Beyond Linearization: Attributed Table Graphs for Table Reasoning
Proposes TabGR using Attributed Table Graph and QG-PPR to improve table reasoning accuracy by 9.7%.
Key Findings
Methodology
This paper introduces the TabGR model, which employs an Attributed Table Graph (ATG) to explicitly encode row-column-cell relations without task-specific training. ATG consists of nodes representing rows and cell values, connected by edges that preserve the table's structure. The Question-Guided Personalized PageRank (QG-PPR) mechanism assigns importance scores to graph triples based on the question, prioritizing key evidence. The process involves extracting relevant subgraphs from ATG, reranking triples via QG-PPR, and applying chain-of-thought reasoning with large models to generate answers. This integration of explicit graph structure and question guidance enhances evidence traceability, robustness against permutations, and interpretability.
Key Results
- On WikiTableQuestions and TabFact benchmarks, TabGR achieves 78.8% and 93.6% accuracy, outperforming the state-of-the-art RoT by 9.7%. It maintains stable performance under random table permutations, demonstrating improved robustness. Experiments with GPT-4 and LLaMA models show significant gains, especially on large and complex tables, validating the method's scalability.
- Ablation studies confirm the importance of question-guided reranking and ATG structure, with performance drops when these components are removed. The model performs well across different table sizes and content variations, indicating strong generalization and adaptability.
- Across multiple datasets and backbone models, TabGR not only boosts accuracy but also enhances evidence traceability and robustness, providing a more transparent reasoning process and addressing key limitations of linearized approaches.
Significance
This work advances table reasoning by explicitly modeling the full structural relationships within tables, overcoming the limitations of linearization. It significantly improves accuracy, robustness, and interpretability, facilitating applications in automated data analysis, knowledge extraction, and intelligent question answering. The integration of graph neural networks with large language models opens new avenues for complex reasoning tasks involving structured data, with broad implications for AI research and industry deployment.
Technical Contribution
The core contribution is the ATG structure that preserves global table topology, combined with QG-PPR for question-aware evidence prioritization. This enables end-to-end reasoning without task-specific training, leveraging graph neural networks for structural encoding and large models for logical inference. The approach offers theoretical guarantees of evidence importance and practical benefits in robustness and efficiency, representing a significant step beyond existing linearization and local graph methods.
Novelty
This is the first application of Attributed Table Graphs combined with question-guided PageRank for end-to-end table reasoning. Unlike prior work relying on linearized sequences or local graphs, ATG maintains full global structure, enabling explicit evidence tracing and permutation robustness. The integration of graph neural networks with large language models in this context is a novel contribution that pushes the boundary of structured data understanding.
Limitations
- Constructing ATG depends on predefined nodes and edges, which may be challenging in highly dynamic or unstructured tables. Scalability to extremely large or deeply nested tables remains computationally intensive.
- QG-PPR relies on question relevance signals that may be less effective in ambiguous or noisy queries, potentially reducing evidence prioritization accuracy.
- While robust to permutations, the method incurs higher computational costs due to graph construction and iterative PageRank calculations, requiring further optimization for real-time applications.
Future Work
Future research will focus on dynamic graph construction for real-time updates, integrating multimodal data (images, text) for richer reasoning, and developing end-to-end training strategies to jointly optimize graph encoding and evidence reranking. Exploring multi-task learning with knowledge graphs and expanding to more complex hierarchical tables are promising directions.
AI Executive Summary
Table reasoning is fundamental for extracting meaningful insights from structured data, yet traditional linearization approaches often dilute critical structural information, leading to reduced accuracy and interpretability. Existing methods, such as Chain-of-Thought prompting over serialized tables, struggle with content permutations and complex layouts, especially in large-scale or nested tables. These limitations hinder the deployment of reliable AI systems in real-world applications like automated data analysis, financial reporting, and knowledge extraction.
In response, this paper introduces TabGR, a novel framework that models tables as Attributed Table Graphs (ATGs). This explicit graph representation preserves the full row-column-cell relationships, enabling more faithful structural reasoning. To address the challenge of evidence prioritization, the authors propose the Question-Guided Personalized PageRank (QG-PPR) mechanism, which assigns importance scores to graph triples based on their relevance to the question. This approach effectively surfaces key evidence regardless of their original position in the table, mitigating the 'lost-in-the-middle' problem.
