SLASH-AND-PLUS.txt

FIRST, AN INCORRECT VIEW OF THE PROBLEM, WITH SOLUTION
------------------------------------------------------

Think about

    (NP-SBJ-80 
      (NP (NNS Modifications) )
      (PP (-NONE- *ICH*-3) ))
    (VP (VBD had) 
      (VP (VBN been) 
        (VP (VBN made) 
          (NP (-NONE- *-80) )
          (PP-3 (TO to) 
            (NP (DT the) (NNP Souper) (NNP Combo) (NN product) ))

where really, just a part of "modifications to the ... product" has been
moved.  This comes out in flattened form as 
    NP-80/PP-3  V NP/NP-80 PP-3
but really the way to think of it is that
the PP got moved out of the NP:
    V NP   --->   V NP/PP-3 PP-3
and then the NP/PP-3 got moved to subject position.  So let's
write   
    [NP-SBJ/PP]-80 V [NP/PP-3]/[NP/PP]-80 PP-3
Indeed, this is much better for the NON-flattened structure,
where there is a VP node, since both the -3 and -80 traces are 
discharged by c-commanding positions.

Notice that the PP trace inside the NP-SBJ is not linked up with the
overt PP in any way.  Moreover, the gap passed up within [NP-SBJ/PP]-80
isn't indexed in any way that tracks the progress of the gap.
So we probably want
    [NP-SBJ/PP-6]-80 V [NP/PP-3]/[NP/PP-6]-80 PP-3
i.e., 
    ([NP-SBJ/PP-6]-80 
      (NP (NNS Modifications))
      (PP/PP-6 0))
    (VP/[NP/PP-6]-80 (VBD had)
      (VP/[NP/PP-6]-80 (VBN been)
        (VP/[NP/PP-6]-80 (VBN made)
          ([NP/PP-3]/[NP/PP-6]-80 0)
          (PP-3 to the Souper Combo Product))))

This still isn't a direct link -- PP/PP-6 is filled by PP-3.
To figure out who is in what role, we have to do index unification at
traces.  In particular, -3 and -6 have to be unified here, just as
-1 and -2 would be unified in NP-1/NP-2.  That yields

    ([NP-SBJ/PP-36]-80 
      (NP (NNS Modifications))
      (PP/PP-36 0))
    (VP/[NP/PP-36]-80 (VBD had)
      (VP/[NP/PP-36]-80 (VBN been)
        (VP/[NP/PP-36]-80 (VBN made)
          ([NP/PP-36]/[NP/PP-36]-80 0)
          (PP-36 to the Souper Combo Product))))

from which nulls can now be deleted if desired:

    ([NP-SBJ/PP-36]-80 
      (NP (NNS Modifications)))
    (VP/[NP/PP-36]-80 (VBD had)
      (VP/[NP/PP-36]-80 (VBN been)
        (VP/[NP/PP-36]-80 (VBN made)
          (PP-36 to the Souper Combo Product))))

Perhaps NP-1/NP-2 should now be [NP/NP-2]-1, for consistency with
[NP-SBJ/PP-36]-80?  No - I don't think so.  NP-1/NP-2 means that NP-1
landed here and then we move a piece of it (actually all of it)
farther as NP-2.  [NP-SBJ/PP-36]-80 means just that [NP-SBJ/PP-36]-80
landed here.  We could then decide to move a piece of it higher;
that would append /FOO-99 to the end.  If we wrote [NP/NP-2]-1, it 
would mean that we had moved an actual trace -- i.e., we'd moved
NP-2 away from its original site and then we'd moved the zero residue
somewhere else!  

If we wanted to undo the moves (assuming that there's been no
unification), we should do as follows: find pairs of constituents X-i
and Y/Z-i that rewrites as -NONE-.  In general the former should
c-command the latter.  Note that i is the index for the whole X or Z,
which may be a slashed category in [].  When such a category is found,
move X-i back to fill the /Z-i gap, and remove /Z-i everywhere it
appears (generally, on some chain of ancestors of Y/Z-i up to a
sibling of X-i).  This may create new constituent pairs of this sort,
which should be treated the same way, i.e., repeat until there's
nothing left to do.

