The SG-CIM Entity Linking Method Based on BERT and Entity Name Embeddings

The current iteration of the SG-CIM (State Grid Common Information Model) requires manual extraction of entity attributes from texts such as design plans and reports. To address the problems of slow update time and the high error rate of manual iteration data, this paper presents an entity linking method combing deep learning and knowledge base. Firstly, the SG-CIM model is used to construct a knowledge base of grid data used as a vector embedding of entities; Secondly, the joint recognition model of BERT-CRF and BERT-ENE(BERT-Entity Name Embeddings) is used for named entity recognition, where the BERT-ENE model can be used for dictionary matching of entity descriptions in the knowledge base; Then BERT-based binary classification model to predict the candidate entities, select the entity with the highest probability as the result, realize the entity disambiguation of alternate entities and new entities; Finally add the found important new entities to the SG-CIM model to realize the automated iteration of SG-CIM model. According to the experimental findings, the Entity Linking approaches have an F1-score of over 80%.