AIObioEnts: All-in-one biomedical entities
Biomedical named-entity recognition following the all-in-one NER (AIONER) scheme introduced by Luo et al.. This is a straightforward Hugging-Face-compatible implementation without using a decoding head for ease of integration with other pipelines.
For full details, see the main GitHub repository
Core biomedical entities
We have followed the original original AIONER training pipeline based on the BioRED dataset along with additional BioRED-compatible datasets:
- Gene: GNormPlus, NLM-Gene, DrugProt
- Disease: BC5CDR, NCBI Disease
- Chemical: BC5CDR, NLM-Chem, DrugProt
- Species: Species-800, Linnaeus
- Variant: tmVar
- Cell line: BioID
using three pre-trained language models as a base. This model corresponds to the implementation based on BiomedBERT-base pre-trained on both abstracts from PubMed and full-texts articles from PubMedCentral
F1 scores
The F1 scores of the current implementation on the BioRED test set are shown below:
| BiomedBERT-base abstract+fulltext | |
|---|---|
| Cell line | 96.91 |
| Chemical | 92.02 |
| Disease | 88.64 |
| Gene | 94.41 |
| Species | 97.59 |
| Variant | 89.58 |
| Overall | 92.44 |
Usage
The model can be directly used from HuggingFace in a NER pipeline. However, we note that:
- The model was trained on sentence-level data, and it works best when the input is split
- Each sentence to tag must be surrounded by the flag corresponding to the entity type one wishes to identify (any of the 6 individual entities or "ALL"), as in:
<entity_type>sentence</entity_type> - Since additional
'O'labels are used in the AIONER scheme, the outputs should be postprocessed before aggregating the tags
We provide helper functions to tag individual texts in the main repository
from tagging_fn import process_one_text
from transformers import pipeline
pipe = pipeline('ner', model='SIRIS-Lab/AIObioEnts-core-pubmedbert-full', aggregation_strategy='none', device=0)
process_one_text(text_to_tag, pipeline=pipe, entity_type='ALL')