MultiSlav MultiSlav-5lang

Multilingual Many2Many MT Model

MultiSlav-5lang is an Encoder-Decoder vanilla transformer model trained on sentence-level Machine Translation task. Model is supporting translation between 5 languages: Czech, English, Polish, Slovak, Slovene. This model is part of the MultiSlav collection. More information will be available soon in our upcoming MultiSlav paper.

Experiments were conducted under research project by Machine Learning Research lab for Allegro.com. Big thanks to laniqo.com for cooperation in the research.

MultiSlav-5lang - translates directly between all supported languages using single Many2Many model as seen on the diagram above.

Model description

• Model name: multislav-5lang
• Source Languages: Czech, English, Polish, Slovak, Slovene
• Target Languages: Czech, English, Polish, Slovak, Slovene
• Model Collection: MultiSlav
• Model type: MarianMTModel Encoder-Decoder
• License: CC BY 4.0 (commercial use allowed)
• Developed by: MLR @ Allegro & Laniqo.com

Supported languages

Using model you must specify target language for translation. Target language tokens are represented as 3-letter ISO 639-3 language codes embedded in a format >>xxx<<. All accepted directions and their respective tokens are listed below. Each of them was added as a special token to Sentence-Piece tokenizer.

Target Language	First token
Czech	>>ces<<
English	>>eng<<
Polish	>>pol<<
Slovak	>>slk<<
Slovene	>>slv<<

Use case quickstart

Example code-snippet to use model. Due to bug the MarianMTModel must be used explicitly.

from transformers import AutoTokenizer, MarianMTModel

model_name = "Allegro/MultiSlav-5lang"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

text = "Allegro to internetowa platforma e-commerce, na której swoje produkty sprzedają średnie i małe firmy, jak również duże marki."
target_languages = ["ces", "eng", "slk", "slv"]
batch_to_translate = [
    f">>{lang}<<" + " " + text for lang in target_languages
]

translations = model.generate(**tokenizer.batch_encode_plus(batch_to_translate, return_tensors="pt"))
decoded_translations = tokenizer.batch_decode(translations, skip_special_tokens=True, clean_up_tokenization_spaces=True)
for trans in decoded_translations:
    print(trans)

Generated outputs:

Czech output:

Allegro je on-line e-commerce platforma, na které své produkty prodávají střední a malé firmy, stejně jako velké značky.

English output:

Allegro is an online e-commerce platform on which medium and small companies as well as large brands sell their products.

Slovak output:

Allegro je internetová e-commerce platforma, na ktorej svoje produkty predávajú stredné a malé podniky, ako aj veľké značky.

Slovene output:

Allegro je spletna platforma za e-poslovanje, na kateri srednje velika in mala podjetja ter velike blagovne znamke prodajajo svoje izdelke.

The model is also capable of translating into Polish language, following the same pattern:

text = ">>pol<<" + " " + "Allegro is an online e-commerce platform on which medium and small companies as well as large brands sell their products."
translation = model.generate(**tokenizer.batch_encode_plus([text], return_tensors="pt"))
decoded_translation = tokenizer.batch_decode(translation, skip_special_tokens=True, clean_up_tokenization_spaces=True)

print(decoded_translation[0])

Generated Polish output:

Allegro to internetowa platforma e-commerce, na której sprzedają swoje produkty średnie i małe firmy, a także duże marki.

Training

SentencePiece tokenizer has a vocab size 80k in total (16k per language). Tokenizer was trained on randomly sampled part of the training corpus. During the training we used the MarianNMT framework. Base marian configuration used: transfromer-big. All training parameters are listed in table below.

Training hyperparameters:

Hyperparameter	Value
Total Parameter Size	258M
Training Examples	578M
Vocab Size	80k
Base Parameters	Marian transfromer-big
Number of Encoding Layers	6
Number of Decoding Layers	6
Model Dimension	1024
FF Dimension	4096
Heads	16
Dropout	0.1
Batch Size	mini batch fit to VRAM
Training Accelerators	4x A100 40GB
Max Length	100 tokens
Optimizer	Adam
Warmup steps	8000
Context	Sentence-level MT
Source Languages Supported	Czech, English, Polish, Slovak, Slovene
Target Languages Supported	Czech, English, Polish, Slovak, Slovene
Precision	float16
Validation Freq	3000 steps
Stop Metric	ChrF
Stop Criterion	20 Validation steps

Training corpora

The main research question was: "How does adding additional, related languages impact the quality of the model?" - we explored it in the Slavic language family. In this model we experimented by additionally adding English <-> Slavic parallel corpora to further increase open-source data-regime. We found that additional data clearly improved performance compared to the bi-directional baseline models, and compared to pivot models and MultiSlav-4slav in the most of the directions. For example in translation from Polish to Czech, this allowed us to expand training data-size from 63M to 578M examples, and from 18M to 578M for Slovak to Slovene translation.

