This repository contains the implementation of the methods described in our paper: "When automated fact-checking meets argumentation: unveiling fake news through argumentative evidence".

Overview

This codebase implements a deep learning framework for argument mining, fake news classification and analysis. It includes:

• Data preprocessing and feature extraction
• Transformer-based models for joint argument mining and fake news classification
• Evaluation metrics and visualization tools

Data

In the data folder you will find LIARArg.csv

The names of the columns are quite self-explanatory, here are some explanations for the less straightforward columns:

• claim_id: Each claim has a unique id to be identified during the relation prediction task.
• premise_id: Each premise has a unique id to be identified during the relation prediction task.
• claim_position: The start and end position of a claim vs. the text in the columnwhole_text.
• premise_position: The start and end position of a premise vs. the text in the columnwhole_text.
• summary: The summary of the whole fact-checking article (see the paper for more details).
• whole_text: The text displayed during the annotation task, useful to extract claims and premises using positions (start and end token indices).
• fullText_based_content: The full-length content of the fact-checking article, useful for summary generation.
• support_relation and similar columns: the dependant and governor of each relation. For instance, [1838, 1828] in the row with id 8249 means that the component with id 1838 supports the component with id 1828. In this specific example, 1838 is a premise and 1828 is a claim.

Installation

To install dependencies:

pip install -r requirements.txt

Usage

Data Preparation

python src/data/prepare_dataset.py --input_path data/raw --output_path data/processed

Model Training

python src/train.py --config configs/transformer_config.yaml

Evaluation

python src/evaluate.py --model_path models/best_model.pt --test_data data/processed/test.json