Few-Shot Supervised OCR Verification with Graph Convolutional Networks on Relational Line Graphs

Optical Character Recognition (OCR) systems are foundational to document digitization, yet their outputs often contain errors that require costly manual verification. Most automated verification methods are supervised, demanding large, meticulously labeled datasets of correct and incorrect text, which presents a significant bottleneck. In this paper, we propose a novel few-shot supervised framework for post-hoc OCR verification using Graph Neural Networks (GNNs). Our primary contribution is a unique graph-based representation of a word. We model each word as a line graph, where each node corresponds to a relational comparison—either between a real character and its synthetic counterpart (\((\text{R}_i)\)-\((\text{S}_i)\)) or between two sequential real characters (\((\text{R}_i)\)-\((\text{R}_{i+1})\)). These nodes are encoded with a rich feature vector combining classic geometric properties (Hu Moments, pixel counts) with deep visual features and a cross-entropy stability metric derived from data augmentation. We frame verification as a few-shot graph classification task, benchmarking GCN, GAT, and GraphSAGE architectures across varying depths on a dataset collected from approximately 50 archival books. Contrary to the assumption that deeper models yield better representations, our results demonstrate that a lightweight 3-layer Graph Convolutional Network (GCN) outperforms deeper 5-layer variants and attention-based models. With as few as 50 labeled examples, this compact model achieves competitive error detection performance, establishing a highly effective and scalable alternative for OCR verification that mitigates the dependence on large-scale labeled data.