EntityLinker returns different CUIs with every run

Hi,

I’m running into a weird issue with NER extraction.

Different runs on the same document gives me identical entities, but different CUIs. I noticed the EntityLinker guide says it uses an approximate nearest neighbor search while linking entities to the KB.

This is the config I’m using:

linker = EntityLinker(resolve_abbreviations=True, name="umls", max_entities_per_mention=1, k=1)

Configuration:

>>> spacy.__version__
'2.3.4'
>>> scispacy.__version__
'0.3.0'

Question: How do I make the results deterministic?

Thanks!

[mnhcorp]

Yup, that works - thanks again!

[dakinggg]

ahh, i see. As a work around, i think you can do something like this, after loading the entity linker on just the core model

linker = model['core'].get_pipe('scispacy_linker')
doc = linker(model['bc5cdr'](text))

I’m not sure if there is a way to add a pipe without creating a copy of it in spacy3, will have to look into it.