Prevailing approaches do not have a computational lexicon

[Mark Johnson]

>... The words are then combined
>via merge, into phrase markers, where all the relevant features are
>"checked" (very roughly equivalent to unified -- which is why some of us
>lurk on these lists). ...
>
I haven't kept up with Minimalism since I took my statistical turn, but
it always seemed to me that MP feature checking doesn't have much to do
with unification, but instead is really analagous to the kind of
cancellation you see in categorial grammar or linear logic; i.e., the
very act of checking a feature consumes or discharges it.

As far as I know, Stabler's formulation of Minimalist Grammars
formalizes feature checking as a kind of cancellation process.

Right now LFG actually incorporates both kinds of feature interaction:
it has feature unification (in the f-structures) and feature
cancellation (in the linear logic semantic interpretation component).
About 5 or so years ago I wondered if it was really necessary to have
both kinds of feature interaction in a single grammar, and came to the
conclusion that most uses of feature unification could equally well be
treated in terms of feature cancellation, and even outlined a version of
LFG that only used feature cancellation called R-LFG (for
"resource-sensitive LFG").

Perhaps more interestingly, there are a few cases in which feature
cancellation seems to provide a simpler and more natural treatment
(e.g., Ingria's examples of the failure of transitivity of agreement).
 Mary and Ron and others came up with some very interesting alternative
analyses that involve making all agreement features into set-valued
features and express agreement as subset rather than equality
constraints.  To my mind at least this is just a simulation of feature
cancellation in a unification grammar (assuming you think subset
constraints ought to be part of a unification grammar).

IMHO I feel that it is better to design your formal theory so that its
formal primitives are as close as possible to the linguistic operations
used in your analyses; i.e., why not construct your theory with feature
cancellation as a primitive, rather than code it up using set values and
subset constraints?  Of course, not everyone agrees with this (HPSG
seems to be designed around the idea of defining linguistic constraints
out of what is ultimately few simple but very general primitives).
 Ultimately I suppose what matters is whether the formal framework
encourages linguists to make interesting empirical discoveries; in that
regard LFG, HPSG and CG have all been productive, so perhaps it really
doesn't matter whether a theory uses feature unification or feature
cancellation.

Sorry for introducing a spurious discussion of LFG into the conversion,
but this is the LFG list (or at least that's where I found the message I
replied to).

Best to everyone,

Mark