Paris - Artificial intelligence is now capable of translating a huge number of different languages, making the old dream of a world where the language barrier is abolished almost attainable.
Meta, the parent company of Facebook, announced Wednesday that its artificial intelligence in this area was now able to translate 200 languages between them, regardless of the combination, compared to 100 until now.
"Many" of the languages covered by this expansion "were not currently accessible to machine translation," Meta founder Mark Zuckerberg boasted in a blog post.
This announcement is an example of the race for machine translation among the big Internet groups, anxious to make their services and products available to anyone on the planet.
In May, Google researchers published a scientific article entitled Building Machine Translation Systems for the Next 1,000 Languages, whose title sums up the specialists' ambition.
The artificial intelligence developed by Google, Microsoft, or Meta is now capable of translating languages for which there is very little parallel data, i.e. languages that have been very little translated into any language.
They allow translation between two very localized languages on the planet, such as Quechua (spoken in Peru) and Peul (a West African language), even though no human has ever attempted this task.
The challenge of machine translation "is particularly important for Facebook, which must succeed in filtering hate messages" that appear all over the world, in all languages, stresses François Yvon, a researcher at the CNRS and specialist in language processing.
The automatic translation can allow moderators in the English language to intervene in content published in other languages, he explains.
It remains to be seen how reliable these tools are. Meta says its new system is capable of "44 percent better" performance than its previous 100-language model.
"For some African and indigenous languages, this difference exceeds 70% compared to recent translation systems," the company also claims.
But for François Yvon, the automatic translations provided by Google or Facebook engines will inevitably remain unequal in quality depending on the language.
One day, speaking in 200 languages
"Languages that are highly translated, such as European languages, will probably always have an advantage," he says.
Vincent Godard, CEO of Systran, a French company that pioneered machine translation and works on 56 languages, has a similar diagnosis.
The technology that this group uses is initially the same as that of Meta and Google, but it has been enriched by the work of real linguists to avoid errors, he says.
"When you're working on the translation of an assembly manual for a fighter jet, you can't afford a single mistake," whereas mistakes can be acceptable when translating a restaurant review, he says.
So, are we close to having automatic speech translation, to be able to speak live with anyone on the planet, for example in the future metavers?
"We're not there yet, but we're working on it," says Antoine Bordes, the managing director of Fair, Meta's artificial intelligence research laboratory.
"We have another project on automatic speech translation, which for the moment works with far fewer languages," he says.
"But the interest will be to connect the two projects so that one day we will be able to speak in 200 languages while keeping the intonations, the emotion, the accents...", he anticipates
Translated with www.DeepL.com/Translator (free version)
No comments:
Post a Comment