Parallel corpora for machine translation training
Sentence-aligned bilingual data built to order: new human translations of your source text, or your existing translated documents aligned, cleaned and checked by translators. Delivered as TMX, TSV or JSONL with the metadata your training pipeline needs.
corpus_en-fr.tmx3 clean pairs, domain: consumer electronics- Sources
- New translations, your translated documents or your translation memories
- Checked by
- Professional translators in each language pair
- Formats
- TMX, TSV, JSONL or CSV, with domain and source metadata
- Rights
- Built from text you own or text cleared for training use
What a parallel corpus is
A parallel corpus is a set of texts paired with their translations, aligned sentence by sentence. Machine translation models and multilingual language models learn from these pairs, so errors in the data turn into errors in the model. Misaligned pairs, duplicates and machine-translated targets are the most common problems in data scraped from the web.
Human-built corpora are smaller than scraped ones, but they're accurate, they fit your domain, and you know where every sentence came from.
What we build
From your content, in your language pairs and domain.
New human translation
Your source sentences translated by professionals, with no machine translation in the target.
Alignment of your documents
Translated manuals, contracts or web pages split into sentences and paired, with every pair checked.
Cleaning existing data
Translation memories and scraped pairs filtered for misalignment, duplicates, wrong language and poor translation.
Domain corpora
Legal, medical, financial or technical pairs, translated by people who work in that field.
Multi-way data
The same source translated into several languages, for multilingual models and fair comparisons.
Test and dev sets
Held-out pairs kept separate from training data, for honest evaluation. More on MT evaluation.
How a corpus project runs
You see a sample file before the full build.
Spec
Language pairs, domain, size, source text and output format.
Sample
A small file per pair for your team to load and check.
Build
Translation or alignment by professional translators.
Check
A second translator reviews a sample of every batch, and automatic checks catch duplicates.
Second review QADelivery
Clean files with metadata and a data sheet describing the corpus.
Related AI language services
Questions about parallel corpora
What is a parallel corpus?
A parallel corpus is a collection of texts and their translations, aligned so that each sentence is paired with its translation. It's the main training data for machine translation models.
Can you build a corpus from our own documents?
Yes. If you have translated documents or translation memories, we align and clean them. If you only have source text, we translate it.
What formats do you deliver?
TMX, TSV, CSV or JSONL, with fields for source, target, language codes, domain and any metadata you need.
Who owns the corpus?
Corpora built from your text for your project are delivered to you under the terms agreed in the contract. We only use text you own or text cleared for training use.
Scope a corpus
Tell us the language pairs, domain and size you need, or book a call. We send a sample file before the full build.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time