README.md

---
license: apache-2.0
datasets:
- projecte-aina/CA-EN_Parallel_Corpus
language:
- en
- ca
metrics:
- bleu
library_name: fairseq
---
## Projecte Aina's English-Catalan machine translation model
  
## Model description

This model was trained from scratch using the [Fairseq toolkit](https://fairseq.readthedocs.io/en/latest/) on a combination of English-Catalan datasets, 
which after filtering and cleaning comprised 30.023.034 sentence pairs. The model was evaluated on the Flores and NTREX evaluation datasets.  

## Intended uses and limitations

You can use this model for machine translation from English to Catalan. 

## How to use

### Usage
Required libraries:

```bash
pip install ctranslate2 pyonmttok
```

Translate a sentence using python 
```python
import ctranslate2
import pyonmttok
from huggingface_hub import snapshot_download
model_dir = snapshot_download(repo_id="projecte-aina/aina-translator-en-ca", revision="main")

tokenizer=pyonmttok.Tokenizer(mode="none", sp_model_path = model_dir + "/spm.model")
tokenized=tokenizer.tokenize("Welcome to the Aina Project!")

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 model was trained on a combination of several datasets, including data collected from [Opus](https://opus.nlpl.eu/), [HPLT](https://hplt-project.org/), 
an internally created [CA-EN Parallel Corpus](https://huggingface.co/datasets/projecte-aina/CA-EN_Parallel_Corpus), and other sources.

### 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](https://huggingface.co/sentence-transformers/LaBSE). 
 The filtered datasets are then concatenated to form a final corpus of 30.023.034 parallel sentences and before training the punctuation 
 is normalized using a modified version of the join-single-file.py script from [SoftCatalà](https://github.com/Softcatala/nmt-models/blob/master/data-processing-tools/join-single-file.py)

#### 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.](https://aclanthology.org/2021.wmt-1.18.pdf)
The following hyperparamenters 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               | 96.000                           |
| Optimizer                          | adam                             |
| Adam betas                         | (0.9, 0.980)                     |
| Clip norm                          | 0.0                              |
| Learning rate                      | 1e-3                             |
| Lr. schedurer                      | inverse sqrt                     |
| Warmup updates                     | 4000                             |
| Dropout                            | 0.1                              |
| Label smoothing                    | 0.1                              |

The model was trained for a total of 16000 updates. Weights were saved every 1000 updates and reported results are the average of the last 6 checkpoints. 

## Evaluation

### Variable and metrics

We use the BLEU score for evaluation on the following test sets: 
[Spanish Constitution (TaCon)](https://elrc-share.eu/repository/browse/tacon-spanish-constitution-mt-test-set/84a96138b98611ec9c1a00155d02670628f3e6857b0f422abd82abc3795ec8c2/), 
[United Nations](https://zenodo.org/record/3888414#.Y33-_tLMIW0), 
[AAPP](https://elrc-share.eu/repository/browse/catalan-spanish-catgov-corpus/8088130a722811ed9c1a00155d02670690607f8261a847549c8a0583cbe729da/),
[European Commission](https://elrc-share.eu/repository/browse/european-commission-corpus/8a419b1758ea11ed9c1a00155d0267069bd085cae124481589b0858e5b274327/), 
[Flores-200](https://github.com/facebookresearch/flores), 
[Cybersecurity](https://elrc-share.eu/repository/browse/cyber-mt-test-set/2bd93faab98c11ec9c1a00155d026706b96a490ed3e140f0a29a80a08c46e91e/), 
[wmt19 biomedical test set](http://www.statmt.org/wmt19/biomedical-translation-task.html), 
[wmt13 news test set](https://elrc-share.eu/repository/browse/catalan-wmt2013-machine-translation-shared-task-test-set/84a96139b98611ec9c1a00155d0267061a0aa1b62e2248e89aab4952f3c230fc/).

### Evaluation results

Below are the evaluation results on the machine translation from English to Catalan compared to [Softcatalà](https://www.softcatala.org/) and 
[Google Translate](https://translate.google.es/?hl=es):


| Test set             | SoftCatalà | Google Translate | aina-translator-en-ca |
|----------------------|------------|------------------|---------------|
| Spanish Constitution | 32,6       | 37,8             | **41,2**      |
| United Nations       | 39,0       | 40,5             | **41,2**      |
| AAPP                 | 46,5       | 51,40            | **51,70**     |
| European Commission  | 49,1       | **52,0**         | 51            |
| Flores 200 dev       | 41,0       | **45,1**         | 43,3          |
| Flores 200 devtest   | 42,1       | **46,0**         | 44,1          |
| Cybersecurity        | 42,5       | **48,1**         | 45,8          |
| wmt 19 biomedical    | 21,7       | 25,5             | **26,7**      |
| wmt 13 news          | 34,9       | **35,7**         | 34,0          |
| **Average**          | 37,9       | **41,34**        | 40,91         |

## 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
[Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0)

### Funding
This work has been promoted and financed by the Generalitat de Catalunya through the [Aina project](https://projecteaina.cat/).

### Disclaimer

<details>
<summary>Click to expand</summary>

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.

</details>