[Paper Review] Orthogonal Hierarchical Decomposition for Structure-Aware Table Understanding with Large Language Models

Bin Cao, Huixian Lu+4 more|arXiv (Cornell University)|Feb 2, 2026

Data Quality and Management0 citations

TL;DR

The paper introduces the Orthogonal Hierarchical Decomposition (OHD) framework with Orthogonal Tree Induction (OTI) and a dual-pathway association protocol to enable structure-aware table understanding for LLMs, achieving state-of-the-art results on AITQA and HiTab.

ABSTRACT

Complex tables with multi-level headers, merged cells and heterogeneous layouts pose persistent challenges for LLMs in both understanding and reasoning. Existing approaches typically rely on table linearization or normalized grid modeling. However, these representations struggle to explicitly capture hierarchical structures and cross-dimensional dependencies, which can lead to misalignment between structural semantics and textual representations for non-standard tables. To address this issue, we propose an Orthogonal Hierarchical Decomposition (OHD) framework that constructs structure-preserving input representations of complex tables for LLMs. OHD introduces an Orthogonal Tree Induction (OTI) method based on spatial--semantic co-constraints, which decomposes irregular tables into a column tree and a row tree to capture vertical and horizontal hierarchical dependencies, respectively. Building on this representation, we design a dual-pathway association protocol to symmetrically reconstruct semantic lineage of each cell, and incorporate an LLM as a semantic arbitrator to align multi-level semantic information. We evaluate OHD framework on two complex table question answering benchmarks, AITQA and HiTab. Experimental results show that OHD consistently outperforms existing representation paradigms across multiple evaluation metrics.

Motivation & Objective

• Address the challenge of understanding complex tables with multi-level headers, merged cells, and irregular layouts.
• Decouple table structure into independent column and row hierarchies to preserve hierarchical semantics.
• Reconstruct semantic lineage of table cells via dual-pathway associations and LLM-based arbitration.
• Demonstrate robustness over non-canonical table layouts on benchmark datasets (AITQA, HiTab).
• Show ablations to validate the contributions of semantic predicates, dual-pathways, and arbitration.

Proposed method

• Introduce Orthogonal Hierarchical Decomposition (OHD) which factorizes a table into a column tree and a row tree guided by Semantic-Spatial Synergy.
• Develop Orthogonal Tree Induction (OTI) with two stages: Header Skeleton Induction and Adaptive Data Anchoring, using semantic predicates and spatial constraints to build trees.
• Apply Dual-Pathway Association Reconstruction to build structured context for each data cell from primary and orthogonal axes, with boundary-aware anchoring.
• Use a Multi-pathway Semantic Arbitration where an LLM refines and synthesizes the final structure-aware representation from dual inputs.
• Provide a final structure-aware textual surrogate via zero-shot prompting to the LLM, optimizing for logical cohesion, completeness, and readability.

Figure 1 : Illustration of table complexity and structural diversity. The examples encompass several challenging non-standard layouts. (a) : Tables featuring multi-level nested column headers and merged data cells; (b) : Tables characterized by deep hierarchical row header structures; (c) : Complex

Experimental results

Research questions

• RQ1 How can complex tables with hierarchical headers and irregular layouts be decomposed into orthogonal structural representations that preserve semantics?
• RQ2 Can orthogonal (row/column) hierarchies plus dual-pathway association improve LLM reasoning over non-standard tables?
• RQ3 What is the impact of semantic predicates and LLM-based arbitration on table QA performance?
• RQ4 How does OHD perform on challenging benchmarks like AITQA and HiTab compared to linearization and schema-based baselines?

Key findings

• OHD achieves superior performance on AITQA and HiTab across EM and LLM-based evaluations compared to baselines.
• Using dual orthogonal trees (column and row) significantly improves robustness to flexible headers and non-canonical layouts.
• Semantic predicates and LLM-based arbitration are critical; removing them degrades performance notably.
• Lineage representations (Markdown/HTML) are substantially less effective than OHD's dual-pathway topology for complex tables.
• Ablation shows full OHD yields the best results across backbones (Qwen2-72B and TableLLaMA-7B).

Figure 2 : Workflow of the OHD framework. The process begins with a Categorized Table Input where each cell is pre-identified as a Row Header, Column Header, or Data unit. The pipeline then proceeds in three stages: (1) Orthogonal Tree Induction (OTI) to decompose the table into independent row and