Zero-shot cross-lingual transfer for disinformation detection in low-resource Indic languages

Disinformation poses a severe threat in low-resource linguistic settings lacking annotated data and robust detection infrastructure. This study investigates the efficacy of multilingual transformers for zero-shot and few-shot cross-lingual disinformation detection, transferring knowledge from English to four Indic languages (Bangla, Hindi, Malayalam, and Tamil). Zero-shot evaluations reveal that large-scale models enable substantial knowledge transfer—with XLM-RoBERTa achieving a peak Macro-F1 of 0.726 on Bangla—while traditional TF-IDF baselines fail entirely, underscoring the necessity of deep contextual embeddings. Furthermore, few-shot adaptation utilizing merely 200 target-language samples yields profound improvements, elevating MuRIL's performance to an average Indic Macro-F1 of 0.805 and 0.945 on Hindi. Post-hoc explainability analyses utilizing LIME and SHAP confirm that these models successfully identify genuine, cross-lingual stylistic cues of deception (e.g., sensationalism and urgency markers) rather than relying on spurious memorization. Ultimately, this research demonstrates that combining multilingual transformers with minimal target supervision provides a highly effective, scalable framework for combating disinformation in resource-constrained environments.