Using Knowledge Graphs and Agentic LLMs for Factuality Text Assessment and Improvement

This paper addresses the challenge of assessing and enhancing the factual accuracy of texts generated by large language models (LLMs). Existing methods often rely on self-reflection or external knowledge sources, validating statements individually and rigidly, thus missing a holistic view. We propose a novel approach utilizing a comprehensive knowledge graph (KG), such as Wikidata, to assess and improve the factuality of generated texts. Our method dynamically retrieves and integrates relevant facts during the assessment process, providing a more interconnected and accurate evaluation. Integrating KG with LLM capabilities enhances the overall factual integrity, leading to more reliable AI-generated content. Our results demonstrate improvements in factual accuracy, highlighting the effectiveness of our approach.