Artificial intelligence has been able to translate languages ​​without a dictionary

A huge breakthrough in machine learning and in research, language teaching.

At this rate, we will have a completely different and completely new language, just look at this Facebook example.

Just recently, two studies show how a neural network computer can translate languages ​​without a dictionary, an important step in language research and teaching.

At this rate, we will have a completely different and completely new artificial intelligence language.

"Just imagine you give a bunch of Chinese books and a bunch of Arabic books - no content overlaps - and that person must learn to translate Chinese into Arabic. feasible right? " , author of one of the two studies, a computer scientist at the Basque University in Spain, Mikel Artetxe said. "But we have shown that a machine can do that."

Until now, computer "teaching" would include a computer process of guessing the answer - getting the right answer - adjusting the analysis process to create the right answer next time; This way is "supervised".

These two new studies focus on a completely different approach: running without supervision, letting it learn properly. The two systems create a bilingual dictionary of their own without anyone correcting them. This is possible because language often has common points between word groups as well as word combinations.

The two systems create a bilingual dictionary of their own without anyone correcting them.

For example, "tables" and "chairs" are often paired together. So computers create a map from a big point, words are big cities, connected with words related by national highways. Gradually, it will find the best way to draw an effective path between cities. So, it had a bilingual dictionary.

In these two studies, scientists use similar methods and can translate the language one by one. They all use two reverse- intelligence artificial training strategies - back translation and noise reduction - denoising.

Now computers can translate languages ​​without human help.

In the method of back translation, a sentence in A language will be translated into B language, and then B will be translated back to A. If the translated sentence is not true to the original, the AI ​​will gradually correct .

Denoising is akin to back translation, but instead of translating back and forth between languages ​​A and B, the AI ​​will add other elements into the sentence (rearrange the word position, remove the words) and try reverse the sentence to the original.

Using these two translation methods, the AI ​​will gain a deeper understanding of the structure of the language.

"It's amazing that computers can translate languages ​​without the help of people," Di He, a computer scientist at Microsoft, has inspired research for both studies, to speak.

The two studies, one from Mikel Artetxe and one from computer scientists at Facebook, Guillaume Lample and colleagues, have all been published in Science. Both teams are happy to show that they can improve their AI by applying new elements from the remaining team.

However, these two studies have not been tested and analyzed thoroughly (peer review).