The project aims to develop methods for automatically verifying and correcting facts in biomedical knowledge graphs using scientific literature as evidence. Biomedical knowledge graphs, widely used in applications such as drug discovery and clinical decision support, are often constructed automatically and may therefore contain incorrect, outdated, or inconsistent information. The objective is to assess the validity of knowledge graph triples by comparing them with information extracted from scientific publications, and to detect unsupported or conflicting claims.

The project explores a unified framework combining information extraction from text, structured knowledge graph representations, and reasoning capabilities of large language models. It investigates different strategies, including textual entailment, structured comparison between extracted relations and graph statements, and retrieval-augmented LLM-based reasoning, while also addressing the challenge of proposing corrections supported by evidence. More broadly, it seeks to evaluate whether structured representations improve the reliability, interpretability, and explainability of fact verification compared to purely generative approaches. Ultimately, the project aims to contribute to the development of reliable, evidence-grounded systems for managing and validating biomedical knowledge at scale.