Unfair to Greenberg

[Larry Trask]

On Tue, 31 Aug 1999 ECOLING at aol.com wrote:

> Concerning the Cambridge use of unrooted trees,
> as compared with Greenberg, both of them draw
> conclusions about relative similarity, hence potentially
> about relative closeness of genetic relationship on a
> probabilitistic basis.

No, not at all.  Greenberg claims absolute relatedness; the Cambridge
group do not.

> Since Greenberg does not reconstruct
> proto-languages, the "roots" of his trees have only the
> weakest of implications, if any at all beyond the usual
> distinctions from a dialect chain or dialect space,
> so that is not an important difference between his
> expression of results and the "unrooted trees" of the
> Cambridge project.

Ask Greenberg if he agrees with this characterization of his work.
I'll bet you he doesn't.

> Greenberg of course is a human being making human
> judgments, therefore not as explicit in what the criteria
> of the judgements are as a computer would be,
> and potentially not as consistent.  Just like any other human
> being doing Multilateral Comparison.

Greenberg says *not one word* about what his criteria are.  Those who
have attempted to scrutinize G's work -- Ringe, Campbell, Nichols --
have been forced to make guesses about what G's criteria might be, since
he himself provides no clues.

This, in my view, is unforgivable.  There is no reason at all why G
should not publish his criteria explicitly -- assuming he has any.

A computer program, of course, does exactly what it's told to do by its
constructor, and it's no better and no worse than its instructions.

> The Cambridge computer algorithms are of course
> applied mechanically, and are therefore completely
> consistent.  In addition they are explicit about what
> their criteria are for decisions.

More precisely, the investigators are explicit about their criteria --
or they should be, as the Cambridge group are.

> That does not mean the Computer algorithms are better.
> Computer algorithms can sometimes be better than an
> individual human for very complex tasks if they can be
> refined over time by many people, and when appropriate,
> conflicting goals can be harmonized or balanced.
> But unless done very very well they can also be inferior to
> human judgements.  Computer algorithms can have biases
> built into them, and more consistently applied biases
> are worse than biases applied less regularly.

But I have never claimed that computer programs are "better" than any
other approach for performing any given task.  That would be absurd,
since it is easy to write a useless program.

All I have done is to point out that the Cambridge group make their
criteria fully explicit, while Greenberg says nothing at all about his
criteria.  This observation has nothing to do with the use or non-use of
computer programs.

> Trask cannot stand the idea that Multilateral Comparison
> done by Greenberg and done by Cambridge has much in
> common.

That's because they have nothing in common that I can think of.
I am moderately familiar with G's work, and I am probably more familiar
with the Cambridge work than are most people.  And I can't see that the
two have anything much in common:

criteria explicit? C yes, G no.
rooted trees? C no, G yes.
proof of absolute relatedness claimed? C no, G yes.

These look to me like pretty huge differences, not like substantial
similarity.

> Trask is making a completely unfair comparison below.

[snip summary of Greenberg and Cambridge]

> The comparison just expressed is not fair by any stretch of the
> imagination.

Why not?  It looks pretty spot-on to me.

> To be fair, Trask could have compared Greenberg with another human
> being doing classification by Multilateral Comparison.

And who might that be?  The only other people using MC are not offering
any competing results to rival G's.  Instead, they buy G's results as
beyond reproach and merely work on other groupings not considered by G.

It would indeed be interesting and illuminating to see two rival
proponents of MC working on much the same languages, but so far that
hasn't happened.

> (We are after all talking about a situation of very distant
> relationships, where are genetic relationships at all, so recurring
> sound correspondences are not likely to be established, and THAT
> sense of "explicit" cannot distinguish Greenberg from another
> human.)

Sorry, but this looks to me like that tired old argument: long-rangers
can't be held to the same standards of evidence as other people.
And I won't buy this.

> Or Trask could have generalized and referred to ANY human
> making judgments, so the burden of the difference would not
> fall selectively on Greenberg.

Er -- what?

I didn't bring up Greenberg, Lloyd -- you did.

> But as stated previously, it is not even
> certain that computer algorithms are in general better than
> human judgments in these matters.

And I've never said anything to the contrary.

Larry Trask
COGS
University of Sussex
Brighton BN1 9QH
UK

larryt at cogs.susx.ac.uk