bangla-bert-base / README.md
1
---
2
language: bn
3
tags:
4
- bert
5
- bengali
6
- bengali-lm
7
- bangla
8
license: mit
9
datasets:
10
- common_crawl
11
- wikipedia
12
- oscar
13
---
14
15
16
# Bangla BERT Base
17
A long way passed. Here is our **Bangla-Bert**! It is now available in huggingface model hub. 
18
19
[Bangla-Bert-Base](https://github.com/sagorbrur/bangla-bert) is a pretrained language model of Bengali language using mask language modeling described in [BERT](https://arxiv.org/abs/1810.04805) and it's github [repository](https://github.com/google-research/bert)
20
21
22
23
## Pretrain Corpus Details
24
Corpus was downloaded from two main sources:
25
26
* Bengali commoncrawl corpus downloaded from [OSCAR](https://oscar-corpus.com/)
27
* [Bengali Wikipedia Dump Dataset](https://dumps.wikimedia.org/bnwiki/latest/)
28
29
After downloading these corpora, we preprocessed it as a Bert format. which is one sentence per line and an extra newline for new documents. 
30
31
```
32
sentence 1
33
sentence 2
34
35
sentence 1
36
sentence 2
37
38
```
39
40
## Building Vocab
41
We used [BNLP](https://github.com/sagorbrur/bnlp) package for training bengali sentencepiece model with vocab size 102025. We preprocess the output vocab file as Bert format.
42
Our final vocab file availabe at [https://github.com/sagorbrur/bangla-bert](https://github.com/sagorbrur/bangla-bert) and also at [huggingface](https://huggingface.co/sagorsarker/bangla-bert-base) model hub.
43
44
## Training Details
45
* Bangla-Bert was trained with code provided in Google BERT's github repository (https://github.com/google-research/bert)
46
* Currently released model follows bert-base-uncased model architecture (12-layer, 768-hidden, 12-heads, 110M parameters)
47
* Total Training Steps: 1 Million
48
* The model was trained on a single Google Cloud TPU 
49
50
## Evaluation Results
51
52
### LM Evaluation Results
53
After training 1 million steps here are the evaluation results. 
54
55
```
56
global_step = 1000000
57
loss = 2.2406516
58
masked_lm_accuracy = 0.60641736
59
masked_lm_loss = 2.201459
60
next_sentence_accuracy = 0.98625
61
next_sentence_loss = 0.040997364
62
perplexity = numpy.exp(2.2406516) = 9.393331287442784
63
Loss for final step: 2.426227
64
65
```
66
67
### Downstream Task Evaluation Results
68
- Evaluation on Bengali Classification Benchmark Datasets
69
70
Huge Thanks to [Nick Doiron](https://twitter.com/mapmeld) for providing evaluation results of the classification task.
71
He used [Bengali Classification Benchmark](https://github.com/rezacsedu/Classification_Benchmarks_Benglai_NLP) datasets for the classification task.
72
Comparing to Nick's [Bengali electra](https://huggingface.co/monsoon-nlp/bangla-electra) and multi-lingual BERT, Bangla BERT Base achieves a state of the art result.
73
Here is the [evaluation script](https://github.com/sagorbrur/bangla-bert/blob/master/notebook/bangla-bert-evaluation-classification-task.ipynb).
74
75
76
| Model | Sentiment Analysis | Hate Speech Task | News Topic Task | Average |
77
| ----- | -------------------| ---------------- | --------------- | ------- |
78
| mBERT | 68.15 | 52.32 | 72.27 | 64.25 |
79
| Bengali Electra | 69.19 | 44.84 | 82.33 | 65.45 |
80
| Bangla BERT Base | 70.37 | 71.83 | 89.19 | 77.13 |
81
82
- Evaluation on [Wikiann](https://huggingface.co/datasets/wikiann) Datasets
83
84
We evaluated `Bangla-BERT-Base` with [Wikiann](https://huggingface.co/datasets/wikiann) Bengali NER datasets along with another benchmark three models(mBERT, XLM-R, Indic-BERT). </br>
85
`Bangla-BERT-Base` got a third-place where `mBERT` got first and `XML-R` got second place after training these models 5 epochs.
