow-up relationship of an emergency based on the Transformer model is proposed.This method first uses the BERT model to obtain the feature vector matrix of the input samples;then the generated feature vector is fused with the four features inside the follow-up relationship of the incident,and the multi-feature fused vector matrix is input to the Transformer model for deep feature extraction.The highly distinguished features of weighted fusion are passed to the CRF model to decode the sequence labeling task.By comparing the experimental results on the CEC experimental corpus with other deep learning models,the model training efficiency is improved by about 58.6%,and the model F value is 87.01%,which verifies the effectiveness of the proposed method.