LongBEL links biomedical entities across whole documents, not sentence by sentence
A new arXiv paper proposes a generative entity-linking framework that keeps a memory of earlier predictions to keep concept labels consistent within a document.
In plain words
Biomedical entity linking means matching phrases in text, like a drug or disease name, to entries in reference databases such as UMLS or SNOMED CT. The LongBEL system reads an entire document and remembers its earlier choices, so the same concept gets the same label even when worded differently. Its authors report better results than sentence-by-sentence methods on five benchmarks.
What was announced
The paper's authors say most existing systems link each mention independently, using only the mention or its surrounding sentence, which ignores dependencies between mentions in the same document and can produce inconsistent predictions when one concept appears under different surface forms. LongBEL addresses this by combining full-document context with a memory of previous predictions. To keep that memory reliable, the model is trained on cross-validated predictions rather than gold labels, which the authors say reduces the mismatch between training and inference and limits cascading errors.
In experiments on five biomedical benchmarks spanning English, French, and Spanish, the authors report that LongBEL improves over sentence-level generative baselines, with the largest gains on datasets where concepts recur often within a document. An ensemble combining local, global, and memory-based variants achieved the best results across all benchmarks. Their analysis suggests the main benefit is document-level consistency for recurring concepts rather than better disambiguation of isolated mentions.
What it does not show
- Evidence is limited to the paper's abstract; no full-text methods, per-benchmark numbers, or external evaluation are available, and the reported results are the authors' own and not independently confirmed.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.