Download the requirements using pip install -r requirements.txt
Replace the model_checkpoint.py from pytorch_lightning package in your conda environment with the model_checkpoint. Just so that the overwrite_existing can be used in the ModelCheckpoint in main.py. Just to prevent saving multiple same models which can take up alot of space. If this is not a problem, can ignore.
- Download the Eraser data
- Preprocess data to train NLI predictor
- Train NLI predictor
- Preprocess base dataset to train explainer and predictor
- Train explainer and predictor
Step 1 - run bash scripts/dl_eraser.sh
Step 2-4 - run bash scripts/preprocess_data.sh
step 5 - run bash scripts/main.sh (includes Full-C , Supervised , Hard and Soft approach for 3 datasets)
ROBUSTNESS TEST To get data to carry out robustness, ie get the adversarial sentences, change the root and data directory path in rationale-robustness/rr/config.py to your specified path.
rationale-robustness is from https://github.com/princeton-nlp/rationale-robustness
ERASER dataset is from https://github.com/jayded/eraserbenchmark
To get averaged results across all seed runs, run get_seed_results.py
Paper Citation
@article{jie2024plausible,
title={Plausible Extractive Rationalization through Semi-Supervised Entailment Signal},
author={Jie, Yeo Wei and Satapathy, Ranjan and Cambria, Erik},
journal={ACL},
year={2024}
}