From best to best-er
Google Translate has grown in the past 10 years to be one of the leading online translation services, with over 100 languages and 140 billion words translated every day. Now, it’s about to get even better.
Beginning in September 2016, Google Translate announced a switch to a new system called Google Neural Machine Translation (GNMT). Previously, Google (and most other online translators) relied on phrase-based translation, essentially the digital equivalent of combing through pages of a language dictionary to search for a word. While phrase-based translation is quick and effective, it lacks a grasp of some linguistic structures and can lead to awkward translations.
*You’re
GNMT, is a smarter, machine-learning translation method with an ability to learn from users. Think of it as a digital version of being corrected for misuse of a word or grammar structure. Just like we learn when people point out our grammar mistakes, GNMT can adapt to make educated guesses about how to translate phrases based on the content, tone and context.
The biggest breakthrough of GNMT was a side-affect though. After the short initial implementation of the new system, Google Translate internally made the task more efficient by inventing a whole new language that could be used as an intermediary in translations. The new language was not programmed by Google’s engineers. Instead, Google Translate developed it automatically to improve its own ability to solve complex problems.
More accurate, just as fast
In a blog post published in November, the team behind GNMT explain how things work. The core of the breakthrough is the ability for the system to “transfer the ‘translation knowledge’ from one language pair to the others.” For example, if the system is trying to translate a phrase from English to Japanese, rather than guess, it can pull from other languages that it has more knowledge of.
The new language, or “interlengua” as Google has chosen to call it, is an all-knowing middleman that can become smarter and more developed than any phrase-based model could ever be. It is also just as quick as other translation systems that do not use an interlengua, meaning you’ll get more accurate translation at the same speed.
Starting off simple
To say the system actually “invented” a new language or went in a direction that the Google engineers did not see coming is a bit of an exaggeration. Rather, the system just worked a lot faster and more effectively than expected. And, interlenguas have been used by human translators for years, so while the GNMT is a breakthrough for online translation, the concept is not new.
Currently, GNMT is only being used for a fraction of the languages Google Translate offers, but the speed the technology has progressed suggests that Google may push things along faster to get it introduced to more languages as well.
An exciting future
What may even be the most exciting, beyond the fact that GNMT can give more accurate translations to language learners and foreign travelers, is a possible application to provide better translations of classic texts or antiquities.
Google has not directly touched on this as a possible benefit of the developed GNMT system, but given the major moves forward it has made in the past few months, the future applications could be huge.