    (NP-SBJ-80 t)
    (VP (VBD had)
      (VP (VBN been)
        (VP (VBN made)
          (NP-80 (NP (NNS Modifications))
	         (PP-3 to the Souper Combo Product))
          (PP-3 t))))

An alternative would be to show the moved consituent in both
positions, specially marking the positions where it's not actually
present.  This could be achieved by taking the above form and redoing
the moves in the reverse order from the order in which we undid them
(so that we don't raise "modifications to the Souper Combo Product"
before we've removed the PP from it):

    (NP-SBJ-80 (NP (NNS Modifications))
	       (/PP-3 to the Souper Combo Product))
    (VP (VBD had)
      (VP (VBN been)
        (VP (VBN made)
          (/NP-80 (NP (NNS Modifications))
	          (/PP-3 to the Souper Combo Product))
          (PP-3 to the Souper Combo Product))))

We could also reproduce the slashes when doing all this:

    ([NP-SBJ/PP-6]-80 
      (NP (NNS Modifications))
      (*PP/PP-6 to the Souper Combo Product))
    (VP/[NP/PP-6]-80 (VBD had)
      (VP/[NP/PP-6]-80 (VBN been)
        (VP/[NP/PP-6]-80 (VBN made)
          (*[NP/PP-3]/[NP/PP-6]-80 
	    (NP (NNS Modifications))
            (*PP/PP-6 to the Souper Combo Product))
          (PP-3 to the Souper Combo Product))))

Something funny about this is that PP-3 is not REALLY an argument of
"made" -- it's merely left behind.  In particular, it is not assigned
any kind of theta role by "made."  Perhaps it should be marked as such
...  when NP is split into NP/PP-i PP-i, it should really be split
into NP/PP-i M-i, where M has a completely free distribution.  If the
category is always the same in this instance (namely *ICH*) anyway --
and it is except occasionally for some minor qualifiers -- then I
don't see any harm in this.     [But see below for a better solution.]

----------------------------------------------------------------------

A BETTER VIEW OF THE PROBLEM
----------------------------

The above doesn't explain the case where an UNMOVED subject ends up
extraposing one of its modifiers down into the VP (generally following
the object).  Some examples of this are below.  This can't have
happened when the subject was in the VP, except perhaps by some
minimalism-style story where the deep position of the subject follows
the surface position of the object.

So maybe I have the order of movements wrong for the example above:
maybe really the NP raised to subject position first, and then
extraposed a modifier back into the VP.  On this story, it's just an
accident that the modifier landed back at its original position.
That's not implausible.  Material moved into a VP often falls at the
start or the end of the adjunct sequence.  In examples like the one
above, the start would be the original position, but the end is also
possible:
   "Modifications were made recently to the product"
   "Modifications were made to the product recently"
are both good, much better than 
   "We made recently modifications to the product"

So I think I just have to bite the bullet and grant that just as
constituents may have holes (/PP) that are unified between parents and
kids, or filled by siblings, they may also have knobs -- extra
material like +PP -- that are unified between parents and kids, or
fill siblings.  

    (NP-SBJ-80/PP-3
      (NP (NNS Modifications) )
      (PP/PP-3 0))
    (VP/NP-SBJ-80+PP-3 (VBD had) 
      (VP/NP-SBJ-80+PP-3 (VBN been) 
        (VP/NP-SBJ-80+PP-3 (VBN made) 
          (NP/NP-SBJ-80 0)
          (+PP-3 (TO to) 
            (NP (DT the) (NNP Souper) (NNP Combo) (NN product) )))))

This is not all that different, in that /NP-SBJ-80+PP-3 is essentially
just /[NP-SBJ-80/PP-3].  The new point is just that if the NP-SBJ
originated in subject position, then we'd have only +PP-3 inside the
verb phrase, with no /NP-SBJ.  We now have a way to express that.

