Datasets:

Swarnadeep-28
/

bn_code_mix_sentiment_dataset

license: mit
task_categories:
  - text-classification
language:
  - bn
  - en
  - hi
pretty_name: Bengali-English Code-Mixed Sentiment Dataset
size_categories:
  - 10K<n<100K

Bengali-English Code-Mixed Sentiment Dataset

Dataset Summary

This dataset contains Bengali–English code-mixed social media text annotated for sentiment classification.
The primary goal is to support research and applications in code-mixed NLP, especially sentiment analysis in low-resource Indic languages.

The dataset combines and cleans multiple publicly available sources:

BnSentMix: Bengali–English code-mixed sentiment dataset
SentMix-3L: Multi-lingual code-mixed dataset (Bengali–English–Hindi subset used)
Kaggle code-mixed sentiment dataset

We unified the label scheme into three sentiment classes:

positive
negative
neutral

Supported Tasks

Text Classification / Sentiment Analysis
Code-Mixed NLP Research
Low-Resource Language Modeling

Languages

Bengali (code-mixed with Roman script)
English

Dataset Structure

Data Fields

text: (string) Input sentence in Bengali-English code-mixed text
label: (string) Sentiment label → positive, negative, neutral, mixed

Splits

train: 16,012 examples
validation: 2,002 examples
test: 2,002 examples

Example

{
  "text": "Aaj movie ta khub bhalo chilo! Totally loved it.",
  "label": "positive"
}

Usage

from datasets import load_dataset
ds = load_dataset("Swarnadeep-28/bn_code_mix_sentiment_dataset")
print(ds["train"][0])

Dataset Creation

Source datasets: BnSentMix, SentMix-3L, Kaggle
Preprocessing: label unification, text cleaning, removal of duplicates
License: MIT (respect original dataset licenses if reused)

Citation

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

@dataset{das2025_bn_code_mix_sentiment, author = {Swarnadeep Das}, title = {Bengali-English Code-Mixed Sentiment Dataset}, year = {2025}, url = {https://huggingface.co/datasets/Swarnadeep-28/bn_code_mix_sentiment_dataset} }

Limitations

Informal Romanized Bengali text may vary widely (spellings/slang).

Small proportion of neutral/mixed cases compared to positive/negative.

Not designed for toxic/abusive language detection.

Acknowledgements

This dataset builds on the work of:

BnSentMix
SentMix-3L
Kaggle Code-Mixed Dataset Contributors