According to a recent research paper conducted by Google, their artificial intelligence is now able to translate from and into languages without being trained in them. The AI reportedly uses something called “interlingua”, which can represent different phrases and their meanings without any prior training and regardless of the language. So, the main principle is that as long as the AI translated the unknown phrases into a familiar language internally, it can then translate it into any existing language. And with acceptable accuracy. Or so Google claims.
A very smart AI
The team over at Google kept wondering how can they make the system better. The idea was to make it translate from Chinese into Mandarin for example, without first making the Chinese-English, English-Mandarin translation. They called this a “zero-shot”. Blind translation. Surprisingly enough, according to Google:
“It can generate reasonable Korean to Japanese translations, even though it has never been taught to do so.”
How does this work? Nobody is sure, but the Google team is positive that the AI developed its own learning system. It may use a three-language model which grasps the same meanings from phrases written in different languages. It then groups them, thus learning new languages. Remember when Stephen Hawking warned us about the self-development of artificial intelligence? This is a real example that he was right. This is the system’s “interlingua”.
Google made some changes
It seems like recently, Google’s translation feature became something known as Google Neural Machine Translation. The company is saying that this system is learning multiple languages from the power of example. Millions of them, to be more precise. This helped with the improvement of the quality of translation. However, there is a big problem here. Considering that there are 103 languages which Google’s system works with, this means that there are over 5 000 possible pairs of languages which must be translated. Imagine just how many examples it would need to learn all of them. It would be almost impossible.
The Google team also conducted an experiment. They put twelve pairs of languages into a model the size of one single pair. The result was astounding. Even with such a reduced code base, the quality of the translation was almost the same with that of a normal pair of languages. This huge and very quick progress of Google’s translation AI can be scary though. It shows how can artificial intelligence develop itself to be more and more prepared for different situations. It also finds new and simpler ways to deal with more complex matters. And in such a short time.
All in all, it looks like the future of professional translators is bleak. If artificial intelligence will start replacing human jobs sometimes in the future, translators might be the first ones to lose their jobs. And it is not an encouraging prospect. Years of hard work and dedication will be completely lost to a computer which can learn in a few days what a human learned in years. Still, this is the future and we must accept it, whether we like it or not.
Image source: here