Note that if we flatten the above, so that everything depends on the verb,
we get

    NP-80/PP-3 AUX AUX V NP/NP-80 +PP-3
   
In both flattened and unflattened cases, the + effectively indicates
that the +PP-3 is not assigned a theta-role by its surface parent (the
verb), even though the verb has to licence its appearance there.
Indeed, even if the structure had originally appeared in the above
flat form, I would still have wanted to write +PP-3 for this reason.
Another advantage is that all cases arise from an edit operation that
adds /PP to one constituent and adds +PP to another (or just inserts
+PP freely).

The same is even true if we move the entire constituent but don't
change its parent, e.g.,
    NP V NP/NP-1 PP +NP-1
since we don't want NP-1 to get its theta role twice (as two different
kids of V).  It should get its theta role only in the first condition.
Note that the Treebank annotation doesn't generally allow us to conclude
that this movement has happened: such a construction would generally look 
like
    NP V PP NP
in the Treebank.

The rule is that an indexed constituent is marked with + unless it
asymmetrically c-commands its trace -- that is, unless it matches
a sibling that constains something other than the trace.  

But perhaps the rule should have something to do with headedness, in
light of the remarks about theta roles above, which I interpret as
meaning dependencies.  If the structure has been flattened, then the
asymmetric c-command definition does apply.  For consider all
the possiblities for LCA(you, your trace):
    you: then you dominate your trace, which shouldn't happen
           (we fix those cases where it does, namely the Treebank's funny
              treatment of parenthetical "he says")
    your trace: then your trace dominates you, which shouldn't happen
    your parent: then either
        (1) you're siblings with your trace, so you c-command it only 
              symmetrically, and get marked with +.  That's good because
              you have the same parent.  
        (2) you asymmetrically c-command your trace, so you don't get marked
              with +.  That's good because you have different parents.
    your grandparent or higher: then you don't c-command your trace at all, and
      get marked with +, which is good because you are merely extraposed and
      must be propagated up as a "knob" until you do c-command your trace
    
In that case, the subjects of passives would be marked with +, because
they haven't changed heads -- the c-command is symmetric in the
flattened case.  But we wouldn't use + for borderline cases where we
passivize out of some subconstituent: "??Sara was taken a liking to by
the teacher," ""??Bill was kept tabs on," and "*Bill was thought that
Mary liked."  These would still be regarded as movement, and they'd be
candidates for multiple theta roles, I guess, although I suspect that
to the extent that these items are grammatical in the first place,
they're still not constrained -- just about anything can be a passive
subject!  (Maybe the reason they're not grammatical is that they'd
expect to get a theta role in subj position if they were, and the
passive verb isn't capable of assigning one.)

I suspect that A+B only occurs if A is a verbal category, such as VP
or S or (in the flattened structure) V.  (Nominalizations probably
don't qualify here: "A group departed that included several
professors," "??Any group's departure that includes several
professors.")

Note the possibility of examples like this made-up one:

    (NP (NP the stocks) 
        (SBAR (WHNP-1 that) 
              (S/NP-1 (NP/PP-2 enthusiasm) 
                      (VP+PP-2/NP-1 (VBD ran) 
				    (ADJP-PRD (JJS highest))
				    (PP-2/NP-1 (P for) (NP/NP-1 0)) 
				    (NP-TMP this week)))))

