license: apache-2.0
Projecte Aina’s Catalan-Chinese machine translation model
Table of Contents
- Model Description
- Intended Uses and Limitations
- How to Use
- Training
- Evaluation
- Additional Information
Model description
This model was trained from scratch using the Fairseq toolkit on a combination of Catalan-Chinese datasets totalling 6.833.114 sentence pairs. 174.507 sentence pairs were parallel data collected from the web while the remaining 6.658.607 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 simplified Chinese.
How to use
Usage
Required libraries:
pip install ctranslate2 pyonmttok
Translate a sentence using python
import ctranslate2
import pyonmttok
import re
def remove_jieba(text):
rejoin_chinese = re.sub(r'(?<!\s)\s(?! )', '', text)
return re.sub(r'\s+', ' ', rejoin_chinese)
from huggingface_hub import snapshot_download
model_dir = snapshot_download(repo_id="projecte-aina/mt-aina-ca-zh", 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]])
translation = tokenizer.detokenize(translated[0][0]['tokens'])
print(remove_jieba(translation))
Training
Training data
The Catalan-Chinese data collected from the web was a combination of the following datasets:
Dataset | Sentences before cleaning |
---|---|
WikiMatrix | 90.643 |
XLENT | 535.803 |
GNOME | 78 |
QED | 3.677 |
TED2020 v1 | 56.269 |
OpenSubtitles | 139.300 |
Total | 882.039 |
The 6.658.607 sentence pairs of synthetic parallel data were created from the following Spanish-Chinese datasets:
Dataset | Sentences before cleaning |
---|---|
UNPC | 17.599.223 |
CCMatrix | 24.051.233 |
MultiParacrawl | 3.410.087 |
Total | 45.060.543 |
Training procedure
Data preparation
The Chinese side of all datasets are passed through the fastlangid language detector and any sentences which are not identified as simplified Chinese are discarded. The datasets are then also 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 to form a final corpus of 6.833.114. The Chinese side of the dataset is tokenized using Jieba 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 17.000 updates. Weights were saved every 1000 updates and reported results are the average of the last 4 checkpoints.
Evaluation
Variable and metrics
We use the BLEU score for evaluation on test sets: Flores-200.
Evaluation results
Below are the evaluation results on the machine translation from Catalan to Chinese compared to Google Translate, M2M 1.2B and NLLB-200's distilled 1.3B variant:
Test set | Google Translate | M2M 1.2B | NLLB 1.3B | mt-aina-ca-zh |
---|---|---|---|---|
Flores Dev | 42,6 | 27,8 | 18,9 | 31,4 |
Flores Devtest | 43,7 | 28,4 | 18,4 | 32,6 |
NTREX | 36,3 | 24,4 | 14,2 | 26,6 |
Average | 41,0 | 26,9 | 17,0 | 30,2 |
Additional information
Author
Language Technologies Unit (LangTech) at the Barcelona Supercomputing Center.
Contact information
For further information, send an email to langtech@bsc.es
Copyright
Copyright Language Technologies Unit at Barcelona Supercomputing Center (2023)
Licensing information
This work is licensed under a Apache License, Version 2.0
Funding
This work was funded by the Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya within the framework of Projecte AINA.