LINGUIST List 9.487

Sun Mar 29 1998

Disc: State of Comparative Linguistics

Editor for this issue: Martin Jacobsen <martylinguistlist.org>

Directory

1. MARK HUBEY, Re: 9.485, Disc: State of Comparative Linguistics
2. manaster, Re: 9.485, Disc: State of Comparative Linguistics

Message 1: Re: 9.485, Disc: State of Comparative Linguistics

On Sun, 29 Mar 1998, The LINGUIST List wrote:

> 
> Date: Sat, 28 Mar 1998 14:13:16 +0000
> From: "Robert R. Ratcliffe" <ratclifffs.tufs.ac.jp>
> Subject: Re: 9.453, Disc: State of Comparative Linguistics
>
> What basis, indeed what right, do we have to draw such inferences?
> The best answer, the correct answer, is or should be that we base
> our inferences on a body of case studies of actually observed
> changes, or on more general principles of change derived from this
> body of evidence. That is, given a particular pattern in the data
> (a sound

The basis of drawing such inferences actually less with a body of case
studies than on rational reasoning or at least one particular method
of reasoning, namely probability theory. The only thing we need to do
is to demonstrate that a state of affairs which cannot be due to
chance or randomness or coincidence exists. IF the reason for such a
state of affairs is not chance/coincidence then the options we have
left are that it is due to some other process such as borrowing, or
descent from the same language.


> correspondence, for example), we conclude that it must have come
> about as a result of one particular process, or could only have come
> about as a result of one of two or three possible processes, because
> in all cases in which this pattern is found in languages whose
> history is known, this is the process, or these are the processes by
> which it has developed.

The "method of regular sound correspondences" is the intuitive way of
guaranteeing that the occurrences were not due to chance. If we see
only a single example of a change we can calculate the probability of
its happening due to chance just the same way as we would if we had
multiple occurrences. The reason for greater faith in multiply
occurring sound correspondences is that the probability of such
occurrence is the product of the probabilities. Thus if the
probability of a particular sound change (in a single word) is p, the
probability that 3 such sound changes are due to chance is p^3 (p
cubed) so that if p is something like 0.01 then p^3 is 0.001.

The trick is that none of this tells us what these sound changes are
due to; we simply have evidence that it is not due to
chance. Therefore a few things should be kept in mind. First, it is
not necessary for the sound change to be exhibited more than
once. Naturally, the more the merrier but it is not
necessary. Probability theory works just as well if the sound changes
are not evidenced in multiples. Secondly, we should know very well
that the longer the time span between languages the smaller the number
of sound changes that we will be able to find. Therefore there should
be no reason to insist that someone find 3 examples (or more) of say
10 sound changes. We can just as easily work with 20 sound changes
with each of them being represented by a single word/lexeme.


>The classic fallacy in historical linguistics is the assumption that
>a particular pattern in the data (VSO word order in Celtic and
>Berber, for example) necessarily or plausibly implies something (that
>the languages are related, for example) which it perhaps only
>possibly implies or doesn't imply at all. Unfortunately historical
>linguistics generally fail to follow this procedure, and too often
>draw inferences on the basis of criteria which are not explicit.

IT is difficult to tell why and which idea is correct or incorrect.
If we wanted to use such data one way to do so would be (as above)
calculate the probability of it occurring due to chance. Since only
six word orders are possible, then the odds that a language would have
a particular word order should be 1/6. Therefore the odds that any two
languages selected at random should have the same word order is
1/6. Nobody in any science ever does science by assuming that unlikely
events happen all the time. Two languages selected at random (if every
word word order is equally likely) should not have the same word order
5/6 of the time, so assuming that languages which share the same order
do so because of some reason (other than chance) is quite rational.

The arguments about reasoning along these lines, for example, that
agglutinating languages comprise about 50% of all languages and hence
agglutination is not indicative of any kind of a genetic link, are
fallacious. About 50% of the world's population is considered white
and no geneticist would argue that humans do not descend genetically
from and display the genetic features of their ancestors. Maybe 90% of
life on earth is insect life and nobody would argue that insects are
not genetically related to each other more closely (in some reasonable
geneticity space) than they are to humans or fish.


> It is I think a grave methodological error to treat historical
> linguistics as an extension of descriptive linguistics. There is a
> big difference between describing a body of data and drawing
> inferences on the basis of the probabilistic implications of that
> data. In practical terms, the procedure outlined by Professor Teeter
> would lead us to treat any features shared by the majority of a set
> of related languages as retentions from the proto-language-- whereas
> actual language history shows that there are two other posssible
> explanations for such features-- namely drift and areal diffusion.

What is at stake is the definition of "geneticity". AT present the
ideas being passed of as scientific truth are held together mostly by
equivocation, and rubber bands. There are basic principles of
measurement which have not been taken into account at all in
historical linguistics. The two fundamental characteristics of
measurement instruments are validity (accuracy) and reliability
(precision).

