liboaccn/nmt-parallel-corpus

Neural Machine Translation parallel corpora

Introduction

We use OpusTools to extract resources from the OPUS project, a renowned platform for parallel corpora, and create a multilingual dataset. Specifically, we collect the parallel corpora from prominent projects within OPUS, including NLLB, CCMatrix, and OpenSubtitles.

This comprehensive data collection process results in a corpus of more than 3T, covering 60 languages and over 1900 language pairs.

Preprocessing

High-quality parallel corpora are essential for training powerful NMT models. However, raw data, such as web-crawled content, often contains significant noise—including length inconsistencies, irrelevant content, and sensitive material—which can negatively impact model performance. To address this, we adopt a six-stage cleaning pipeline inspired by previous work to ensure high-quality multilingual parallel corpora.

1. Text Extraction and Preprocessing

• Compressed files for each target language pair are obtained from the OPUS corpus.
• After decompression, only customized plain-text files in Moses format are retained.
• For example, in English-Chinese translation, only .en and .zh files are preserved.

2. Proportion of Characters

To remove noisy or irrelevant sentences, we apply the following filters:

• Punctuation ratio filtering: Sentences with punctuation exceeding 50% are discarded.
• Rule-based filtering: Sentences with only spaces, invalid UTF8 characters, or excessively long tokens (e.g., DNA-like sequences) are removed.
• Character ratio filtering: Sentences with a low proportion of target language characters are eliminated to ensure relevance.

3. Data Length Filtering

Sentence length inconsistencies are controlled as follows:

• Tokenization: Both source and target texts are tokenized using SentencePiece.
• Length ratio filtering: Sentence pairs with one side exceeding three times the length of the other are removed.
• Short text removal: Documents with average line lengths < 10 words or total lengths > 250 characters are discarded.

4. Sensitive Word Filtering

To prevent the model from learning harmful language:

• A predefined list of sensitive words is used.
• Sentences with high-frequency sensitive words (e.g., frequency > 0.5) are removed.

5. Duplication Removal

• Duplicate sentence pairs are identified and removed using a deduplication script.
• Only the first occurrence of each pair is retained.

6. Normalization

• Text is normalized to unify punctuation, numbers, and space formatting.
• Unicode standardization ensures consistent symbol encoding.
• All quotation marks are normalized to a single standard format to reduce vocabulary size and improve model efficiency.

Citation Information

You can cite our paper https://arxiv.org/abs/2505.14256

@misc{zhu2025fuximtsparsifyinglargelanguage,
      title={FuxiMT: Sparsifying Large Language Models for Chinese-Centric Multilingual Machine Translation}, 
      author={Shaolin Zhu and Tianyu Dong and Bo Li and Deyi Xiong},
      year={2025},
      eprint={2505.14256},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.14256}, 
}