Tran1312/MachineTranslation_en_vi

English–Vietnamese Machine Translation Dataset 2025

Dataset Description

This dataset is a large-scale English–Vietnamese parallel corpus designed for training and evaluating neural machine translation systems.

The corpus was constructed by aggregating sentence pairs from several publicly available English–Vietnamese datasets and applying a multi-stage filtering pipeline to reduce malformed samples, duplicated content, language mismatches, and semantically misaligned sentence pairs.

The resulting dataset contains between 10 million and 100 million parallel sentence pairs after aggregation and filtering.

Supported Task

• Machine Translation
  • English → Vietnamese
  • Vietnamese → English

Languages

• English (en)
• Vietnamese (vi)

---

Data Sources

The dataset was constructed from multiple publicly available English–Vietnamese parallel corpora:

Source	Description / Link
CCMatrix	https://opus.nlpl.eu/CCMatrix/en&vi/v1/CCMatrix
OpenSubtitles	https://opus.nlpl.eu/OpenSubtitles/en&vi/v2024/OpenSubtitles
MultiHPLT	https://opus.nlpl.eu/MultiHPLT/en&vi/v2/MultiHPLT
CCAligned	https://opus.nlpl.eu/CCAligned/en&vi/v1/CCAligned
ParaCrawl	https://opus.nlpl.eu/ParaCrawl-Bonus/en&vi/v9/ParaCrawl-Bonus
PhoMT	https://huggingface.co/datasets/ura-hcmut/PhoMT
VietAI MT / MTet	https://huggingface.co/datasets/wanhin/VietAI_MTet

These sources provide parallel text from heterogeneous domains such as web content, subtitles, crawled documents, and curated machine translation corpora.

---

Data Processing Pipeline

The raw corpora were passed through several stages of rule-based and model-based filtering.

The overall pipeline can be summarized as:

Raw Parallel Corpora
        │
        ▼
Text Normalization
        │
        ▼
Structural / Consistency Filtering
        │
        ▼
Deduplication
        │
        ▼
Language Identification
        │
        ▼
Semantic Alignment Filtering
        │
        ▼
Filtered English–Vietnamese Corpus

---

Basic Data Filtering

Basic filtering was applied to remove malformed, noisy, or obviously misaligned sentence pairs before more expensive model-based filtering.

Alphabetic Ratio Filtering

Both source and target sentences were required to contain sufficient alphabetic content.

The minimum alphabetic-character ratio was:

0.7

Sentence pairs in which either side fell below this threshold were removed.

This step helps eliminate samples dominated by punctuation, markup, numbers, symbols, or corrupted text.

Special-Character De-escaping

HTML entities and escaped TSV characters were converted back to their normal textual representations.

Examples include:

&   → &
&lt;    → <
&gt;    → >
\t      → tab / normalized whitespace

Whitespace Normalization

Repeated whitespace characters were collapsed into a single space.

For example:

"This   is    a sentence."

becomes:

"This is a sentence."

Empty-Sentence Removal

Sentence pairs were discarded when either the English or Vietnamese sequence was empty after normalization.

Numerical Consistency Filtering

Sentence pairs containing inconsistent numerical information between source and target were removed.

This helps eliminate alignment errors in which important quantities, dates, measurements, percentages, or other numerical values differ between the two sentences.

Currency Consistency Filtering

Samples containing inconsistent currency information between source and target were discarded.

This filtering reduces translation pairs where financial values or currency symbols were incorrectly aligned.

URL Consistency Filtering

URLs appearing in one side of a sentence pair were checked against URLs appearing in the corresponding translation.

Pairs containing inconsistent URL information were removed.

Email Consistency Filtering

Email addresses were similarly checked between the source and target sentences.

Pairs containing mismatched email addresses were discarded.

Control-Character Removal

Non-printable and unwanted control characters were removed from the text.

Deduplication

Duplicate source–target sentence pairs were detected using hash-based deduplication and removed.

This reduces unnecessary repetition and prevents highly duplicated samples from disproportionately influencing the training distribution.

---

Language Identification

After basic filtering, language identification was performed using a FastText language identification model.

The expected language configuration is:

Source: English
Target: Vietnamese

Sentence pairs were removed when the detected languages did not match the expected English–Vietnamese language pair.

This stage is particularly useful for large web-crawled corpora, where incorrectly classified languages and multilingual fragments may occur frequently.

---

Semantic Alignment Filtering

Language correctness alone does not guarantee that two sentences are valid translations of each other.

To detect semantically misaligned pairs, the dataset uses LaBSE — Language-agnostic BERT Sentence Embedding.

Each English sentence and its corresponding Vietnamese sentence are encoded into a shared multilingual embedding space.

Conceptually:

English sentence ──► LaBSE ──► embedding_en
                                  │
                                  │ cosine similarity
                                  ▼
Vietnamese sentence ─► LaBSE ─► embedding_vi

The semantic similarity between both embeddings is then measured.

Only sentence pairs satisfying:

similarity >= 0.8

are retained.

This stage removes sentence pairs that may contain valid English and Vietnamese text individually but do not express sufficiently similar semantic content.

---

Filtering Strategy

The filtering process follows a coarse-to-fine design:

Cheap deterministic filters
          │
          ▼
Remove obvious noise
          │
          ▼
Language identification
          │
          ▼
Semantic embedding model
          │
          ▼
High-quality parallel pairs

Computationally inexpensive operations such as normalization, numerical checks, and deduplication are performed before FastText and LaBSE inference.

This reduces the number of samples that need to pass through the more computationally expensive neural filtering stages.

---

Intended Uses

The dataset is primarily intended for:

• English → Vietnamese machine translation
• Vietnamese → English machine translation
• Neural Machine Translation pretraining
• Translation model fine-tuning
• Sequence-to-sequence modeling
• Cross-lingual representation learning
• English–Vietnamese alignment research
• Data filtering and machine translation experiments

It can be used with architectures such as:

Transformer Encoder–Decoder
        │
        ├── Transformer-base / Transformer-big
        ├── mBART
        ├── mT5
        ├── NLLB
        └── other multilingual sequence-to-sequence models

---

Data Quality Considerations

The filtering pipeline significantly reduces common sources of noise, but it does not guarantee perfect translation equivalence.

Residual issues may include:

• subtle semantic mismatches;
• incomplete translations;
• domain-specific terminology errors;
• named-entity inconsistencies not captured by the filtering rules;
• unusual formatting inherited from source datasets;
• naturally noisy subtitle or web-crawled text;
• duplicate or near-duplicate sentences not detected by exact hash-based deduplication.

The LaBSE threshold of 0.8 prioritizes stronger semantic alignment but may also remove legitimate translation pairs whose lexical or structural forms differ substantially across languages.

---

Data Sources, Attribution, and Licensing

This dataset was constructed by aggregating and filtering English–Vietnamese parallel data obtained from multiple external datasets.

The inclusion of a dataset in this corpus does not imply that all constituent datasets share the same license or redistribution terms. Each upstream dataset remains subject to its own license, terms of use, attribution requirements, and other applicable restrictions.

Users should review the original source documentation before redistributing the corpus or using it in settings that may be subject to additional legal or commercial requirements.

Source	Upstream Project / Publication	License / Redistribution Notes	Citation / Acknowledgment
CCMatrix	OPUS / CCMatrix	Consult the corpus-specific OPUS documentation and the original CCMatrix distribution terms before redistribution.	Cite OPUS and the relevant CCMatrix work where applicable.
OpenSubtitles	OPUS / OpenSubtitles	Subtitle text may originate from copyrighted audiovisual works or user-contributed subtitles. Availability through OPUS should not be interpreted as transferring copyright ownership of the underlying text. Review corpus-specific terms before redistribution.	Cite OPUS and acknowledge OpenSubtitles as the upstream source.
MultiHPLT	HPLT / OPUS	The OPUS MultiHPLT v2 packaging is released under CC0. OPUS requests acknowledgment of both the HPLT project and OPUS.	Acknowledge HPLT and OPUS.
CCAligned	OPUS / CCAligned	Consult the upstream corpus documentation for source-specific licensing and redistribution conditions.	Cite OPUS and relevant CCAligned publications where applicable.
ParaCrawl	ParaCrawl / OPUS	Consult the ParaCrawl release and corpus-specific OPUS documentation for the exact redistribution conditions of the version used.	Cite the relevant ParaCrawl release and OPUS where applicable.
PhoMT	VinAI Research	The referenced Hugging Face copy declares CC BY-NC-ND 4.0. Because this license includes NonCommercial and NoDerivatives restrictions, redistribution of modified, filtered, or incorporated copies should be reviewed carefully against the terms of the original release.	Cite the original PhoMT paper.
MTet / VietAI MT	VietAI	The associated MTet publication is available under CC BY 4.0; however, the license of the paper should not automatically be assumed to be the license of the dataset itself. Dataset distribution terms should be verified separately.	Cite the MTet publication and acknowledge VietAI.

