A Classification of Debunking in Indonesian Fact-Checking Platforms Using NLP and Machine Learning : A Mixed-Methods Approach with Corpus Analysis and IndoBERT: Klasifikasi Pembantahan dalam Platform Pengecekan Fakta Indonesia Menggunakan NLP dan Machine Learning: Pendekatan Metode Campuran dengan Corpus Analysis dan IndoBERT

Bayu Hartono (Universitas Pertahanan Republik Indonesia)
Riduan Riduan (Indonesia Defense University)
Rudy Agus Gemilang Gultom (Indonesia Defense University)
Hondor Saragih (Indonesia Defense University)

Journal of Data Insights
Vol 4 No 1 (2026): Journal of Data Insights

The rapid spread of disinformation through digital platforms constitutes a serious threat to social cohesion and public health. Debunking—the systematic refutation of false information using verified evidence—has emerged as a key countermeasure, yet manual identification and classification of debunking strategies is labor-intensive and difficult to scale. This study addresses this gap through a mixed-methods design integrating qualitative corpus analysis with automated machine learning (ML) classification. A corpus of 120 debunking articles published by three leading Indonesian fact-checking institutions (Kominfo AIS, Mafindo, and Cek Fakta Kompas, 2022–2024) was first manually annotated by two trained coders (Cohen's κ = 0.82) to identify four dominant debunking strategies: (1) contextual correction with emotional narrative framing; (2) source authority endorsement; (3) visual verification and reverse image search; and (4) myth-versus-fact inoculation format. This annotated corpus was subsequently used as a training dataset to develop and benchmark five NLP-based text classification models: TF-IDF + Support Vector Machine (SVM), TF-IDF + Random Forest, IndoBERT fine-tuned, IndoBERT with data augmentation (IndoBERT-Aug), and XGBoost with linguistic features. The IndoBERT-Aug model achieved the highest overall performance (macro-averaged F1 = 0.847, Precision = 0.851, Recall = 0.843), substantially outperforming the SVM baseline (F1 = 0.612). Logistic regression analysis further identified three significant moderators of debunking effectiveness: correction timeliness within 6 hours (OR=2.80, p<0.01), content readability (OR=0.68, p<0.01), and multi-platform distribution (OR=1.84, p<0.05), with the full model explaining 41% of variance (Nagelkerke R²=0.41). These contributions are formalized into the Indonesian Debunking Effectiveness Model (IDEM), a framework integrating automated strategy detection with evidence-based deployment guidelines for scalable counter-disinformation operations.