This project compares two IR approaches for claim verification. We use the ClimateFever dataset, which contains claims and evidences that support or refute a claim.

The data/ folder contains raw and processed data:

• raw/ The initial dataset in csv and json form
• processed/ Separated evidences, claims and the mapping between the claims and the evidences
• embeddings/ Claim- and Evidence embeddings for retrieval in Approach 2

The src/ folder contains the python file and jupyter notebooks.

• a1/ Files for approach 1. The pipeline.ipynb can be used to start the pipelin
• a2/ Files for approach 2. The pipeline.ipynb can be used to start the pipelin
• data/ Files for preprocessing the data. This includes cleaning the dataset and creating the files in `data/processed/
• eval/ Files for the evaluation of both approaches and their comparison
• config.py general settings.

Download the necessary models:

cd src/a1 && python3 bart_large_pull

cd src/a2 && python3 climatebert_pull

Start pipelines for each approach (pipeline.ipynb).

Run the demo to get a interactive Gradio Based interface.

python demo.py

Access the interface at:

http://127.0.0.1:7860