Hmmm. I think there's a problem with your use of the word semantics . .
. . There is a huge difference between labelling an object, which young
children do quite early, and dealing with concepts (even fairly concrete
ones). There is an even larger difference between correlating
co-occurrences, which is all that information retrieval and text
classification systems do, and actually dealing with meaning.
----- Original Message -----
From: "Matt Mahoney" <[EMAIL PROTECTED]>
To: <[email protected]>
Sent: Tuesday, May 01, 2007 4:28 PM
Subject: Re: [agi] rule-based NL system
--- Mark Waser <[EMAIL PROTECTED]> wrote:
> Not really. Semantics is an easier problem.
If so, then why "When you write a compiler, you develop it in this order:
lexical, syntax, semantics."
To point out the difference from the way children learn language: lexical,
semantics, syntax.
This is why the compiler approach doesn't work for natural language. An
artificial language has a fixed and precise grammar and semantics. Such a
language cannot be learnable, because language learning implies that the
language changes over time, which implies that the meanings of sentences
do
also. This implies that the language must be ambiguous.
In order for people to communicate successfully, they must share (almost)
the
same language. This happens because humans learn language in the course
of
normal communication, so that the listener's language evolves toward that
of
the speaker. If this did not happen, then the members of a social group
would
not evolve toward a common language.
Language has a structure that allows both incremental learning and
compensation for ambiguity using redundancy. Most of the meaning of a
message
is captured in single words, and to a lesser extent on groups or patterns
of
words over larger windows. The meanings are additive. In a pure
text-based
system, the meaning of word or phrase is the set of words or phrases that
appeared near it in past communication. In a grounded system, words may
also
be associated with nonverbal patterns by the same learning mechanism.
The developmental learning sequence is over increasingly large windows:
letters, syllables, words, phrases, sentences, paragraphs.
> Information retrieval and text
> classification systems work pretty well by ignoring word order.
Semantics is defined as the study of meaning. Information retrieval and
text classification systems do not understand the meaning of what they
return. They do work . . . . but their job isn't semantics.
Yes it is, except that their semantics are not grounded.
-- Matt Mahoney, [EMAIL PROTECTED]
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
