��ɫ��Ƶ

The journey began when one of NetApp’s regions requested more local-language content, but didn’t have the budget for translation. It was this request that spurred NetApp to begin exploring MT as a way to offer greater support to its global business. 

This was before the advent of neural machine translation (NMT). NetApp evaluated several options and chose ��ɫ��Ƶ, whose statistical MT solution met all their requirements, including security. NetApp already knew ��ɫ��Ƶ as a trusted technology partner, because they were already using ��ɫ��Ƶ content management and translation technologies.

NetApp’s knowledge base articles were the first materials to be translated using MT, and the first target languages were Japanese and Simplified Chinese – two of the most challenging. As a result, it took significant time and effort to customize the solution to deliver the desired quality.

Working with ��ɫ��Ƶ to continually train and improve its custom MT engines, and with authors to create more MT-ready content, NetApp steadily began to realize the potential of this new way of translating. So much so, that the business began looking at its content strategy from an MT-first perspective, considering what content types could be translated using MT and what languages should be the next targets for custom training.

By the time NMT proliferated on the scene, NetApp was using MT to translate both knowledge base articles and product documentation with no post-editing at all, and had added English-German, English-Spanish and English-French to its two original custom language pairs. This experience put NetApp ahead of the curve, ready to start seeing the benefits of NMT as soon as they migrated to it.

Continuing to partner with ��ɫ��Ƶ, NetApp migrated to its cloud-based neural machine translation solution, Language Weaver – and quickly saw the difference.

“Because we want the translation to sound natural and human, custom training remains just as important to us, even with neural machine translation,” says Edith. “The difference is that it now takes a lot less resource to achieve good quality, so we’ve really been able to realize our MT-first vision.”

This vision has been realized to the extent that 80-90% of NetApp’s localized content is translated by MT, almost all of it with no post-editing. The team has such confidence in the quality that for many content types the only review is occasional spot-checks, with feedback loops in place in case issues are discovered by those using the content. There are very few content types that don’t go through MT at all. The company has also added five more target MT languages, bringing the total to ten custom language pairs built and maintained by the Language Weaver team.

While MT has enabled NetApp to translate content they wouldn’t have had the budget to translate before, the volume of content still being localized through human translators hasn’t changed.

“We’re always testing to see how far we can push machine translation,” says Edith, “but from my experience you still need human translation as a quality touchstone, to show what good quality looks like and to deliver the data to continue to train and improve the MT output. Nothing we do stands still, and we need humans to tell the machines how we’re evolving our products, language and style.”

A big part of NetApp’s MT-first strategy is to integrate MT directly into its content systems, creating an automated localization workflow for content creators that doesn’t require them to use any other system. Integrations already exist with the company’s knowledge base system, with GitHub for software and documentation localization, and with a vendor’s video localization workflow. Edith’s team continually looks for new integrations to improve globalization for the business and ultimately deliver a better service and experience to NetApp’s customers.

To maximize the benefits of secure machine translation for the business, anybody in NetApp who wants access to Language Weaver for internal use can ask for it. They can then translate what they need to without using one of the many free, publicly available MT services that could potentially expose sensitive information. Beyond the custom language pairs used by NetApp to localize from English into other languages, Language Weaver supports 3,000+ language combinations based on 150+ out-of-the-box language pairs.

Edith is often asked for advice about implementing a globalization operation that includes MT. 

“I encourage any business to go ahead because I really believe in the value,” she says. “I tell them to find and evaluate vendors and identify a content type to start with. Technical or product content is usually a good candidate. Focus on your use case and determine your quality expectations and the workflows you need. Start small and build it out to more languages. Take it step by step; it’s definitely worth it.”