Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Graph-based Verification Framework for Fact-Checking

Yani Huang, Richong Zhang, Zhijie Nie, Junfan Chen, Xuefeng Zhang

TL;DR

GraphFC introduces a graph-based fact-checking framework that reformulates claims as a bipartite triplet graph and grounds evidence via corresponding graphs. Its three components—graph construction, graph-guided planning, and graph-guided checking—enable fine-grained, context-preserving verification and improved handling of unknown entities through planning-driven grounding and completion. Across HOVER, FEVEROUS, and SciFact, GraphFC yields state-of-the-art macro-F1, with particular strength on multi-hop reasoning and robustness in Open Book settings. The work highlights the importance of structured evidence and planning in complex verification tasks, while acknowledging limitations in simple-claim gains and sensitivity to the graph-construction model, suggesting avenues for lighter-weight components and efficiency optimizations.

Abstract

Fact-checking plays a crucial role in combating misinformation. Existing methods using large language models (LLMs) for claim decomposition face two key limitations: (1) insufficient decomposition, introducing unnecessary complexity to the verification process, and (2) ambiguity of mentions, leading to incorrect verification results. To address these challenges, we suggest introducing a claim graph consisting of triplets to address the insufficient decomposition problem and reduce mention ambiguity through graph structure. Based on this core idea, we propose a graph-based framework, GraphFC, for fact-checking. The framework features three key components: graph construction, which builds both claim and evidence graphs; graph-guided planning, which prioritizes the triplet verification order; and graph-guided checking, which verifies the triples one by one between claim and evidence graphs. Extensive experiments show that GraphFC enables fine-grained decomposition while resolving referential ambiguities through relational constraints, achieving state-of-the-art performance across three datasets.

A Graph-based Verification Framework for Fact-Checking

TL;DR

GraphFC introduces a graph-based fact-checking framework that reformulates claims as a bipartite triplet graph and grounds evidence via corresponding graphs. Its three components—graph construction, graph-guided planning, and graph-guided checking—enable fine-grained, context-preserving verification and improved handling of unknown entities through planning-driven grounding and completion. Across HOVER, FEVEROUS, and SciFact, GraphFC yields state-of-the-art macro-F1, with particular strength on multi-hop reasoning and robustness in Open Book settings. The work highlights the importance of structured evidence and planning in complex verification tasks, while acknowledging limitations in simple-claim gains and sensitivity to the graph-construction model, suggesting avenues for lighter-weight components and efficiency optimizations.

Abstract

Fact-checking plays a crucial role in combating misinformation. Existing methods using large language models (LLMs) for claim decomposition face two key limitations: (1) insufficient decomposition, introducing unnecessary complexity to the verification process, and (2) ambiguity of mentions, leading to incorrect verification results. To address these challenges, we suggest introducing a claim graph consisting of triplets to address the insufficient decomposition problem and reduce mention ambiguity through graph structure. Based on this core idea, we propose a graph-based framework, GraphFC, for fact-checking. The framework features three key components: graph construction, which builds both claim and evidence graphs; graph-guided planning, which prioritizes the triplet verification order; and graph-guided checking, which verifies the triples one by one between claim and evidence graphs. Extensive experiments show that GraphFC enables fine-grained decomposition while resolving referential ambiguities through relational constraints, achieving state-of-the-art performance across three datasets.

Paper Structure

This paper contains 32 sections, 10 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Fact Checking
  4. PLM-based Fact Checking
  5. LLM-based Fact-Checking
  6. Method
  7. Fact-Checking over Graphs
  8. Task Definition
  9. GraphFC
  10. Graph Construction
  11. Claim Graph Construction
  12. Evidence Graph Construction
  13. Graph-Guided Planning
  14. Graph-Guided Checking
  15. Graph Match
  16. ...and 17 more sections

Figures (4)

  • Figure 1: An illustration demonstrating the differences between existing textual sub-claims and our graph-based sub-claims. The example is from the HOVER dataset jiang2020hover, and the decomposition results are from GPT-3.5-Turbo.
  • Figure 2: Overeview of GraphFC. The framework consists of three components: (1) graph construction (&\ref{['sec:Graph Construction']}), where claim graph and corresponding evidence graph are constructed for each claim; (2) graph-guided planning (&\ref{['sec:Graph-Guided Planning']}), which determines the verification order based on unknown entity count and plans the sequence of verification tasks; and (3) graph-guided checking (&\ref{['sec:Graph-Guided Checking']}), which executes either graph match for verifying the known entity triples or graph completion for inferring incomplete entity.
  • Figure 3: Ablation studies of GraphFC.
  • Figure 4: Distribution of task proportions for graph match and graph completion on HOVER