This dataset is released for non-commercial research use only (CC BY-NC 4.0). Part of the text comes from public social media and remains the property of its original authors. If you are an author of any text here and want it removed, contact yeashahmed23@gmail.com. By requesting access you agree to use the data for non-commercial research only and to cite this dataset.

Bangla_Satire

A parallel Banglish (romanized Bangla) and Bangla-script satire sentence dataset. Some sentences are closer to humor or sarcasm than to satire (see Satire, humor and sarcasm).

Curated by: Yeas, Mowmita
Language: Bangla (bn), in Bangla script and in romanized form (Banglish)
License: CC BY-NC 4.0 (non-commercial research use, attribution required)
Size: 1,362 sentence pairs

Dataset Structure

Column	Description
`id`	Row index (1 to 1362), assigned after random shuffling
`sentence`	Banglish (romanized Bangla) text, lowercased, punctuation removed
`sentence_bn`	The same sentence in Bangla script, punctuation removed
`satire`	Label (1 = satire)

from datasets import load_dataset

ds = load_dataset("Yeashhh/Bangla_Satire")
print(ds["train"][0])

Dataset Creation

Source Data

The sentences come from three sources:

Self-written (40%): written by the dataset authors.
Social media (40%): collected manually from public Facebook pages and comments (Switzerland probashi, Earki, and other pages). The authors selected only text they judged to be satirical, cleaned it, and in some cases modified it before adding it to the dataset.
LLM-generated (~20%): generated with ChatGPT (ChatGPT Go).

Annotation

Labels come from the authors' own judgment: a sentence was included only if the authors considered it satire. To measure consistency, two annotators labeled a sample of 500 sentences and agreed on 72% of them (raw percentage agreement; a chance-corrected score such as Cohen's kappa was not computed). Disagreements were resolved by asking a third annotator to check and taking the majority vote.

Satire, humor and sarcasm

The satire column marks every row as satire. Some sentences also lean toward humor or sarcasm. These are not annotated separately, so users who want to use the data for humor or sarcasm tasks can do so, but should verify or annotate the relevant rows themselves.

Preprocessing

All text was lowercased, punctuation was removed (hyphens and dashes were replaced with a space, decimal points inside numbers were kept), whitespace was normalized, and rows were randomly shuffled (seed 42). No other changes were made to the wording.

Personal and Sensitive Information

Personally identifying information (such as usernames and names of private individuals) was removed by the authors. Some sentences refer to public figures or institutions in a satirical context.

Considerations for Use

Text collected from social media remains the property of its original authors. This dataset is released for non-commercial research use only.
Satire is subjective and annotator agreement was moderate (72% on a 500-sentence sample), so some labels may be noisy.
LLM-generated text is subject to the provider's terms of use.
If you are an author of any text in this dataset and want it removed, contact yeashahmed23@gmail.com and it will be taken down.
All rows are labeled satire = 1; the dataset contains no non-satire (negative) examples.

Citation

If you use this dataset, please cite:

@misc{yeas_mowmita_2026_bangla_satire,
  title        = {Bangla_Satire},
  author       = {Yeas and Mowmita},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Yeashhh/Bangla_Satire}}
}

Authors

Yeas (@Yeashhh)
Mowmita (@RownokMowmita)

Downloads last month: -

Total file size:

642 kB