Countering Social Media Misinformation: A Sentence-Level Feature Approach to Fake News Classification

Closed

Rakhmat Arianto, Rudy Ariyanto, Ahmadi Yuli Ananta, Imam Fahrur Rozi, Erfan Rohadi, Natasha Dwi Pramudita, Bagus Winarko

2025 ICoAIT 2025 - 1st International Conference on Artificial Intelligence Technology - Artificial Intelligence: Driving Prosperity and Sustainability in the Modern World Conference paper Cited by 0 Quartile

Abstract

This paper introduces a hybrid framework for fake news classification in Indonesian-Language social media by integrating document-level similarity with sentence-level linguistic analysis. The approach begins with cosine similarity computation between user-submitted headlines and a curated hoax corpus represented through TF-IDF vectors. Headlines that exceed the threshold are immediately labeled as fake, while those falling below undergo a second stage of analysis. In this stage, relevant tweets are retrieved in real time and examined using six carefully selected sentence-level features: average sentence length, punctuation frequency, function word usage, phrase structure count, sentiment polarity, and the type-token ratio of content words. These features are designed to capture the syntactic and stylistic patterns commonly found in misinformation. The dataset, collected from TurnBackHoax.id, Komdigi, and Kompas, consists of 32,865 labeled entries. A stratified 10-fold cross-validation was employed to evaluate five machine learning classifiers. Results demonstrate that the Support Vector Machine (SVM) with an RBF kernel achieved the strongest performance, recording an F1-score of 84.4% and surpassing MLP, KNN, Decision Tree, and Naive Bayes. Validation on 15 real news headlines further confirmed the robustness of the framework in low-similarity cases. These findings underscore that the integration of vector-based similarity with optimized sentence-level features enhances detection accuracy while preserving transparency and adaptability. The proposed model offers a lightweight and domain-flexible solution that is particularly suitable for real-time misinformation mitigation in low-resource contexts. © 2025 IEEE.

Affiliations

Politeknik Negeri Malang, Department of Information Technology, Malang, Indonesia; Universitas Pendidikan Indonesia, Faculty of Economics and Business Education, Department of Management, Bandung, Indonesia