rate of change

[manaster at umich.edu]

On Fri, 29 Jan 1999, Larry Trask wrote:

> *First* we have to agree on what *exactly* it is that we want to try to
> measure.  "Measuring the rate of language change" is every bit as
> diffuse a concept as "measuring the rate of social change".

[snip]

But there is well-known work on specific rates of specific kinds
of change, e.g., the rate at which words are replaced in the
100-item Swadesh list.  We know for certain, even Swadesh
towards the end conceded, that this rate is not the same
for all languages,  There are examples of the rate being
much slower than Swadesh's norm (Icelandic) and ones of it being
much faster (Eastern Greenlandic).  This much is or should
be widely known.

However, it is meaningful to ask whether what I referred
to as the norm is still useful to know about, because
perhaps it ALMOST always holds.

Let me add that there are many areas of science and
math where ALMOST always is a useful concept.  Theoretical
computer science is a good example of such an area.
There is considerable interest in algorithms which
almost always work right (but not quite). Why? Because
there are many problems which cannot be solved (in
practice, at least) by algorithms which always work
right but can by ones which almost always do.  The
small rate of error is acceptable because otherwise
we have nothing.  Practical computer work of course
is another such area, since few if any real-life
systems or applications work right ALL the time.
We would have no computer anything if we insisted on
programs that ALWAYS work.

Anyway, if it were true that Swadesh's norm was
ALMOST always right, with only very occasional
exceptions, then of course we could still use
his methods of glottochronology and lexicostatistics
but the results would always be subject to a small
amount of uncertainty.

I conjecture that this is in fact the
case.

AMR

[ Moderator's comment:
  I was under the impression--given by a supporter of glottochronology, Dyen--
  that G/L dates for the Romance languages, for example, are wildly off when
  compared to the known history.  Given that no testable languages have ever
  agreed with Swadesh's hypothesis, can we really treat this method as "almost
  always right" with regard to those languages we cannot otherwise date?
  --rma ]