Patients don't write "fatigue." They write "wiping me out."
That gap breaks medical concept normalization and most tools need manual annotation or NER to bridge it.
Our preprint benchmarks 27 language models on closing it, end to end. 🧵
https://t.co/QwRzvNK0E5
Then 12 instruction-tuned LLMs as upstream correctors.
SapBERT led the encoders at 63.8% Top-5 on SNOMED CT. Adding Qwen 2 Instruct pushed it to 69.4%, matching models 10× its size at a fraction of the compute.
4. Then 12 instruction-tuned LLMs as upstream correctors.
SapBERT led the encoders at 63.8% Top-5 on SNOMED CT. Adding Qwen 2 Instruct pushed it to 69.4%, matching models 10× its size at a fraction of the compute.
NATUS VINCERE JUST BEAT VITALITY IN A COUNTER-STRIKE MATCH‼️‼️‼️ I REPEAT NATUS VINCERE JUST BEAT VITALITY IN A COUNTER-STRIKE MATCH 😭😭❤️ FINALLLLYYYYYYYY