>>4927688Deepl is run by google.
Google translate uses a traditional method of algorithmic machine translation. It is not designed to translate whole sentences and is a simple word-for-word substitution. These kinds of systems are stumbled by words whose meaning changes based on context, languages that don't have good or extensive translations available in its databases because google didn't want to hire anybody, and when translating between two languages with dramatically different grammar structures.
Deepl uses a "deep learning" neutral network trained by feeding it analogous words and phrases. This allows it to differentiate between an individual word, and a phrase where context changes the meaning of a word so it's generally much better at handling sentences--though still far from perfect.
Deepl is still 'experimental' because the actual technology itself is still under r&d. It's still only a minor upgrade over google translate in that it handles thing google translate is explicitly not developed to handle, but the technology behind it should hypothetically be capable of consistently near-perfect translation and it's not there yet. Also, despite the better technology behind deepl, a root issue is just that google translate lacks robusticity because of the way that its language database has been expanded (eg. very slowly and unevenly, relying mostly on community contributions for most non-european languages) whereas no expense was spared training deepl. The fact that deepl is ONLY this much better is actually a bit of a problem.
