From: AAAI Technical Report SS-02-06. Compilation copyright © 2002, AAAI (www.aaai.org). All rights reserved.
Mining Answers from Texts and KnowledgeBases: Our Position
Bruce Porter,
Ken Barker,
Paul Navratil, Dan Tecuci,
JamesFan,
Peter
Peter Yeh
Department
of Computer
Sciences
Universityof Texasat Austin
pofter@cs.utexas.edu
Knowledge
SystemsGroup
BoeingMathematicsand Computing
Technology
peter.e.clark@boeing.corn
Recent advances in question answering from text have
shown that information retrieval,
natural language
processing and machinelearning techniques can go a long
wayin retrieving answersto certain types of questionsfrom
large bodies of text. Questions requiring morereasoning
and inference, or those whoseanswersrequire synthesis or
explanation are more difficult. Systemsthat reason over
domin-specific knowledge bases are capable of more
sophisticated behavior than answerretrieval systems, but
are expensive in terms of their knowledgerequirements.
The problem of answering difficult questions from the
knowledgeexix’essed in text can be attacked from both
ends: by improving answer retrieval from large corpora,
and by makingit possible for formal representations of
knowledgecontained in text to be authored more quickly
and easily.
Research
Interests
And Experience
Ourresearch group has interests in manyaspects across the
spectrum of this problem. Wehave experience in the
knowledgerepresentation issues involvedin building large
knowledgebases that capture knowledgecontained in text
as well as in simplifying the process of knowledgecapture
(Barker, Porter, and Clark 2001;Clark et al. 2001; Clark.
Thompson,and Porter 2000, Clark and Porter 1997; Fan et
al. 2001). Wehave workedin natural language generation
of explanations from knowledgebases (Lester and Porter
1997) and reasoning for question answering (Rickel and
Porter 1997; Clark, Thompsonand Porter 1999). Wehave
also investigated the relationship betweentext’s linguistic
form and its meaning (Barker 1998; Barker and
Szpakowicz1998).
Knowledge Capture Tools
Experts
Clark
quickly. One of the ultimate goals of our research is to
make it as easy for authors of text to encode formal
representationsof their knowledge
as it is for themto build
webpagesof it. Byputting intuitive knowledgeengineering
tools in the hands of the authors of text, we avoid the
bottleneck of knowledgeengineering having to go through
knowledge engineers. And by bringing the rate of
generating small knowledgebases closer to the rate of
producing textual documents, we increase the amountof
formally represented knowledgeavailable for specific
domains, allowing more sophisticated reasoning for
question answering.
Our work on accelerating the process of knowledge
engineering has three main parts: 1) building generic,
reusable representations to seed the knowledgeengineering
task; 2) developingtools for intelligent knowledgeaccess
and integration; 3) investigating methodsfor revising and
augmenting knowledge from information gleaned from
text.
We are building a library of reu~ble, cobble,
domain-independent knowledge components (Barker,
Porter, and Clark 2001). The library contains a small
numberof generic Entities, Events and Roles (Fan et al.
2001) and a restricted language for combiningthese to
represent more complex knowledge. As more complex
conceptsare encoded,these becomepart of the library to be
used as building blocks for morecomplexconcepts still.
Wecontinue to expand the coverage and granularity of
componentsto allow users to express moreknowledgewith
less effort.
Knowledge Integration
In order to ensure that this systemis intuitive to users
unfamiliar with knowledgeengineering, we are stressing
the importanceof allowing users to express themselvesin
familiar ways. They should not be burdened with the
representational requirementsof formal reasoning. To that
end, we are investigating waysto bridge the gap betweena
user’s unrestricted vocabularyand the restricted vocabulary
of our component library. In our development of the
componentiibrmy we have made use of dictionaries and
other linguistic resources to makethe library components
intuitive. Wehave encodedlinks to WordNet
(Miller 1990)
For Domain
Weare currently doing research on a project under
DARPA’s
Rapid KnowledgeFormation program that will
allow domainexperts to build knowledgebases easily and
~ght0 2000,American
Asm~ation
for Artificial lnteUigence
(w~vw.mmi.ors).
Allright8rc~rv~l.
8O
for each cogent, allowing the system to guide the user
to appropriate componentsthrough the WordNet
hierarchy.
