sapbert-ncbi-taxonomy-no-ab3p

sapbert-ncbi-taxonomy-no-ab3p

Biomedical Entity Mention Linking for UMLS. We use this model for species since NCBI Taxonomy is contained in UMLS:

• Model: cambridgeltl/SapBERT-from-PubMedBERT-fulltext
• Dictionary: NCBI Taxonomy (See FTP)

NOTE: This model variant does not perform abbreviation resolution via A3bP

Demo: How to use in Flair

Requires:

• Flair>=0.14.0 (pip install flair or pip install git+https://github.com/flairNLP/flair.git)

from flair.data import Sentence
from flair.models import Classifier, EntityMentionLinker
from flair.tokenization import SciSpacyTokenizer

sentence = Sentence(
    "The mutation in the ABCD1 gene causes X-linked adrenoleukodystrophy, "
    "a neurodegenerative disease, which is exacerbated by exposure to high "
    "levels of mercury in dolphin populations.",
    use_tokenizer=SciSpacyTokenizer()
)

# load hunflair to detect the entity mentions we want to link.
tagger = Classifier.load("hunflair-species")
tagger.predict(sentence)

# load the linker and dictionary
linker = EntityMentionLinker.load("species-linker-no-abbres")
linker.predict(sentence)

# print the results for each entity mention:
for span in sentence.get_spans(tagger.label_type):
    for link in span.get_labels(linker.label_type):
      print(f"{span.text} -> {link.value}")

As an alternative to downloading the already precomputed model (much storage). You can also build the model and compute the embeddings for the dataset using:

from flair.models.entity_mention_linking import BioSynEntityPreprocessor
linker = EntityMentionLinker.build("cambridgeltl/SapBERT-from-PubMedBERT-fulltext", dictionary_name_or_path="ncbi-taxonomy", entity_type="species", preprocessor=BioSynEntityPreprocessor(), hybrid_search=False)

This will reduce the download requirements, at the cost of computation. Note hybrid_search=False as SapBERT unlike BioSyn is trained only for dense retrieval.