September 20, 2006

"Talking" Elephants, "Signing" Robots

Bill Poser comments on an article about a "talking" pachyderm, in which a journalist wisely remembered to mention "the difference between making word-like sounds and associating sounds with meanings". I wish the folks at Wired Magazine would be that careful.

Wired is otherwise wonderful; my whole family looks forward to its arrival each month. But every now and then they go a bit over the top when it comes to language technology. The most recent tooth-grinder: a blurb about a robotics display at their upcoming NextFest, concerning a robotic hand from the University of Tsukuba. The blurb says it "learns movement as an infant might, by watching and mimicking human behavior", and can "talk in sign language".

Robotic hand

Now, as far as I can tell, the researchers themselves are nowhere near as unrestrained. They say their goal is to "understand the dexterity of the human hand and imitate that with robot hands". After browsing around their publications a bit, I don't see any references to "talking" or to claims about emulating infant learning. I'm at a loss to understand why the hype was added -- isn't really cool robotics cool enough? (Ok, yes, there may also be an element of sour grapes in my writing this: we natural language processing folks always have a hard time convincing people that what we work on is "sexy", compared to our brethren in robotics, computer graphics, etc.!)

A word to the wise: if someone says a computer system reads/talks/understands language or processes language "the way people do", you should assume either that they're being sloppy, or that they're trying a little too hard to sell you something.

Posted by Philip Resnik at September 20, 2006 11:41 AM