Epicurus at SemEval-2023 Task 4

Project description

This repository is created for reproducing the approach and results in the following paper:

Epicurus at SemEval-2023 Task 4: Improving Prediction of Human Values from Arguments by Leveraging Their Definitions.

Set up environment

conda create --name epicurus python=3.10 pip
conda activate epicurus
pip install -r requirements.txt
cd ./src

Download data

Please download the challenge data from this link to the folder ./data/raw/touche23

Model training

python train.py --batch_size 32 --gradient_step_size 4 --definition description --weighted_loss 'weighted' --test_mode no
python train.py --batch_size 32 --gradient_step_size 4 --definition description --weighted_loss 'not_weighted' --test_mode no
python train.py --batch_size 32 --gradient_step_size 4 --definition survey --weighted_loss 'weighted' --test_mode no
python train.py --batch_size 32 --gradient_step_size 4 --definition survey --weighted_loss 'not_weighted' --test_mode no

Get predictions

python predict.py --definition description --weighted_loss weighted --model_number [best model number] --test_mode no
python predict.py --definition description --weighted_loss not_weighted --model_number [best model number] --test_mode no
python predict.py --definition survey --weighted_loss weighted --model_number [best model number] --test_mode no
python predict.py --definition survey --weighted_loss not_weighted --model_number [best model number] --test_mode no

License

This project is licensed under the terms of the MIT License.