Here, PP-2/NP-1 exports both a knob and a slash, yielding +PP-2 and
/NP-1 on the VP.  We might think that the PP-2/NP-1 should travel
together as a funny-shaped knob, giving VP+[PP-2/NP-1].  I can't
imagine what this analysis would predict differently, however.  We
might expect that PP-2/NP-1 would have to be matched by /(PP-2/NP-1),
but that's not true in the above example.  Or we might expect that the
NP-1 part would have to be filled first, but that's not true in the
above example either: in fact the above example needs a rule to
combine /PP-2 and +[PP-2/NP-1] into /NP-1, rather like CCG
composition.  Or we could even try to predict the reverse, that 
the PP-2 part would have to be filled first, but I don't think that's
so either -- there might be other reasons to extract from the 
PP within the parent VP, as in 

              (S (NP/PP-2 investors' enthusiasm)    <-- but not *investors' spirits
                 (VP+PP-2 (VP+PP-2/NP-1 (VP+PP-2/NP-1 (VBD ran) 
		                                      (ADJP-PRD (JJ high))
			                              (PP-2/NP-1 (P for) (NP/NP-1 0)) 
                                        ,
                                        and
                                        (VP/NP-1 eventually damaged)
                                        ,))
                          (NP-1 the very stocks that had been recommended by the magazine)
                       
The above example shows a funny property of knobs: because they're just
sitting in the middle of a constituent, they don't conjoin in the
usual way.  Similarly, "The factory floor was nearly bare, and
dustiest WHERE IT WAS BAREST."  "The factory floor was dusty WHERE IT
WAS BARE, and looked like it hadn't been walked on for years."  

   But wait - for extraposition, they do have to match in conjunction:
   "It is possible that John is happy, and certain that he is
   content."  The first SBAR cannot be present without the second.
   This is basically a distributed reading of "possible" that gives
   two sentences.  Is the difference that it's extraposition, or is it
   something else?  

A related case is how we might show the successive transformations
giving tough-movement.  (This is not the Treebank analysis, which puts
an operator into the SBAR and attaches it to "tough."  That analysis
is lacking because it makes it difficult to see why "this boat" is the
object of "row.")  

    For John to row this boat is tough.
    --> It is tough for John to row this boat.
    --> This boat is tough for John to row.

    (S (SBAR/SBAR-2+NP-SBJ-1 this boat)   
       (VP+SBAR-2/NP-1 (VBZ is)
           (ADJP-PRD tough)
           (SBAR-2/NP-1 (IN for)
			(S/NP-1 (NP-SBJ John)
				(VP/NP-1 (TO to)
					 (VB row)
					 (NP/NP-1 0))))))

    This analysis bears a close relation to a particular CCG-style
    analysis, in which "this boat" gets category S/(S/NP) through
    non-harmonic type-raising, and "is tough" takes an S in subject
    position:

          (SBAR/(SBAR/NP) this boat) (S\S is tough) (SBAR/NP for John to row)
          (S/(SBAR/NP) this boat is tough)          (SBAR/NP for John to row)

    Really the CCG analysis is equivalent to a version where "for John
    to row" attaches higher.  To bury it inside the VP, we need +, which
    CCG doesn't have.  Hence CCG would have trouble with 
    "this boat is tougher for John to row than it used to be,"
    or "this boat has been tough for John to row these days."

    The new analysis (or its CCG version) predicts that tough movement
    should be possible for ANY predicate that can take a sentential
    subject.  I think that's roughly true, but I'm not sure:  

       "Slime is disgusting for John to eat."  
       "?Slime would be surprising for John to eat."  
       "?Slime would be surprising to eat."     but * on the reading "surprising to me for anyone to eat"
       "Slime would be surprising to find in one's bed."

    But there may be stronger-than-usual island constraints for
    the NP extraction.  "Green cheese would be ??surprising/?easy for
    the earth to be made of."  "*This rock is possible that John
    lifted."  (These are fine with no NP movement, or with wh-movement
    of the NP.)

    Oh - non-adjectival examples suggest it's NOT true anyway:
    
       "For John to eat bricks would astonish me."
       "It would astonish me for John to eat bricks."
       "*Bricks would astonish me for John to eat."

