Year: 2025
Type: article
Abstract: Entity linking in visually rich documents (VRDs) is critical for industrial automation but faces challenges from complex layouts and computational inefficiency in existing models. Traditional approaches relying on pre-trained transformers or graph networks struggle with noisy OCR outputs, a high number of parameters, and invalid edge predictions in industrial VRDs. To Vision Information Extraction (VIE) for Industrial VRDs, we introduce a lightweight multi-rule graph construction network for entity linking, integrating text and layout embeddings as graph nodes. A multi-rule filtering method reduces invalid edges using node distance, link interference, and standardization rules inspired by document production logic and reading habit. A node relation enhancement module with Graph Attention Networks (GAT) enhances nodes through multi-rule edges and attention scores, enabling robust reasoning for noisy and complex layouts. Evaluated on FUNSD and SIBR datasets, our model achieves F1 scores of 58.96% and 65.08% with only 18M parameters, outperforming non-pretrained baselines while maintaining deployability for resource-constrained environments.
Authors: Yinhuan Zheng, Jiabao Chen, Weiru Zhang
Institutions: Wuhan University of Technology, Wuhan Textile University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4409778351