xlmindic-base-uniscript

XLMIndic Base Uniscript

This model is pretrained on a subset of the OSCAR corpus spanning 14 Indo-Aryan languages. Before pretraining this model we transliterate the text to ISO-15919 format using the Aksharamukha library. A demo of Aksharamukha library is hosted here where you can transliterate your text and use it on our model on the inference widget.

Model description

This model has the same configuration as the ALBERT Base v2 model. 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 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.
• Uses a Brahmic script.

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

Transliteration

The unique component of this model is that it takes in ISO-15919 transliterated text.

The motivation behind this is this. When two languages share vocabularies, a machine learning model can exploit that to learn good cross-lingual representations. However if these two languages use different writing scripts it is difficult for a model to make the connection. Thus if if we can write the two languages in a single script then it is easier for the model to learn good cross-lingual representation.

For many of the scripts currently in use, there are standard transliteration schemes to convert to the Latin script. In particular, for the Indic scripts the ISO-15919 transliteration scheme is designed to consistently transliterate texts written in different Indic scripts to the Latin script.

An example of ISO-15919 transliteration for a piece of Bangla text is the following:

Original: "রবীন্দ্রনাথ ঠাকুর এফআরএএস (৭ মে ১৮৬১ - ৭ আগস্ট ১৯৪১; ২৫ বৈশাখ ১২৬৮ - ২২ শ্রাবণ ১৩৪৮ বঙ্গাব্দ) ছিলেন অগ্রণী বাঙালি কবি, ঔপন্যাসিক, সংগীতস্রষ্টা, নাট্যকার, চিত্রকর, ছোটগল্পকার, প্রাবন্ধিক, অভিনেতা, কণ্ঠশিল্পী ও দার্শনিক।"

Transliterated: 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli kabi, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika.'

Another example for a piece of Hindi text is the following:

Original: "चूंकि मानव परिवार के सभी सदस्यों के जन्मजात गौरव और समान तथा अविच्छिन्न अधिकार की स्वीकृति ही विश्व-शान्ति, न्याय और स्वतन्त्रता की बुनियाद है"

Transliterated: "cūṁki mānava parivāra kē sabhī sadasyōṁ kē janmajāta gaurava aura samāna tathā avicchinna adhikāra kī svīkr̥ti hī viśva-śānti, nyāya aura svatantratā kī buniyāda hai"

Training procedure

Preprocessing

The texts are transliterated to ISO-15919 format using the Aksharamukha library. Then these 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.

>>> AutoModel.from_pretrained('ibraheemmoosa/xlmindic-base-uniscript', 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 benchmark dataset. We report the mean and standard deviation of nine fine-tuning runs for this model. We compare with an ablation model that do not use transliteration and is instead trained on original scripts.

IndicGLUE

Intended uses & limitations

This model is pretrained on Indo-Aryan languages. Thus it is intended to be used for downstream tasks on these languages. However, since Dravidian languages such as Malayalam, Telegu, Kannada etc share a lot of vocabulary with the Indo-Aryan languages, this model can potentially be used on those languages too (after transliterating the text to ISO-15919).

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 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

To use this model you will need to first install the Aksharamukha library.

pip install aksharamukha

Using this library you can transliterate any text wriiten in Indic scripts in the following way:

>>> from aksharamukha import transliterate
>>> text = "चूंकि मानव परिवार के सभी सदस्यों के जन्मजात गौरव और समान तथा अविच्छिन्न अधिकार की स्वीकृति ही विश्व-शान्ति, न्याय और स्वतन्त्रता की बुनियाद है"
>>> transliterated_text = transliterate.process('autodetect', 'ISO', text)
>>> transliterated_text
"cūṁki mānava parivāra kē sabhī sadasyōṁ kē janmajāta gaurava aura samāna tathā avicchinna adhikāra kī svīkr̥ti hī viśva-śānti, nyāya aura svatantratā kī buniyāda hai"

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

>>> from transformers import pipeline
>>> from aksharamukha import transliterate
>>> unmasker = pipeline('fill-mask', model='ibraheemmoosa/xlmindic-base-uniscript')
>>> 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।"
>>> transliterated_text = transliterate.process('Bengali', 'ISO', text)
>>> transliterated_text
'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli [MASK], aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama [MASK] puraskāra lābha karēna.'
>>> unmasker(transliterated_text)
[{'score': 0.39705055952072144,
  'token': 1500,
  'token_str': 'abhinētā',
  'sequence': 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli abhinētā, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama nōbēla puraskāra lābha karēna.'},
 {'score': 0.20499080419540405,
  'token': 3585,
  'token_str': 'kabi',
  'sequence': 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli kabi, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama nōbēla puraskāra lābha karēna.'},
 {'score': 0.1314290314912796,
  'token': 15402,
  'token_str': 'rājanētā',
  'sequence': 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli rājanētā, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama nōbēla puraskāra lābha karēna.'},
 {'score': 0.060830358415842056,
  'token': 3212,
  'token_str': 'kalākāra',
  'sequence': 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli kalākāra, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama nōbēla puraskāra lābha karēna.'},
 {'score': 0.035522934049367905,
  'token': 11586,
  'token_str': 'sāhityakāra',
  'sequence': 'rabīndranātha ṭhākura ēphaāraēēsa (7 mē 1861 - 7 āgasṭa 1941; 25 baiśākha 1268 - 22 śrābaṇa 1348 baṅgābda) chilēna agraṇī bāṅāli sāhityakāra, aupanyāsika, saṁgītasraṣṭā, nāṭyakāra, citrakara, chōṭagalpakāra, prābandhika, abhinētā, kaṇṭhaśilpī ō dārśanika. 1913 sālē gītāñjali kābyagranthēra iṁrēji anubādēra janya tini ēśīẏadēra madhyē sāhityē prathama nōbēla puraskāra lābha karēna.'}]

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 ([email protected])
• Mahmud Elahi Akhter ([email protected])
• Ashfia Binte Habib

BibTeX entry and citation info

@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}
}