projecte-aina/aina-translator-ca-zh

Projecte Aina’s Catalan-Chinese machine translation model

Table of Contents

Click to expand

• Model description
• Intended uses and limitations
• How to use
• Limitations and bias
• Training
• Evaluation
• Additional information

Model description

This machine translation model is built upon the M2M100 1.2B, fine-tuned specifically for Catalan-Chinese translation. It is trained on a combination of Catalan-Chinese datasets totalling 94,187,858 sentence pairs. 113,305 sentence pairs were parallel data collected from the web, while the remaining 94,074,553 sentence pairs were parallel synthetic data created using the Aina Project's Spanish-Catalan machine translation model and the Aina Project's English-Catalan machine translation model.

Following the fine-tuning phase, Contrastive Preference Optimization (CPO) was applied to further refine the model's outputs. CPO training involved pairs of "chosen" and "rejected" translations for a total of 4,006 sentences. These sentences were sourced from the Flores development set (997 sentences), the Flores devtest set (1,012 sentences), and the NTREX set (1,997 sentences).

The model was evaluated on the Projecte Aina's Catalan-Chinese evaluation dataset (unpublished), achieving results comparable to those of Google Translate.

Intended uses and limitations

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

How to use

Usage

Translate a sentence using python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "projecte-aina/aina-translator-ca-zh"

model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

sentence = "Benvingut al projecte Aina!"

input_ids = tokenizer(sentence, return_tensors="pt").input_ids
output_ids = model.generate(input_ids, max_length=200, num_beams=5)

generated_translation= tokenizer.decode(output_ids[0], skip_special_tokens=True, spaces_between_special_tokens = False).strip()
print(generated_translation)
#欢迎来到 Aina 项目！

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-Chinese data collected from the web was a combination of the following datasets:

Dataset	Sentences before cleaning
OpenSubtitles	139,300
WikiMatrix	90,643
Wikipedia	68,623
Total	298,566

94,074,553 sentence pairs of synthetic parallel data were created from the following Spanish-Chinese datasets and English-Chinese datasets:

Spanish-Chinese:

Dataset	Sentences before cleaning
NLLB	24,051,233
UNPC	17,599,223
MultiUN	9,847,770
OpenSubtitles	9,319,658
MultiParaCrawl	3,410,087
MultiCCAligned	3,006,694
WikiMatrix	1,214,322
News Commentary	375,982
Tatoeba	9,404
Total	68,834,373

English-Chinese:

Dataset	Sentences before cleaning
NLLB	71,383,325
CCAligned	15,181,415
Paracrawl	14,170,869
WikiMatrix	2,595,119
Total	103,330,728

Training procedure

Data preparation

Catalan-Chinese parallel data

The Chinese side of all datasets were first processed using the Hanzi Identifier to detect Traditional Chinese, which was subsequently converted to Simplified Chinese using OpenCC.

All data was then filtered according to two specific criteria:

• Alignment: sentence level alignments were calculated using LaBSE and sentence pairs with a score below 0.75 were discarded.

• Language identification: the probability of being the target language was calculated using Lingua.py and sentences with a language probability score below 0.5 were discarded.

Next, Spanish data was translated into Catalan using the Aina Project's Spanish-Catalan machine translation model, while English data was translated into Catalan using the Aina Project's English-Catalan machine translation model.

The filtered and translated datasets are then concatenated and deduplicated to form a final corpus of 94,187,858.

Catalan-Chinese Contrastive Preference Optimization dataset

The CPO dataset is built by comparing the quality of translations across four distinct sources:

• Reference translation: Chinese sentences from Flores test set, Flores devtest set, and NTREX dataset.
• aina-translator-ca-zh: A specialized bilingual model for Catalan-Chinese translations.
• Google Translate: A widely-used general-purpose machine translation system.
• OpenAI GPT-4: A large-scale language model capable of performing a wide range of tasks in conversational settings, including high-quality translation.

To evaluate the quality of translations without relying on human annotations, we employ two reference-free evaluation models:

• Unbabel/wmt23-cometkiwi-da-xxl
• Unbabel/XCOMET-XXL

These models provide direct assessment scores for each translation. The scores from both models are averaged to determine the relative quality of each translation. Based on this evaluation, the highest-scoring ("chosen") and lowest-scoring ("rejected") translations are identified for each source sentence, forming contrastive pairs. The CPO dataset comprises a total of 4,006 such pairs of "chosen" and "rejected" translations.

Training

The training was executed on NVIDIA GPUs utilizing the Hugging Face Transformers framework. The model was trained for 245,000 updates.

Following fine-tuning on the M2M100 1.2B model, Contrastive Preference Optimization (CPO) was performed using our CPO dataset and the Hugging Face CPO Trainer. This phase involved 1,500 updates.

Evaluation

Variable and metrics

Below are the evaluation results on the Projecte Aina's Catalan-Chinese test set (unpublished), compared to Google Translate for the CA-ZH direction. The evaluation was conducted using tower-eval following the standard setting (beam search with beam size 5, limiting the translation length to 200 tokens). We report the following metrics:

• BLEU: Sacrebleu implementation, version:2.4.0
• ChrF: Sacrebleu implementation.
• Comet: Model checkpoint: "Unbabel/wmt22-comet-da".
• Comet-kiwi: Model checkpoint: "Unbabel/wmt22-cometkiwi-da".

Evaluation results

Below are the evaluation results on the machine translation from Catalan to Chinese compared to Google Translate:

Projecte Aina's Catalan-Chinese evaluation dataset

	Bleu ↑	ChrF ↑	Comet ↑	Comet-kiwi ↑
aina-translator-ca-zh	43.88	40.19	0.87	0.81
Google Translate	44.64	41.15	0.87	0.80

Additional information

Author

The Language Technologies Unit from Barcelona Supercomputing Center.

Contact

For further information, please send an email to [email protected].

Copyright

Copyright(c) 2023 by Language Technologies Unit, Barcelona Supercomputing Center.

License

Apache License, Version 2.0

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.