Research And Application Of Event Extraction Method In Disaster Field Based On Short Text

With the development of mobile Internet,microblog and other short text messages are increasing rapidly.Compared with the traditional long text messages represented by newspapers,short texts have the advantages of strong timeliness and rapid dissemination.These short texts are mostly uploaded spontaneously by ordinary users,which contain a lot of effective information and have great application value.The event extraction of short texts in a disaster field can obtain structured information and provide data support for practitioners in related fields.This dissertation mainly studies event extraction methods and applications in the field of disasters.It is divided into three tasks: named entity recognition,event type detection,and event relationship recognition.The focus of disaster event research is to obtain the time,place and type of disaster.Therefore,it is very important to identify the core information in disaster sentences and judge the relationship between event sentences.Crawler technology is used to obtain short text data from Weibo to solve the problem of insufficient data.Finally,the research data set is constructed by manual annotation.For the task of named entity recognition,sequence labeling method based on the BERT pre-training model is used to predict the label of each character.In order to combine the needs of practical applications,this dissertation further analyzes the extracted time and location information,and obtains the standardized expression of them,which achieves a good application effect.For the task of event type detection,a type detection model that combines the characteristics of trigger words and the description of the event type is proposed.The model combines the similarity between trigger words and disaster type descriptions,and transforms the type detection task into a feature matching task.This model has a certain improvement compared with the benchmark model.For the task of event relationship recognition,trigger word tags and disaster type description information are used to expand text semantics.A relationship recognition model that combines trigger word features and event type descriptions is proposed.This model transforms the relation recognition task into a feature matching task between two sentences.Trigger words and disaster type descriptions can characterize the text in different dimensions.This model also achieved better results compared with the benchmark.Based on this research work,this dissertation applies the proposed model in practice,and builds an intelligent analysis mining system for meteorological disaster information.The system uses the data crawler module as the data support,and comprehensively describes disasters through the entity recognition module.It uses the type detection module to determine the type of disasters,and counts the number of various disaster information.The complete data display process within the system has been implemented.