A Collaborative Framework for Evidence-Grounded Misinformation Detection in Bengali Using Retrieval-Augmented Large Language Models
Mahadi Hasan, Abdulla Masud, Sajida Kabir, John Gomes, Nafees Mansoor
Conference Paper. IEEE International WIE Conference on Electrical and Computer Engineering/WIECON-ECE, no. 2025, pp. 943–948 (2025).
Abstract
The rapid proliferation of misinformation across digital platforms has heightened the need for interpretable and language-inclusive verification systems. In low-resource contexts such as Bengali, the lack of multimodal datasets and benchmark frameworks significantly constrains progress in automated fact-checking. This study presents an in-depth evaluation of Large Language Models (LLMs) for veracity detection in Bengali and introduces interpretability-driven strategies that couple reasoning transparency with factual grounding. To facilitate this research, a manually curated dataset, BD-FakeDetect, comprising 5,929 Bengali news items labeled across eight veracity categories, was developed from verified fact-checking portals. A comprehensive analysis of linguistic distributions and category imbalance underscores the complexities of Bengali misinformation. Multiple LLMs, including DeepSeek-7B, BanglaLLaMA, and Mixtral, were fine-tuned using parameter-efficient fine-tuning on textual data, while supervised vision-based fine-tuning was applied to GPT-4o, utilizing multimodal inputs, while a Retrieval-Augmented Generation (RAG) pipeline was integrated to reduce hallucination and improve factual consistency. GPT-4o achieved a validation accuracy of 81.8%, outperforming all text-only baselines, and, constitutes the initialization multimodal misinformation detection study in the Bengali language. These findings highlight the promise of combining reasoning-aware LLMs with retrieval-based grounding to strengthen factual reliability in low-resource misinformation detection. The work contributes a publicly usable Bengali dataset, empirical evaluation of multilingual LLMs, and an interpretable framework supporting human-AI collaboration in fact-checking. © 2025 IEEE.
Keywords
Bengali NLP, Dataset Creation, FactChecking, Interpretability, Large Language Models, Low-Resource Languages, Misinformation Detection, Retrieval-Augmented Generation