Mike Maxwell: taxonomic phonemics (reply to Dan Everett)

[Martha McGinnis]

Having raised this issue (or at least having responded to Dan Everett's
mention of it in this list), I guess I should say something about it.
Actually, when there was no response to my original query (concerning a more
suitable forum) for several days, I decided it must not be of general
interest to the members of this mailing list, and off-lined a response to
Dan Everett.

But perhaps it is of interest here after all, so let me try to summarize
what I said to Dan, and (as best I can) his response to me.

Martha requested that the term "taxonomic phonemes" be defined.  I trust the
following will at least clarify, if not define it.  As to the term itself,
as Dan mentioned, it was a perjorative term--the American Structuralists
themselves would just have said "phonemes."  The problem is that the word
"phoneme" has been used in different ways, and in fact that is the core of
my query: I think that the Lexical Phonologists' use of "phoneme" (and for
that matter, the European Structuralists' use) is different enough from the
American Structuralists' use that they are really different concepts.

In the most general sense, quoting Dan's msg (of 11/21), a phoneme

>is a set of sounds, not features, which the native
>speaker hears as a single sound

It's when you try to pin things down a bit more that the trouble starts.

The leading idea among the structuralists was that phonemes were defined in
terms of contrast and complementary distribution (and free variation).  (I'm
going to assume that these terms don't need definition here, but feel free
to tell me I'm wrong.)  So in English, /p/ and /b/ are phonemes because they
are in contrast, whereas [ph] (the 'h' is superscripted, but you can't see
it on your machine :-)) and [p] are in complementry distribution (CD).

The difficulty with the definition of 'phoneme' is in the situation where a
contrast is neutralized (i.e. two phones are in contrast in one environ, but
in CD in another). The difficulty is particularly severe where the
neutralization can be observed in alternations in the morphology.  (The
contrast between /p/ and /b/ happens to be neutralized in certain
environments, where we get only [p].  But that wasn't perceived as a problem
for the theory, because there are no alternations: the environment only
occurs root-internal.)

An example of the neutralization (from Morris Halle) is Russian voicing.
Given that there is a contrast in Russian between /t/ and /d/, but not
between [c] and [j] (the 'c' and 'j' here are alveopalatal affricates), the
voicing of [c] to [j] before a voiced obstruent must be analyzed under
Taxonomic Phonemics as an allophonic rule, while the voicing of /t/ to /d/
must be analyzed as a morphophonemic rule.  Similar examples abound, of
course.

There are several approaches one could take to counter Halle's argument:

(1) Argue that there really is a difference in the two rules, with
native speakers being aware of the morphophonemic one (/t/ --> /d/), but
unaware of the allophonic one (/c/ --> [j]).  (One might look at whether the
degree of voicing is the same in both cases.)

(2) Give up the idea that allophonic rules can only choose an allophone of a
phoneme, and that they cannot change one phoneme into another.  Thus there
would be an allophonic rule (or rules, depending on how you count them)
/t/ --> [d] and /c/ --> [j].  But that would have been striking at the very
center of Taxonomic Phonemics, namely the idea of contrast vs. complementary
distribution.

(3) A variant of both these ideas would be to say that there are two voicing
processes, one of which is morphophonemic (and applies to change /t/ to
/d/, but does not affect /c/), the other of which is 'allophonic'--but
applies to convert 't' to 'd' AND 'c' to 'j' (you'll notice I've left off
the brackets).  The former applies in some situations (such as
word-internally), while the latter applies in other situations
(word-externally, perhaps).  This of course underlies the LP notion of
lexical vs. post-lexical rules.  (See e.g. Zsiga "An acoustic and
electropalatographic study of lexical and postlexical palatalization in
American English" p. 282-302 in _Phonology and Phonetic Evidence: Papers in
Laboratory Phonology IV_, edited by Bruce Connell and Amalia Arvaniti.


I think the structuralists, to the extent that they actually tried to
respond to this argument on empirical grounds, took approach (1).  Dan tells
me that Pike taught that phonemes could share allophones, which I take to be
(2), more or less.  (Pike was not your average structuralist--the American
Structuralists saw that, but the generativists either didn't, or ignored
it.)

All of this to say that the terms 'phoneme', 'allophone', and 'allophonic
rule' don't mean the same in LP as they did in American Structuralist
theories.  In particular, the notion that phonemes are defined in terms of
contrast and CD doesn't work in LP: in neutralizing environments, an
allophone can belong to more than one phoneme, a no-no for the American
Structuralists.  (There are also differences that have to do with the use of
features in place of atomic phonemes, but I think that's a fairly minor
difference, in comparison.  Your milage may differ, of course.)

There might, however, be an 'out'.  Suppose you could show that there was
some difference between the phones of the supposedly neutralized phonemes.
That is, suppose that neutralization was never complete, at least in the
sense of types (as opposed to tokens).  For example, maybe there was a
statistically detectable difference between the [d] allophone of /d/ and the
[d] allophone of /t/.  The difference might be that sometimes the voicing of
/t/ was incomplete, particularly if you asked native speakers to pronounce
the words carefully.  (For a similar idea with a different phonological
process, see e.g. Johnson, Flemming and Wright "The Hyperspace Effect:
Phonetic Targets are Hyperarticulated" _Language_ 69: 505-528.  The notion
is that what shows up acoustically may not always be what the speaker really
intends to say--mental phones are different from acoustic ones.)  Or the
difference might be detectable only by its effects on adjacent sounds, e.g.
by (statistical) length differences on the preceding vowels.  (Pike and
others called this 'displaced contrast'.)  Of course, to be convincing you'd
have to show this for for _every_ neutralizing environment in _every_
language (or at least an interesting sample of them!).  In that case, you
might be justified in saying that while one might transcribe both /t/ and
/d/ in the neutralizing environment with [d], in fact there are differences
in the two kinds of [d]s, at least on a statistical basis.  The result would
be that neutralization by allophonic rules (unlike morphophonemic rules) is
never complete.

Well, this has gotten to be a long-winded response.  I'll appreciate any
responses you might have--

      Mike Maxwell
      Linguistic Data Consortium
      maxwell at ldc.upenn.edu