SNOMED CT Entity Linking Benchmark

A benchmark for linking text in medical notes to entities in SNOMED Clinical Taxonomy. #health

Benchmark
Open
134 joined

Model: Residual Linker
Published

TeamSTLabs

Abstract

We present a detect and link model that builds on the KIRI super dictionary as the base. One shared MPNet encoder over section-aware tiles feeds three heads: a boundary head proposing candidate extents, a null head giving a per-span linkability score, and a retrieval head applying a learned map to produce a bi-encoder cosine against a frozen description index max-pooled to 230K concepts. On spans the dictionary gets right, the model independently recovers the same concept 92.8% of the time given the correct extent; replacing the dictionary nonetheless scores worse than deferring to it, so the model is confined to concepts the dictionary predicts nowhere, at extents the dictionary did not claim.

Submissions (3)

Submissions
Name Macro char IoUSupport-weighted char IoU
Residual Linker
22h 54min ago
0.4631 0.6263
Residual Linker
1d 17h ago
0.4636 0.6256
Residual Linker
2d 15h ago
0.4605 0.6162