OWLはオントロジーの課題を

本当に解決するのか？山口　 高平

（静岡大学）

人工知能学会研究会資料SIG-SWO-A201-02

Semantic Web の９階層

オントロジーに関する知識工学と次世代 Web の流れ

知識工学のトレンド

• 90-: 概念化の明示的仕様（ Tom Gruber オントロジーの定義）オントロジー記述言語 (Ontolingua)知識交換言語 (KIF)Generic OntologyCYC, WordNet, EDR…PSMTask Ontologyオントロジー構築方法論．．．

次世代Webのトレンド

• 95-97: XML as arbitrary structures

• 97-98: RDF

• 98-99: RDFS (schema) as a frame-like system

• 00-01: DAML+OIL

• 02-07: OWL

OWL への道のり

• On-To-Knowledge が OIL を定義• 重要 EU プロジェクトのいくつかが OIL を採

用• DAML プロジェクトと OIL がリンク• 定義： DAML+OIL• 改定： DAML+OIL• W ３ C へ DAML+OIL を提出• WebOnt Working Group 　の立ち上げ

⇒ OWL のドラフト

Jan’00Med’00Sep’00Dec’00Mar’01Aug’01Oct’01

Jan’00Med’00Sep’00Dec’00Mar’01Aug’01Oct’01

RDF （ S) と OWL の関連

• class-def• subclass-of• slot-def• subslot-of• domain• range

• class-def• subclass-of• slot-def• subslot-of• domain• range

• class-expressions• AND, OR, NOT

• slot-constraints• has-value, value-type• cardinality

• slot-properties• trans, symm

• class-expressions• AND, OR, NOT

• slot-constraints• has-value, value-type• cardinality

• slot-properties• trans, symm

RDF(S)DAML+OILDAML+OIL

≒≒OWLOWL

Ontologies by Keyword

Keyword URI

academic department

http://www.cs.umd.edu/projects/plus/DAML/onts/cs1.0.daml

academic department

http://www.cs.umd.edu/projects/plus/DAML/onts/cs1.1.daml

Academic Positions http://www.daml.ri.cmu.edu/ont/homework/cmu-ri-employmenttypes-ont.daml

access control primitives

http://www.w3.org/2000/10/swap/pim/doc.rdf

acronym http://orlando.drc.com/daml/Ontology/Thesaurus/CALL/current/

activity http://www.kestrel.edu/DAML/2000/12/OPERATION.daml

Actors http://opencyc.sourceforge.net/daml/cyc.daml

Actors http://www.cyc.com/2002/04/08/cyc.daml

Actors http://www.cyc.com/cyc-2-1/cyc-vocab.daml

address book http://www.w3.org/2000/10/swap/pim/contact.rdf

agenda http://www.daml.org/2001/10/agenda/agenda-ont

DAML Ontology Library (Ontology's by Keyword) http://www.daml.org/ontologies/keyword.html

<rdf:RDF xmlns="http://www.daml.org/2001/08/baseball/baseball-ont#" xmlns:daml="http://www.daml.org/2001/03/daml+oil#" xmlns:oiled="http://img.cs.man.ac.uk/oil/oiled#" xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#"> <daml:Ontology rdf:about=""> <rdfs:comment>baseball</rdfs:comment> <rdfs:comment>baseball ontology</rdfs:comment>

<daml:versionInfo>"1.0"</daml:versionInfo> </daml:Ontology> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#PostSeasonGame"><rdfs:label>PostSeasonGame</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:55:22 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Game"></rdfs:Class> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Inning"> <rdfs:label>Inning</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:22:49 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#AggregateEvent"></rdfs:Class> </rdfs:subClassOf> <rdfs:subClassOf> <daml:Restriction> <daml:onProperty rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#number"></daml:onProperty> <daml:toClass> <rdfs:Class rdf:about="#http://www.w3.org/TR/xmlschema-2/#string"></rdfs:Class> </daml:toClass> </daml:Restriction> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#League">

<rdfs:label>League</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:35:38 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <daml:Restriction> <daml:onProperty rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#division"></daml:onProperty> <daml:toClass> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Division"></rdfs:Class> </daml:toClass> </daml:Restriction> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Person">

