opus-mt-tc-base-en-sh

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 English (en) to Serbo-Croatian (sh).

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-align)
Release: 2021-04-20
License: CC-BY-4.0
Language(s):
- Source Language(s): eng
- Target Language(s): bos_Latn hbs hrv srp_Cyrl srp_Latn
- Language Pair(s): eng-bos_Latn eng-hbs eng-hrv eng-srp_Cyrl eng-srp_Latn
- Valid Target Language Labels: >>bos_Cyrl<< >>bos_Latn<< >>cnr<< >>cnr_Latn<< >>hbs<< >>hbs_Cyrl<< >>hrv<< >>srp_Cyrl<< >>srp_Latn<<
Original Model: opus+bt-2021-04-20.zip
Resources for more information:
- OPUS-MT-train GitHub Repo
- More information about released models for this language pair: OPUS-MT eng-hbs 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. >>bos_Latn<<

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 = [
    ">>hrv<< You're about to make a very serious mistake.",
    ">>hbs<< I've just been too busy."
]

model_name = "pytorch-models/opus-mt-tc-base-en-sh"
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:
#     Ti si o tome napraviti vrlo ozbiljnu pogrešku.
#     [4]

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-base-en-sh")
print(pipe(">>hrv<< You're about to make a very serious mistake."))

# expected output: Ti si o tome napraviti vrlo ozbiljnu pogrešku.

Training

Data: opus+bt (source)
Pre-processing: SentencePiece (spm32k,spm32k)
Model Type: transformer-align
Original MarianNMT Model: opus+bt-2021-04-20.zip
Training Scripts: GitHub Repo

Evaluation

test set translations: opus+bt-2021-04-20.test.txt
test set scores: opus+bt-2021-04-20.eval.txt
benchmark results: benchmark_results.txt
benchmark output: benchmark_translations.zip

langpair	testset	chr-F	BLEU	#sent	#words
eng-bos_Latn	tatoeba-test-v2021-08-07	0.666	46.3	301	1650
eng-hbs	tatoeba-test-v2021-08-07	0.631	42.1	10017	63927
eng-hrv	tatoeba-test-v2021-08-07	0.691	49.7	1480	9396
eng-srp_Cyrl	tatoeba-test-v2021-08-07	0.645	45.1	1580	9152
eng-srp_Latn	tatoeba-test-v2021-08-07	0.613	39.8	6656	43729
eng-hrv	flores101-devtest	0.586	28.7	1012	22423
eng-hrv	flores200-dev	0.57963	28.1	997	21567
eng-hrv	flores200-devtest	0.58652	28.9	1012	22423
eng-srp_Cyrl	flores101-devtest	0.59874	31.7	1012	23456
eng-srp_Cyrl	flores200-dev	0.60096	32.2	997	22384
eng-srp_Cyrl	flores200-devtest	0.59874	31.7	1012	23456

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: e2a6299
port time: Tue Oct 11 10:14:32 CEST 2022
port machine: LM0-400-22516.local

Helsinki-NLP
/

opus-mt-tc-base-en-sh