LINGUIST List 5.93

Thu 27 Jan 1994

Disc: The Last Phonological Rule (author's response)

Editor for this issue: <>


  1. John Goldsmith, Re: 5.83 Book discussion: Goldsmith, The Last Phonological Rule

Message 1: Re: 5.83 Book discussion: Goldsmith, The Last Phonological Rule

Date: Sun, 23 Jan 94 13:20:29 GMRe: 5.83 Book discussion: Goldsmith, The Last Phonological Rule
From: John Goldsmith <>
Subject: Re: 5.83 Book discussion: Goldsmith, The Last Phonological Rule

[Moderator's note: Regarding the opening line below, we
want to reiterate that, however unconventional it may be
for an author to respond to a written review, we do encourage
it on the net. In fact, the opportunity to "talk" to the author
seems to us one of the potential advantages of book discussions
on this medium--it lets all of us ask the author the kind of
questions that friends woulask thif they ran into him or
her in the halls. . . . Of course, we expect that the tone of
our net "conversations" will be equally friendly--but that's
never been a problem on LINGUIST. --Helen & Anthony]

While it's pretty much unheard of for an author or editor to respond to a
printed review, it seems to me that the interactive nature of the Internet
invites this sort of thing, so I'm taking the opportunity to add some
comments to Alex Monaghan's ("AM", hereafter) remarks about The
 Last Phonological Rule (University of Chicago Press, 1993).

Speaking as the editor, I'd like to add the note that there was nothing in
either the intent or in the process leading to the book that justifies calling
these papers "working papers". They were reviewed anonymously by reviewers
unknown to me (but selected by the editor at the University of Chicago Press),
and reviewed both in their original form and in a revised form, as
university presses will often require; they were, of course,
edited by me as well. For the most part, they were papers which
made sense appearing in a volume of this sort rather than in a journal.
My paper is of that sort: its first half is written in a discursive
style that is much more appropriate for a book than for a journal. Most of the
other articles (though each for different reasons) fits better into a book than
into a journal article.

I think AM juxtaposes two things that shouldn't be: here --

=In Chapter 2, Goldsmith fleshes out this approach by giving the fullest
=account to date of his theory of Harmonic Phonology. He proposes three
=levels of "phonological" representation: the morphonemic level (M), the
=phonemic level (W) and the phonetic level (P), together with intra-level
=and cross-level rules. These levels are crucially unordered, so that M
=is as close to W as it is to P, allowing M-P rules. The central thesis
=is stated thus (p.22): "all rules which apply at a particular level have
=the explicit function of moving a representation as far as possible
=toward meeting the phonotactics of that level".
=This approach seems to me to suffer from the same problems as beset
=connectionist approaches, to syntax, i.e. that the mechanisms proposed do
=not bear much relation to our intuitions about how language works, and
=the combination of these simple mechanisms rapidly makes it difficult to
=tell how they are achieving their results, thus obscuring the nature of
=the very phenomena which they are intended to explain.

Two things I'd like to point out.

First, there's an implication that I, at least,
read in this comment -- that the MWP harmonic phonology model that
suggest here is connectionist. It isn't, as I'll explain in a moment. But
then, for better or worse, I do turn around in the final section of thepaper
and actually propose a connectionist model, complete with connection
weights and equations. It's important to keep these two things apart.
I _believe_ that there's a connection between the two models, for
reasons that I go into in my paper, but the model of harmonic phonology
that I discuss at greatest length isn't connectionist -- it's pretty much
symbolic like SPE phonology, or most anything else that's come down
the pike during the 1970s or 1980s.

Second: I think AM's comment is meant to bear on the connectionist
model, not the harmonic phonology model. And I'd like to explain
briefly why I think that the connectionist model that I explore in the
final section of the paper is not as vulnerable to the sort of meta-
theoretical criticisms that have (rightly, it seems to me) been leveled
against a lot of connectionist work on syntax.

The connectionist-style model that I discuss is the model that I
had looked at in a paper several years back on Indonesian
stress and the cycle, and more recently in a paper on
a typology of stress systems, to appear in a volume edited
by Jennifer Cole and Charles Kisseberth from CSLI. I did
a good deal more work on the model in collaboration with

Gary Larson, whose dissertation (available from our department,
by the way) discusses this material in depth; we have a book
in the works for the University of Chicago Press entitled
Cognitive Neurophonology.

It seems to me there arefour basic points to this model:

a. The effects of iterative rule application can profitably be viewed
as the result of local distributed computational devices, rather than
a central processing unit;

b. The effects that have traditionally been attributed to ordered rules
(these effects being precedence relationships between the
generalizations embodied in symbolic phonological rules)
can better be handled by allowing connections to be weighted,
permitting some effects to drown others out;

c. the categorical (i.e., the binary yes/no) effects familiar to us
in traditional phonology can profitably be understood in many
cases as being the result of threshold conditions placed on the
output of continuous-valued computations (rather than discrete computations);

d. Perhaps most important of all: by viewing a grammar as a point
in a finite-dimensional space, the space of connection weights, it
is possible to set up, and to use, a simple learning theory which
actually assigns connection weights purely on the basis of the data
fed into it. This is the Linguists' Stone, in a way -- the device
we've said we've been looking for. I go into some detail in my
chapter explaining what's going on here. And it is of some
real importance, it seems to me, that for any given set of data
that is input to the device, there is a substantial subregion of
the space of grammars, any point in which may be "found" by
the learning algorithm we establish. This suggests, in line with
the suspicions of many linguists, that internal grammars may
vary in greater or lesser ways, and still be determined in a
straightforwardly mathematical way by the input data.

I'd go on at greater length here, but there is, after all, a book
recently published that discusses this point -- The Last
Phonological Rule.
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue