Jomar T. Cuñado, January Venice V. Toledo, Christine D. Bandalan
The rise of fake news threatens public discourse, especially in regions with low-resource languages like Filipino (Tagalog). This study benchmarks five compact transformer models: TinyBERT, DistilBERT, MobileBERT, MiniLMv2, and ELECTRA-small, using the Fake News Filipino dataset. DistilBERT achieved the best individual performance (0.969 F1), while a weighted soft voting ensemble with MobileBERT offered a negligible 0.000005 gain. Efficiency benchmarking indicated MobileBERT was the fastest (47.95 ms latency) and most resource-efficient (94.38 MB model size). In contrast, DistilBERT and the ensemble were more computationally demanding for only marginal improvements. These results suggest that while ensembles can offer slight gains, a single compact model, particularly MobileBERT, is preferable in resource-constrained environments, reinforcing the practicality of lightweight transformers for mobile and rural deployment. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
Department of Computer, Information Sciences and Mathematics, University of San Carlos, Talamban Campus, Cebu City, 6000, Philippines
