A Distant Supervision Corpus for Extracting Biomedical Relationships Between Chemicals, Diseases and Genes

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1073–1082 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1073 A Distant Supervision Corpus for Extracting Biomedical Relationships Between Chemicals, Diseases and Genes Dongxu Zhang*,1, Sunil Mohan*,2, Michaela Torkar2, Andrew McCallum1 1 University of Massachusetts, Amherst, Massachusetts, USA 2 Chan Zuckerberg Initiative, Redwood City, California, USA {dongxuzhang, mccallum}@cs.umass.edu smohan@chanzuckerberg.com, michaela.torkar@contractor.chanzuckerberg.com Abstract We introduce ChemDisGene, a new dataset for training and evaluating multi-class multi-label document-level biomedical relation extraction models. Our dataset contains 80k biomedical research abstracts labeled with mentions of chemicals, diseases, and genes, portions of which human experts labeled with 18 types of biomedical relationships between these entities (intended for evaluation), and the remainder of which (intended for training) has been distantly labeled via the CTD database with approximately 78% accuracy. In comparison to similar preexisting datasets, ours is both substantially larger and cleaner; it also includes annotations linking mentions to their entities. We also provide three baseline deep neural network relation extraction models trained and evaluated on our new dataset. Keywords: Corpus, Information Extraction, Linked Data, Weakly-supervised Learning, Relation Extraction 1. Introduction Biomedical researchers have used systems of exper- imentally confirmed interactions between chemicals, diseases, genes/proteins and other entities, for under- standing disease mechanisms for diagnosis, e.g. Lee et al. (2019), for drug repurposing (Morselli Gysi et al., 2021), and even for understanding the health hazards associated with spaceflight (Nelson et al., 2021). These knowledge graphs (KGs) are often built by integrating manually curated databases like CTD1 and DrugBank2, who use domain experts to extract observed interac- tions from research publications and other sources. While the information in these databases is high in pre- cision, with the growing publication rate their recall is low (Baumgartner et al., 2007). To improve cover- age, researchers have resorted to automated mining of biomedical interactions from research texts, to supple- ment their KG (Himmelstein et al., 2017), or even to build the entire KG, e.g. (Crichton et al., 2020). The bioinformatics community recognized that ma- chine learning Relation Extraction (RE) models could help the manual curation task, and the BioCreative workshops introduced the first shared task and man- ually labeled ‘gold standard’ dataset for training and evaluating models for extracting protein-protein inter- actions from full text articles in 2006 (Krallinger et al., 2006). Several such labeled corpora have followed, primarily focusing on extracting relationships from ab- stracts. However, labeling of relationships requires do- main experts and is slow and expensive. Consequently, most labeled corpora are small, and focus on a small number of entity types and relationships. * Equal contribution 1CTD: Comparative Toxicogenomics Database 2https://go.drugbank.com In this paper, we introduce ChemDisGene3, a new dataset of biomedical research abstracts labeled with pairwise interactions between Chemicals, Diseases and Genes/Gene-products. It contains two sub-corpora: • A large corpus of ∼ 80k abstracts with distant la- beling of 14 relation types. This corpus is auto- matically derived from CTD (Davis et al., 2020), thus allowing for a larger size …

Content truncated to 500 words