3 ms·

> But if someone said "95% of paint splatters OCR as Welsh words" and it turned out that's because the OCR is looking for words, it doesn't say much interesting

> But if someone said "95% of paint splatters OCR as Welsh words" and it turned out that's because the OCR is looking for words, it doesn't say much interesting about Welsh like the title implies. I think it depends on the claim, right? IMO, a more analogous statement might be "95% of paint splatters OCR as grammatically-correct Welsh sentences." Unless the software is using something like GPT-3, that finding would be notable regardless of software tuning. Even if it tells us more about Welsh grammar than it does paint splatters. > Look at the whole gallery though. Most of the OCR results are just a string of letters or a number. Those being valid is very mundane. Yes, thanks, my mistake!

Dylan16807 5y ago
> IMO, a more analogous statement might be "95% of paint splatters OCR as grammatically-correct Welsh sentences." Unless the software is using something like GPT-3, that finding would be notable regardless of software tuning. Even if it tells us more about Welsh grammar than it does paint splatters. Well, I'd argue that "potato" is a grammatically correct sentence in most languages. (Imagine someone asked you what kind of soup you're holding.) And so is "7" and so is "Fififz". And that's literally what's happening with Perl with the vast majority of these. It's treating a single word/number as a statement. If you want to say something notable about grammar, you need to be generating sentences that are longer than a single noun.