I'm taking Thanuir's suggestion to start a new thread based on danepo's post about Clozemaster, which uses sentences from the Tatoeba corpus ( https://tatoeba.org/eng/wall/sh...#message_32447 ). danepo wrote:

> Here are som extracts from user reviews on the Android app store:
>> I was very excited at first about this app, until several native french speakers told me that phrases I'd learned on Clozemaster aren't correct French. Discovering this after having paid for the premium version was extremely disappointing (20181228)
>> The idea of the app is nice, but as I was testing it in Italian i realised there are a lot of mistakes in the translations, at least from Spanish, and that would be only confusing for my students. That's why I'm giving the app just 2 stars (20171102)
> Does the above comments mean that the French sentences on Tatoeba aren't correct French and that there are a lot mistakes in the translations from Spanish into Italian?

My response:

While it's potentially useful to see what people are saying about our corpus (or a subset of it), there's a limit to what we can learn given the absence of examples and statistics, and the difficulty of following up with the people who left the comments. For instance, we don't know whether the sentences criticized by the native French speakers are wrong in the sense of being something no native speaker would ever say, or whether they're something native speakers might say, but only informally.

In my experience with English-to-Russian on Clozemaster, all of the Russian sentences and English sentences that I've seen seem to be correct, and I've seen between 1000 and 1500 sentences altogether. I've complained about a poor match between a Russian sentence and an English translation perhaps 10-15 times. More frequently, I've seen sentences misclassified by assigning the wrong part of speech to a Russian word (for instance, treating the adjective готов like a noun because it has the ов ending characteristic of many nouns in the genitive plural). However, that's not a problem coming from the Tatoeba corpus.

The biggest problem I've seen at Clozemaster has to do with the major theme of the past two weeks, which is the lack of diversity in the Tatoeba corpus, and hence in the third-party apps that take their sentences from it. Since Clozemaster has no sentences with typically Russian names, which are the only ones that are declined, I get no practice in this important part of the language.