Adjacency-Aware Table Graph Modeling for Enhanced Logical Structure Prediction
Angran Sun, Bo Quan Jiang, Hong Liang Xu · 2025
Tables serve as essential structures for organizing and conveying structured information across a wide range of domains. Early methods for table structure recognition often model a table either as a markup sequence or an adjacency matrix between cells, which neglects the logical location of each cell—such as being in the first row and second column—that is vital for downstream understanding. To address this, TGRNet introduces a table graph reconstruction framework that explicitly models the logical location of table cells through a graph neural network (GNN). By predicting the start and end positions of rows and columns, TGRNet achieves superior performance in logical structure recognition tasks. Despite its effectiveness, TGRNet heavily relies on clean, well-segmented table images, and its performance degrades significantly when encountering noisy inputs, distorted tables, or incomplete cell boundaries. Moreover, the use of fully connected graph structures limits its ability to leverage local structural priors effectively. In this paper, we propose a structure-aware enhancement module with an attention mechanism that improves the robustness and generalization of TGRNet by modifying the graph construction process and introducing structure-constrained priors. We further evaluate the model on synthetically degraded table images and demonstrate that our method significantly improves its robustness in challenging scenarios without requiring additional supervision. On the TableGraph-24K dataset, our method achieves an improvement in logical location accuracy compared to the baseline TGRNet, highlighting the effectiveness of introducing structure-aware graph construction.