Hi And...

Thanks for the quick reply!

>> (0)   The dog Joe thinks John hit died
>>
>> (1)  * The dog I screamed when Dave hit died
>>
>> (2)  * Joe seemed likely that John would go
>>
>> (3)  * It acted likely that John would go
>>
>> (4)    It seemed likely that John would go
>>
>>
>> I wonder how the difference between 0 vs. 1, and the difference
>> between 2-3 vs. 4, are handled in word grammar?
>
>
> (0) vs (1): (1) violates a constraint that the link from "hit" to "dog" is
> not valid if the path from "dog" to "hit" contains a 'tensed adjunct', i.e.
> an adunct ("when") with a tensed complement ("hit").

This is interesting....

So the link from "dog" to "hit" (a B link, in the link grammar
dictionary , which would otherwise be judged valid, is judged as
invalid because of the C link from "when" to "Dave" (and then the S
link from "Dave" to "hit")

This would seem to be solvable within the plain old link grammar
dictionary framework.  But to make it less messy, we would want to
make the C link (or something like it) point from "when" to "hit"
rather than "when" to "Dave" [and possibly this has been done in the
most recent link parser dictionary, I'm looking at an old version
right now]

We could then make this sentence ungrammatical just by editing the
disjunct involving the B link from "dog" to "hit", to specify that it
can't be formed when there is a "C" type link from some word (in this
case "when") from "when" to "hit"

> (2) vs (4). I don't remember how exactly Dick analyses (4), but (2) is out
> because no rules generate it in the first place. In (2), the "that" is licit
> only if its accompanying "it" is present.

I see.  So probably in link grammar, this case could be handled via
editing the appropriate link parser dictionary entries, so that some
of the disjuncts
involving "that" require the link to the accompanying "it"

> (3) is not ungrammatical, just semantically anomalous. "Act" requires its
> subject to be the actor in the acting. "It" is either meaningless or
> coreferential with "that (John would go)"; either way, it's not something
> that can be conceptualized as a plausible actor.

This sounds sensible, yet I wonder why Sleator & Temperley thought
differently...

Hmm...

Overall, your reply suggests to me that the postprocessing rules may
well be comparable to the fat links that Linas removed.

That is: they can probably be dealt with via just editing the link
grammar dictionary appropriately, and not in such incredibly perverted
ways.   Some cleanup may be required to the way coordination oriented
links are handled.

The theoretical implication, in quasi-Chomskian language, would be
that Closure type phenomena  can be handled via Merge type operations,
which is very much in the spirit of the Minimalist programme.   (I
view link grammar's piecing together of words based on their
connectors as an implementation of Merge; and the link grammar
postprocessor as a probably unnecessary additional implementation of
closure related constraints....)

To dig deeper I'd have to come up with more examples by running them
through the current link parser and seeing which sentences WOULD be
judged grammatical if not for the postprocessor...

But if this analysis is correct then it means from the standpoint of
unsupervised learning of a link grammar dictionary, we can ignore the
postprocessor.   If the postprocessor is unnecessary like the fat
links were, then an effective lexicon learning system would just learn
the appropriate lexicon to take care of the closure-type phenomena the
postprocessor handles, without need for any messy postprocessor...

Thanks again...

-- ben