FaiyazAbdullah114708/BanglaVerse

Many Dialects, Many Languages, One Cultural Lens: Evaluating Multilingual VLMs for Bengali Culture Understanding Across Historically Linked Languages and Regional Dialects

Abstract: Bangla culture is richly expressed through region, dialect, history, food, politics, media, and everyday visual life, yet it remains underrepresented in multimodal evaluation. To address this gap, we introduce BanglaVerse, a culturally grounded benchmark for evaluating multilingual vision–language models (VLMs) on Bengali culture across historically linked languages and regional dialects. Built from 1,152 manually curated images across nine domains, the benchmark supports visual question answering and captioning, and is expanded into four languages and five Bangla dialects, yielding ∼32.3K artifacts. Our experiments show that evaluating only standard Bangla overestimates true model capability: performance drops under dialectal variation, especially for caption generation, while historically linked languages such as Hindi and Urdu retain some cultural meaning but remain weaker for structured reasoning. Across domains, the main bottleneck is missing cultural knowledge rather than visual grounding alone, with knowledge-intensive categories. These findings position BanglaVerse as a more realistic test bed for measuring culturally grounded multimodal understanding under linguistic variation.

---

Methodology

Figure 1: Overview of the BanglaVerse dataset and experimental setup. The figure shows the two task types, example annotations for each task, artifacts generation and evaluation pipeline with multiple metrics.

---

Dataset Summary

The benchmark is built from 1,152 manually curated images spanning nine culturally rich domains. It is expanded into four languages and five major Bangla dialects, resulting in approximately 32.3K total artifacts.

Core Statistics

• Images: 1,152 unique source images.
• Tasks: Supports Visual Question Answering (VQA) and Image Captioning (CAP).
• Total Artifacts: 10,377 captions and 20,727 VQA pairs (32,256 total).
• Languages: Bangla, English, Hindi, and Urdu.
• Bangla Dialects: Barishal, Chittagong, Noakhali, Rangpur, and Sylhet.

Cultural Domains

The dataset encapsulates Bengali identity through nine carefully selected domains: Culture, Food, History, Media & Movies, National Achievements, Nature, Personalities, Politics, and Sports.

---

Key Findings

• Dialectal Sensitivity: Evaluating only on standard Bangla overestimates model capability; performance drops consistently under dialectal variation, particularly for free-form caption generation.
• Cross-lingual Preservation: Historically linked languages like Hindi and Urdu preserve more cultural meaning in descriptive tasks than standard translation baselines suggest.
• The Knowledge Bottleneck: The primary limitation in understanding is missing cultural knowledge (e.g., named entities and specific context) rather than simple visual grounding or linguistic variation alone.

---

💻 All the codes, resources, and evaluation scripts are available on the Project Website.

💾 Downloading the Full Corpus

You can easily download and use the BANGLA VERSE dataset in multiple ways:

🔹 Option 1: Using the 🤗 Datasets Library

from datasets import load_dataset

# Load the dataset from Hugging Face
dataset = load_dataset("FaiyazAbdullah114708/BanglaVerse")

# Check available splits
print(dataset)

🔹 Option 2: Clone Directly from Hugging Face

You can also clone the repository directly using Git LFS:

git lfs install
git clone https://huggingface.co/datasets/FaiyazAbdullah114708/BanglaVerse

🔹 Option 3: Manual Download

Visit the dataset page and use the “Download Dataset” button: 👉 https://huggingface.co/datasets/FaiyazAbdullah114708/BanglaVerse

---

Citation

If you use this dataset or benchmark in your research, please cite:

@article{sayeedi2026many,
  title={Many Dialects, Many Languages, One Cultural Lens: Evaluating Multilingual VLMs for Bengali Culture Understanding Across Historically Linked Languages and Regional Dialects},
  author={Sayeedi, Nurul Labib and Sayeedi, Md Faiyaz Abdullah and Dipta, Shubhashis Roy and Tabassum, Rubaya and Hridoy, Ariful Ekraj and Mahmood, Mehraj and Sobhani, Mahbub E and Hasan, Md Tarek and Shatabda, Swakkhar},
  journal={arXiv preprint arXiv:2603.21165},
  year={2026}
}