Hacker News Reader

Comment by baroffoos | Hacker News Reader

I recently had an experience getting software translated to Japanese. Just to test things while the translators were working I ran the all the yml files through google translate and then when we got the proper translations back I had a look at some of the things that changed that I would not expect to.

One interesting one is we had translations for status indicators (Red, Amber, Green) The real translator had translated these to the Japanese versions of Red, Yellow and Green but google translated had picked the word for "fossilized resin" rather than the color.

Since google translate is not using any context it has no way to know if I meant the color or the material.

View on HN

FYI, google translate for Japanese is pretty unreliable for lone uncommon words. It seems to rely heavily on internet-trained ML, and does pretty well with phrases and sentences, but for single words I see it return nonsense fairly often.

As an alternative, https://jisho.org/ is super easy to use and hasn't let me down yet.

In the other direction, I've seen gtl invert the meaning of sentences on subtle variations of the なければ pattern, which is probably a fairly small error in the problem space but, well, inverts the meaning of the sentence. A good reminder that parallel corpus machine learning does not see the world the same way you do.

Personally, when I'm checking a translation, I like to use WWWJDIC's text glossing mode: https://www.edrdg.org/cgi-bin/wwwjdic/wwwjdic?9T

The thing with my Japanese skills is that my grammar is excellent, but my vocabulary is limited, so a gloss of the kanji in a sentence is _massively_ useful to me (and kanji is a double-edged sword: I recognize far more words in hiragana (or when spoken) than in kanji, but kanji is pretty much necessary for a dictionary to have any decent idea of what the word is).

I'll have to give Jisho a try though.

Fun question: did the translators use green or blue for the green translation? Green (midori) is a newish word and semaphores, traditionally, use blue (aoi) when referring to the green light.

This is something I assume translators would be able to catch that would be much more difficult for translation software.

Reply on news.ycombinator.com