Emotion-aware fake news detection via BERT and NRC emotion lexicon

Abstract

This paper addresses the automated detection of fake news by developing an emotion-aware credibility scoring model that leverages the complementarity of affective signals and transformer-based text representations, with particular attention to low-resource computing environments characteristic of Sub-Saharan African digital contexts. A three-stage pipeline integrates: (i) subjectivity and polarity scores from TextBlob, (ii) eight-dimensional emotion intensity profiles from the NRC Emotion Lexicon, and (iii) a contextual probability from a fine-tuned BERT-base-uncased classifier. The three signals are fused into a unified credibility score S₁ ∈ [0,1] through a weighted formula calibrated by grid search on validation data. Evaluation is performed on the FakeNewsNet benchmark (3,720 articles) with cross-domain validation on the LIAR dataset. The proposed model achieves F₁ = 0.77 and AUC = 0.84 on FakeNewsNet, outperforming all evaluated baselines including standalone fine-tuned BERT (F₁ = 0.74). An ablation study confirms that the NRC emotion module provides the largest marginal gain (−0.04 when removed), with anger, fear, and disgust identified as the strongest discriminative markers. Cross-domain evaluation on LIAR yields F₁ = 0.71 without retraining, demonstrating domain-agnostic robustness. Integrating lexicon-based emotion features with contextual transformer representations yields measurable improvements over BERT alone and produces a model that is both effective and computationally affordable. The model has been integrated into the EyeAndEars platform for Cameroonian deployment, providing an accessible misinformation detection tool for journalists, educators, and citizens in environments where dedicated fact-checking infrastructure is limited.