Semantic Search, Question Answering systems, inferencing

Knowledge-Intensive Query Processing

Barbara Starr Vinay K. Chaudhri
SAIC Artificial Intelligence Center
San Diego SRI International
Barbara.H.Starr@cpmx.saic.com vinay@ai.sri.com

Adam Farquhar Richard Waldinger
Knowledge Systems Laboratory Artificial Intelligence Center
Stanford University SRI International
axf@hpp.stanford.edu waldinger@ai.sri.com

1 Introduction We conclude the paper by listing a few of the research
challenges faced in building such a system. The work
Innovative query interfaces to knowledge and database
described in the paper is preliminary and is aimed at
systems must go beyond simply returning the re-
suggesting directions for future work rather than at
quested information. They must be capable of produc-
describing in-depth technical results.
ing intentional answers when a description improves
the understanding of an answer [Mot94], producing
conditional answers when no one answer matches the 2 Crisis Management Benchmark
conditions of a query, and using ontological informa- The Crisis Management Benchmark (CMB) defines a
tion in processing a query. They should be able to collection of approximately 100 queries, their expected
call upon stand-alone reasoning modules that are most answers, and the knowledge sources that can be used
suitable for a given query. When answering a question to answer those queries [IACR97]. The CMB has been
involves reasoning beyond a simple lookup, the system motivated by the needs of a crisis analyst who is moni-
must be able to explain the answer to the user. toring a part of the world with the objective of predict-
We are building a question answering system with ing a crisis. The CMB has been defined in the context
these objectives. The heart of the system is a knowl- of a scenario involving a conflict among Persian Gulf
edge base (KB) and a collection of reasoning meth- nations. Many of the queries in the CMB are, however,
ods. The KB is being constructed by a combination of of general interest. Two features distinguishing the
manual and semiautomatic methods. The reasoning CMB from other benchmarks for measuring database
methods include conventional database query process- performance [Gra93] are as follows: (1) it measures the
ing, frame-based reasoning, and full first-order theo- knowledge content of the system and not just time to
rem proving. The performance of this system will be process a query, and (2) it is designed for queries that
tested on the Crisis Management Benchmark (CMB), require processing beyond a simple lookup or join of
which defines a collection of queries of interest to a relations.
crisis analyst.
We begin the paper by a description of the CMB. 2.1 Knowledge Benchmark
We describe the architecture of our system and then
For a system to be useful to a crisis analyst, it must
sketch some design ideas for two of its components.
have access to sufficiently broad geo-political knowl-
edge. The knowledge benchmark represents a user’s
The copyright of this paper belongs to the paper’s authors. Per-
mission to copy without fee all or part of this material is granted view of the system in the sense that it defines the do-
provided that the copies are not made or distributed for direct mains in which a user is interested. The CMB does not
commercial advantage. require the system to necessarily collect all the knowl-
Proceedings of the 5th KRDB Workshop edge in one place. We list here the domain areas in
Seattle, WA, 31-May-1998 which the knowledge must be encoded. For each do-
(A. Borgida, V. Chaudhri, M. Staudt, eds.) main area we give an example question from the CMB
http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS/Vol-10/ that is representative of that type of knowledge.

