Biomedical Entity Linking via Contextualized Token Retrieval

Results

We store all predictions and gold identifiers in data.

To reproduce the results from the paper create a virtual env and install:

pip install pandas
pip install bioc

Performance on BELB:

python evaluate.py --target belb

Performance on partial BELB with refined test set:

python evaluate.py --target belb-refined-retrieve

Performance on partial BELB with LLM-reranking:

python evaluate.py --target belb-refined-rerank

Performance on BioRED (end-to-end document-level):

python evaluate.py --target biored

Train

Install torch) (hardware-dependent)

Install other deps:

# datasets
pip install git+https://github.com/sg-wbi/belb.git

pip install -e .

Run (see ./config/ for hyperparameters):

# for multi-GPU
# accelerate launch train.py 
python train.py \
    run=ncbi-disease \
    hps.train=ncbi-disease.ctd-diseases \
    hps.test=ncbi-disease.ctd-diseases \
    hps.kb_entity_based=false \
#   hps.qe=false \ # add [MASK]
#   hps.qe_train=false  \ train [MASK]