    So the adjectives really do control it ... of course, we could say
    that the SBAR/NP extraposes into ADJP, and that it's easier to do this
    kind of extraction from an ADJP than from a VP ...

A funny case for knobs:

   VP = draw t_i and send pictures_i to us

   which Debbie claims people often say, while agreeing that it feels
   ungrammatical.  (She has a parallel example that Cathy wrote in the
   Amherst Alumni newsletter -- contains the word endeavor.)  I
   suppose it gets slighly better (but not much!) if "to us" is replaced by
   something more adjuncty:

      he killed and grieved over the eagle daily

   This would presumably be parsed as 

      (VP (VP/NP-1 draw) and (VP+NP-1 send (NP-1 pictures) to us))

   so that we get a VP by conjoining two VPs, where part of one
   VP has been extraposed into the other VP.  Note that this knob,
   "pictures," gets theta roles in both positions.  

   This analysis can be extended, of course, to the "normal"
   right-node-raising cases.  But I wouldn't like to encourage it, I
   don't think.

Parasitic gaps:

      (WHNP which papers_i) (SQ did you (VP file t_i (PP without (S-NOM PRO reading_i))))

   turns into 
 
      (WHNP/NP-1 which papers) (SQ/NP-1 did you (VP/NP-1 file NP-1/NP-1 (PP/NP-1 without (S-NOM/NP-1 PRO (VP/NP-1 reading)))))

   (Is this handled correctly by current scripts?  Probably -- it's
   basically the same slash unification mechanism as in conjunction.)


Maybe think of oblique arguments as involving knobs:

   (NP the apple) was devoured (PP+NP by (NP Sara))
   I want (S+NP (NP Sara) to devour more apples)

   

   That is, "Sara" is simultaneously the object of "by" and the subject
   of "devour."  Just as in "I want Sara to devour apples," where is
   is the object of "want" and the subject of "devour."

----------------------------------------------------------------------

INTERESTING EXAMPLES
--------------------

Some examples of the basic phenomenon of extraposition of a modifier
or appositive out of subject position.  In some cases, like the third,
the whole subject (which is some kind of S) may be extraposed and
replaced by an expletive.  In the fourth and fifth cases, the
extraposition isn't actually out of subject position, but rather out
of an operator that controls subject position, or an NP modified by an
operator that controls object position.

       (NP-SBJ 
	 (NP (NNS Areas) )
	 (PP (IN of) 
	   (NP (DT the) (NN factory) ))
	 (SBAR (-NONE- *ICH*-2) ))
       (VP (VBD were) 
	 (ADJP-PRD (RB particularly) (JJ dusty) )
	 (SBAR-2 
	   (WHADVP-1 (WRB where) )
	   (S 
	     (NP-SBJ-8 (DT the) (NN crocidolite) )
	     (VP (VBD was) 
	       (VP (VBN used) 
		 (NP (-NONE- *-8) )
		 (ADVP-LOC (-NONE- *T*-1) ))))))


      (NP-SBJ 
        (NP 
          (NP (NNP Time) (POS 's) )
          (NN ad) (NN rate) )
        (PP (-NONE- *ICH*-2) ))
      (VP (MD will) 
        (VP (VB be) 
          (ADVP (RB effectively) )
          (ADJP-PRD 
            (NP-ADV (CD 7.5) (NN %) )
            (JJR higher) )
          (PP-2 (IN per) 
            (NP (NN subscriber) )))))


      (NP-SBJ 
	(NP (PRP It) )
	(SBAR (-NONE- *EXP*-3) ))
      (VP (VBZ is) 
	(ADJP-PRD (JJ possible) )
	(, ,) 
	(PP (IN of) 
	  (NP (NN course) ))
	(, ,) 
	(SBAR-3 (IN that) ...))



