Re: [Corpora-List] Incidence of MWEs

From: Rob Freeman (lists@chaoticlanguage.com)
Date: Mon Mar 20 2006 - 09:44:00 MET

  • Next message: Flint, John M.: "[Corpora-List] Date: Mon, 20 Mar 2006 15:50:40 -0000"

    Hi Mike,

    Thanks for this. You have thought about my post, which means I may be able to
    explain my position.

    On Monday 20 March 2006 03:55, Mike Maxwell wrote:
    > Rob Freeman wrote:
    > > Surely the question is are tags sensible parameters of language in
    > > the first place.
    >
    > I am not sure what you mean by "parameters".

    I mean the same as you mean. The difference in our thinking is not in the
    meaning of the word "parameters". The difference is in the nature of the
    parameters we think are appropriate to language.

    Briefly...

    You think generalizations of usage parametrize language.

    I think usage parametrizes language by virtue of various generalizations.

    They are not the same.

    Remember I said my differentiating point was "...it is not enough that your
    model be based on generalizations of usage, you must also allow for the
    possibility of discontinuous change in those generalizations."

    To push your rope metaphor, wound and unwound are not just two ways of looking
    at the same thing. Sometimes the way you wind the rope makes a difference.

    Your metaphor suggests there is only one way of winding language usage into a
    "rope", and the rope itself, wound in this way, is a sufficient parameter. In
    contrast I think there are many different ways of winding that rope, and we
    need them all.

    In my opinion this potential of different "windings" is the single greatest
    misapprehension preventing language engineering (in particular) moving
    forward.

    One "winding" for each MWE.

    > > Instead of worrying where MWE's start and stop, let's accept that
    > > MWE's cover all of language. All language is an MWE.
    >
    > Except for this, which isn't an MWE. And except for your posting, which
    > isn't an MWE either (at least not one that I've ever seen before).

    This is easy to refute. At least, it is standard theory in some areas of
    linguistics.

    "Except for this" is an MWE. Replacing "except" with "also" what does "also
    for this" mean? Replacing "for" with "to benefit", what does "except to
    benefit this" mean? Replacing "this" with "it" is "except for it" acceptable
    usage? Is "except for here" acceptable usage, or should that just be "except
    here"? Is "excepting for this" proper usage? If so, is "including for this"
    meaningful?

    "Except for this" is an MWE, and so is everything else. I could go on. Pawley
    and Syder is a classic reference on this.

    > > Explain MWE's in terms of generalizations over usage and let's start
    > > thinking about how we can use these generalizations over usage
    >
    > Uh, let's see. Here's a generalization over usage: the MWE "kick the
    > bucket" has a distribution much like the MWE "fire off a shot", which
    > has a distribution much like the MWE "pick up the pace", etc. Let's
    > make up a label for these MWEs that obey this generalization--I dunno,
    > maybe "VP".
    > ...

    I think I covered this (different "windings") above, but it is worth
    repeating.

    This is the important bit. This is the core of the difference between us.

    This is exactly the mistake which is made by the popular modern sub-field of
    Grammatical Inference, and statistical NLP in general. This is where they
    fail.

    You are making generalizations, and assuming these generalizations can be
    treated as sufficient parameters.

    The problem is there is no single complete set of generalizations of this
    kind to be made. You can make generalizations, but the generalizations are
    necessarily multiple and inconsistent.

    Of historical interest, I understand this is related to the result which
    Chomsky observed (for phonology) and used to force the abandonment of context
    distributions as a parameter of grammatical inference in the '50s.

    Chomsky used this result to force the abandonment of rules based on
    generalizations over usage. We must now use this result (in contrast to
    Chomsky) to give precedence back to generalizations over usage, and refute
    the idea of a single consistent set of rules, instead.

    Whatever we do, we must not ignore the result (inconsistent generalizations)
    itself anymore. That would be reason for Chomsky to (continue to) hold us in
    contempt, indeed.

    -Rob



    This archive was generated by hypermail 2b29 : Mon Mar 20 2006 - 09:36:03 MET