Learning Ontologies from RDF Annotations

1
Learning Ontologies from RDF Annotations

• Alexandre Delteil, Catherine Faron-Zucker, Rose
  Dieng
• ACACIA project, INRIA, 2004
• Sophia Antipolis, France

2
TOC

• Introduction
• RDF RDFS Background
• Ontology Example
• Approach to Ontology Learning
• Conclusion
• Future Work

3
Introduction

• Build ontologies from information extracted from
  RDF annotations
• We have a method to learn ontologies from RDF
  annotations by systematically generating the most
  specific generalization of all the possible sets
  of resources.

4
RDF Annotation

• Triplet statement (resource, property, value),
  (Njal, type, Cat)
• Easily represented as a graph
• XML syntax provided

5
XML Serialization of RDF Annotation

• ltrdfDescription aboutNjalgt
• ltrdftype resourceCat /gt
• ltlivesIngt
• ltrdfDescriptiongt
• ltrdftype resourceHouse /gt
• ltownedBy rdfresouceCatherine /gt
• lt/rdfDescriptiongt
• lt/livesIngt
• lt/rdfDescriptiongt

6
RDF Schema (RDFS)

• RDFS -gt schema specification language
• Specifies ontological knowledge used in RDF
  statements
• Consists of a set of declarations of classes and
  properties
• Defines class and property hierarchies
• Multiple inheritance

7
RDFS Metamodel and RDFS Ontology
8
Pieces of Knowledge and Descriptions

• Piece of Knowledge -gt set of nodes directly
  connected with the resource
• Descriptionn -gt largest set of nodes connected
  with the resource and having a path length lt n
• Complete Description -gt the set of nodes
  connected to the resource through all possible
  properties

9
Piece of Knowledge Relative to Njal
10
D1(Njal) Description
11
D2(Njal) Description
12
Ontology Learning

• Systematically consider all concepts covering a
  set of resource nodes
• RDF graph resource extraction techniques
  preliminary first step
• Group concepts and resources based on intensions
  and extensions
• Incrementally build generalization hierarchy

13
Building of S1
14
D1(Njal) Concept Hierarchy
15
Hierarch Based on Descriptions of Length N

• Construct triples of intensions and related
  extensions
• Iteratively join triple L1 with triple in path
• Join all possible triples and paths
• Construct intensions of length n
• Build sets Sn from inclusion relations between
  node extensions

16
Building of S2 from S1
17
D2(Njal) Concept Hierarchy
18
Conclusion

• Lacks clarity
• Gaps in logic in explanation, S1 -gt Ontology
• Relies on RDF annotations previously generated
• Result complexity can increase exponentially
• Requires no training data
• Little or no user input
• Implemented and tested inside European IST Comma
  Project

19
Future Work

• Inclusion of heuristics
• Insertion of domain specific criteria
• Graphical UI
• Bounding methods to reduce complexity
• RDF annotation generator