Research On Entity Relationship Extraction Technology For Long Tail Distribution Of Urban Public Safety Data

The amount of urban public safety data is complicated and huge,and human resources cannot extract information from these data efficiently and accurately.Knowledge graphs enable more efficient and more accurate urban public safety management by embedding things in the city as entities and extracting the relationships between the entities.Entity relationship extraction is the core part of building knowledge graphs,and there is a relative lack of tail class instances in the long-tail distribution of urban public safety data,which causes unsatisfactory accuracy of entity relationship extraction due to the problem of few samples,so how to improve the accuracy of entity relationship extraction with few samples has become an important issue that needs to be solved urgently.According to the characteristics of entity relationship extraction in data long-tail distribution,the topic is divided into vertical domain entity relationship extraction study and cross-domain entity relationship extraction study.In vertical domain scenarios,this article proposes the bidirectional matching aggregation model,which improves the accuracy of the model for tail instance relationship extraction using bidirectional training.In the cross-domain scenario,this article proposes the bidirectional global transformation model,which improves the model’s ability to identify relationship types in relationship obfuscation through bidirectional global transformation,thus enhancing the accuracy of the model for tail instance relationship extraction.The main work is as follows :(1)The bidirectional matching aggregation network model BMAN based on few-shot learning is proposed,which exploits the symmetry of data to learn more features of relational prototypes.The model first predicts the relations of query instances using the prototype network,and then calculates the relational prototypes of query instances using the obtained relations to verify the accuracy of the obtained relations and to correct the deviations of the model predictions.(2)The data enhancement method is designed,which uses different mapping methods for different task scenarios to ensure that each query instance participates in the calculation of the inverse relationship prototype,obtains the connection between forward and inverse training,and at the same time expands the number of instance samples within the range of instance relationship types to avoid overfitting in bidirectional matching.(3)The bidirectional global transformation network model BGTN based on fewshot learning is proposed,which uses global transformation to learn discriminative information between candidate classes for transferring more knowledge in cross-domain task scenarios,and also uses representation learning in the spherical coordinate embedding space to aggregate instance-embedded representations within classes and separate instance-embedded representations between classes in the embedding space.