B. Starr, V. Chaudhri, A. Farquhar, R. Waldinger 18-1

World trade and economic information. How phase, the KB and the scenario will be kept fixed and
much oil does Japan purchase from the Persian Gulf new questions will be asked. In the second phase, the
states? KB will be fixed and the scenario will be changed. In
Geography and demographics. Which states bor- the third phase, new knowledge will be added to the
der the Persian Gulf? KB to address the new questions and the changed sce-
History of international behavior. Has Japan ever nario.
refused to trade with some other country?
Country policies. What is the US policy on illegal 3 System Architecture
immigrants?
Country capabilities. Is Iraq capable of refusing The components in our system can be classified into
inspection by UN officials? three categories: user interface, knowledge services,
International organizations. What is the Interna- and question answering. Many of the components have
tional Monetary Fund and who are its members? been developed by different research groups working
independently.
2.2 Processing Benchmark An overview of our system architecture is shown in
Figure 1. The components are held together by a com-
The processing benchmark represents an implemen- mon application programming interface: Open Knowl-
tor’s view of the system as it defines processing ca- edge Base Connectivity (OKBC) [CFF+ 98]. OKBC
pabilities that are necessary for answering the CMB interfaces for some of the components are already func-
questions. The CMB does not mandate any specific tional, whereas others are being constructed.
reasoning method that must be used for answering a Two kinds of user interface are supported by our
given question. The CMB queries can classified into system. HIKE is a form-based interface that allows a
the following categories. user to construct queries by using pull-down menus.
Retrieval queries. At what fraction of its current Forms are provided to construct any of the queries
sustainable capacity is Iran producing oil? in the CMB. START (Syntactic Analysis using Re-
What-if queries. Assuming constant production by versible Transformations) is a natural language inter-
Iran and others, would a 5% increase in production face that accepts queries in English [Kat97]. START
by Saudi Arabia have a positive or negative effect on generates a formal representation of a query and trans-
the economy of Iran? Would a 5% increase by Kuwait mits it for evaluation to one of the question answering
have as large an effect? systems. Automatic generation of a formal representa-
Analysis queries. Is Iraq capable of refusing inspection for an English query is supported only for a subset
tion by UN officials? of queries. Even though the goal of START is to ac-
cept arbitrary questions expressed in English, during
2.3 Operation of the Benchmark the course of the current project, the use of START
The initial evaluation criteria for the CMB are qual- will be restricted to the questions defined in the CMB.
itative, and the answers produced by the system will Knowledge services are provided by three compo-
be judged by a team of experts. Each answer will be nents: GKB-Editor, WebKB, and Ontolingua. GKB-
tested on the following criteria. Is the answer correct Editor is a graphical tool for browsing and editing
and accurate? Does the answer include any correct, large knowledge bases [KCP98]. It is primarily used for
nontrivial analysis which was not obvious? Are the manual knowledge acquisition. WebKB is a semiauto-
assumptions behind the knowledge appropriate? Does matic tool for extracting information from the World-
it constitute a realistic model for the question’s pur- wide Web (WWW) [CDF+ 98]. Given an ontology, and
pose? Are simplifications appropriate? Is the level a few examples of the information to be extracted, We-
of generality or detail appropriate? Is an explanation bKB can extract objects, relations and probabilistic
offered? Is it intelligible to a nonexpert? The bench- rules from the text sources on the Internet. The ex-
mark is being refined to include quantitative measures traction of knowledge is done in a semiautomatic fash-
to evaluate these aspects of the answers. ion. Ontolingua is the knowledge server and stores all
The system will be tested over a period of three the knowledge in the system [FFR97]. Since it has a
years. Each year will consist of a development pe- focal role in our architecture, we discuss it in detail.
riod of 11 months and a testing period of one month. Ontolingua is a tool for creating, evaluating, ac-
Each year will begin with a specific scenario and a cessing, using, and maintaining reusable ontologies. It
list of questions that are relevant for it. During the contains a collection of tools and services to support
11-month development period, a system will be devel- not only the development of ontologies by individuals,
oped to satisfy this scenario. During the final month, but also the process of achieving consensus on common
the system will be tested in three phases. In the first ontologies by distributed groups. These tools include a


SPOOK ATP
ATP
SPOOK
WWW
WWW

OKBC BUS
Training
Training WebKB
WebKB SNARK
SNARK
Data
Data

GKB
GKB SKC
SKC
Editor
Editor

Knowledge Ontolingua
Ontolingua
Engineer
HIKE
HIKE
START
START
WWW
WWW
Analyst

Figure 1: System Architecture: ATP – Abstract Theorem Prover, GKB-Editor – Generic KB Editor, HIKE –
form-based GUI, OKBC – Open Knowledge Base Connectivity, Ontolingua – Knowledge Server, SKC – Scalable
Knowledge Composition, SNARK – SRI’s New Automated Reasoning Kit, SPOOK – System for Probabilistic
Object-Oriented Knowledge

