LINGUIST List 5.544

Thu 12 May 1994

Disc: Greenberg

Editor for this issue: <>

Directory

1. Ecological Linguistics,Anderson,PRT, Sem change, Greenberg
2. Jacques Guy, More Fun with Greenberg

On the discussion of Greenberg simulations:

It is indeed true that very small variations in assumptions can have absolutely
enormous effects on results of theoretical experiments. And the assumptions
which do this are often unconscious and unknown. Put another way, statistics
is infallible, or even reasonably valid *when pushed to extreme extrapolations*
only in artificially defined contexts with all parameters known.

With reference to two recent comments:

>> I would urge those
>> that are doing statistical studies of Greenberg's techniques to consider
>> various ways to model the approach that you believe he uses. Small variation
>> in how you model the process may have important effects on your conclusions.
>>

>As I have shown in this and the previous postings, large variations in the
>modelall yield the same result: chance resemblances have (vulgarly speaking)
>infinitely greater probabilities of happening than Greenberg claims. "Large"
>variations: from allowing no semantic shifts at all, to allowing roughly as
>many as Greenberg allowed himself.


>if you allow, as Greenberg does, a semantic shift
<reordered and grouped by me- LAnderson>

chew - suck - breast - udder - milk - to milk

suck - swallow - to drink - throat - neck - nape of the neck

>and thereby define a *closed* semantic domain, you are at the
>same time disallowing such semantic shifts as breast-nipple,
>throat-throttle-gag-stench, etc.

Exactly the point is here. Greeberg like ***any*** linguist will at times use
semantic shifts which other linguists would disapprove of. The only solution
for the field is to attempt to accumulate known cases of semantic shifts. For
those who are not aware of this, there is extensive literature on semantic
fields. I have myself published two articles giving maps of semantic space,
partly simply for the knowledge of what kinds of human meanings can change into
what other meanings, and partly precisely to try to objectify controls on such
reconstructions of meaning change.

I operate with the following general guidelines:

*********************************************************************

1. There are no "closed" semantic domains. There is an highly complex web of
semantic interconnections, with highly probable paths of semantic change,
rivers along which change naturally drifts. Such an account must be
empirically founded on paths of change observed in two or more unrelated cases,
not on any abstract or logician's theory of semantic change. Our intuitive
assumed knowledge of what is a reasonable semantic change must always be highly
suspect, and constantly corrected, educated, and extended by experience of the
field as a whole.
 I have written two papers laying out maps of parts of semantic spaces
(noted just below). There is an extensive literature of lexical meaning spaces
(cf. James Matisoff on body part terms in Southeast Asia for one highly
integrative work). Such maps of spaces need to be factored into any
theoretical extrapolation of statistics to estimate probabilities of meaning
shifts. The process cannot be done as a general mathematical formula. It must
be done based on enormous quantities of tabular data, the kind which is
absorbed by any good practicing historical linguist, but with the leavening
that they must also have the personality to recognize that their particular
experience with the languages they have worked on cannot be a restrictive guide
to what is probable in languages of the world.


"The "Perfect" as a universal and as a language-particular category"
IN: Tense Aspect between semantics and pragmatics, ed. Paul Hopper, John
Benjamins, pp.227-264 (includes general maps of tense and aspect spaces as well
as of categories related to the Perfect) 1982

"Evidentials, paths of change, and mental maps: typologically regular
asymmetries" In Evidentiality, the Linguistic Coding of Epistemology. eds.
Wally Chafe and Johanna Nichols, Ablex Publishing Corp. 1982


Compare phonological changes which are "surprising" for a parallel, such as the
Japanese glottal stop to /w/, or typologically recurring asymmetries, such as
the relative weaker attestations of the opposite corners (p) and (g) in the
theoretically neat pattern
 p t k
 b d g.
Theoretical patterns are simply not relevant here, neither in phonology nor in
semantics, as the basis for extrapolating to extremes.


2.There are also idiosyncratic examples where oddities of culture history
produced unique results (note I say "unique", with empirical reference, rather
than "unpredictable", with reference to an abstract theory.)


3. And there is the very strong possibility that a less dense packing of words
into a given semantic space can have an effect on how meaning shifts occur,
that is, change may have been more common in the past which are less common
now. Once again, a limit on our ability to theorize or extrapolate.


4. When attempting to extrapolate to greater time depths, we can use our
knowledge of observed semantic changes in useful ways. We cannot develop an
airtight theory.
 It is in principle possible to assign quantified values to a given
hypothesized change from meaning M to meaning N, depending on how many times
this change has been observed, and how quickly it occurred (that is, when such
information is available, how quickly after a word arrived at meaning M from
its previous meaning L did it shift to its next meaning N).


5. Given all of the above, it is not possible to produce axiomatic theoretical
proofs of anything, pro or con Greenberg's or any similar hypotheses. Most
attempts to provide even plausible estimates of probability either way also
founder hopelessly on oversimple assumptions and models. What we can do, and
that is a point which again and again is not appreciated, is get some estimates
of which language connections seem more or less probable relative to each
other, not relative to true or false. All of this with the caveats that
certain kinds of sound or meaning changes may cause the brains of our linguist
analysts to see or to fail to see resemblances in either sound or meaning
either more or less easily, distorting the results we get. The process of
reporting and analyzing the data is very fragile, susceptible to our particular
personal experiences or lack thereof.

**********************************************************************
So why don't people get to dealing with the specifics, improving Greenberg's
data and seeing how that affects his results, improving estimates of sound
change and meaning change, and seeing how that affects his results, etc. etc.?

This is, you may note, exactly the same criticism made by critics of Greenberg
who say one should only work within known language families or by already known
"methods" (notice they do not say techniques, because that word is positive,
"methods" is morally critical). Here this criticism is directed at those who
want a theoretical disproof of his works.

My answer is that people do not do this for the same reason in both cases, it
is long hard work, and no one person can do all of it, or satisfy all of the
requirements which another may attempt to place upon their work, especially
another investigator with a very different personality and approach to the
material.

We will make progress on it, gradually pushing deeper. That progress will be
due to both the rigid traditionalists, working within their own narrowly
defined fields where they can believe they are more secure in their answers
will try new things, often unaware of how great the risks are of being flat
wrong. It is not that one group cares too much about making mistakes, the
other too little. It is that both approaches have always been necessary in all
fields of discovery. What ***would*** be nice is greater respect of each side
for the other, and a cessation of theoretical attempts to prove anyone right or
wrong when the hidden unkown assumptions behind the models are what determines
the answer in each case, and the assumptions are mostly wrong when pushed
beyond their extremes.

Lloyd Anderson

Have you noticed that Greenberg allows for metathesis? Scientific
American, Nov. 92, p.64: Kutenai /u'mqolh/ "to swallow", Faai
/mekeli/ "nape of neck", Kaliana /imukulali/ "throat", Mixe
/amu'ul/ all evidence for *malq'a.

Metathesis between the second and third consonants.

This doubles the probability of chance resemblances,
and does far, far more than doubling the number of
expected spurious cases of evidence for reconstruction.

Meanwhile, the documentation of the simulation algorithms
mimicking semantic shifts is turning into a full-fledged
paper. Tell you what you can do in the meantime. The
Pascal program I posted does not allow for semantic
shifts. But there is no question of semantic shifts
when using grammatical (eg. SVO order) or phonological
(tones vs no tones) features to classify languages.
So simulate 20 languages each represented by 12
features ("words") with a fudge factor of course
of zero. If you consider SVO order the probability
of chance resemblance is one in 6 (SVO, SOV, OSV,
VSO, VOS, OVS), tone, one in 2 (tones or not).
I know, they are not equally probable, but the assumption
of equiprobability does in fact result *always* in an understimate,
often drastic, of the true probability of accidental
resemblance (Greenberg got that bit right).

So all in all, a probability of one in 5 should be about
fair, shouldn't it? Try and see. Now try the same
but with 100 languages this time. See? Now think about
how many Jane Nichols used. Three hundred I seem to
remember.