Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Linearization: Attributed Table Graphs for Table Reasoning

Yuxiang Wang, Junhao Gan, Shengxiang Gao, Shenghao Ye, Zhengyi Yang, Jianzhong Qi

TL;DR

This work tackles the limitations of linearized table reasoning by introducing TabGR, a training-free framework that represents tables as Attributed Table Graphs to preserve row-column-cell structure. It couples ATG-based table decomposition with a Question-Guided Personalized PageRank mechanism (QG-PPR) to surface the most evidence-grounded triples and guide an LLM to generate explicit, fine-grained reasoning paths. Empirical results on WikiTableQuestions and TabFact show consistent accuracy gains over state-of-the-art methods, with added robustness to table permutations and improved token efficiency. The approach also demonstrates extensibility to complex HiTab structures and yields interpretable reasoning traces that align with table topology, enabling more reliable and explainable table reasoning in real-world applications.

Abstract

Table reasoning, a task to answer questions by reasoning over data presented in tables, is an important topic due to the prevalence of knowledge stored in tabular formats. Recent solutions use Large Language Models (LLMs), exploiting the semantic understanding and reasoning capabilities of LLMs. A common paradigm of such solutions linearizes tables to form plain texts that are served as input to LLMs. This paradigm has critical issues. It loses table structures, lacks explicit reasoning paths for result explainability, and is subject to the "lost-in-the-middle" issue. To address these issues, we propose Table Graph Reasoner (TABGR), a training-free model that represents tables as an Attributed Table Graph (ATG). The ATG explicitly preserves row-column-cell structures while enabling graph-based reasoning for explainability. We further propose a Question-Guided Personalized PageRank (QG-PPR) mechanism to rerank tabular data and mitigate the lost-in-the-middle issue. Extensive experiments on two commonly used benchmarks show that TABGR consistently outperforms state-of-the-art models by up to 9.7% in accuracy. Our code will be made publicly available upon publication.

Beyond Linearization: Attributed Table Graphs for Table Reasoning

TL;DR

This work tackles the limitations of linearized table reasoning by introducing TabGR, a training-free framework that represents tables as Attributed Table Graphs to preserve row-column-cell structure. It couples ATG-based table decomposition with a Question-Guided Personalized PageRank mechanism (QG-PPR) to surface the most evidence-grounded triples and guide an LLM to generate explicit, fine-grained reasoning paths. Empirical results on WikiTableQuestions and TabFact show consistent accuracy gains over state-of-the-art methods, with added robustness to table permutations and improved token efficiency. The approach also demonstrates extensibility to complex HiTab structures and yields interpretable reasoning traces that align with table topology, enabling more reliable and explainable table reasoning in real-world applications.

Abstract

Table reasoning, a task to answer questions by reasoning over data presented in tables, is an important topic due to the prevalence of knowledge stored in tabular formats. Recent solutions use Large Language Models (LLMs), exploiting the semantic understanding and reasoning capabilities of LLMs. A common paradigm of such solutions linearizes tables to form plain texts that are served as input to LLMs. This paradigm has critical issues. It loses table structures, lacks explicit reasoning paths for result explainability, and is subject to the "lost-in-the-middle" issue. To address these issues, we propose Table Graph Reasoner (TABGR), a training-free model that represents tables as an Attributed Table Graph (ATG). The ATG explicitly preserves row-column-cell structures while enabling graph-based reasoning for explainability. We further propose a Question-Guided Personalized PageRank (QG-PPR) mechanism to rerank tabular data and mitigate the lost-in-the-middle issue. Extensive experiments on two commonly used benchmarks show that TABGR consistently outperforms state-of-the-art models by up to 9.7% in accuracy. Our code will be made publicly available upon publication.
Paper Structure (43 sections, 7 equations, 3 figures, 15 tables)

This paper contains 43 sections, 7 equations, 3 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Linearization-based Table Reasoning
  4. Graph Modeling for Tabular Tasks
  5. TabGR
  6. Attributed Table Graph
  7. ATG-Based Table Decomposition
  8. Question-Guided Personalized PageRank
  9. Tripe Salience Scores
  10. Personalization Vector
  11. Propagation Matrix
  12. Triple Reranking
  13. Answer Generation
  14. Experiments
  15. Experimental Setup
  16. ...and 28 more sections

Figures (3)

  • Figure 1: Motivating example. (a) Limitations of linearization-based methods. (b) Performance of different methods on WikiTableQuestions pasupat under random table permutations.
  • Figure 2: Overview of TabGR, a graph-based framework for table reasoning. The input table is first transformed into an Attributed Table Graph (ATG), which supports Question-Guided Personalized PageRank (QG-PPR) for estimating the salience of table triples and reranking evidence. The reranked evidence then guides an LLM to perform reasoning over the ATG, producing an explicit and fine-grained reasoning path.
  • Figure 3: Case study comparing TabGR with strong baselines: Table-Critic and RoT.