semiformal representation language that supports the ponents. The components may query the knowledge
description of terms in a representation language that server (Ontolingua) during query evaluation. The re-
is an extension of the Knowledge Interchange Format sulting answer is returned to HIKE and/or START.
(KIF) [GF92], browsing and retrieval of ontologies,
and facilities for translating ontologies into multiple 4 Knowledge Services
representation languages.
Two knowledge bases have been playing an active role
Question answering is supported by multiple rea-
in our work so far — the HPKB upper-level ontol-
soning methods. SNARK is a first order theorem
ogy (HPKB-UL) and the World Fact Book Knowledge
prover [SWL+ 94]. SPOOK is a reasoner based on
Base (WFBKB).
Bayes nets [KP97]. When answers are returned by
multiple reasoning methods or by using alternative
knowledge sources, the answers should be ranked be- 4.1 Upper Ontology
fore being presented to the user. Such a ranking will The HPKB-UL is the upper-level ontology of the Cyc
be done by SKC [WG95]. SKC is not yet a part of KB [LG89] augmented by some links to the Sensus on-
our system and will be integrated in the later phases tology [KL94]. It is being used in DARPA’s High Per-
of development. Finally, some of the question answer- formance Knowledge Bases (HPKB) program. HPKB-
ing is done by START that uses information retrieval UL provides a taxonomy of about 3000 terms and rela-
methods to process a query expressed in English on tions for general terms such as tangible-object, action,
a text-based source. We will illustrate a sample an- and transportation. It also defines some relations be-
swer produced by one of these components later in the tween them, such as the starting time of an event, the
paper. relationship between an object and its parts, and the
Let us consider an example of how a question is an- borders of a geographic region. The X3T2 working
swered by our system. A user poses her query by us- group of ANSI has adopted HPKB-UL as the current
ing either HIKE or START. In either case, the query draft for a standard upper ontology.
is mapped to a formal representation expressed in a There are at least two advantages in using HPKB-
language which is an extension of KIF. The query is UL. First, for formalizing the CMB questions, we need
shipped to one or more of the question answering com- vocabulary. HPKB-UL provides a significant subset


of the vocabulary necessary for this purpose. Sec- a broad range of information about the countries and
ond, since our knowledge base is being developed by territories of the world.1 The fact book includes geo-
multiple research groups, we need a terminology that graphical, economic, demographic, and some historical
will help us in combining KBs developed by different information.
groups. Using a common ontology makes the task of We have been augmenting HPKB-UL and linking it
sharing knowledge easier. to WFBKB. HPKB-UL stops well above many of the
For formalizing the CMB queries, we used the concrete terms that appear in the fact book, such as
following approach: map nouns to concepts, verbs bauxite mining, the food-and-beverage industry, sandy
to events and actions, and adjectives to quality at- beaches, ethnic minorities, and spoken languages. We
tributes, identify individuals, and finally; specify tem- are working to construct a richer ontology that spans
poral and spatial information. If a necessary term is the substantial gaps between the HPKB-UL concepts
not found in HPKB-UL, we extend HPKB-UL appro- and the terms that are introduced by the World Fact
priately. As an example, consider the question: Book. We expect the WFBKB to provide knowledge
necessary to answer many of the CMB questions.
Has post-Shah Iran launched ballistic
missiles in wartime? 5 Question Answering
Our formalization of this query is as follows. For the analysis queries defined in the CMB, a sim-
ple yes or no answer is not appropriate. Instead, the
(and system must return a descriptive answer and provide
(attack ?act) some justification for that answer. Let us consider an
(performed-by ?act Iran) example of a conditional answer for an analysis ques-
(device-used ?act ballistic-missile) tion. This answer was produced using SNARK, which
(later-than (start-of ?act) is one of the question answering components.
(start-of post-shah-iran))) Consider the following question: What will be the
likely position of Iraq on allowing inspection by UN
In this formalization, performed-by, device-used, officials? For an analysis question like this, usually we
later-than, and start-of are predicates defined in want to know more than what the question literally
HPKB-UL. For example, device-used relates an ac- requires. For this question, on one hand, we mean that
tion to the device that was used in performing it. if Iraq is likely to refuse inspection, we would like to
The predicate attack represents a collection of ac- know the reasons. On the other hand, if Iraq is likely to
tions in which an agent attacks another agent, and go along with such an inspection, we need confirmation
post-shah-iran is a constant representing the time of it; simply failing to find a proof that Iraq can refuse
interval after Shah of Iran. This formalization does UN inspection does not necessarily imply that it will—
not explicitly represent wartime as it assumes that an there are many true things we cannot prove. In either
attack is performed only during wartime. case, we would like some indication of the reasoning
Since all the question answering components sub- by which the conclusion was reached.
scribe to the same ontology, the above question can Suppose we have a domain knowledge rule in the
be sent for evaluation to any of them. A component KB which states that Iraq is likely to refuse the in-
may decide to transform this representation into an spection if it has political support from Russia: if (sup-
alternative representation which is more efficient for ports Russia Iraq) then (refuses Iraq UN-Inspection)
evaluation by it. else (delays Iraq UN-Inspection)
In finding a proof for this question, SNARK per-
4.2 World Fact Book KB forms a case analysis, depending on whether Iraq has
The World Fact Book KB, being developed by the support from Russia. If Iraq has support from Russia,
Knowledge Systems Laboratory at Stanford is a sub- it returns an answer (refuses Iraq UN-Inspection). If
stantial knowledge base covering basic geographic, eco- Iraq does not have support from Russia, it returns an
nomic, political, and demographic knowledge about answer (delays Iraq UN-Inspection). If it cannot deter-
the world’s nations. The goal of the project is to pro- mine whether Iraq has support from Russia, it returns
vide a useful knowledge resource, to explore ways for a conditional answer stating that if Iraq has support
structuring large knowledge bases, and to develop a from Russia, it can refuse UN inspection; otherwise, it
knowledge base large enough to stress existing knowl- will try to delay it. The capability to produce condi-
edge representation systems. tional answers is not supported in any of the existing
The primary source for the World Fact Book knowl- 1 For more information on the World Fact Book, see

