MedSETNER

Research question

How can dataset names be identified reliably in medical research papers?

My work

For my master’s research at Novosibirsk State University, I led corpus development and annotation across 691 articles containing 9,376 dataset-name mentions. Inter-annotator agreement was Fleiss’ κ = 0.85.

I compared rules, CRF, SciBERT, ModernBERT, BiomedBERT, and GLiNER2 using article-level splits and a separate 151-article out-of-distribution test set. BiomedBERT achieved mean exact-span F1 of 60.7% on same-pool data and 60.8% on the out-of-distribution set.

The paper was accepted at AIST 2026, CCIS track. See the publication details.