opus-mt-tc-big-gmq-zlw

Table of Contents

• Model Details
• Uses
• Risks, Limitations and Biases
• How to Get Started With the Model
• Training
• Evaluation
• Citation Information
• Acknowledgements

Model Details

Neural machine translation model for translating from North Germanic languages (gmq) to West Slavic languages (zlw).

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. Model Description:

• Developed by: Language Technology Research Group at the University of Helsinki
• Model Type: Translation (transformer-big)
• Release: 2022-08-03
• License: CC-BY-4.0
• Language(s):
  • Source Language(s): dan nob nor swe
  • Target Language(s): ces pol
  • Language Pair(s): dan-ces nob-ces swe-ces swe-pol
  • Valid Target Language Labels: >>ces<< >>csb<< >>czk<< >>dsb<< >>hsb<< >>pol<< >>pox<< >>slk<< >>szl<<
• Original Model: opusTCv20210807_transformer-big_2022-08-03.zip
• Resources for more information:
  • OPUS-MT-train GitHub Repo
  • More information about released models for this language pair: OPUS-MT gmq-zlw README
  • More information about MarianNMT models in the transformers library
  • [Tatoeba Translation Challenge](https://github.com/Helsinki-NLP/Tatoeba-Challenge/

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. >>ces<<

Uses

This model can be used for translation and text-to-text generation.

Risks, Limitations and Biases

CONTENT WARNING: Readers should be aware that the model is trained on various public data sets that may contain content that is disturbing, offensive, and can propagate historical and current stereotypes.

Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al. (2021) and Bender et al. (2021)).

How to Get Started With the Model

A short example code:

from transformers import MarianMTModel, MarianTokenizer

src_text = [
    ">>ces<< Normalt er jeg hjemme hele weekenden.",
    ">>pol<< Lev ditt liv."
]

model_name = "pytorch-models/opus-mt-tc-big-gmq-zlw"
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:
#     Většinou jsem doma celý víkend.
#     Żyj swoim życiem.

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-gmq-zlw")
print(pipe(">>ces<< Normalt er jeg hjemme hele weekenden."))

# expected output: Většinou jsem doma celý víkend.

Training

• Data: opusTCv20210807 (source)
• Pre-processing: SentencePiece (spm32k,spm32k)
• Model Type: transformer-big
• Original MarianNMT Model: opusTCv20210807_transformer-big_2022-08-03.zip
• Training Scripts: GitHub Repo

Evaluation

• test set translations: opusTCv20210807_transformer-big_2022-08-03.test.txt
• test set scores: opusTCv20210807_transformer-big_2022-08-03.eval.txt
• benchmark results: benchmark_results.txt
• benchmark output: benchmark_translations.zip

langpair	testset	chr-F	BLEU	#sent	#words
swe-pol	tatoeba-test-v2021-08-07	0.66326	46.2	1392	8157
dan-ces	flores101-devtest	0.54065	26.7	1012	22101
dan-pol	flores101-devtest	0.48389	18.8	1012	22520
isl-ces	flores101-devtest	0.43582	17.7	1012	22101
isl-pol	flores101-devtest	0.41929	13.9	1012	22520
nob-ces	flores101-devtest	0.50336	22.3	1012	22101
nob-pol	flores101-devtest	0.46130	16.3	1012	22520
swe-ces	flores101-devtest	0.53188	25.7	1012	22101
swe-pol	flores101-devtest	0.48163	18.6	1012	22520

Citation Information

• 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",
}

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: 8b9f0b0
• port time: Sat Aug 13 00:02:29 EEST 2022
• port machine: LM0-400-22516.local