eswc2011phd-schneid

Reasoning in Expressive Extensions
of the RDF Semantics

Michael Schneider (FZI Karlsruhe, Germany)
ESWC 2011 PhD Symposium
Heraklion (Greece), 31 May 2011

WIR FORSCHEN FÜR SIE

RDF Semantics and Semantic Extensions
• RDF Semantics:
– Part of W3C RDF Specification (Hayes, 2004)
– Defines formal meaning of RDF graphs (as a model-theory)
– Includes four increasingly expressive semantics:
Simple Entailment, RDF, RDFS, and D-Entailment
– Characteristics:
• all RDF graphs are valid and have a semantic meaning
• Semantics is defined on the level of RDF triples and sets of triples
• all nodes represent resources (aka individuals)

• Semantic Extensions of the RDF Semantics:
– Semantics that builds on top of RDF(S) or D-Entailment:
• all parts of semantics of weaker language are reused and extended
• Syntax is all RDF graphs
– Example: RDFS is a semantic extension of RDF
– Example: OWL 2 Full is a semantic extension of RDFS (or D) 2

Semantic Web Ontology Languages:
Syntactic Flexibility vs. Semantic Expressivity

?
Syntactic Flexibility (RDF)

OWL 2 ? OWL 2
RDFS
RL/RDF Full

?
OWL 2
DL
OWL
Lite

Semantic Expressivity

• Unclear: Differences of OWL 2 Full to OWL 2 DL and OWL 2 RL/RDF?
• Unclear: Implementability of OWL 2 Full (or any expressive RDF extension)?

3

OWL 2 Full vs. OWL 2 DL:
Enhanced Syntactic Flexibility in RDF
OWL 2 DL tools typically cannot properly deal with every RDF graph:

Use of RDF(S) Entity Types
dcels:title rdf:type rdf:Property
dcterms:title rdf:type rdf:Property
dcterms:title rdfs:subPropertyOf dcels:title

OWL API 3.2 read/write roundtrip:
re-declaration of both properties
as OWL annotation properties

Result (after read/write roundtrip)
dcels:title rdf:type owl:AnnotationProperty
dcterms:title rdf:type owl:AnnotationProperty
dcterms:title rdfs:subPropertyOf dcels:title 4

OWL 2 Full vs. OWL 2 RL/RDF Rules:
Enhanced Semantic Expressivity
RDF entailment rule reasoning not always sufficient:

Vocabulary (GoodRelations):
gr:condition rdfs:domain [ owl:unionOf ( gr:Offering gr:ProductOrService ) ] .
gr:eligibleRegions rdfs:domain [ owl:unionOf ( gr:Offering gr:DeliveryChargeSpecification) ] .
gr:DeliveryChargeSpecification rdfs:subClassOf gr:PriceSpecification .
gr:PriceSpecification owl:disjointWith gr:ProductOrService .
Data (invented example):
ex:myThingy gr:condition "old but fine"^^xsd:string .
ex:myThingy gr:eligibleRegions "de"^^xsd:string .
Expected Result (OWL 2 Full/DL):
ex:myThingy rdf:type gr:Offering .

Beyond the OWL 2 RL/RDF rules ! 5

OWL 2 Full vs. OWL 2 DL & RL/RDF Rules:
Enhanced Modeling & Reasoning Capabilities

• Metamodeling
e.g. reasoning upon zoological hierarchies: Harry → Eagle → Species

• Cyclic relationships
e.g. detection of circular chemical molecules

• Macros, conditional semantics, etc.
e.g. custom entity types

6

Usage Scenarios for OWL 2 Full Reasoners

• Complementing RDF entailment-rule reasoners:
– much stronger in terminological reasoning
– RDF rule reasoners advantage: faster, better scalability
– fully compatible with RDFS and OWL 2 RL/RDF rules:
OWL 2 Full reasoner can safely operate in parallel

• Complementing description-logic reasoners:
– basically compatible due to „correspondence theorem“
– robust on weakly-structured data (typical for LOD cloud)
– „trans-DL“ reasoning (metamodeling, cyclic structures, …)
– DL reasoners advantage: better on valid OWL 2 DL input
7

Prior Art in OWL Full Reasoning

• Fikes, McGuinness, Waldinger: A First-Order Logic Semantics
for Semantic Web Markup Languages. TR, Stanford, 2002.
– translation of specifications of precursers of OWL and RDF into first-order
logic (FOL) theory, and application of FOL reasoners.
– focus: checking for technical issues in specifications (less on inferencing)
• Hayes: Translating Semantic Web Languages into Common Logic.
TR, Pensacola (Florida), 2005.
– translation of OWL 1 Full into Common Logic
– no report on reasoning experiments
• Hawke: Surnia. 2003. URL: http://www.w3.org/2003/08/surnia
– OWL 1 Full reasoner based on FOL translation using Otter FOL reasoner
– did not perform well on W3C OWL 1 test suite
– ad hoc implementation: does not properly follow specification; many flaws

Research Questions

1. What are the distinctive features of OWL 2 Full
compared to other approaches used for
Semantic Web reasoning?

2. To which degree and how can reasoning in
OWL 2 Full be implemented?

9

Approach
• „Feature Analysis“ (addresses 1st research question):
– Building up catalogs of distinctive pragmatic features of OWL 2 Full
– „distinctive“: not supported by either OWL 2 DL or OWL 2 RL/RDF rules
– will cover both syntactic (parsing) and semantic (reasoning) aspects:
• syntactic aspect example: disjoint annotation properties (SKOS)
• semantic aspect example: entailments from metamodeling (vs. „punning“)

• „Implementability Analysis“ (addresses 2nd research question) :
– Focus: in-deph investigation of „naive“ FOL translation approach:
• Translation of OWL 2 Full semantics into a first-order logic (FOL) theory
• Translation of RDF graphs into FOL formulae
• Applying FOL reasoners (theorem provers, model finders) for reasoning

• Evaluation:
– Collecting evidence for all identified OWL 2 Full features (empirical)
– Evaluating FOL-based reasoner prototype w.r.t. identified features
10

Feature Analysis: First Results
• Created: Catalog of syntactic-aspect features for OWL /1/ Full
– identified 14 feature categories and 90 features
– Example feature: “Anonymous Individuals with Cyclic Relationships”
– Example category: “Unrestricted Use of Blank Nodes“

• Usage: Evaluation of ontology engineering tools in EU Project SEALS
– per identified feature: created one small example ontology („spot test“)
– for each example ontology: analyzed read/write roundtrip for tool under test

• Results:
– OWL DL tools (OWL API 3.1, Protege 4, …) had many difficulties:
• almost all test ontologies were changed during read/write roundtrips
• in many cases, the changes were significant or even severe
– see SEALS deliverable D-10.3, specifically Appendix A for detailed analysis 11

Implementability Analysis: First Results
• Test suite: 32 characteristic OWL 2 Full conclusions („Fullish Testsuite“)
– „characteristic“: either OWL 2 DL reasoner or OWL 2 RL/RDF rule reasoner expected to fail
– Example test: „{} |= owl:equivalentClass rdfs:subPropertyOf rdfs:subClassOf“

• Results:
1. OWL 2 DL reasoner Pellet: 9 correct, 22 wrong, 1 system error
2. OWL 2 RL/RDF rule reasoner OWLIM : 9 correct, 23 wrong
3. ATP iProver-SInE, complete OWL 2 Full axiomatization: 28 correct, 4 timeouts (median: 5.31s)
4. ATP iProver-SInE, small subset of sufficient axioms per test case: all correct (median: 0.08s)

0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
Pellet 2.2.2 + + + - - - - - + + - - - - + - - - - + + - - - - + - - ? - - -
BigOWLIM 3.4 + - - + - - + + - - + + - - + - - + + - - - - - - - - - - - - -
iProver 0.8, all axioms + + + + + + + + + + + ? ? + + + + + + ? ? + + + + + + + + + + +
iProver 0.8, sufficient + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

+ success (termination with correct result)
- failure (termination with wrong result)
? unknown (timeout, system error, etc.) 12

Conclusions and Future Work

• OWL 2 Full has many distinguishing features and potential benefits

• OWL 2 Full reasoning generally works with FOL reasoners 
but there is a serious efficiency issue due to the large FOL axiomatization 

Report on the results of all reasoning experiments (to appear):
Michael Schneider, Geoff Sutcliffe: Reasoning in the OWL 2 Full Ontology Language
using First-Order Automated Theorem Proving. CADE 2011.

• FOL-translation approach is very flexible:
– applies to arbitrary extension of RDF semantics (including complete RDFS)
– enables rule-style extensions (e.g. RIF+OWL-Full combination)

• Future work: finish feature analysis (syntactic and semantic features)

• Future work: address main efficiency issue: method to remove irrelevant axioms

• Future work: investigate query answering (towards SPARQL 1.1) 13

Thank You !

Questions ?

