TabBridge

← Back

TabBridge: Bridging Structure and Context for Accurate Table Reasoning

SURGeLLM Workshop at ACL 2026, Best Paper Award · Paper Code Slides

Problem

Text-to-SQL-based table reasoning depends on schema linking: mapping the question's terms to the table's column headers. This fails when the two differ on the surface even while meaning the same thing. If a table's column is named gross but the question asks about earnings, the model anchors on the question's vocabulary and generates SQL over an earnings column that does not exist, producing a query that cannot execute at all. We observed this same failure pattern consistently across multiple models, which pointed away from a model-capability explanation. Our diagnosis: the pipeline itself processes the table's structural information and the question's contextual information separately, so nothing in the design ever forces the two to meet.

Approach

The framework draws on Kintsch's Construction–Integration model of human text comprehension, which describes understanding in two phases: first constructing all plausible meaning representations independent of context (reading bank, both "financial institution" and "riverbank" activate), then integrating with context to reinforce the fitting meaning and suppress the rest. TabBridge maps this two-phase structure onto table reasoning.

Stage 1: Structural Encoding. Without seeing any question, the model interprets the table and encodes its information into TabSpec, a textual specification of the table's structure and semantics. To verify that TabSpec faithfully captures the table, we designed reconstruction-based evaluation: a model regenerates the table from the TabSpec alone, and the reconstruction is scored against the source on Header Exact Match and Cell Similarity. This validates structural fidelity before TabSpec is ever used downstream.

Stage 2: Contextual Reasoning. Given a question, the model generates SQL grounded in the TabSpec, resolving the question's context to headers that actually exist; earnings now finds gross.

Results

73.94% on WikiTableQuestions, a +5.3pp improvement over prior SOTA, where SQL-only approaches reach around 50%. Performance was comparable to SOTA on TabFact and FeTaQA. Across open-source models, TabBridge delivered consistent gains of roughly 10–20%. Since WTQ uses raw, unnormalized tables, the results also indicate the approach handles non-standardized table formats.

Limitations. Structure–context alignment succeeded, but SQL generation itself remains a bottleneck: cases exist where the correct header is identified yet the generated SQL is still wrong, pointing to SQL generation quality as the next problem.

This work builds on three earlier studies on LLM table interpretation: few-shot prompting with data augmentation (HCLT 2024), least-to-most prompting (KICS 2025), and open- vs. closed-source model comparison (ISIS 2025). The model comparisons from this line of work also led to an open-source tool for running and comparing these approaches directly on local LLMs via Ollama.