Machine Translation for Turkic Languages

Introduction

What is Turkic Interlingua (TIL)?

Turkic Interlingua (TIL) is a community of researchers, Machine Learning (ML) engineers, language enthusiasts and community leaders whose mission is to develop language technologies (from spell checkers to translation models), collect diverse datasets, and explore linguistic phenomena through the lens of academic research for Turkic languages.

What is TIL Corpus?

TIL Corpus is a large-scale parallel corpus combining most of the public datasets for 22 Turkic languages. The current version of the corpus yields data for almost 400 language directions with development and test sets available for >300 of them. You can learn more about the corpus and how to use it here.

What is this repo?

This repo is a collection of resources for training and researching Machine Translation models. It includes links and scripts to download the TIL Corpus, X-WMT test sets, multilingual models and more. The repo also allows researchers to replicate our results from our publications in the past.

Model results on the X-WMT test set

Pairs	Train size	Bilingual		MNMT		Improvement
		Chrf	BLEU	Chrf	BLEU	▲ Chrf	▲ BLEU
ru-ba	523.7K	0.59	24.31	0.56	23.13	-0.03	-1.18
ba-ru	523.7K	0.58	24.39	0.57	24.57	-0.02	0.18
en-tr	35.8M	0.55	16.04	0.56	26.74	0.01	10.70
tr-en	35.8M	0.51	20.39	0.55	24.66	0.04	4.27
az-tr	410.1K	0.43	10.61	0.48	19.63	0.05	9.02
kk-en	564.8K	0.42	12.00	0.44	15.71	0.02	3.71
az-en	548.9K	0.41	12.01	0.49	20.41	0.08	8.40
ru-uz	1.3M	0.41	6.58	0.42	6.70	0.01	0.12
en-kk	564.8K	0.40	7.82	0.43	9.92	0.03	2.10
en-uz	529.6K	0.40	6.34	0.42	9.89	0.02	3.55
tr-az	410.1K	0.39	7.78	0.42	8.21	0.03	0.43
en-az	548.9K	0.38	6.79	0.43	9.71	0.05	2.92
uz-ru	1.3M	0.36	6.08	0.39	9.16	0.03	3.08
ru-ky	293.7K	0.35	4.42	0.45	10.35	0.10	5.93
ky-ru	293.7K	0.30	5.23	0.44	14.08	0.14	8.85
ky-en	312.6K	0.29	4.65	0.39	10.87	0.10	6.22
en-ky	312.6K	0.27	2.33	0.34	4.64	0.07	2.31
uz-en	529.6K	0.24	4.81	0.45	14.45	0.21	9.64
kaa-en	17.1K	0.21	1.04	0.38	10.21	0.17	9.17
sah-en	8.1K	0.21	0.16	0.24	3.38	0.03	3.22
ba-en	34.3K	0.19	0.32	0.40	10.55	0.21	10.23
en-kaa	17.1K	0.19	0.31	0.27	2.82	0.08	2.51
sah-ru	9.2K	0.18	0.42	0.25	4.41	0.07	3.99
ru-sah	9.2K	0.16	0.12	0.17	4.64	0.01	4.52
en-ba	34.3K	0.14	0.16	0.34	8.35	0.20	8.19
en-sah	8.1K	0.14	0.04	0.12	3.46	-0.02	3.42

How to cite

If you are using the TIL Corpus or X-WMT test sets, please cite:

@inproceedings{mirzakhalov2021large,
  title={A Large-Scale Study of Machine Translation in Turkic Languages},
  author={Mirzakhalov, Jamshidbek and Babu, Anoop and Ataman, Duygu and Kariev, Sherzod and Tyers, Francis and Abduraufov, Otabek and Hajili, Mammad and Ivanova, Sardana and Khaytbaev, Abror and Laverghetta Jr, Antonio and others},
  booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  pages={5876--5890},
  year={2021}
}

If you are using the MNMT model in your research, please cite:

@article{mirzakhalov2021evaluating, title={Evaluating Multiway Multilingual NMT in the Turkic Languages}, author={Mirzakhalov, Jamshidbek and Babu, Anoop and Kunafin, Aigiz and Wahab, Ahsan and Moydinboyev, Behzod and Ivanova, Sardana and Uzokova, Mokhiyakhon and Pulatova, Shaxnoza and Ataman, Duygu and Kreutzer, Julia and others}, journal={arXiv preprint arXiv:2109.06262}, year={2021} }

If you want to talk about the Turkic Interlingua (TIL) community overall, please cite:

@phdthesis{mirzakhalov2021turkic,
  title={Turkic Interlingua: A Case Study of Machine Translation in Low-resource Languages},
  author={Mirzakhalov, Jamshidbek},
  year={2021},
  school={University of South Florida}
}

Contributors

This project was carried out with the help and contributions from dozens of individuals and organizations. We acknowledge and greatly appreciate each and every one of them:

Authors on the publications (in alphabetical order)

Abror Khaytbaev
Ahsan Wahab
Aigiz Kunafin
Anoop Babu
Antonio Laverghetta Jr.
Behzodbek Moydinboyev
Dr. Duygu Ataman
Esra Onal
Dr. Francis Tyers
Jamshidbek Mirzakhalov
Dr. John Licato
Dr. Julia Kreutzer
Mammad Hajili
Mokhiyakhon Uzokova
Dr. Orhan Firat
Otabek Abduraufov
Sardana Ivanova
Shaxnoza Pulatova
Sherzod Kariev
Dr. Sriram Chellappan

Translators, annotators and dataset contributors (in alphabetical order)

Abilxayr Zholdybai
Aigiz Kunafin
Akylbek Khamitov
Alperen Cantez
Aydos Muxammadiyarov
Doniyorbek Rafikjonov
Erkinbek Vokhabov
Ipek Baris
Iskander Shakirov
Madina Zokirjonova
Mohiyaxon Uzoqova
Mukhammadbektosh Khaydarov
Nurlan Maharramli
Petr Popov
Rasul Karimov
Sariya Kagarmanova
Ziyodabonu Qobiljon qizi

Industry supporters

Google Cloud
Khan Academy Oʻzbek
The Foundation for the Preservation and Development of the Bashkir Language

turkic-interlingua/til-mt