Sufficient Facts

This work is based on the paper Fact Checking with Insufficient Evidence, by Atanasova et al. as NLP Project for M2 IASD

by Oumaima BENDRAMEZ, Rida LALI, Hamza TOUZANI

Installation

To install the required packages to run the code, you should run the following command in your environment :

pip install focal_loss_torch tqdm torch transformers numpy sklearn

Dataset files

• "full.jsonl" is the 3 datasets combined (Hover+fever+vitaminc)
• the "full2.jsonl" file was computed from "full.jsonl" file to rearrange and rebalance the dataset and train/val/test split.

Run

Here is an example command to Train and Evaluate RoBERTa model on the full2.jsonl dataset:

python model.py --epochs 300 --dataset_dir sufficient_facts --dataset full2.jsonl --batch_size 8 --lr 1e-6 --model roberta

• epochs flag is for the number of epochs
• dataset_dir flag is to give the local directory of the dataset
• dataset flag is to give the dataset file
• batch_size and lr are learning parameters
• model is to select between BERT and RoBERTa models (bert or roberta)