classifiers and number marking

[MORAVCSIK Edit]

The point that David Gil made yesterday seems to me to be very important.
Taking off from Johanna Nichols' earlier remark, David pointed out that
when testing implicational universals, it is not enough that the
sample be genetically and areally balanced but one must make sure that
the features serving as implicans and implicatum are not themselves
areally (or genetically) clustered.

In other words: if one wants to test a typological implication,
the language sample has to be appropriate to the particular features
whose distributional relationship is to be tested, rather than
simply being a sample of all the languages in the world. This
is because, while implicational universals are "restricted universals"
relative to the set of all human languages, they are
"unrestricted universals" with respect to a proper subset of human
languages - the universe defined by the implicans. It then follows
that the sample of languages relevant for testing has to be a
sample of THAT subuniverse of languages. The requirement that
the sample be areally and genetically balanced then pertains to
this sample.

For an example, let us suppose that we want to see what the
crosslinguistic distribution is of two features - say clicks and
object-verb agreement. Is it OK to do this by looking at a balanced
sample of all the languages of the world? No - such a sample would
be right for testing unrestricted universals only. For the clicks and
object-verb agreement case, we need a genetically and areally
balanced sample of those languages THAT HAVE CLICKS and a
similarly balanced sample of those languaes THAT HAVE V-O AGREEMENT.
If such samples cannot be constructed - as for clicks, being
an areally restricted feature, it would not be possible - this
already indicates that any implicational relationship for the two
features that might emerge will be suspect as a typological claim.

But what this means is that one cannot use the same (areally and
genetically balanced) language sample for the testing of ALL
typological claims. Instead, one needs to construct separate
(balanced) samples for EACH pair of features that figure in the
hypothesized implication! (Ooooh...! Can THIS be right?)

Edith Moravcsik