Understanding an erroneous Google translation

Question

I do not know if this question is outside the realm of the FSE. If this is so I will erase it. I am trying to understand why Google cannot give a proper translation of the following sentence:

How much work does it do, how much heat is discarded and what is the efficiency of the engine?

Combien de temps cela fait-il (!), combien de chaleur est rejetée et quelle est l'efficacité du moteur ?

For comparison DeepL works very well here :

Quelle est la quantité de travail, quelle est la quantité de chaleur rejetée et quel est le rendement du moteur ?

I tried several (free) online translating services of various sorts and I have not seen the like of Google's strange behavior (translator). Any ideas ?

Answer 1

No one outside google really knows how the translation system works. It's obviously based on machine learning but we do not know the parameters or the architecture they use exactly.

They updated their Google's Neural Machine Translation System GNMT in 2016, to reduce the gap between Human and machine translation by including the context.

[...] Google has been able to continually improve quality of translations by enabling their systems to take into consideration not only source words and phrases, but also broader contexts of where they appear in sentences, and what are the other words and phrases around them. (How does Google translate work?)

Like all machine learning systems, it is not perfect. That is why you can find a certified translation for more easier -> plus facile even if it is not correct in English. They may have used some corpus, speeches to determine that. Or someone added manually the translation via community tab.

And because DeepL uses different architecture, different sources, they have not got the same result.

Machine learning is really complicated, it is pure mathematical and it depends on word vectorization, number of queries,...