In today's world, more than 7000 different languages coexist. However, the Internet brings us together in a virtual space in which we interact. This small miracle would not have been possible without the monumental progress made by the online translation services offered by Google and some of its rivals.
In an exciting study, researchers from Amazon Web Services and the University of California looked at online translations. The authors believe that despite the enormous progress made, many translations still leave something to be desired.
To see more clearly, scientists sifted through nearly 6 billion sentences translated into at least two languages. The idea was to assess quality. Concretely, the more the expression is translated, the more its quality leaves something to be desired.
Unsurprisingly, certain languages, which are not very widespread, fare quite poorly, notably African dialects. Researchers explain these poor results by the lack of data. Indeed, language models need a sufficient quantity to function correctly, and this is not the case for some languages.
Big mistakes that make task
This is a real concern, because it means that new AIs such as ChatGPT and its rivals may have difficulty working correctly for some languages if the translations are not up to par.
In certain cases, bad translations can also cause big problems. Our colleagues from Digital Information World cite an example of medical advice to Armenians where taking ibuprofen to relieve pain becomes taking an anti-tank missile to relieve pain”.
Other examples have punctuated recent news. We remember in particular that in the context of the conflict between Israel and Hamas, Instagram's translation function automatically added the word “terrorist” to biographies containing the word “Palestinian”.
This bug was quickly corrected, and Meta immediately split a message of apology through a spokesperson: “We fixed an issue that briefly caused inaccurate Arabic translations in some of our products”. Mark Zuckerberg's group, however, has not explained the origin of this very annoying bug.
In any case, it illustrates the shortcomings of these translation tools. Although they prove very useful in breaking down the language barrier, they still need to be perfected to avoid such blunders.
📍 To not miss any news from Presse-citron, follow us on Google News and WhatsApp.