From: Ben Goertzel (email@example.com)
Date: Sat Jul 28 2001 - 15:20:22 MDT
There is no doubt in my mind that an AI, to get really smart really fast,
will have to learn human language so it can learn from us, conversationally
and through reading texts.
There are three ways to approach this:
1) Build a smart system and teach it human language
2) Build a smart system and hope it learns human language on its own
3) Build a smart system and *wire knowledge of human language into it*, in
such a way that this knowledge can be mutated and improved freely by the
system as time goes on
4) Build a smart system and carefully engineer the structure so that
learning human language will be relatively easy, then teach it human
In Webmind Inc., we experimented with approaches 2) and 3), inconclusively.
2 didn't work, but there were known shortcomings in the system's cognition
modules (it wasn't that smart!).
3 didn't work, but we didn't get through more than 30% of the work it would
have taken to make it work. What you need to do is take all the parts of
the NLP pipeline and implement them using completely flexible, learnable
"schema": tokenization, morphological analysis, parsing (using lexicalized
feature structure grammars, context-free grammars, or whatever framework),
semantic mapping, discourse planning, sentence generation, reference
resolution, etc. There's a lot of stuff there, and a lot of knowledge in
the NLP community about how to do it.
Our current plan is approach 4. From all our work on approaches 2 and 3, we
understand pretty well what kinds of cognitive structures are needed to make
language learning work, so we can 'prime' the system for language learning
in various subtle ways, without building in any specific linguistic
knowledge, rules, etc. We believe the human brain comes primed for language
learning in similar ways.
If you're going to study computational linguistics and are interested in
real AI, my advice would be to focus on *unsupervised language learning* and
on *combining syntax and semantics in the course of processing*. These are
cutting-edge topics which are closer to a "real AI" approach to language
than plain vanilla computational linguistics. Most computational
linguistics is based on either hand-coded rules, or supervised learning from
corpuses, and while this stuff is interesting, it doesn't teach you that
much about how to make software that *really understands* human language.
-- Ben G
> -----Original Message-----
> From: firstname.lastname@example.org [mailto:email@example.com]On Behalf
> Of Gordon Worley
> Sent: Saturday, July 28, 2001 3:30 PM
> To: firstname.lastname@example.org
> Subject: Usefullness of computational linguists in reaching the
> This fall, I enter college for my formal education in the ways of
> computer science and linguistics. So, my question is for some of you
> with real life AI experience, how can a computational linguist be
> useful in AI projects? I'd like to know so that I don't waste time
> taking courses that might seem useful to me but end up not being.
> I have thought about this some, so let me throw my ideas out and
> please add to/correct them. First, I know that this kind of role is
> useful in gathering information for the information backing that a
> good seed AI is going to need to get started, but we're already about
> 20 years into doing this and I project that in maybe 10 years or
> sooner pretty much all human knowledge is going to be archived and
> we'll know how to best have an AI interface with that knowledge base.
> Second, I have thought that the AI may just start spurting out to us
> in vis own language that ve develops (this seems unlikely and I don't
> expect it to happen, but there could always be a bug that causes the
> AI to develop communication skills before ve has anything to
> communicate with). In this case, having a linguist on hand would be
> very useful. Similarly, when we have two or more AIs running and
> able to talk to each other, they are going to develop a new language
> that will be more efficient for them (we may not even recognize it as
> a language, but they'll surely develop some means of communication
> more efficient than English, especially since anything they come up
> doesn't need to be spoken, eliminating a lot of the bulk found in
> natural human languages). While the AIs will be very smart, I think
> a linguist may still be needed to, if nothing else, teach AIs how to
> analyze language and develop a means of translation (especially if it
> turns out that AIs are really bad at pattern recognition compared to
> humans and it will take them a few real years to develop algorithms
> to do it properly).
> Anyway, I'm curious what the list might have to say on the subject,
> but beyond a few posts I think it will not be of interest to the list
> in general, so you should probably direct replies to me unless you
> are sure the list will find it interesting in some way (I think
> flames count as interesting, or at least can ;-)).
> Gordon Worley `When I use a word,' Humpty Dumpty
> http://www.rbisland.cx/ said, `it means just what I choose
> email@example.com it to mean--neither more nor less.'
> PGP: 0xBBD3B003 --Lewis Carroll
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT