Ontology Quality and the Semantic Web Chris Welty IBM Watson Research Center

Ontology Quality and the Semantic Web

Chris Welty

IBM Watson Research Center

Outline

• Welcome, opening joke

• History of web and hypertext

• Semantic Web overview

• Ontology Engineering and Quality

• Summary and Closing joke

History of Hypertext

• 1945: Vannevar Bush’s Memex– Associative Indexing and links

• 1965: Ted Nelson coins hypertext– “Nonsequential writing”

• 1967: Andries van Dam’s Hypertext Editing System (sponsored by IBM).

• 1985: Janet Walker’s Symbolics Document Examiner• 1987: Bill Atkinson’s Hypercard on the Mac• 1991: Tim Berners-Lee proposes HTTP, HTML, & URL

– Genesis c. 1989

• 1993: Mark Andreesen releases Mosaic for Mac, Unix, Windows…

Hypertext Research

• Dating back at least to the late 60s• Many foci

– Technology (mouse, software, protocols)– User interaction– Aesthetic– Post-modern– Engineering

• Largely ignored by web developers– Especially in the early days of the web (93-96)

Grassroots to the Web

• Early web dominated by “what it looks like” in Mosaic

• Focus on spreading the word, not doing it right• Many early web pages didn’t have links in text at

all– “Catalog” pages with lists of links– “Text” pages with few or no links– Embedded images more interesting than links

• Just do it rather than do it right• But…

– When the web became serious, the research started to matter

Semantic Web

• Defined, to date, by RDF and OWL

• Genesis c. 2000

• Still in the “early days”– Faster adoption (so far) than early web– FOAF the most widely used SW Ontology

Agent

Person

Organization

Group

Document Image

http://xmlns.com/foaf/0.1/

Ontology Research

• Dating back…• Multiple foci

– Technology (logics, reasoners…)– Meta-physics (what there is)– Knowledge Acquisition– NLP– Engineering

• Largely ignored by SW developers– Web 2.0, groundswell– Specifically criticized by some SW pundits

A little semantics…

• The SW catchphrase– “A little semantics goes a long way”

• Sometimes strengthened– A lot of semantics is too much– 80/20 rule

• Double-edged sword– FOAF doesn’t look like even 1%– The simplicity of FOAF hides any serious value

proposition for SW– SW not for people, for data– Important to get it right?

Some evidence

• Does quality matter?• Good quality ontologies cost more

– Required for some applications

• Improvements in quality can improve performance [Welty, et al, 2004]– 18% f-improvement in search– Cleanup cost ~1mw/3000 classes– BUT … low quality ontology still improved

base

Dimensions of Quality

• Coverage, correctness, richness, commitment [Kashyap, 2003]

• Organization, modularity [Rector, 2002]• Relation to reality [Smith & Welty, 2001]• Making meaning clear [Guarino, 1998]• Meta-level consistency [Guarino & Welty,

2000]• Captures the invariant structure of the

domain [Welty & Guarino, 2001]

Making Meaning Clear

• Part-of relates parts to their wholes– E.g. part-of(engine,car)

• Part-of is irreflexive

• Part-of is anti-symmetric

• Nothing can have only one part

Reduction of unintended models

• Generally, involves more axioms• Typically requires negation

– Disjointness

• Positive axioms– Also makes meaning clear, e.g.

• Clear significance for ontology alignment

Mammal

Horse

Chess Piece

Horse

Meta-Level Consistency with OntoClean

• Identity• Unity• Rigidity• Dependence• Actuality• Permanence

• Note on terminology: property is a unary relation (aka class), meta-property is a property of a class

Identity

• The foundation of ontology, conceptual analysis, etc• The criteria under which equivalence is determined

– Or under which difference is determined

• Already accepted practice in RDBs, OOP• When you conceive of a class, ask “What makes each

instance unique?”– Note for SW: uniqueness not assumed

• Meta-property– Is there an identity criterion for this class (+I)– Not always productive to specify the precise condition

• Esp. if this results in artificial attributes

– -I +I

Unity Criteria• An object x is a whole under iff is an

equivalence relation that binds together all the parts of x, such that

P(y,x) (P(z,x) y,z))

but not

y,z) x(P(y,x) P(z,x))

• P is the part-of relation can be seen as a generalized indirect

connection

Unity Meta-Properties

• If all instances of a propertyare wholes under the same relation itcarries unity (+U)

• When at least one instance of a property is not a whole, or when two instances are wholes under different relations, it does not carry unity (-U)

• When no instance of a property is a whole, itcarries anti-unity (~U)

• -U +U• +U ~U

Rigidity

• An essential property of an entity is a property that must necessarily (always) hold

• A rigid property is a property that is essential to all possible instances (+R)

• A non-rigid property is a property that is not rigid (-R)

• An anti-rigid property is a property that is not essential to all possible instances (~R)

• +R ~R

Formal Rigidity

is rigid (+R): x (x) � (x)

– e.g. Person, Apple

is non-rigid (-R): x (x) ¬ � (x)– e.g. Red, Male

is anti-rigid (~R): x (x) ¬ � (x)

– e.g. Student, Agent

(what about time?)

Rigidity Constraint

+R ~R

• Why?

� x P(x) Q(x)

Q~R

P+R

O10

Which one is better?

Computer

has-part

MemoryDisk Drive

Computer Part

Memory PartDisk Part

Computer Part

Disk Drive Memory

Computer

has-part

Due to: Guizzardi, et al, 2004.

-I~R-U

+I+R+U +I+R~U

-I~R-U

+I+R+U

+I~R~U+I~R-U

+I+R~U

Ontology Alignment

• Most automatic alignment tools would say yes

• Let’s take a closer look

Food

Apple

Food

Apple Caterpillar

Are these the same?

Ontology Alignment

• Different meta-properties for Food

• Different intended meaning

• Should not be aligned

• Meta-level analysis helps make meaning more clear

Food

Apple

Food

Apple Caterpillar

+I~U+D~R+I+U-D+R

A formal ontology of properties

Property

Non-sortal-I

Role~R+D

Sortal+I

Formal Role

Attribution -R-D

Category +R

Mixin -D

Type +O

Quasi-type -O

Non-rigid-R

Rigid+R

Material roleAnti-rigid~R Phased sortal -D

+L

The Backbone TaxonomyAssumption: no entity without identity

Quine, 1969

• Since identity is supplied by types, every entity must instantiate a type

• The taxonomy of types spans the whole domain• Together with categories, types form the

backbone taxonomy, which represents the invariant structure of a domain (rigid properties spanning the whole domain)

Entity

Physical object

Amount of matter Group

Organization

Location

Living being

Person

Animal

Social entity

Agent

Apple

Fruit

Food Legal agent

Group of people

Red apple

Red

Vertebrate

Caterpillar ButterflyCountry

Geographical Region

Lepidopteran

Entity

Physical object

Amount of matter Group

Organization

Location

Living being

Person

Animal

Social entity

Apple

Fruit

Group of people

Vertebrate

CountryGeographical

Region

Lepidopteran

Upper-Level Backbone

• The upper level backbone accounts for 5% of an ontology and spans the domain

• In empirical work, this is the most important layer [Fan et al, 2003]

• Some value in providing upper level ontologies to establish the basic distinctions

Backbone of quality

• Conjecture: the primary purpose of an ontology is to specify the backbone taxonomy, which is the invariant structure of the domain

• Bad ontologies:– “folksonomies”, – Subject hierarchies– Thesauri

Summary

• Good ontologies should:– Clarify meaning

• Add constraints to eliminate unintended models

– Have clear identity criteria– Have consistent meta-level properties– Specify the invariant structure of a domain

Use OntoClean for all your

ontology cleaning needs!