Prompting large language models to extract chemical‒disease relation precisely and comprehensively at the document level

Currently, the intricate relationships between chemicals and diseases have increasingly been revealed and documented in many academic studies. However, constrained by the scarcity of high-quality annotated data, document-level relation extraction techniques based on deep learning are plagued by ambiguity and one-sidedness, severely impeding further automatic in-depth analysis of these relationships. In this study, we harness the advanced reading comprehension capabilities and extensive world knowledge of LLMs to innovatively construct precise and comprehensive zero-shot prompt workflows for extracting chemical‒disease relationships on the basis of the patterns of LLM-based relation extraction, the attributes of chemical‒disease relationships, and the linguistic features of biomedical literature. Evaluations of the enhanced chemical‒disease relation (CDR) datasets demonstrate that the F1 scores for the workflows in precise extraction can reach 87% and can reach 73% for comprehensive extraction. Additionally, we investigate the pivotal factors that influence LLMs' ability to extract complex relations, offering crucial insights for relation extraction across diverse fields.