Since the competence model linguists choose to model the languages they
work with with is simply the embodiment of some set of idealisations the
paradigm they as language scientists work (and publish) within is
underpinned by, necessarily some strong and interesting regularities
within and observations about the structures the language people learn
over eight years of their life and use every day expresses are missed.
Take two words at random out of the above paragraph, and the similar
length one above it, and see how easy reconstruction of meaning is.
Note that too much centre-embedding will mess up the statistical
predictive power of an individual word instance (or pair of word
instances), since the words around them will be more dependent on
high-level syntactic structure (*). This would be bad for understanding
noisy text, and probably bad for acquiring language structure from
observing local regularities too.
Steve.
(*) One observation, about verbs and prepositions, for example, is that
far less often then chance to they follow themselves (VB VB; P P) in
sentences. Centre-embedded constructions are almost the only exception
to this exceedingly strong and useful local statistical regularity.
-----------------------------------------------------------------------
Steve Finch, | Phone: 301-548-4093
| Fax: 301-527-4080
Thomson Technical Labs/NLP |
1375, Piccard Drive, Suite 250 | email: sfinch@thomtech.com
Rockville, MD 20850 |
-----------------------------------------------------------------------
______________________________ Reply Separator _________________________________
Subject: Re: center embedding of relative clauses
Author: Mark Johnson <Mark.Johnson@grenoble.rxrc.xerox.com> at INTERNET
Date: 3/12/96 4:23 AM
> don't we have an odd sort of science going on here?
As I am sure Ted Dunning knows, the justification for describing
language in terms of two systems, viz. a ``competence'' grammar (which
generates the sentences Ted claims would never be uttered) and
``performance'' constraints (which determine just which sentences are
uttered, and in which contexts) is the claim that this constitutes the
simplest and ``most satisfying'' overall account of language behaviour.
Such a division seems to be conceptually necessary given a categorical
grammar that generates an infinite number of sentences (most of which
will never be uttered).
But it is not clear that the competence/performance division is
necessary if the grammar is conceived of as probabalistic in nature.
(Of course, one still has to come up with the model!)
Best
Mark Johnson
---------------------------------------------------------------------
Rank Xerox Research Centre Tel: (33) 76 61 50 37
6, chemin de Maupertuis (33) 76 61 50 50
F38240 Meylan
FRANCE Fax: (33) 76 61 50 99
---------------------------------------------------------------------