Datasets:

kishormorol
/

BanglaPoliticalStance

language:
  - bn
license: cc-by-nc-4.0
task_categories:
  - text-classification
  - image-classification
  - zero-shot-classification
task_ids:
  - multi-class-classification
  - sentiment-analysis
tags:
  - political-stance-detection
  - bangla
  - bengali
  - multimodal
  - news
  - south-asia
  - bangladesh
  - nlp
  - vision-language
size_categories:
  - 10K<n<100K
pretty_name: 'BanglaPoliticalStance: Bangla Multimodal Political Stance Detection Dataset'
dataset_info:
  features:
    - name: item_id
      dtype: string
    - name: headline
      dtype: string
    - name: source_url
      dtype: string
    - name: outlet
      dtype: string
    - name: image
      dtype: image
    - name: date
      dtype: string
    - name: section
      dtype: string
    - name: label
      dtype:
        class_label:
          names:
            '0': govt_critique
            '1': neutral
            '2': govt_leaning
  splits:
    - name: annotated
      num_examples: 198
    - name: unannotated
      num_examples: 14521
configs:
  - config_name: default
    data_files:
      - split: annotated
        path: data/annotated-*
      - split: unannotated
        path: data/unannotated-*
citation: |
  @dataset{morol2026banglapoliticalstance,
    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}
  }

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:

Label	ID	Description
`govt_critique`	0	Critical of the government
`neutral`	1	Neutral reporting
`govt_leaning`	2	Favorable toward the government

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

Feature	Type	Description
`item_id`	string	Unique identifier (MD5 hash of URL)
`headline`	string	News headline in Bangla (median ~8 words)
`source_url`	string	Original article URL
`outlet`	string	News outlet name
`image`	image	Accompanying news photograph
`date`	string	Publication date (when available)
`section`	string	News section (politics, national, etc.)
`label`	ClassLabel	Stance label (annotated split only)

Source Outlets (Top 20)

Outlet	Articles
The Daily Star	996
Prothom Alo	604
Bangladesh Sangbad Sangstha (BSS)	464
BBC Bangla	230
Jugantor	182
Bangladesh Pratidin	166
Daily Ittefaq	142
The Business Standard	118
Jago News	112
bdnews24	102
Samakal	82
Kaler Kantho	79
Naya Diganta	74
Ajker Patrika	71
Bangla Tribune	67
NTV	65
Risingbd	64
Dhaka Post	96
Dhaka Mail	99
Daily Inqilab	98

...and 370+ more outlets.

Important Notes

The text is headlines, not articles. Median headline length is ~8 words. Results should be described as headline stance classification.
Label IDs are frozen. 0 = govt_critique, 1 = neutral, 2 = govt_leaning. Do not renumber.
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.
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.