laituan245 / rescnn_bioel

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Biomedical Entity Linking

For reproducing our results on COMETA, please use the branch cometa. Basically, we turn on this mode called hard_negatives_training when experimenting with COMETA.

This repo provides the code for the paper BERT might be Overkill: A Tiny but Effective Biomedical Entity Linker based on Residual Convolutional Neural Networks (EMNLP 2021 Findings).

Download the pretrained embedding layer from this link. And set this line to the path of the downloaded file.

Basic running instructions

pip install -r requirements.txt
python cg_trainer.py --dataset bc5cdr-chemical

Please refer to the file constants.py for the list of all supported datasets. Note that for COMETA, you need to download the dataset from https://github.com/cambridgeltl/cometa.

Note that for ncbi-disease, bc5cdr-disease, and bc5cdr-chemical, we follow the protocol of BioSyn. We use development (dev) set to search the hyperparameters, and train on traindev (train+dev) set to report the final performance.

We are cleaning the codebase and we will add more running instructions soon.

About


Languages

Language:Python 99.7%Language:Shell 0.3%