In recent years, computers have become much more adept at translating from one language to another thanks to the use of neural networks. However, AI training usually requires a lot of human-translated content for computers.
Mikel Arteks, computer scientist at the University of the Basque Country (UPV) and author of one of these works, compares the situation to giving someone different books in Chinese and different books in Arabic, without any of the same texts overlapping each other. It would be very difficult for a human to learn how to translate from Chinese to Arabic in this scenario, but a computer can.
In a typical machine learning process, an AI system is monitored. This means that when the AI tries to find the correct answer to any given problem, the person will say if it is correct or not, and, as necessary, the AI will make adjustments.
Instead, AI learns how words are related in the same way in different languages in a new way - for example, the words "table" and "chair" are often used together, regardless of dialect. By comparing these combinations for each language and then comparing them, you can get a good idea of which terms are related to each other.
These systems can be used to translate complete sentences, not just single words, using two complementary learning strategies. Reverse translation assumes that a sentence written in one language is roughly translated into another, and then back to the original language, in case of a mismatch, the AI sets up its protocols differently. Noise reduction is a process similar to the same process in radio engineering, but with words that are removed or added to the sentence. Synchronizing these methods helps the machine better understand how the language works.
Test vocabulary
Both systems - one developed by UPV and the other by Facebook computer scientist Guillaume Lampler - have yet to receive expert evaluation, but both have shown promising results in preliminary testing.
Promotional video:
To assess their ability to translate text from English into French and vice versa, approximately 30 million sentences were proposed for translation. Both AIs managed to get 15 points. The AI Google Translate, which uses supervised machine learning, has a score of 40, while human translators can score up to 50.
Both researchers agree that each can improve their self-learning AI system by building on the work of the other. AI could be made more capable by introducing several thousand parallel sentences into their curriculum, which would reduce the time it takes to master the language.
Serg kite