Bangla BERT Base

A long way passed. Here is our Bangla-Bert ! It is now available in huggingface model hub.

Bangla-Bert-Base is a pretrained language model of Bengali language using mask language modeling described in BERT and it's github repository

Pretrain Corpus Details

Corpus was downloaded from two main sources:

• Bengali commoncrawl corpus downloaded from OSCAR
• Bengali Wikipedia Dump Dataset

After downloading these corpora, we preprocessed it as a Bert format. which is one sentence per line and an extra newline for new documents.

sentence 1
sentence 2

sentence 1
sentence 2

Building Vocab

We used BNLP package for training bengali sentencepiece model with vocab size 102025. We preprocess the output vocab file as Bert format. Our final vocab file availabe at https://github.com/sagorbrur/bangla-bert and also at huggingface model hub.

Training Details

• Bangla-Bert was trained with code provided in Google BERT's github repository ( https://github.com/google-research/bert )
• Currently released model follows bert-base-uncased model architecture (12-layer, 768-hidden, 12-heads, 110M parameters)
• Total Training Steps: 1 Million
• The model was trained on a single Google Cloud GPU

Evaluation Results

LM Evaluation Results

After training 1 million steps here are the evaluation results.

global_step = 1000000
loss = 2.2406516
masked_lm_accuracy = 0.60641736
masked_lm_loss = 2.201459
next_sentence_accuracy = 0.98625
next_sentence_loss = 0.040997364
perplexity = numpy.exp(2.2406516) = 9.393331287442784
Loss for final step: 2.426227

Downstream Task Evaluation Results

• Evaluation on Bengali Classification Benchmark Datasets

Huge Thanks to Nick Doiron for providing evaluation results of the classification task. He used Bengali Classification Benchmark datasets for the classification task. Comparing to Nick's Bengali electra and multi-lingual BERT, Bangla BERT Base achieves a state of the art result. Here is the evaluation script .

• Evaluation on Wikiann Datasets

We evaluated Bangla-BERT-Base with Wikiann Bengali NER datasets along with another benchmark three models(mBERT, XLM-R, Indic-BERT). Bangla-BERT-Base got a third-place where mBERT got first and XML-R got second place after training these models 5 epochs.

All four model trained with transformers-token-classification notebook. You can find all models evaluation results here

Also, you can check the below paper list. They used this model on their datasets.

• DeepHateExplainer: Explainable Hate Speech Detection in Under-resourced Bengali Language
• Emotion Classification in a Resource Constrained Language Using Transformer-based Approach
• A Review of Bangla Natural Language Processing Tasks and the Utility of Transformer Models

NB: If you use this model for any NLP task please share evaluation results with us. We will add it here.

Limitations and Biases

How to Use

Bangla BERT Tokenizer

from transformers import AutoTokenizer, AutoModel

bnbert_tokenizer = AutoTokenizer.from_pretrained("sagorsarker/bangla-bert-base")
text = "আমি বাংলায় গান গাই।"
bnbert_tokenizer.tokenize(text)
# ['আমি', 'বাংলা', '##য', 'গান', 'গাই', '।']

MASK Generation

You can use this model directly with a pipeline for masked language modeling:

from transformers import BertForMaskedLM, BertTokenizer, pipeline

model = BertForMaskedLM.from_pretrained("sagorsarker/bangla-bert-base")
tokenizer = BertTokenizer.from_pretrained("sagorsarker/bangla-bert-base")
nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
for pred in nlp(f"আমি বাংলায় {nlp.tokenizer.mask_token} গাই।"):
  print(pred)

# {'sequence': '[CLS] আমি বাংলায গান গাই । [SEP]', 'score': 0.13404667377471924, 'token': 2552, 'token_str': 'গান'}

Author

Sagor Sarker

Reference

• https://github.com/google-research/bert

Citation

If you find this model helpful, please cite.

@misc{Sagor_2020,
  title   = {BanglaBERT: Bengali Mask Language Model for Bengali Language Understading},
  author  = {Sagor Sarker},
  year    = {2020},
  url    = {https://github.com/sagorbrur/bangla-bert}
}