File size: 6,794 Bytes
3d3e2f6
 
66112b8
 
 
02ca275
 
66112b8
 
f064dc4
66112b8
 
81f8b41
66112b8
 
 
8ff622e
66112b8
 
8aa37f3
66112b8
 
 
 
3f415c3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66112b8
 
 
8aa37f3
c2e4bdf
66112b8
 
 
 
 
 
 
 
8aa37f3
66112b8
8aa37f3
66112b8
 
3f415c3
81f8b41
 
3f415c3
 
66112b8
 
 
 
 
 
 
 
3f415c3
66112b8
 
 
81f8b41
66112b8
 
 
8aa37f3
 
 
 
66112b8
 
 
 
8ff622e
66112b8
8ff622e
 
66112b8
 
81f8b41
66112b8
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
---
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. [A more interactive & comparison demo is available here](https://huggingface.co/spaces/flax-community/roberta-hindi).

> This is part of the
[Flax/Jax Community Week](https://discuss.huggingface.co/t/pretrain-roberta-from-scratch-in-hindi/7091), organized by [Hugging Face](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(a combination of **mc4, oscar and indic-nlp** datasets)

### 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("हम आपके सुखद <mask> की कामना करते हैं")
[{'score': 0.3310680091381073,
  'sequence': 'हम आपके सुखद सफर की कामना करते हैं',
  'token': 1349,
  'token_str': ' सफर'},
 {'score': 0.15317578613758087,
  'sequence': 'हम आपके सुखद पल की कामना करते हैं',
  'token': 848,
  'token_str': ' पल'},
 {'score': 0.07826550304889679,
  'sequence': 'हम आपके सुखद समय की कामना करते हैं',
  'token': 453,
  'token_str': ' समय'},
 {'score': 0.06304813921451569,
  'sequence': 'हम आपके सुखद पहल की कामना करते हैं',
  'token': 404,
  'token_str': ' पहल'},
 {'score': 0.058322224766016006,
  'sequence': 'हम आपके सुखद अवसर की कामना करते हैं',
  'token': 857,
  'token_str': ' अवसर'}]
```

## 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 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>`. 
- We had to perform cleanup of **mC4** and **oscar** datasets by removing all non hindi (non Devanagari) characters from the datasets. 
- We tried to filter out evaluation set of WikiNER of [IndicGlue](https://indicnlp.ai4bharat.org/indic-glue/) benchmark by [manual labelling](https://github.com/amankhandelia/roberta_hindi/blob/master/wikiner_incorrect_eval_set.csv)  where the actual labels were not correct and modifying the [downstream evaluation dataset](https://github.com/amankhandelia/roberta_hindi/blob/master/utils.py). 


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).A randomized shuffle of combined dataset of **mC4, oscar** and other datasets listed above was used to train the model. Training logs are present in [wandb](https://wandb.ai/wandb/hf-flax-roberta-hindi).

## Evaluation Results

RoBERTa Hindi is evaluated on various 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
- Aman K ([amankhandelia](https://huggingface.co/amankhandelia))
- Haswanth Aekula ([hassiahk](https://huggingface.co/hassiahk))
- Kartik Godawat ([dk-crazydiv](https://huggingface.co/dk-crazydiv))
- Prateek Agrawal ([prateekagrawal](https://huggingface.co/prateekagrawal))
- Rahul Dev ([mlkorra](https://huggingface.co/mlkorra))


## Credits
Huge thanks to Hugging Face 🤗 & Google Jax/Flax team for such a wonderful community week, especially for providing such massive computing resources. 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>