Fake News Detection Framework using Natural Language Processing
The digital age has led to a rapid spread of fake news, undermining democracy and social cohesion. The emergence of convincing AI-generated misinformation has intensified this challenge, as existing detection systems either depend on GPU-intensive transformer models or rely on shallow surface features that fail to capture deeper semantic context. This study aims to design, develop, and evaluate a computationally efficient NLP-based framework for detecting both human-authored and AI-generated fake news, with emphasis on deployability in resource-constrained environments such as Nigerian newsrooms. A hybrid feature extraction pipeline was applied to a combined corpus of 73,349 articles, comprising 71,349 articles from the WELFake benchmark and 2,000 purpose-built AI-generated articles. The pipeline concatenated 50,000 statistical TF-IDF features with 384-dimensional semantic sentence embeddings (all-MiniLM-L6-v2). A Random Forest classifier was optimised via 5-fold cross-validated grid search and evaluated on a stratified held-out test set of 11,003 articles on a CPU-only environment (Intel Xeon, 12.7 GB RAM). The framework achieved an accuracy of 93.76%, a weighted F1-score of 93.75%, and a ROC-AUC of 98.68%, completing training in 160.8 minutes without GPU acceleration. A feature importance mechanism ensures alignment with Explainable AI (XAI) principles, demonstrating that principled hybrid feature engineering delivers scalable, interpretable, and hardware-friendly fake news detection.
Download Paper