Multilateral comparison

[ECOLING at aol.com]

This message is devoted to clarifying what
multilateral comparison is, how it has been successfully
used as a heuristic in the past, and its role relative
to proofs of genetic relationship.
Both the method and its role are often misunderstood
or misrepresented.

Sections:

1.  Draft consensus statement on the value of Multilateral
     Comparison as heuristic.
2.  How this can contribute to civility in our field.

3.  Examples of successful uses of Multilateral Comparison --
          some validated cases and as-yet-pending cases.
4.  Greenberg's view that M.C. is sufficient proof --
           his extrapolation beyond the validated cases.
5.  Classification BOTH as a precursor to the Comparative Method
          AND as a result of the Comparative Method!
6.  How to find a null hypothesis opposite of M.C. results?
          Other ways of testing the results of Multilateral Comparison.
7.  Resistance of Multilateral Comparison to "noise" in the data,
          where bilateral comparison (the Compartive Method)
          is not resistant to as great a degree.
8.  Importance of refining judgements of "similar" or "closely
          related", beyond current intuitive or superficial definitions.
9.  So what has Greenberg actually done?
          Something very similar to the Cambridge team.

***

1.  Draft consensus statement on the value of Multilateral
     Comparison as heuristic.

Here is a summary statement which I think can be supported
by many without reservation, at least if the terms are understood
with the meanings clear from the following context.
Any suggestions for improvements in such a statement
will be most welcome, so that it is even more balanced and
moderate. Those of us who do NOT wish to be assaulted
by the obviously political battles can do something about it.
Expressions of simple agreement with this,
from moderates in these debates, might also be useful.
(Sometimes moderates have to take a stand to constrain extremists.)

? Consensus statement ?

Multilateral Comparison is a heuristic which has been successfully used
many times in the past and may be used in the future to throw up
promising hypotheses to be further investigated also by other methods,
especially the Comparative Method which can handle the data
with more detail and precision.

There has as yet been insufficient empirical work to estimate to
what depths Multilateral Comparison can be useful.
There can also be more empirical work to estimate what
refinements or extensions of the Comparative Method
may be able to penetrate greater depths.

***

2.  How this can contribute to civility in our field.

>From an active supporter of the "conservative"
tradition of comparative reconstruction,
whose contributions I value in that regard,
I recently received the following statement.
I have not received permission to quote the source,
so must leave it unattributed:

>...I myself find it [hard] to muster any enthusiasm for [multilateral]
>comparison, except perhaps as a heuristic for throwing up promising
>hypotheses to be further investigated by other methods.

(The words replaced in brackets were in the original "harder"
and "mass".  Since "mass comparison" and
"wide-rangers" and other such terms have a pejorative cast,
I am using more neutral terminology.)

The "except" in the quoted statement is crucial,
assuming as is normal in science
that heuristic discovery procedures are VALUED
(and of course are distinguished from proofs).

I think we would have a very civil field of historical and comparative
linguistics, at least on these issues, if we could reach some greater
consensus.
Those who practice comparative reconstruction and for
that or any other reason dislike Greenberg's work should be
more willing to say such things more publicly, preferably in a simple
positive form something like the statement near the beginning of
this message.
Greenberg and advocates should be more willing to grant
that the conclusions from only this one method
are not as solid as conclusions using also other methods.

***

3.  Examples of successful uses of Multilateral Comparison --
          some validated cases and borderline cases.

I think there are a LARGE number of us who accept
Greenberg's methods as a heuristic, and as a heuristic only.
I am certainly one, and have always been.
I have always been agnostic about claims to "proof" using
only multilateral comparison, simply because it has not been
sufficiently tested.

Here are some examples where it has been successfully used:

A.  The great collection of Eurasian vocabularies done under
Catherine the Great of Russia in the 1700's
arranged those vocabularies
to place the most similar ones adjacent to each other.
My understanding is that the classification so established
has remained valid, has not been superseded by
more recent family trees, trees "proven" valid
by the Comparative Method or in other way.
This should then be counted as a successful use of multilateral
comparison.

B.  Greenberg's work on African language classification
years ago can probably be seen in the same light.
Although there were precursors
(which we need neither minimize nor exaggerate),
and he created new larger classifications
(which we need neither minimize nor exaggerate),
my understanding from debates pro and con some years ago
was that his classifications have remained valid.
(Same caveats as above for number 1.)
This also should be counted as a successful use of multilateral
comparison.

