1 AUTOMATIC GENERATION OF RECIPROCAL QUES-ROCK BAND, WE FOUND THAT H...
5.1 Automatic Generation of Reciprocal Ques-
rock band, we found that he is not only that, but also
tions
the community investment manager of the English
While not done in the work reported here, we are
conglomerate Whitbread, the executive director of
looking at generating reciprocal questions automati-
the U.S. Figure Skating Association, a writer for
cally. Consider the following transformations:
New Scientist, an Australian medical advisor to the
WHO, and the general sales manager of Houseman,
“What is the capital of California?” -> “Of what
a supplier of water treatment systems. Thus the
state is <candidate> the capital?”
problem of word sense disambiguation has returned
in a particularly nasty form. To be fully effective,
“What is Frank Sinatra’s nickname?” ->
“Whose (or what person’s) nickname is <can-
QDC must be configured not just to find a consistent
didate>?”
set of properties, but a number of independent sets
that together cover the highest-confidence returned
“How deep is Crater Lake?” -> “What (or what
answers
3
. Altogether, we see that some of the very
lake) is <candidate> deep?”
problems we aimed to skirt are still present and need
to be addressed. However, we have shown that even
“Who won the Oscar for best actor in 1970?”
disregarding these issues, QDC was able to provide
-> “In what year did <candidate> win the
substantial improvement in accuracy.
Oscar for best actor?” (and/or “What award
did <candidate> win in 1970?”)
6 Summary
These are precisely the transformations necessary
We have presented a method to improve the accu-
to generate the auxiliary reciprocal questions from
racy of a QA system by asking auxiliary questions
the given original questions and candidate answers
for which natural constraints exist. Using these con-
to them. Such a process requires identifying an en-
straints, sets of mutually consistent answers can be
tity in the question that belongs to a known class,
generated. We have explored questions in the bio-
and substituting the class name for the entity. This
graphical areas, and identified other areas of appli-
entity is made the subject of the question, the previ-
cability. We have found that our methodology
ous subject (or trace) being replaced by the candi-
exhibits a double advantage: not only can it im-
date answer. We are looking at parse-tree rather
than string transformations to achieve this. This
3
Possibly the smallest number of sets that provide such cover-