A Relation Extraction Method for GIS Defect Text Based on Self-attention
In the process of daily inspection and maintenance of power distribution networks,gas insulated switchgear(GIS)has accumulated a large number of defective texts.Structured extraction of the knowledge in these texts can effectively improve the efficiency of equipment operation and maintenance.To this end,this paper proposes a self-attention based relational extraction method for GIS defective text.First,a pre-trained language representation model is utilized to obtain word embedding vectors of input sentences as well as entity pairs.Subsequently,the extracted sentence embeddings are analyzed using a self-attentive neural network to extract sentence feature vectors.Then,entity feature vectors are obtained by performing a maximum pooling operation on the embeddings of the entity pairs,and positional features are obtained by encoding the position of the entity in the sentence.Finally,the sentence features,entity features,and positional features are classified to obtain the relationships of the entity pairs.The experimental results confirm the effectiveness of the proposed algorithm and achieve more than 70%accuracy on the dataset,which provides a powerful tool for constructing large-scale O&M knowledge graphs.
relation extractionneural networkGISnatural language processingknowledge extraction
万方数据
维普