      (WHNP-41 
        (WHNP (WP What) )
        (PP (-NONE- *ICH*-1) ))
      (S 
        (NP-SBJ (-NONE- *T*-41) )
        (VP (VBZ 's) 
          (ADJP-PRD (RB so) (JJ wild) )
          (PP-1 (IN about) 
            (NP 
              (NP (DT the) (NNS funds) (POS ') )
              (NN frenzy) ))
          (ADVP-TMP (RB right) (RB now) ))))


      (NP-SBJ 
	(NP 
	  (NP (DT the) (NN enthusiasm) )
	  (PP (-NONE- *ICH*-2) ))
	(SBAR 
	  (WHNP-1 (-NONE- 0) )
	  (S 
	    (NP-SBJ (NNS investors) )
	    (VP (VBD showed) 
	      (NP (-NONE- *T*-1) )
	      (PP-2 (IN for) 
		(NP (NNS stocks) ))

    (NP-SBJ 
      (NP (DT An) (RB entirely) (JJ new) (NN band) )
      (SBAR (-NONE- *ICH*-2) ))
    (VP (VBZ rings) 
      (NP-TMP (NN today) )
      (PP-LOC (IN at) 
        (NP (JJ Great) (NNP Torrington) ))
      (, ,) 
      (SBAR-2 
        (WHNP-1 
          (NP (JJ several) )
          (WHPP (IN of) 
            (WHNP (WP whom) )))
        (S 
          (NP-SBJ (-NONE- *T*-1) )
          (VP (VBP are) 
            (NP-PRD 
              (NP (NNS members) )
              (PP (IN of) 
                (NP (DT the) (NN congregation) )))))))


      (NP-SBJ 
        (NP 
          (QP (DT A) (NN total) (IN of) (CD 139) )
          (NNS companies) )
        (ADJP (-NONE- *ICH*-1) ))
      (VP (VBD raised) 
        (NP (NNS dividends) )
        (PP-TMP (IN in) 
          (NP (NNP October) ))
        (, ,) 
        (ADJP-1 (RB basically) (JJ unchanged) 
          (PP (IN from) 
            (NP 
              (NP (CD 138) )
              (ADVP-TMP 
                (NP (DT a) (NN year) )
                (RB ago) ))))))

 
      (NP-SBJ 
	(NP 
	  (NP (DT a) (NN scene) )
	  (PP (IN of) 
	    (NP (NN devastation) )))
	(NP (-NONE- *ICH*-1) ))
      (VP (VBZ emerges) (: :) 
	(NP-1 
	  (NP 
	    (NP (DT a) (NN freeway) )
	    (VP (VBN crumbled) 
	      (NP (-NONE- *) )
	      (PP-CLR (IN into) 
		(NP (DT a) (JJ concrete) (NN sandwich) ))))
	  (, ,) 
	  (NP 
	    (NP (VBZ hoses) )
	    (VP (VBG pumping) 
	      (NP (NN water) )
	      (PP-DIR (IN into) 
		(NP (JJ once-fashionable) (NNS apartments) ))))
	  (, ,) 
	  (NP (VBN abandoned) (NNS autos) )))


Passivization examples like the one I started with:

      (NP-SBJ-161 
        (NP (RB Much) (RBR less) 
          (-LRB- -LCB-)
          (NN index-arbitrage) (NN activity) 
          (-RRB- -RCB-) )
        (PP (-NONE- *ICH*-1) ))
      (VP (VBZ is) 
        (VP (VBN done) 
          (NP (-NONE- *-161) )
          (PP-LOC (IN over) 
            (NP (RB here) ))
          (PP-1 (IN than) 
            (PP-LOC (IN in) 
              (NP (DT the) (NNP U.S.) )))))

    (NP-SBJ-2 
      (NP (JJ Dark) (NNS hints) )
      (SBAR (-NONE- *ICH*-1) ))
    (VP (MD would) 
      (VP (VB be) 
        (VP (VBN raised) 
          (NP (-NONE- *-2) )
          (SBAR-1 (IN that) 
            (S 
              (NP-SBJ 
                (NP (NNS parts) )
                (PP (IN of) 
                  (NP (DT the) (NN administration) )))
              (VP (VBP hope) 
                (SBAR (-NONE- 0) 
                  (S 
                    (NP-SBJ (NNP Mr.) (NNP Gorbachev) )
                    (VP (MD would) 
                      (VP (VB fail) ))))))))))


The following shows that the problem can't be solved just by
eliminating VP level, because the +PP is propagated down through TWO
lexical verbs.  (Interestingly, the /PP is also propagated down
through two lexical nouns.)  

