Techniques for segmenting large description logic ontologies

1
Techniques for segmenting large description logic
ontologies

• Julian Seidenberg, Alan Rector
• julian.seidenberg_at_cs.man.ac.uk
• Project CO-ODE / HyOntUse
• http//www.co-ode.org

2
Ontology Labyrinth
3
Segment Corridor
4
Overview

• Need for segmentation
• 3 styles of segmentation
• Segmentation by traversal
• Evaluation

5
Ontology?!

• Definition
• An ontology describes concepts in a domain of
  interest and the relationships that hold between
  them.

6
Impossible to understand
Stack overflow! Out of memory!
Where am I? What do I do?
7
Segment!

• Browse segments
• Query segments
• Annotate segments
• Compare segments
• Evaluate segments
• Discuss segments
• Publish segments
• Transform segments
• Provenance using segments
• Plug segments into application

8
Query-based methods
?
?
9
Network partitioning
10
Extraction by traversal
?
11
Intra-taxonomy relationships Properties link
between Concepts in a hierarchy

• Heart ?
• is-a InternalOrgan
• isPartOf CardiovascularSystem

12
Basic segmentation algorithm
13
Filtering segmentationInitial state

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

14
Filtering segmentationAdded reciprocal links

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

15
Side noteHow reciprocals are different from
inverses

• Heart isPartOf Torso
• Torso hasPart Heart

• Heart isPartOf Torso
• Heart hasPart Torso (!)

16
Filtering segmentationFiltered out isPartOf
relations

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

17
Filtering segmentationFollowing links (1)

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

18
Filtering segmentationFollowing links (2)

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

19
Filtering segmentationFollowing links (3)

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

20
Filtering segmentationFollowing links (4)

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

21
Filtering segmentationComplete

1. Add missing reciprocal links
2. Filter out isPartOf relations
3. Follow links to select classes to extract

22
Boundary extract
23
Test case

• Big (23,000 concepts)
• Complex (30,000 relationships)
• Representative (basis of many future systems)

24
Basic segment size
25
Property filtered size
26
Property filtered classification
27
Boundary depth vs. size
28
Summary

• 3 different styles of segmentation
• Query, network partitioning traversal
• Basic extract by traversal
• Reduces size
• Non-destructive
• Multi-purpose
• Property filtering
• Focuses subject-area
• Increases tractability
• Facilitates ontology analysis
• Boundary limiting
• Accurately reduces size

29
Questions?