GraphCheck: Breaking Long-Term Text Barriers with Extracted Knowledge Graph-Powered Fact-Checking

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Large language models (LLMs) frequently produce subtle factual errors, particularly in extensive texts.
• These inaccuracies pose significant risks in specialized fields like medicine.
• Current fact-checking methods struggle with complex, multi-hop reasoning in long documents and are computationally expensive.

Purpose of the Study:

• To introduce GraphCheck, a novel fact-checking framework designed to improve the accuracy of LLMs.
• To address the limitations of existing fact-checking methods, including their difficulty with multi-hop reasoning and high computational demands.

Main Methods:

• GraphCheck extracts knowledge graphs to enrich text representations.
• Graph Neural Networks process these graphs as soft prompts for LLMs, integrating structured knowledge.
• The framework employs graph-based reasoning to capture intricate reasoning chains.

Main Results:

• GraphCheck demonstrated up to a 7.1% overall improvement across seven benchmarks in general and medical domains.
• The framework effectively captures multi-hop reasoning chains often missed by other methods.
• GraphCheck achieved performance comparable to state-of-the-art LLMs with significantly fewer parameters.

Conclusions:

• GraphCheck offers a precise and efficient solution for LLM fact-checking.
• The framework significantly enhances factual accuracy, especially in domains sensitive to errors.
• GraphCheck presents a computationally efficient alternative to existing specialized fact-checking models.