NMT & TMS: How Do They Work Together?

Report written by Özge Ünlü.

It is a given nowadays that when organizations expand internationally or consolidate their existing global presence, they need to invest in localization. This does not simply mean linguistic resources but also technological investments that contribute to translation productivity among other efficiencies. When speaking about translation productivity, two technologies are an absolute must —translation management systems (TMS) and machine translation (MT). While a TMS helps optimize processes around linguistic tasks and project management, MT represents a time gain by contributing to the scalability and speed of translation. Indeed, these two tools are often integrated together and work seamlessly side by side. For this reason, it seems to us to be important to analyze and explain the various common ways they are combined and what this means for the shaping of language technology’s future.

What is NMT and why is it used?

Neural machine translation (NMT) uses machine learning technology to teach systems how to produce translations with a close to human-like quality and fluidity. Thanks to algorithms, systems learn from pools of corpora to translate texts in the best way possible. It is the latest MT technology and the most widely used today.

Overview of the most common NMT solutions

NMT engines can be generic, domain-based, customized, or adaptive.

Generic MT is a non-customized, non-specialized system — free publicly available systems such as Google Translate fall into this category. But MT engines can also be trained on a domain or customized for a specific organization. In this case, we refer to them, respectively, as domain-based and customized MT.

Since the NMT market is continually expanding, there’s an ever-growing choice of generic MT solutions on the market such as Microsoft Translator, DeepL, and the like. Generic engines don’t have a specific domain or specialization. Therefore, they often lack context, specific terminology, and/or style features. The vocabulary might sometimes be imprecise and/or inconsistent, for example. This kind of MT usually requires thorough human post-editing. For this reason, a well-trained customized engine might be a better fit than a generic MT system, even if the latter is provided by a big name like Google or Microsoft.

Source: Nimdzi Language Technology Atlas 2022

This report has been researched and written by Nimdzi's Localization Researcher, Özge Ünlü. If you want to learn more about this topic, reach out to Özge at [email protected].