CliniQG4QA

Introduction

This repository provides source code for CliniQG4QA: Generating Diverse Questions for Domain Adaptation of Clinical Question Answering accepted by Machine Learning for Health Workshop (ML4H) at NeurIPS 2020.

Please cite our paper if you use the codes from this repo and/or our released test set:

@article{cliniqg4qa2020,
      title={CliniQG4QA: Generating Diverse Questions for Domain Adaptation of Clinical Question Answering}, 
      author={Xiang Yue and Xinliang Frederick Zhang and Ziyu Yao and Simon Lin and Huan Sun},
      journal={arXiv preprint arXiv:2010.16021},
      year={2020}
}  

@misc{https://doi.org/10.13026/j0y6-bw05,
  doi = {10.13026/J0Y6-BW05},
  url = {https://physionet.org/content/mimic-iii-question-answer/1.0.0/},
  author = {Xiang Yue and Xinliang Frederick Zhang and Huan Sun},
  title = {Annotated Question-Answer Pairs for Clinical Notes in the MIMIC-III Database},
  publisher = {PhysioNet},
  year = {2021}
}

We are still refactoring and cleaning the codes, please stay tuned and check back later!

Set up

Run the following commands to clone the repository.

$ git clone https://github.com/sunlab-osu/CliniQG4QA.git

To run our codes, it requires Python 3.6 or higher (Python 3.6.8 preferred). It also requires installing PyTorch version 1.2.0 or higher and Tensorflow version 1.1 or higher.

Dataset Preparation

emrQA: Follow instructions listed on https://github.com/panushri25/emrQA to download raw data and convert them into emrQA dataset.

MIMIC-III: Follow instructions listed on https://mimic.physionet.org/gettingstarted/access/ to get the access to MIMIC-III dataset.

After you get both datasets, please place them under Data directory.

Note that We do not have the rights to include these two datasets in this repo.

Preprocessing

Run the following scripts to convert json files into plain texts (train/dev/test split ratio: 8:1:1):

cd Data
python json_cleaner.py
python text_generator.py
cd ../

Run the following scripts to get cleaned-up input files (and append lexical features if desired) and prepare necessary files for QPP module:

cd ./QG/Data/relation_5
python pre_pre_preprocessing.py --dataset xxx
python QW_statistics --dataset xxx
cd ../../../

Replace xxx with train/dev/test to do preprocessing for each dataset

QPP-enhanced Question Generation

QPP-enhanced QG training:

In order to train a QG model, for example QPP-enhanced NQG++ model, please follow the following steps:

cd ./QG/NQG++
python preprocessing.py
python train.py -ext test

In order to save efforts and speed up the training process, we combine the training of QG and QPP by adding up losses.

QPP-enhanced QG testing/inference:

In order to do inference on trained model, for example QPP-enhanced NQG++ model, please follow the following steps:

python eval.py -model XXX -out ../prediction/test.txt -beam 1 -alpha 1.0

XXX referes to the trained QG model. XXX checkpoint is saved under ./output directory

In order to evaluate the performance of generated corpus, please follow the following steps:

cd ../qgevalcap
For relevance: ./eval.py --out_file ./prediction/test.txt --src_file ../Data/relation_5/test.src.ref --tgt_file ../Data/relation_5/test.tgt (run with Python2)
For diversity: python Distinct.py -out ../prediction/test.txt
cd ../

Answer Evidence Extractor:

Follow the instuctions listed on https://github.com/huggingface/transformers/tree/master/examples/token-classification to extract raw answer evidences. Then, run the following script to polish the extracted answer evidences (i.e. our designed heuristic rules for post-processing):

cd ./AEE
sh transform.sh		
cd ../

Question Answering:

QA training:

DrQA: Follow the instrcutions listed on https://github.com/facebookresearch/DrQA/tree/master/scripts/reader --Training to do QA model training.

ClinicalBERT: Follow the instrcutions listed on https://github.com/google-research/bert (under SQuAD 1.1 section) to do QA model training.

In order to accommodate our Clinical Setting, please download the pretrained ClinicalBERT model (Details can be found here: https://github.com/EmilyAlsentzer/clinicalBERT).

QA testing:

DrQA: Follow the instrcutions listed on https://github.com/facebookresearch/DrQA/tree/master/scripts/reader --Predicting to do QA model testing.

ClinicalBERT: Follow the instrcutions listed on https://github.com/google-research/bert (under SQuAD 1.1 section) to do QA model testing.

Please run the follwowing block, adapted from offical eval script of SQuAD v1.1 (https://worksheets.codalab.org/rest/bundles/0x5d2ee15e3f2d4e34bb864df4955274e8/contents/blob/evaluate.py), to evaluate QA model performance.

cd ./QA
python ./evaluate-v1.1_human_generated.py $Dataset $Prediction	
python ./evaluate-v1.1_human_verified.py $Dataset $Prediction		
python ./evaluate-v1.1_overall.py $Dataset $Prediction		
cd ../