Addressing Treatment-Relevance in Biomedical Relation Extraction

Explore a conference talk on addressing treatment-relevance in biomedical relation extraction presented at DeclMed'23. Delve into the challenges of aligning upstream Natural Language Processing (NLP) corpus annotations with the needs of downstream treatment tasks (DTTs) like drug repurposing and precision medicine. Learn about the proposed multi-task training approach to flag treatment-relevant relations, tested on the BioRed corpus as part of the NIH LitCoin Challenge. Discover how a majority voting ensemble of BioBERT models was used to predict document-level relation types and a novel relation modifier for treatment relevance. Gain insights into the team's performance, achieving a top-ranking F1 score of 0.49 on the testing set, with the highest individual model accuracy of 88.81% for novelty and 87.74% for relation finding.

[DeclMed'23] Addressing Treatment-Relevance in Biomedical Relation Extraction