Year: 2026
Type: preprint
Abstract: However, automated detection of misinformation is difficult in low-resource languages because of the sheer volume of misinformation in these languages. This study aimed to detect fake news in the Kurdish language, using the Kurdish Dataset for Fake News Detection (KDFND), which is composed of 100,966 labeled articles. This work compared a variety of models, ranging from traditional machine learning (ML) (logistic regression (LR), random forest (RF), XGBoost, and support vector machines (SVM)) to ensemble models (voting and stacking classifiers) and deep learning (DL) architectures (CNN and BiLSTM-Attention). Another important contribution is the source-aware feature representation (based on the concept of source verification in the Qur'an, Surah Al-Hujurat 49:6), which combines information from news sources with the text of news articles, as it demonstrates the empirical validation of the critical role of source credibility in the detection of misinformation. The results show that the source-aware models come close to perfection in terms of performance, with CNN and BiLSTM-Attention reaching a 100% F1-score. In terms of accuracy, the source-aware CNN and BiLSTM-Attention models both achieved 100%, while the stacking and XGBoost classifiers followed closely at 99.98%. The average performance of the models drops significantly by 18.44% without source information, with the best model using only content (CNN) dropping to 82.97% F1. The accuracy of the best content-only model (CNN) was 82.43%. The performance gap empirically validates the Qur'anic command that information cannot be evaluated on its own without corroboration of its source. The CNN model was the best solution for real-world deployment, with perfect accuracy and the fastest training time (10 seconds) and inference time (0.01 seconds). This study introduces four contributions: (1) comprehensive benchmarking of Kurdish fake news detection: ML, ensemble, and DL; (2) empirical validation of the Qur'anic principle of source verification; (3) optimal model for Kurdish fake news detection in a resource-constrained environment: CNN; and (4) baseline performance metrics for Kurdish NLP research. The methodology is language-independent and can be generalized to other low-resource languages for which source credibility can provide meaningful cues to detect misinformation.
Source: Research Square
Authors: azad Karim, Bryar Hassan
Institutions: Charmo University, University of Kurdistan Hewler
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225430232