Research On Knowledge Graph Construction Technology For Power Text

With the further development of China’s electric power industry,a large number of electric power data appear.However,the current power data mainly takes the traditional unstructured text as the representation form,which is not conducive to computer processing.How to obtain the structured knowledge that is easier to be processed and understood from the power text and build the corresponding knowledge base is an important work to realize the intelligent upgrade of the power industry.Therefore,this thesis focuses on the entity and relationship extraction method for the power text,and constructs the power knowledge graph.The main research work includes the following three aspects.(1)Research on entity extraction methods for power text.During the entity extraction process based on deep learning,the traditional word vector models are static models,taking the corresponding word vector is fixed,character expression of polysemy is not very accurate,inadequate expression of the features in the statement,this thesis presents a method of electric text entity extraction based on BERT-BiLSTM-ATT-CRF model,dynamic generation of sequences of character vectors by using the BERT pretrained model,can vary depending on the context,characterize the polysemy of the characters,then we extract the vector sequence features after the BiLSTM model,then an attention mechanism is introduced to further extract internal features,finally,the sequence tag results are output through the CRF model.The experiments show that the entity extraction model designed in this thesis reaches 95.82%,94.68% and 95.25% in precision,recall rate and F1 value respectively,which are better than the traditional entity extraction model.(2)Research on relationship extraction methods for power text.In the process of relationship extraction based on deep learning,the representation of complex Chinese sentences by a single model can not fully mine the internal information of sentences.In response to the above issues,this thesis proposes a method for extracting power text relationships based on the BERT-BiGRU-PCNN model.Firstly,the BERT pre-trained model input statements are encoded to generate a sequence of text semantic vectors.Then,the sequence information is processed by the BiGRU model to fully capture the sequence information of longer sentences.Finally,the PCNN model is processed to mine prominent relationship features,this model not only has the ability of BiGRU to capture feature information of longer sentences,but also has the ability of PCNN model to mine local prominent relationship features,complementing the advantages of the two,and ultimately obtaining entity relationships.The experimental results show that the accuracy rate,recall rate and F1 value of the relational extraction model designed in this thesis reach 93.82%,92.32% and 93.06%,respectively,and all achieve good results.(3)Research the storage of power knowledge and complete the construction of the knowledge graph.On the basis of entity and entity relationship extraction,the Neo4 j graph database is used to store the ternary information and build the power knowledge graph,and the visualization interface is used to search and query the power information through the Cypher language provided by Neo4j.