Researchers at Amazon Web Services (AWS) AI lab have discovered that a significant portion of online content is translated by machines. This phenomenon is observed across various languages, but the quality of the translations is often subpar.
The findings underscore the critical importance of using high-quality data and carefully considering its sources when training large language models. The researchers noted that machine-translated content is prevalent, particularly for languages lacking extensive resources, comprising a substantial portion of the online content users encounter.
Mehak Dhaliwal, a former intern at AWS who is now a PhD student at the University of California, Santa Barbara, said they got interested in this because colleagues noticed that a lot of stuff in their own language seemed to be translated by machines.
Also Read: Secret Alliances Exposed! Tech Giants and AI Startups Caught in Antitrust Crossfire
They made a big collection of sentences in different languages to understand more about what machine-translated content looks like.
They found that a lot of web content gets translated into many languages, mostly by machines.
This is especially true for languages with fewer resources. They also noticed that the kind of content translated into many languages often seems to be chosen to make money from ads.
The researchers say that while machine translation has gotten better, it’s still not as good as human translation. A lot of the stuff translated by machines on the web is probably not very good quality.
This could make language models less accurate, especially since the data chosen might not be very good even without translation errors. Good data is really important when training language models, and usually, people use high-quality stuff like books and Wikipedia articles.
 AI lab found out that a lot of stuff online is actually translated by machines.){kind=link}