Dick:
And:
[...]
#>Most linguistic categories, at least at the rarefied end of things where
#>formal linguists work, are ones that can be reduced to a small number of
#>properties. 
### Yes - so much the worse for these analyses. Formal linguists simply
#ignore everything they can't easily formalise. 

Well what I said seems to me to apply to WG too. Both WG and, say, HPSG
can take a sentence and provide a complete specification of its structure
and a complete specification of the grammar involved in building it.

Furthermore, I believe that a formal model of grammar can (and should) be 
integrated unchanged into a larger model of knowledge of language. That is,
knowledge of language can be modelled modularly, with formal grammar as
one module providing output to other modules, such as the stylistic-
sociolinguistic module. By 'modular' I mean the way a computer program
is modular. 

But when I say this, Joe & Jasp (and you?) object that there is something
wrong with this picture, where one module contains the rules for matching
sounds and meanings, and the other contains memory of usage.

#In an earlier message you asked how come formal grammars work so well. One
#answer that I have some sympathy with is that they work so well because
#they define their own goals so that they do work; and that these goals have
#evolved in something like the way that chess has evolved because of the
#social history of grammar which historically has concentrated on:
#a. writing,
#b. standard languages and/or
#c. dead languages.
#Remember in Ancient Greece grammar was invented partly to help 'modern' (4c
#BC) Greeks to understand the archaic language of Homer, and partly to help
#people to write (gramma = letter). If you're learning a dead written
#language you don't have much context to pay attention to, nor do you have
#to worry about the kinds of variation that plays havoc with grammars of
#unwritten living languages. Moreover, like chess (but not football) the
#writing system is totally digital, whereas some parts of spoken language
#are analog. The Greek philosophers were hand in glove with their
#grammarians and helped to encourage this idea of the abstract
#decontextualised system.  If they had been trying to understand a purely
#spoken and unstandardised spoken language they would almost certainly have
#come up with a different explanation. At least, that's the conclusion I've
#come to from my very limited understanding of history; I'm sure Chet has a
#much more informed view.

Right, and Jasp has said something similar.

Now, are you saying that they work well as models of *language in its
entirety* only because they redefine language to mean only those bits
that their models work well for? If so, then we agree on this. But in
my view the only sin of the formalists is the terminological one of
expropriating the term 'language' and redefining it to mean 'formal
grammar'. But the formal grammar itself is impeccable. And the way to
rectify the sin is, as I said above, to add a knowledge of usage module
that takes sentences, matches them to usage and assigns them properties
to do with style, register, interpretation, and so on. In other words, 
formal grammar is correct in itself, but forms only one component of 
language as a whole.

Or, are you saying that formal grammar can be binned in the way that 
the proper fate of elaborate theories of the humours and of the celestial 
spheres, which were formal systems constructed on empirically weak 
foundations by highly rational minds, was to be junked in their entirety? 

I have had the impression that Jasp has been trying to say the latter,
which of course I emphatically disagree with. But if you are only trying
to say the former, then this is not much of a critique of formalism: the
critique amounts only to it failing to account for everything you want it
to account for, a fault that can easily be rectified in the way I have
described.

#>### I don't think I've ever said that WG requires us to abstract away from
#>#(cont)textual things. On the contrary, with my sociolinguistic hat on I
#>#explicitly deny it - e.g. we know that BONNY is used by Scots, and that
#>#ATTEMPT is high style. 
#>
#>You haven't ever said that WG requires us to abstract away from contextual
#>things. But the fact remains that if you try to abstract away from 
#>contextual things the result is successful; the level at which language is
#>highly systematic is a highly 'schematic' level from which contextual
#>things have been abstracted away. 
### I think it depends how you measure success. Sure, when it's talking
#about things that are contextually neutral it is successful; but there's a
#whole lot of other stuff that it tends to ignore because it's too deeply
#embedded in context. Take those Mad Magazine sentences (What, me work in
#the evening?!) or instructions on bottles (Take twice a day.). Or even the
#French passé simple, which is only ever used in formal writing - or English
#pied-piped prepositions (much better with some preps than with others -
#e.g. "for which" vs "except which"). etc etc. 

Okay, so you're not criticizing what a formal model actually does say about
such sentences (-- I assume that what it does say is the grammatical structure
of those sentences), you're criticizing its failure to say anything about their
register. So in fact in principle you could actually agree with every detail
of those formal models' analysis! 

#>And tho yes, of course, we have knowledge like "BONNY is used by Scots",
#>well-formedness constraints don't involve such knowledge and nor are there
#>rules like "Verb X subcategorizes for a word used only by Scots", "Nonfinite
#>verbs must agree in register with their subjects", or whatever.
### Well, there is my favourite example of the verb GO which can (you) or
#must (me) take as complement a non-verbal action: The train went [whistle].

I agree that's a moderately problematic case for a highly abstracted model
of grammar, but it's not the same sort of thing as the pseudorules I gave.

#>### Why not? What kind of information couldn't be modelled explicitly and
#>#formally? I can see the problem of building perceptual knowledge (e.g.
#>#visual images) into a network model, but this comes up with abstract
#>#categories like Bike as well. Anything that we can perceive in day-to-day
#>#life could, in principle, be stored and used as a permanent fact about a
#>#general concept, so I certainly don't WANT to limit WG networks to a
#>#limited range of abstract features.
#>
#>Yes, you can make statements like "bike has 2 wheels", etc., but for each of 
#>us the prototypical bike is the sum or average of our different individual 
#>experiences. I don't see how it is possible to model that prototype with 
#>completeness. 
### Can you model an abstract grammar, including the vocabulary, with
#completeness?

Surely yes. What makes you think one can't? Okay, we can't yet, because we
haven't done enough work on the problem so far, and because we give priority
to making our models better over making them complete. But I don't see any
obstacle other than brain-hours.

#>The task of modelling the prototypical bike is pretty
#>much like the task of modelling a bike, with the added requirement that
#>properties have different degrees of strength/typicality. I don't mean to
#>say that modelling the prototypical bike is beyond the powers of human
#>research, but I do think that the methods required are not those of
#>formal linguistics.
### Precisely - that's what I think we all agree on. The methods of formal
#linguistics do not work in all areas. But I don't think we should use the
#methods of f. l. to define an area of reality: the abstract language (which
#can be studied using those methods) and other things (which can't). 

Why not, though? If FL works so well for abstract language, why not apply
it to abstract language?

Oh I see what you're saying. But FL doesn't define abstract language as
"that which can be studied using FL methods". It defines abstract language
as "the rules that map sounds to meanings; the rules that generate
sentences". It just so happens that when abstract language is so defined,
FL methods work incredibly well on it.

And WG does use FL methods to model grammar! Or so I had thought. All WG
adds is extra contextual info that most formal models don't bother with.
(Tho I can well imagine HPSG bothering with it, if it doesn't already.)
WG doesn't claim to be able to use its methods to adequately model bikes,
ants or grins, but it does claim to be able to adequately model grammatical
categories. No?

--And.