IKRIS Challenges, Schedule, and Goals

advertisement
Interoperable Knowledge Representation
for Intelligence Support (IKRIS)
A challenge problem project on knowledge representation
sponsored by DTO
Technical Team Leaders
Prof. Richard FikesDr. Christopher Welty
Knowledge Structures Group
Artificial Intelligence Laboratory (KSL)
Stanford UniversityIBM Corporation
Knowledge Systems,
T. J. Watson Research Center
Northeast Regional Research Center Leaders
Dr. Brant Cheikes (MITRE)
Dr. Mark Maybury (MITRE)
Government Champions
Steve Cook (NSA)Jean-Michel Pomarede (CIA)
John Donelan (CIA)John Walker (NSA)
2/7/06
Knowledge Representation and Reasoning

Knowledge Representation

Encoding descriptions –
> That correspond in some coherent way to a world of interest
> Are usable by a computer to make conclusions about that world

Primary areas of activity:
> Developing declarative formalisms for expressing knowledge
– Mostly “general-purpose” languages (e.g., First-order logic)
> Encoding knowledge (knowledge engineering)
– Mostly identifying and describing conceptual vocabularies (ontologies)

Reasoning

Automating coherent creation of new knowledge from existing knowledge

Primary areas of activity:
> Development and analysis of computational reasoning methods
– Task-specific methods such as planning, scheduling, diagnosis, …
– Methods for managing reasoning such as hybrid reasoning, …
2
Challenge Problems for the IC

DTO (Disruptive Technology Office) funded challenge problem projects

Focus is on problems that require collaboration to solve

DTO recognizes knowledge representation (KR) as a critical technology

IKRIS is addressing two KR challenges

Enabling interoperability of KR technologies
> Developed by multiple contractors
> Designed to perform different tasks

Interoperable representations of scenarios and contextualized knowledge
> To support automated analytical reasoning about alternative hypotheses
3
Hypothesis Modeling and Analysis

Tools for modeling and analyzing alternative hypothetical scenarios
What
happened?
…
What’s the current
situation?
What’s going
to happen?

Models enable automated reasoning to accelerate and deepen analysis


Consistency and plausibility checking, deductive question-answering,
hypothesis generation, …
Requires sophisticated knowledge representation technology

Actions, events, “abnormal” cases, alternatives, open-ended domains, …
4
Interoperable KR Technology


No one representation language is suitable for all purposes

Technology development necessarily involves exploring alternatives

Differing tasks require differing representation languages
So, modules using differing KR languages need to be interoperable


Requires enabling modules to use each other’s knowledge
The IKRIS approach to achieving interoperability –

Select and refine a standard knowledge interchange language
> Called IKRIS Knowledge Language (IKL)


Develop translators to and from IKL
Each system module will then –

Use its own KR language internally

Use IKL for inter-module communication

Translate knowledge to and from IKL as needed
5
IKRIS Organization

Prime Contractor – MITRE, Brant Cheikes and Mark Maybury

Technical Team Leads – Fikes (Stanford KSL) and Welty (IBM Watson)

Working Groups

Interoperability – Pat Hayes, University of West Florida
Chris Menzel, Michael Witbrock, John Sowa, Bill Andersen, Deb McGuinness, …

Scenarios – Jerry Hobbs, Information Sciences Institute
Michael Gruninger, Drew McDermott, David Martin, Selmer Bringsjord, …

Contexts – Selene Makarios, Stanford KSL
Danny Bobrow, Valeria de Paiva, Charles Klein, David Israel, …

Evaluation – Dave Thurman, Battelle Memorial Institute

Technology Transfer – Paula Cowley, Pacific Northwest National Laboratory

Translation technology and example translators – Stanford KSL

Government Champions –
Steve Cook, John Donelan, Jean-Michel Pomarede, John Walker
6
IKRIS Project Schedule

Preparation – January - April, 2005

Kickoff Meeting – April 2005


Established working groups and their charters

Developed work plan and began work in each group
Working groups – April 2005 through April 2006


Evaluation – January through September 2006



Producing results and planning technology transfer
Iterative evaluation of workshop results
Second face-to-face workshop – April 2006

Finalize and coordinate results of working groups

Finalize plans for technology transition and for completing evaluation
Technology transition – April through September 2006

Initiation of planned transition activities
7
FOL Knowledge Interchange Languages



KIF (Knowledge Interchange Format)

ASCII Lisp-style syntax

No formal model theory

Pre-WWW/XML/Unicode

Included a set theory, definition language, etc.

