TableMaster: A Recipe to Advance Table Understanding with Language Models

TL;DR

TableMaster introduces a unified recipe for table understanding with language models, addressing four core challenges: locating target data, semantic deficiency, numerical inaccuracy, and symbolic reasoning rigidity. It combines table structure understanding (table-of-focus), table content understanding (verbalization and reconstruction), and adaptive reasoning (textual and text-guided symbolic) to enable robust, scalable QA over tabular data. Across WikiTQ, TabFact, and FetaQA, TableMaster achieves state-of-the-art results and demonstrates consistent gains across backbones, with ablations showing the critical importance of structure, content, and especially textual reasoning. The framework is designed to be broadly applicable to web tables, spreadsheets, and databases, offering a practical, efficient approach for real-world table reasoning tasks with LMs.

Abstract

Tables serve as a fundamental format for representing structured relational data. While current language models (LMs) excel at many text-based tasks, they still face challenges in table understanding due to the complex characteristics of tabular data, such as their structured nature. In this paper, we aim to enhance LMs for improved table understanding. We identify four key challenges: 1) difficulty in locating target data, 2) deficiency in table semantics, 3) numerical inaccuracies in textual reasoning, and 4) semantic inflexibility in symbolic reasoning. To address these issues, we propose TableMaster, a recipe and comprehensive framework that integrates multiple solutions to overcome these obstacles. TableMaster first extracts relevant table content and verbalizes it with enriched semantic context. Additionally, we introduce adaptive reasoning, a flexible approach that dynamically adjusts between textual and symbolic reasoning, tailoring the reasoning process to each query. Extensive analyses and experiments demonstrate our findings and the effectiveness of TableMaster. On the WikiTQ dataset, TableMaster achieves an accuracy of 78.13% using GPT-4o-mini, surpassing existing baselines.

Figures (29)

• Figure 1: Overview of the challenges and proposed solutions in this work. Tabular data is inherently structured, dense, concise, and numerical. Based on these characteristics, we identify four key challenges. To address them, we propose four targeted solutions. The gray arrows between the characteristics and challenges represent the potential causes of these challenges stemming from specific characteristics. Each proposed solution corresponds to the challenge presented on the left in the same row. TableMaster is a unified recipe developed based on these findings.
• Figure 2: Experimental analysis of challenges in table understanding with language models. (a) Impact of table size on task difficulty. (b) Effect of verbalized tables with enriched semantic context. (c) Performance comparison of different reasoning methods on calculation-required versus non-calculation questions. (d) Performance differences when processing normalized versus noisy tables.
• Figure 3: The framework of TableMaster. It comprises three stages: (1) table structure understanding, where the table’s structure is analyzed, and a table-of-focus is constructed through row and column lookup; (2) table content understanding, where the table-of-focus is reconstructed based on the question, and its information is verbalized to enhance the semantic context; and (3) table reasoning for question answering, where an adaptive reasoning strategy determines whether to use textual reasoning or text-guided symbolic reasoning to derive the final answer. The dashed arrows indicate optional workflows, such as the table-of-focus re-construction and incorporating text-guided symbolic reasoning.
• Figure 4: An example (fetaqa-164) from the FetaQA dataset where the result is accurate, but the evaluation metric assigns a low score.
• Figure 5: Performance Comparison Across Table Sizes (Row Count, Column Count, Area Size, Token Count).