MULTI-ROLES.txt

# Note that movement chains may have more than two positions.  
# For example (leaving out some levels for readability):
#
# (SBAR (WHNP-1 which) 
#       (S\NP-1 (NP-2\NP-1 0) (VP\NP-2 struggled (S\NP-2 (NP\NP-2 0) (VP to live)))))

# Remarks:
# 
#    "struggled" doesn't merely agree to pass on the gap from its
#    complement S, but actively recruits a complement S that has such
#    a gap.  Similarly, "to live" (because it is untensed) actively
#    recruits a gapped subject (in fact, a completely null one) and
#    then passes up the gap.
#
#    there are 3 influences on the lexical realization of this 3-step
#    chain: its position as head of the SBAR, and its government by
#    the verbs "struggled" and "live," where it appears as subject.
#    These correspond to the origin and landing sites of NP-1 and NP-2
#    (note that NP-2\NP-1 is both an origin and a landing).  
#    Put another way, they correspond to WHNP-1, NP-2, and the NP that
#    is slashed away by NP-2.  By contrast, the slashes in VP\NP-2
#    and S\NP-2 are merely transmissions -- the transmission may be
#    subject to island effects, but there is no influence here on
#    the lexical realization of NP-2.  This is generally the right
#    attitude, except for passives (the NP fills the subject position,
#    but not the subject theta role, extraposition/expletives (the landing site
#    has no theta role), ... !!!
#
#    Ignoring the second movement, struggled appears in the frame
#    S --> NP-2 struggled S\NP-2.  Notice that NP-2 plays two roles
#    here.  Object control verbs should come out the same way
#    (S --> NP convince NP-2 S\NP-2), but the Treebank doesn't
#    bracket them as object control.  (Might try fixing this.)
#
#    WHNP-1 is really an abbreviation for S/(S\NP-1).  
#

When generating fillers for traces in the probability model -- if we
have kids Z-1 and Y/X-1, expand the Y/X-1 part first, until we fill
some descendant X/X-1 with 0; then fill the Z-1 in such a way that it
would also make a good X for the X/X-1's parent.  Another way of
thinking of this is that we fill X/X-1 with something that's
reasonable for X and then immediately move it in such a way that it
will make a good Z-1 for Z-1's parent.

Perhaps a gap like X/X-1 should be marked on the surface as an complete
gap rather than an X with a partial gap.  There are probably edit ops
that are willing to gap a full constituent but not a partial one.


----(an attempt to reproduce the above when it was thought lost)----

When we do lexical insertion into a chain, we'd like to have
multiple lexical influences -- we're inserting into all the
landing sites "at once."  What's the right generative model for this?

We should be able to handle it by generating in the right order:

Ignore knobs for the moment.  When we generate X/Y-i and its sister
filler Z-i, we should pursue expanding X/Y-i until we have reached
some X'/Y-i that expands as 0.  Then we can do lexical insertion into
the top position Z-i, conditioned by the context of the bottom position.

!!! apply this to multiple movements as above ...

!!! think about knobs ...