This model is designed to perform transliteration from Banglish (Romanized Bengali) to Bengali script using the
facebook/mbart-large-50-many-to-many-mmt model. The training was conducted using the dataset
SKNahin/bengali-transliteration-data.
The notebook used for training can be found here:
Kaggle Notebook.
The model is intended for direct transliteration of Banglish text to Bengali script.
It can be integrated into NLP applications where transliteration from Banglish to Bengali is required, such as chatbots, text normalization, and digital content processing.
The model is not designed for language translation beyond transliteration, and it may not perform well on text containing mixed languages or code-switching.
Users should validate outputs, especially for critical applications, and consider further fine-tuning if necessary.
1from transformers import MBartForConditionalGeneration, MBartTokenizer
2
3model_name = "facebook/mbart-large-50-many-to-many-mmt"
4tokenizer = MBartTokenizer.from_pretrained(model_name)
5model = MBartForConditionalGeneration.from_pretrained(model_name)
6
7text = "ami tomake bhalobashi"
8inputs = tokenizer(text, return_tensors="pt")
9
10translated_tokens = model.generate(**inputs)
11output = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
12
13print(output) # Expected Bengali transliteration
14
The dataset used for training is
SKNahin/bengali-transliteration-data, which contains pairs of Banglish (Romanized Bengali) and corresponding Bengali script.
The model follows the Transformer-based Seq2Seq architecture from mBART.
1@inproceedings{SKNahin2023,
2 author = {SK Nahin},
3 title = {Bengali Transliteration Dataset},
4 year = {2023},
5 publisher = {Hugging Face Datasets},
6 url = {https://huggingface.co/datasets/SKNahin/bengali-transliteration-data}
7}
8
9@article{lewis2020mbart,
10 title={mBART: Multilingual Denoising Pre-training for Neural Machine Translation},
11 author={Lewis, Mike and others},
12 journal={arXiv preprint arXiv:2001.08210},
13 year={2020}
14}
