answer and you find your meal, I’d say the
answer is yes.
Even Knight himself, in his own article, mentions new natural-language-processing capabilities in question
answering and image captioning; rapid
advances in speech recognition and translation between languages; and parsers
that untangle the relationships among
words in a sentence. He might also have
pointed to search engines, sentiment and
opinion analysis, information extraction
and summarization from text, assistive
devices, and more. For certain dialects
within certain languages, we’ve seen massive progress in language technologies.
The struggle we researchers face is
to consistently explain what it means to
“understand” formally enough to guide
algorithm design and measure how close
we’re getting to human-level understanding. (And this shouldn’t be surprising—
everyone has a brain, but our intuitions
about how it works don’t qualify us to perform brain surgery. Why would it be any
different with our capacity for language
So how will we know that we’ve made
AI systems that truly understand us?
Maybe it’s less complicated than we make
to pick up your kid from day care, you’ll
know it “understood” you when she’s delivered safely to your door. If a lawyer asks
her digital assistant to draft a brief for a
case, she’ll know it “understood” if she
Write MIT Technology Review
One Main Street, 13th Floor
Cambridge, MA 02142
Please include your address,
telephone number, and e-mail address.
Letters and comments may be edited for
both clarity and length.
Give AI a Little Bit of Credit, Please
Will Knight’s “AI’s Language Problem” is
more or less a call to action to anybody
(like me) who works in artificial intelligence. It’s time, Knight declares, for AI
to grapple seriously with language understanding. If AI systems can learn to win a
difficult game like Go, why can’t they yet
understand language—which the average
child learns effortlessly?
But Knight’s not really giving us as
much credit as we deserve. And he’s
being a little too prescriptive about what
it means to “understand language”—a
debate that’s raged at least since Turing’s
day. Does your phone “understand” you
when you ask it where the nearest sushi
is? If it responds with a good, humanlike
doesn’t need to revise it much (and when
she wins the case).
We’re making progress, even if it’s not
quite as fast as Knight would like it to happen. We’re seeing advances in tools whose
(at least partial) language understanding
complements and augments what humans
are capable of processing and understanding. As one of my students once told me, “I
want to understand what’s in documents
without reading them.” In finance, health,
law, politics, education, science, and many
other pursuits, natural language processing is giving us that power.
—Noah A. Smith
Associate professor of computer science
and engineering, University of Washington
Will Knight responds:
Professor Smith makes a good point
about progress in language understanding in recent years. He’s also right when
he says machines already “understand” a
great deal of what we say. Still, limitations
are revealed whenever you try to hold a
meaningful conversation with a system
like Apple’s Siri or Amazon’s Alexa, and
these show there’s a long way to go. The
problem will get even more serious as we
depend more heavily on machines that
have essentially programmed themselves.
To take Smith’s example, if the driverless
car failed to deliver your daughter to the
right place, it would be nice to be able to
ask what went wrong and to hold a dialogue so the car could get it right the next
time. If our goal is AI systems that do our
bidding most of the time, we’re doing fine.
But if we want machines that we can collaborate and coöperate with, they’ll need
a more sophisticated grasp of language.
Letters and Comments
MIT Technology Review
Volume 119, Number 5
Does your phone “understand” you when you ask it where
the nearest sushi is? If you find your meal, the answer is yes.