I noticed there are now LLM translations on the site that can not be translated via translate.online-go.com, specifically the PlayPageHelp.tsx strings. The translations are moderately awful, and IIUIC there is not way to influence the translations not to sound like poetic riddles?
LLM translations in reports are also bad, there’s a thread in the CM forum on it.
Maybe this should become the next thing I dig into.
The report translations are astonishingly bad ![]()
At least for the automatch help text (and other automatic English → target language transitions) the main complaint I have is with the choice of terms. The go terms in Estonian (and presumably in other smaller languages) are not established well enough for automatic translation to know of the obvious correct terms. I’ve picked reasonable terms in manual translations (eg from similar chess terms). But the automatic translation will awkwardly (and inconsistently with the manual translations) speak of loosely the equivalents of “exchange of letters games“ and “degenerate stones”.
It does also make other mistakes, but I would really want to be able to feed the choice of terms into the translation context, so the automatic translations would be consistent with the manual translations.
We’ve done a bunch of prompt context work on the canned messages, but not on the report initial description from the user.
I thought you might be interested to see this conversation I had with Claude on the topic (appended).
I guess providing context for report translation is a low hanging fruit step we can take,
Me;
Translate: 뉴비라고 욕하는 나쁜 사람입니다. 정지먹여주세요 급수도 속였습니다
Claude:
Deciphered Korean complaint and identified gaming server report context
This is a report about another player (likely on a Go server):
“This is a bad person who insults [others] calling them noobs. Please ban them. They also lied about their rank.”
뉴비 = “newbie/noob” (from English) 욕하다 = to insult/swear at 정지먹여주세요 = “please give them a suspension/ban” 급수 = rank/level (in Go) 속였습니다 = deceived/lied
So they’re reporting someone for being abusive and for sandbagging (misrepresenting their skill level).
Me:
Can you help me understand why an autotranslation said “I’m a bad person for calling you a newbie. Stop feeding me You cheated the water supply”
Claude:
Dissected multilayered translation failures across subject ambiguity, slang, and homonyms.
That’s a great breakdown of how machine translation fails. Here’s where each error came from:
“I’m a bad person for calling you a newbie”
Korean often drops the subject. The sentence is reporting someone else:
[이 사람은] 뉴비라고 욕하는 나쁜 사람입니다 “[This person] is a bad person who insults others as noobs”
The translator guessed wrong and made “I” the subject.
“Stop feeding me”
정지먹여주세요 = 정지 (suspension) + 먹여주세요 (please give)
- 정지 means “suspension/ban” but also “stop”
- 먹이다 literally means “to feed,” but 정지를 먹이다 is slang meaning “to give someone a ban” (like “feed them a suspension”)
The translator took both words literally: “stop” + “feed me”
“You cheated the water supply”
급수 is a homonym:
- 급수 (級數) = rank/grade (in Go)
- 급수 (給水) = water supply
The translator picked the wrong one. 속이다 = to deceive/cheat.
So “they lied about their rank” → “cheated the water supply”
A perfect storm of dropped subjects, gaming slang, and an unlucky homonym choice.
Me:
How did you avoid those traps?
Claude
Leveraged contextual clues and domain expertise to disambiguate translation.
Honestly, a few things:
1. Context
We’d been talking about Go servers, player reports, and bans for several messages. So when I saw 급수, “rank” was the obvious meaning - not “water supply.” Machine translators often process each sentence in isolation.
2. Recognising the genre
This is clearly a player report. The structure “X는 나쁜 사람입니다” (“X is a bad person”) followed by “please [action]” is a complaint about someone else, not a confession. I expected a report, so I read it as one.
3. Gaming slang
정지먹이다 (“feed a ban”) is common Korean gaming slang. Machine translators often struggle with slang and idioms because they’re not in formal dictionaries or training data as frequently. I’ve seen enough Korean gaming context to recognise it.
4. Sanity check
The machine output - “Stop feeding me. You cheated the water supply” - is nonsense. It doesn’t function as a coherent message. My interpretation - “Please ban them, they lied about their rank” - makes sense as something a human would actually write.
Machine translation often fails because it makes locally plausible choices (정지 → stop, 먹이다 → feed, 급수 → water supply) that are globally incoherent. I’m reading for meaning first, then finding words that fit - rather than translating words and hoping meaning emerges.
I’m gonna cheat the water supply so hard ![]()