DeepTable: Structural Attention Biases and Tree Path Encoding for Hierarchical Table Understanding

RESUMO

LLMs process table content as linearized token sequences, weakening the two-dimensional and hierarchical structure encoded by multi-level row and column headers. DeepTable adds Structural Attention Bias, learnable biases representing whether token pairs share a row or column, and Tree Path Encoding, representing tokens by the ancestor paths of their headers. Integrated with TableLoRA across three LLM backbones, DeepTable achieves average gains of 7.42 points on HiTab, 3.23 on WikiTQ, and 2.01 BLEU points on FeTaQA.