Translating GZDoom's text content. Read if you want to help

[Professor Hastig]

I actually have seen similar results from actual Human translators over the years. The results of someone having to translate a text they do not really understand but being forced to come up with a result can truly create ludicrous results, regardless whether it's man or machine doing the translation.

Indeed. I also have seen it many times that if you give a Human translator some highly specific text the translated result is often just as unusable as a machine generated translation - where you may even have more luck with the machine if it has seen some training in the specific field the text is made for. Care has to be taken, of course when you translate between languages with highly different gender rules - then it can quickly happen with a machine to get these wrong - just like the aforementioned example of formal/informal address mode in German and French.

[Gez]

Care has to be taken, of course when you translate between languages with highly different gender rules - then it can quickly happen with a machine to get these wrong - just like the aforementioned example of formal/informal address mode in German and French.

And this, right there, is always going to be a problem with AI translators, forever and ever and ever, because all machine translation works on the same principle, which is:

1. Translate from input language to English
2. Translate from English to output language

And English lacks gender representation, it surviving only in a few pronouns, as well as the T-V distinction. So these things are always lost.

If the AI translator worked directly from German to French, there would be no problem. But it goes from German to English, and then from English to French.

That use of English as the intermediary language for machine translation is born entirely out of English's predominant status as the language that everyone and everything should use by default every time, and not out of any sort of fitness for the purpose. Ideally, you'd use specialized translators for each language pair, but that's a ton of work. Alternatively, design a conlang meant to carry as many semantic and grammatical marker as possible with every word so that it can offer lossless translation. What we have is the equivalent of using grayscale as the intermediary step for converting image colors between RGB and HSV.

[Professor Hastig]

You are correct, of course that a translation AI may never be able to replace Human translators - but there is no doubt that it will make the job a lot easier. Even if it just eliminates half the work it will be a major gain in productivity.

Besides, I am quite confident that this English as an intermediate problem will be solved eventually. Like you said, it is currently one of the biggest roadblocks and whoever manages to overcome it first will dominate the business so I am fairly confident that significant research is being done here to improve matters.
Automated translations with the quality that deepl delivers were unthought of a few years ago and I do not think this is the end.

[Graf Zahl]

And this, right there, is always going to be a problem with AI translators, forever and ever and ever,

Yeah, right. Remember how often this was said about technical limitations in IT? Only to see someone overcoming it eventually? What on Earth makes you think that this one will be the ulitmate exception to the rule?

Automated translations with the quality that deepl delivers were unthought of a few years ago and I do not think this is the end.

And wanna bet that 5 or 10 years from now people will laugh about the quality of deepl because it is so much worse than what they will be used to?

[wildweasel]

We are at a point where machine translation, even with AI, has progressed past the point of producing obviously wrong translations with grammatical errors, and started producing translations that look convincingly correct but carry the wrong intent. If anything, that's far worse than the obviously incorrect translations and could easily lead to major communication breakdown.

[Graf Zahl]

You get that from Human translators as well. You wouldn't believe some of the shit I have seen over the years working on localizing software.

[phantombeta]

You get that from Human translators as well. You wouldn't believe some of the shit I have seen over the years working on localizing software.

Case in point

[Professor Hastig]

Touché.

While I can understand the reservations some people have with automated translations it is easy to forget that Human made translations can vary strongly in quality and it is very easy to get poor results as well.
So either way a careful review of the translated text is an absolute necessity

[Gez]

Yeah, right. Remember how often this was said about technical limitations in IT? Only to see someone overcoming it eventually? What on Earth makes you think that this one will be the ulitmate exception to the rule?

Because that's not a technical limitation, more a cultural or even ideological one.

[Professor Hastig]

If this was all American companies, I might agree, but deepl is German, so the basic conditions are very different to begin with. The relatively small list of their languages suggests that they don't want to go all out on the broadest support possible but to do it right. No, currently they clearly are not - or at least not fully - because sometimes there's visible artifacts of an intermediate English step if you translate between two other languages.

Nothing of all this suggests in any way there's no way to overcome it, though.

[Graf Zahl]

Just a thing I noticed by pure chance:

In Strife Veteran Edition a new character called "Master Smithy" is added. And for some reason it was translated to "master smith" (as in master of smiths) in ALL languages (except German which I did myself) I guess this is one of those places where all Human control did not help - the machine translators also got this consistently wrong in all languages I checked.
I have to admit I am a bit puzzled why nobody realized that Smithy in this case was just a name, not a profession.

[Graf Zahl]

Because that's not a technical limitation, more a cultural or even ideological one.

And I ask you again: Why do you think this is an insurmountable limitation? "Just because it was always done like that." does not count for a good reason.

[Xeotroid]

Oops, that was an error. I definitely should have noticed it in the dialogue line due to the capitalisation of the S and obviously it being strange to call a person Smithy (the building), not Smith (the person). However, to my understanding there should be commas or dashes surrounding the name too, then: "Talk to the master, Smithy, if you have questions." This is ambiguous, however, as it might be confused for the character, for some reason, calling the player Smithy. Alternatively, the definite article could just be dropped: "Talk to master Smithy if you have questions."
With how the original string is written, it sounds like "master Smithy" is a title – not a title and name combination – and just an odd, incorrect spelling of "master smith". As for the name tag, all of them are capitalised ("Warehouse Guard"), so there was no clear distinction to catch on to.

[Gez]

And I ask you again: Why do you think this is an insurmountable limitation? "Just because it was always done like that." does not count for a good reason.

Because the people who work on this stuff do not even realize there is a problem.

[Graf Zahl]

Somehow I doubt that if even laymen like us can see the problem. I was just checking up on some translations today, and when converting them to German I could immediately spot some residual English idioms in the translated text, despite both languages not being English and the original text definitely not containing these idioms. But when mentally translating to English all things fell in place. This is the #1 reason why machine translation is not what it could be and you REALLY believe that these people do not care about it? I'd rather say they have a problem of incomplete databases that are yet to be built with English as intermediate only being a stopgap measure.