Will machine translation ever be on par with humans?
Machine translation has improved considerably in recent years thanks to a new technology known as “NMT". So, what does the future hold for this rapidly advancing field and what does it all mean for the human translators trying to keep up?
AI-driven online translation tools have improved almost out of sight over the last few years. Websites that were once a reliable source of clunky texts, mixed-meanings and cross-cultural gaffes now produce readable, smooth sentences.
Google Translate now understands almost any bizarre slang you can throw at it. Meanwhile, new players like German company DeepL show incredible accuracy time after time.
The change is thanks to the recent rise of neural machine translation (NMT), a comparatively new method of machine translation which produces more natural-sounding translations.
NMT-driven online translation sites require masses of data
| © Pexels
“You have to think of it like a child's brain,” he told German website Gruenderszene.de in March of this year. “This child sees many things during its development, experiences the world and learns from the experiences.”
“Our neural networks function in much the same way. They see a lot of translations and learn to translate a sentence in a certain way if it is constructed in a certain way.”
“A mathematical sequence is run which trains the model,” he adds. “This makes the procedure so generic: it can be applied to different languages.”
NMT-run online translation sites require masses of data, mainly in the form of paired sentences that the system can extrapolate from. DeepL, although only a few years old, is the successor to popular online language dictionary Linguee, and uses that data to produce excellent translations of large texts.
“We used to have statistical machine translation - which is data in, data out,” she explains. “It took the text and just gave you the equivalent of the words.”
“The benefit was: it left nothing out. It might have sounded awful but it didn’t leave anything out.”
NMT in comparison has a completely different science behind it, she says, and may leave out words that don’t need to be translated individually, leading to a much more natural-sounding translation.
“This is why people go, ‘Oh, this has got a lot better, it doesn’t sound like rubbish anymore’,” she says.
Asking for directions in the local language is about to get a whole lot easier
| © Colourbox
“It’s very good at learning to create good-sounding English but sometimes it might miss some nuance at the input, in order to make it good-sounding English,” he explains.
“There’s nothing really in the way that you create the model that’s telling it to do that, or not to do that, it’s just sort of an emergent property and that’s a real problem.”
Aside from accuracy, there are security issues with online machine translation websites too. Information entered into the websites isn’t always secure and is often being saved, to help train the system to learn.
A memorable case in Norway in 2017 saw confidential documents from oil company Statoil appear on the internet, after an employee entered them into a free online translation service.
Norway's oil company Statoil had a data leak after using an online translation service
| © Colourbox / Carsten Gulbrandsen
He also thinks that machine translation technology will start to form part of virtual assistants, such as Apple’s Siri and Amazon’s Alexa. Cohn believes developers will boost the devices to facilitate multi-lingual communication with someone else or retrieve documents from the internet written in a foreign language.
“I think translation is a key part of that puzzle of having decent dialogue systems,” he says.
So, do the impressive developments in machine translation technology spell doom for the human translator? Berlin-based translator Martin Haynes doesn’t think so. He expects machine translation to be used more in multi-lingual website creation and digital media, but doesn’t think it poses a risk to his livelihood.
“I think you should never shun new tech,” he says. “Instead, try and work with it to find out how it can improve your client experience.”
“Translators may have to adapt their skill sets to include localisation, ‘transcreation’ and content creation or copywriting as a result.”
Tea Dietterich also thinks that humans have something to add to translation work, that machines can never provide: “The ability of taking two different concepts and creating something new: that’s something only we humans still can do.”
Learn more about Trevor Cohn's views on the future of creative AI here.
Google Translate now understands almost any bizarre slang you can throw at it. Meanwhile, new players like German company DeepL show incredible accuracy time after time.
The change is thanks to the recent rise of neural machine translation (NMT), a comparatively new method of machine translation which produces more natural-sounding translations.
NMT-driven online translation sites require masses of data
| © Pexels
NMT explained
The new technology uses an artificial neural network to predict the likelihood of a sequence of words, often in the form of whole sentences. If that sounds a little abstract, let DeepL CEO Jaroslaw Kutylowski explain it a bit more simply.“You have to think of it like a child's brain,” he told German website Gruenderszene.de in March of this year. “This child sees many things during its development, experiences the world and learns from the experiences.”
“Our neural networks function in much the same way. They see a lot of translations and learn to translate a sentence in a certain way if it is constructed in a certain way.”
“A mathematical sequence is run which trains the model,” he adds. “This makes the procedure so generic: it can be applied to different languages.”
NMT-run online translation sites require masses of data, mainly in the form of paired sentences that the system can extrapolate from. DeepL, although only a few years old, is the successor to popular online language dictionary Linguee, and uses that data to produce excellent translations of large texts.
“Doesn’t sound like rubbish anymore”
Tea Dietterich, whose company 2M Language Services offers tailored NMT services to some of its bigger clients, says that the new translation technology has clearly led to smoother translations in certain language combinations.“We used to have statistical machine translation - which is data in, data out,” she explains. “It took the text and just gave you the equivalent of the words.”
“The benefit was: it left nothing out. It might have sounded awful but it didn’t leave anything out.”
NMT in comparison has a completely different science behind it, she says, and may leave out words that don’t need to be translated individually, leading to a much more natural-sounding translation.
“This is why people go, ‘Oh, this has got a lot better, it doesn’t sound like rubbish anymore’,” she says.
Asking for directions in the local language is about to get a whole lot easier
| © Colourbox
Technology with limitations
But, NMT still has its shortcomings. Professor Trevor Cohn, from the University of Melbourne’s School of Computing and Information Systems, says that while the texts can sound fluent, sometimes they are just plain wrong.“It’s very good at learning to create good-sounding English but sometimes it might miss some nuance at the input, in order to make it good-sounding English,” he explains.
“There’s nothing really in the way that you create the model that’s telling it to do that, or not to do that, it’s just sort of an emergent property and that’s a real problem.”
Aside from accuracy, there are security issues with online machine translation websites too. Information entered into the websites isn’t always secure and is often being saved, to help train the system to learn.
A memorable case in Norway in 2017 saw confidential documents from oil company Statoil appear on the internet, after an employee entered them into a free online translation service.
Norway's oil company Statoil had a data leak after using an online translation service
| © Colourbox / Carsten Gulbrandsen
A bright future
Cohn says that one way machine translation could be used more in the future is to help people learn a second language. Certain language apps are already employing technologies from machine learning and language processing and trying to build them into their products.He also thinks that machine translation technology will start to form part of virtual assistants, such as Apple’s Siri and Amazon’s Alexa. Cohn believes developers will boost the devices to facilitate multi-lingual communication with someone else or retrieve documents from the internet written in a foreign language.
“I think translation is a key part of that puzzle of having decent dialogue systems,” he says.
So, do the impressive developments in machine translation technology spell doom for the human translator? Berlin-based translator Martin Haynes doesn’t think so. He expects machine translation to be used more in multi-lingual website creation and digital media, but doesn’t think it poses a risk to his livelihood.
“I think you should never shun new tech,” he says. “Instead, try and work with it to find out how it can improve your client experience.”
“Translators may have to adapt their skill sets to include localisation, ‘transcreation’ and content creation or copywriting as a result.”
Tea Dietterich also thinks that humans have something to add to translation work, that machines can never provide: “The ability of taking two different concepts and creating something new: that’s something only we humans still can do.”
Learn more about Trevor Cohn's views on the future of creative AI here.