File size: 7,677 Bytes
4ab520c
ee11621
b1fc2b6
 
 
 
 
 
4ab520c
b1fc2b6
3af3027
 
 
b1fc2b6
5176ee4
 
3af3027
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b1fc2b6
3af3027
 
 
 
 
 
 
 
 
b1fc2b6
 
 
 
3af3027
 
 
 
5176ee4
 
3af3027
 
 
 
 
b1fc2b6
 
3af3027
b1fc2b6
5176ee4
3af3027
 
 
 
b1fc2b6
 
3af3027
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b1fc2b6
 
3af3027
 
 
 
 
5176ee4
 
b1fc2b6
 
 
 
 
3af3027
 
 
b1fc2b6
 
3af3027
b1fc2b6
3af3027
f968ade
35c82de
 
 
f968ade
35c82de
 
 
3af3027
 
 
 
b1fc2b6
3af3027
b1fc2b6
 
3af3027
 
b1fc2b6
3af3027
b1fc2b6
3af3027
 
 
b1fc2b6
3af3027
b1fc2b6
3af3027
 
 
 
b1fc2b6
 
 
 
 
 
 
3af3027
b1fc2b6
 
3af3027
b1fc2b6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
---
license: apache-2.0
language:
- ca
- es
metrics:
- bleu
library_name: fairseq
---
## Aina Project's Catalan-Spanish 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 Catalan-Spanish datasets, 
up to 92 million sentences before cleaning and filtering. Additionally, the model is evaluated on several public datasets comprising 5 different domains 
(general, adminstrative, technology, biomedical, and news).  

## Intended uses and limitations

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

## 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-ca-es", revision="main")

tokenizer=pyonmttok.Tokenizer(mode="none", sp_model_path = model_dir + "/spm.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 model was trained on a combination of several datasets, totalling around 92 million parallel sentences before filtering and cleaning. 
The trainig data includes corpora collected from [Opus](https://opus.nlpl.eu/), internally created parallel datsets, and corpora from other sources.

### Training procedure

### Data preparation

 All datasets are concatenated and filtered using the [mBERT Gencata parallel filter](https://huggingface.co/projecte-aina/mbert-base-gencata) and 
 cleaned using the clean-corpus-n.pl script from [moses](https://github.com/moses-smt/mosesdecoder), allowing sentences between 5 and 150 words.

 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 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_bi |
| 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 using shards of 10 million sentences, for a total of 13.000 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 test sets:
[Flores-101](https://github.com/facebookresearch/flores), 
[TaCon](https://elrc-share.eu/repository/browse/tacon-spanish-constitution-mt-test-set/84a96138b98611ec9c1a00155d02670628f3e6857b0f422abd82abc3795ec8c2/), 
[United Nations](https://zenodo.org/record/3888414#.Y33-_tLMIW0), 
[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 Catalan to Spanish compared to [Softcatalà](https://www.softcatala.org/) 
and [Google Translate](https://translate.google.es/?hl=es):

| Test set             | SoftCatalà | Google Translate | aina-translator-ca-es |
|----------------------|------------|------------------|---------------|
| Spanish Constitution | 70,7       | 77,1             | **88,3**      |
| United Nations       | 78,1       | 84,3             | **87,3**      |
| Flores 101 dev       | 23,5       | 24               | **24,2**      |
| Flores 101 devtest   | 24,1       | 24,2             | **24,7**      |
| Cybersecurity        | 67,3       | 76,9             | **78,2**      |
| wmt 19 biomedical    | 60,4       | 62,7             | **64,1**      |
| wmt 13 news          | 22,5       | 23,1             | **23,7**      |
| Average              | 53,4       | 53,2             | **55,1**      |

## 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>