Misinformation model research paper

Abstract

We introduce a modular pipeline for automated fact‑checking that integrates neural text understanding with retrieval‑based ranking (e.g., BM25). Claims from four public corpora, FEVER, LIAR, PolitiFact, and GossipCop, are unified into a three‑class FEVER label scheme (SUPPORTS, REFUTES, NOT ENOUGH INFO), ensuring a balanced training pool. We fine‑tune a BERT‑based verifier on claim–evidence pairs for semantic verification. Simultaneously, we train a lightweight CNN on claim‑only text for complementary classification. Their probability outputs are then fused in a stacked ensemble via a logistic regression meta‑classifier. 

On a 600‑claim validation set, individual model accuracies reach 47.6% (BERT), 45.0% (CNN), and 39.6% (BM25). The ensemble of BERT and CNN achieves an accuracy of 55.0% (macro-F1 = 0.550), representing a 7.4-point improvement over the best single model. Confusion‑matrix analysis shows REFUTES statements are detected most reliably, while SUPPORTS and NOT ENOUGH INFO remain challenging. Our findings confirm that a simple, interpretable ensemble can effectively leverage complementary strengths of neural models and retrieval methods, providing a strong foundation for scalable fact‑checking.