Skip to content

Latest commit

 

History

History
209 lines (165 loc) · 7.52 KB

README.md

File metadata and controls

209 lines (165 loc) · 7.52 KB
Raw

Patient-friendly Clinical Notes: Towards a new Text Simplification Dataset

This repository provides the code for following paper:

Jan Trienes, Jörg Schlötterer, Hans-Ulrich Schildhaus, and Christin Seifert. 2022. Patient-friendly Clinical Notes: Towards a new Text Simplification Dataset. In Proceedings of the Workshop on Text Simplification, Accessibility, and Readability (TSAR-2022), pages 19–27, Abu Dhabi, United Arab Emirates (Virtual). Association for Computational Linguistics.

Sharing of the dataset is currently underway. For the time being, the code in the repository can be run with the paragraph-level simplification dataset by Devaraj et al., (NAACL 2021).

Contents:

  1. Computational Environment
  2. Data Format
  3. Corpus Statistics
  4. Model Training and Evaluation
  5. Experiments on Cochrane Data (English, Medical)
  6. Development
  7. Citation
  8. Contact

Computational Environment

conda update -f environment.yml
conda activate simple-patho

pip install -r requirements-dev.txt
pip install -e .

# if on GPU, install CUDA toolkit
conda install cudatoolkit=11.6.0

Data Format

Data is expected to be split into train/val/test sets and to have one sample per line in source/target.

tree data/raw/$YOUR_DATASET
├── test.source
├── test.target
├── train.source
├── train.target
├── val.source
└── val.target

For training seq2seq models, convert the data into jsonlines format.

python -m scripts.convert_data \
    --raw_path data/raw/$YOUR_DATASET \
    --out_path data/processed/$YOUR_DATASET

# Should give
tree data/processed/$YOUR_DATASET
├── test.json
├── train.json
└── val.json

Corpus Statistics

python -m simplepatho.corpus_statistics \
    --data_path data/processed/$YOUR_DATASET \
    --language [english|german] \
    --spacy_model [en_core_web_sm|de_core_news_sm]

Model Training and Evaluation

In the paper, we run experiments with three models: Bert2Bert, Bert2Share, and mBART. Please refer to the files in scripts/simplepatho for training/inference parameters and how to use your own dataset.

# Training
sh scripts/simplepatho/{bert2bert,bert2share,mbart}/train.sh
# Inference
sh scripts/simplepatho/{bert2bert,bert2share,mbart}/predict.sh

Model predictions can be evaluated with below command.

# --sys_paths takes multiple files, each being the predictions of one system.
# Below, the first run corresponds to the identity baseline.
python -m simplepatho.evaluate \
    --orig_path data/processed/$YOUR_DATASET/test.source \
    --refs_path data/processed/$YOUR_DATASET/test.target \
    --sys_paths \
        data/processed/$YOUR_DATASET/test.source \
        output/$YOUR_DATASET/bert2bert/predictions_test.txt \
        output/$YOUR_DATASET/bert2bert-shared/predictions_test.txt \
        output/$YOUR_DATASET/mbart/predictions_test.txt \
    --compact \
    --calculate_metrics R-1 R-2 R-L BLEU SARI "|y|_w" "CMP_w" "CMP_s" "Nov. (n=1)" "Nov. (n=2)" \
    --n_jobs 4

Experiments on Cochrane Data (English, Medical)

The code in this repository can be used to reproduce the BART-based model described in Paragraph-level Simplification of Medical Texts (Devaraj et al., NAACL 2021). We reimplement the official code with following changes:

  • Evaluation is based on easse (BLEU, SARI) and rouge_score (ROUGE). The original evaluation script is not available as of 10/2022.
  • We use the Huggingface seq2seq trainer rather than PyTorch lightning. See simplepatho/run_translation.py.
  • No support for unlikelihood loss (corresponds to bart-no-ul in Devaraj et al., 2021)

Download and preprocess raw data:

make cochrane

Corpus statistics:

python -m simplepatho.corpus_statistics \
    --data_path data/processed/cochrane \
    --language english \
    --spacy_model en_core_web_sm

# Expect following output:
                             Source     Target
Docs                          4,459      4,459
Tokens                    1,844,617  1,052,902
Sentences                    63,539     43,631
Types (raw)                  29,577     20,844
Types (lemmas)               25,534     17,074
Avg. text length (words)        414        236
Avg. text length (sents)         14         10
Avg. TTR                       0.38       0.51
Avg. Flesch-Reading Ease      46.73      42.49
Avg. word length               4.69       4.91
Avg. sent length                 29         24
Novelty                              40/70/82%
CMP                                       0.59

Train models:

sh scripts/cochrane/{bart,bart-xsum,bert2bert,mbart}/train.sh
sh scripts/cochrane/{bart,bart-xsum,bert2bert,mbart}/predict.sh

Evaluate predictions:

python -m simplepatho.evaluate \
    --orig_path data/raw/cochrane/test.source \
    --refs_path data/raw/cochrane/test.target \
    --sys_paths \
        data/raw/cochrane/test.source \
        data/raw/cochrane/test.target \
        output/cochrane/bart/predictions_test.txt \
        output/cochrane/bart-xsum/predictions_test.txt \
        output/cochrane/bert2bert/predictions_test.txt \
        output/cochrane/mbart/predictions_test.txt \
    --compact \
    --calculate_metrics R-1 R-2 R-L BLEU SARI "|y|_w" "CMP_w" "Nov. (n=1)" "Nov. (n=2)" \
    --n_jobs 6

Results are given in the table below.

R-1 R-2 R-L BLEU SARI |y|_w CMP_w Nov. (n=1) Nov. (n=2)
gold - - - - - 203.52 0.47 0.37 0.69
identity 43.82 19.63 41.34 14.47 9.62 411.42 0.00 0.00 0.00
bart-no-ul* 40.00 15.00 37.00 44.00 38.00 228.27 - 0.05 0.11
Reproduction results.
bart 45.36 18.43 42.42 16.22 35.24 227.93 0.42 0.07 0.15
bart-xsum (corresponds to bart-no-ul) 45.13 18.38 42.27 16.48 34.23 240.92 0.39 0.05 0.14
bert2bert 31.84 5.71 29.54 3.41 38.40 213.88 0.44 0.60 0.92
mbart 43.77 19.16 41.14 14.59 20.64 372.61 0.08 0.04 0.06

*Taken from (Devaraj et al. 2021), Table 6.

Development

make format
make test
make lint

Citation

If you use the resources in this repository, please cite:

@inproceedings{trienes-etal-2022-patient,
    title = "Patient-friendly Clinical Notes: Towards a new Text Simplification Dataset",
    author = {Trienes, Jan  and
      Schl{\"o}tterer, J{\"o}rg  and
      Schildhaus, Hans-Ulrich  and
      Seifert, Christin},
    booktitle = "Proceedings of the Workshop on Text Simplification, Accessibility, and Readability (TSAR-2022)",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates (Virtual)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.tsar-1.3",
    pages = "19--27",
}

Contact

If you have any question, please contact Jan Trienes at jan.trienes [AT] gmail.com.