Holmes: Automated Fact Check with Large Language Models

Abstract: The rise of Internet connectivity has accelerated the spread of disinformation, threatening societal trust, decision-making, and national security. Disinformation has evolved from simple text to complex multimodal forms combining images and text, challenging existing detection methods. Traditional deep learning models struggle to capture the complexity of multimodal disinformation. Inspired by advances in AI, this study explores using LLMs for automated disinformation detection. The empirical study shows that (1) LLMs alone cannot reliably assess the truthfulness of claims; (2) providing relevant evidence significantly improves their performance; (3) however, LLMs cannot autonomously search for accurate evidence. To address this, we propose Holmes, an end-to-end framework featuring a novel evidence retrieval method that assists LLMs in collecting high-quality evidence. Our approach uses (1) LLM-powered summarization to extract key information from open sources and (2) a new algorithm and metrics to evaluate evidence quality. Holmes enables LLMs to verify claims and generate justifications effectively. Experiments show Holmes achieves 88.3% accuracy on two open-source datasets and 90.2% in real-time verification tasks. Notably, our improved evidence retrieval boosts fact-checking accuracy by 30.8% over existing methods

Automated Fact-Checking with LLMs: Holmes

The paper "Holmes: Automated Fact Check with LLMs" presents a sophisticated framework designed to address the growing complexity of disinformation, especially in multimodal formats encompassing text and images. In an age where misinformation threatens societal trust and decision-making, this study leverages advances in AI, particularly LLMs, for disinformation detection and verification.

Key Findings and Contributions

The paper begins with an empirical assessment of LLMs' capabilities in verifying multimodal disinformation. The study identifies several limitations in existing LLM-based models: 1) they struggle to assess the truthfulness of claims without additional context, 2) they cannot autonomously retrieve the necessary evidence, and 3) when provided with sufficient evidence, LLMs demonstrate improved accuracy in verification tasks.

To address these deficiencies, the paper introduces Holmes, an end-to-end framework designed to enhance the fact-checking process using LLMs. Holmes integrates a novel evidence retrieval methodology, emphasizing two main innovations:

1. Summarization Leveraging LLMs: Holmes utilizes LLMs' summarization capabilities to extract key insights from open-source information. The approach focuses on identifying fundamental elements of narrative construction such as the primary actors, events, locations, and implications.
2. Evidence Quality Evaluation: A dynamic algorithm and metrics are proposed to evaluate and ensure the quality of extracted evidence. This ensures that only reliable and relevant information is utilized during the verification process.

These innovations are crucial for advancing disinformation detection, as they shift the focus from merely identifying misleading text to comprehensively analyzing multimodal narratives. Through rigorous experimentation, Holmes demonstrated high accuracy rates: 88.3% on open-source datasets and 90.2% on the real-time verification task, significantly enhancing fact-checking accuracy over existing baseline methods.

Implications and Future Directions

Holmes showcases the potential of integrating LLMs into automated fact-checking systems, offering promising implications for both practical applications and theoretical advancements. Practically, Holmes can significantly reduce the time resource human fact-checkers traditionally require, as it offers a scalable alternative to manual verification processes. Theoretically, this research highlights the robustness of LLMs in summarization and evidence evaluation tasks, proposing new avenues for exploration in LLM-driven analysis tasks beyond conventional applications.

The paper stimulates discussion towards enhancing model capabilities for autonomous evidence retrieval—critical for the autonomous detection of disinformation. Future research might explore expanding model capacities to accommodate complex multimodal inputs beyond text and images, including audio and video, which are increasingly prevalent in digital communications.

Conclusion

In summary, "Holmes: Automated Fact Check with LLMs" makes valuable contributions to disinformation detection literature by establishing a robust framework leveraging LLMs for automated fact-checking. By improving evidence retrieval and evaluation processes, the framework addresses current limitations in LLM applications, offering substantial gains in accuracy and efficiency for multimodal misinformation detection. The study lays the foundation for future advancements in AI-driven fact-checking, emphasizing the importance of high-quality evidence and reasoning in content verification tasks.