metadata

language:
  - multilingual
  - af
  - am
  - ar
  - ast
  - az
  - ba
  - be
  - bg
  - bn
  - br
  - bs
  - ca
  - ceb
  - cs
  - cy
  - da
  - de
  - el
  - en
  - es
  - et
  - fa
  - ff
  - fi
  - fr
  - fy
  - ga
  - gd
  - gl
  - gu
  - ha
  - he
  - hi
  - hr
  - ht
  - hu
  - hy
  - id
  - ig
  - ilo
  - is
  - it
  - ja
  - jv
  - ka
  - kk
  - km
  - kn
  - ko
  - lb
  - lg
  - ln
  - lo
  - lt
  - lv
  - mg
  - mk
  - ml
  - mn
  - mr
  - ms
  - my
  - ne
  - nl
  - 'no'
  - ns
  - oc
  - or
  - pa
  - pl
  - ps
  - pt
  - ro
  - ru
  - sd
  - si
  - sk
  - sl
  - so
  - sq
  - sr
  - ss
  - su
  - sv
  - sw
  - ta
  - th
  - tl
  - tn
  - tr
  - uk
  - ur
  - uz
  - vi
  - wo
  - xh
  - yi
  - yo
  - zh
  - zu
license: mit
tags:
  - small100
  - translation
  - flores101
  - gsarti/flores_101
  - tico19
  - gmnlp/tico19
  - tatoeba

SMALL-100 Model

SMaLL-100 is a compact and fast massively multilingual machine translation model covering more than 10K language pairs, that achieves competitive results with M2M-100 while being much smaller and faster. It is introduced in this paper(accepted to EMNLP2022), and initially released in this repository.

The model architecture and config are the same as M2M-100 implementation, but the tokenizer is modified to adjust language codes. So, you should load the tokenizer locally from tokenization_small100.py file for the moment.

Note: SMALL100Tokenizer requires sentencepiece, so make sure to install it by:

pip install sentencepiece

Supervised Training

SMaLL-100 is a seq-to-seq model for the translation task. The input to the model is source:[tgt_lang_code] + src_tokens + [EOS] and target: tgt_tokens + [EOS].

An example of supervised training is shown below:

from transformers import M2M100ForConditionalGeneration
from tokenization_small100 import SMALL100Tokenizer

model = M2M100ForConditionalGeneration.from_pretrained("alirezamsh/small100")
tokenizer = M2M100Tokenizer.from_pretrained("alirezamsh/small100", tgt_lang="fr")

src_text = "Life is like a box of chocolates."
tgt_text = "La vie est comme une boîte de chocolat."

model_inputs = tokenizer(src_text, text_target=tgt_text, return_tensors="pt")

loss = model(**model_inputs).loss  # forward pass

Training data can be provided upon request.

Generation

Beam size of 5, and maximum target length of 256 is used for the generation.

from transformers import M2M100ForConditionalGeneration
from tokenization_small100 import SMALL100Tokenizer

hi_text = "जीवन एक चॉकलेट बॉक्स की तरह है।"
chinese_text = "生活就像一盒巧克力。"

model = M2M100ForConditionalGeneration.from_pretrained("alirezamsh/small100")
tokenizer = SMALL100Tokenizer.from_pretrained("alirezamsh/small100")

# translate Hindi to French
tokenizer.tgt_lang = "fr"
encoded_hi = tokenizer(hi_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi, max_length=256, num_beams=5)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "La vie est comme une boîte de chocolat."

# translate Chinese to English
tokenizer.tgt_lang = "en"
encoded_zh = tokenizer(chinese_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_zh, max_length=256, num_beams=5)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Life is like a box of chocolate."

Evaluation

Please refer to original repository for spBLEU computation.

Languages Covered

Afrikaans (af), Amharic (am), Arabic (ar), Asturian (ast), Azerbaijani (az), Bashkir (ba), Belarusian (be), Bulgarian (bg), Bengali (bn), Breton (br), Bosnian (bs), Catalan; Valencian (ca), Cebuano (ceb), Czech (cs), Welsh (cy), Danish (da), German (de), Greeek (el), English (en), Spanish (es), Estonian (et), Persian (fa), Fulah (ff), Finnish (fi), French (fr), Western Frisian (fy), Irish (ga), Gaelic; Scottish Gaelic (gd), Galician (gl), Gujarati (gu), Hausa (ha), Hebrew (he), Hindi (hi), Croatian (hr), Haitian; Haitian Creole (ht), Hungarian (hu), Armenian (hy), Indonesian (id), Igbo (ig), Iloko (ilo), Icelandic (is), Italian (it), Japanese (ja), Javanese (jv), Georgian (ka), Kazakh (kk), Central Khmer (km), Kannada (kn), Korean (ko), Luxembourgish; Letzeburgesch (lb), Ganda (lg), Lingala (ln), Lao (lo), Lithuanian (lt), Latvian (lv), Malagasy (mg), Macedonian (mk), Malayalam (ml), Mongolian (mn), Marathi (mr), Malay (ms), Burmese (my), Nepali (ne), Dutch; Flemish (nl), Norwegian (no), Northern Sotho (ns), Occitan (post 1500) (oc), Oriya (or), Panjabi; Punjabi (pa), Polish (pl), Pushto; Pashto (ps), Portuguese (pt), Romanian; Moldavian; Moldovan (ro), Russian (ru), Sindhi (sd), Sinhala; Sinhalese (si), Slovak (sk), Slovenian (sl), Somali (so), Albanian (sq), Serbian (sr), Swati (ss), Sundanese (su), Swedish (sv), Swahili (sw), Tamil (ta), Thai (th), Tagalog (tl), Tswana (tn), Turkish (tr), Ukrainian (uk), Urdu (ur), Uzbek (uz), Vietnamese (vi), Wolof (wo), Xhosa (xh), Yiddish (yi), Yoruba (yo), Chinese (zh), Zulu (zu)

Citation

If you use this model for your research, please cite the following work:

@misc{mohammadshahi2022small100,
    title={SMaLL-100: Introducing Shallow Multilingual Machine Translation Model for Low-Resource Languages},
    author={Alireza Mohammadshahi and Vassilina Nikoulina and Alexandre Berard and Caroline Brun and James Henderson and Laurent Besacier},
    year={2022},
    eprint={2210.11621},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

@misc{mohammadshahi2022compressed,
    title={What Do Compressed Multilingual Machine Translation Models Forget?},
    author={Alireza Mohammadshahi and Vassilina Nikoulina and Alexandre Berard and Caroline Brun and James Henderson and Laurent Besacier},
    year={2022},
    eprint={2205.10828},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}