GTR: Graph-Table-RAG for Cross-Table Question Answering

2025-04-02

AI Summary

To address the challenge of multi-table knowledge integration in cross-table question answering—exacerbated by scarce real-world training data—this paper proposes Graph-Table-RAG, a novel retrieval-augmented generation framework. It pioneers heterogeneous table modeling as a semantic graph and synergistically integrates graph neural network–enhanced hierarchical retrieval (coarse-grained table-level → fine-grained cell-level) with graph-structure-guided LLM prompting for efficient and faithful reasoning. Key contributions include: (1) MultiTableQA, the first large-scale, realistic multi-table QA benchmark comprising 60K tables and 25K natural-language questions; (2) a new graph-table co-retrieval RAG paradigm that unifies structural and semantic knowledge; and (3) state-of-the-art performance achieving optimal trade-offs between accuracy and latency—outperforming prior methods by significant margins and demonstrating strong viability for industrial deployment.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Knowledge Representation and Reasoning: Knowledge Acquisition
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Beyond pure text, a substantial amount of knowledge is stored in tables. In real-world scenarios, user questions often require retrieving answers that are distributed across multiple tables. GraphRAG has recently attracted much attention for enhancing LLMs' reasoning capabilities by organizing external knowledge to address ad-hoc and complex questions, exemplifying a promising direction for cross-table question answering. In this paper, to address the current gap in available data, we first introduce a multi-table benchmark, MutliTableQA, comprising 60k tables and 25k user queries collected from real-world sources. Then, we propose the first Graph-Table-RAG framework, namely GTR, which reorganizes table corpora into a heterogeneous graph, employs a hierarchical coarse-to-fine retrieval process to extract the most relevant tables, and integrates graph-aware prompting for downstream LLMs' tabular reasoning. Extensive experiments show that GTR exhibits superior cross-table question-answering performance while maintaining high deployment efficiency, demonstrating its real-world practical applicability.

Problem

• Develops a benchmark for cross-table question answering
• Proposes a Graph-Table-RAG framework for multi-table queries
• Enhances LLMs' reasoning with graph-aware prompting

Innovation

• Reorganizes tables into heterogeneous graph
• Hierarchical coarse-to-fine retrieval process
• Graph-aware prompting for LLMs reasoning