Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models

Han Cao, Lingwei Wei, Wei Zhou, Songlin Hu

TL;DR

This paper introduces LLM-SKAN, a novel framework for multi-hop fact verification that leverages Large Language Models (LLMs) to extract fine-grained entity-relation triplets. By augmenting these triplets with a knowledge graph and a graph fusion module, LLM-SKAN significantly improves the model's ability to understand complex, multi-step reasoning required for verifying claims based on multiple pieces of evidence.

What Problem Does This Paper Address?

The proliferation of misinformation on social media necessitates robust fact-verification systems. While single-hop fact verification (checking a claim against a single piece of evidence) has seen progress, multi-hop fact verification presents a greater challenge. This task requires models to synthesize information from multiple, often interconnected, pieces of evidence, each with its own complex internal logic and relationships between entities. Existing methods often struggle to capture these intricate, multi-step dependencies, sometimes resorting to shortcuts rather than deep understanding. Furthermore, they tend to overlook the crucial role of fine-grained relationships between entities within the evidence, which are vital for accurate multi-step reasoning.

Methods

The proposed LLM-SKAN framework is built around four key components:

  1. Structured Knowledge-Augmented LLM: Instead of using LLMs directly for prediction, the authors fine-tune a Llama2 model on a document-level entity relation extraction dataset (DocRED-FE) using the LoRA mechanism. This specialized fine-tuning aims to enhance the LLM's ability to extract structured knowledge, specifically entity-relation-entity (ErE) triplets.
  2. LLM-driven Knowledge Extractor: This component utilizes the fine-tuned LLM to extract fine-grained information, focusing on entities and their relationships from both the claim and evidence. It prompts the LLM to output structured triplets in the format (entity, relation, entity). This approach is designed to be more efficient than traditional knowledge graph-based methods.
  3. Knowledge-Augmented Relation Graph Fusion: To model the complex dependencies between claims, evidence, and extracted entities, a heterogeneous graph is constructed. Nodes represent claims, evidence, and entities, while edges represent their relationships. This graph is then processed by a Graph Neural Network (GNN) to learn comprehensive representations by iteratively updating node embeddings, capturing intricate interactions.
  4. Fact Verification: The fused claim-evidence representation from the graph fusion module is fed into a simple classifier (a 2-layered fully connected network) to predict the final verdict (e.g., Supported, Refuted).

Datasets

The authors evaluate LLM-SKAN on four publicly available datasets:

  • FEVER: A large-scale dataset for single-hop fact verification, containing over 180,000 human-generated claims with evidence from Wikipedia. Claims are categorized as Supported, Refuted, or NEI (Not Enough Info).
  • HOVER (2-, 3-, and 4-hop): Datasets specifically designed for multi-hop fact verification, containing over 15,000 claims that require multi-step reasoning. Evidence is sourced from Wikipedia, and claims are labeled as Supported or Not-Supported.

Key Insights and Ideas

  • LLMs as Relation Extractors, Not Predictors: The paper argues that LLMs excel at understanding and extracting structured information (like relations) but are less adept at complex reasoning for prediction tasks, especially in multi-hop scenarios. This insight leads to their novel approach of using LLMs primarily for knowledge extraction.
  • The Power of Fine-Grained Entity Relations: The study highlights that capturing intricate relationships between entities (ErE triplets) is crucial for understanding the "inner logic" of evidence and enabling effective multi-step reasoning in fact verification.
  • Graph Structures for Multi-Hop Reasoning: The research reinforces the idea that graph-based representations are highly effective for modeling complex interdependencies in multi-hop tasks, outperforming simpler fusion methods like concatenation or sequential attention.
  • Structured Knowledge Augmentation is Key: By combining LLM-extracted fine-grained knowledge with graph structures, LLM-SKAN demonstrates a significant improvement in handling the complexity of multi-hop fact verification.

Key Results

  • LLM-SKAN consistently outperforms various baseline methods, including graph-based and LLM-based approaches, on both multi-hop and single-hop fact verification tasks across the FEVER and HOVER datasets.
  • Ablation studies show that removing the LLM-driven Knowledge Extractor leads to a dramatic decline in performance, underscoring the importance of fine-grained entity information.
  • The Knowledge-Augmented Relation Graph Fusion module also proves critical, with performance degrading when replaced by simpler fusion techniques.
  • The model achieves competitive performance on single-hop tasks, demonstrating its versatility.

Limitations and Future Work

  • LLM Reasoning Limitations: The authors acknowledge that LLMs can still struggle with retaining contextual information over long documents, which can hinder their reasoning capabilities. This is why they opt to use LLMs for extraction rather than direct prediction.
  • Fine-tuning LLMs: While fine-tuning LLMs for relation extraction proves effective, it requires specialized datasets and computational resources.
  • Future Directions: The paper suggests that further exploration of LLM capabilities in fact verification, potentially through more advanced fine-tuning strategies or hybrid approaches, could yield even better results. The effectiveness of structured data and graph fusion for complex reasoning is also a promising avenue for future research.

Summary updated

Abstract

The rapid development of social platforms exacerbates the dissemination of misinformation, which stimulates the research in fact verification. Recent studies tend to leverage semantic features to solve this problem as a single-hop task. However, the process of verifying a claim requires several pieces of evidence with complicated inner logic and relations to verify the given claim in real-world situations. Recent studies attempt to improve both understanding and reasoning abilities to enhance the performance, but they overlook the crucial relations between entities that benefit models to understand better and facilitate the prediction. To emphasize the significance of relations, we resort to Large Language Models (LLMs) considering their excellent understanding ability. Instead of other methods using LLMs as the predictor, we take them as relation extractors, for they do better in understanding rather than reasoning according to the experimental results. Thus, to solve the challenges above, we propose a novel Structured Knowledge-Augmented LLM-based Network (LLM-SKAN) for multi-hop fact verification. Specifically, we utilize an LLM-driven Knowledge Extractor to capture fine-grained information, including entities and their complicated relations. Besides, we leverage a Knowledge-Augmented Relation Graph Fusion module to interact with each node and learn better claim-evidence representations comprehensively. The experimental results on four common-used datasets demonstrate the effectiveness and superiority of our model.
BibTeX
@inproceedings{cao2025enhancing,
  author = {Han Cao and Lingwei Wei and Wei Zhou and Songlin Hu},
  title = {{Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=AzQjbGEHfm},
}
Figure