Explanation for Query Answering

Download Report

Transcript Explanation for Query Answering

Explanation for Query
Answering (and Ontology Env
Highlights)
Deborah McGuinness
Associate Director and Senior Research Scientist
Knowledge Systems Laboratory
Stanford University
[email protected]
With support from Richard Fikes, Jessica Jenkins, Gleb
Frank, Yulin Li, Steve Wilder
1
McGuinness
Mar 26, 2002
Outline
 Motivation
 Logical Foundation
 Explanation Needs
 Explanation Issues
 Explanation Status in some implemented
systems
 Ontology Evolution Environments
2
McGuinness
Mar 26, 2002
What is an explanation?
From Merriam Webster:
1 a : to make known
b : to make plain or understandable
2 : to give the reason for or cause of
3 : to show the logical development or relationships
of intransitive senses : to make something plain or
understandable
For operational computer systems, we can consider
an explanation to be a justification of system belief
or a justification of a statement.*
3
McGuinness
Mar 26, 2002
A Few Simple Queries
 Find (a/some) school(s) in Santa Clara County
 Find (a/some) university(ies) in Santa Clara County
 Find (a/some) source(s) of advanced degrees in Santa Clara County
 Filter by criteria (best ranking, cheapest tuition, closest
geography,…)
one view is why is xx an instance (or subclass) of a compound noun
phrase
4
McGuinness
Mar 26, 2002
5
McGuinness
Mar 26, 2002
Name
matches
but is
provably
too
specific
6
McGuinness
Mar 26, 2002
Good needs a defn
– prayed for,
ranking of students,
samaritan deeds,
…?
7
McGuinness
Mar 26, 2002
Structured Information can help…
 Search in finding the direct (structured) piece(s) of information
desired
 Providing foundation for
 Declarative basis for justification
 Comparison of answers
 Pruning by matching answers
 Clearly defining terms
…
8
McGuinness
Mar 26, 2002
Simple Motivating Example
 Why is Stanford a school in Santa Clara County
(or
why is Stanford an instance of the class school
Stanford -> School
and why is Stanford in Santa Clara County)?
Stanford -> (mincardQ 1 hasLocation SantaClaraCounty)
 Why is Stanford a (good/highly ranked) school in Santa Clara
County
beyond above, what makes the system believe that it is “good” or
“highly ranked”
9
McGuinness
Mar 26, 2002
Some Possible “Justifications”
 Because I told you so and I am always right
 Because (authoritative) source XYZ says so directly
 Because source XYZ and PQR says so (& no awareness of
contradictions)
 Because authoritative reasoner deduced it
 Because authoritative reasoner deduced it as follows:
Stanford is believed to be a University (stated in in authoritative
source S1) and
all universities are schools (stated in authoritative source S2)
and
if x isa y and y is a z, then x isa z (from axiom set A1)
10
McGuinness
Mar 26, 2002
Justifications
 Necessary components for a simple justification foundation:
 Inference rule
 Bindings for variables in the inference rule
 Source of information
 Sample Inference Rule
Inheritance
X => Y
Y=> Z
X => Z
 Note: Follow-up questions follow from logical form of rule
11
McGuinness
Mar 26, 2002
Example
Inheritance
X => Y
Y=> Z
X => Z
(inference list available for some systems: CLASSIC, DAML+OIL, …)
Why is Stanford an instance of School?
Sample Application:
Rule: Inheritance
Bindings: from query X=Stanford, Z=School,
Y=University
Source
X=>Y
DB/KB 1
Y=>Z
DAML ontology library organizations
12
McGuinness
Mar 26, 2002
One Approach
 For any statement expressible in your kb, allow a user to ask why
the system believes it
 Some system (possibly a separate explainer system) generates a
justification for the statement
(alternative, present a list of inferences(SHAKEN), present some
form of (abstracted) proof tree,…)
 Justifications are first class objects so that follow-up questions
may be asked about statements (and about inference rules)
 Justifications may be viewed as portable, exchangeable objects
that facilitate system interaction
13
McGuinness
Mar 26, 2002
Some history
 Explanation system for normalize/compare description logic-
based reasoner (CLASSIC)
