LINGUIST List 4.1099

Sun 26 Dec 1993

Disc: Internal/external evidence

Editor for this issue: <>

Directory

1. , internal/external evidence

Some time ago, Joe Stemberger noted that the value of psycholinguistic
evidence (and, I would add, other kinds of external ev.) is that it is a
different kind of ev. from that used to come up with whatever the new
ev. is being used to test. I think he's hit the nail on the head, and
I would like to try to drive it all the way in.

But first, an anecdote that will, I trust, make the no-difference-between-
internal-and-external folks unconfortable. At the Linguistic Institute
last summer, Donca Steriade, during a talk at the Phonology Workshop,
made the claim that a word in an African language (don't remember which)
had an (intervocalic) prenasalized stop. An audience member questioned
her claim, suggesting that it might be a sequence of nasal stop + oral
stop. Donca gave the standard arg. -- no (other?) clusters, only open
syllables (I don't remember if the language allowed word-initial
prenasals), but the (now sizable number of) sceptics were not convinced.
[I don't know for sure, but I suspect that their number included no
native speakers of languages with prenasals!] Then I pointed out the
work of Hombert and others on (apparent) syllable-reversing language
games in OTHER (related) lgs. (not even the lg. in question), in which
prenasals don't get split up. THAT WAS THE END OF THE DISCUSSION, and
Donca went on with her talk. So the question is: why did this quiet
the sceptics? Were half the phonologists in the world guilty of failing
to appreciate an elementary methodological point? Or is external ev.
really different? [Incidentally, it appeared to me that there was a
general atmosphere in the audience that the issue had been decisively
resolved. I'd be interested in finding out if others who were there
shared my perception.]

Back to Stemberger's nail. (In fact, I guess I've already taken a whack
or two at it!) It really does help to have new kinds of data.
Philosophers of science frequently suggest this, but then abandon the
issue because it's so hard to make precise the notion "new kind of
data". So let's take another approach, which came to my attention in
Wesley Salmon's "The Foundations of Scientific Inference", and was
pursued by me in my diss. (better hurry -- Garland's clearing it out!)
wrt arguments using external ev. -- a Bayesian approach to inference.
Basically, it says: the probability of a hypothesis given a piece of
evidence increases to the extent that the ev. was unlikely unless the H
was true. How is this relevant to int/ext? Well, suppose I've got an
analysis that says I'm dealing with prenasals, not sequences of two
phonemes, based on Donca-style reasoning. How do I know if I'm right?
Look harder, and see if any clusters have escaped my notice? And what
if I still don't find any? Isn't this new evidence? It is -- but it's
not UNLIKELY even if my analysis is wrong, knowing what we do about the
nature of phonological change. This is always going to be a problem
with phonological analysis -- we can never be sure that a given
generalization is not, synchronically, simply an accident. (Kenstowicz
and Kisseberth are admirably insistent on the necessity of showing that
this "null hypothesis" is not true, in their 1979 textbook.) [Geoff
Pullum has a paper that appeared in the mid70s in York (i think) WPL in
which he gives 2 gloriously accidental linguistic generalizations --
grist for the Topic ... Comment mill, Geoff? I can't remember them,
though, so I'll offer the example of the combined grad/undergrad
morphology class I taught in which all (two) of the undergrads had
surnames beginning with "M". Accidents don't get their just due among
linguists!] So what if I find a native speaker who can "talk backwards"
with some fluency (for an Eng. example involving affricates and
diphthongs, see the paper by Cowan and Leavitt in the 1981 CLS
paravolume), and s/he doesn't reverse the nasal and oral "parts",
despite the fact that every other phoneme sequence is reversed? This
really IS unlikely unless we don't have a sequence of two phonemes in
the prenasals (granted, language change could play a role here, too --
especially if the reversing ability has been taught -- but it's less
likely; and one can always get around that possibility by teaching
people a novel way of reversing (just ask Hombert, or Ohala)). And
that's why new kinds data (better, unlikely ones without the H) are
important. So I hope the nail is all the way in now.

But are int. and ext. ev. fundamentally different? I think they are, at
least if one takes the Chomskyan approach that one of the things that
linguists ought to do is explain language acquisition. For if so,
analytic principles ought to be such that a child could use them to
acquire lg. BASED ON THE KIND OF DATA AVAILABLE TO HIM/HER. This kind
of data is, by and large, internal evidence -- kids don't have access to
things like psy-ling. experiments or future language changes, and they
can't use some of the stuff that they do have access to, like speech
errors or, I would argue, game data. But in order to get the right
principles, we have to know exactly what it is that kids are acquiring,
since we can't inspect their LADs and find the principles that way.
I.e., we have to use data that are external to the child's corpus (one
should keep in mind that the non-elliptical expressions are
CORPUS-internal/external ev.). So in part, I think, we're all right.
Data is data, as far as finding the right principles is concerned, but
the principles have to be such that they work when all we have is
internal evidence.