it automatically converted “Russian Federation” to “Mordor”

March 6, 2023 9:24 AM

A major advance in translation technology means that Ukrainians can inform and debunk in real time. The first Google Translate used phrase-based SMT [statistical machine translation] — phrase-based, because it translates one phrase at a time, without considering the context of the phrase.... “Milkport” — from the Turkish süt liman, an idiom akin to “smooth sailing” — became Turkish shorthand for an amalgam of ludicrous machine translation and fake news.
posted by spamandkimchi (22 comments total)

Hard for me not to approach this as an indicator that the demand for language model hype is now so insatiable that Politico has to do it, with an article about an innovation from more than half a decade ago. The updated Google Translate definitely was a significant milestone in retrospect, though, and it’s interesting how low-key its rollout was at the time.
posted by atoxyl at 9:59 AM on March 6, 2023

On Google Translate's limits.

Google Says Google Translate Can’t Replace Human Translators. Immigration Officials Have Used It to Vet Refugees.
An internal manual produced by U.S. Citizenship and Immigration Services, the federal agency charged with admitting immigrants, instructs officers who sift through non-English social media posts of refugees that “the most efficient approach to translate foreign language contents is to utilize one of the many free online language translation services provided by Google, Yahoo, Bing, and other search engines.” The manual includes step-by-step instructions for Google Translate.
posted by spamandkimchi at 10:30 AM on March 6, 2023 [4 favorites]

TIL that WRT LLM, its vaunted 'attention' is for sussing context cues...

Obvious, but yeay for our NLP overlords??
posted by Heywood Mogroot III at 10:41 AM on March 6, 2023

Ok, Heywood, let's give this a shot... I had Google Translate take your obfuscated comment and translate into Ukrainian, then back to English. The result:

BEFORE WRT LLM, his vaunted "attention" is for identifying contextual cues...

Obviously, but so for our NLP masters??

Not bad...
posted by gwint at 10:52 AM on March 6, 2023

The article made me wonder: As machine translation becomes ubiquitous, will that mean the age of English as the world's lingua francasoon be over? Like, why would a non-native English speaker need to bother?
posted by gwint at 10:54 AM on March 6, 2023

Like, why would a non-native English speaker need to bother?

International air traffic control requires both pilots and controllers to be able to speak English - AFAIAA French airline landing at a French airport can speak French if both sides are fluent, but they must be able to fall back on English to a standardized degree of fluency. It’s not a context where anybody is going to let an unreasoning language model play an intermediary role.

Vastly more programming documentation and examples are written in English than any other language, mostly because the people who created the language itself did so in English (though I suspect the amount of simple device driver examples written in Mandarin is now higher). While very simple programs can be written by language models if a fairly high number of functionally identical tasks are included in its training set (modern models can pick up both the syntactic uniformity and the structural similarities), the second you stray from common tasks it will go completely off the rails because there is no reasoning at work in language models - just an insanely complex graph of which words are most likely to follow which words, and the first 90% of programming is abstract reasoning (the second 90% is swearing, at which language models excel when permitted).

A hybrid of current language models and reinforcement learning could (and eventually probably will) solve the programming side of the above, with a couple decades of further research.
posted by Ryvar at 11:48 AM on March 6, 2023 [1 favorite]

The machine model can also be rigged to provide deliberate mistranslations: For a time, for example, it automatically converted “Russian Federation” to “Mordor,” “Russians” to “occupiers,” and the name of Russia’s foreign minister, Sergey Lavrov, to “sad little horse.” But Google Translate is used by too many people, daily, for fraud to be sustained.

I wish the article had explained this part--how was it rigged? Why do lots of people using it make the fraud unsustainable?
posted by polecat at 1:55 PM on March 6, 2023

The statement of whether or not Google Translate can replace a human translator is that it depends on who the human is.
posted by scolbath at 2:14 PM on March 6, 2023 [1 favorite]

Here's a 2016 BBC piece on the Mordor thing, but no actual explanation.
posted by spamandkimchi at 2:34 PM on March 6, 2023

Back when NMT was first becoming a thing, I thought that the knowledge professionals cheering for the resulting "reduced translation costs" (aka ability to force translators into desperate poverty) were whistling past the graveyard. For example, to be an effective legal translator I already needed to know lawyers' jobs pretty well, in addition to my own. Once it was possible to do a passable job of automating the translation of professional work product, automating the work product itself couldn't be far behind.

These days my income is a mixture of legal and translation work, so I would prefer to have been wrong. But I guess I wasn't. Translators were the proverbial canaries in the coal mine -- but the miners were too busy carving us up to realize the implications for themselves.

Anyway, I think that talking about whether MT can replace a human translator is somewhat beside the point, just as it is for other professions. There was a discussion here recently about some seemingly unrelated topic -- music maybe? -- where the point was made that having to compete against your own stolen work product will tend to depress prices, even if customers would prefer to buy the genuine, non-stolen stuff. That's substantially the effect that MT has had, and that "AI agents" are very likely to have, on the industries they impact.
posted by Not A Thing at 3:06 PM on March 6, 2023 [8 favorites]

Haven’t translation models been translating more than a word at a time for well over a decade?
posted by MisantropicPainforest at 5:07 PM on March 6, 2023

I wish the article had explained this part--how was it rigged? Why do lots of people using it make the fraud unsustainable?
Maybe by large numbers of people correcting Google Translate and it learns from them? If so, it would only work for any length of time through sustained effort, because it would keep learning from later 'correct' translations.
posted by dg at 8:49 PM on March 6, 2023

Politco is literally sponsored by Lockheed Martin. They run endless propaganda for endless war. So here we're made to think Europe united because Google Translate, and not because it's a colony of America. It's also not mentioned that Europe is not united, people are protesting, and most people would rather have heat.

Meanwhile the US has blown up Europe's energy supplies, is making a ruin of Ukraine while arms dealers help themselves, and you dummies cheer it on, just like the Iraq War and all the wars before.

At long last, have you no shame?
posted by indica at 2:30 AM on March 7, 2023

It's also not mentioned that Europe is not united, people are protesting, and most people would rather have heat.

Most people would rather have heat than not have heat, that's for sure.

But if you're saying that people would rather have heat than stand by Ukraine then no, that's emphatically not true. As a Polish citizen I'd rather have no heat than see Ukraine defeated and Ukrainians slaughtered.

Let's not repeat Russian talking points.
posted by M. at 7:27 AM on March 7, 2023 [12 favorites]

Meanwhile the US has blown up Europe's energy supplies, is making a ruin of Ukraine

I would have thought that the country making a ruin of Ukraine is the one that invaded it and is in the process of murdering its citizens and bombing its cities. Or are you suggesting that Putin is secretly under US control?
posted by tdismukes at 11:26 AM on March 7, 2023 [9 favorites]

That's because you have a connection to reality rather than tankie-world.
posted by tavella at 11:57 AM on March 7, 2023 [3 favorites]

Europe united because Google Translate, and not because it's a colony of America

I can assure you, through direct experience as well as ample study of history, that what's going on with Poland-Baltics-Belarus-Ukraine-Russia right now has very deep roots and is very personal. And aside from the injection of $$$ would be very similar with or without the US in the mix.

NATO is probably actually a constraint on Poland, I would think if they were not entangled there would likely be massive "volunteer" type Polish brigades or Friekorps on the frontlines.
posted by Meatbomb at 11:57 AM on March 7, 2023 [3 favorites]

Europe united because Google Translate, and not because it's a colony of America

Right yes, nobody outside of America is a real person with agency, or governments, or any kind of existence at all, except for how it relates to and is controlled by the US (and maybe Russia or China). Of course.
posted by Dysk at 8:25 AM on March 8, 2023 [7 favorites]

I’d add Finland too to that mix of its personal, its historic, and its complicated, complex, and contradictory, meatbomb
posted by infini at 11:09 AM on March 9, 2023 [2 favorites]

Right yes, nobody outside of America is a real person with agency, or governments, or any kind of existence at all, except for how it relates to and is controlled by the US (and maybe Russia or China). Of course.

For the longest time, I thought that way too... but events over the past 7 or 8 years have completely disabused me of this notion.
posted by infini at 11:12 AM on March 9, 2023

... nobody outside of America is a real person with agency, or governments, or any kind of existence at all, except for how it relates to and is controlled by the US (and maybe Russia or China). Of course.
Well, that's the way America (the country) acts and the way the overwhelming majority of Americans act, in my observation. I mean, the idea that Europe is a colony of America is obviously absurd, but there's a kernel of truth there in the way that Americans (not all, but a LOT) view the entire world.
posted by dg at 1:13 PM on March 9, 2023

I am aware of that. The offensively ludicrous nature of that was what I was pointing out, through the medium of sarcasm. That it exists is not a reason to indulge it engage in it here, as with any other form of agency-denying othering.
posted by Dysk at 2:23 AM on March 10, 2023