Vision-Language AI Model Detects Bengali Fake News with Explainable Reasoning

Fake news in Bengali digital media rarely arrives as plain text. It spreads through memes, screenshots, posters, and photographs overlaid with Bengali script, which means detection systems that read only text or only images miss much of the evidence. A new study addresses this gap with a framework called the Context-Aware Multimodal Reasoning Framework, or CAMRF, which jointly analyzes images and embedded Bengali text and, unusually for this field, explains its verdicts in natural language.

The framework, described in an open-access paper published on 31 August 2026, chains together several established components in a novel arrangement. EasyOCR first extracts Bengali text embedded within each image. Two visual encoders then process the picture itself: ResNet-50 captures local features such as edges, textures, and spatial patterns, while a Vision Transformer uses self-attention to model global, long-range relationships across image regions. The OCR-derived text is converted into dense 768-dimensional semantic embeddings using the pretrained IndicBERT encoder.

The methodological centerpiece is the Context-Aware Multimodal Fusion Module, or CAFM. Rather than simply concatenating the three evidence streams, CAFM maps them into a common latent feature space and computes a trainable contribution score for each modality on every individual sample. Softmax normalization turns these scores into weights that sum to one, so the relative influence of local visual detail, global visual context, and extracted Bengali text adapts sample by sample. The authors argue this matters because the reliability of OCR text, image details, and overall visual context varies considerably across memes, posters, and screenshots.

The fused representation then feeds into Qwen2-VL-2B-Instruct, a vision-language model employed as a frozen reasoning component. It performs cross-modal semantic reasoning over the combined evidence, checking for consistency or contradiction between what the text asserts and what the image shows, and produces two outputs: a binary authenticity label and a human-readable explanation of the reasoning behind it. The authors position this explainability as central to building user trust in automated misinformation judgments, moving beyond a bare fake-or-real verdict.

Evaluation used the Multimodal Bengali Fake News Dataset, a collection of 1,003 instances balanced between 501 fake and 502 authentic items, compiled from reliable Bengali media sources and web scraping. The team applied perceptual hashing and exact string matching to remove duplicates, and maintained stratified five-fold cross-validation strictly at the unique post or event level, keeping variant re-posts of the same event within a single fold to prevent test-set contamination. Reported metrics are averages across the five evaluation folds.

CAMRF achieved 97.00% accuracy, with precision, recall, and F1-score all at 0.97, and a standard deviation of only 0.38 percentage points in accuracy across folds. It outperformed a series of visual-only baselines evaluated under identical splits: a core CNN at 89.00%, MobileNet at 92.00%, ResNet-50 at 91.00%, ViT at 93.00%, and a hybrid ResNet-50 plus ViT model at 94.00%. An ablation analysis traced the gains stepwise: adding OCR-extracted Bengali text raised accuracy to 95.00%, incorporating Qwen2-VL reasoning brought it to 96.00%, and the full CAFM fusion reached 97.00%. The area under the ROC curve was 0.98.

The gains held up under statistical testing. A paired t-test against the best baseline, the ResNet-50 plus ViT hybrid, showed a 3.00% accuracy gain with p = 0.021; the Wilcoxon signed-rank test gave W = 0.0 with p = 0.018; and McNemar’s test on binary outputs yielded χ² = 5.33 with p = 0.021. Comparisons with general-purpose vision-language models evaluated zero-shot on the same dataset also favored the task-specific pipeline: standalone Qwen2-VL-2B-Instruct prompting reached an F1-score of 0.94, LLaVA-1.5 scored 0.90, and BLIP-2 scored 0.88, which the authors attribute to the difficulty these models have with complex Bengali script and high-resource-language-centric pretraining.

The explainability claims received their own scrutiny. In a blind evaluation of 100 randomly selected test samples by two independent domain experts, the generated explanations scored 4.62 out of 5 for faithfulness, 4.78 for comprehensibility, and 4.55 for evidence alignment, with inter-rater agreement of Cohen’s κ = 0.82. Grad-CAM heat maps on the visual branch indicated that the encoders focused on altered image regions, face splices, and Bengali text overlays. Two case studies, a lunar eclipse post correctly labeled real and a misleading post about a public figure correctly labeled fake, illustrated the pipeline end to end.

The authors are candid about limitations. The framework was tested on a single dataset of 1,003 samples, its binary classifier cannot handle satire, humor, or partially false content, and OCR accuracy degrades on stylized, low-contrast, or handwritten calligraphy. The two-billion-parameter vision-language model demands GPU resources, and explanation quality was assessed by expert opinion rather than against gold-standard references. An error analysis of the six misclassified test instances identified three failure modes: OCR failures on decorative typography, sarcasm conveyed through metaphor, and heavy JPEG compression in images smaller than 150 by 150 pixels.

Even so, the study marks a concrete step for misinformation detection in low-resource languages, where Bengali, one of the world’s most widely spoken languages, has seen far less research than English. The authors suggest the approach could extend to other Indic languages and point toward future work on multilingual detection, integration with fact-checking knowledge bases, and lightweight deployment for resource-constrained environments. For now, the result demonstrates that combining OCR, complementary visual encoders, adaptive fusion, and vision-language reasoning can outperform image-only classification while telling users why it reached its conclusion.

Subject of Research: A context-aware multimodal reasoning framework for explainable Bengali fake news detection using vision-language models.

Article Title: Context-aware multimodal reasoning for explainable Bengali fake news detection using vision-language models

Article References: Mishrra, S., Bal, S., Dhar, A., Mondal, S., Ghosh, A., & Guha, S. (2026). Context-aware multimodal reasoning for explainable Bengali fake news detection using vision-language models. Discover Networks, 2(1), Article 32. https://doi.org/10.1007/s44354-026-00047-w

Image Credits: AI Generated

DOI: 10.1007/s44354-026-00047-w

Keywords: fake news detection, Bengali, multimodal learning, vision-language models, explainable AI, OCR, machine learning, misinformation, deep learning, low-resource languages, Context-aware, multimodal