2002), answer types in QA systems typically corre-
the usefulness of the entire approach.
spond to the types identifiable by their named-entity
recognizer (NER). There is no agreed-upon number
As a consequence, we have identified the need for a
of classes for an NER system, even approximately.
component whose sole purpose is to establish the
It turns out that for best coverage by our
equivalence, or generally the kind of relationship,
C
ONSTRAINTS M
ODULE, it is advantageous to have a
between two terms. It is clear that the processing
relatively large number of types. It was mentioned
will be very type-dependent – for example, if two
in Section 4.2 that certain questions were not invert-
populations are being compared, then a numerical
ible because no terms in them were of a recogniz-
difference of 5% (say) might not be considered a
able type. Even when questions did have typed
difference at all; for “Where” questions, there are
terms, if the types were very high-level then creating
issues of granularity and physical proximity, and so
a meaningful inverted question was problematic.
on. More examples of this problem were given in
For example, for QA without Constraints it is not
(Prager et al. 2004a). Moriceau (2006) reports a
necessary to know the type of “MTV” in “When
system that addresses part of this problem by trying
was MTV started?”, but if it is only known to be a
to rationalize different but “similar” answers to the
Name then the inverted question “What <Name>
user, but does not extend to a general-purpose
was started in 1980?” could be too general to be ef-
equivalence identifier.
fective.
6 Summary
Bạn đang xem 2002) - BÁO CÁO KHOA HỌC IMPROVING QA ACCURACY BY QUESTION INVERSION DOCX