Subset became de facto AI/KR standard

Subset developed as a proposed ANSI standard
CL (Common Logic)

Based on KIF

Formal model theory (based on Menzel/Hayes)

Abstract syntax

“Web savvy”

In final stages of becoming an ISO standard
IKL (IKRIS Knowledge Language)

Extension of CL

Extensions include propositions, quoting
8
CLIF Syntax for IKL

Designed for use on an open network

Names are made globally unique by –
> Including a URI as part of the name
> Using the XML namespace conventions to abbreviate names

Universal quantifiers can be restricted by a unary predicate
E.g., “All humans own a car.”
(forall ((x isHuman)) (exists ((y Car)) (Owns x y)))

Existential quantifiers can be restricted by a number
Cool!
E.g., “All humans have as parts 10 toes.”
(forall ((x isHuman))
(exists 10 (y) (and (Toe y) (PartOf y x))))
9
Examples of CL/IKL Expressivity

Relations and functions are in the universe of discourse
E.g.,

(owl:inverseOf parent child)
A relation or function can be represented by a term
E.g.,
x)))
(forall (x y r) (iff (r x y) ((owl:inverseOf r) y
Given the above axiom,
((owl:inverseOf parent) Arthur Ygrain)
is equivalent to –
(child Arthur Ygrain)
and entails
(parent Ygrain Arthur)
1
Examples of CL/IKL Expressivity

A unary relation could be allowed to take multiple arguments

So that, e.g., (isHuman Fred Bill Mary)
abbreviates
(and (isHuman Fred) (isHuman Bill) (isHuman Mary))

We might call such relations “Predicative”
E.g., assert (Predicative isHuman)

What it means to be Predicative could be axiomatized as follows –
(forall (r) (if (Predicative r)
(forall (x y z) (iff (r x y z)
(and (r x) (r y) (r z))))))

Predicative itself could be Predicative –
WOW!
(Predicative Predicative)
allowing such abbreviations as
(Predicative isHuman isAnimal isFish)
1
Examples of CL/IKL Expressivity

Sequence names


Allows a sentence to stand for an infinite number of sentences, each
obtained by replacing each sequence name by a finite sequence of names
A sequence name is any constant beginning with “…”
E.g., the general axiom for Predicative is as follows:
(forall (r) (if (Predicative r)
(forall (x y ...) (iff (r x y ...)
(and (r x) (r y ...))))))

Function “list” and relation “isList” are predefined as follows:
(forall (...) (isList (list ...)))
1
Extending CL to Include Propositions

Goal: Support representation of contextualized and modal knowledge

Achieved by making propositions first-class entities in IKL
> Refer to them by name, quantify over them, have relations between them and
other entities, define functions that apply to them, …
> Technically, a proposition is a 0-arity relation
BAM!

The operator that is used to denote propositions

that takes a sentence as an argument
E.g., (that (Married Ygrain Uther))

A that expression denotes the proposition expressed by its argument
E.g., (that (Married Ygrain Uther))
is a name, denoting the proposition that Ygrain and Uther are married

Issue: When are two propositions equivalent?
E.g., does (and a b) name the same proposition as (and b a)?
 IKL provides a propositional equivalence relation, but does not build it in
 General propositional equivalence is undecidable
1
3
Relativizing Names in IKL

In some cases, the denotation of logical names needs to be
relativized
(believes Mary
(that (forall (x) (if (Child x Joe) (Male x))))
… but what if Mary thinks Frank is Joe?