   It would still be possible to use a solution where PP-5 is assumed
   to attach at a higher level where it would c-command *ICH*-5; that
   would be blocked by a version to which we'd added "in coming years"
   (which sounds fine).  Some such examples are shown in a moment.

        Few insurers have announced 
	(WHNP-3                                                  (WHNP-3/PP-5                                       
	  (WHNP (WDT what) (NN amount) )		           (WHNP/PP-5 (WDT what) (NN amount) )           
	  (PP (IN of) 					           (PP/PP-5 (IN of)                              
	    (NP 					             (NP/PP-5                                    
	      (NP (NNS losses) )			               (NP (NNS losses) )                        
	      (PP (-NONE- *ICH*-5) ))))			               (PP/PP-5 0))))                            
	(S 						         (S/NP-3+PP-5                                    
	  (NP-SBJ-1 (PRP they) )			           (NP-SBJ-1 (PRP they) )                        
	  (VP (VBP expect) 				           (VP\NP-SBJ-1/NP-3+PP-5 (VBP expect)           
	    (S 						             (S\NP-SBJ-1/NP-3+PP-5                       
	      (NP-SBJ (-NONE- *-1) )			               (NP-SBJ\NP-SBJ-1 0)                       
	      (VP (TO to) 				               (VP/NP-3+PP-5 (TO to)                     
		(VP (VB see) 				                 (VP/NP-3+PP-5 (VB see)                  
		  (NP (-NONE- *T*-3) )			                   (NP/NP-3 0)                           
		  (PP-5 (IN from) 			                   (PP-5 (IN from)                       
		    (NP (DT the) (NN earthquake) )))))))                     (NP (DT the) (NN earthquake) )))))))




These examples show that we can't solve the problem by attaching
the extraposed constituent at a higher level, at least not without
moving other material too!  The second example is remarkable for
doing such extraposition twice -- a PP is extraposed from the subject,
and an appositive NP is extraposed from inside that PP.

	( (S 
	    (NP-SBJ-1 
	      (NP (DT A) (NN hearing) )
	      (PP (-NONE- *ICH*-2) ))
	    (VP (VBZ is) 
	      (VP (VBN scheduled) 
		(NP (-NONE- *-1) )
		(PP-2 (IN on) 
		  (NP (DT the) (NN issue) ))
		(NP-TMP (NN today) )))

	    (. .) ))

	( (S 
	    (NP-SBJ-1 
	      (NP (NN Work) )
	      (PP (-NONE- *ICH*-2) ))
	    (VP (VBD was) 
	      (VP (VBN begun) 
		(NP (-NONE- *-1) )
		(PP-2 (IN on) 
		  (NP 
		    (NP (JJ single) (NN family) (NNS homes) )
		    (PRN (: --) 
		      (NP 
			(NP (DT the) (NN core) )
			(PP (IN of) 
			  (NP (DT the) (NN housing) (NN market) )))
		      (: --) )))
		(PP (IN at) 
		  (NP 
		    (NP 
		      (NP (DT an) (JJ annual) (NN rate) )
		      (PP (IN of) 
			(NP (CD 971,000) )))
		    (NP (-NONE- *ICH*-3) )))
		(PP-TMP (IN in) 
		  (NP (NNP September) ))
		(, ,) 
		(NP-3 
		  (NP (DT a) (NN drop) )
		  (PP (IN of) 
		    (NP (CD 2.1) (NN %) ))
		  (PP-TMP (IN from) 
		    (NP (DT the) (JJ previous) (NN month) )))))
	    (. .) ))




It should be noted that some *ICH*'s are c-commanded properly,
because they're extracted from object position.  Of course,
we still want to mark these with +.

