csebuetnlp/banglat5_small

BanglaT5

This repository contains the pretrained checkpoint of the model BanglaT5 (small). This is a sequence to sequence transformer model pretrained with the "Span Corruption" objective. Finetuned models using this checkpoint achieve state-of-the-art results on many of the NLG tasks in bengali.

For finetuning on different downstream tasks such as Machine Translation, Abstractive Text Summarization, Question Answering etc., refer to the scripts in the official GitHub repository.

Note: This model was pretrained using a specific normalization pipeline available here. All finetuning scripts in the official GitHub repository use this normalization by default. If you need to adapt the pretrained model for a different task make sure the text units are normalized using this pipeline before tokenizing to get best results. A basic example is given below:

Using this model in transformers (tested on 4.11.0.dev0)

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from normalizer import normalize # pip install git+https://github.com/csebuetnlp/normalizer

model = AutoModelForSeq2SeqLM.from_pretrained("csebuetnlp/banglat5_small")
tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/banglat5_small", use_fast=False)

input_sentence = ""
input_ids = tokenizer(normalize(input_sentence), return_tensors="pt").input_ids
generated_tokens = model.generate(input_ids)
decoded_tokens = tokenizer.batch_decode(generated_tokens)[0]

print(decoded_tokens)

Benchmarks

• Supervised fine-tuning

Model	Params	MT (SacreBLEU)	TS (ROUGE-2)	QA (EM/F1)	MD (SacreBLEU-1)	NHG (ROUGE-2)	XLS (ROUGE-2)	BNLG score
mT5 (base)	582M	36.6/22.5	10.3	59.0/65.3	17.5	9.6	2.7/0.7	24.9
XLM-ProphetNet	616M	23.3/16.4	7.8	53.0/57.3	20.0	9.5	6.2/2.7	21.8
mBART-50	611M	23.6/16.7	10.4	53.4/58.9	18.5	11.2	5.4/3.7	22.4
IndicBART	244M	22.7/13.1	8.1	53.3/58.8	14.8	7.9	6.3/2.5	20.8
BanglaT5	247M	38.8/25.2	13.7	68.5/74.8	19.0	13.8	6.4/4.0	29.4

The benchmarking datasets are as follows:

• MT: Machine Translation
• TS: Abstractive Text Summarization
• QA: Question Answering
• MD: Multi Turn Dialogue Generation
• NHG: News Headline Generation
• XLS: Cross-lingual Summarization

Citation

If you use this model, please cite the following paper:

@article{bhattacharjee2022banglanlg,
  author    = {Abhik Bhattacharjee and Tahmid Hasan and Wasi Uddin Ahmad and Rifat Shahriyar},
  title     = {BanglaNLG: Benchmarks and Resources for Evaluating Low-Resource Natural Language Generation in Bangla},
  journal   = {CoRR},
  volume    = {abs/2205.11081},
  year      = {2022},
  url       = {https://arxiv.org/abs/2205.11081},
  eprinttype = {arXiv},
  eprint    = {2205.11081}
}

If you use the normalization module, please cite the following paper:

@inproceedings{hasan-etal-2020-low,
    title = "Not Low-Resource Anymore: Aligner Ensembling, Batch Filtering, and New Datasets for {B}engali-{E}nglish Machine Translation",
    author = "Hasan, Tahmid  and
      Bhattacharjee, Abhik  and
      Samin, Kazi  and
      Hasan, Masum  and
      Basak, Madhusudan  and
      Rahman, M. Sohel  and
      Shahriyar, Rifat",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.207",
    doi = "10.18653/v1/2020.emnlp-main.207",
    pages = "2612--2623",
    abstract = "Despite being the seventh most widely spoken language in the world, Bengali has received much less attention in machine translation literature due to being low in resources. Most publicly available parallel corpora for Bengali are not large enough; and have rather poor quality, mostly because of incorrect sentence alignments resulting from erroneous sentence segmentation, and also because of a high volume of noise present in them. In this work, we build a customized sentence segmenter for Bengali and propose two novel methods for parallel corpus creation on low-resource setups: aligner ensembling and batch filtering. With the segmenter and the two methods combined, we compile a high-quality Bengali-English parallel corpus comprising of 2.75 million sentence pairs, more than 2 million of which were not available before. Training on neural models, we achieve an improvement of more than 9 BLEU score over previous approaches to Bengali-English machine translation. We also evaluate on a new test set of 1000 pairs made with extensive quality control. We release the segmenter, parallel corpus, and the evaluation set, thus elevating Bengali from its low-resource status. To the best of our knowledge, this is the first ever large scale study on Bengali-English machine translation. We believe our study will pave the way for future research on Bengali-English machine translation as well as other low-resource languages. Our data and code are available at https://github.com/csebuetnlp/banglanmt.",
}