Indic BERT

Website | Downloads | Paper

Indic bert is a multilingual ALBERT model that exclusively covers 12 major Indian languages. It is pre-trained on our novel corpus of around 9 billion tokens and evaluated on a set of diverse tasks. Indic-bert has around 10x fewer parameters than other popular publicly available multilingual models while it also achieves a performance on-par or better than these models.

We also introduce IGLUE - a set of standard evaluation tasks that can be used to measure the NLU performance of monolingual and multilingual models on Indian languages. Along with IGLUE, we also compile a list of additional evaluation tasks. This repository contains code for running all these evaluation tasks on indic-bert and other bert-like models.

Table of Contents

• Introduction
• Setting up the Code
• Running Experiments
• Pretraining Corpus
• IndicGLUE
• Additional Evaluation Tasks
• Evaluation Results
• Downloads
• Citing
• License
• Contributors
• Contact

Introduction

The Indic BERT model is based on the ALBERT model, a recent derivative of BERT. It is pre-trained on 12 Indian languages: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, Telugu.

The easiest way to use Indic BERT is through the Huggingface transformers library. It can be simply loaded like this:

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('ai4bharat/indic-bert')
model = AutoModel.from_pretrained('ai4bharat/indic-bert')

Setting up the Code

The code can be run on GPU, TPU or on Google's Colab platform. If you want to run it on Colab, you can simply use our fine-tuning notebook . For running it in your own VM, start with running the following commands:

git clone https://github.com/AI4Bharat/indic-bert
cd indic-bert
sudo pip3 install -r requirements.txt

By default, the installation will use GPU. For TPU support, first update your .bashrc with the following variables:

export PYTHONPATH="${PYTHONPATH}:/usr/share/tpu/models:<path to this repo"
export PYTHONIOENCODING=utf-8
export TPU_IP_ADDRESS="<TPU Internal Address"
export TPU_NAME="grpc://$TPU_IP_ADDRESS:8470"
export XRT_TPU_CONFIG="tpu_worker;0;$TPU_IP_ADDRESS:8470"
export LD_LIBRARY_PATH="/usr/local/lib"

Then, install pytorch-xla:

curl https://raw.githubusercontent.com/pytorch/xla/master/contrib/scripts/env-setup.py -o pytorch-xla-env-setup.py
sudo python3 pytorch-xla-env-setup.py --version nightly --apt-packages libomp5 libopenblas-dev

Running Experiments

To get help, simply run:

python3 -m fine_tune.cli --help

To evaluate a specific model with default hyper-parameters, execute:

python3 -m fine_tune.cli --model <model name> --dataset <dataset name> --lang <iso lang code> --iglue_dir <base path to indic glue dir> --output_dir <output dir>

Pretraining Corpus

We pre-trained indic-bert on AI4Bharat's monolingual corpus. The corpus has the following distribution of languages:

Language	as	bn	en	gu	hi	kn
No. of Tokens	36.9M	815M	1.34B	724M	1.84B	712M
Language	ml	mr	or	pa	ta	te	all
No. of Tokens	767M	560M	104M	814M	549M	671M	8.9B

IndicGLUE

IGLUE is a natural language understanding benchmark for Indian languages that we propose. While building this benchmark, our objective was also to cover most of the 11 Indian languages for each task. It consists of the following tasks:

News Category Classification

Predict the genre of a given news article. The dataset contains around 125k news articles across 9 Indian languages. Example:

Article Snippet:

கர்நாடக சட்டப் பேரவையில் வெற்றி பெற்ற எம்எல்ஏக்கள் இன்று பதவியேற்றுக் கொண்ட நிலையில் , காங்கிரஸ் எம்எல்ஏ ஆனந்த் சிங் க்கள் ஆப்சென்ட் ஆகி அதிர்ச்சியை ஏற்படுத்தியுள்ளார் . உச்சநீதிமன்ற உத்தரவுப்படி இன்று மாலை முதலமைச்சர் எடியூரப்பா இன்று நம்பிக்கை வாக்கெடுப்பு நடத்தி பெரும்பான்மையை நிரூபிக்க உச்சநீதிமன்றம் உத்தரவிட்டது . 

Category: Politics

Named Entity Recognition

Recognize entities and their coarse types in a sequence of words. The dataset contains around 787k examples across 11 Indian languages.