When we make a measurement of a space dimension by using a ruler
nobody ever thinks that anything can go wrong. First the ruler does
not change its length, or we assume so. In reality it does, and
surveyers who have to measure long distances using a tape have to take
the temperature expansion into account. But in principle we are all
quite sure that the ruler does measure length; that is what it's
supposed to do. When we use a thermometer to measure heat, we are less
sure but even then we can all see that the hotter it gets (as felt by
our bodies) the higher the thermometer reading. However although
everyone would know immediately if 100 cm is twice 50 cm, most people
cannot tell if 100C (celsius) is twice 50C. In fact, it turns out that
it is not. That is because the Celsius temperature scale is not an
absolute (ratio) scale. Most people would have no idea if a certain
change in the volume setting (loudness) of the stereo set is twice
another setting (loudness).

In any case, it is inconceivable in any of these measurements that an
instrument can be accurate but not precise or precise and yet not
accurate. This was something that popped up when psychologists started
creating instruments (questionnaires) to measure various [alleged]
attributes of humans. The most [in]famous of these is probably the IQ
test. The same person's scores might vary by about 10% (I am guessing)
depending on the day and mood. Meanwhile even a worse problem crops
up; does the test really measure intelligence?

So in principle we can envision some psychological measure that is
accurate to 1% error but how do we know that it really measures what
it is supposed to be measuring? Does the XYZ hostility test really
measure hostility or aggression? Is there are difference? OR does it
measure aggressiveness?

That is the main problem with the methods of historical linguistics as
it is being practiced. First, an instrument has been designed (a la
Swadesh) which is less formal than any psychological test. Even
supposing that the method can be fixed up so that it generates a
number in some standard range, say 0 to 1 (i.e. in the interval [0,1])
can we say with certainty that this test measures "geneticity"?

Is the instrument accurate and precise? That is where the equivocation
comes in. It is circular. The test creates geneticity and geneticity
is verified by the test. IN what range are the numbers produced by
this test? Suppose the test (the comparative method) was fixed up so
that it produced numbers about 0.8 or 0.9 for the various IE
languages. Does this mean that there are no other ways of knowing
about family relationships?

We know that we can stick a glass tube with a liquid in it to measure
temperature in some range, say from less than freezing to greater than
boiling of water. But can we stick that glass tube into a blue flame
of a blow torch and measure its temperature? No obviously. Does this
mean that (1) the blow torch has not temperature or (2) that it is
impossible to measure the temperature of the blow torch?

The answers to both are obviously "no!". We can measure temperatures
as high as 100 million degrees (which is required for fusion) and also
thousands of degrees for measuring the temperature of boiling metals,
etc.

So it is an act of faith to say that the comparative method is the
only way to measure geneticity when geneticity has been circularly
defined in terms of the comparative method. The first thing to do is
to clearly define a mathematical model of language change so that what
we mean by geneticity is not ambiguous. At the same time we can use
probability theory to make calculations with which we can create a
fuller and more mature theory of language change.

The present method suffers from lots of flaws. The main one can be
understood with the example of "morphing". We have all seen on TV
commercials and sci-fi movies the morphing of objects and faces into
other objects and faces. Suppose we had one of those morping programs
for the PC so we can run a small experiment. We start with the face of
Clinton and then specify the final face to be Mao Tse Dong, and also
Eddie Murphy. We then specify the number of frames in which this
metamorphosis will take place. Now if we look at the any frame, say
the nth frame and compare that to the n-1 frame and n+1 fram we will
see that they will look almost like the nth frame. So if we look at
these frames at random (statically) we can see that from one frame to
the other there is such a small change that we can always say that
this person on the n+1 frame is the same person as the one on the nth
frame. By using induction on this we will then obviously conclude
that Clinton, Eddie Murphy and Mao are the same person. This is the
modern version of the sorites paradox (which comes in various flavors
including the Neurath's boat version). This is what linguists have
been doing all along when they were creating language families. That
is what you will get if you only trace some 100 words accross time and
space. This is the result if you ignore everything about language
except some 100 or 200 specially selected words. This is what you get
if you ignore every aspect of language except some 100 words.

In order to get a grasp we should first clearly define what language
is? Clearly it is more than 100 words. That is why the present state
is held together by equivocation and scotch tape (or was it rubber
bands?).

With this kind of thinking we could demonstrate that an elephant and a
dragon fly are genetically related. It's easy, start with a picture of
an elephant and morph it to a dragon fly!

How is this idea of geneticity of language related to biological
genetics? IN genetics the father's contribution is not called
"borrowing". In genetics, we are not called fish although we are
descended from them. In genetics, the genetic line of a person is not
decided only via the mtDNA (the mother's line).


> As far as classification is concerned, there is a simple
> systematic procedure for demonstrating a genetic relationship among
> languages that will be universally accepted. 1) Given that there
> are some points of similarity between two (or more) languages, it
> has to be shown that these are not due to chance.