<rdfs:label>Person</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>13:00:13 26.08.2001</oiled:creationDate> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Player"> <rdfs:label>Player</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:25:42 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Employee"></rdfs:Class> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Relief"> <rdfs:label>Relief</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>18:31:03 28.08.2001</oiled:creationDate> <rdfs:subClassOf> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#LineupEvent"></rdfs:Class> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Season">

<rdfs:label>Season</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:13:04 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <daml:Restriction> <daml:onProperty rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#year"></daml:onProperty> <daml:toClass> <rdfs:Class rdf:about="#http://www.w3.org/TR/xmlschema-2/#string"></rdfs:Class> </daml:toClass> </daml:Restriction> </rdfs:subClassOf> </rdfs:Class> <rdfs:Class rdf:about="http://www.daml.org/2001/08/baseball/baseball-ont#Series">

<rdfs:label>Series</rdfs:label> <rdfs:comment></rdfs:comment> <oiled:creationDate>12:56:25 26.08.2001</oiled:creationDate> <rdfs:subClassOf> <daml:Restriction> <daml:onProperty rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#home"></daml:onProperty> <daml:toClass rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#Team"/> </daml:Restriction> </rdfs:subClassOf> <rdfs:subClassOf> <daml:Restriction> <daml:onProperty rdf:resource="http://www.daml.org/2001/08/baseball/baseball-ont#away"></daml:onProperty> <daml:toClass

Baseball Ontology

ソフトウェア開発プロセスのためのオントロジー開発 (人手）経験

IssueIssue ApproachApproach

A Process-Centered Software Engineering A Process-Centered Software Engineering Environment Using OntologiesEnvironment Using Ontologies

GoalGoal

Real software processeschange dynamically...

need the facility to correspondto changing process dynamically

very difficult to maintain

develop software process ontologiesthrough the domain analysis usingtwo case studies

First,

Afterwards,design the system which generatesoftware processes interactivelywith a user

Development of Software Process Ontologies

Step1. Extract real software processes and objects fromthe two cases

InventInvent

JoinJoin

AppendAppend

ExtractExtract

C

A+BA

B

inputinput

referencereference

Process

B’orA’

Aαα+

outputoutput

Step2. Organize extracted processes and objects into distinctgroups such that there are more similarity in meaning,while defining the software process scheme

Step3. Give detailed definitions to each processes using software process scheme

activityactivity

InventInvent

JoinJoin

AppendAppend

ExtractExtract

Invent with referencesInvent with references

Invent without referencesInvent without references

Join without referencesJoin without references

Extract without referencesExtract without references

Join with referencesJoin with references

Extract with referencesExtract with references

EvaluateEvaluateReviewReviewExamineExamine

Focus on input,output,and reference roles

Focus on input,output,and reference roles

Focus on reference,tool and agent rolesFocus on reference,tool and agent roles

Process OntologyProcess Ontology

Software Process Ontologies(conceptual hierarchy)

Object Ontology includingabout 400 concepts

Process Ontology including about 250 concepts

Process Ontology Object Ontology

GoalObjectGoal

Object??

??

??

??

??Start

ObjectStart

Object

Retrieve processessatisfied with constraints

Software Process Automation using ontologies

A query about SＥ P

What SEP between the following input and output ?What SEP between the following input and output ?

Input : Requirement Specification from S/W Subsystems Output : Detailed Design of Each Component of S/W Subsystems

