Parallel documents are pairs of documents where one is the translation of the other. One document in the pair contains sentences in the source language and the other document contains these sentences translated into the target language. It doesn't matter which language is marked as "source" and which language is marked as "target" – a parallel document can be used to train a translation system in either direction.
You need a minimum of 10,000 unique aligned parallel sentences to train a system. This limitation is a safety net to ensure your parallel sentences contain enough unique vocabulary to successfully train a translation model. As a best practice, continuously add more parallel content and retrain to improve the quality of your translation system. For more information, see Sentence Alignment.
Microsoft requires that documents uploaded to the Custom Translator don't violate a third-party copyright or intellectual properties. For more information, please see the Terms of Use. Uploading a document using the portal doesn't alter the ownership of the intellectual property in the document itself.
Parallel documents are used by the system:
As a best practice, make sure that there's a 1:1 sentence correspondence between the source and target language versions of the documents.
If your project is domain (category) specific, your documents should be consistent in terminology within that category. The quality of the resulting translation system depends on the number of sentences in your document set and the quality of the sentences. The more examples your documents contain with diverse usages for a word specific to your category, the better job the system can do during translation.
Documents uploaded are private to each workspace and can be used in as many projects or trainings as you like. Sentences extracted from your documents are stored separately in your repository as plain Unicode text files and are available for you to delete. Don't use the Custom Translator as a document repository, you can't download the documents in the same format that was uploaded.