Weare also working on a system to mapfrom a user’s
casual linguistic su’ueturesto the precise structures required
for reasoning in the knowledge base. To integrate
knowledgeexpressed by the user into a growingknowledge
base, we plan to allow users to express knowledge
imprecisely and to translate that impreciseexpressioninto
the exact form required in the knowledge base. For
example,we expectthe user to be able to refer to "’airplane
flaps" instead of the moreprecise "flap pan of the wing
part of an airplane"; "laser scalpel" instead of "laser
fulfilling the purposeof scalpel as instrumentof cutting".
Our knowledge
integration research will also investigate
howto use prior knowledge
to help in the interpretation of
imprecisely expressedknowledge.Integration also requires
the ability to expandor modifyexisting knowledge
based
ona user’sinput.
Text, KnowledgeModels AndQuestion
Answering
Anotherarea of our interest is in the interplay betweenthe
tasks of information extraction from text, model
construction, and question answering. Weare developing
an architecture in whichansweringsophisticated questions
is treated fundamentallyas a task of modelconstruction (as
opposedto informationretrieval), and in whichthese three
tasks are tightly integrated (as opposedto a "waterfall"
approach, in which information extraction results in a
model, and then the model is subsequently used for
question answering). In this architecture, question
answering provides requirements for a model of the
scenario of interest; background knowledge provides
candidate componentsfrom which that modelcan be built;
data suggests which of these candidate componentsare
relevant; and the partially built modelitself suggests new
questionsto pose to the text data. In other words,questions
guide informationretrieval; informationretrieval suggests
modelcomponents;and models suggest further questions.
Throughthis cycle, a coherent picture of a u:enario can
thus be built.
References
Barker, K., Porter, B., and Clark, P. 2001. A Library of
Genetic Concepts for ComposingKnowledgeBases. First
International Conference on KnowledgeCapture, 14-21.
Victoria.
Barker, K. 1998. Semi-AutomaticRecognition of Semantic
Relationships in English Technical Texts. PhD. diss.,
School of Information Technology and Engineering,
University of Ottawa.
Barker, K. and S. Szpakowicz 1998. Semi-Automatic
Recognition of Noun Modifier Relationships.
In
Proceedings of COLING-A
CL ’98, 96-102. Montr~l.
Clark, P., Thompson,
J., Barker, K., Porter, B., Chandhri,
V., Rodriguez,A., Thom~r~,
J., Mishra,S., Gil, Y., Hayes,
P., Reichherzer, T. 2001. Knowledge Entry as the
Graphical Assemblyof Components.First international
Conferenceon KnowledgeCapture, 22-29. Victoria.
Clark, P., Thompson,J., and Porter, B. 2000. Knowledge
Patterns.
Knowledge Representation
Conference
( KR’2000).
Clark, P., Thompson, J., and Porter, B. 1999. A
Knowledge-Based
Approachto Question-Answering.In the
AAAl’99 Fall Symposiumon Question.Answering Systems,
43-51, CA:AAAI
Press.
Clark, P. and Porter, B. 1997. Building Concept
Representations from Reusable Components.In AAAi’97,
369-376, CA:AAAI
Press.
Fan, J., Barker, IL, Porter, B., and Clark, P. 2001.
Representing Roles and Purpose. First lmemational
Conferenceon KnowledgeCapture, 38-43. Victoria.
J. Lester and potter, B. 1997. Developingand Empirically
Evaluating Robust Explanation Generators: The KNIGHT
Experiments.Computational
Linguistics 23( 1 ):65-101.
G. Miller ed. 1990. WordNet: An On-Line Lexical
Database.International Journalof Lexicography3(4).
J. Rickel and Porter, B. 1997. AutomatedModeling of
Complex Systems to Answer Prediction Questions.
Artificial Intelligence Journal93( 1-2):201-260.
Mining Answers FromTexts And Knowledge
Bases
Theresearch interests of our groupare a close matchto the
aims of the "Mining Answersfrom Texts and Knowledge
Bases" symposium. Wehave experience in acquiring
knowledgefrom text, in reasoning over knowledgebases
for answer/explanationgeneration and in building tools to
help end users (such as authors of specialized texts) build
knowledgebases quickly and easily. Webelieve all three of
these elementsare essential to the development
of the next
generation of question answeringsystems.
81