The power of neural network machine learning is helping Google offer faster and more accurate translations when you’re lost for words.
The key to solving complex computing challenges isn’t always to simply throw more processing power at the problem but instead to actually change the way computers think about the problem. That’s where neural networks come into play, attempting to mimic the steps which the human brain uses to solve problems, rather than simply relying on statistical analysis and data crunching.
In late 2015, Google’s TensorFlow neural network machine learning team decided to apply the technology to the Google Translate service, which had already been operating for 10 years using phrase-based machine learning based on statistical analysis.
Only 13 months later Google revealed that it was using neural networks to overhaul its Word Lens feature, allowing it to support both Simplified Chinese and the more complex Traditional Chinese when you point your smartphone’s camera at a sign to translate the text in real time.
Lend me your ears
Today 26 Google Translate languages are underpinned by neural network machine learning, offering a significant performance boost compared to the old phrase-based system. The improvement in Chinese/English translations over the last year was a greater leap than the progress made by automated translation tools in the last 10 years, says Google machine learning research scientist Mike Schuster.
Translations of most Asian languages have seen a similar improvement thanks to neural networks. Rather than being taught by humans, Google Translate teaches itself by scouring the web for translated texts to analyse, a bit like the Rosetta stone which helped decipher hieroglyphics because it contained the same text translated into Demotic script and Ancient Greek.
As a result the accuracy of Google Translate’s efforts today is less related to the complexity of the languages involved and more to do with the availability of online translations to study and their accuracy. Just like a person, Google Translate can pick up bad habits if it’s trained on poor translations.
Google rates the accuracy of its translations based on user feedback, with a score out of six. The results vary from language to language but in the graph above you can see the improvement based on old phrase-based machine learning and how close neural networks are coming to matching human translations. According to Google Translate user ratings, these days neural network French to English translations are considered almost as good as your average human translation and while Chinese lags behind you can see the significant improvement thanks to the introduction of neural networks.
At the same time Google has also improved the speed of translations thanks to enhanced algorithms and new hardware. A 10-word sentence previously took 10 seconds to translate but now it’s down to .2 seconds even though Google Translate handles more than 1 billion translations each day with 500 million active users across 103 languages.
The people have spoken
Google eventually plans to shift all 103 languages to neural network translations, but a separate neural network model is required to translate each language to each other language – which would require more than 10,000 models to handle 103 languages. The next big step for Google is to build multilingual models capable of translating to and from several languages.
Google’s first multilingual model is based on English, Korean and Japanese, designed to translate either Asian language to and from English but not between the two Asian languages. Despite this, it also allows for “zero shot” translations between Japanese and Korean — which means translating between two languages even though the model isn’t specifically trained to support that combination of languages. It’s one of the first zero shot translational efforts to offer intelligible results, Schuster says, although the results aren’t yet as good as using a neural network model trained specifically to translate from Japanese to Korean.
While Schuster won’t be drawn on a time frame, he says it’s only a matter of time before neural network machine language translations are better than the average human translation.
Google is also looking at the way that sentences with the same meaning can be mapped to similar regions of a language’s structural map, in another step towards the holy grail of building a universal translator. It will remain the stuff of science fiction for the foreseeable future, but at this rate neural network machine learning is looking like the best shot for translating every language in the world into your native tongue.
Adam Turner travelled to the Google I/O conference in California as a guest of Google.