We only used explicitly open-source data to ensure open-source license of our model. Datasets were downloaded via MT-Data library. Number of total examples post filtering and deduplication: 578M.

The datasets used and data amount prior to filtering and deduplication:

Corpus	Data Size
paracrawl	246407901
opensubtitles	167583218
multiparacrawl	52388826
dgt	36403859
elrc	29687222
xlent	18375223
wikititles	12936394
wmt	11074816
wikimatrix	10435588
dcep	10239150
ELRC	7609067
tildemodel	6309369
europarl	6088362
eesc	5604672
eubookshop	3732718
emea	3482661
jrc_acquis	2920805
ema	1881408
qed	1835208
elitr_eca	1398536
EU-dcep	1132950
rapid	1016905
ecb	885442
kde4	541944
news_commentary	498432
kde	473269
bible_uedin	429692
europat	358911
elra	357696
wikipedia	352118
wikimedia	201088
tatoeba	91251
globalvoices	69736
euconst	65507
ubuntu	47301
php	44031
ecdc	21154
eac	20224
eac_reference	10099
gnome	4466
EU-eac	2925
books	2816
EU-ecdc	2210
newsdev	1953
khresmoi_summary	889
czechtourism	832
khresmoi_summary_dev	455
worldbank	189

Evaluation

Evaluation of the models was performed on Flores200 dataset. The table below compares performance of the open-source models and all applicable models from our collection. Metrics BLEU, ChrF2, and Unbabel/wmt22-comet-da.

Translation results on translation from Polish to Czech (Slavic direction with the highest data-regime):

Model	Comet22	BLEU	ChrF	Model Size
M2M−100	89.6	19.8	47.7	1.2B
NLLB−200	89.4	19.2	46.7	1.3B
Opus Sla-Sla	82.9	14.6	42.6	64M
BiDi-ces-pol (baseline)	90.0	20.3	48.5	209M
P4-pol ◊	90.2	20.2	48.5	2x 242M
P5-eng ◊	89.0	19.9	48.3	2x 258M
P5-ces ◊	90.3	20.2	48.6	2x 258M
MultiSlav-4slav	90.2	20.6	48.7	242M
MultiSlav-5lang *	90.4	20.7	48.9	258M

Translation results on translation from Slovak to Slovene (Slavic direction with the lowest data-regime):

Model	Comet22	BLEU	ChrF	Model Size
M2M−100	89.6	26.6	55.0	1.2B
NLLB−200	88.8	23.3	42.0	1.3B
BiDi-slk-slv (baseline)	89.4	26.6	55.4	209M
P4-pol ◊	88.4	24.8	53.2	2x 242M
P5-eng ◊	88.5	25.6	54.6	2x 258M
P5-ces ◊	89.8	26.6	55.3	2x 258M
MultiSlav-4slav	90.1	27.1	55.7	242M
MultiSlav-5lang *	90.2	27.1	55.7	258M

* this model

◊ system of 2 models Many2XXX and XXX2Many, see P5-ces2many

Limitations and Biases

We did not evaluate inherent bias contained in training datasets. It is advised to validate bias of our models in perspective domain. This might be especially problematic in translation from English to Slavic languages, which require explicitly indicated gender and might hallucinate based on bias present in training data.

License

The model is licensed under CC BY 4.0, which allows for commercial use.

Citation

TO BE UPDATED SOON 🤗

Contact Options

Authors:

• MLR @ Allegro: Artur Kot, Mikołaj Koszowski, Wojciech Chojnowski, Mieszko Rutkowski
• Laniqo.com: Artur Nowakowski, Kamil Guttmann, Mikołaj Pokrywka

Please don't hesitate to contact authors if you have any questions or suggestions:

• e-mail: [email protected] or [email protected]
• LinkedIn: Artur Kot or Mikołaj Koszowski