---

Citations

OPUS

Several source corpora in this dataset were obtained through OPUS — the Open Parallel Corpus.

When using data distributed through OPUS, please cite:

@inproceedings{tiedemann2012opus,
  author    = {J{\"o}rg Tiedemann},
  title     = {Parallel Data, Tools and Interfaces in OPUS},
  booktitle = {Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12)},
  year      = {2012},
  address   = {Istanbul, Turkey},
  publisher = {European Language Resources Association (ELRA)}
}

Corpus-specific publications and acknowledgments should additionally be cited where required by the corresponding upstream dataset.

PhoMT

Please cite the original PhoMT publication:

@inproceedings{doan2021phomt,
  title     = {PhoMT: A High-Quality and Large-Scale Benchmark Dataset for Vietnamese-English Machine Translation},
  author    = {Doan, Long and Nguyen, Linh The and Tran, Nguyen Luong and Hoang, Thai and Nguyen, Dat Quoc},
  booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  year      = {2021},
  pages     = {4495--4503}
}

MTet

For MTet / VietAI data, please cite:

@article{ngo2022mtet,
  title   = {MTet: Multi-domain Translation for English and Vietnamese},
  author  = {Ngo, Chinh and Trinh, Trieu H. and Phan, Long and Tran, Hieu and Dang, Tai and Nguyen, Hieu and Nguyen, Minh and Luong, Minh-Thang},
  journal = {arXiv preprint arXiv:2210.05610},
  year    = {2022}
}

---

Dataset Provenance and Copyright Notice

This dataset is an aggregated and filtered derivative corpus created from multiple independently published datasets.

The dataset maintainers do not claim ownership of the original textual content contained in the upstream corpora.

Copyright, database rights, neighboring rights, contractual restrictions, and other rights associated with individual source datasets or their underlying documents remain with their respective rights holders where applicable.

Processing performed in this dataset includes normalization, filtering, deduplication, language identification, and semantic-alignment filtering. These operations improve corpus quality but do not supersede or replace the terms governing the original source data.

Licensing Status

Because the corpus combines data originating from sources with heterogeneous and, in some cases, potentially restrictive licensing conditions, no single permissive license is currently asserted over the entire aggregated corpus.

Accordingly, the dataset metadata uses:

license: unknown

This designation should not be interpreted as meaning that the dataset is free of copyright restrictions or equivalent to public-domain material.

Users are responsible for determining whether their intended use, including research use, commercial use, modification, redistribution, or publication of derived datasets, complies with the licenses and terms of the corresponding upstream sources.

Redistribution Notice

Particular care should be taken when redistributing the processed corpus.

Some upstream resources may permit research use while imposing restrictions on redistribution, commercial use, or the distribution of modified versions.

Where redistribution rights for a source cannot be established with sufficient confidence, a safer distribution strategy is to release:

• the filtering and preprocessing code;
• source identifiers and provenance metadata;
• scripts for reconstructing the processed dataset from legally obtained upstream copies;
• hashes or sample identifiers used for filtering;

rather than redistributing the underlying sentence pairs themselves.

No Legal Warranty

The provenance and licensing information provided here is intended for documentation purposes and does not constitute legal advice.

Users should consult the original license and terms of each upstream dataset when determining whether a particular form of use or redistribution is permitted.

---

Keywords

English-Vietnamese
Machine Translation
Parallel Corpus
Neural Machine Translation
NMT
NLP
OPUS
FastText
LaBSE
Semantic Filtering
Language Identification
Data Cleaning
Cross-lingual Learning