Explainable Misinformation Detection: A Hybrid Bert And Large Language Model (LLM) Framework for Interpretable Fake News Classification
Main Article Content
Abstract
The rapid growth of online news and social-media platforms has increased the spread of misinformation and created a need for automated detection methods that are both effective and interpretable. This study presents a compact hybrid framework for explainable misinformation detection that combines traditional machine-learning models, Transformer-based classification, post-hoc explainability, and controlled Large Language Model (LLM) interpretation. The LIAR-PLUS dataset is transformed into a binary classification task, with statements categorized as Real or Misinformation. TF-IDF-based Support Vector Machine (SVM) and Logistic Regression (LR), along with DistilBERT and BERT, are evaluated using Accuracy, Precision, Recall, F1-score, and ROC-AUC. BERT is subsequently used as the primary classifier for SHAP and LIME-based local explanations. To examine explanation behaviour and model sensitivity, SHAP–LIME feature overlap, SHAP-based feature deletion, and prediction stability are analysed. The study further employs a controlled local Qwen2.5-0.5B-Instruct model to convert SHAP-derived evidence into human-readable explanations without using the LLM as an independent fact-checker. The experimental results show that traditional and Transformer-based models achieve comparable performance under the evaluated configuration, while SHAP and LIME provide complementary local feature-level interpretations. Feature deletion demonstrates increasing prediction perturbation as more influential features are removed, while most predictions remain stable under the applied perturbation. The controlled LLM can generate structured explanations from model evidence, but unsupported semantic or contextual interpretations may still occur. Therefore, grounding and safety validation remain important when LLMs are used for explainable misinformation detection.