Research On Relation Extraction And Knowledge Graph Construction In The Filed Of Electric Power

With the improvement of knowledge graph construction technology,more and more applications based on knowledge graph begin to appear.People's attention began to shift from the generic domain knowledge map to the vertical domain knowledge map.Vertical domain knowledge graphs are domain-specific,with different data types and data patterns depending on business requirements.Relationship extraction is a sub-task of information extraction task,and it is one of the key steps to construct knowledge graph.It is intended to obtain the relationship of identified entities from unstructured data to fill the knowledge graph.This paper mainly studies the task of relationship extraction in the construction of knowledge map.For different types of power texts,different relationship extraction methods are used to extract entity triples in the field of power and draw the map.This paper studies from two aspects:(1)For the text of power dispatch management regulation,this paper uses the method of dependency syntax analysis to extract entity triples.Through the completion of the extracted entity and the processing of the juxtaposing structure of the long and difficult sentences,the more accurate entity triples around the core verbs are automatically extracted based on the basic Chinese grammatical structure.The results show that the relationship extraction method based on the dependency syntax analysis is still effective for the text of power dispatch and management regulation with obvious structure but complex sentence structure.Finally,the automatic extraction results are combined with the structured data in the text to draw a knowledge graph.(2)Aiming at the large amount of data available on the Internet,this paper constructs a data set for entity relationship extraction in the power field by distant supervision method.The effects of bidirectional long and short term memory network,convolutional neural network and piecewise convolutional neural network on sentence representation were compared.In addition,for the problem of noise in datasets built by distant supervision,the effects of multi-instance learning and attention mechanisms on noise reduction in data sets are compared.Conclusion it is proved that the convolutional neural network is slightly better than the bidirectional short and long term memory network in sentence representation on the self-built training set,and the multi-instance learning and attention mechanism can significantly reduce the noise of the data set.There is little difference between the piecewise convolutional neural network and the traditional convolutional neural network in the task of relation extraction.