BanglaGEC: A Large-Scale Parallel Corpus for Bangla Grammatical Error Correction

BanglaGEC is a large-scale parallel corpus of 7,074,425 (~7.1M) sentence pairs for Bangla (Bengali) Grammatical Error Correction (GEC). Each pair maps a grammatically erroneous Bangla sentence to its grammatically correct counterpart, along with the error type, making it directly usable for training and evaluating sequence-to-sequence models, transformers, and large language models on the Bangla grammar correction task. The corpus covers ten distinct grammatical error types.

Dataset Summary

Property	Value
Language	Bangla / Bengali (bn)
Task	Grammatical Error Correction (text-to-text / seq2seq)
Total pairs	7,074,425 (~7.1M)
Train / Validation / Test	5,730,275 / 636,702 / 707,448
Error types	10 (see below)
Max sentence length	50 tokens

Languages

The dataset is entirely in Bangla (Bengali), ISO code bn, written in the Bengali script.

Dataset Structure

Split	File	# Instances
Train	train.csv	5,730,275 (~5.73M)
Validation	validation.csv	636,702 (~636.7K)
Test	test.csv	707,448 (~707.4K)

Data Fields

Each CSV has three columns:

• Correct (string): the grammatically correct Bangla sentence (reference / target output).
• Erroneous (string): the grammatically incorrect Bangla sentence (model input).
• ErrorType (string): the category of the introduced error (e.g., Bivokti, Bochon, Misplaced, Subject Missing).

Data Instance

{
  "Correct": "তৈমুরের দাপটে এলাকার ধনীরা আতংকিত হয়ে থাকতো।",
  "Erroneous": "তৈমুরের দাপটে এলাকাকে ধনীরা আতংকিত হয়ে থাকতো।",
  "ErrorType": "Bivokti"
}

Error Types

The ErrorType column labels each pair with one of ten grammatical error categories. Category distribution across the corpus:

Error type	# Instances	Percentage
Verb inflection	1,804,721	25.51%
Sentence structure	1,795,641	25.38%
Main verb missing	1,258,072	17.78%
Number (Bochon)	709,480	10.03%
Punctuation	523,255	7.40%
Auxiliary verb missing	237,741	3.36%
Subject–verb agreement	227,534	3.22%
Subject missing	197,540	2.79%
Verb (from the dictionary) missing	172,704	2.44%
Word choice (homonym error)	147,737	2.09%
Total	7,074,425	100%

How to Use

from datasets import load_dataset

dataset = load_dataset("nahid-hub/BanglaGEC")

train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

for row in train.select(range(3)):
    print("Erroneous:", row["Erroneous"])
    print("Correct:  ", row["Correct"])
    print("Type:     ", row["ErrorType"])
    print("---")

Citation

If you use the BanglaGEC corpus in your research, please cite:

@article{BanglaGEC,
  author={Hossain, Nahid and Bijoy, Mehedi Hasan and Islam, Salekul and Shatabda, Swakkhar},
  title={Panini: a transformer-based grammatical error correction method for Bangla},
  journal={Neural Computing and Applications},
  year={2024},
  month={Mar},
  day={01},
  volume={36},
  number={7},
  pages={3463-3477},
  issn={1433-3058},
  doi={10.1007/s00521-023-09211-7},
  url={https://doi.org/10.1007/s00521-023-09211-7}
}

License

This card lists cc-by-4.0 (attribution required).