Towards Trustworthy Explanation: On Causal Rationalization

(Notice: we have updated the results of a baseline method folded-rationalisation (FR)) in our most recent version in arxiv)

Getting Started

Firstly, create the Python environment and activate it

conda create --name pytorch_py38 python=3.8

source activate pytorch_py38

To install the dependencies, run the following command

cd rationale-causal

# Install all python dependencies
pip install -r requirements.txt

You can download Beer and Hotel review dataset from https://github.com/YujiaBao/R2A and then put datasets in the data folder.

Before running experiments for Beer review data, utilize cr/data-processing.ipynb to get the short and the noise version data.

Training commands for causal rationalization method:

# real data
./scripts/beer/run_beer_aroma.sh causal-rationale

# synthetic data
./scripts/beer_noise/run_beer_aroma.sh causal-rationale

We thank the authors of Can Rationalization Improve Robustness? for their implementations of the baseline methods