flyingmothman/bionlp

The models and framework used in the BioNLP 2023 paper titled "Comparing and combining some popular NER approaches on Biomedical tasks" can be found here !

Feel Free to create an Issue and chat with me on the Issue -- I will reply in at most 48 hours.

This repo contains all the code(preprocessing, postprocessing, modeling, training) used in the BioNLP paper "Comparing and combining some popular NER approaches on Biomedical tasks".
The code for all models is in ./models.
Utility code is in ./utils.
Code for preprocessing datasets (getting data ready for training) can be found in ./preprocessing
Configurations for all experiments, models, and datasets are in ./configs.

Table of Contents

• Requirements
• Preprocessed Data
• Training
  • Training the three models
  • Training the Meta model
• Preprocessing
  • Preprocessing data for NER models
  • Preprocessing data for Meta

Requirements

Works with Python version 3.10 and above ; We use an Anaconda python 3.10 environment. You can create an anaconda python 3.10 environment called my_env with:

conda create --name my_env python=3.10

• allennlp : Install with:

# On Debian and Ubuntu
sudo apt-get -y install pkg-config
sudo apt-get -y install build-essential

pip install allennlp

All remaining dependencies can be installed by running ./setup_env.sh, which simply includes:

echo "Install Requirements"
pip install spacy
pip install pudb
pip install colorama
pip install gatenlp
pip install pandas
pip install transformers
pip install flair
pip install benepar
pip install ipython
pip install overrides

echo "Installing Pytorch"
pip install torch torchvision torchaudio --upgrade

Preprocessed data

Run script ./download_preprocessed_data.sh to download all preprocessed training data into ./preprocessed_data.

Training

train.py is the main script for running experiments -- experiments involve training and evaluating models. Running python train.py (no arguments) will start a test-run, which will train the models on a very small subset of the data and then evaluate them on the validation set. The models' performance results are stored in ./training_results/performance.

Training and evaluating SEQ, SpanPred, and SeqCRF models.

To train SEQ, SpanPred, and SeqCRF models on SocialDisNER and evaluate on test data. Run:

python train.py --production --test --experiment=social_dis_ner_experiment

--production: indicates that all the training data should be used.
--test: indicates that the test data should also be used to evaluate the models(other than the validation data).
The details of the experiment can be found in ./configs/experiment_configs/social_dis_ner_experiment.py


Similarly, for training the three models on GENIA, run :

python train.py --production --test --experiment=genia

For NCBI-Diease, run:

python train.py --production --test --experiment=ncbi

For LivingNER, run:

python train.py --production --test --experiment=living_ner

Training and evaluating the Meta model.

For SocialDisNER, run:

python train.py --production --test --experiment=social_dis_ner_meta

For GENIA, run:

python train.py --production --test --experiment=genia_meta

For NCBI Disease, run:

python train.py --production --test --experiment=ncbi_disease_meta

For LivingNER, run:

python train.py --production --test --experiment=living_ner_meta

Preprocessing

Preprocessing involves preparing training data for models.

Preprocessing data for NER models

Use the available prepreocessing configurations for NER models in all_preprocessing_configs.py to prepare training data for the 3 NER models.
For example, ./preprocess.py shows how to preprocess the Genia dataset.

Preprocessing data for Meta

Use the Meta preprocessing configurations in all_preprocessing_configs.py to prepare data for Meta. For example, ./preprocess_meta.py uses the predictions made by SEQ and SpanPred(generated by this framework) for GENIA, which are located in ./raw_data_for_meta/genia, to prepare data for Meta.