Ontology Modularization for Knowledge Selection: Experiments and Evaluations

1
Ontology Modularization for Knowledge
SelectionExperiments and Evaluations

• Mathieu dAquin, Anne Schlicht, Heiner
  Stuckenschmidt and Marta Sabou
• The Knowledge Media Institute, The Open
  University
• m.daquin,r.m.sabou_at_open.ac.uk
• University of Mannheim, Germany
• anne, heiner_at_informatik.uni-mannheim.de

2
The semantic web
The Semantic Web, an extension of the current
Web in which information is given well-defined
meaning, better enabling computers and people to
work in cooperation, is becoming more and more a
reality The amount of knowledge published on the
Semantic Web is rapidly increasing.
Needs for tools to facilitate the management of
this large amount of knowledge.
3
Examplehttp//www.co-ode.org/ontologies/test/brea
ksmetrics.owl

4
Ontology Modularization

• Ontology modularization refers to the task of
  extracting from a potentially large ontology, one
  or several significant sub-parts (components,
  modules)
• The interests of ontology modularization for
  knowledge engineering include facilitating
  knowledge reuse, life-cycle management, ontology
  evolution, visualization, etc.
• Ontology modularization has also been used for
  improving the performance of ontology-based tools
  (e.g. reasoners).

5
Example, modularized
6
Agenda

• Considering a specific scenario for ontology
  modularization Knowledge selection
• Identify existing modularization techniques.
• Identify evaluation criteria.
• Evaluate the techniques.
• Analyse issues and limitations in the field of
  ontology modularization.
• Hypothesis The adequate modularization technique
  and evaluation criteria are dependent on the
  application scenario.

7
Considered ScenarioKnowledge Selection
The ideal world (Web)
The real world (Web)
Knowledge Selection
Knowledge Selection
8
Two types of modularization techniques
t1, t2, t4
Module Extraction
Ontology Partitioning
t1
t5
tn
t2
t3
t4
t1
t4
t2
9
Partitioning Technique SWOOP

• The Partitioning tool integrated into the SWOOP
  ontology editor
• Divide the currently edited ontology into
  partitions
• Fully automatic
• Focus on the logical completeness of the
  resulting partitions

10
Partitioning Technique PATO

• Partitioning relying on the dependency graph
  between ontology entities
• Use measures inspired from the field of network
  analysis for clustering entities
• Try to minimize interconnectivity between modules

11
Extraction Technique Prompt

• Extraction feature of the Prompt Toolkit, a
  Protégé Plugin
• Takes a concept and extract the related entities
  by traversing the ontology relations
• This tool is
• Interactive (the user have to select class and
  relations to consider)
• incremental (the current module can extended)
• The generated module is integrated with the
  ontology in Protégé

12
Extraction Technique Galen Segmentation Service

• Module extraction service for the GALEN ontology
• Takes a set of concept names from GALEN and
  extracts the corresponding module
• Also rely on traversal techniques
• Additional filters for reducing the size of the
  generated module

13
Extraction Technique KMi

• Fully automatic
• Exploits inferences during modularization
• Takes shortcuts in the hierarchy to reduce the
  size of the module
• Can be used on simple taxonomies as well as on
  complex OWL structures
• The technique is itself modular, so easy to adapt
  to particular needs

14
Evaluation criteria

• Criteria evaluating the tool
• Assumption on the ontology (works with various
  ontologies)
• Level of user interaction
• Use of module (availability of tools to
  manipulate modules)
• Performance (usable in run-time)
• Criteria evaluating the modularization
• Size (simple, but effective evaluation criteria)
• Distance (concept of interest can easily be
  considered together)
• Other (not considered) criteria
• Logical criteria (correctness and completeness),
  relations between partitions (connectedness,
  redundancy)

15
Experiments

• Simulate the knowledge selection process on two
  examples
• Example 1 (simple)
• Based on the news stories for a research lab
• Keywords Student, University, Researcher
• Example 2 (less simple)
• Based on a news story about the queen birthday
  dinner
• Keywords Queen, Birthday, Dinner
• Select ontologies for both examples
• Modularize these ontologies using the presented
  techniques
• Evaluate the results

16
Experiments Selected ontologies

• Example 1
• ISWC http//annotation.semanticweb.org/
  iswc/iswc.owl
• KA http//protege.stanford.edu/plugins/owl/owl-
  library/ka.owl
• Portal http//www.aktors.org/ontology/portal
• Example 2
• OntoSem http//morpheus.cs.umbc.edu/aks1/ontosem.
  owl
• Tap http//athena.ics.forth.gr9090/RDF/VRP/Examp
  les/ tap.rdf
• Mid-level http//reliant.teknowledge.com/DAML/Mid
  -level-ontology.owl
• Example 1 ontologies are small, well defined
  ontologies, while Example 2 ontologies are bigger
  and more heterogeneous in quality

17
Results application criteria
18
Results performance

• Example 1
• Small well defined ontologies
• All techniques provided modularization in less
  than a few second
• Example 2
• Bigger ontologies that are heterogeneous in
  quality and richness of the definition
• Big variations
• The worst case big ontologies like Tap, with KMi
  and Pato.
• The best case SWOOP but still requires several
  minutes.
• Most of the time is dedicated to loading the
  ontology

19
Results size
In Example 2 There is even more differences
between SWOOP (100 of the ontologies in the
modules) and the other techniques (only a few
entities in the modules)
20
Results distance

• The KMi tool includes mechanisms that take
  shortcut in the class hierarchy, thus reducing
  the distance between concepts of interest.
• Example

Person
Person
Student
AcademicStaff
Student
Researcher
Researcher
In the module resulting from KMi
In the portal ontology
21
Conclusions

• Our experiments show important variations in the
  way ontology modularization techniques behave.
• Each of them is based on its own intuition
• Each of them is relevant for a particular
  scenario
• KMi performs best in the case of the knowledge
  selection scenario -)
• But our basic conclusions are rather negative
• There is No Universal Modularization
• Existing criteria are insufficient
• Techniques need to be improved

22
Conclusions

• Need for a benchmark for ontology modularization
  techniques
• Helping developers in selecting the appropriate
  technique for their applications
• Need for parametric modularization
• Allowing developers to specify the modularization
  technique so that it matches their application
  requirements.

23
Thank you!
24
Example Application (Power)Magpie
25
Application scenario Magpie
Knowledge Selection
tn

t1
t2
t3
t4
t5
t2
t5
Term extraction
t1
tn
t3
t4
Magpie