Example:

Token	चाणक्य	पुरी	को	यहाँ	देखने	हेतु	यहाँ	क्लिक	करें
Type	B-LOC	I-LOC	O	O	O	O	O	O	O

Headline Prediction

Predict the correct headline for a news article from a given list of four candidate headlines. The dataset contains around 880k examples across 11 Indian languages. Example:

News Article:

 ರಾಷ್ಟ್ರೀಯ\nಪುಣೆ: 23 ವರ್ಷದ ಇನ್ಫೋಸಿಸ್ ಮಹಿಳಾ ಟೆಕ್ಕಿಯೊಬ್ಬರನ್ನು ನಡು ರಸ್ತೆಯಲ್ಲಿಯೇ ಮಾರಾಕಾಸ್ತ್ರಗಳಿಂದ ಬರ್ಬರವಾಗಿ ಹತ್ಯೆ ಮಾಡಿರುವ ಘಟನೆ ಪುಣೆಯಲ್ಲಿ ಶನಿವಾರ ರಾತ್ರಿ ನಡೆದಿದೆ.\nಅಂತರ ದಾಸ್ ಕೊಲೆಯಾದ ಮಹಿಳಾ ಟೆಕ್ಕಿಯಾಗಿದ್ದಾರೆ. ಅಂತರಾ ಅವರು ಪಶ್ಚಿಮ ಬಂಗಾಳದ ಮೂಲದವರಾಗಿದ್ದಾರೆ. ಕಳೆದ ರಾತ್ರಿ 8.00 ಗಂಟೆ ಸುಮಾರಿಗೆ ಕೆಲಸ ಮುಗಿಸಿ ಮನೆಗೆ ತೆರಳುತ್ತಿದ್ದ ಸಂದರ್ಭದಲ್ಲಿ ಅಂತರಾ ಅವರ ಮೇಲೆ ದಾಳಿ ಮಾಡಿರುವ ದುಷ್ಕರ್ಮಿಗಳು ಮಾರಾಕಾಸ್ತ್ರಗಳಿಂದ ಹಲ್ಲೆ ನಡೆಸಿದ್ದಾರೆಂದು ಪೊಲೀಸರು ಹೇಳಿದ್ದಾರೆ.\nದಾಳಿ ನಡೆಸಿದ ನಂತರ ರಕ್ತದ ಮಡುವಿನಲ್ಲಿ ಬಿದ್ದು ಒದ್ದಾಡುತ್ತಿದ್ದ ಅಂತರಾ ಅವರನ್ನು ಸ್ಥಳೀಯರು ಆಸ್ಪತ್ರೆಗೆ ದಾಳಸಿದ್ದಾರೆ. ಆದರೆ, ಆಸ್ಪತ್ರೆಗೆ ದಾಖಲಿಸುವಷ್ಟರಲ್ಲಿ ಅಂತರಾ ಅವರು ಸಾವನ್ನಪ್ಪಿದ್ದಾರೆಂದು ಅವರು ಹೇಳಿದ್ದಾರೆ.\nಪ್ರಕರಣ ದಾಖಲಿಸಿಕೊಂಡಿರುವ ಪೊಲೀಸರು ತನಿಖೆ ಆರಂಭಿಸಿದ್ದಾರೆ",

Candidate 1: ಇನ್ಫೋಸಿಸ್ ಮಹಿಳಾ ಟೆಕ್ಕಿಯ ಬರ್ಬರ ಹತ್ಯೆ [correct answer] Candidate 2: ಮಾನಸಿಕ ಅಸ್ವಸ್ಥೆ ಮೇಲೆ ಮಕ್ಕಳ ಕಳ್ಳಿ ಎಂದು ಭೀಕರ ಹಲ್ಲೆ Candidate 3: ಕಸಬ ಬೆಂಗ್ರೆಯಲ್ಲಿ ಮುಸುಕುಧಾರಿಗಳ ತಂಡದಿಂದ ಮೂವರು ಯುವಕರ ಮೇಲೆ ಹಲ್ಲೆ : ಓರ್ವ ಗಂಭೀರ Candidate 4: ಕಣಿವೆ ರಾಜ್ಯದಲ್ಲಿ mobile ಬಂದ್, ಪ್ರಿಂಟಿಂಗ್ ಪ್ರೆಸ್ ಮೇಲೆ ದಾಳಿ

Wikipedia Section Title Prediction

Predict the correct title for a Wikipedia section from a given list of four candidate titles. The dataset has 400k examples across 11 Indian languages.

