While most of us by now have had some experience of using machine translation, namely through Google’s easy-to-use Google Translate feature, this programme's widely criticised lack of accuracy in the past has rendered it unfit for professional usage.
Upon its launch in 2006, Google Translate used statistical translation – a flawed technology that translated one word at a time without considering the context of each word within the sentence. This often resulted in text that read awkwardly to native speakers. But now a new system has emerged, called neural machine translation, and language translation heavyweights such as Systran and Google are confident that it’s a game changer.
“A lot of things that were not possible a few years ago are now possible,” says Mike Schuster, a scientist who previously worked on the Google Brain team to help develop the system. “Think of translation of ad campaigns, translation of social media network content, translation using a camera just looking at text, translation of newspaper articles and other written content, and of course just for daily communication using a smart phone.”
Speed and accuracy are touted as the biggest benefits of neural machine translation systems.
In short, unlike its predecessor, neural machine translation (NMT) translates less like a machine. It’s a deep learning technology that translates whole sentences at a time and within context, rather than just word-by-word, which ultimately results in much more natural, accurate and fluent translations.
Google launched its neural machine translation system (GNMT) in late 2016 via its Google Cloud Translation API. It offers more accurate translation of a growing list of languages to and from English. “GNMT has achieved more translation quality improvements in a single leap than in the last 10 years combined, with 50-80% quality improvement across many languages,” says Barak Turovsky, head of product for Google Translate and Machine Intelligence at Google Inc.
Google is not the only tech giant to release an NMT implementation. Amazon Translate was made generally available in April this year and in Asia, companies including Alibaba, Baidu, Tencent, NetEase-Youdao, Sogou and iFlytek have also deployed neural machine translation. All are seeking a competitive advantage as machine translation continues to evolve.
Speed and accuracy are touted as the biggest benefits of neural machine translation systems. Pioneers behind the tech are confident that this increased reliability and pace will smash language barriers and be the key to a new “global economy”. “Neural Machine Translation is going to change the economy by giving more businesses a language capability they can use to communicate and understand in real time,” said Denis Gachot, CEO of leading machine translation company Systran, which is headquartered in Seoul.
With the capacity to immediately translate documents and websites into multiple languages with reliable accuracy and precision, some brands are already convinced of the benefits.
Given how far systems like Google’s Neural Translation Machine and Amazon Translate have already come, some experts are confident that within a few years they could be more reliable than human translators
Matthew Fryer, VP and chief data science officer at leading online accommodation booking website Hotels.com, believes that neural machine transition systems like Amazon Translate offer a quick, efficient and, most importantly, accurate solution for accelerating their translation workflows. “At Hotels.com, we operate 90 localised websites in 41 languages. We have more than 25M customer reviews and more are coming in every day, making us a great candidate for machine translation.
“We want to take advantage of the latest advances in the transition to neural systems to further personalise and localise our reviews, and generally improve our customer experience.”
In addition to being accessed by millions of people every day on translate websites, neural machine translation is built into social media sites like Facebook and Instagram to enable analysis of multilingual social content, and is even being used to speed up professional translation of things like patents and software. “Companies use GNMT to localise their marketing and sales materials, translate user reviews (e.g. Airbnb, Tripadvisor etc.), translate in-game characters, customer service calls and many other scenarios,” says Google’s Turovsky.
Overall, NMT's ability to speed up entry into new markets and enhance globalisation seems undisputable — and the numbers speak for themselves. Currently, Google’s GNMT reports 140 billion words translated every single day, over 500 million monthly active users, with 92% of all users coming from outside of the U.S.
Keen to further break down language barriers and make the internet more inclusive, in April 2017 Google introduced neural machine translation between English and nine widely-used Indian languages. With around 400 million internet users in India, and with only 20% of the population fluent in English, previously most internet users in the country faced significant language barriers to getting full value out of the Internet.
“One of the biggest challenges of Indian languages is that due to historical and other reasons, most of Indian internet content is in English, and there is very little translated content between English and Indic languages on the internet,” says Turovsky. “Translation quality of Indian languages experienced remarkable improvement with the introduction of GNMT. The results are truly astonishing – translation quality for many Indic languages shot up from barely usable to pretty good. As a result, Google Translate usage in India is exploding across many surfaces. Usage of Google Translate in India increased 10 times in last 4 years, and usage of translations on Google Search shot up almost 60 times.”
Keen to further expand the reach of its neural machine translation system, Turovsky says Google’s goal is to eventually roll out neural machine translation to all the 103 languages and surfaces where you can access Google Translate (which still uses the older phrase-based machine translation model.)
As with the example of Google in India, since 2016, neural machine translation systems have greatly improved machine translation, moving usability from unusable to usable for many language pairs (for example, Korean to English, and there are many more). Given how far systems like Google’s Neural Translation Machine and Amazon Translate have already come, some experts are confident that within a few years they could be more reliable than human translators. “On average most machine translation systems are better than most bilingual humans given arbitrary text to translate,” says Schuster. “Within a few years almost everybody will trust machine translation more than any human translator.”
However, not everyone is as convinced as Schuster. Maarit Koponen, an expert in language technology at the University of Turku in Finland, says that although there is evidence for improvements in quality, neural MT is still far from perfect. “The claims we have seen of neural MT reaching human translation quality levels appear somewhat overstated,” says Koponen. She warns that neural MT systems can still be inaccurate and unpredictable. “One tendency that has been noted in more than one quality evaluation of neural MT systems is that they tend to omit words. As a simple example of what that may mean, if a word like "not" gets omitted, the sentence becomes the opposite of what was intended.”
Koponen believes that while these systems do offer potential for wider reach and communication, there needs to be some caution with regard to quality and an understanding of situations and text type for which machine translation is suited, and for which it is not.
Using machine translation on social media like Facebook, or for some technical help and support forums like Microsoft, or for customer service chats, where machine translation could enable communication between a customer and a service representative in different languages, can all be useful says Koponen. But for other cases where the guaranteed accuracy is mission critical, there will always need to be a human in the loop.
“I recall a presentation by a translation agency representative who noted that they never recommend machine translation to their clients in cases where the content being translated is critical to the brand image or marketing material,” says Koponen. “Even the best machine translation systems available are able to only deal with the words, but targeting material to a new market in a new language also needs to consider cultural issues and how the message should be constructed for this culture. Failing to account for cultural differences may lead the messages to be anything from potentially offensive material to simply ineffective.”
Overall, Kopenen believes that it would be ill-advised for any brand or marketer attempting to reach new markets by using any machine translation system (whether neural or other technology) to translate content that then gets published without any checking by a human. “In my opinion, neural MT offers potential – and since it is still a relatively new approach in machine translation, it will be interesting to see how the technology develops – but it should be used with some caution and understanding that it is neither 100% reliable nor suited for all situations.”