Mike Lorrey wrote:
>
> John Clark wrote:
>
> > Interesting! I haven't heard anything about CYC for a while and I was starting to
> > worry the entire thing was a failure. It seems to me that even if it's not the key to
> > AI CYC could be very useful in translation and speech recognition. I think the
> > trouble was that up to now programs had no idea what the words it heard
> > meant, I'll bet even a human would be wrong at least 10% of the time in
> > distinguishing between "I scream" and "ice cream" if he didn't have other
> > words in a sentence like "whenever I see a rat I scream" to provide context
> > and give clues. When a program can hear the words "this book has been red,
> > but now it's painted blue" and know I meant "red" not "read" I'll be impressed.
>
> At which point, it will be better than human, if its based on vocal input. Note: Eliezer
> told me on our walk back from that burger joint that getting computers to understand
> words was one of the biggest problems in AI. If CYC has it licked, then we must be moving
> along just fine.
Actually, what I said was that Cyc didn't understand the sentences that
were being programmed in, because it didn't understand the words. It
doesn't help to correctly parse the syntax of the sentence "the ball is
red" as "apply(adjective: 'red', noun: 'ball')" if the AI doesn't have a
visual cortex, an abstracted symbol for 'red'ness, and a prototype image
for 'ball'. Cyc conceivably has abstract knowledge about which symbols
tend to be associated with each other, and occasionally, how those symbols
fit into more abstract network-like properties such as causal
relationships. But the symbols are still pretty much devoid of content by
human standards. Cyc has a tremendous amount of syntax but still very
little in the way of semantics, except for those symbols that are
fortunate enough to refer to fundamentally network-like or logic-like
referents. So Cyc *might* understand "before", "after", "cause", or even
"supercategory", but not "ball" or "red".
-- -- -- -- --
Eliezer S. Yudkowsky http://singinst.org/
Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2b30 : Fri Oct 12 2001 - 14:39:44 MDT