edge base is the CIA World Fact Book, which collects http://www.odci.gov/cia/publications/nsolo/wfb-all.htm.


query processing systems. • What are the principles and techniques for de-
SNARK also prints an explanation for the answer signing a large knowledge base that would enable
produced by it. The explanation shows all the axioms knowledge reuse?
used and the intermediate inference steps. Each in-
ference step shows the axioms, inference method, and • What is a good interface for allowing a user to
rewrites used in that step, the conclusion derived, and construct queries by using ontological informa-
the current answer term. If an English description of tion?
an axiom is available, it is shown. The output is pro- • What techniques are useful for explaining the an-
duced in the HTML format making it easier for a user swers of a system that uses multiple reasoning
to navigate amongst several inference steps. methods?
6 Summary and Conclusions Acknowledgments
We have presented initial design ideas of an innovative At Science Applications International Corporation
query interface for a knowledge and database system. (SAIC), this work was supported by a DARPA/SSC
The system is capable of producing conditional an- contract entitled Knowledge-Base Technology Devel-
swers when no one answer matches the conditions of a opment and Integration (Contract Number N66007-
query, and uses ontological information in processing 97-C-8546). At SRI International, it was supported
a query. The heart of the system is a knowledge base by a DARPA contract entitled Ontology Construction
(KB) and a collection of reasoning methods. The KB Toolkit (Contract Number N66001-97-C-8550). At
is being constructed by a combination of manual and Stanford University, it was supported by a DARPA
semiautomatic methods. The reasoning methods in- contract entitled Large-Scale Repositories of Highly
clude conventional database query processing, frame- Expressive Reusable Knowledge (Contract Number
based reasoning, and full first-order theorem proving. N66001-97-C-8554).
The performance of this system will be tested on the
Crisis Management Benchmark (CMB) that defines a References
collection of queries which are of interest to a crisis
analyst. [CDF+ 98] M. Craven, D. DiPasquo, D. Freitag,
We believe that innovative query interfaces such as A. McCallum, T. Mitchell, K. Nigam, and
the one described here represent a major advance in S. Slattery. Learning to Extract Symbolic
the query processing capabilities of current knowledge Knowledge from the World Wide Web. In
and database systems. They open up several challeng- Proceedings of the National Conference on
ing research problems that must be addressed. We Artificial Intelligence, July 1998. To ap-
believe that the following problems are fundamental pear.
to enabling the construction of such interfaces.
[CFF+ 98] Vinay K. Chaudhri, Adam Farquhar,
• How can one take a KB such as WFBKB or Richard Fikes, Peter D. Karp, and
HPKB-UL developed for one purpose and use it James P. Rice. OKBC: A Foundation for
in a different context? Knowledge Base Interoperability. In Pro-
ceedings of the National Conference on Ar-
• How can we reformulate a KB into a form that tificial Intelligence, July 1998. To appear.
allows efficient evaluation of a query with a new
reasoner? [FFR97] Adam Farquhar, Richard Fikes, and
James P. Rice. A Collaborative Tool
• When, and in what forms are conditional answers for Ontology Construction. Interna-
useful? tional Journal of Human Computer Stud-
• How effective can be a generic API such as OKBC ies, 46:707–727, 1997.
in integrating diverse technology components into
[GF92] Michael R. Genesereth and Richard E.
one system?
Fikes. Knowledge Interchange Format,
• Given a query, on what basis should it be dis- Version 3.0 Reference Manual. Technical
patched to a component subsystem and how to Report Logic-92-1, Computer Science De-
combine the results of sub-queries returned by dif- partment, Stanford University, Stanford,
ferent systems? CA, 1992.

