MISINFORMATION DETECTION IN LOW-RESOURCE LANGUAGES USING MULTILINGUAL TRANSFORMERS AND EXPLAINABLE CLASSIFICATION

Abstract

Misinformation detection in low-resource languages remains a major challenge because most existing fake-news classification systems are trained on high-resource languages with large annotated datasets. This paper presents a multilingual transformer-based framework for detecting misinformation in low-resource language settings using explainable text classification. The proposed approach evaluates multilingual transformer models, including mBERT, XLM-RoBERTa, and domain-adapted multilingual variants, on multilingual misinformation datasets containing diverse linguistic structures and limited labeled samples. To improve robustness, the study integrates transfer learning, cross-lingual fine-tuning, class-balanced training, and explainability techniques such as attention visualization and SHAP-based feature interpretation. The results show that multilingual transformers outperform conventional machine learning baselines by capturing contextual, semantic, and cross-lingual cues more effectively. XLM-RoBERTa achieved the strongest overall performance, especially in languages with limited training examples, demonstrating the benefit of large-scale multilingual pretraining. Explainability analysis further revealed that the model relied on emotionally charged terms, misleading claims, named entities, and uncertainty markers when identifying misinformation. These findings highlight the potential of multilingual and explainable AI systems to support misinformation monitoring in linguistically diverse online environments. The study contributes a practical framework for low-resource misinformation detection while emphasizing transparency, fairness, and adaptability across languages.