NEURAL ARCHITECTURE COMPARISON FOR FACT VERIFICATION ON FEVER DATASET

Abstract

The exponential growth of misinformation and fake news across digital platforms poses unprecedented challenges to information integrity, requiring sophisticated automated fact-checking systems capable of verifying claims against reliable evidence sources with high accuracy and computational efficiency. This study aims to evaluate and compare four hybrid neural architectures (BiLSTM-CNN, BiLSTM-RNN, BiLSTM-GRU, and BiLSTM-GNN) for automated fact verification using the FEVER dataset, investigating their effectiveness in claim-evidence verification under GPU memory constraints while analyzing training dynamics and generalization capabilities. The following results are obtained: The BiLSTM-CNN architecture achieved optimal performance with 79.5% accuracy, 79.5% recall, 77.9% F1-score, and 93.4% AUC-ROC, followed by BiLSTM-GNN (78.9% accuracy, 93.3% AUC-ROC) and BiLSTM-GRU (77.9% accuracy, 92.2% AUC-ROC), while BiLSTM-RNN exhibited catastrophic failure (33.3% accuracy). All successful architectures demonstrated significant overfitting with 15-17% train-validation accuracy gaps, indicating systematic generalization challenges with limited training data (40,000 samples). Conclusion. Multi-kernel convolutional feature extraction proves most effective for local pattern recognition in fact verification, while graph-inspired approaches show promising potential for relational reasoning. The consistent overfitting across architectures highlights the critical need for enhanced regularization, data augmentation, and ensemble methods to achieve robust performance in automated fact-checking systems under computational constraints.