Section Text:

2005માં, જેકમેન નિર્માણ કંપની, સીડ પ્રોડકશન્સ ઊભી કરવા તેના લાંબાસમયના મદદનીશ જહોન પાલેર્મો સાથે જોડાયા, જેમનો પ્રથમ પ્રોજેકટ 2007માં વિવા લાફલિન હતો. જેકમેનની અભિનેત્રી પત્ની ડેબોરા-લી ફર્નેસ પણ કંપનીમાં જોડાઈ, અને પાલેર્મોએ પોતાના, ફર્નેસ અને જેકમેન માટે “ યુનિટી ” અર્થવાળા લખાણની આ ત્રણ વીંટીઓ બનાવી.[૨૭] ત્રણેયના સહયોગ અંગે જેકમેને જણાવ્યું કે “ મારી જિંદગીમાં જેમની સાથે મેં કામ કર્યું તે ભાગીદારો અંગે ડેબ અને જહોન પાલેર્મો અંગે હું ખૂબ નસીબદાર છું. ખરેખર તેથી કામ થયું. અમારી પાસે જુદું જુદું સાર્મથ્ય હતું. હું તે પસંદ કરતો હતો. I love it. તે ખૂબ ઉત્તેજક છે. ”[૨૮]ફોકસ આધારિત સીડ લેબલ, આમન્ડા સ્કિવેઈટઝર, કેથરિન ટેમ્બલિન, એલન મંડેલબમ અને જોય મરિનો તેમજ સાથે સિડની આધારિત નિર્માણ કચેરીનું સંચાલન કરનાર અલાના ફ્રીનો સમાવેશ થતાં કદમાં વિસ્તૃત બની. આ કંપીનોનો ઉદ્દેશ જેકમેનના વતનના દેશની સ્થાનિક પ્રતિભાને કામે લેવા મધ્યમ બજેટવાળી ફિલ્મો બનાવવાનો છે. 

Candidate 1: એકસ-મેન

Candidate 2: કારકીર્દિ

Candidate 3: નિર્માણ કંપન [correct answer]

Candidate 4: ઓસ્ટ્રેલિય

Cloze-style Question Answering (WCQA)

Given a text with an entity randomly masked, the task is to predict that masked entity from a list of 4 candidate entities. The dataset contains around 239k examples across 11 languages. Example:

Text

ਹੋਮੀ ਭਾਬਾ ਦਾ ਜਨਮ 1949 ਈ ਨੂਂ ਮੁੰਬਈ ਵਿੱਚ ਪਾਰਸੀ ਪਰਿਵਾਰ ਵਿੱਚ ਹੋਇਆ । ਸੇਂਟ ਮੇਰੀ ਤੋਂ ਮੁਢਲੀ ਸਿਖਿਆ ਪ੍ਰਾਪਤ ਕਰਕੇ ਉਹ ਬੰਬੇ ਯੂਨੀਵਰਸਿਟੀ ਗ੍ਰੈਜੁਏਸ਼ਨ ਲਈ ਚਲਾ ਗਿਆ । ਇਸ ਤੋਂ ਬਾਅਦ ਉਹ ਉਚੇਰੀ ਸਿਖਿਆ ਲਈ <MASK> ਚਲਾ ਗਿਆ । ਉਸਨੇ ਓਥੇ ਆਕਸਫੋਰਡ ਯੂਨੀਵਰਸਿਟੀ ਤੋਂ ਐਮ.ਏ ਅਤੇ ਐਮ ਫਿਲ ਦੀਆਂ ਡਿਗਰੀਆਂ ਪ੍ਰਾਪਤ ਕੀਤੀਆਂ । ਤਕਰੀਬਨ ਦਸ ਸਾਲ ਤਕ ਉਸਨੇ ਸੁਸੈਕਸ ਯੂਨੀਵਰਸਿਟੀ ਦੇ ਅੰਗਰੇਜ਼ੀ ਵਿਭਾਗ ਵਿੱਚ ਬਤੌਰ ਲੈਕਚਰਾਰ ਕਾਰਜ ਨਿਭਾਇਆ । ਇਸਤੋਂ ਇਲਾਵਾ ਹੋਮੀ ਭਾਬਾ ਪੈਨਸੁਲਵੇਨਿਆ , ਸ਼ਿਕਾਗੋ ਅਤੇ ਅਮਰੀਕਾ ਦੀ ਹਾਰਵਰਡ ਯੂਨੀਵਰਸਿਟੀ ਵਿੱਚ ਵੀ ਪ੍ਰੋਫ਼ੇਸਰ ਦੇ ਆਹੁਦੇ ਤੇ ਰਿਹਾ ।