14

Enhanced Modeling & Reasoning:
Metamodeling Example
Meta-Classes

Species

Endangered Species NonEndangeredSpecies
= { BaldEagle, Tiger } = ¬ EndangeredSpecies п Species
(mutually disjoint)
Classes

Tiger BaldEagle GoldenEagle Dog

rdf:type
Individuals

Harry

ex:hasMetaClass owl:propertyChainAxiom ( rdf:type rdf:type )
16

Cyclic Relationship Example
ex:r owl:propertyChainAxiom (
ex:hasRelative
HasRelativeAsBoss [ owl:inverseOf ex:hasBoss ] ) .
Coincidences

ex:HasRelativeAsBoss owl:equivalentClass [
rdf:type owl:Restriction ;
rdf:type

owl:onProperty ex:r ;
hasRelative owl:hasSelf "true"^^xsd:boolean ] .

alice bob ex:alice ex:hasRelative ex:bob .
ex:alice ex:hasBoss ex:bob .
hasBoss |=
ex:alice rdf:type ex:HasRelativeAsBoss .
Cycles

Basic Cycle Complex Cycle
(Coincidence with Inverse) (Coincidence with
17
intermediate Nodes)

Macros Example
Modeling Aim: Define the „custom entity type“ PersonAttribute
as the class of all functional data properties that have class foaf:Person as their domain.
Premise (Definition and Data) Expected Conclusion
Definition: ex:name rdf:type owl:DatatypeProperty .
foaf:Person rdf:type owl:Class . ex:name rdf:type owl:FunctionalProperty .
ex:PersonAttribute ex:alice rdf:type foaf:Person .
owl:intersectionOf (
owl:DatatypeProperty
owl:FunctionalProperty
[ rdf:type owl:Restriction ;
owl:onProperty rdfs:domain ;
owl:hasValue foaf:Person ] ) .

Data:
ex:name rdf:type ex:PersonAttribute .
ex:alice ex:name „Alice" . 18

Syntactic-Aspect Feature Categories

19

Syntactic Aspect Feature Analysis:
Evaluation of OWL API 3.1 (coarse)
• Application of concrete example OWL Full ontologies to OWL API 3.1
• Observation: most test ontologies were modified („repaired“)
• Note: the differences have been analysed in detail (not shown)

HR CR TR TC NT ME DP AP DT LT BN CP LS LR + isomorphic RDF graph
reconstruction
01 - - - - + - - + - - - - - -
- different RDF graph
02 - - - - + - - + - - - - -
03 - - - - + - - - - - - - - X processing error
04 - - - + - - - - - - - -
05 - - - + - - - + - - -
06 - - - - - - - - -
07 - - - - -
08 - - - -
09 X - -
10 - - -
11 - -
12 -
20

Syntactic Aspect Feature Analysis:
Evaluation of OWL API 3.1 (fine-grained)
OWL (2) Full Perspective OWL (2) DL Perspective
H C T T N M D A D L B C L L H C T T N M D A D L B C L L
R R R C T E P P T T N P S R R R R C T E P P T T N P S R

0 # - # # + # # + # - - # # # 0 - ! # - - # - + - - ! - - -
1 1

0 # # # # + # # + # - - - # 0 # # / - - # - + - - ! ! -
2 2

0 - / # # + # # # # - - # # 0 ! # # - - # - - - ! ! - -
3 3

0 ! # / + # # # / - - # / 0 - / - - # - - - - - - -
4 4

0 # # / + # # ! + # - / 0 # # - + # - - + - - -
5 5

0 # # # # # - # - - 0 / - # # - ! - - -
6 6

0 # # - - # 0 # # ! - -
7 7

0 ! # - # 0 - # - -
8 8

0 X / # 0 X # -
9 9

1 - / # 1 - # -
0 0

1 / # 1 # -
1 1

1 # 1 #
2 2

21

(Entailment Checking)
OWL 2 Full Reasoning
FOL Translation Approach
FOL-Translations of
Semantic Conditions &

FOL-Translation of
Premise Graph

negated
FOL-Translation of
&

&
FOL
Reasoner
(ATP) { TRUE
FALSE
UNKNOWN }
&
Conclusion Graph
Semantic Conditions
FOL-Translation of

model-theoretic OWL 2 Full semantic condition
corresponding FOL formula (TPTP)
FOL-Translation of
RDF Graphs

RDF graph (Turtle) 22
corresponding FOL formula (TPTP)

rdfbased-sem Test Suite:
Language Coverage: (no datatypes)
Pellet 2.2.2 (OWL API 3.1) 237 168 6

