kishormorol/BanglaPoliticalStance

BanglaPoliticalStance

A large-scale Bangla multimodal dataset for political stance detection, containing 14,719 news items (headlines + photos) from 392 Bangladeshi news outlets.

Dataset Summary

BanglaPoliticalStance is the first large-scale multimodal Bangla political stance detection dataset. Each item consists of a news headline in Bangla and its accompanying photograph, collected from major Bangladeshi news portals. The dataset supports three-way stance classification:

Dataset Structure

The dataset has two splits:

annotated (198 items)

Expert-annotated by 3 human annotators with inter-annotator agreement of κ=0.73 (Cohen's kappa between the two primary annotators). These items include gold-standard labels for benchmarking.

unannotated (14,521 items)

Recently collected headlines and images from 392 Bangla news outlets, ready for annotation. These items do not have stance labels yet.

We welcome community contributions to annotate this data.

Features

Source Outlets (Top 20)

...and 370+ more outlets.

Important Notes

1. 1.The text is headlines, not articles. Median headline length is ~8 words. Results should be described as headline stance classification.

1. 1.Label IDs are frozen. 0 = govt_critique, 1 = neutral, 2 = govt_leaning. Do not renumber.

1. 1.Image-text stance can differ. In the annotated subset, article-level and image-level labels agree on only 47.4% of items — the photo often carries a different stance from the headline. This gap is the core argument for multimodal approaches.

1. 1.The annotated split has class imbalance. govt_critique (103) > neutral (53) > govt_leaning (42). This reflects the real distribution and should not be artificially balanced for evaluation.

Usage

from datasets import load_dataset

# Load annotated split (with labels)
ds = load_dataset("kishormorol/BanglaPoliticalStance", split="annotated")

# Load unannotated split (for annotation or self-supervised pretraining)
ds_new = load_dataset("kishormorol/BanglaPoliticalStance", split="unannotated")

# Example
print(ds[0]["headline"])  # Bangla headline
print(ds[0]["label"])     # 0, 1, or 2
ds[0]["image"].show()     # PIL Image

Citation

If you use this dataset, please cite:

@dataset{bangla_political_stance_2026,
  title={BanglaPoliticalStance: A Large-Scale Bangla Multimodal Political Stance Detection Dataset},
  author={Kishor Morol},
  year={2026},
  url={https://huggingface.co/datasets/kishormorol/BanglaPoliticalStance},
  note={14,719 Bangla news headlines with photographs from 392 outlets}
}

License

The annotations and metadata are released under CC BY-NC 4.0. The headlines and photographs belong to their respective news outlets — source_url records the provenance of each item. Please check each outlet's terms before redistributing article text or images.

Contributing

We welcome contributions to:

• —Annotate items in the unannotated split
• —Validate existing annotations
• —Report issues with data quality

Please open a discussion on this dataset's page if you'd like to contribute.