Gizmodo fires Spanish staff amid switch to AI translator

Maybe Japanese is not a good target language for this type of test, because its very.. high context, that is to say. the natural way of expressing the ideas of the origional text in japanese, might leave some of the meaning to context, and not have it explicitly marked or encoded in words left on the page, then when that is translated back to english that part is lost. also it doesnt provide a good test of if the japanese in the middle was any good.. like in this case it probably had a lot of "watashi"s in place of "I" where this woud not be natural in japanese.

The fun part of chatgpt is you can ask it to do non-standard translations like

"can you give me the direct non idiomatic translation of the following Japanese into English, to help me remember these phrases in English words but the Japanese way of saying it"

Sure, I can provide a direct, non-idiomatic translation of the Japanese text into English. This kind of task is straightforward for me, so no worries.

"Big forest's right next to, poor wood-cutter and his wife, and two children were living. Boy was Hansel, girl was Gretel called. He had almost no food, once big famine attacked the country, daily bread even could no longer obtain. Night, in bed this thing thinking, anxious tossing and turning, he groaned, to wife said: 'We what become? Poor children to how food provide, ourselves even nothing when.' 'Husband, tomorrow morning early, children to forest's thickest place take let's go. There for them fire light, each one more piece of bread fragment give, then we to work go, them alone leave. They home return way find not become, we them let go can.' 'No, wife,' man said, 'that can't do, why I children in forest alone leave can? Wild animals soon come, them tear apart will.' 'Oh, fool!' she said, 'then we four people all die of hunger, our coffins' plank plane might as well.' And, she him agree until him peace not did. 'But, still poor children very sorry,' man said."

I've used GPT-4 extensively for translation to Vietnamese and it's _incredible_. It's the one use case where, after my wonderment wore off, I'm still consistently amazed. Vietnamese is extremely difficult to translate to and from English - every other system I've used sucks badly.

It doesn't give us a rough first draft. It gives us the equivalent of a second or third draft. The equivalent to the final result we'd get back from translators after a lot more back and forth.

Honestly, it's better or equivalent to any human service I've used. Usually when it translates badly, a human would also translate badly and it means that section needs rewriting.

It handles all kinds of things that I wouldn't expect, idioms, jokes, technical writing. There are very few mistakes.

I can only imagine it's even better at Spanish.

We don't pay for human translation anymore. The workload is reduced enough that we do the final editing ourselves.

We still use human translation for legal stuff because that has to be done by an official translator. So I guess that's safe.

Wonder how it'll handle Korean?

A friend who translates Korean has pointed out that it's kind of special, in that the symbols used are exactly the same for many different things. So the only way to know the actual meaning is from context, which is what then needs translating.

Apparently online translators from Korean to other languages are truly, horribly bad due to this.

Korean is the language I know. What I said is the same for it.

Hangul(Korean writing system) is not a logograph like Chinese/Japanese but an alphabet. Many words do have different meanings but all languages have this to an extent. I don't think that's the reason Korean is context dependent.

The bigger reason is that Korean is very pro drop.

Strictly speaking 친구 있어요 ? means Friend Have ?

And you would say exactly that for any of these English equivalents

"Do they have friends ?"

"Do you have a friend ?"

"Do you have friends ?"

"Does he have a friend ?"

And so on.

Plurality as a special marker doesn't really exist. Possessive and Article Determiners don't really exist either. Articles don't exist at all. Possessive markers are few and only used when surrounding context might not be enough to convey the needed information.

Subject/Object/Topic markers exist but are regularly omitted.

You might ask, well how would you know which of those equivalents 친구 있어요 means ?

There is no way to tell before reading or hearing the surrounding context.

Traditional models are unable to decipher this context and translations are only good with close language pairs. Google translate can be very good for English to French and Japanese to Korean but fail on Korean to English.

This is not a good test of translation systems. It's like pointing out that compressing a zip file lead to a larger file, or converting a JPG to a PNG and then saving it as a JPG again lead to more artifacting.