PharmaCoNER

PharmaCoNER

Linguistically Informed Relation Extraction and Neural Architectures for Nested Named Entity Recognition in BioNLP-OST 2019

Setting-up

Install requirements.

pip install -r requirements.txt

Data

Install docker, for Ubuntu.

sudo curl -L "https://github.com/docker/compose/releases/download/1.24.1/docker-compose-$(uname -s)-$(uname -m)" -o /usr/local/bin/docker-compose
sudo chmod +x /usr/local/bin/docker-compose
sudo apt install docker.io

Download MedTagger and run the docker daemon.

sudo service docker start
cd PATH_TO/MedTagger
sudo docker-compose up

You can skip the following steps if you are using the pre-processed version of dataset, included in this repository.

Download train, development and test set from PharmaCoNER's webpage. Extract the content of zip files in PharmaCoNER/dataset/raw/ directory (e.g. path of S0004-06142005000500011-1.txt relative to this project would be PharmaCoNER/dataset/raw/train/S0004-06142005000500011-1.txt).

Prepare the data by running this command.

python data.py

Embeddings

Download Spanish Medical embeddings from here and store them under resources/spanish-embs/

Train

python train.py --train dataset/data-strategy=1/train.txt --dev dataset/data-strategy=1/dev.txt --word_lstm_dim 200 --word_dim 100 --pos_dim 50 --ortho_dim 50 --pre_emb resources/spanish-embs/fasttext-100d-train_dev_test.vec --pre_emb_1 resources/spanish-embs/fasttext-100d-train_dev.vec --ranking_loss 1

Evaluation

python evaluate.py -model models/data-strategy=1/best_model/ -data dataset/raw/test/ -span dataset/test-token_span.pkl

python ensemble.py -model1 models/data-strategy=1/best_model/ -model2 models/data-strategy=2/best_model/ -model3 models/data-strategy=3/best_model/ -data dataset/raw/test/ -span dataset/test-token_span.pkl

Prediction

python predict.py -model models/data-strategy=1/best_model/ -text 'consideró el diagnostico de metástasis de carcinoma de mama y el Servicio de Oncología Médica inició tratamiento con epirubicina y ciclofosfamida.'

consideró	 O
el	         O
diagnostico	 O
de	         O
metástasis	 O
de	         O
carcinoma	 O
de	         O
mama	     O
y	         O
el	         O
Servicio	 O
de	         O
Oncología	 O
Médica	     O
inició	     O
tratamiento	 O
con	         O
epirubicina	 B-NORMALIZABLES
y	         O
ciclofosfamida	B-NORMALIZABLES
.	         O

Results

Test Set

Model	P	R	F1
---------------------------	----------	------	------
Data Split 1	0.8894	0.8165	0.8514
Data Split 2	0.9046	0.8186	0.8595
Data Split 3	0.8905	0.8175	0.8525
Ensemble (DS1 + DS2 + DS3)	0.9060	0.8273	0.8649

Acknowledgements

LSTM-CRF implementation is based on https://github.com/glample/tagger.

Citation

If you found this codebase useful than please consider citing our work.

@inproceedings{DBLP:conf/bionlp/GuptaYS19,
  author    = {Usama Yaseen and
               Pankaj Gupta and
               Hinrich Sch{\"{u}}tze},
  editor    = {Jin{-}Dong Kim and
               Claire N{\'{e}}dellec and
               Robert Bossy and
               Louise Del{\'{e}}ger},
  title     = {Linguistically Informed Relation Extraction and Neural Architectures
               for Nested Named Entity Recognition in BioNLP-OST 2019},
  booktitle = {Proceedings of The 5th Workshop on BioNLP Open Shared Tasks, BioNLP-OST@EMNLP-IJNCLP
               2019, Hong Kong, China, November 4, 2019},
  pages     = {132--142},
  publisher = {Association for Computational Linguistics},
  year      = {2019}
}