Addressing Treatment-Relevance in Biomedical Relation Extraction
summary
Fine-tuned encoder models on the BioRed corpus (NIH LitCoin) use entity-marked, modified abstracts and the CLS embedding in a multitask setup to predict four relation types plus a treatment-relevance modifier; a majority-voting BioBERT ensemble places among top teams but yields F1 0.49 while the best single model reports 88.81% novelty and 87.74% relation accuracy, highlighting severe class imbalance and noisy KGs; a proposed alternative uses zero-shot LLM prompting to evaluate relation context and treatment applicability.
AI-written summary · about
chapters (8)
- 0:00Treatment relevance in biomedical extraction
- 1:11Limitations of multitask training
- 3:38Modifying abstracts for prediction
- 6:26Performance challenges in extraction
- 8:10Content completeness issues
- 10:02LLMs for context hypothesis
- 13:01LLM method for treatment relevance
- 14:25Audience questions and discussion
related talks
01h 30mRelation extraction from texts
20mBERT Model for Real World Healthcare Data
26mNatural Language Processing for Healthcare: Current Trends...
46mBeyond BLEU & ROUGE — Evaluating LLMs and AI Systems
34mExtreme Multilabel Classification in the Biomedical NLP Domain
26mExtracting Medical Attributes from Clinical Trials
58mRelation Extraction from Online PH News Source
35mBuilding Named Entity Recognition Models Efficiently Using NERDS