Case Study

Initial Software Process Plan CandidatesInitial Software Process Plan Candidates

Requirement Specification

fromS/W Subsystems

Detailed Design ofEach Components of

S/W Subsystems

Design BasicComponents of

S/W Subsystems

Design BasicComponents ofI/F Subsystems

Design LogicalAspects of DB

Review AllBasic

DesignsTogether

Evaluate AllBasic Designs

Design Physical

Aspects of DB

Specify TestingS/W Integration

Design DetailedComponents of

S/W Subsystems

Design DetailedComponents ofI/F Subsystems

Specify TestingUnits of S/W

Evaluate AllDetailed Designs

Review AllDetailed Designs

Together

OutputOutput

InputInput

Interim Software Process Plan CandidatesInterim Software Process Plan Candidates

Requirement Specification

fromS/W Subsystems

Detailed Design ofEach Components of

S/W Subsystems

Design BasicComponents of

S/W Subsystems

Design BasicComponents ofI/F Subsystems

Design LogicalAspects of DB

Evaluate AllBasic Designs

Design Physical

Aspects of DB

Specify TestingS/W Integration

Design DetailedComponents of

S/W Subsystems

Design DetailedComponents ofI/F Subsystems

Specify TestingUnits of S/W

Evaluate AllDetailed Designs

Review AllDetailed Designs

Together

OutputOutput

InputInput

Final Software Process Plan CandidatesFinal Software Process Plan Candidates

Requirement Specification

fromS/W Subsystems

Detailed Design ofEach Components of

S/W Subsystems

Design BasicComponents of

S/W Subsystems

Design BasicComponents ofI/F Subsystems

Design LogicalAspects of DB

Evaluate AllBasic Designs

Design Physical

Aspects of DB

Specify TestingS/W Integration

Design DetailedComponents of

S/W Subsystems

Design DetailedComponents ofI/F Subsystems

Specify TestingUnits of S/W

Evaluate AllDetailed Designs

Review AllDetailed Designs

Together

OutputOutput

InputInput

Evaluation

Issues and Future work

•The environment can’t allocate human and time resource•The environment has not the facility that supports the user to add user’s specific processes•The methodology for building ontologies has no theoretical aspects,such as soundness

Applicability

It turned out that the environment generated SPP candidates good for the user’s query with user interaction

計算機可読型辞書とテキストコーパスからの

オントロジー開発 (半自動）経験

How to Build up Domain Ontologies with Less Cost

• Taking Existing Information Resources

DODDLE Project (a Domain Ontology rapiD DeveLopment Environment)

exploiting a MRDConstructing up just a conceptual hierarchy

Why not taking MRD to build up domain ontologies ?• Good News:

MRD have many concepts.

• Bad News:The concepts have been defined from the point of natural language processing (common sense) and so there are some concept drift between MRD and specific domain ontologies.

DomainDomainＢＢDomain Domain ＡＡ

Concept Drift

DomainMRD

Reusable Part

No reusable partbecause ofconcept drift

Selection of Best-match

Spell-Matched Results

Spell-Match

Concept HierarchyConcept Hierarchy

Domain Expert

Extension of Modification

•Matched-Results Analysis •Trimmed- Results Analysis

Trimmed Model

Trimming

Initial Model

A Set of A Set of Input Input TermsTerms

WordNet

DODDLEDODDLE

Best-Match

Internal Node

Root

MoveMove

Move

StayStay

Matched-Results Analysis

A

B C

D

Trimmed part

Initial Model

00

3

B C D

Trimmed Model

A

B C

AUser re-constructsthe sub-tree.

D

Trimmed-Results Analysis

Acquiring not only Conceptual Hierarchy but Concept Definitions

DODDLE II Project (Extending DODDLE)

Info. Resource:Domain-Specific TextsTechnique: Co-occurence information WordSpace

non-TaxonomicRelationships

TaxonomicRelationships

TaxonomicRelationshipAcquisition

Module

Concept SpecificationTemplate

Concept SpecificationTemplate

extension & modification

WordNet

MRD

DODDLE-DODDLE-IIIIOverviewOverview

Domain SpecificTexts

Domain SpecificTexts

DomainOntology

non-TaxonomicRelationship

LearningModule

a Set of Domain Termsa Set of Domain Terms

• Words and phrases in texts can be expressed by vector representation containing co-occurrence statistics

Extracting Concept Relationshipsfrom Domain-Specific Texts

WordSpace　（Marti A. Hearst, Hinrich Schutze）

… wi … wj …C1 … wk …… wi … wj …C2 … wk …

Context Similarity between concepts C1 and C2

• Inner products among the vectors work as the similarity between the words and phrases.

high value with similar words coming up around the concepts

A sum of w in Texts

A sum of W in same synset

Constructing WordSpace1. Extract Word 4-grams with high-

frequency2. Collocation Matrix3. Context Vectors4. Word Vectors ( WordSpace is a set of word

vectors )5. Vector representations of all concepts…

.. f8 f4 f3 f7 f8

f4 f1 f3 f8 f9 f2

f5 f1 f7 f1 f5 ..…

….. f8 f4 f3 f7 f8

f4 f1 f3 f4 f9 f2

f5 f1 f7 f1 f5 ..…

Collocation Matrix

4-gram vector4-gram vector

f1

f1

fn

fnf2

f4 2

…

…

..