Great Idea.


> Proving that something is not due to chance is a mathematical
> problem and it has to be formulated in explicitly mathematical
> terms.

Great idea.

> 2) If a similarity is not due to chance, it is either due to
> historical circumstance, or to universal properties of language
> systems. In order to make a decision at this point we need to know

This makes probabilistic computations more difficult because it is
difficult to know what is due to universals without knowing some kind
of a bound on all "theoretically possible languages". How many
theoretically possible languages are there? Is it 1,000? If that is so
then the very existence of only a few types (isolating, fusional, etc)
itself speaks volumes. How do we go about calculating the number of
possible languages?


> more about language universals. I believe that one reason Altaic
> became a question again is because of the work by Greenberg and
> others on implicational universals. So what were thought of as, say,
> five separate word-order features shared by these languages:-- SOV,
> AdjN, Postpositions, RelN, GN--are now reduced to a single shared
> feature.

It could be the other way. Suppose I create an "index of fusion" (for
example as in Comrie). Then if I create this range from 1 to 100 I now
have 100 different types of languages. Or I can simply use this index
along with all sorts of other indices so that probability
calculations, correlation-regression analysis can be run. But all of
this requires that we start to measure things. That requires what are
disdainfully called "quantitative" approaches. IT is the only way of
the future. PRetty soon all the world's languages' dictionaries will
be available on CDROM. Anyone can write a program to analyze all that
data.

> enough about language universals, we don't know enough about
> contact, and we haven't advanced far enough in establishing the
> mathematical foundations of historical linguistics. And that finally
> may be another


If the definition of geneticity is changed to become more like that of
biological genetics, then "borrowings" would no longer be considered
mere noise but rather the contribution of another language to the
creation of a language. The Latin words are as much a part of the
English language now as the Germanic ones. The Gaelic speakers were
probably responsible for not learning the case system of German and
creating this isolating language. They also dropped some of the front
rounded vowels of Germanic (u-umlaut). That is certainly a genetic
component of English as much as the Germanic and Latin words in it
presently. IF we keep thinking that languages change type the same way
people change underwear or the way bored housewives re-arrange the
furniture, historical linguistics will stay in the same rut that it
has been in most of this century.


> reason why classification has perhaps diminished as a research
> field. For some linguists, at least, these three areas of research--
> universals, contact phenomena, and the mathematicization of
> probability claims-- are all more interesting than the ostensible
> 'goal' of classification.



Mark Hubey
Montclair State University
Probability theory is common sense reduced to calculation.
					Pascal

Message 2: Re: 9.485, Disc: State of Comparative Linguistics

I do not object to everything in Robert R. Ratcliffe's posting. It
certainly is true that there is no formal or even merely rigorous
exposition of the foundations of comparative linguistics--but of
course the same is true of all branches of linguistics and most
sciences (maybe all). It is true that we do not know enough and could
know more--again that is true everywhere, with THIS difference: that
there is a vocal group of scholars (to whom Dr. Ratcliffe may or may
not wish to belong) within linguistics itself who are trying to forbid
or prevent further progress in this field (esp. as regards
classification but apparently also reconstruction) and there is no
such group actively pursuing the demolition of any other branch of
linguistics or any other science I know of.

Moreover, when Dr. Ratcliffe says:

" And that finally may be another reason why classification has
perhaps diminished as a research field. For some linguists, at least,
these three areas of research-- universals, contact phenomena, and the
mathematicization of probability claims-- are all more interesting
than the ostensible 'goal' of classification.

he is presupposing something entirely false, namely, that
classification HAS diminished as a research field. Nothing could be
further from the truth. It is true that those of us who work in this
field face a campaign of disinformation unparalleled since the days of
Lysenko and Marr, but that has not SO FAR prevented work getting done,
and more so than ever before. I do fear that this may not last
because of the gradual but accelerating pace at which the whole
subject of comparative linguistics is being "diminished" in
linguistics departments, esp. in North America.

But fortunately anthropology departments or in my own case computer
science ones have not been nearly so determined to do away with the
subject. So at least for a time we are OK and we may indeed survive.

In addition, speaking as someone who has done extensive work in
universals and mathematical linguistics and borrowings, I am puzzled
by the suggestion that these are more interesting than classification.
But I think Dr. Ratcliffe is being coy here. Esp. his reference to
"the mathematicization of probability claims" seems to me to suggest
that he is advocating the acceptance of the mathematically illiterate
claims of authors who try to abuse probability theory as a way of
closing off progress in comparative linguistics. If so, he should
perhaps consult the published work where these problems were disposed
of some time ago--or find out what probabilists have had to say on the
subject.

In short, GENUINE work on universals, contact phenomena, and
mathematical models of language change is by no means inimical to
further progress in comparative linguistics. Indeed, the opposite is
the case. Those who ignore the results of comparative linguistics
cannot ipso facto do successful work in any of these areas.

Alexis MR