license: apache-2.0
datasets:
- projecte-aina/CA-DE_Parallel_Corpus
language:
- ca
- de
metrics:
- bleu
library_name: fairseq
Projecte Aina’s Catalan-German machine translation model
Model description
This model was trained from scratch using the Fairseq toolkit on a combination of Catalan-German datasets, totalling 100.000.000 sentence pairs. 6.258.272 sentence pairs were parallel data collected from the web while the remaining 93.741.728 sentence pairs were parallel synthetic data created using the ES-CA translator of PlanTL. The model was evaluated on the Flores and NTREX evaluation datasets.
Intended uses and limitations
You can use this model for machine translation from Catalan to German.
How to use
Usage
Required libraries:
pip install ctranslate2 pyonmttok
Translate a sentence using python
import ctranslate2
import pyonmttok
from huggingface_hub import snapshot_download
model_dir = snapshot_download(repo_id="projecte-aina/aina-translator-ca-de", revision="main")
tokenizer=pyonmttok.Tokenizer(mode="none", sp_model_path = model_dir + "/spm.50k.model")
tokenized=tokenizer.tokenize("Benvingut al projecte Aina!")
translator = ctranslate2.Translator(model_dir)
translated = translator.translate_batch([tokenized[0]])
print(tokenizer.detokenize(translated[0][0]['tokens']))
Limitations and bias
At the time of submission, no measures have been taken to estimate the bias and toxicity embedded in the model. However, we are well aware that our models may be biased. We intend to conduct research in these areas in the future, and if completed, this model card will be updated.
Training
Training data
The Catalan-German data collected from the web was a combination of the following datasets:
| Dataset | Sentences | Sentences after Cleaning |
|---|---|---|
| Multi CCAligned | 1.478.152 | 1.027.481 |
| WikiMatrix | 180.322 | 125.811 |
| GNOME | 12.333 | 1.241 |
| KDE4 | 165.439 | 105.098 |
| QED | 63.041 | 49.181 |
| TED2020 v1 | 46.680 | 38.428 |
| OpenSubtitles | 303.329 | 171.376 |
| GlobalVoices | 4.636 | 3.578 |
| Tatoeba | 732 | 655 |
| Books | 4.445 | 2049 |
| Europarl | 1.734.643 | 1.734.643 |
| Tilde | 3.434.091 | 3.434.091 |
| Total | 7.427.843 | 6.258.272 |
All corpora except Europarl and Tilde were collected from Opus. The Europarl and Tilde corpora are synthetic parallel corpora created from the original Spanish-German corpora by SoftCatalà.
The 93.741.728 sentence pairs of synthetic parallel data were created from the following Spanish-German datasets:
| Dataset | Sentences before cleaning |
|---|---|
| globalvoices_es-de_20230901 | 70.097 |
| multiparacrawl_es-de_20230901 | 56.873.541 |
| dgt_es-de_20240129 | 4.899.734 |
| eubookshop_es-de_20240129 | 4.750.170 |
| nllb_es-de_20240129 | 112.444.838 |
| opensubtitles_es-de_20240129 | 18.951.214 |
| Total | 197.989.594 |
Training procedure
Data preparation
All datasets are deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75. This is done using sentence embeddings calculated using LaBSE. The filtered datasets are then concatenated and before training the punctuation is normalized using a modified version of the join-single-file.py script from SoftCatalà
Tokenization
All data is tokenized using sentencepiece, with a 50 thousand token sentencepiece model learned from the combination of all filtered training data. This model is included.
Hyperparameters
The model is based on the Transformer-XLarge proposed by Subramanian et al. The following hyperparameters were set on the Fairseq toolkit:
| Hyperparameter | Value |
|---|---|
| Architecture | transformer_vaswani_wmt_en_de_big |
| Embedding size | 1024 |
| Feedforward size | 4096 |
| Number of heads | 16 |
| Encoder layers | 24 |
| Decoder layers | 6 |
| Normalize before attention | True |
| --share-decoder-input-output-embed | True |
| --share-all-embeddings | True |
| Effective batch size | 48.000 |
| Optimizer | adam |
| Adam betas | (0.9, 0.980) |
| Clip norm | 0.0 |
| Learning rate | 5e-4 |
| Lr. schedurer | inverse sqrt |
| Warmup updates | 8000 |
| Dropout | 0.1 |
| Label smoothing | 0.1 |
The model was trained for a total of 22.000 updates. Weights were saved every 1000 updates and reported results are the average of the last 3 checkpoints.
Evaluation
Variable and metrics
We use the BLEU score for evaluation on the Flores-101 and NTREX test sets.
Evaluation results
Below are the evaluation results on the machine translation from Catalan to German compared to Softcatalà and Google Translate:
| Test set | SoftCatalà | Google Translate | aina-translator-ca-de |
|---|---|---|---|
| Flores 101 dev | 26,2 | 34,8 | 34,1 |
| Flores 101 devtest | 26,3 | 34,0 | 33,3 |
| NTREX | 21,7 | 28,8 | 27,8 |
| Average | 24,7 | 32,5 | 31,7 |
Additional information
Author
The Language Technologies Unit from Barcelona Supercomputing Center.
Contact
For further information, please send an email to langtech@bsc.es.
Copyright
Copyright(c) 2023 by Language Technologies Unit, Barcelona Supercomputing Center.
License
Funding
This work has been promoted and financed by the Generalitat de Catalunya through the Aina project.
Disclaimer
Click to expand
The model published in this repository is intended for a generalist purpose and is available to third parties under a permissive Apache License, Version 2.0.
Be aware that the model may have biases and/or any other undesirable distortions.
When third parties deploy or provide systems and/or services to other parties using this model (or any system based on it) or become users of the model, they should note that it is their responsibility to mitigate the risks arising from its use and, in any event, to comply with applicable regulations, including regulations regarding the use of Artificial Intelligence.
In no event shall the owner and creator of the model (Barcelona Supercomputing Center) be liable for any results arising from the use made by third parties.