Abstract :

False And Misleading News Now Moves Through Social Platforms Faster Than Editors Or Fact-checkers Can Respond, And The Classifiers Built To Flag It Are Often Hard To Question. Transformer Models Such As BERT And RoBERTa Reach High Accuracy On Standard Benchmarks, Yet They Give A Reader Little Reason To Believe Any Particular Verdict. This Paper Describes A Hybrid Detector That Looks At A News Article In Three Complementary Ways: Contextual Representations From XLNet, Sub-word FastText Embeddings Passed Through A Convolutional Neural Network, And A Small Set Of Stylometric Cues Such As Sentence Length, Vocabulary Richness And Emotive Wording. The Three Feature Groups Are Concatenated, Scaled And Classified By A Weighted Soft-voting Ensemble, And Every Prediction Comes With SHAP Token Attributions That Show Which Words Moved The Decision. On A Held-out Set Of 800 Articles The Model Reached 98.00% Accuracy, 98.31% Precision, 97.84% Recall, An F1-score Of 98.08% And An Area Under The ROC Curve Of 0.9975, Against 82.00% Accuracy For The Global-local Ensemble Detector Used As The Reference Baseline. Standalone Transformer Classifiers Scored Higher On Larger Test Partitions, And We Explain Why Those Figures Are Not Directly Comparable. The Explanations Also Proved Useful As A Diagnostic, Exposing A Reliance On Publisher Names That A Headline Accuracy Figure Would Have Hidden. Index Terms—fake News Detection, Misinformation, XLNet, FastText, Convolutional Neural Network, Ensemble Learning, Explainable Artificial Intelligence, SHAP

Published:

03-11-2026

Issue:

Vol. 26 No. 11 (2026)

Page Nos:

66-70

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite