A Table Classification Approach Utilizing Graph Attention Network

This paper presents an advanced table classification approach that integrates graph neural networks (GNNs) with linear attention mechanisms to address challenges in complex table semantic analysis. Traditional methods relying on stylistic features or shallow context modeling exhibit limited performance in handling irregular tables with merged cells or heterogeneous layouts. To overcome these limitations, we propose a dual-branch feature fusion model, DFCGA, which synergistically combines structural and semantic modeling. The model comprises three core modules: SLA-DWA dynamically adjusts attention windows to balance global semantic correlation modeling and local detail capture, addressing long-range dependencies and irregular table boundaries; MHCAF integrates structural features from graph attention networks and semantic features from SLA-DWA through channel-level attention, resolving feature fragmentation and enhancing multi-scale representation; GGF representations by fusing layout-aware pooling and semantic-guided weighting (prioritizing critical regions like headers). Experimental results on the WCC and Salt Cavern Gas Storage datasets demonstrate that DFCGA achieves superior performance compared to baseline models, with an 86.02% F1-score on WCC. Ablation studies confirm the effectiveness of SGF and GGF in improving classification accuracy, particularly for irregular tables. The proposed method provides a robust solution for real-world complex table analysis tasks.