How Google translations are getting more natural
Mumbai: Researchers are increasingly striving to help machines translate words from one language to another the way professional translators would. This implies that machines must understand the context of words and sentences, and make sense of idioms, phrases and jokes.
However, despite the fact that billions of words are being translated daily by multilingual machine translation services like Google Translate, Microsoft Translator, Systran’s Pure Neural Machine Translator, WordLingo, SDL FreeTranslation, China’s Baidu, Russia’s Yandex or Babel Fish, machines have a long way to go before they can function as fluently as humans do when speaking in, and translating, different tongues.
Barak Turovsky, product lead at Google Translate—a free multilingual machine translation service from Google Inc.—understands this dilemma well. “Today, translation by machines can be likened to my five-year-old son speaking Russian. Since I speak fluent Russian, I know the mistakes he makes and how he forms words,” he says.
ALSO READ : Google Translate’s India focus
Turovsky asserts, though, that neural machine translation, or NMT, is accelerating the pace of machine translation. He describes Google Translate as the “largest artificial intelligence (AI) project in the world that uses cutting-edge machine learning models and custom Google hardware—Tensor Processing Units (TPUs) and cloud TPUs—to serve one billion Translate users by doing more than 140 billion translations a day”.
From words to sentences
Machine translation systems are applications or online services that typically use machine learning—an artificial intelligence (AI) technology—to translate text from one language to another. Till even a couple of years back, major translation service providers including Google Translate and Microsoft Translation were using the statistical machine translation (SMT) system.
For instance, when Google announced the launch of its Translate service 11 years ago, they were using the phrase-based machine translation (PBMT) as the key algorithm behind this service. The PBMT algorithm, considered to be the simplest and most popular version of SMT, breaks an input sentence into words and phrases to be translated largely independently.
Turovsky likens SMT to “old school machine learning”. The approach, he explains, was to take advantage of the same technologies that Google built for search “…and crawl the web, find content that was already translated by humans, put it in a giant index--and if you have billions and billions of combinations, the machine would start seeing statistical patterns. And that is how machines learn the language.”
In other words, SMT uses advanced statistical analysis to make sense of words. However, SMT had its limitations.
“We had to break the sentence into chunks—not more than five words. It became unmanageable,” Turovsky recalls.
NMT changed the picture dramatically. The advent of the deep neural network (DNN)-based translation, coupled with the increase in computing power due to more powerful chips, resulted in better quality translation.
NMT uses DNNs that have nodes which are loosely modelled on the human brain and relate to each other.
These nodes build relationships with each other based on bilingual texts with which you train the system.
“The biggest breakthrough of NMT is that it works analogous to the brain. That allows us to translate a whole sentence at a time without breaking it. Since we can take the sentence in its context, the accuracy increases. Besides, since we translate the whole sentence, its fluency increases, which means that the sentence sounds more natural,” Turvosky explains.
Last November, Google announced that it would begin using the Google Neural Machine Translation system (GNMT). Unlike PBMT, NMT considers the entire input sentence as a unit for translation. Turovsky hopes to extend NMT to all the 103 languages (from the current 41) that Google Translates, by the year end.
But NMT, which other machine translation services such as Systran and Microsoft Translation have also begun using, needs a lot of computing power.
“Typically, it takes a human about 15 seconds to read a sentence. But it takes a lot of mental power and another minute to understand the context. Our brain also uses very less energy but can do complex calculations. Not so with computers. They require a lot of energy,” Turovsky notes. To address the issue, GNMT uses Google’s machine learning toolkit TensorFlow and its own Tensor Processing Units (TPUs) that were specifically developed for machine learning.
“We are probably the biggest internal clients of TPUs for Google,” Turovsky says, adding, “Today, all neural models stop at one sentence. We will soon take an entire paragraph or entire documents, which will increase the efficiency of translation.”
The Google Translate app, for one, already lets you snap a photo of text and get a translation for it in English. With Word Lens, that Google introduced in 2015, you can use the Translate app, point your camera at the text, and the English translations will appear (use of augmented reality) as an overlay on the screen—even if you are not connected to the internet.
Second, computers still can easily get fooled by idioms and jokes “because they are culturally sensitive”, Turovsky acknowledges. Google uses “crowd sourcing” to solve this problem in part. He cited the example of the people of Kazakhstan who “were very interested in enabling Kazakh on Google Translate but did not have much training data”. The people of Kazakhstan were insistent, he recalls, that they would produce the necessary data.
“We were sceptical. We created a micro task—sent it to 300-400 people asking them to produce two million validated translations. To our surprise, we started getting 200,000 contributions a day. We wondered how. Apparently, there was a press conference in the office of the President who called everyone who knew English in Kazakhstan, and cajoled them to contribute. That’s how we got Kazakh,” Turovsky says, adding that about 14 languages were enabled on Google Translate in a similar manner, with the help of Google’s so-called ‘Translate Community’.
“This is the best way to translate idioms, jokes and songs, which machines are not very good at,” Turovsky insists, adding that his team is also working “very closely with other Google teams to enable speech translation”.
GNMT has, so far, succeeded in reducing translation errors by an average of 60% compared to its own phrase-based production system. The goal is to completely bridge the gap between human and machine translation. When will this happen?
“There will always be limitations to machine translation,” Turovsky says.
“We are currently not focusing on complex negotiations like legal transactions, for which you will require professional translators. Casual communication is typically used in 90% cases. So the focus now is on translation for basic communication needs like travel, commerce, and the like.”
•India has 234 million Indian language users online, compared to 175 million English users
•The Indian language user base will continue to grow at 18% annually to reach 534 million in the next four years
•Nine out of 10 new internet users coming online today are Indian language users
•Tamil, Hindi, Kannada, Bengali and Marathi-speaking users have the highest adoption of online services, followed by those who speak Telugu, Gujarati and Malayalam
•In the next four years, Hindi-speaking users alone will overtake English-speaking users and will make Hindi the most used language on the internet in India
•Marathi, Bengali, Tamil and Telugu-speaking internet users will form 30% of the total Indian language internet user base
•99% of Indian language users access internet through mobile devices. The overall share of internet users in India accessing internet through mobile devices is 78%
•68% internet users consider local language digital content to be more reliable than that in English
•35% of Indian language internet users access government services, classifieds, news and payment services exclusively online
•Language-enabled preloaded applications and web browsers see higher adoption among first time/new Indian language internet users
•88% of Indian language internet users are more likely to respond to a digital ad in their local language vis-a-vis an ad in English
Source: Indian Languages- Defining India’s Internet, a report by Google and KPMG India, April 2017