MultiSlav BiDi Models

Multilingual BiDi MT Models

BiDi is a collection of Encoder-Decoder vanilla transformer models trained on sentence-level Machine Translation task. Each model is supporting Bi-Directional translation.

BiDi models are 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.

Graphic above provides an example of an BiDi model - BiDi-ces-pol to translate from Polish to Czech language. BiDi-ces-pol is a bi-directional model supporting translation both form Czech to Polish and from Polish to Czech directions.

Supported languages

To use a BiDi 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. Note that, for each model only two directions are available. 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<<`

Bi-Di models available

We provided 10 BiDi models, allowing to translate between 20 languages.

Bi-Di model	Languages supported	HF repository
BiDi-ces-eng	Czech ↔ English	allegro/BiDi-ces-eng
BiDi-ces-pol	Czech ↔ Polish	allegro/BiDi-ces-pol
BiDi-ces-slk	Czech ↔ Slovak	allegro/BiDi-ces-slk
BiDi-ces-slv	Czech ↔ Slovene	allegro/BiDi-ces-slv
BiDi-eng-pol	English ↔ Polish	allegro/BiDi-eng-pol
BiDi-eng-slk	English ↔ Slovak	allegro/BiDi-eng-slk
BiDi-eng-slv	English ↔ Slovene	allegro/BiDi-eng-slv
BiDi-pol-slk	Polish ↔ Slovak	allegro/BiDi-pol-slk
BiDi-pol-slv	Polish ↔ Slovene	allegro/BiDi-pol-slv
BiDi-slk-slv	Slovak ↔ Slovene	allegro/BiDi-slk-slv

Use case quickstart

Example code-snippet to use model. Due to bug the MarianMTModel must be used explicitly. Remember to adjust source and target languages to your use-case.

from transformers import AutoTokenizer, MarianMTModel

source_lang = "pol"
target_lang = "ces"
first_lang, second_lang = sorted([source_lang, target_lang])
model_name = f"Allegro/BiDi-{first_lang}-{second_lang}"

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

text = f">>{target_lang}<<" + " " + "Allegro to internetowa platforma e-commerce, na której swoje produkty sprzedają średnie i małe firmy, jak również duże marki."

batch_to_translate = [text]
translations = model.generate(**tokenizer.batch_encode_plus(batch_to_translate, return_tensors="pt"))
decoded_translation = tokenizer.batch_decode(translations, skip_special_tokens=True, clean_up_tokenization_spaces=True)[0]

print(decoded_translation)

Generated Czech output:

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

Training

SentencePiece tokenizer has a vocab size 32k 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	209M
Vocab Size	32k
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
Languages Supported	See Bi-Di models available
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. BiDi models are our baseline before expanding the data-regime by using higher-level multilinguality.

Datasets were downloaded via MT-Data library. The number of total examples post filtering and deduplication varies, depending on languages supported, see the table below.

Language pair	Number of training examples
Czech ↔ Polish	63M
Czech ↔ Slovak	30M
Czech ↔ Slovene	25M
Polish ↔ Slovak	26M
Polish ↔ Slovene	23M
Slovak ↔ Slovene	18M
----------------	-------------------------------
Czech ↔ English	151M
English ↔ Polish	150M
English ↔ Slovak	52M
English ↔ Slovene	40M

The datasets used (only applicable to specific directions):

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. Metric used: Unbabel/wmt22-comet-da.

Direction	CES → ENG	CES → POL	CES → SLK	CES → SLV	ENG → CES	ENG → POL	ENG → SLK	ENG → SLV	POL → CES	POL → ENG	POL → SLK	POL → SLV	SLK → CES	SLK → ENG	SLK → POL	SLK → SLV	SLV → CES	SLV → ENG	SLV → POL	SLV → SLK
M2M-100	87.0	89.0	92.1	89.7	88.6	86.4	88.4	87.3	89.6	84.6	89.4	88.4	92.7	86.8	89.1	89.6	90.3	86.4	88.7	90.1
NLLB-200	88.1	88.9	91.2	88.6	90.4	88.5	90.1	88.8	89.4	85.8	88.9	87.7	91.8	88.2	88.9	88.8	90.0	87.5	88.6	89.4
Seamless-M4T	87.5	80.9	90.8	82.0	90.7	88.5	90.6	89.6	79.6	85.4	80.0	76.4	91.5	87.2	81.2	82.9	80.9	87.3	76.7	81.0
OPUS-MT Sla-Sla	88.2	82.8	-	83.4	89.1	85.6	-	84.5	82.9	82.2	-	81.2	-	-	-	-	83.5	84.1	80.8	-
OPUS-MT SK-EN	-	-	-	-	-	-	89.5	-	-	-	-	-	-	88.4	-	-	-	-	-	-
Our contributions:
BiDi Models*	87.5	89.4	92.4	89.8	87.8	86.2	87.2	86.6	90.0	85.0	89.1	88.4	92.9	87.3	88.8	89.4	90.0	86.9	88.1	89.1
P4-pol◊	-	89.6	90.8	88.7	-	-	-	-	90.2	-	89.8	88.7	91.0	-	89.3	88.4	89.3	-	88.7	88.5
P5-eng◊	88.0	89.0	90.7	89.0	88.8	87.3	88.4	87.5	89.0	85.7	88.5	87.8	91.0	88.2	88.6	88.5	89.6	87.2	88.4	88.9
P5-ces◊	87.9	89.6	92.5	89.9	88.4	85.0	87.9	85.9	90.3	84.5	89.5	88.0	93.0	87.8	89.4	89.8	90.3	85.7	87.9	89.8
MultiSlav-4slav	-	89.7	92.5	90.0	-	-	-	-	90.2	-	89.6	88.7	92.9	-	89.4	90.1	90.6	-	88.9	90.2
MultiSlav-5lang	87.8	89.8	92.5	90.1	88.9	86.9	88.0	87.3	90.4	85.4	89.8	88.9	92.9	87.8	89.6	90.2	90.6	87.0	89.2	90.2

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

* results combined for all bi-directional models; each values for applicable model

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

allegro
/

BiDi-pol-slv