[Prev][Next][Index]

[xtag-meeting]: Extraposition -- follow up to meeting


So here are some facts about the extraposition examples that I didn't have
with me at the meeting today, but which some of you were interested in.

Fei's algorithm extracted a bunch of examples of non-local dependencies
from the treebank (Treebank 1). She then hand-classified these in terms of
what kind of dependency it was -- there were 7 main types. There were 337
that looked like extraposition from NP. Out of these, many could be set
aside because they could be analyzed as local or not true examples of
extraposition.

Some examples were considered local because the host NP was inside a PP
whose preposition should be considered a co-anchor of the elementary tree.
for example:

\ex
Mr. Allen quickly ran up against the liberal {\bf establishment} again
\HCone{which somehow elevated the vague concept of Indian rights above the
rights of individual Indians}. (\#2793)

\ex
Ms. Garzarelli said she was swamped with {\bf phone calls} over the
weekend \HCone{from nervous shareholders}. (\#2984)

Other examples didn't seem like true examples of extraposition. for
example:

\ex
Max Gitter, another GAF defense attorney, said yesterday, ``As we go in
for
the third time, Yogi Berra's famous {\bf line} *ICH* is apt : \HCone{`It's
deja vu all over again.'}'' (\#2023)
\end{exe}

here the extraposition is supposed to be between the quote and "line". In
any case -- this is the single clause case.

Out of the real extraposition examples, I only looked at extraposition of
full relative clauses (as opposed to PPs, reduced relatives, etc.) There
were 41 examples of extraposed full relatives that I thought were real
extrapositions. Out of these, all but 4 were tree-local in the sense that
the host NP and the modifying rel clause attached (adjoined or
substituted) into a single elementary tree. Here are the 4 non-local
examples.

Two were cases where the NP was embedded in a partitive:

\ex
Mr. Black said that he doesn't see {\em a lot of} {\bf excesses} out there
\HCone{that would tilt us into recession}. (\#266)

\ex
The FCC plans to hear {\em a day of} {\bf testimony} Nov. 16 \HCone{on the
plight of AM radio}. (\#1640)

and these are the other two:

\begin{exe}
\ex
Supply troubles were on {\em the minds of} {\bf Treasury investors}
yesterday
\HCone{who worried about the flood of new government securities}. (\#2638)

\ex
The candy business had {\em sales of} {\bf about \$154 million} last year,
\HCone{which was roughly 12\% of total revenue...} (\#1068)

\end{exe}

The question is whether the rel clause is really modifying "sales" or
whether one of our alternative solutions would work in this case (i.e.,
the single variable approach, etc.).

Also, keep in mind that there may be PP, reduced relative, etc. -cases
that still need to be examined. So there may be other relevant examples.

The result of the XTAG meeting was that we should implement the "local"
cases of extraposition for the time-being. So the question regarding these
non-local cases is not about how or if to implement them now, but more of
a theoretical or in principle issue.


Tonia
Follow-Ups: