And Rosta wrote:
> It seems to me that the categorization of both Smudge and BEHIND is
> driven by the criterion of maximizing inheritable generalizations and
> minimizing stipulated overriding.
Yes. I thought that was evident in what I wrote.
> I'm not sure if you're also asking how come the classificatory
> criteria for words are largely distributional.
This is exactly what I am interested in. Or, to put it another way, I am
interested whether this is correct. E.g. can the distributional facts be
attributed to other information?
Most categorisation elsewhere in cognition is gestaltwise. If
distribution is the correct way of classifying words and lexemes, then
in language categorisation isn't gestaltwise.
For example, it is possible to construct an argument such that nouns
prototypically denote objects, verbs prototypically denote events &c.
(For example, any predicative noun has a subject whose referent isa its
sense. Even a relational, event denoting, noun links like this in
predicative structures, and you could therefore advance the argument
that these nouns therefore denote a reification of an event, which is
why there is never any grammatically encoded linking of the participants
in the apparent event they denote.) But if you construct such an
argument, you end up with the situtation where that fact is immaterial
to the classification of LEXEME as a noun. Because it's the distribution
that matters.
So why do I mind? Because if linguistic categorisation is limited to
distributional information, it is different from other kinds of
classification. Word classes are defined by the relations that they are
the potential values of. Other categories are defined by the relations
that they are the potential arguments of (as well as perhaps the
relations they are the potential arguments of, but I bet that these are
less important).
If linguistic categorisation is different from other kinds of
categorisation, because the relationship of the nodes to the relations
is different, then isn't language different from the rest of cognition?
Words, and word classes, can support a whole bunch of properties, but if
distribution is the criterion for categorisation then only a limited set
of those properties is salient in classification. This isn't
best-fit-wise, because it ignores properties (and exactly those
properties which words are the arguments rather than the values of) as
unimportant.
I wouldn't mind, if it were possible to attribute distributional facts
to properties supported by the words, or word classes. E.g., if
distribution were part of a gestalt, it might be said (accurately
enough, I think) that the property of being an argument follows from the
property of being referential, and ergo nouns are arguments (but you
have to set up the semantics right). But in WG, that's not what gets
said. I also wouldn't mind if WG said that language *is* different from
the rest of cognition. But ....
Nik
> My first stab at an answer to that question would be simply that if
> you factored out distributional properties then there'd be no need for
> categorization into word classes.
>
> --And.
>
> [PS We won our local trade union battle, and I am hoping for the
> imminent resurrection of my intellectual life...]
>
>
>
> Nikolas Gisborne, On 28/06/2006 11:17:
>
>> Dear WG,
>>
>> I am having a problem with categorisation. On the one hand, we
>> classify by attributes. We have a classification - call it C - and we
>> know that instances of C fit the classification by best fit, because
>> they have enough matching attributes. So, Smudge was a cat, because
>> Smudge had whiskers, fur, a wagging tail when angry, a purr when
>> happy, retractile claws etc. Whiskers, fur, a wagging tail, a purr, &
>> retractile claws are all attributes of Smudge which fit the category
>> Cat & we expect all instances of Cat to have the same attributes.
>> Because we classify in this way, even Manx cats - which are tailless
>> - belong in the Cat category by best fit.
>>
>> On the other hand, we don't seem to classify in language this way.
>> BEHIND is a preposition, not because it has a spatial and relational
>> meaning (both of which might be prototypical for prepositions) but
>> because it has a preposition's distrubtion. But having a distribution
>> is not an attribute of the the category - it's actually the property
>> of being some other entities attribute. So, instances of BEHIND can
>> be complements of verbs of motion and location. in (1) ...
>>
>> (1) Jan ran behind the bike shed
>>
>> ... _behind_ is the value of "complement-of" &_ran_ is it's argument.
>> Now, being the value of something else's attribute doesn't make that
>> attribute your property - so how come this is relevant to the
>> categorisation of BEHIND?
>>
>> Crucially, we use what-it-can-be-the-value-of as the main diagnostic
>> of word class. Now, given the attribute business above, this makes no
>> sense to me. I am absolutely sure that I am missing something
>> blindingly obvious, but in the usual way of these things I can't see
>> what it is!
>>
>> Nik.
>>
>>
|