Two new papers indicate that it’s possible to develop an AI system that does not need to rely on parallel texts, instead, the system is able to figure things out for itself.
When it comes to translating one language into another, computers are quite advanced and will have no problem with this task. During the typical machine-learning process the AI is supervised, when it attempts to translate one language into another, a human will tell it whether or not that’s correct. However, these two new papers have focused their attention on mapping connections for each language. They highlight that regardless of the dialect, words are frequently used together, such as, ‘table’ and ‘chair’, ‘shoe’ and ‘sock’ are the like. If the AI is able to figure out the connections between each language, it can learn independently of a human.
The Polytechnic University of Valencia, developed a system which uses back translation – this is where the AI sees a sentence written in new language that is roughly translated into the other, then back to the first language again. A similar process known as denoising was developed by Facebook computer scientist Guillaume Lample. This system adds or removes words being added to the sentence for different translations. When the two systems are combined, these methods help machines comprehend how language is constructed.
The two systems have not yet been peer reviewed, but they have shown signs of promise in the early stages. The researchers hope to draw on each other’s work and improve the system, so brace yourself, because it looks like AI is about to get a whole lot smarter.
Image Credit: Copyright: maxxyustas / 123RF Stock Photo