FactCG: Enhancing Fact Checkers with Graph-Based Multi-Hop Data

2025-01-28

AI Summary

Existing fact-checking models exhibit limited long-text processing capability, high computational overhead, and strong dependence on large foundation language models for document-level LLM hallucination detection. To address these limitations, this paper proposes FactCG: a novel framework that first introduces CG2C, a context-graph-to-claim synthetic data generation method, which extracts context graphs from source documents and explicitly models multi-hop reasoning paths; it then constructs a graph-enhanced fact classification model that leverages structured graph reasoning to strengthen the multi-step logical discrimination capacity of compact models. Evaluated on the LLM-Aggregate benchmark, FactCG achieves superior factual consistency classification performance over GPT-4-o—despite using orders-of-magnitude fewer parameters—marking the first such result for a lightweight model. This demonstrates that synergistic context graph modeling and multi-hop reasoning significantly enhance both robustness and efficiency in fact verification.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

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

Abstract

Prior research on training grounded factuality classification models to detect hallucinations in large language models (LLMs) has relied on public natural language inference (NLI) data and synthetic data. However, conventional NLI datasets are not well-suited for document-level reasoning, which is critical for detecting LLM hallucinations. Recent approaches to document-level synthetic data generation involve iteratively removing sentences from documents and annotating factuality using LLM-based prompts. While effective, this method is computationally expensive for long documents and limited by the LLM's capabilities. In this work, we analyze the differences between existing synthetic training data used in state-of-the-art models and real LLM output claims. Based on our findings, we propose a novel approach for synthetic data generation, CG2C, that leverages multi-hop reasoning on context graphs extracted from documents. Our fact checker model, FactCG, demonstrates improved performance with more connected reasoning, using the same backbone models. Experiments show it even outperforms GPT-4-o on the LLM-Aggrefact benchmark with much smaller model size.

Problem

• Fact-Checking
• Large Language Models
• Computational Efficiency

Innovation

• CG2C
• FactCG
• Error Detection in LLMs