A multilingual Grammatical Error Correction (GEC) model fine-tuned from mT5-small for Bangla (Bengali). Developed as part of the BHASHA 2025 Shared Task 1: IndicGEC.

Developed by: Manav Dhamecha, Gaurav Damor, Sunil Choudhary, Pruthwik Mishra
License: MIT
Base model: google/mt5-small
Paper: Team Horizon at BHASHA Task 1
Repository: manavdhamecha77/IndicGEC2025
GitHub.io: Multilingual IndicGEC

What it does

Given a grammatically incorrect Bangla sentence, the model outputs a corrected version. It handles errors across spelling, grammar (tense, person, number, gender, case), punctuation, missing/extra words, and semantic issues.

GLEU Score on Bangla test set: 82.69

Quick Start

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

model_name = "manavdhamecha77/GEC-mT5-Small-Bangla"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

sentences = [
    "আমি স্কুলে যাই",
    "সে বাড়ি গেছিলাম",
]

inputs = ["correct this: " + s for s in sentences]

encoded = tokenizer(
    inputs,
    return_tensors="pt",
    padding=True,
    truncation=True,
    max_length=128
).to(device)

outputs = model.generate(**encoded, max_length=128, num_beams=4)
corrected = tokenizer.batch_decode(outputs, skip_special_tokens=True)

for orig, corr in zip(sentences, corrected):
    print(f"Input:     {orig}")
    print(f"Corrected: {corr}\n")

Training Details

The model was fine-tuned using a sequence-to-sequence objective on parallel noisy–clean sentence pairs. Training data was expanded from ~598 annotated pairs to ~10k pairs using a synthetic error injection pipeline that introduces realistic errors across 10 linguistic categories.

Parameter	Value
Optimizer	AdamW
Learning Rate	5e-5
Batch Size	16–32
Epochs	10–15
Max Sequence Length	128
Early Stopping	Based on GLEU (dev set)

Input format: "correct this: <incorrect sentence>"

Evaluation

Language	Model	GLEU
Bangla	mT5-small	82.69

Limitations

Performance may degrade on heavy code-mixing, informal slang, or dialectal text.
Trained primarily on formal written Bangla; may not generalize to all domains.
Evaluation uses automatic metrics (GLEU) only; human evaluation not conducted.

Citation

bibtex

@inproceedings{dhamecha2025horizon,
  title     = {Team Horizon at {BHASHA} Task 1: Multilingual {IndicGEC} with Transformer-based Grammatical Error Correction Models},
  author    = {Dhamecha, Manav and Damor, Gaurav and Choudhary, Sunil and Mishra, Pruthwik},
  booktitle = {Proceedings of the 1st Workshop on Benchmarks, Harmonization, Annotation, and Standardization for Human-Centric AI in Indian Languages (BHASHA 2025)},
  year      = {2025},
  url       = {https://aclanthology.org/2025.bhasha-1.14/}
}