                  (NP-SBJ (EX there) )
                  (VP (VBP are) 
                    (NP-PRD 
                      (NP (JJR more) (NNS risks) )
                      (PP (IN for) 
                        (NP (DT the) (NN pound) ))
                      (SBAR (-NONE- *ICH*-4) ))
                    (PP-LOC (IN under) 
                      (NP (JJ Major) ))
                    (SBAR-4 (IN than) 
                      (S 
                        (NP-SBJ (EX there) )
                        (VP (VBD were) 
                          (NP-PRD (-NONE- *?*) )
                          (PP-LOC (IN under) 
                            (NP (NNP Lawson) ))))))


                (S 
                  (NP-SBJ (PRP we) )
                  (VP (MD can) 
                    (VP (VB bring) 
                      (NP 
                        (NP (DT the) (NN cost) )
                        (PP (-NONE- *ICH*-2) ))
                      (PRT (RP down) )
                      (PP-2 (IN of) 
                        (NP (DT these) (NNS systems) ))
                      (PP-EXT (IN by) 
                        (NP 
                          (NP 
                            (QP (CD 30) (NN %) (TO to) (CD 40) )
                            (NN %) )
                          (NP-ADV (DT a) (NN year) ))))))

This one is interesting for extracting two ICH's in a row.  

( (S 
    (NP-SBJ-3 (NNS Banks) )
    (ADVP (RB also) )
    (VP (VBP want) 
      (S 
        (NP-SBJ-6 (-NONE- *-3) )
        (VP (TO to) 
          (VP (VB be) 
            (ADJP-PRD (JJ able) 
              (S 
                (NP-SBJ (-NONE- *-6) )
                (VP (TO to) 
                  (VP (VB call) 
                    (NP 
                      (NP (NN someone) )
                      (SBAR 
                        (WHNP-5 (-NONE- 0) )
                        (S (-NONE- *ICH*-4) ))
                      (NP (-NONE- *ICH*-1) ))
                    (PP-MNR (IN on) 
                      (NP (DT the) (NN telephone) ))
                    (S-4 
                      (NP-SBJ (-NONE- *T*-5) )
                      (VP (TO to) 
                        (VP (VB fix) 
                          (NP 
                            (NP (DT a) (NN problem) )
                            (PP (IN with) 
                              (NP 
                                (NP (DT a) (NN deal) )
                                (SBAR 
                                  (WHNP-2 (WDT that) )
                                  (S 
                                    (NP-SBJ (-NONE- *T*-2) )
                                    (VP (VBZ goes) 
                                      (ADJP-PRD (JJ bad) ))))))))))
                    (: --) 
                    (NP-1 
                      (NP (RB preferably) (NN someone) )
                      (ADJP (JJ other) 
                        (PP (IN than) 
                          (NP (DT a) (NN union) (NN leader) ))))))))))))
    (. .) ))

Also amusing for its lots of traces.

    (SBARQ-4 
      (WHADVP-1 (WRB Why) )
      (SQ (VBP do) 
        (NP-SBJ (NNS people) )
        (VP (VB play) 
          (NP (DT the) (NN lottery) )
          (SBAR-TMP 
            (WHADVP-2 (WRB when) )
            (S 
              (NP-SBJ 
                (NP (DT the) (NNS odds) )
                (PP (-NONE- *ICH*-3) ))
              (VP (VBP are) 
                (ADJP-PRD (JJ great) )
                (PP-3 (IN against) 
                  (NP (PRP them) ))
                (ADVP-TMP (-NONE- *T*-2) ))))
          (ADVP-PRP (-NONE- *T*-1) )))
      (. ?) )