MultiSlav P4-many2pol

Multilingual Many-to-Polish MT Model

P4-many2pol is an Encoder-Decoder vanilla transformer model trained on sentence-level Machine Translation task. Model is supporting translation from 3 languages: Czech, Slovak, and Slovene to Polish. 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.

P4-many2pol - Many-to-Polish model translating from Slavic languages to Polish. This model and P4-pol2many combine into P4-pol pivot system translating between 4 Slavic languages. P4-pol translates all supported slavic languages using Many2One model to Polish bridge sentence and next using the One2Many model from Polish bridge sentence to target Slavic language.

Model description

• Model name: P4-many2pol
• Source Languages: Czech, Slovak, Slovene
• Target Language: Polish
• Model Collection: MultiSlav
• Model type: MarianMTModel Encoder-Decoder
• License: CC BY 4.0 (commercial use allowed)
• Developed by: MLR @ Allegro & Laniqo.com

Supported languages

To use the model, you must provide the 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.

Source Language	First token
Czech	>>ces<<
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

m2o_model_name = "Allegro/P4-many2pol"

m2o_tokenizer = AutoTokenizer.from_pretrained(m2o_model_name)
m2o_model = MarianMTModel.from_pretrained(m2o_model_name)

text = ">>ces<<" + " " + "Allegro je on-line e-commerce platforma, na které své produkty prodávají střední a malé firmy, stejně jako velké značky."

translations = m2o_model.generate(**m2o_tokenizer.batch_encode_plus([text], return_tensors="pt"))
bridge_translation = m2o_tokenizer.batch_decode(translations, skip_special_tokens=True, clean_up_tokenization_spaces=True)
print(bridge_translation[0])

Generated bridge Polish output:

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

To pivot-translate to other languages via bridge Polish sentence, we need One2Many model. One2Many model requires explicit target language token as well:

o2m_model_name = "Allegro/P4-pol2many"

o2m_tokenizer = AutoTokenizer.from_pretrained(o2m_model_name)
o2m_model = MarianMTModel.from_pretrained(o2m_model_name)

texts_to_translate = [
    ">>slk<<" + bridge_translation[0],
    ">>slv<<" + bridge_translation[0]
]
translation = o2m_model.generate(**o2m_tokenizer.batch_encode_plus(texts_to_translate, return_tensors="pt"))
decoded_translations = o2m_tokenizer.batch_decode(translation, skip_special_tokens=True, clean_up_tokenization_spaces=True)

for trans in decoded_translations:
    print(trans)

Generated Czech to Slovak pivot translation via Polish:

Allegro je online platforma elektronického obchodu, na ktorej svoje produkty predávajú stredné a malé podniky, ako aj veľké značky.

Generated Czech to Slovene pivot translation via Polish:

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

Training

SentencePiece tokenizer has a vocab size 64k 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	242M
Training Examples	112M
Vocab Size	64k
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, Slovak, Slovene
Target Languages Supported	Polish
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 with expanding data-regime by using data from multiple source languages. We found that additional fluency data clearly improved compared to the bi-directional baseline models. For example in translation from Czech to Polish, this allowed us to expand training data-size from 63M to 112M examples, and from 23M to 112M for Slovene to Polish 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: 112M.

The datasets used:

Corpus
paracrawl
opensubtitles
multiparacrawl
dgt
elrc
xlent
wikititles
wmt
wikimatrix
dcep
ELRC
tildemodel
europarl
eesc
eubookshop
emea
jrc_acquis
ema
qed
elitr_eca
EU-dcep
rapid
ecb
kde4
news_commentary
kde
bible_uedin
europat
elra
wikipedia
wikimedia
tatoeba
globalvoices
euconst
ubuntu
php
ecdc
eac
eac_reference
gnome
EU-eac
books
EU-ecdc
newsdev
khresmoi_summary
czechtourism
khresmoi_summary_dev
worldbank

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 Czech to Polish (Slavic direction to Polish with the highest data-regime):

Model	Comet22	BLEU	ChrF	Model Size
M2M−100	89.0	18.3	48.0	1.2B
NLLB−200	88.9	27.5	47.3	1.3B
Opus Sla-Sla	82.8	13.6	43.5	64M
BiDi-ces-pol (baseline)	89.4	19.2	49.2	209M
P4-many2pol *	89.6	19.3	49.5	242M
P5-eng ◊	89.0	18.5	48.7	2x 258M
P5-ces ◊	89.6	19.0	49.0	2x 258M
MultiSlav-4slav	89.7	18.9	49.2	242M
MultiSlav-5lang	89.8	19.0	49.3	258M

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

Model	Comet22	BLEU	ChrF	Model Size
M2M−100	88.7	17.8	47.3	1.2B
NLLB−200	88.6	17.0	46.3	1.3B
Opus Sla-Sla	80.8	12.3	41.8	64M
BiDi-pol-slv (baseline)	88.1	17.0	47.3	209M
P4-many2pol *	88.7	17.7	48.0	242M
P5-eng ◊	88.4	17.6	47.8	2x 258M
P5-ces ◊	87.9	16.7	46.7	2x 258M
MultiSlav-4slav	88.9	17.8	48.0	242M
MultiSlav-5lang	89.2	18.0	47.9	258M

* this model; it is Many2One a part of the P4-pol pivot system.

◊ system of 2 models Many2XXX and XXX2Many.

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: artur.kot@allegro.com or mikolaj.koszowski@allegro.com
• LinkedIn: Artur Kot or Mikołaj Koszowski