Skip to main content
Translator

Microsoft Translator Blog

Customize your neural translations with the new Translator custom feature

Translate your company’s content better than ever by customizing Microsoft Translator’s state-of-the-art Neural Machine Translation (NMT) with the new Translator custom feature, announced this week at Microsoft Build.

Use the new custom feature with the new Translator text translation API (V3) or the preview of the new unified Speech service. With the custom feature, you can now customize not only text translation, but speech translation as well.

With as few as 10,000 parallel sentences, i.e. human-generated translations, you can begin customizing your translations. The more data you have, the better the quality. For instance, below are a few examples of BLEU score (an industry standard translation quality measure) improvements for various datasets size and content type. The gain over the default NMT system can vary from go from a noticeable few BLEU points to a remarkable jump in quality of over 10 BLEU points.

 

To make training easier, the system also accepts data that is parallel at the document but is not yet aligned at the sentence level, so if you only have versions of the same content in multiple languages in separate documents (e.g. training manuals, web pages, leaflets, etc.) the custom feature will be able to automatically match sentences across documents. Also, if you have monolingual data in either or both languages, this data can be used to complement the parallel data to improve the translations.

Learn how the Translator custom feature works below and read the in-depth article on the Microsoft Research Blog.

 

To begin adding customized translations to your apps, workflows, and websites, sign up for the newest version of the Translator Text API or the preview of the new unified Speech service. Then get started at http://portal.customtranslator.azure.ai

This new custom feature comes with other Translator product releases for Microsoft Translator:

 

Learn more