Automated extraction of structured information from digital documents

Abstract

We study information extraction from digital documents, including page images and PDF files, with special emphasis on energy invoices, which motivate the practical problem addressed in this work. We organize the pipeline into five stages: element detection, optical character recognition (OCR), key-value pair (KVP) linking, table structure recognition, and semantic normalization. The main methodological contribution is a compact Graph Neural Network (GNN) for KVP linking that extends Doc2Graph with selective edge generation, attention-based message passing, and focal loss. The second contribution is a multi-view normalization module that maps heterogeneous invoice fields to a standardized schema with 27 categories. The evaluation combines public tests with an invoice-oriented semantic normalization study and uses a cross-dataset train/test design to separate in-domain performance, transfer to a different dataset, and combined-training behavior. On FUNSD, the proposed GNN reaches 𝐹1 = 0.77 with 890 K parameters, while BROS reaches 𝐹1 = 0.80 with 111 M parameters. Across the full public KVP cross-dataset evaluation, the proposed GNN increases the average cross-dataset 𝐹1 from 0.19 (Doc2Graph) to 0.61 and matches BROS’s average cross-dataset performance with roughly 124 times fewer parameters. For table structure recognition, the domain-pretrained Table Transformer reaches 𝐹1 = 0.96, compared with 𝐹1 = 0.83 for a generic DETR detector. For semantic normalization, the mixed-invoice split reaches top-1/top-3 = 0.77/0.88 with ground-truth candidates but drops to 0.43/0.52 when fed KVP+TSR candidates. When invoices are trained and tested on providerspecific layouts, the corresponding provider mean remains much higher at 0.83/0.90. These results show that the main gains come from task-specific graph construction and from an explicit normalization stage that converts heterogeneous invoice fields into a common schema.