|
--- |
|
widget: |
|
- text: "मुझे उनसे बात करना <mask> अच्छा लगा" |
|
- text: "हम आपके सुखद <mask> की कामना करते हैं" |
|
- text: "सभी अच्छी चीजों का एक <mask> होता है" |
|
--- |
|
|
|
# RoBERTa base model for Hindi language |
|
|
|
Pretrained model on Hindi language using a masked language modeling (MLM) objective. RoBERTa was introduced in |
|
[this paper](https://arxiv.org/abs/1907.11692) and first released in |
|
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/roberta). |
|
|
|
> This is part of the |
|
[Flax/Jax Community Week](https://discuss.huggingface.co/t/pretrain-roberta-from-scratch-in-hindi/7091), organized by [HuggingFace](https://huggingface.co/) and TPU usage sponsored by Google. |
|
|
|
## Model description |
|
|
|
RoBERTa Hindi is a transformers model pretrained on a large corpus of Hindi data in a self-supervised fashion. |
|
|
|
### How to use |
|
|
|
You can use this model directly with a pipeline for masked language modeling: |
|
```python |
|
>>> from transformers import pipeline |
|
>>> unmasker = pipeline('fill-mask', model='flax-community/roberta-hindi') |
|
>>> unmasker("मुझे उनसे बात करना <mask> अच्छा लगा") |
|
|
|
[{'score': 0.2096337080001831, |
|
'sequence': 'मुझे उनसे बात करना एकदम अच्छा लगा', |
|
'token': 1462, |
|
'token_str': ' एकदम'}, |
|
{'score': 0.17915162444114685, |
|
'sequence': 'मुझे उनसे बात करना तब अच्छा लगा', |
|
'token': 594, |
|
'token_str': ' तब'}, |
|
{'score': 0.15887945890426636, |
|
'sequence': 'मुझे उनसे बात करना और अच्छा लगा', |
|
'token': 324, |
|
'token_str': ' और'}, |
|
{'score': 0.12024253606796265, |
|
'sequence': 'मुझे उनसे बात करना लगभग अच्छा लगा', |
|
'token': 743, |
|
'token_str': ' लगभग'}, |
|
{'score': 0.07114479690790176, |
|
'sequence': 'मुझे उनसे बात करना कब अच्छा लगा', |
|
'token': 672, |
|
'token_str': ' कब'}] |
|
``` |
|
|
|
## Training data |
|
|
|
The RoBERTa Hindi model was pretrained on the reunion of the following datasets: |
|
- [OSCAR](https://huggingface.co/datasets/oscar) is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus using the goclassy architecture. |
|
- [mC4](https://huggingface.co/datasets/mc4) is a multilingual colossal, cleaned version of Common Crawl's web crawl corpus. |
|
- [IndicGLUE](https://indicnlp.ai4bharat.org/indic-glue/) is a natural language understanding benchmark. |
|
- [Samanantar](https://indicnlp.ai4bharat.org/samanantar/) is a parallel corpora collection for Indic language. |
|
- [Hindi Wikipedia Articles - 172k](https://www.kaggle.com/disisbig/hindi-wikipedia-articles-172k) is a dataset with cleaned 172k Wikipedia articles. |
|
- [Hindi Text Short and Large Summarization Corpus](https://www.kaggle.com/disisbig/hindi-text-short-and-large-summarization-corpus) is a collection of ~180k articles with their headlines and summary collected from Hindi News Websites. |
|
- [Hindi Text Short Summarization Corpus](https://www.kaggle.com/disisbig/hindi-text-short-summarization-corpus) is a collection of ~330k articles with their headlines collected from Hindi News Websites. |
|
- [Old Newspapers Hindi](https://www.kaggle.com/crazydiv/oldnewspapershindi) is a cleaned subset of HC Corpora newspapers. |
|
|
|
## Training procedure |
|
### Preprocessing |
|
|
|
The texts are tokenized using a byte version of Byte-Pair Encoding (BPE) and a vocabulary size of 50265. The inputs of |
|
the model take pieces of 512 contiguous token that may span over documents. The beginning of a new document is marked |
|
with `<s>` and the end of one by `</s>` |
|
The details of the masking procedure for each sentence are the following: |
|
- 15% of the tokens are masked. |
|
- In 80% of the cases, the masked tokens are replaced by `<mask>`. |
|
- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. |
|
- In the 10% remaining cases, the masked tokens are left as is. |
|
Contrary to BERT, the masking is done dynamically during pretraining (e.g., it changes at each epoch and is not fixed). |
|
|
|
### Pretraining |
|
The model was trained on Google Cloud Engine TPUv3-8 machine (with 335 GB of RAM, 1000 GB of hard drive, 96 CPU cores) **8 v3 TPU cores** for 42K steps with a batch size of 128 and a sequence length of 128. The |
|
optimizer used is Adam with a learning rate of 6e-4, \\(\beta_{1} = 0.9\\), \\(\beta_{2} = 0.98\\) and |
|
\\(\epsilon = 1e-6\\), a weight decay of 0.01, learning rate warmup for 24,000 steps and linear decay of the learning |
|
rate after. |
|
|
|
## Evaluation Results |
|
|
|
RoBERTa Hindi is evaluated on downstream tasks. The results are summarized below. |
|
|
|
| Task | Task Type | IndicBERT | HindiBERTa | Indic Transformers Hindi BERT | RoBERTa Hindi Guj San | RoBERTa Hindi | |
|
|-------------------------|----------------------|-----------|------------|-------------------------------|-----------------------|---------------| |
|
| BBC News Classification | Genre Classification | **76.44** | 66.86 | **77.6** | 64.9 | 73.67 | |
|
| WikiNER | Token Classification | - | 90.68 | **95.09** | 89.61 | **92.76** | |
|
| IITP Product Reviews | Sentiment Analysis | **78.01** | 73.23 | **78.39** | 66.16 | 75.53 | |
|
| IITP Movie Reviews | Sentiment Analysis | 60.97 | 52.26 | **70.65** | 49.35 | **61.29** | |
|
|
|
## Team Members |
|
- Kartik Godawat ([dk-crazydiv](https://huggingface.co/dk-crazydiv)) |
|
- Aman K ([amankhandelia](https://huggingface.co/amankhandelia)) |
|
- Haswanth Aekula ([hassiahk](https://huggingface.co/hassiahk)) |
|
- Rahul Dev ([mlkorra](https://huggingface.co/mlkorra)) |
|
- Prateek Agrawal ([prateekagrawal](https://huggingface.co/prateekagrawal)) |
|
|
|
## Credits |
|
Huge thanks to Huggingface 🤗 & Google Jax/Flax team for such a wonderful community week. Especially for providing such massive computing resource. Big thanks to [Suraj Patil](https://huggingface.co/valhalla) & [Patrick von Platen](https://huggingface.co/patrickvonplaten) for mentoring during the whole week. |
|
|
|
<img src=https://pbs.twimg.com/media/E443fPjX0AY1BsR.jpg:medium> |