There are obviously many other cases like these.
Pioneers in Polynesian or Austronesian and in many other
language families have used Multilateral Comparison,
implicitly or explicitly.  They have ALSO at various stages
used the Comparative Method or its precursor fragments.

C.  In his Language in the Americas, Greenberg gave a table of
basic vocabulary from languages of Europe, showing that the
method of Multilateral Comparison correctly classifies those
into major families, where those are known via the Comparative
Method.  This should be counted as a successful illustration
of Multilateral Comparison.
(Question: did it include the difficult case of Albanian?
Albanian-with-other-IndoEuropean is a fruitful field for testing
both Multilateral Comparison and the Comparative Method.)

D.  In his Language in the Americas, Multilateral Comparison
clearly singles out Eskimo-Aleut and Na-Dene (Athabaskan etc.)
as two families, and does NOT segment the rest in the same way.
To the extent of those two families, Mutlilateral Comparison is
at least congruent with knowledge resulting from use of the
Comparative Method.
(Does anyhone think Greenberg fudged his results to get this,
or that it would not result from Multilateral Comparison alone?)

What about the rest?

E.  What is additional in Greenberg's Language in the Americas
is particularly the following two points.
These are not yet confirmed or disconfirmed, to the level of "proof",
so we cannot yet estimate whether Multilateral Comparison has
been valid in this extended application at presumably greater
time depths.

E-1.  Neither Eskimo-Aleut nor Na-Dene are grouped by
Multilateral Comparison with other languages or families of the
Americas.  This is a negative statement.

E-2.  Multilateral Comparison does not lead to the conclusion
that there are other cleavages in the Americas between families
as divergent as Eskimo-Aleut and Na-Dene are from each other
or from any other language families of the Americas.
This also is a negative statement.

What are possible conclusions we might draw, if we for the sake
of argument grant that the results of Multilateral Comparison
so far may be valid for case E as well as for cases A-B-C-D?

It may be that the remaining Amerind, or much of it,
formed a family at a much greater time depth;
or that there has been sufficient borrowing back and forth
in many directions to blur any deep cleavages
which may once have existed,
or ???)

Notice the above statements do NOT conclude either "related"
or "not related" concerning the remaining Amerind.  I have
tried to word them very precisely to reflect exactly what
Multilateral Comparison CAN contribute AS A HEURISTIC.
As a heuristic, what it suggests is that we look for additional
deep genetic relationships, which we may be able eventually
do prove by refinements of Comparative Method,
WITHIN remaining Amerind, but not between parts of
Amerind and Na-Dene or Eskimo-Aleut.

This heuristic conclusion does not seem highly improbable
to me.  It may or may not be true, but that is quite another matter,
a matter for the future.

***

4.  Greenberg's view that M.C. is sufficient proof --
           his extrapolation beyond the validated cases.

Greenberg obviously believes that his methods establish firm knowledge.
This appears to be based on an extrapolation from the clear cases like
3A-B-C-D.  Whether such an extrapolation is valid remains to be tested,
since presumably Multilateral Comparison has not been used on data
sets with such potentially gigantic time depths as Amerind.

***

5.  Classification BOTH as a precursor to the Comparative Method
          AND as a result of the Comparative Method!

Greenberg believes Classification is a PRECURSOR to the
Comparative Method, which it may very well be in terms of heuristics
and the history of how useful hypotheses have been discovered.
Look at the History of Science to study this question,
not at the Philosophy of Science.

However, Classification is a RESULT of the Comparative Method
when viewed from the point of view of establishing secure knowledge,
"proofs" of relationship.  Look at the Philsophy of Science to study
notions of "proof".

Much of the battles over Multilateral Comparison and the
Comparative Method seem really to be turf battles, practitioners
of one method who feel sleighted by or who sleight the practitioners
of another method.  If we recognize that the same word can have
two very different logical statuses from different turf perspectives,
perhaps we can help to avoid such battles or at least to shorten them,
to tell the combatants to lay off, we need both, and they themselves
USE both even if they don't like to admit it (in both directions, here).

***

6.  How to find a null hypothesis opposite of M.C. results?

The results of Multilateral Comparison are a statement of roughly
the following form:
"The following Classification tree of given individuals (languages)
appears to reflect their likely historical close relations vs.
divergencies better than any other Classification tree."
And in cases where an algorithm is used, there may be a measure
of the security of the conclusion, based on how much better this
particular result is than alternative Classification trees would be
for the same data.

