Research On Key Technologies Of Emergency Extraction And Evolution

Emergency will damage the safety of life and property,public order and social environment.Therefore,it is urgent to establish an information extraction system for emergencies,extract the structured information and implied relationship from unstructured text,so as to quickly understand the information and evolution process of emergency,and provide data and technical support for the field of emergency decision-making.According to the characteristics of emergency,this thesis divides them into four categories: event occurrence,event description,event impact and post-processing.An event representation framework suitable for describing emergency is proposed,and the key issues of extraction and evolution of emergencies are studied based on the framework.The main research contents of this thesis are as follows:(1)An event sentence recognition method based on unknown trigger words is proposed.In order to improve the learning effect of semantic features,the BERT pre training model is used to optimize the word vector,and the syntactic and part of speech features are added as the input.The context semantic features of the word level are extracted by CNN,and the semantic features of the whole sentence sequence are extracted by Bi LSTM and attention mechanism to improve the recognition rate of event sentences.The event recognition effect of this method is verified by comparative experiments.(2)An event feature extraction method based on entity features is proposed.Based on the relationship between entity and event elements,entity feature vector is added as input to the word vector,and sequence features are extracted by Bi LSTM and CRF to label the elements in the event sentence.(3)Based on syntactic analysis and event representation framework,a completion algorithm is proposed for incomplete extraction of event elements.According to the dependency syntactic relationship between the elements in the event sentence,the rules and algorithms of elements completion are designed to provide a more complete information source for the joint analysis and evolutionary analysis of events.(4)Aiming at the relationship between events in the field of emergency,this thesis proposes an event relation extraction model which integrates the association feature of event sentences and relational pattern information.After splicing the two event sentences in order,the Bi LSTM joint coding is used to extract the association features between the two event sentences,and then the relational pattern features of event pairs are integrated to help the relationship classification,which breaks through the limitations of independent feature extraction of two event sentences and further improves the effect of event relationship extraction.(5)In order to make up for the deficiency of event relationship in emergency evolution analysis,a model of emergency evolution relationship extraction based on dual attention is proposed.Attention calculation is done for the scenario,event and their sentences to capture the correlation feature between scenario and event and analyzes the impact of scenario elements on emergencies.It provides the basis for the prediction of the development trend of the event.Finally,the design and implementation of the verification system is completed,which integrates various model results for event text to realize data visualization,and help people to see event information and evolution process more intuitively.