--- Vladimir Nesov <[EMAIL PROTECTED]> wrote:

> Adult can do tricks not accessible to a child, increasing efficiency
> of language learning, and this process can make use of completely
> different information. Explicit learning of difference between
> categories is faster than unsupervised learning of such distinction.
> Syntax can be regarded as such caricature of natural language that
> helps in forming new categories relevant for it. Sometimes, categories
> required for obtaining expertise in certain domain may never form
> without supervised learning (it's hard to invent something, but easy
> to learn). For example[*], Japanese speakers can't distinguish 'R' and
> 'L' sounds, as in Japanese these are interchangeable and pronounced
> closer to each other. The best way to teach the distinction is by
> exaggerating both sounds, so that categories can be formed, which then
> allows differentiation between normal sounds.
> 
> 
> * JL Mcclelland, 2006. How Far Can You Go with Hebbian Learning, and
> When Does it Lead you Astray?

Like English speakers learning Hindu cannot learn to speak the 3 different
versions of the 'k' sound because they sound the same.  Or the Spanish speaker
who pronounces "sit" as "seat" because there is no equivalent vowel sound in
Spanish and both words sound the same.  The part of the brain responsible for
auditory phoneme recognition becomes read-only by age 6.  So we all speak
foreign languages learned at later ages with an accent.

People do not learn grammar by being given grammatical rules, because we still
don't know what they are.  Grammar rules seem to have a Zipf distribution,
like vocabulary.  About 200 words account for half of the tokens in text, and
then it gets complicated.  Likewise, a small number of rules cover a lot of
the cases, like (S ::= NP VP, NP := det adj noun, etc).  But after this point,
you run into huge number (nobody know exactly how many) of idioms and special
cases like "what/why/how in the world?" that makes language immensely complex.
 By the time you are old enough to learn the concept of grammar, you have
already learned more of the rules than anyone has ever been able to write
down.

Understanding how children learn language helps one understand why rule-based
language models like parsers have been a failure, and statistical methods have
been a relative success (as measured by data compression, and applications to
speech recognition and OCR).  Language has evolved to be learnable by the
human brain.  If a construct isn't learnable, it disappears.  Children learn
by example and usage, not explicit rules.  Each level is learnable from the
earlier levels, namely phonemes shortly after birth, word segmentation at 7-10
months, semantics beginning at 12 months, and grammar at 2-3 years.


-- Matt Mahoney, [EMAIL PROTECTED]

-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription: 
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com

Reply via email to