How Cold War Computers Failed At Russian – MT #2 – with CompChompTony wyaad September 10, 2019 36 Comments
Last time on CompChomp:
Cold War computers were showing real promise at translating Russian into English. Georgetown
University staged a successful tech demo and the press was impressed. Times were good for
machine translation. Okay, now that we’re back, the money starts
rolling in because if the public supports something, politicians can’t lose by throwing
money at it, but people started taking sides. Round 1, fight!
In this corner, you had your linguistically minded perfectionists. They wanted to feed
computers all the detailed grammatical rules about every language. And in this corner,
there were your brute force crowd. Brute forcers were all about the evidence. You set computers
loose on real sentences and they should be able to find the patterns on their own.
Unfortunately, progress from both sides just wasn’t living up to the bold claims made after
Georgetown and the public, they started to notice. In 1962, Harper’s ran an article titled
“The Trouble with Translation”, and with a title like that you know it was going to be
fair and objective. The author recalled a demonstration of machine translation where
the computer was given the scriptural nugget, “The spirit is willing but the flesh is weak.”
The output, we’re told, is laughable: the liquor is holding out all right but the meat
has spoiled. I just want to pause for a moment. The demo reported there didn’t happen, but
this story in Harper’s really illustrates how far machine translation had fallen. A
few years earlier, those exact same reporters were just heaping praise on it and now it
was the butt of their magazine jokes. The scientists stepped in for some damage
control. Do you remember Bar Hillel, our first full-time machine translation man? He argued
that our expectations were just unreasonable. Natural language could be ambiguous and computers
cannot handle ambiguity. We shouldn’t expect them to translate languages for us. We should
just get them to help us translate for ourselves. Translating languages was way more complex
than even the experts could’ve guessed – it was literally turning out to be harder than
rocket science. Now, in the face of souring public opinion,
the US government – because, again, they only like throwing money at popular things – they
commissioned the ALPAC. Josh, is that a llama? Wait, no, sorry. It’s not a llama. It’s actually
an elite group of seven experts tasked with deciding the usefulness or uselessness of
machine translation and computational linguistics as a whole. That’s… quite a task!
Their verdict? Oh, we’ll… we’ll get into that. But first there’s something you need
to understand. Throughout the history of Artificial Intelligence,
people get sold on promises far beyond their wildest dreams and it happens so often that
there’s a term for it: the AI Hype Cycle. And when things go wrong, a chill sets in.
A chill with lasting effects. It’s called an AI Winter. (shivers) And no amount of jackets
can protect you from that chill. So, the verdict. The 1966 ALPAC report concluded
that “we do not have useful machine translation and there is no immediate or predictable prospect
of useful machine translation.” Ouch! Computers, it turned out, were twice as expensive
as human translators and they were doing way, way worse. The officials were convinced and
machine translation was defunded. The first AI winter set in.
Hmm. Computers and language obviously had a rough
start, but I promise it gets better. Eventually. Subscribe, go eat your borscht and, uh, write
some code. Chomp!