MicrobioRel: A Set of Datasets for Microbiome Relation Extraction

Biomedical knowledge curation relies on a variety of Natural Language Processing tasks, including biomedical entity recognition and document-level relation extraction. With the growing size and capabilities of Language Models, effectively deploying them in specific and specialised domains remains a persistent challenge, highlighting the need for high-quality, domain-adapted datasets. In this work, we present MicrobioRel, a corpus of two datasets to study the relations between biological entities in the gut microbiome. The first dataset, MicrobioRel-cur, is a document-level labelled corpus, corresponding to paragraphs from journal articles that were manually annotated with different types of relations between biomedical concepts in the gut microbiome domain. We describe its creation process, annotation guidelines, and key statistics. On this dataset, we evaluated different architectures for relation extraction and identify PubMedBERT as the most effective model for this task. We also created a second dataset, MicrobioRel-pred, by generating relation predictions on other journal articles using the fine-tuned PubMedBERT model. We demonstrate its potential to extract meaningful interactions. The MicrobioRel is a crucial resource for advancing tasks like automatic knowledge extraction in specialised domains such as the gut microbiome, facilitating hypothesis generation and supporting scientific discovery.