Pre-CoFactv3

Code for AAAI 2024 Workshop Paper: “Team Trifecta at Factify5WQA: Setting the Standard in Fact Verification with Fine-Tuning”

In this paper, we present Pre-CoFactv3, a comprehensive framework comprised of Question Answering and Text Classification components for fact verification. Leveraging In-Context Learning, Fine-tuned Large Language Models (LLMs), and the FakeNet model, we address the challenges of fact verification. Our experiments explore diverse approaches, comparing different Pre-trained LLMs, introducing FakeNet, and implementing various ensemble methods. Notably, our team, Trifecta, secured first place in the AAAI-24 Factify 3.0 Workshop, surpassing the baseline accuracy by 103% and maintaining a 70% lead over the second competitor. This success underscores the efficacy of our approach and its potential contributions to advancing fact verification research.

💡 Usage

Setup environment

1. Clone or download this repo

   git clone https://github.com/AndyChiangSH/Pre-CoFactv3.git
   

2. Move into this repo

   cd Pre-CoFactv3
   

3. Setup the virtual environment

   conda env create -f environment.yaml
   

4. Activate the virtual environment

   conda activate pre_cofactv3

Question Answering

Fine-tuning Large Language Models (LLMs)

1. Change the arguments in question_answering/config.yaml

2. To fine-tune the question answering model, run

   python question_answering/finetune.py
   

   The fine-tuned model will be saved in question_answering/model/finetune/<model name>/

3. To generate the answer, run

   python question_answering/generate_answer.py
   

   The generated answer will be saved in question_answering/answer/<model name>/

4. To evaluate the answer, run

   python question_answering/evaluate_answer.py
   

   The evaluation result will be saved in question_answering/evaluate/<model name>/

Text Classification

FakeNet

1. Create the config in text_classification/fakenet/config/<model name>.yaml

2. To train the FakeNet, run

   bash generate_label/fakenet/train.sh <model name>

   The FakeNet will be saved in text_classification/fakenet/model/<model name>/

3. To generate the label, run

   python generate_label/fakenet/generate_label.py --model=<model name> --mode=<train or val or test>
   

   The generated label will be saved in text_classification/fakenet/label/<model name>/

4. To evaluate the label, run

   python generate_label/fakenet/evaluate_label.py --model=<model name> --mode=<train or val or test>
   

   The evaluation result will be saved in text_classification/fakenet/evaluate/<model name>/

5. To extract features, run

   python text_classification/fakenet/feature_extractor/feature_extraction.py
   

Fine-tuning Large Language Models (LLMs)

1. Create the config in text_classification/finetune/config/<id>.yaml

2. To fine-tune the text classification model, run

   python finetune/finetune.py --id <id>
   

   The fine-tuned model will be saved in text_classification/finetune/model/<model name>/

3. If you want to fine-tune the model sequentially, run

   bash text_classification/finetune/finetune.sh
   

4. To generate the label, run

   python generate_label/finetune/generate_label.py --model=<model name> --mode=<train or val or test> --device=<device name>
   

   The generated label will be saved in text_classification/finetune/label/<model name>/

5. To evaluate the label, run

   python generate_label/finetune/evaluate_label.py --model=<model name> --mode=<train or val or test>
   

   The evaluation result will be saved in text_classification/finetune/evaluate/<model name>/

Ensemble

1. Put the models that you want to ensemble in text_classification/ensemble/model/<model name>/<train or val or test>_prob.json, which will be generated in text_classification/fakenet/label/<model name>/ or text_classification/finetune/label/<model name>/

2. To ensemble by weighted sum with labels, run

   python ensemble/ensemble_1.py --model_1=<model_1 name> --model_2=<model_2 name> --mode=<train or val or test>
   

   The ensemble result will be saved in text_classification/ensemble/ensemble_1/<model_1 name>+<model_2 name>/

3. To ensemble by power weighted sum with labels, run

   python ensemble/ensemble_2.py --model_1=<model_1 name> --model_2=<model_2 name> --mode=<train or val or test>
   

   The ensemble result will be saved in text_classification/ensemble/ensemble_2/<model_1 name>+<model_2 name>/

4. To ensemble by power weighted sum with two models, run

   python ensemble/ensemble_3.py --model_1=<model_1 name> --model_2=<model_2 name> --mode=<train or val or test>
   

   The ensemble result will be saved in text_classification/ensemble/ensemble_3/<model_1 name>+<model_2 name>/

5. To ensemble by power weighted sum with three models, run

   python ensemble/ensemble_4.py --model_1=<model_1 name> --model_2=<model_2 name> --model_3=<model_3 name> --mode=<train or val or test>
   

   The ensemble result will be saved in text_classification/ensemble/ensemble_4/<model_1 name>+<model_2 name>+<model_3 name>/

In-Context Learning

1. Add your own ChatGPT API key in in_context_learning/key.txt

2. To generate labels by In-Context Learning, run

   python in_context_learning/main.py
   

3. When enough data is collected, run

   python in_context_learning/compare.py
   

🤖 Models

Question Answering

AndyChiang/Pre-CoFactv3-Question-Answering on Hugging Face.

Text Classification

AndyChiang/Pre-CoFactv3-Text-Classification on Hugging Face.

💾 Datasets

We utilize the dataset FACTIFY5WQA provided by the AAAI-24 Workshop Factify 3.0, saved in data/.

This dataset is designed for fact verification, with the task of determining the veracity of a claim based on the given evidence.

• claim: the statement to be verified.
• evidence: the facts to verify the claim.
• question: the questions generated from the claim by the 5W framework (who, what, when, where, and why).
• claim_answer: the answers derived from the claim.
• evidence_answer: the answers derived from the evidence.
• label: the veracity of the claim based on the given evidence, which is one of three categories: Support, Neutral, or Refute.

	Training	Validation	Testing	Total
Support	3500	750	750	5000
Neutral	3500	750	750	5000
Refute	3500	750	750	5000
Total	10500	2250	2250	15000

🏆 Leaderboard

Team Name	Accuracy
Team Trifecta	0.695556
SRL_Fact_QA	0.455111
Jiankang Han	0.454667
Baseline	0.342222

📌 Citation

😀 Author

• Shang-Hsuan Chiang (andy10801@gmail.com)
• Ming-Chih Lo (max230620089@gmail.com)
• Lin-Wei Chao (william09172000@gmail.com)
• Wen-Chih Peng (wcpeng@cs.nycu.edu.tw)

⭐ Star History