Install PyTorch following the instructions on the (https://pytorch.org/). The code has been tested over PyTorch 1.8.1 versions.
Then install the other dependencies.
conda env create -f environment.yml
The dataset preprocessing code is in data/Sentires-preprocessing.py with randomized train|validation|test split in 80|10|10 ratios.
The training and evaluation code can be executed with:
python src/run.py
Default parameters are not the best performing-hyper-parameters. Hyper-parameter selection on the validation set has been specified in Subsection 5.1.4 in our paper and needs to be done through the command line arguments.