opus-mt-tc-big-zls-zle

Neural machine translation model for translating from South Slavic languages (zls) to East Slavic languages (zle).

This model is part of the OPUS-MT project, an effort to make neural machine translation models widely available and accessible for many languages in the world. All models are originally trained using the amazing framework of Marian NMT, an efficient NMT implementation written in pure C++. The models have been converted to pyTorch using the transformers library by huggingface. Training data is taken from OPUS and training pipelines use the procedures of OPUS-MT-train.

• Publications: OPUS-MT – Building open translation services for the World and The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT (Please, cite if you use this model.)

@inproceedings{tiedemann-thottingal-2020-opus,
    title = "{OPUS}-{MT} {--} Building open translation services for the World",
    author = {Tiedemann, J{\"o}rg  and Thottingal, Santhosh},
    booktitle = "Proceedings of the 22nd Annual Conference of the European Association for Machine Translation",
    month = nov,
    year = "2020",
    address = "Lisboa, Portugal",
    publisher = "European Association for Machine Translation",
    url = "https://aclanthology.org/2020.eamt-1.61",
    pages = "479--480",
}

@inproceedings{tiedemann-2020-tatoeba,
    title = "The Tatoeba Translation Challenge {--} Realistic Data Sets for Low Resource and Multilingual {MT}",
    author = {Tiedemann, J{\"o}rg},
    booktitle = "Proceedings of the Fifth Conference on Machine Translation",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.wmt-1.139",
    pages = "1174--1182",
}

Model info

• Release: 2022-03-23
• source language(s): bul hbs hrv slv srp_Cyrl srp_Latn
• target language(s): bel rus ukr
• valid target language labels: >>bel<< >>rus<< >>ukr<<
• model: transformer-big
• data: opusTCv20210807+bt (source)
• tokenization: SentencePiece (spm32k,spm32k)
• original model: opusTCv20210807+bt_transformer-big_2022-03-23.zip
• more information released models: OPUS-MT zls-zle README
• more information about the model: MarianMT

This is a multilingual translation model with multiple target languages. A sentence initial language token is required in the form of >>id<< (id = valid target language ID), e.g. >>bel<<

Usage

A short example code:

from transformers import MarianMTModel, MarianTokenizer

src_text = [
    ">>rus<< Gdje je brigadir?",
    ">>ukr<< Zovem se Seli."
]

model_name = "pytorch-models/opus-mt-tc-big-zls-zle"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)
translated = model.generate(**tokenizer(src_text, return_tensors="pt", padding=True))

for t in translated:
    print( tokenizer.decode(t, skip_special_tokens=True) )

# expected output:
#     Где бригадир?
#     Мене звати Саллі.

You can also use OPUS-MT models with the transformers pipelines, for example:

from transformers import pipeline
pipe = pipeline("translation", model="Helsinki-NLP/opus-mt-tc-big-zls-zle")
print(pipe(">>rus<< Gdje je brigadir?"))

# expected output: Где бригадир?

Benchmarks

• test set translations: opusTCv20210807+bt_transformer-big_2022-03-23.test.txt
• test set scores: opusTCv20210807+bt_transformer-big_2022-03-23.eval.txt
• benchmark results: benchmark_results.txt
• benchmark output: benchmark_translations.zip

langpair	testset	chr-F	BLEU	#sent	#words
bul-rus	tatoeba-test-v2021-08-07	0.71467	52.6	1247	7870
bul-ukr	tatoeba-test-v2021-08-07	0.71757	53.3	1020	4932
hbs-rus	tatoeba-test-v2021-08-07	0.74593	58.5	2500	14213
hbs-ukr	tatoeba-test-v2021-08-07	0.70244	52.3	942	4961
hrv-ukr	tatoeba-test-v2021-08-07	0.68931	50.0	389	2232
slv-rus	tatoeba-test-v2021-08-07	0.42255	27.3	657	4056
srp_Cyrl-rus	tatoeba-test-v2021-08-07	0.74112	56.2	881	5117
srp_Cyrl-ukr	tatoeba-test-v2021-08-07	0.68915	51.8	205	1061
srp_Latn-rus	tatoeba-test-v2021-08-07	0.75340	60.1	1483	8311
srp_Latn-ukr	tatoeba-test-v2021-08-07	0.73106	55.8	348	1668
bul-rus	flores101-devtest	0.54226	24.6	1012	23295
bul-ukr	flores101-devtest	0.53382	22.9	1012	22810
hrv-rus	flores101-devtest	0.51726	23.5	1012	23295
hrv-ukr	flores101-devtest	0.51011	21.9	1012	22810
mkd-bel	flores101-devtest	0.40885	10.7	1012	24829
mkd-rus	flores101-devtest	0.52509	24.3	1012	23295
mkd-ukr	flores101-devtest	0.52021	22.5	1012	22810
slv-rus	flores101-devtest	0.50349	22.0	1012	23295
slv-ukr	flores101-devtest	0.49156	20.2	1012	22810
srp_Cyrl-rus	flores101-devtest	0.53656	25.7	1012	23295
srp_Cyrl-ukr	flores101-devtest	0.53623	24.4	1012	22810

Acknowledgements

The work is supported by the European Language Grid as pilot project 2866, by the FoTran project, funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 771113), and the MeMAD project, funded by the European Union’s Horizon 2020 Research and Innovation Programme under grant agreement No 780069. We are also grateful for the generous computational resources and IT infrastructure provided by CSC -- IT Center for Science, Finland.

Model conversion info

• transformers version: 4.16.2
• OPUS-MT git hash: 1bdabf7
• port time: Thu Mar 24 04:08:51 EET 2022
• port machine: LM0-400-22516.local