> BG> However, including the NLP aspects will make it a LOT harder,
> because the
> BG> app would require pretty sophisticated NLP understanding, or
> it will be
> BG> incredibly annoying to phrase things in ways the system understands.
>
> Yes, the NLP aspect is tricky.  However, in thinking about this I had
> in mind one of the earliest "impressive" pseudo-NLP systems, SHRDLU:
> http://hci.stanford.edu/~winograd/shrdlu/ -- I'm sure you're familiar
> with the demo conversation at the bottom of that page.

The problem is, this kind of system's *best* conversations look nifty to
show off, but their *average* conversations really suck...

> The rough analogy with SHRDLU would be that you're communicating about
> a structured domain in which certain rules apply (paragraphs consist
> of sentences, sentences consist of words and punctuation, ...) and the
> concepts are, although more extensive than SHRDLU's, still extremely
> limited compared to full human-level conversation.  In the end you're
> just manipulating serial text -- you can move, copy, delete, insert,
> and format it...every operation should break down into combinations of
> those at root.

Yes, I understand, and I think you could make it work.  It would involve a
hell of a lot of NLP engineering, of a sort that has next to nothing to do
with AGI.  But this narrow-AI NLP component could be hooked up to an
AGI-based component that learned what actions were desired in what
contexts...

I actually have some experience building NLP systems like this, and I know
how much work it is, and how deep you get into the details of
hand-engineering grammar rules and semantic mapping rules.  It's fun but has
little to do with AGI.

One product that looks useful for doing this sort of work is the NLP++
development environment, available from www.textanalysis.com.  I played
around with it a bit, and it seemed cool for certain kinds of narrow-AI NLP
development.


-- Ben G

-------
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/

Reply via email to