Microsoft reaches new breakthrough in translation

[manpo]

https://blogs.microsoft.com/ai/machine-translation-news-test-set-human-parity/

long story, short; Microsoft believes that they have made the first system that can translate Chinese news articles to English with the same quality and accuracy of a professional translator.

Apparently this is not some program that searches words in a dictionary and try to merge the results together, It's more like a self - learning AI( one of its learning method is that once it translates a sentence, it translates it back and sees how much it has held its meaning and improve its database like that)

Of course this may mean much since the material was only chosen from news articles and so far it has only done Chinese to English; even the team has stated "No one knows whether machine translation systems will ever get good enough to translate any text in any language pair with the accuracy and lyricism of a human translator." ,but even so, this is still a great achievement since their test results matched that of professional translators.

Now the team is feeding their system with wider contents, not just news articles, and the bright point is this can work between any two language, since it's a self-learning AI.

If you know Chinese, you can test the translation here: https://translator.microsoft.com/neural/

[Funyarinpa]

Fucking finally, I can read all the Japanese tweets I get on my timeline

[manpo]

 

3 minutes ago, Funyarinpa said:

Fucking finally, I can read all the Japanese tweets I get on my timeline

Don't get too hyped, this is still far from complete

[Asonn]

Within 10 years all translators will lose their job : ) another plus point is that we don't have to deal with bad TL's like Moenovel. Although I suspect the editor's workload will increase. 

[manpo]

3 minutes ago, Asonn said:

we don't have to deal with bad TL's like Moenovel.

couldn't agree more on that

[Funyarinpa]

4 minutes ago, Asonn said:

Within 10 years all translators will lose their job : ) another plus point is that we don't have to deal with bad TL's like Moenovel. Although I suspect the editor's workload will increase. 

This is actually an interesting point to bring up (and to contrast with the threat of automation to many assembly line jobs). But I don't think it'll be the death knell of translation. One, there simply won't be proficient AI translations between all two given languages, so perhaps only translators who translate to/from English are at risk. Two, translation can survive by focusing on stylization- if AI translation is the invention of the camera, then translators are still painters. Does the analogy make sense?

[EastCoastDrifter]

I'm still a little skeptical about this. Sure, it's just a test with one language, but until I see some actual progress with other languages, and if their results are accurate and correct, then I'll have more faith in machine TLs.

[Plk_Lesiak]

This might be another neat tool to speed up professional translation process and when technology like this gets cheap and dependable enough, it will partially replace human translators, especially in places where cutting costs is the priority. People actually having skills in translating will always be needed though in more specialized fields (technical and academic spheres, for example), projects in which quality is the main focus and for quality control. Even the machine learning mechanisms will most likely need to serious oversight and constant tweaks from translation professionals, you can't trust all of it to algorithms. Just as other rumours about machines replacing us in various fields, this is quite premature to talk about human translators becoming obsolete.

[manpo]

8 minutes ago, Plk_Lesiak said:

This might be another neat tool to speed up professional translation process and when technology like this gets cheap and dependable enough, it will partially replace human translators, especially in places where cutting costs is the priority. People actually having skills in translating will always be needed though in more specialized fields (technical and academic spheres, for example), projects in which quality is the main focus and for quality control. Even the machine learning mechanisms will most likely need to serious oversight and constant tweaks from translation professionals, you can't trust all of it to algorithms. Just as other rumours about machines replacing us in various fields, this is quite premature to talk about human translators becoming obsolete.

Actually Microsoft is in competition with Google on this; since they made a breakthrough first their article tends to be a bit exaggerated.

PS: I was just taking a breaking from studying kanji when I came upon this. lol

[VirginSmasher]

And some people said translation by computers would never be perfected. It'll be a long time, but I can imagine that in 5 to 10 years, we'll have AI translation almost perfected.

[Clephas]

1 hour ago, VirginSmasher said:

And some people said translation by computers would never be perfected. It'll be a long time, but I can imagine that in 5 to 10 years, we'll have AI translation almost perfected.

Probably only between compatible languages.  I wouldn't get your hopes up for perfect machine translations from Japanese to English, and I seriously doubt that this will work perfectly for Chinese. 

[VirginSmasher]

11 minutes ago, Clephas said:

Probably only between compatible languages.  I wouldn't get your hopes up for perfect machine translations from Japanese to English, and I seriously doubt that this will work perfectly for Chinese. 

Well, I can only hope there's a future where machine TLs end up being really good and not an incomprehensible mess, but that'll be a long time from now.

[Clephas]

1 hour ago, VirginSmasher said:

Well, I can only hope there's a future where machine TLs end up being really good and not an incomprehensible mess, but that'll be a long time from now.

We'll probably see perfect machine translations between romance and germanic languages in relatively little time (even as soon as next year, given the basic similarities between the languages and their usage).  However, I give it at least thirty years before they manage to perfectly manage the transition from an Asian language to English.

[VirginSmasher]

3 minutes ago, Clephas said:

We'll probably see perfect machine translations between romance and germanic languages in relatively little time (even as soon as next year, given the basic similarities between the languages and their usage).  However, I give it at least thirty years before they manage to perfectly manage the transition from an Asian language to English.

And by then, a majority of the good VNs will have come out in the West by then. 

Not like I would be reading VNs anymore at 48.

[ittaku]

If you know the two languages, believe it or not doing Chinese to English is infinitely easier than doing Japanese. Japanese could well be the LAST language where translation to English becomes accurate. Sorry but it's still time for the rest of you to learn Japanese. As a software engineer that knows Japanese, I'd say give it another, oh, say 30 years (which is what @Clephas said about all Asian languages though I disagree)?

[Yuuko]

Still can't translate stuff that human can't translate either though that is MOE and MEMES

[Zakamutt]

I totally disagree with Asonn's timeline (but he's obviously memeing), but this part is pretty silly. Sure it's bad for the translators if they're out of a job, but MTL actually getting good enough to use for good translations (it really isn't even close) would be an amazing asset for mankind and arguably mean we're close to or already at artificial general intelligence. Which admittedly might kill us all. Good luck everyone

[Dreamysyu]

Well, even if we consider that this system works really well, I still wouldn't trust it to translate fiction. There are generally too many things to keep track of, stylistically and plot-wise. Remember that part in Ever17 where incompetent translation managed to almost spoil an important plot point? On the other hand, it could be used to accelerate the process of manual translation. You could use it to get a general outline of the text and then just compare the original text to it to see if it's correct.

[Clephas]

We will probably be able to trust machines to translate most languages' objective writing (non-prose, informational stuff) relatively early on.  However, realistically, machines are unlikely to be able to handle cultural and emotional context and the nuance of prose before we manage to create a perfect human emulator AI, lol.