Candidate 1: ਬਰਤਾਨੀਆ [correct answer] Candidate 2: ਭਾਰਤ Candidate 3: ਸ਼ਿਕਾਗੋ Candidate 4: ਪਾਕਿਸਤਾਨ

Cross-lingual Sentence Retrieval (XSR)

Given a sentence in language $L_1$ the task is to retrieve its translation from a set of candidate sentences in language $L_2$. The dataset contains around 39k parallel sentence pairs across 8 Indian languages. Example:

Input Sentence

In the health sector the nation has now moved ahead from the conventional approach.

Retrieve the following translation from a set of 4886 sentences:

ആരോഗ്യമേഖലയില് ഇന്ന് രാജ്യം പരമ്പരാഗത രീതികളില് നിന്ന് മുന്നേറിക്കഴിഞ്ഞു.

Additional Evaluation Tasks

Natural Language Inference

• Winnograd Natural Language Inference (WNLI)
• Choice of Plausible Alternatives (COPA)

Sentiment Analysis

• IITP Movie Reviews Sentiment
• IITP Product Reviews
• ACTSA Sentiment Classifcation

Genre Classification

• Soham Articles Genre Classification
• iNLTK Headlines Genre Classifcation
• BBC News Articles

Discourse Analysis

• MIDAS Discourse

Evaluation Results

IndicGLUE

Task	mBERT	XLM-R	IndicBERT
News Article Headline Prediction	89.58	95.52	95.87
Wikipedia Section Title Prediction	73.66	66.33	73.31
Cloze-style multiple-choice QA	39.16	27.98	41.87
Article Genre Classification	90.63	97.03	97.34
Named Entity Recognition (F1-score)	73.24	65.93	64.47
Cross-Lingual Sentence Retrieval Task	21.46	13.74	27.12
Average	64.62	61.09	66.66

Additional Tasks

Task	Task Type	mBERT	XLM-R	IndicBERT
BBC News Classification	Genre Classification	60.55	75.52	74.60
IIT Product Reviews	Sentiment Analysis	74.57	78.97	71.32
IITP Movie Reviews	Sentiment Analaysis	56.77	61.61	59.03
Soham News Article	Genre Classification	80.23	87.6	78.45
Midas Discourse	Discourse Analysis	71.20	79.94	78.44
iNLTK Headlines Classification	Genre Classification	87.95	93.38	94.52
ACTSA Sentiment Analysis	Sentiment Analysis	48.53	59.33	61.18
Winograd NLI	Natural Language Inference	56.34	55.87	56.34
Choice of Plausible Alternative (COPA)	Natural Language Inference	54.92	51.13	58.33
Amrita Exact Paraphrase	Paraphrase Detection	93.81	93.02	93.75
Amrita Rough Paraphrase	Paraphrase Detection	83.38	82.20	84.33
Average		69.84	74.42	73.66

* Note: all models have been restricted to a max_seq_length of 128.

Download Model

The model can be downloaded here. Both tf checkpoints and pytorch binaries are included in the archive. Alternatively, you can also download it from Huggingface.

Citing

If you are using any of the resources, please cite the following article:

@inproceedings{kakwani2020indicnlpsuite,
    title={{IndicNLPSuite: Monolingual Corpora, Evaluation Benchmarks and Pre-trained Multilingual Language Models for Indian Languages}},
    author={Divyanshu Kakwani and Anoop Kunchukuttan and Satish Golla and Gokul N.C. and Avik Bhattacharyya and Mitesh M. Khapra and Pratyush Kumar},
    year={2020},
    booktitle={Findings of EMNLP},
}

We would like to hear from you if:

• You are using our resources. Please let us know how you are putting these resources to use.
• You have any feedback on these resources.

License

IndicNLP Corpus is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Contributors

• Divyanshu Kakwani
• Anoop Kunchukuttan
• Gokul NC
• Satish Golla
• Avik Bhattacharyya
• Mitesh Khapra
• Pratyush Kumar

This work is the outcome of a volunteer effort as part of AI4Bharat initiative.

Contact

• Anoop Kunchukuttan ([email protected])
• Mitesh Khapra ([email protected])
• Pratyush Kumar ([email protected])