empirical basis, typology, and gradience

[Luis Casillas]

On Wed, May 02, 2001 at 12:49:17AM -0500, Raúl Aranovich wrote:

> What I had in mind as the subject of my comment is something like
> this: let's say someone introduces a new feature in an analysis of a
> particular phenomenon in a particular language. For the sake of
> argument, let's say that an attribute CLTS is proposed, the value of
> which is a list of pronominal NPs, and that this feature makes an
> analysis of clitic placement in Romanian work, and accounts for the
> facts. That's great, but it could be taken one more step ahead: Is
> this feature universal? Do all verbs have a feature CLTS, which in
> some languages can only have the empty list as its value, or is this
> feature only appropriate for Romanian verbs? How about other
> languages that have clitics?  Does the same feature exist in the
> descriptions of those languages' verbs? How much variation in the
> appropriateness of features and values does the theory allow? These
> are the questions that for me need to be asked, and this can be done
> without sacrificing empirical rigor.

I apologize in advance for the lentgh of this email.

These are very interesting questions, IMHO, because they run straight
into the heart of something which I've never seen a very precise
discussion about (though I'd be happy if somebody pointed out that
this is due to my ignorance and general newbieness more than anything
else).

HPSG grammars are sets of statements, which constrain certain classes
of models.  This much is made explicit by the formalization of the
framework.  But the framework doesn't really tell you what the models
are models for.  The idea is that the objects described by the theory
correspond to "linguistic entities" or "the empirical domain of
language" (P&S 94), but the nature of these entities is, as far as I
can see (maybe not far enough?), left fairly open.  (Which, in my
opinion, is not a bad choice at all-- it decouples the framework from
the positions one may take on this problem).

Two quick (and hopefully coherent) ideas I have as to how one could
interpret these models:

* A psychological interpretation: the models for our theories are
  intended to correspond in a fairly transparent way to features of the
  actual human language faculty.

* An "informational" interpretation: here, the objects described by
  the theory are required merely to register in some way the
  information the signs bear.  The features of these informational
  objects must stand in some relation to the actual language faculty,
  since people, as is generally known, are indeed attuned to the
  informational content of the signs in their language.  However, the
  relation between these objects and the language faculty need not be
  in any way transparent.

This distinction is not an absolute one in any sense; presumably, all
of the theories which make all the right predictions about English
would be informationally equivalent, but not all of these would be The
Ideal Psychologically Interpretable Theory of English Grammar (a.k.a.
"TIPITEG").  And even among the unwashed masses of unTIPITEG theories,
some would be in some sense "more TIPITEGish" than others.

Now, to your feature CLTS.  Let's assume that the feature must somehow
be there in TIPITEG; for instance, let's imagine that tomorrow, using
a newly-invented technique, the CART-scan ("CART" as in "Cartesian"),
not only we uncovered uncontroversial evidence of the existence of the
language organ, but also, we observed, of all things, that it has a
nook in which fit precisely the values of your feature CLTS.

But, on the other hand, let's also assume that for determining which
things are signs of English and which are not, the state of the nook
is completely irrelevant.  It could be at some fixed state for every
sign (as in your suggestion of the empty list), or its state could be
completely free yet inconsequential.  The generalization here would be
that, for any sign in the language, the set of possible states the for
the CLTS nook in the language organ should be exactly the same-- and
thus, no sign ever provides any information about its state, nor its
state provide any information about the set of signs that could be in
present consideration by the notorious language organ.

Thus, here comes The Point: it is possible that on one kind of
interpretation of the models (e.g. the "psychological" one I suggest)
some features could be motivated which, in another interpretation
(e.g. the strictly informational one) would be completely unmotivated.
However, I don't think this means that one of the interpretations is
"right" and the other one "wrong"-- they're just *different*, and in
my toy example, presumably they're different in a insightful way.  (Of
course, I mean "insightful" relative to precisely the right sort of
toy world in which my toy example is insightful.)

How is all this relevant to theories of language variation and
universals in HPSG?  I really can't think clear about this by this
point in this email.  I was thinking that corresponding to each
different way of interpreting the models there may be corresponing
notions of what "universal" and "variation" mean.  What "a linguistic
universal" would mean under the informational interpretation ("an
informational invariant across languages"?) is quite "abstract".
(Given all the recent controversy about the uses of "abstract", I
hereby clarify what I presently mean with it: "I have no clue how one
would even start thinking about this.")

Enough.

--
Luis Casillas
Department of Linguistics
Stanford University