Research and Application of Document Key Information Extraction Based on Pre-trained Models in the Digital Handover of Power Grids

Addressing the need to rapidly and efficiently extract various key field information from documents during the digital handover of power grids, this paper introduces a method for document key information extraction based on pre-trained models. It extracts key information paragraphs through heuristic rules based on keywords or chapter levels and uses a unified information extraction model based on ERNIE3.0 to extract the required key field information from these paragraphs. In the unified information extraction model, this paper proposes the integration of a pointer network with global semantic information and a data augmentation method based on the large language model. This significantly enhances the model’s capability to extract key field information. Experimental results show that the F1 score of the unified information extraction model based on the pre-trained model introduced in this paper is approximately 95.2, an increase of 2.2 percentage points compared to the open-source UIE-Paddle model from Baidu, which validates the effectiveness of the proposed method.