One Fourth Labs
We deliver courseware in AI and related areas
Samanantar: The Largest Publicly Available Parallel Corpora Collection for 11 Indic Languages
Gowtham Ramesh, Sumanth Doddapaneni, Aravinth Bheemaraj, Mayank Jobanputra, Raghavan AK, Ajitesh Sharma, Sujit Sahoo, Harshita Diddee, Mahalakshmi J, Divyanshu Kakwani, Navneet Kumar, Aswin Pradeep, Kumar Deepak, Vivek Raghavan, Anoop Kunchukuttan, Pratyush Kumar, Mitesh Shantadevi Khapra
Given India's rich linguistic diversity, language technology is absolutely essential to magnify reach and impact of AI solutions
বা ગુ हि ಕ മ म ने ਪੰ த తె ار
বা ગુ हि ಕ മ म ने ਪੰ த తె ار
বা ગુ हि ಕ മ म ने ਪੰ த తె ار
The social view
Touch Points: Digital
Users: Multlingual
Language technology is absolutely essential to magnify reach and impact in the social sector
The commercial view
Multilingual chatbots
Sentiment Analysis
Content Moderation
Code mixed song search
Speech
QA
Multilingual Authoring Tools
... And the demand for these tools is increasing
A concrete use case
original judgements written in English
Translation essential for better information dissemination
Manual translation is tedious and time consuming
বা
हि
ಕ
म
ଓ
த
മ
తె
অ
कॉ
ગુ
ने
कों
सं
ਪੰ
सिं
اُر
ब
मै
22 constitutional languages
What's the problem?
বা
हि
ಕ
म
ଓ
த
മ
తె
অ
कॉ
ગુ
ने
कों
सं
ਪੰ
सिं
اُر
ब
मै
Well, just use an automatic translation system!
No free, open-source, accurate translation system for Indic languages!
What would it take to solve this problem?
A large number of parallel sentences between En and Indic languages
DATA
MODELS
COMPUTE*
Modern infrastructure to train large scale models with millions of data points
This talk
Large scale models with (minor) innovations specific to Indic languages
*Most large scale models (T5, mT5, mBART) come out of deep tech companies
हि
En
How did we address the data problem?
Principle: Curate data from the web, manual collection is too expensive and time consuming
WEB SOURCES
Comparable
Non-Comparable
(Monolingual, billion tokens)
Machine Readable
हि
En
Non-Machine Readable
हि
En
En
हि
ગુ
... ...
How did we address the data problem?
Shared multilingual vector space*
* Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, Wei Wang, Language-agnostic BERT Sentence Embedding, arXiv 2020
24 such news sources considered in this work with data from 2010 onwards
En
https://mykhel.com/
हि
https://hindi.mykhel.com/
Jan 2020
How did we address the data problem?
https://tn.gov.in/
* Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, Wei Wang, Language-agnostic BERT Sentence Embedding, arXiv 2020
Legislative proceedings from Tamil Nadu, Andhra Pradesh, Telangana, West Bengal, Bangladesh
en_budget_2020.pdf
ta_budget_2020.pdf
https://tn.gov.in/
OCR
OCR
En
த
Shared multilingual vector space*
How did we address the data problem?
Jeff Johnson, Matthijs Douze, Hervé Jégou, Billion-scale similarity search with GPUs, arXiv, 2019
IndicCorp contains 1.3M (Assamese) to 100M (English) monolingual sentences for 11 Indic languages
FAISS Index* for efficient indexing, clustering, semantic matching, and retrieval of dense vectors (1000 sent./sec)
En
100M
Hi
64M
Brute force search (100M x 64M) is infeasible)
How much data did we collect?
33M parallel sentences mined from the web (3X improvement)
Qualitative Analysis
10000 samples manually evaluated using 30+ annotators across 11 languages
Average rating of sentence pairs around 4.17 (min:3.83, max:4.82)
Quality depends on resource size (lowest for As, Or; highest for Hi, Bn)
LaBSE alignment scores is negatively correlated with sentence length
What is the model that we use?
હું
અહીં
છું
Multi-headed self attention
Add and Normalise
હું અહીં છું
હું
અહીં
છું
I am here
हुं अहीं छुं
Decoder
Decoder
Decoder
Decoder
Decoder
Decoder
Transformer
Highlights
Joint multilingual model for 11 Indic languages
3 models: En-X, X-En and X-X
Single script (enables transfer, reduces vocabulary)
6 encoder layers, 6 decoder layers, 16 heads/layer
wide models (ff_dim=4096, embedding_dim=1536)
বা
हि
ಕ
म
ଓ
த
മ
తె
অ
ગુ
ਪੰ
How do our models perform?
Highlights
State of the art performance
Gains are higher for low resource languages
Impact
Deployed for translating Supreme Court judgements
“The quality of translations is significantly improved. I would say this is more so for the legal document where there were complicated sentences/cards which were translated very well. The syntax mostly did not falter even in the face of multiple ideas/information contained in one sentence.”
“The amount of time spent on correcting/improving the translation has dropped.”
“THIS IS VERY PROMISING. AMAZED BY THE SPEED.”
“I TOOK A PRINTOUT AND WENT THROUGH EVERY LINE. THE TRANSLATION IS 98% ACCURATE AND HIGHLY SATISFIED.”
Summary
Samanantar: Largest Parallel Corpus for Indic Languages
IndicTrans: State of the art translation models for En-X and X-En
https://github.com/AI4Bharat/indicTrans
https://indicnlp.ai4bharat.org/samanantar/
Contact Us
Let's make India ready for the AI age
Anoop Kunchukuttan
Mitesh M. Khapra
Pratyush Kumar
anoop.kunchukuttan@gmail.com
miteshk@cse.iitm.ac.in
pratyush@cse.iitm.ac.in
Samanantar
By One Fourth Labs
Samanantar
A hub for Indic NLP resources, datasets and tools
