I have been using this website for a week already and the amount of incorrect translations and sentences its making me worry. If I did not have any knowledge of German I would be learning some bad translations.
Even correcting the sentences seems like I am making the work and taking the effort that developers or the team should be taking.
Will they pay me for the amount of corrections that I am doing?
Like seriously man.
Out of interest, in which collections do you find these incorrect translations? More specifically, is it in the “new” Fast Track or the “legacy” collections (sourced from Tatoeba)?
And could you give some examples?
I’ve corrected the translation for at least 10 phrases but didn’t keep a full record, as I found it too tedious to save each one just to report them later. (I’ve already fixed them myself.)
However, today I came across a clear example: a list of preposition exercises that incorrectly included adjectives, mixing up grammar and providing straight-up wrong translations.
I don’t plan to list every incorrect sentence, as it would be too time-consuming. My question is: Does this happen with other languages as well, or is it mainly a problem with German and Spanish?
For instance, I just opened an exercise and saw this German phrase:
“Die Milch vor dem Gebrauch unbedingt kochen!”
It was translated as:
“Forzosamente se debe cocer la leche antes de su consumo.”
This translation is problematic for three reasons:
A more natural and accurate translation would be:
“¡Es indispensable hervir la leche antes de consumirla!” or “¡Hervir la leche antes de su consumo!”
The point is, translations should sound natural and clear, not just be technically correct word for word.
I understand that learning a language requires active engagement from the user, and that occasionally checking translations is part of the process. However, as a paying Pro user, I expected a higher standard of quality. Constantly encountering errors that need correction shifts the responsibility from the service to the user. Instead of focusing on learning, I find myself spending time identifying and overlooking mistakes, which diminishes the value of the subscription.
Don’t get me wrong, but having to doubt the accuracy of phrases completely changes the experience for me.
Since many of the phrases seem to come from Tatoeba.org, I suspect the issue might stem from the quality of the source material itself.
The sentence you mention does come from Tatoeba. I posted a comment on it:
https://tatoeba.org/en/sentences/show/959873#comment-1546100
For what it’s worth, the person who translated from the German to the Spanish self-identifies as a native Spanish speaker.
I don’t plan to list every incorrect sentence, as it would be too time-consuming.
I agree that listing every sentence would be too time consuming. But your chances of having someone pay attention to your complaint will increase if you list several other sentences, especially here in this thread.
From my own experience, responses to complaints filed using the red “Report” button vary widely by language. I frequently receive notifications that sentences in Portuguese, Hebrew, and French have been updated in response to my feedback, but although I report sentences in Russian quite often (especially pertaining to the pronunciation field), I haven’t received a notification of an update since July 2022. Have you received responses to your feedback?
My guess is that English/X and X/English pairs have the best coverage and quality because they have the greatest number of Tatoeba members who know both languages well, especially when X is a widely-spoken language.
Before you explained that your example was from Tatoeba, and many others, too, I was expecting this to be an AI problem. There was another thread, months back, where someone pointed out that Google Translate goes to English behind the scenes when translating between any other pair of languages. (Chances are that many other services do the same thing.) That would lead to frequent screw-ups in a situation like this.
Pity it isn’t an AI problem. If it were, then maybe the safest move would be to use a German/English course to learn German instead. That may still be the best move right now, on the other hand…
I have also seen plenty of exercises at Clozemaster focused on a part of speech (like prepositions in your example) where many individual clozes do not represent that part of speech. That is a purely Clozemaster (not Tatoeba) problem caused by reliance on part-of-speech classification tools that have a significant rate of error, coupled with insufficient verification of the result by humans.
In order to correctly identify the part of speech of a word in a sentence, you potentially need a substantial amount of information from the rest of the sentence. I can easily imagine tools that cannot collect or make use of that information. That’s where humans need to come in, and they need to be good at sifting through errors without getting overwhelmed.