File size: 13,243 Bytes
f974d3f
 
 
 
 
 
 
 
 
 
 
c2fd454
 
 
 
 
f974d3f
 
 
 
 
 
 
 
 
6e57129
f974d3f
6e57129
 
 
 
 
 
 
4c90553
6e57129
 
 
 
f974d3f
 
3afa5a6
f974d3f
6e57129
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64aa11e
6e57129
 
 
64aa11e
 
 
 
 
 
6e57129
 
 
 
 
 
 
0476568
6e57129
 
 
 
0476568
 
6e57129
 
0476568
 
6e57129
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64aa11e
6e57129
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
055f61d
 
 
 
 
 
 
 
72a9dd3
 
 
 
 
 
 
 
 
 
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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
---
language:
- as
- bn
- gu
- hi
- mr
- ne
- or
- pa
- si
- sa
- bpy
- bh
- gom
- mai
license: apache-2.0
datasets:
- oscar
tags:
- multilingual
- albert
- masked-language-modeling
- sentence-order-prediction
- fill-mask
- xlmindic
- nlp
- indoaryan
- indicnlp
- iso15919
widget:
- text : 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি [MASK], ঔপন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'

co2_eq_emissions:
      emissions: 28.53
      source: "calculated using this webstie https://mlco2.github.io/impact/#compute"
      training_type: "pretraining"
      geographical_location: "NA"
      hardware_used: "TPUv3-8 for about 180 hours or 7.5 days"
---

# XLMIndic Base Multiscript

This model is identical in all aspects to [this model](https://huggingface.co/ibraheemmoosa/xlmindic-base-uniscript) except that we do not perform the ISO-15919 transliteration. Thus it is intended to serve as an ablation model for our study. See [this](https://huggingface.co/ibraheemmoosa/xlmindic-base-uniscript) to understand the details.

## Model description
This model has the same configuration as the [ALBERT Base v2 model](https://huggingface.co/albert-base-v2/). Specifically, this model has the following configuration:
- 12 repeating layers
- 128 embedding dimension
- 768 hidden dimension
- 12 attention heads
- 11M parameters
- 512 sequence length

## Training data
This model was pretrained on the [OSCAR](https://huggingface.co/datasets/oscar) dataset which is a medium sized multilingual corpus containing text from 163 languages. We select a subset of 14 languages based on the following criteria:
 - Belongs to the [Indo-Aryan language family](https://en.wikipedia.org/wiki/Indo-Aryan_languages).
 - Uses a [Brahmic script](https://en.wikipedia.org/wiki/Brahmic_scripts).
 
These are the 14 languages we pretrain this model on:
- Assamese
- Bangla
- Bihari
- Bishnupriya Manipuri
- Goan Konkani
- Gujarati
- Hindi
- Maithili
- Marathi
- Nepali
- Oriya
- Panjabi
- Sanskrit
- Sinhala

## Training procedure
### Preprocessing
The texts are  tokenized using SentencePiece and a vocabulary size of 50,000. The inputs of the model are
then of the form:
```
[CLS] Sentence A [SEP] Sentence B [SEP]
```
### Training
Training objective is the same as the original ALBERT. 
.
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.

The details of the sentence order prediction example generation procedure for each sentence are the following:
- Split the sentence into two parts A and B at a random index.
- With 50% probability swap the two parts.  

The model was pretrained on TPUv3-8 for 1M steps. We have checkpoints available at every 100k pretraining steps. These are available at different branches of this repository. You can load these checkpoints by passing the `revision` parameter. For example to load the checkpoint at 500k you can use the following code.

```python
>>> AutoModel.from_pretrained('ibraheemmoosa/xlmindic-base-multiscript', revision='checkpoint_500k')
```

## Evaluation results
We evaluated this model on the Indo-Aryan subset of languages (Panjabi, Oriya, Assamese, Bangla, Hindi, Marathi, Gujarati) from the [IndicGLUE](https://huggingface.co/datasets/indic_glue) benchmark dataset. We report the mean and standard deviation of nine fine-tuning runs for this model.

