Skip to content

Latest commit

 

History

History
80 lines (68 loc) · 2.82 KB

README.md

File metadata and controls

80 lines (68 loc) · 2.82 KB

Hyperparameters

The table shows the used hyper-parameters for fine-tuning BERT models.

Dataset Epochs Batch Size Maximum Length
NCBI-Disease 10 8 320
BC5CDR-Disease 10 8 320
BioRED-Disease 10 8 384
BC5CDR-Chemical 10 16 256
BioRED-Chem 10 32 128

Fine-tuned HF 🤗

The table provides links to the fine-tuned models as an output of our experimental work. As I mentioned previously, I couldn't change my username on Hugging Face, therefore I can't share the links because they contain one of the co-authors' names. However, they are available publicly on HF.

Dataset Orignal WeLT
NCBI-Disease model_name_or_path [model_name_or_path] ()
BC5CDR-Disease model_name_or_path model_name_or_path
BioRED-Disease model_name_or_path model_name_or_path
BC5CDR-Chemical model_name_or_path model_name_or_path
BioRED-Chem model_name_or_path model_name_or_path

Usage example for WeLT finetuning

This is an example of fine-tuning NCBI-Disease over BioBERT using WeLT weight scheme

cd named-entity-recognition
./preprocess.sh

export SAVE_DIR=./output
export DATA_DIR=../datasets/NER

export MAX_LENGTH=320
export BATCH_SIZE=8
export NUM_EPOCHS=10
export SAVE_STEPS=1000
export ENTITY=NCBI-disease
export SEED=1

python run_weight_scheme.py \
    --data_dir ${DATA_DIR}/${ENTITY}/ \
    --labels ${DATA_DIR}/${ENTITY}/labels.txt \
    --model_name_or_path dmis-lab/biobert-v1.1 \
   --output_dir ${ENTITY}-WELT-${MAX_LENGTH}-BioBERT \
    --max_seq_length ${MAX_LENGTH} \
    --num_train_epochs ${NUM_EPOCHS} \
    --weight_scheme WELT \
    --per_device_train_batch_size ${BATCH_SIZE} \
    --save_steps ${SAVE_STEPS} \
    --seed ${SEED} \
    --do_train \
    --do_eval \
    --do_predict \
    --overwrite_output_dir

Usage example for predicting disease entities using WeLT

This is an example of fine-tuning NCBI-Disease over BioBERT using an WELT fine-tuned model on HF

cd named-entity-recognition

python3 ner.py \
--xmlfilepath path/unannotatedxmls/NCBItestset_corpus_noannotations.xml \
--model_name_or_path [path] \
--NERType Disease \
--outputfilepath Re-scaling-class-distribution-for-fine-tuning-BERT-based-models/predictedpath/NCBI-WELT-BioBERT-example.xml

Usage example for strict evaluation of NCBI-Disease predicted file using WeLT

Assuming that download.sh has been already executed. Adjust the paths to the corresponding working directories.

cd BC7T2-evaluation_v3
python3 evaluate.py --reference_path path/referencepath/NCBItestset_corpus.xml \
 --prediction_path path/predictedpath/NCBI-WELT-BioBERT.xml \
 --evaluation_type span \
 --evaluation_method strict \
 --annotation_type Disease