SECTION 3.3. A BENEFIT, SINCE IT GIVES MORE OPPORTUNITY FOR ENFORC-
39-41, 1995.
due to training limitations we focused on two re-
Moldovan, D. and Novischi, A, “Lexical Chains for
stricted evaluations. In the first we used a fixed
Question Answering”, COLING 2002.
question type, and showed that the error rate was
reduced by 36% and 30% on two very different cor-
Moldovan, D. and Rus, V., “Logic Form Transfor-
pora. In the second evaluation we focused on ques-
mation of WordNet and its Applicability to Ques-
tions whose direct answers were correct in the
tion Answering”, Proceedings of the ACL, 2001.
second position. 43% of these questions were sub-
Moriceau, V. “Numerical Data Integration for Co-
sequently judged correct, at a cost of only 3.7% of
operative Question-Answering”, in EACL Work-
originally correct questions. While in the future we
shop on Knowledge and Reasoning for Language
would like to extend the Constraints process to the
Processing (KRAQ’06), Trento, Italy, 2006.
entire answer candidate list, we have shown that ap-
plying it only to the top two can be beneficial as
Prager, J.M., Chu-Carroll, J. and Czuba, K. "Ques-
long as the second-place answers are at least a tenth
tion Answering using Constraint Satisfaction:
as numerous as first-place answers. We also showed
QA-by-Dossier-with-Constraints", Proc. 42nd
that the application of Constraints can improve the
ACL, pp. 575-582, Barcelona, Spain, 2004(a).
system’s confidence in its answers.
Prager, J.M., Chu-Carroll, J. and Czuba, K. "A
Multi-Strategy, Multi-Question Approach to
We have identified several areas where improve-
Question Answering" in New Directions in Ques-
ment to our system would make the Constraints
tion-Answering, Maybury, M. (Ed.), AAAI Press,
process more effective, thus getting a double benefit.
2004(b).
In particular we feel that much more attention
should be paid to the problem of determining if two
Prager, J., "A Curriculum-Based Approach to a QA
entities are the same (or “close enough”).
Roadmap"' LREC 2002 Workshop on Question
Answering: Strategy and Resources, Las Palmas,
7 Acknowledgments
May 2002.
This work was supported in part by the Disruptive
Radev, D., Prager, J. and Samn, V. "Ranking Sus-
Technology Office (DTO)’s Advanced Question
pected Answers to Natural Language Questions
Answering for Intelligence (AQUAINT) Program
using Predictive Annotation", Proceedings of
under contract number H98230-04-C-1577. We
ANLP 2000, pp. 150-157, Seattle, WA.
would like to thank the anonymous reviewers
Voorhees, E. “Overview of the TREC 2002 Ques-
for their helpful comments.
tion Answering Track”, Proceedings of the 11
th