 Explanation design for tableaux reasoning system (FACT/CRACK)
 Explanation design and implementation for model elimination
theorem prover (ATP)
 Explanation design for hybrid reasoning system (JTP) with partial
implementation
14
McGuinness
Mar 26, 2002
Needs for Justifications*
 People need a justification in order to:
 Trust the validity of a statement
 (Re)use an answer
 Share answers and/or share deductions across systems **
 Debug surprising information
 Maintain knowledge bases
 Train people on knowledge bases
 Train programs (e.g., Disciple)
 Provide customer service support
 Present contradictions more usefully (defend how system determined
contradictions)
 Support for kb updates due to information changes
 Distinguish between types of knowledge (e.g., lookup vs. deduction)
 Support adjudication between hypotheses (helps in handling non-
monotonically changing kbs, conflicting kbs, etc.)
*input from RKF program
McGuinness
Mar 26, 2002
15
Challenges/Issues
 Proof traces quickly become hard to understand and large
 Although techniques exist to help manage large presentations of
proofs (atomic justifications, stepwise explanation, automatic
follow-up questions, …) these may not be enough
 Management techniques for presenting underlying reasoning may
not be appropriate (e.g., tableaux methods are unintuitive….)
 Explanations may be needed at many levels of abstractions – for
novice users, advanced users, simple agents, sophisticated
agents, etc.
 Portable proofs need to understand the other systems inference
rules
 Extra challenges from contradictions, distributed knowledge
sources, distributed authoring, sound & complete vs. incomplete
vs. heuristic strategies, …
…
16
McGuinness
Mar 26, 2002
Architecture
Reasoning
(JTP) module
Proof
Reasoner-specific
proof
•Abstraction
Applications
•Proof Pruning
•Context-specific
info
Reasoner-independent
portable proof
Uses DAML+OIL
Ontology
Human
user
DAML-enabled
application
DAML proof
17
McGuinness
Mar 26, 2002
Proof Object
<daml:Class rdf:ID="Proof">
<rdfs:comment> A Proof object contains information about an entire proof. </rdfs:comment>
<daml:intersectionOf rdf:parseType="daml:collection">
<daml:Restriction>
<daml:onProperty rdf:resource="#conclusion"/>
<daml:toClass rdf:resource="http://www.ksl.Stanford.EDU/projects/DAML/Proof/pce.daml#Sentence"/>
</daml:Restriction>
<daml:Restriction>
<daml:onProperty rdf:resource="#conclusion"/>
<daml:cardinality> 1 </daml:cardinality>
</daml:Restriction>
<daml:Restriction>
<daml:onProperty rdf:resource="#inference"/>
<daml:toClass rdf:resource="#Inference"/>
</daml:Restriction>
<daml:Restriction>
<daml:onProperty rdf:resource="#inference"/>
<daml:cardinality> 1 </daml:cardinality>
</daml:Restriction>
<daml:Restriction>
<daml:onProperty rdf:resource="#subProof"/>
<daml:toClass rdf:resource="#Proof"/>
</daml:Restriction>
</daml:intersectionOf>
</daml:Class>
*http://www.ksl.Stanford.EDU/projects/DAML/Proof/portable-proof.daml
McGuinness
Mar 26, 2002
18
Status
•JTP provides proof trees and step-wise viewer (aimed at
developers)
•Prototype viewer available for demonstration of
reasoner-independent proofs
•Tested on some DAML+OIL ontologies (wines KB)
•DAML+OIL proof ontology in process (uses
DAML+OIL axioms)
•Sample proofs available using proof ontology
•Internal testing continuing
19
McGuinness
Mar 26, 2002
Chimaera – A Ontology
Environment Tool
An interactive web-based tool aimed at supporting:
•Ontology analysis (correctness, completeness, style, …)
•Merging of ontological terms from varied sources
•Maintaining ontologies over time
•Validation of input
• Features: multiple I/O languages, loading and merging into multiple
namespaces, collaborative distributed environment support, integrated
browsing/editing environment, extensible diagnostic rule language
• Used in commercial and academic environments, basis of some
commercial re-implementations (Ontobuilder/Ontoserver,…)
• Available as a hosted service from www-ksl-svc.stanford.edu
• Information: www.ksl.stanford.edu/software/chimaera
McGuinness
Mar 26, 2002
20
Our KB Analysis Task
 Review KBs that:

Were developed using differing standards

May be syntactically but not semantically validated

May use differing modeling representations
 Produce KB logs (in interactive environments)
 Identify provable problems
 Suggest possible problems in style and/or modeling
 Are extensible by being user programmable
 Are explainable
21
McGuinness
Mar 26, 2002
Chimaera finds terms
that were used but
not defined
Chimaera warns of
cycles in class-subclass
hierarchies
Chimaera warns of
missing
documentation
24
McGuinness
Mar 26, 2002
Chimaera notes values
that don’t or may not
satisfy type constraints
25
McGuinness
Mar 26, 2002
Chimaera as a KB Merging Tool
Combine independently developed
KBs with overlapping content
Car 
 Mining
 Coal-Mining
Auto 
 Mining
Automobile   Coal-Mining
(Add relationships)
(Merge terms)
 
Recognize and explain Inconsistencies
identify redundancies,
Check for coherence
Point out possibly related terms
(similar names, structure, etc.)
And otherwise focus attention of user on points of high potential payoff
26
McGuinness
Mar 26, 2002
Chimaera Simple Merging
Other things like acronym expansion,
subclass link suggestions, etc.
27
McGuinness
Mar 26, 2002
Conclusion
 Explanation is becoming more critical in many settings
 Logical Foundation (Proof-theoretic) can be used as a foundation
on which to provide:
 Interactive
 Portable
 Prunable
 Follow-up ready
explanations
 Ontologies can be exploited in explanations as well as intelligent
applications
 Ontology environments are becoming ready for use by non KR
experts
28
McGuinness
Mar 26, 2002
Some Pointers
 Explanation Foundations:
http://www.research.att.com/~dlm/papers/thesis-abstract.html
http://www.research.att.com/~dlm/papers/explain-abstract.html
http://www.ksl.stanford.edu/people/dlm/papers/explainingALC.ps
 Ontologies Come of Age Paper:
http://www.ksl.stanford.edu/people/dlm/papers/ontologies-come-of-ageabstract.html
 Ontologies and Online Commerce Paper:
http://www.ksl.stanford.edu/people/dlm/papers/ontologies-and-onlinecommerce-abstract.html
 Chimaera:
http://www.ksl.stanford.edu/people/dlm/papers/kr00-abstract.html
http://www.ksl.stanford.edu/people/dlm/papers/aaai00-abstract.html
 DAML+OIL: http://www.daml.org/
http://www.w3.org/TR/daml+oil-reference
http://www.w3.org/TR/daml+oil-walkthru/
McGuinness
Mar 26, 2002
29
Merging
 Heuristic Set of techniques for suggesting merges and new
relationships:
 Acronym expansion
 X-Y is likely to be a subclass of Y
 Other name similarity
 Similarity / Difference options
30
McGuinness
Mar 26, 2002
Diagnostic results are
saved for use when
needed
32
McGuinness
Mar 26, 2002
(rdfs:subClassOf ?csub ?cl)
(rdf:type ?inst ?csub))
(rdfs:subClassOf ?csub ?cl)
(rdf:type ?inst ?csub)
{{
[/.3.1]:
(<= (rdf:type ?inst ?cl)
(rdfs:subClassOf ?csub ?cl)
(rdf:type ?inst ?csub))
<<
Direct Assertion
[/.3.2]:
(rdf:type wines-short2.daml:CRAB wines-short2.daml:NON-OYSTER-SHELLFISH)
<<
Direct Assertion
[/.3.3]:
(rdfs:subClassOf wines-short2.daml:NON-OYSTER-SHELLFISH
wines-short2.daml:SHELLFISH)
<<
Direct Assertion
33
McGuinness
Mar 26, 2002
Extras
34
McGuinness
Mar 26, 2002
35
McGuinness
Mar 26, 2002
Next Steps
 Working Group?
 Maintain Needs and Issues list and prioritize
 Maintain RKF researcher list – input from Clark, interest from
Tecuci, Mehrotra
 Publish inference ruleset(s) for portable proofs
…
36
McGuinness
Mar 26, 2002
Sometimes it is better… Google
Searched the web for universities in santa clara county. Results 1 - 10
of about 13,500.
Planning ... San Jose State University; Santa Clara University; Silicon
Valley ... Sonoma County. Empire College; Santa Rosa Junior College; ...
Colleges and Universities Beyond the Bay ...
**
Santa Clara County Network for a Hate-Free Community ... *Universities
include Stanford University, Santa Clara University, San Jose State ...
Copyright 2001 Santa Clara County ISD Web Development Team.
College and University Links - Santa Clara County Library ... University
Links This guide to colleges and universities can be sorted by name,
state, and type ... Santa Clara County Library, 1095 N. 7th Street, **
Master Gardeners of Santa Clara County – Mission ... system was
established to encourage state universities to better serve the
agricultural ... Santa Clara County As volunteer representatives of the
Related Links - Santa Clara Transportation Authority ... Clara, County of
37
Santa Clara County Office of
Education
Saratoga
...
Page
Downtown
San
McGuinness Mar 26, 2002
Sometimes it is worse… Google
Searched the web for sources of advanced degrees in santa clara
county. Results 1 - 10 of about 1,560.
Catholic Charities Santa Clara ... by professional counselors with
advanced degrees in social work or ... 2002 Catholic Charities of Santa
Clara County, 2625 Zanker Road, San Jose ...
Community Health Library of Los Gatos - Web Resources ... Santa Clara
County Department of Environmental ... a jump station for sources of
information on ... research and offering advanced degrees in the ...
San Francisco East Bay Education and Research ... for graduate degrees.
... second to Santa Clara and that ... laboratories, Alameda County is a ...
energy sources, combustion technology ... electronics, advanced
materials ...
Santa Clara University : Convocation Speech 2000 : The Valley ... ... The
reason Santa Clara has started the ... can become sources of threat and
... come here for advanced degrees, immigrants make ... idyllic Marin
County don't ...
38
McGuinness
Mar 26, 2002
Geog-Containmnt:
GeogContains(A,C)
GeogContains(C,D)
GeogContains(A,D)
39
McGuinness
Mar 26, 2002