Extreme Multilabel Classification in the Biomedical NLP Domain
summary
A transformers‑based NLP model is built to tag biomedical grants with 29K MeSH tags, trained on a corpus of about 14 million publications, and it uses techniques that make extreme label sets tractable: sparse parameters, streaming data, negative sampling, binary‑tree pruning, threshold optimization and snapshot ensembles. A multi-label attention head per label scales to all 29K labels, reaches roughly 63% F1 for the full MeSH set, trains in ~4 days on eight GPUs and is deployed for batch grant tagging.
AI-written summary · about
chapters (4)
related talks
23mLeveraging Generative AI and Transformers in Healthcare
38mLeverage Reviews Data for Multi Label Topics Classification
33mLessons Learned Applying Large Language Models in Healthcare
26mExtracting Medical Attributes from Clinical Trials
26mNatural Language Processing for Healthcare: Current Trends...
20mDiagnostic Medical application of ML
33mScalably Internationalizing Millions of Latent Semantic Labels
13mMEDebiaser: A Human-AI Feedback System for Mitigating Bias in Multi-label Medical Image Classific...