The core technical innovation lies in combining graph neural networks for structure encoding with large language models for logical reasoning. The process involves extracting a relevant subgraph from ATG, reranking triples via QG-PPR, and then performing chain-of-thought reasoning grounded in these prioritized triples. Extensive experiments on benchmarks like WikiTableQuestions and TabFact demonstrate that TabGR outperforms state-of-the-art models, achieving accuracy improvements of up to 9.7%. Notably, the model exhibits strong robustness to table permutations and structural variations, making it suitable for complex real-world scenarios.
This work significantly advances the field by providing a scalable, interpretable, and robust approach to table reasoning. Its ability to maintain structural integrity and explicitly trace evidence paths offers new opportunities for explainable AI and automated data understanding. Future directions include optimizing graph construction for larger tables, integrating multimodal data, and exploring end-to-end training strategies to further enhance performance and efficiency.
Deep Analysis
Background
表格推理作为理解和利用结构化数据的关键任务，经历了从符号规则到深度学习的演变。早期方法依赖规则和符号推理，代表性工作如TableBERT、TaBERT在编码表格结构方面取得一定成功，但在大规模复杂表格中表现不足。线性化表示虽简便，但导致信息稀释和路径难追踪，尤其在内容变换时性能波动明显。图模型如超图和实体关系图尝试捕获局部依赖，但缺乏全局结构表达，且多依赖预定义规则或任务特定训练，限制了泛化能力。近年来，结合图神经网络（GNN）与预训练大模型的研究逐渐兴起，旨在同时保持结构完整性和推理能力，推动表格理解向更深层次发展。
Core Problem
核心问题在于线性化表示导致信息稀释和推理路径难以追踪，尤其在内容排列变化或复杂结构场景下表现不佳。传统方法在大规模表格中容易丢失关键信息，推理路径不透明，难以解释。现有图模型多局限于局部关系，缺乏全局结构表达，且多依赖预定义规则或任务特定训练，限制模型的泛化能力。解决这一问题需要一种既能保持全局结构，又能突出关键证据的表示与推理机制，以提升推理的准确性和透明性。
Innovation
本研究提出Attributed Table Graph（ATG），显式编码表格中的行列关系，保持全局结构，避免信息稀释。引入Question-Guided Personalized PageRank（QG-PPR），根据问题引导对图中三元组进行排序，突出关键证据，缓解“中间证据丢失”问题。结合图神经网络和大模型链式推理，构建无任务训练的端到端推理流程。此创新融合了结构化表示与引导机制，显著提升推理性能、鲁棒性和可解释性，为表格理解提供新思路。
Methodology
- �� 构建Attributed Table Graph（ATG）：将表格中的行节点、值节点和边编码为图结构，保持完整的行列关系。• 提取结构子图：从ATG中筛选与问题相关的三元组，形成推理子图。• QG-PPR机制：利用问引导的个性化PageRank对三元组进行排序，赋予重要性分数。• 重排序：根据得分将关键证据提前，增强模型关注度。• 结合链式推理：在重排序的三元组上进行推理，生成答案和路径。• 证据追踪：路径映射回图中三元组，实现推理的可解释性。
Experiments
采用WikiTableQuestions、TabFact等数据集，比较多种基线，包括线性化和图模型。评估指标为准确率和鲁棒性。超参数如QG-PPR的α设为0.15或0.35，模型在不同大模型（GPT-4、LLaMA）上验证。进行消融实验，验证问引导和ATG结构的重要性。测试不同表格大小和内容变换，确保模型在复杂场景中的表现。结果显示，TabGR在准确率和鲁棒性方面均优于对比模型，验证了其有效性和实用性。
Results
TabGR在WikiTQ和TabFact上分别达到78.8%和93.6%的最高准确率，超越最先进的RoT模型，提升幅度达9.7%。在表格内容随机排列的测试中，性能波动显著减小，验证鲁棒性增强。消融实验表明，问引导机制和ATG结构对性能提升至关重要。模型在大规模、多层结构表格中的表现优异，验证了其在复杂场景中的适应性和扩展性。
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) for their semantic understanding and reasoning capabilities. 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 requires LLMs to infer row-column-cell relations from serialized inputs, makes evidence paths harder to trace, and is subject to the "lost-in-the-middle" issue. To address these issues, we propose Table Graph Reasoner (TabGR), a model that represents tables as an Attributed Table Graph (ATG) without task-specific training. The ATG explicitly preserves row-column-cell structure while enabling graph-based reasoning over traceable evidence paths. 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 across multiple table reasoning benchmarks show that TabGR consistently outperforms state-of-the-art models by up to 9.7% in accuracy. Our code is available at: https://github.com/yxw-11/TabGR.