..

w

w

Texts (4-gram array) Texts (4-gram array) Texts (4-gram vector array)

Vector representation for each 4-gram

….. f8 f4 f3 f7 f8

f4 f1 f3 f4 w f2

f5 f1 f7 f1 f5 ..…

A sum of 4-gram vectors around w

WordNet synset

w1

w3

w2

w4

C

Word Vector W

Context vector w

Vector representation of a concept C

collocation scope context scope

non-TAXONOMY?

: Ca

TAXONOMY

: Cb

non-TAXONOMY?

: Cc

non-TAXONOMY?

: Cd

Concept Specification Template

Ci

ConstructingConcept Specification Template

Ci Ca

Ci Cb

Ci Cc

Ci Cd

a Set of Concept Pairsa Set of Concept Pairsfrom non-TR Learning Modulefrom non-TR Learning Module

Ci

Cb

Ckancestor, descendant and sibling

Taxonomic RelationshipsTaxonomic Relationshipsfrom TR Acquisition Modulefrom TR Acquisition Module

DODDLE-IIDODDLE-IIalphaalpha on Perl/Tkon Perl/Tk ((UNIX UNIX platformplatform))

The Target Domain

Contracts for the International Sale of Goods(CISG)

The Target Domain

Contracts for the International Sale of Goods(CISG)

A Case Study

Input Terms to TR module 46 Legal Terms from CISG Part-II

Domain-Specific Texts to NonTR modulefull text of CISG (about 10,000 words)

Results

46Terms

The paths from Spell Matched Nodes

to the Root of WordNet

Trimmed Model Legal Ontology

Trimming

Initial Model

56Terms 61Terms

377Terms113Terms

Select Best Matches

Modification

Taxonomic Relationships for input termsconstructed from TR Acquisition Module

Support ratio: Support ratio: How match is included in final domain ontology the intermediate products at each DODDLE activity.

0.0

10.0

20.0

30.0

40.0

50.0

60.0

70.0

80.0

90.0

100.0

Com

pone

nt r

ate

（％）

Trimmed Model Extended TrimmedModel

Internal DomainOntology1

Internal DomainOntology2

Domain Ontology

DODDLE's process

WN's structure(52.2%) Small DT(8.0%) Strategy1 (5.3%)Strategy2 (6.9%) Correction by user (21.5%) Addition by user (6. %)１

Support Ratio

ExtractingConcept Relationships

Extraction frequency of 4-gram (times) 8

Collocation scope (4-grams) 10

Context scope (4-grams) 60

setting value for WordSpace

the concept pairs extracted according to context similarity (threshold 0.9993)

Domain Experts Modifying Concept Specification Templates

assentnon-TAXONOMY? : offeror

TAXONOMY : act

non-TAXONOMY? : effect

non-TAXONOMY? : offer

non-TAXONOMY? : person

non-TAXONOMY? : offeree

non-TAXONOMY? : withdrawal

non-TAXONOMY? : time

TAXONOMY : proposal

Concept Specification Template

assent

non-Taxonomic Relationshipsfrom the template

ex) non-Taxonomic Relationships for “assent”

: offerLEGAL-SEQUENCE

: personAGENT

: withdrawalANTONYM

DE Modification

non-taxonomic relationship:

person, offer, withdrawal

taxonomic relationship:

act, proposal

inheritance: offeror, offeree

unnecessary: effect, time

Recall & Precision of Concept Specification Templates

recall

precision

The number of extracted concept pairs

ConclusionsDODDLE II: A Domain Ontology Construction SupportEnvironment using MRD and Domain-Specific Texts

Legal experts appreciate DODDLE II to some extent.

•How to Decide Threshold for CS• How to Identify NTR•Another DM method instead of WordSpace• Another Case Studies with Large Scale

Future WorkFuture WorkFuture WorkFuture Work

OWLとオントロジー開発

• OWL とオントロジー開発⇔ UML とソフトウェア開発

• OWL ＋ RDF(S) ＋ RDF⇒ 中味のあるオントロジー

• オントロジー構築支援ツールが必要• オントロジーの重量⇔スケーラビリテ

ィ