Northwest IE attributes

[Sean Crist]

On Thu, 20 Jan 2000, petegray wrote:

> As I understand the debate about NW attributes, it has reduced to an
> either-or which I believe is misleading.   People appear to be arguing
> either that the NW group split off early, in which case satemisation is a
> problem, or that the language groups concerned remained in contact with I-I,
> in which case we cannot say there is a NW group.

> I think, on the contrary, that we can talk of a group even within a
> spectrum.   We can talk of the three colours at the red end of the rainbow
> as a "group" in a meaningful sense (they are the "warm" colours, etc);  we
> can talk of the three "blue" colours at the other end as another group.
> Likewise, I see no problem in talking of a group within a spectrum of
> languages.  Within the Romance spectrum from Portugal to the toe of Italy,
> we could talk meaningfully of a group from Catalan to Piedmont, or a
> Galician-portuguese group, or a Southern Italian group.

> Therefore I think the debate is misguided.   To talk of  a "NW group" does
> not imply that these languages must have split off early from the other IE
> dialects.   The boundaries can also remain fuzzy:   Celtic and Germanic can
> be seen as central to the group, with Baltic and Slavonic on the one hand,
> and Italic on the other, sharing some characteristics but not others.

This is the old debate between "Stammbaumtheorie" and "Wellentheorie"
(tree theory vs. wave theory): should our representations of the relations
between the IE branches look like a tree, or like a a group of overlapping
set boundaries?

Notice that the set of possible trees is a fairly small _proper subset_ of
the set of possible wave models.  You can represent any tree using a wave
model, but not vice versa.  The wave representation of a tree will have
the special property that all of the set boundaries will be nested: you
won't get any overlapping lines.  Wave models in general do allow
overlapping lines; we can define trees as that subset of wave models where
there are no overlapping lines.

As a matter of scientific economy, we should always choose the most
restrictive theory that the data will allow.  Tree representations are
much more restrictive than wave representations; so if the data will allow
us to claim that all language relations are properly represented in this
more restrictive model, that's the claim we should make.  The _empirical_
question is whether the IE languages will allow a tree representation.

So now a further question arises: what does it mean for the data to
"allow" a tree representation?  We've discussed this on the list before,
so I'll just make reference to the page
http://www.cis.upenn.edu/~histling/home.html , where you can click on "A
short course in Phylogeny".

An algorithm was recently developed to get around the NP-completeness
problem of computing the optimal phylogeny over a set of character-based
data.  As we've discussed at length on this list, Ringe, Warnow, and
Taylor made use of this algorithm to compute the family tree for
Indo-European.  There is certainly some noise in the data, and there are a
few problematic areas (notably, the placement of Germanic and of
Albanian), but on the whole, the same tree structure come up run after
run.

If it were the case that the proper model of the relations between the IE
languages were really a fully general wave model, then this isn't the
expected result; what you'd get in that case would be wildly different
trees with every run of the algorithm, with very poor scores each time for
how closely the tree comes to a perfect phylogeny.  This is in fact what
happened when the team tried to compute a phylogeny of the West Germanic
languages: the languages developed in close contact and shared innovations
in ways which can't be captured in a tree.  But it's not what you get with
the IE family in general, and this is very unlikely to be an accident.

So, what I'm saying is that the IE family is such that a tree
representation of the family is possible; and since this is the more
restructive formalism, it is the one we should choose.

  \/ __ __    _\_     --Sean Crist  (kurisuto at unagi.cis.upenn.edu)
 ---  |  |    \ /     http://www.ling.upenn.edu/~kurisuto/
  _| ,| ,|   -----
  _| ,| ,|    [_]
   |  |  |    [_]