biomedical-slot-filling

Biomedical Slot filling with Dense Passage Retrieval Accompanying code for our paper [1]. This code is only meant to be used so as to replicate our results.

Data

The data, indices and BioSF slot filling dataset can be found here.

The BioSF dataset has been built using the following publicly available RE datasets:

• Biocreative ChemProt
• Biocreative DDI-2013
• Biocreative CDR

Models

Our finetuned models are available through huggingface:

• https://huggingface.co/healx/biomedical-dpr-qry-encoder
• https://huggingface.co/healx/biomedical-dpr-ctx-encoder
• https://huggingface.co/healx/biomedical-slot-filling-reader-base
• https://huggingface.co/healx/biomedical-slot-filling-reader-large

Replicating results

Although we provide also the code to train a DPR or a reader model, we here focus only in replicating our evaluation experiments. Assuming that we 've created a virtual environment and install requirements, we can then run either

PYTHONPATH=. python biomedical_slot_filling/scripts/train_eval_reader --eval

or

PYTHONPATH=. python biomedical_slot_filling/scripts/evaluate_retrieval

The above scripts will download the neccessary files from GDrive, the finetuned models from HF and perform the relevant evaluation.

To train a slot filling reader, we can run:

PYTHONPATH=. python biomedical_slot_filling/scripts/train_eval_reader.py --model-name-or-path dmis-lab/biobert-base-cased-v1.2 --train

Note: A large portion of our code is adapted from the following two repositories:

• https://github.com/facebookresearch/DPR
• https://github.com/IBM/kgi-slot-filling

Cite our work

[1] Slot Filling for Biomedical Information Extraction