• How can we quantitatively measure the perfor- [Gra93] James N. Gray. The Benchmark Handbook
mance of answering analysis questions? for Database and Transaction Processing


Systems. Morgan Kaufmann Publishers,
1993.
[IACR97] IET, Alphatech, Paul Cohen, and Pa-
cific Sierra Research. HPKB year 1
end-to-end challenge problem specification,
version 1.1. Technical report, Informa-
tion Extraction and Transport (IET) Inc.,
Rosslyn, Virginia, December 1997. See
http://www.iet.com/Projects/HPKB/.
[Kat97] Boris Katz. From Sentence Processing to
Information Access on the World Wide
Web. In AAAI Spring Symposium on Nat-
ural Language Processing for the World
Wide Web, Stanford, CA, 1997.
[KCP98] Peter D. Karp, Vinay K. Chaudhri, and
Suzanne M. Paley. A Collaborative En-
vironment for Authoring Large Knowledge
Bases. Journal of Intelligent Information
Systems, 1998. To appear.
[KL94] K. Knight and S. Luk. Building a Large-
Scale Knowledge Base for Machine Trans-
lation. In Proceedings of the National Con-
ference on Artificial Intelligence, Seattle,
WA, August 1994.
[KP97] D. Koller and A. Pfeffer. Object-Oriented
Bayesian Networks. In Proceedings of the
13th Annual Conference on Uncertainity in
AI (UAI), Providence, RI, August 1997.
[LG89] Douglas B. Lenat and R.V. Guha. Building
Large Knowledge-based Systems: Repre-
sentation and Inference in the Cyc Project.
Reading, MA, Addison-Wesley Publishing
Co., 1989.
[Mot94] Amihai Motro. Intensional Answers
to Database Queries. IEEE Transac-
tions on Knowledge and Data Engineering,
6(3):444–454, 1994.
[SWL+ 94] M. Stickel, R. Waldinger, M. Lowry,
T. Pressburger, and I. Underwood. De-
ductive Composition of Astronomical Soft-
ware from Subroutine Libraries. In Pro-
ceedings of the Twelfth International Con-
ference on Automated Deduction (CADE-
12), pages 341–355, June 1994.
[WG95] Gio Widerhold and Michael Genesereth.
The Conceptual Basis for Mediation Ser-
vices. In Proceedings of the Interna-
tional Conference on Cooperative Informa-
tion Systems, Vienna, Austria, May 1995.


Semantic Search, Question Answering systems, inferencing

More Related Content

Similar to Semantic Search, Question Answering systems, inferencing (20)

More from Barbara Starr (18)

Recently uploaded (20)

Semantic Search, Question Answering systems, inferencing