Need to talk about “mary’s version of Joe”


Special class of functions: quoted names
‘name’ is a function that returns the “right thing”
> (‘Joe’) is just Joe
> (‘Joe’ Mary) would be Frank (what ‘Joe’ denotes to Mary)
> E.g.
(believes (Mary
(forall (x) (if (Child x (‘Joe’ Mary)) (Male x))))
1
4
IKRIS Language Translators
 Developing
2-way IKL translators for several KR
languages

OWL, RDF, KIF, CycL, Slate/MSL
 API
for parsing/generating IKL
 Design goal: “round trip” compliance

Significant new work in KR
 Major challenge to round trip OWL
> Simple “embedding” in IKL
> Requires “axiom patterns” and meta-data
– (forall (P Q)
(=>
(forall (x) (=> (P x) (Q x)))
(owl:subclassOf P Q)))
1
5
Interoperable Scenarios

IKRIS is addressing two KR challenges

Enabling interoperability of KR technologies
> Developed by multiple contractors
> Designed to perform different tasks

Interoperable representations of scenarios and contextualized knowledge
> To support automated analytical reasoning about alternative hypotheses

Developing an interoperable representation for processes

Includes –
> Time points, time intervals, durations, clock time, and calendar dates
> Events and relationships that overlap in time and interact
> Process constructs, preconditions, states, etc.
1
An Interlingua for Processes
PSL
SWSL/
FLOWS
OWL-S
inter-theory
DONE!
SPARK
ResearchCyc
1
The Scenarios Inter-Theory (ISIT)


The Scenarios Working Group is producing an IKL inter-theory

vocabulary

Bridging axioms to other vocabularies

Trigger axioms for making optional representational commitments
The inter-theory vocabulary includes –

The OWL time ontology
> Terminology for clock time, calendars, intervals, points, etc.

Terms such as the following to describe processes:
> Eventuality
> Precondition
> Event
> EventType
> EventualityType
> PreconditionToken
> State
> FluentFor
> Subevent
> Effect
> StateType
1
ISIT Bridging Axioms

Example bridging axioms to Cyc for Event and EventType:

“For every EventType x, there is a Cyc subclass of cyc:Event that has the
same instances as x”
(forall ((x EventType)))
(exists (y) (and (cyc:genls y cyc:Event)
(forall (e) (iff (cyc:isa e y)
(instanceOf e x)))))))

“For every subclass y of Cyc:Event, there is an EventType that has the
same instances as y”
(forall (y) (if (cyc:genls y cyc:Event)
(exists (x) (and (EventType x)
(forall (e)
(iff (cyc:isa e y)
(instanceOf e x)))))))
1
ISIT Trigger Axioms

Example trigger axioms for Cyc event/token distinction

In Cyc, EventTypes are classes and events are individuals
> The inter-theory is neutral on the issue
> A commitment can be made on this issue using a triggering axioms
“If the TypesAreClasses trigger is true, EventTypes and the subclasses of
Cyc:Events are equivalent”
(forall (x) (if (TypesAreClasses)
(iff (cyc:genls x cyc:Event) (EventType x))))
2
ISIT Modules

Pre/Post conditions



Causality




Can an event cause an event?
Expected outcomes…
Triggering axioms identify the distinction
Inputs/Outputs


Classic AI-planning descriptions
Triggering axioms for situations vs. flows
Processes (esp. information processing) can have inputs and
outputs (different from pre/post conditions)
Control Flow


Are if/then/while important to model logically?
Still under discussion
2
IS IT an Ontology?
 ISIT
includes the five ontologies
 New vocabulary for generalizations of
common terms
 Trigger axioms exclude parts of the Inter
Theory under certain conditions
 In
a strict sense, it is not an ontology, but an
amalgem of existing ontologies…
 Pan-ontology?
2
Interoperable Contextualized Knowledge

IKRIS is addressing two KR challenges

Enabling interoperability of KR technologies
> Developed by multiple contractors
> Designed to perform different tasks

Interoperable representations of scenarios and contextualized
knowledge
> To support automated analytical reasoning about alternative
hypotheses
2
Contextualized Knowledge is Pervasive

The circumstances surrounding a specific activity
E.g., In this conversation, ‘the suspect’ refers to Faris.

A published document
E.g., Based on the schedule, the Holland Queen will arrive in Boston sometime on April
29, and depart there sometime on May 1.

An intelligence report
E.g., Pakes is listed, according to a certain source, on the crew roster of the Holland
Queen.

A database
E.g., Pakes is assumed, based on certain records, to not be a citizen of USA.

An assumption
E.g., Pakes’s presence on board the Holland Queen is assumed to be typical (i.e. he
does not behave abnormally).

A set of beliefs
E.g., In the belief system of Abu Musab al Zarqawi, democracy is evil.
2
Interoperable Contextualized Knowledge

IKRIS is producing –

A context logic with a formal model theory
> Called IKRIS Context Logic (ICL)

Recommended ways of using the logic for IC applications
E.g., to represent alternative hypothetical scenarios


Methodology for translating into and out of IKL

Methodology for automated reasoning

The model theory supports configurable entailments
Three immediate customers

PARC, Cycorp, KANI
2
Context Logic

In McCarthy’s context logic –

Contexts are primitive entities

Propositions can be asserted with respect to a context
> (ist c ) means that proposition  is true in context c
E.g., (ist CM (forall (x) (implies (P x) (G x)))); (ist C0 (P Fred))

How can automated reasoning be done with ist sentences?
E.g., assert (= CM C0) and derive (ist C0 (G Fred))

Contextualize constants rather than sentences

Constants in ist sentences are interpreted with respect to the context
E.g., Fred in (ist C0 (P Fred)) is interpreted with respect to C0

Replace each constant with a function of the context and the constant
E.g., { (forall (x) (implies (P (iso CM x)) (G (iso CM x))));
(P (iso C0 Fred)) }

Use a first-order reasoner to make deductions
2
Whoa!
KANI’s Hypothesis Graph
N1
S1: There will be a coordinated event.
S2: The event will occur on April 30.
S3: Pakes is a participant.
S4: Ramazi is a participant.
S5: Goba is a participant.
…
N2 S8: The event is a face-to-face meeting.
N3
S9: The event is at Select
Gourmet Foods.
New
hypothesi
s added
by the
analyst
N4 S10: The event is in Atlanta.
N5 S11: Pakes is in Boston on April 30.
2
Conflict Detected by KANI
N1
S1: There will be a coordinated event.
S2: The event will occur on April 30.
S3: Pakes is a participant.
S4: Ramazi is a participant.
S5: Goba is a participant.
…
N2 S8: The event is a face-to-face meeting.
N3
S9: The event is at Select
Gourmet Foods.
N4 S10: The event is in Atlanta.
N5 S11: Pakes is in Boston on April 30.
2
Helping Resolve Inconsistencies
N1
Event will not
occur on April 30
N1.1 ~S2,S3
N2.1
N3.1
S8
N5.1
N2.2
N3.2
S9
N4.1
N1.2 S2,~S3
S10
S11
N5.2
Pakes is
not a
participant
N3.3
S10
S11
N1.3
N2.3 ~S8
S8
S9
N4.2
S1,S4,S5,…
S9
N4.3
N5.3
S10
S11
S2,S3
Event is
N2 S8
not a
face-toface
N3.3 S9
meeting
N4.4 ~S10 N4
Event is
not in
Atlanta
N5.5
Pakes is
not in
Boston on
April 30
2
S10
~S11
Evaluation and Tech Transfer

Evaluation

Goals:
> Demonstrate the practical usability of results on IC-relevant problems
> Provide functionality goals, scoping, and feedback for results

Evaluation will be informal using sample IC tasks

Tests will include –
> Round trip translations into and out of IKL
> Inter-system knowledge exchange using IKL.

Tech Transfer

Goal: Transition results into DTO programs and the IC at large

Producing “showcase” presentations of results for transition audiences

Being advised and facilitated by our government champions and MITRE
3
Using CS4 to Demonstrate IKRIS Technology


Our demonstration shows interoperability and
collaboration among three selected NIMD technologies:
KANI, SLATE, and Noöscape
Two motivations for interoperation

Different (overlapping) data
> The CS4 was carefully enhanced and partitioned so no system by itself
had sufficient knowledge to “solve” CS4



Different (overlapping) capabilities
To be successful, each had to call upon the resources of
the others.
Translators are being developed to support the knowledge
representation languages needed to support those
systems and to enable knowledge sharing.
3
Summary

IKRIS is enabling progress to be made on significant KR&R problems

We are addressing two KR challenges relevant to the IC

Enabling interoperability of KR technologies
> Developed by multiple contractors
> Designed to perform different tasks

Interoperable representations of scenarios and contextualized knowledge
> To support automated analytical reasoning about alternative hypotheses

Initial versions of the technical results have been completed

For more information, check out the IKRIS Web site

http://nrrc.mitre.org/NRRC/ikris.htm
3
2
Download