86
87
| Base Pre-trained Model | F1 Score | Accuracy |
88
| ----- | -------------------| ---------------- |
89
| [mBERT-uncased](https://huggingface.co/bert-base-multilingual-uncased) | 97.11 | 97.68 |
90
| [XLM-R](https://huggingface.co/xlm-roberta-base) | 96.22 | 97.03 |
91
| [Indic-BERT](https://huggingface.co/ai4bharat/indic-bert)| 92.66 | 94.74 |
92
| Bangla-BERT-Base | 95.57 | 97.49 |
93
94
All four model trained with [transformers-token-classification](https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/token_classification.ipynb) notebook.
95
You can find all models evaluation results [here](https://github.com/sagorbrur/bangla-bert/tree/master/evaluations/wikiann)
96
97
Also, you can check the below paper list. They used this model on their datasets.
98
* [DeepHateExplainer: Explainable Hate Speech Detection in Under-resourced Bengali Language](https://arxiv.org/abs/2012.14353)
99
* [Emotion Classification in a Resource Constrained Language Using Transformer-based Approach](https://arxiv.org/abs/2104.08613)
100
* [A Review of Bangla Natural Language Processing Tasks and the Utility of Transformer Models](https://arxiv.org/abs/2107.03844)
101
* [BanglaBERT: Combating Embedding Barrier in Multilingual Models for Low-Resource Language Understanding](https://arxiv.org/abs/2101.00204)
102
103
**NB: If you use this model for any NLP task please share evaluation results with us. We will add it here.** 
104
105
## Limitations and Biases
106
107
## How to Use
108
109
**Bangla BERT Tokenizer**
110
111
```py
112
from transformers import AutoTokenizer, AutoModel
113
114
bnbert_tokenizer = AutoTokenizer.from_pretrained("sagorsarker/bangla-bert-base")
115
text = "আমি বাংলায় গান গাই।"
116
bnbert_tokenizer.tokenize(text)
117
# ['আমি', 'বাংলা', '##য', 'গান', 'গাই', '।']
118
```
119
120
121
**MASK Generation**
122
123
You can use this model directly with a pipeline for masked language modeling:
124
125
```py
126
from transformers import BertForMaskedLM, BertTokenizer, pipeline
127
128
model = BertForMaskedLM.from_pretrained("sagorsarker/bangla-bert-base")
129
tokenizer = BertTokenizer.from_pretrained("sagorsarker/bangla-bert-base")
130
nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
131
for pred in nlp(f"আমি বাংলায় {nlp.tokenizer.mask_token} গাই।"):
132
  print(pred)
133
134
# {'sequence': '[CLS] আমি বাংলায গান গাই । [SEP]', 'score': 0.13404667377471924, 'token': 2552, 'token_str': 'গান'}
135
136
```
137
138
## Author
139
[Sagor Sarker](https://github.com/sagorbrur)
140
141
## Acknowledgements
142
143
* Thanks to Google [TensorFlow Research Cloud (TFRC)](https://www.tensorflow.org/tfrc) for providing the free TPU credits - thank you!
144
* Thank to all the people around, who always helping us to build something for Bengali.
145
146
## Reference
147
* https://github.com/google-research/bert
148
149
## Citation
150
If you find this model helpful, please cite.
151
152
```
153
@misc{Sagor_2020,
154
  title   = {BanglaBERT: Bengali Mask Language Model for Bengali Language Understading},
155
  author  = {Sagor Sarker},
156
  year    = {2020},
157
  url    = {https://github.com/sagorbrur/bangla-bert}
158
}
159
160
```
161
162
163
164
165
166