Deepl and the discovery of America

Apparently deepl thinks that the discovery of America cannot have taken place in 1492 and autocorrects the year to 1942!

17 Likes

Is that the director of Harry Potter?

2 Likes

Yeah. Also, I’m increasingly noticing that when translating larger blocks of text, Deepl will occasionally replace a sentence’s translation with the translation of the sentence next to it, and it’ll keep using that translation even when I’ve deleted the adjacent sentence.

For example:

Deepl, only one of those bullet points actually reads “Do not use if you have scars, rashes, or other skin problems.” And if I delete that bullet point from the original text, it simply removes one of the “Do not use…” lines from the translation.

It’s doing me a concern.

7 Likes

Seriously, what is up with that number lmao

18 Likes

Off topic. I’ve never used deepl, how does it compare to Google’s translator?

EDIT: I’ve noticed now that it doesn’t allow translation from English to Japanese, only the contrary, so it can’t fully replace Google as of now

1 Like

Toss all the numbers in a bag, shake it up, and draw them out at random.

It’s several orders of magnitude better, with the biggest caveat being that it tends to favour a natural-sounding translation over an accurate one. So, sometimes it… guesses.

Ohhh, I noticed that too! So I’m not insane!

The trick is that you need to set the “from” language to English first. Then Japanese appears in the “to” language list.

5 Likes

This could be that they trained the network on texts that contained mistakes like that, and so now the network has learned the mistake! :slight_smile:

1 Like

Thanks, that’s very useful to know.

2 Likes

1942 was the battle of Midway so maybe it just parsed a lot of text containing that year.

1 Like

its not a translater, it’s a fancy statistical language model :stuck_out_tongue:

3 Likes

Yeah but that’s pretty much all of us as well. lol

6 Likes

lets not repeat the AGI debate we had before (fun as it was) :upside_down_face:

4 Likes

I like that it goes for a more natural-sounding translation over a literal one, but Google has the romaji along the bottom which is really nice if I am copy/pasting a chunk of text I can’t read. It’s not perfect, but it is nice.