### IndicGLUE
Task | mBERT | XLM-R | IndicBERT-Base | XLMIndic-Base-Uniscript | XLMIndic-Base-Multiscript (This Model)
-----| ----- | ----- | ------ | ------- | --------
Wikipedia Section Title Prediction | 71.90 | 65.45 | 69.40 | **81.78 ± 0.60** | 77.17 ± 0.76
Article Genre Classification | 88.64 | 96.61 | 97.72 | **98.70 ± 0.29** | 98.30 ± 0.26
Named Entity Recognition (F1-score) | 71.29 | 62.18 | 56.69 | **89.85 ± 1.14** |  83.19 ± 1.58
BBC Hindi News Article Classification | 60.55 | 75.52 | 74.60 | **79.14 ± 0.60** | 77.28 ± 1.50
Soham Bangla News Article Classification | 80.23 | 87.6 | 78.45 | **93.89 ± 0.48** | 93.22 ± 0.49
INLTK Gujarati Headlines Genre Classification | - | - | **92.91** | 90.73 ± 0.75 | 90.41 ± 0.69
INLTK Marathi Headlines Genre Classification | - | - | **94.30** | 92.04 ± 0.47 | 92.21 ± 0.23
IITP Hindi Product Reviews Sentiment Classification | 74.57 | **78.97** | 71.32 | 77.18 ± 0.77 | 76.33 ± 0.84
IITP Hindi Movie Reviews Sentiment Classification | 56.77 | 61.61 | 59.03 | **66.34 ± 0.16** | 65.91 ± 2.20
MIDAS Hindi Discourse Type Classification | 71.20 | **79.94** | 78.44 | 78.54 ± 0.91 | 78.39 ± 0.33
Cloze Style Question Answering (Fill-mask task) | - | - | 37.16 | **41.54** | 38.21

## Intended uses & limitations
This model is pretrained on Indo-Aryan languages. Thus it is intended to be used for downstream tasks on these languages.
You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=xlmindic) to look for
fine-tuned versions on a task that interests you.
Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)
to make decisions, such as sequence classification, token classification or question answering. For tasks such as text
generation you should look at model like GPT2.

### How to use

Then you can use this model directly with a pipeline for masked language modeling:
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='ibraheemmoosa/xlmindic-base-multiscript')
>>> text = "রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি [MASK], ঔপন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।"
>>> unmasker(text)
[{'score': 0.34163928031921387,
  'token': 5399,
  'token_str': 'কবি',
  'sequence': 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি কবি, পন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'},
 {'score': 0.30519795417785645,
  'token': 33436,
  'token_str': 'people',
  'sequence': 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি people, পন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'},
 {'score': 0.29130080342292786,
  'token': 30476,
  'token_str': 'সাহিত্যিক',
  'sequence': 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি সাহিত্যিক, পন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'},
 {'score': 0.031051287427544594,
  'token': 6139,
  'token_str': 'লেখক',
  'sequence': 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি লেখক, পন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'},
 {'score': 0.002705035964027047,
  'token': 38443,
  'token_str': 'শিল্পীরা',
  'sequence': 'রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি শিল্পীরা, পন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক। ১৯১৩ সালে গীতাঞ্জলি কাব্যগ্রন্থের ইংরেজি অনুবাদের জন্য তিনি এশীয়দের মধ্যে সাহিত্যে প্রথম নোবেল পুরস্কার লাভ করেন।'}]
```
### Limitations and bias
Even though we pretrain on a comparatively large multilingual corpus the model may exhibit harmful gender, ethnic and political bias. If you fine-tune this model on a task where these issues are important you should take special care when relying on the model to make decisions.

## Contact
Feel free to contact us if you have any ideas or if you want to know more about our models.
- Ibraheem Muhammad Moosa (ibraheemmoosa1347@gmail.com)
- Mahmud Elahi Akhter (mahmud.akhter01@northsouth.edu)
- Ashfia Binte Habib

## BibTeX entry and citation info

```bibtex
@article{Moosa2022DoesTH,
  title={Does Transliteration Help Multilingual Language Modeling?},
  author={Ibraheem Muhammad Moosa and Mahmuda Akhter and Ashfia Binte Habib},
  journal={ArXiv},
  year={2022},
  volume={abs/2201.12501}
}
```