Attribute Extraction of Geographic Entities Based on the LERT Pre-trained Model

Geographic information systems(GIS) are now changing from two-dimensional expression to three-dimensional expression, and the geographic data gradually transforms into a more intuitive and more applied method of expression. In order to support the data expression in GIS, attributes extraction from unstructured Chinese geographic texts is needed. Manual extraction or pattern-matching methods have been insufficient to get geographic attributes, therefore deep learning methods for attribute extraction are required to meet the needs of the supplementary for entity properties. This paper took LERT for pre-training in the word-embedding layer and the BiLSTM module combined with attention mechanism for the feature extraction layer, finally obtained the attribute-value pairs as the output. The experimental results show that the accuracy, recall rate and F1 value of the proposed model on the test set improved compared with the existing model, proving the effectiveness of the proposed model.