Bert approaches for fake news detection
Loading...
Files
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Unuversity Setif 1 Ferhat Abbas . Faculty of Sciences
Abstract
The rapid growth of fake news across digital platforms poses serious risks to public dis-course and democratic stability, particularly during geopolitical crises in which false reports spread virally before verification can occur. This thesis presents a hybrid deep learning framework integrating BERT with Large Language Models (LLMs) to enhance fake news detection across English domain.
The proposed methodology fine-tunes RoBERTa for binary classification on the WELFake dataset — comprising 72,134 labelled English news articles — and evaluates the resulting system on a curated subset of FakeNewsNet PolitiFact under zero-shot cross-domain conditions. LLMs contribute through zero-shot prompting to generate structured authenticity features, including manipulation intensity scores, semantic vagueness assessments, and nu-merical evidence indicators, which are fused with RoBERTa's contextual embeddings via a gated late-fusion mechanism.
Experimental results demonstrate that the hybrid architecture consistently outperforms the RoBERTa-only baseline across all classification metrics, achieving accuracy and weighted F1-score improvements of approximately two percentage points on the in-domain test set, with stronger relative improvements under cross-domain evaluation. An ablation study con-firms that both the gated fusion architecture and the LLM inference-time scores contribute independently to the observed gains.
