Automated Fact-checker

Problem

Build an ML system to verify the veracity of claims.

Dataset

PUBHEALTH dataset has an associated veracity label (true, false, unproven, mixture). Each instance in the dataset has an explanation text field. The explanation is a justification for which the claim has been assigned a particular veracity label.

source: https://huggingface.co/datasets/health_fact

Important files:

• BERT_fact_checker.ipynb : describes the steps of implementation
• src/ bertClassifier.py : contains class and functions to initialize and train the BERT model

Installed libraries

• transformers
• datasets
• sentence_transformers
• umap-learn

Important libraries

import sklearn
from transformers import *
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix, classification_report
import torch
from transformers import PegasusForConditionalGeneration, PegasusTokenizer
from src.bertClassifier import *

Implementation steps

1. Load Data
2. Preprocess Data
3. Build the Model (BERT)
4. Predict & Evaluate (63% Acc.)
5. Data Augmentation + Predict & Evaluate (65% Acc.)
6. Issues for consideration
7. ANNEX - Data visualization

Please see 'BERT_fact_checker.ipynb' for details.