plausible-nlp-explanations

Code and data of the paper:

Lucas Resck, Marcos M. Raimundo, Jorge Poco. "Exploring the Trade-off Between Model Performance and Explanation Plausibility of Text Classifiers Using Human Rationales," NAACL Findings 2024.

Check out the paper!

Instructions

Thank you for your interest in our paper and code.

Steps to reproduce our experiments:

1. Setup: Conda environment, data download, and optional model access request.
2. Experiments: raw experimental data generation with experiments/ code;
3. Figures and Tables: final results generation with code at notebooks/.

1. Setup

1.1. Conda Environment

Clone and enter the repository:

git clone https://github.com/visual-ds/plausible-nlp-explanations.git
cd plausible-nlp-explanations

Ensure anaconda or miniconda is installed. Then, create and activate the Conda environment:

cd conda
conda env create -f export.yml
conda activate plausible-nlp-explanations

In case of package inconsistencies, you may also create the environment from no_builds.yml (which contains package versions without builds) or history.yml (package installation history).

Install moopt:

cd ../..
git clone https://github.com/marcosmrai/moopt.git
cd moopt
git checkout e2ab0d7b25e8d7072899a38dd2458e65e392c9f0
python setup.py install

This is a necessary package for multi-objective optimization. For details, check its repository.

1.2. Data

We point the reader to the original dataset sources and their respective licenses:

• Download HateXplain from authors' GitHub and save it into data/hatexplain/.
• Download Movie Reviews from authors' webpage and extract it into data/movie_reviews/.
• Download Tweet Sentiment Extraction from Kaggle and extract it into data/tweet_sentiment_extraction/.
• Download HatEval dataset at the author's instantaneous data request form and extract it into data/hateval2019/. If the link is unavailable, we can share it under CC BY-NC 4.0—contact us.

data should look like this:

explainability-experiments
├── conda
├── data
│   ├── hateval2019
│   ├── hatexplain
│   ├── movie_reviews
│   └── tweet_sentiment_extraction
├── experiments
├── ...

1.3 Optional Models Access Request

Request access to the models fine-tuned on HateXplain dataset at Hugging Face: DistilBERT and BERT-Mini. Make sure to have set up a Hugging Face user access token and added it to your machine running

huggingface-cli login

We do not release HateXplain–fine-tuned models publicly because of ethical concerns. These models are only necessary for specific experiments though.

2. Experiments

2.1. Main Experiments

To generate the raw experimental data, you can use the following command:

python experiments.py --datasets hatexplain movie_reviews \
    --explainers lime shap --models tf_idf distilbert \
    --negative_rationales 2 5 --experiments_path data/experiments/ \
    --performance_metrics accuracy recall \
    --explainability_metrics auprc sufficiency comprehensiveness \
    --random_state 42 --gpu_device 0 --batch_size 128

You can change the parameters as you wish. For example, to run the same experiment but with BERT-Mini, you can change --models tf_idf distilbert to --models bert_mini. For details, please check the help:

python experiments.py -h

2.2. BERT-HateXplain Comparison

To run the BERT-HateXplain comparison, you can use the following command:

python comparison.py --datasets hatexplain_all --explainers lime \
    --models bert_attention --negative_rationales 2 \
    --experiments_path data/experiments/ \
    --performance_metrics accuracy recall \
    --explainability_metrics auprc sufficiency comprehensiveness \
    --random_state 42 --gpu_device 0 --batch_size 128 --n_jobs 1

You can also change the parameters as you wish. For details, please check the help:

python comparison.py -h

2.3. Out-of-Distribution Results

To run the out-of-distribution experiments, simply run

python out_of_distribution.py

3. Figures and Tables

To generate the final figures and tables, run the Jupyter notebooks at the folder notebooks/. It is necessary to have LaTeX installed on your machine to generate the figures. For example, to generate Figure 3 (DistilBERT and HateXplain's trade-offs), run the notebook notebooks/complete_graphic.ipynb. For details, please check the notebooks.

Adaptation

A new dataset should be a new Python module at experiments/datasets/ following the Dataset base class at experiments/dataset.py. Then, it should be easy to add it to experiments.py. The same is true for new models and explainers, and similar for metrics.py. Feel free to open issues and pull requests.

License

We release this code under the MIT License.

Citation

@inproceedings{resck_exploring_2024,
	address = {Mexico City, Mexico},
	title = {Exploring the {Trade}-off {Between} {Model} {Performance} and {Explanation} {Plausibility} of {Text} {Classifiers} {Using} {Human} {Rationales}},
	url = {https://aclanthology.org/2024.findings-naacl.262},
	doi = {10.18653/v1/2024.findings-naacl.262},
	booktitle = {Findings of the {Association} for {Computational} {Linguistics}: {NAACL} 2024},
	publisher = {Association for Computational Linguistics},
	author = {Resck, Lucas and Raimundo, Marcos M. and Poco, Jorge},
	editor = {Duh, Kevin and Gomez, Helena and Bethard, Steven},
	month = jun,
	year = {2024},
	note = {Also presented as a poster at the LatinX in NLP at NAACL 2024 workshop},
	pages = {4190--4216},
}