That result is a hypothesis about the real world.
WHAT is the null hypothesis corresponding to such a hypothesis?
Presumably that some (other) algorithm which we also take to be a valid
application of classification principles would NOT replicate the
same results?  Or that another human judge applying the same method
would not replicate the same results.  (But see below.)

We all know that this is indeed a hypothesis,
Yet if we are responding to it carefully,
we have trouble, if the specifics are good, trying to falsify it.
Why?

One reason is that the hypothesis is entirely post-hoc,
there is (at the stage it is formulated) no additional evidence
available for use, all the evidence used to develop the hypothesis
is the same evidence which would at that stage be used to test it.
So if we used a computer algorithm to evaluate, the result would
be the same again.
Yet it IS an empirical hypothesis.
If the human judge or the algorithm is a good one,
we are not likely to disagree with the conclusions,
though we may sometimes think there is not enough convincing data.

The hypothesis can be tested by enlarging the data set,
to see whether the results are still the same, or by eliminating
certain languages from the data set, to see whether the results
are robust against such changes for the languages which remain
in the set, or by using slightly different algorithms,
to see whether they yield the same results,
or by applying the Comparative Method if possible,
to see what it yields.
Are there other radically different ways of testing it?

***

7.  Resistance of Multilateral Comparison to "noise" in the data,
          where bilateral comparison is not resistant to as great a degree.

This is often not understood.

Because Multilateral Comparison is IN PRINCIPLE only concerned
with which are CLOSER vs. MORE DIVERGENT, a matter of degree,
Multilateral Comparison with informal judgements of similarity
does not, except in the clearest cases,
lead to secure conclusions about genetic relationships per se,
neither yes nor no.

But a strength of Multilateral Comparison is that random noise in
the data will, if it is truly random, not usually affect the results in
terms of rankings of DEGREES of closeness vs. divergence.
Random noise will eliminate similarities in more or less
EQUAL PROPORTIONS from each potential comparison of closeness.

(If the amount of evidence for closeness
is reduced to near the quantum level of single units of evidence,
then the amount of noise is too great for the method,
noise would completely swamp any conceivable residue of
earlier relationships.  But that is quite another matter.)

By contrast, bilateral comparison (many applications of
the Comparative Method) depends more essentially
on the absolute quantity of evidence available.
For this reason, practitioners of the Comparative Method maintain
that it cannot penetrate beyond some limit (which they may give
as 2000 or 4000 or even 10,000 years, rarely more,
rarely even that much), because the residue of evidence for
genetic relation decays fast enough to leave too few traces.
Multilateral comparison can work with RELATIVELY FEWER
traces of earlier relationship, though just how many fewer is a
matter for empirical study.

A similar potential stabilizing factor in Multilateral Comparison is that it
may easily lead to recognition of loan words from language family Z
which could interfere with comparisons between languages X and Y
both of whom borrowed the loan word from Z.  A specialist
comparing only X and Y might not be aware of the word in family Z.

***

8.  Importance of refining judgements of "similar" or "closely
          related", beyond current intuitive or superficial definitions.

Increases in the power of both Multilateral Comparison and the
Comparative Method will come as we build an empirical basis,
like the Handbook of Physical and Chemical Constants,
for explicit conscious judgments of "similarity", or rather, changes,
rather than the current informal judgments or systematic
computerization of informal judgments:

what sounds more often change proximally into what other
sounds, in languages of which kind of phonetic structure,
and
what meanings more often change proximally into what other
meanings, in languages of which kind of semantic structure?

***

9.  So what has Greenberg actually done?

Well, the preceding has been my attempt at a synthesis.

I have never believed that Greenberg had "proven" his conclusions.
I have always believed that Multilateral Comparison was a valuable
heuristic, from the moment I understood the concept.
I see no reason for any turf specialists to trash others whose
contributions can be made use of.

What Greenberg has DONE with Multilaterial Comparison
is not necessarily what he says he has done.
What he has done is to organize a large body of data to
suggest paths for future research, by suggesting relative
degrees of similarity vs. divergence among languages and
language families.  In this regard, if we discount his excess
claims to have proven his conclusions, rather than to have
used a heuristic to suggest them, it is very much like
what the Cambridge group has been doing, as described by
Larry Trask, which is also a useful heuristic, for the same
reasons.

Greenberg's work, like anyone else's,
should be used for its positive value,
with preliminary comments as needed to specify
that his statements of "proof" are premature since the tool has
not been validated at such presumably great time depths,

but that any such unsubstantiated claims
for the method of Multilateral Comparison
in no way demonstrate it is not a valuable heuristic.

Best wishes,
Lloyd Anderson