HermiT 1.3.2 (OWL API 3.1) 246 157 8

FaCT++ 1.5.0 (OWL API 3.1) 190 45 176

BigOWLIM 3.4 (owl2-rl) 282 129 0

Jena 2.6.4 (OWL_MEM_RULE_INF) 129 282 0

Parliament 2.6.9 (default config) 14 373 24

Vampire 0.6 (all axioms) 349 0 62

iProver-SInE 0.8 (all axioms) 383 028

iProver-SInE 0.8 (sufficient axioms) 396 0
15

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Success Failure Unknown 23

rdfbased-sem Test Suite: Performance

Vampire 0.6 iProver-SInE 0.8
complete axiomset complete axiomset
min 0.01 s 0.05 s
max N/A N/A
max succ 27.57 s 278.71 s
Q1 (1st quantil) 0.03 s 0.09 s
Q2 (median) 0.35 s 0.29 s
Q3 (3rd quantil) 0.56 s 5.21 s
mean succ 0.42 s 14.59 s
StD succ 2.06 s 36.45 s
24

Fullish Testsuite Evaluation Results:
Selected Semantic Web Reasoners

0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
Pellet 2.2.2 (OWLAPI) + + + - - - - - + + - - - - + - - - - + + - - - - + - - ? - - -
Hermit 1.3.2 + ? + - - ? - + + + - - - - + - - - - + + - - + ? + - - ? - - -
Fact++ 1.5.0 + ? ? ? ? ? ? - ? + - - - ? + ? - - - + + ? ? ? ? + - ? ? - - ?
BigOWLIM 3.4 (owl2-rl) + - - + - - + + - - + + - - + - - + + - - - - - - - - - - - - -
Jena 2.6.4 (OWL) + - - - - + + + - - + - - - - - - + - - - - + - - + - - - - - +
Parliament (default) + - - - - - - + - - ? - - - - - - - ? - - - - - - - - - - ? ? -

? unknown (timeout, system error, unsupported, etc.)

25

Fullish Testsuite Evaluation Results:
ATPs on small sufficient Axiomsets

0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
E-Prover 1.2 + + + + + + + + + + + + + + + + + + + ? + + + + + + + + + + + +
Equinox 5.0 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
iProver 0.8 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Metis 2.3 + + + + + + + + + + + + ? + + + + + + ? + + + + + + + + + + + +
Prover9 0908 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
SPASS 3.5 + + + + + + + + + + + + ? + + + + + + + + + + + + + + + + + + +
Vampire 0.6 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

? unknown (timeout, system error, unsupported, etc.)

26

Fullish Test Suite: Performance

iProver-SInE 0.8 iProver-SInE 0.8 iProver-SInE 0.8
complete axiomset small axiomsets 1M bulk RDF data
min 0.08 s 0.04 s 21.73 s
max N/A 164.20 s N/A
max succ 123.01 s 164.20 s 63.10 s
Q1 (1st quantil) 0.72 s 0.05 s 21.91 s
Q2 (median) 5.31 s 0.08 s 22.07 s
Q3 (3rd quantil) 89.45 s 0.14 s 22.50 s
mean succ 30.63 s 7.82 s 24.76 s
StD succ 43.41 s 29.82 s 10.02 s 27

Model-Finding Experiments
• Task: Detection of consistency of ontologies and non-entailments
• Prerequisite: detection of satisfiability for whole axiomatization

Results (Summary):
• OWL 2 Full:
– No FOL model-finder confirmed satisfiability of axiomatization (timeouts) 
– Fortunately: no theorem prover confirmed unsatisfiability!
– Good: all „small-sufficient“ sub-axiomatizations of test cases satisfiable! 
• ALCO Full (undecidable fragment of OWL 2 Full [Motik05]):
– Consistency checking for axiomatization successful ! 
– Non-entailment checking often successful ! (but still some failures)
– Performance: median ~18s with model-finder Paradox
• RDFS (actually: RDFS-EXT, Sec. 4.2 of RDF Semantics):
– Consistency and non-entailment checking always successful ! 
– pretty fast: ~1/10s for most experiments with model-finder DarwinFM
28

Model-Finding Experiments:
Consistency / Non-Entailment Detection

Language: RDFS-EXT Language: ALCO Full
Testsuite: Fullish Testsuite: Fullish
ATP: DarwinFM 1.4.5 ATP: Paradox 4.0
min 0.01 s 8.21 s
max 7.35 s N/A
max succ 7.35 s 89.21 s
Q1 (1st quantil) 0.05 s 13.60 s
Q2 (median) 0.07 s 17.62 s
Q3 (3rd quantil) 0.12 s N/A
mean succ 0.71 s 19.18 s
StD succ 1.86 s 18.99 s
29

eswc2011phd-schneid

More Related Content

What's hot (7)

Viewers also liked (20)

Similar to eswc2011phd-schneid (20)

Recently uploaded (20)

eswc2011phd-schneid