E-CaTCH — Event-Centric Misinformation Detection
- PyTorch
- multimodal
- attention
- misinformation
What it is
Official implementation of E-CaTCH. Instead of judging posts in isolation, it clusters them into pseudo-events (by textual similarity and temporal proximity) and reasons over each event as a coherent narrative.
Within an event it extracts text and image features with BERT and ResNet-152, refines them with intra-modal self-attention, aligns them through bidirectional cross-modal attention with soft gating, and tracks how the story evolves with a trend-aware LSTM using semantic-shift and momentum signals. Severe class imbalance is handled with adaptive class weighting and hard-example mining.
Reported results: ~95.5% accuracy on Fakeddit, with strong cross-dataset generalization to India-Elections and COVID-19 MISINFOGRAPH.
Stack
PyTorch · BERT · ResNet-152 · LSTM · YAML-configured training
Links
- GitHub repository
- Related paper: arXiv:2508.11197