Filler-gap mismatches

[Robert Levine]

>
> Hi Bob,
>
> In all fairness, my recollection is that even though G/HPSG folks were
> thinking about these kinds of mismatches in the early-to-mid 1980's,
> we were thinking about them because LFG folks had called attention
> to them.
>

I actually first came across these kinds of facts in a manuscript by
Polly Jacobson that I was using for introductory graduate syntax at
one point; if I recall correctly, some of these surface in a paper she
published on connectivity in NLLT in 1984---and as Georgia pointed out
to me, Postal mentioned these kinds of facts in `On Raising'. My
impression is that there was a general awareness of this kind of data
from early on, but they weren't regarded as the thing to focus on; the
complementary cases, where you do have fairly tight connectivity, was
what you wanted to capture, but with a mechanism which allowed a
relaxation of that tightness in a plausible way.

> I don't see how they bear on any empirical distinction
> between LFG hypotheses about extraction vs. HPSG hypotheses...
> >>
>
> Well, if the linkage is treated as equality between paths of functions
> and functions at the ends of those paths don't make reference to
> part of speech, then you are not bothered by part-of-speech mismatches;
> whereas if you handle the linkage in terms of equality of _local_
> feature bundles which DO mention part of speech, then you have more
> work to do if something that you claim is a UDC doesn't preserve
> part of speech. Conversely, functional equations ALONE won't explain
> the cases where part-of-speech identity is required. So both
> frame works have to bring in other devices (besides the
> principal explanatory mechanism).
>

Right, but in a sense that's the point: if you only had strongest
connectivity in all UDCs, all the time, then a framework that failed
to capture that---that *allowed* such universal strong connectivity
but failed to enforce it---would clearly be missing something; or if
connectivity varied over a wide spectrum of strength in the majority
of cases, then a framework which enforced strongest connectivity but
had to resort to macabre acrobatics to relax it in most cases would
also be missing something. Neither of those polar situations hold, so
the issue seems to have come down to what one takes as one's point of
departure. There's actually quite a bit of flexibility in HPSG on this
point: weak UDCs can be accomodated by the familiar P&S94 mechanisms,
part-of-speech disconnectivity by special lexical entries for the
cases in question along the lines I sketched, and so on, and it's not
clear that either the function-equality approach and the
morphosyntactic feature-bundle equality approach to the definition of
connectivity is going to emerge as clearly preferable. It just seems
to me important to make it clear that while category disconnectivity may
enter into a comparison between LFG and HPSG, there really isn't any
clear sense in which it gives even a slight competitive advantage in the
theoretical marketplace to either of them...

cheers,

Bob