Integrating large language models with knowledge graphs for the purpose of scientific paper reviewing

Using AI in peer review aims to enhance efficiency and accuracy, yet grapples with challenges in semantic understanding and consistency. Balancing automation with human input to achieve human-level review quality remains challenging, given the complexities in defining a “good” review and collecting annotated data. While AI-generated reviews may cover more paper aspects, they often lack constructiveness and factuality compared to human reviews. This project seeks to augment the generation of factual review statements by integrating structural information from research papers and employing techniques like argument mining to extract and analyze the paper’s reasoning and evidence for summarization. This objective is realized through the integration of knowledge graphs, incorporating external information to refine the assessment of a paper’s impact and contribution to the research field.