Transcript Ontology Mapping needs context & approximation

Ontology mapping
needs
context & approximation
Frank van Harmelen
Vrije Universiteit Amsterdam
Or:
How to make ontology-mapping
less like data-base integration
and
more like a social conversation
2
Three
Two obvious intuitions
The Semantic Web needs
ontology mapping
Ontology mapping needs
background knowledge
Ontology mapping needs approximation
3
Which Semantic Web?
Version 1:
"Semantic Web as Web of Data" (TBL)
recipe:
expose databases on the web,
use RDF, integrate
meta-data from:

expressing DB schema semantics
in machine interpretable ways
enable integration and unexpected re-use
4
Which Semantic Web?
Version 2:
“Enrichment of the current Web”
recipe:
Annotate, classify, index
meta-data from:

automatically producing markup:
named-entity recognition,
concept extraction, tagging, etc.
enable personalisation, search, browse,..
5
Which Semantic Web?
Version 1:
“Semantic Web as Web of Data”
Version 2:
“Enrichment of the current Web”
 Different use-cases
 Different techniques
 Different users
data-oriented
user-oriented
6
Which Semantic Web?
Version 1:
“Semantic Web as Web of Data”
Version 2:
“Enrichment of the current Web”
 But both need ontologies
for semantic agreement
between sources
between source & user
7
Ontology research is
almost done..
we know what they are

“consensual, formalised models of a domain”
we know how to make and maintain them
(methods, tools, experience)
 we know how to deploy them
(search, personalisation, data-integration, …)
Main remaining open questions
 Automatic construction (learning)
 Automatic mapping (integration)
8
Three obvious intuitions
The Semantic Web needs ontology mapping
Ontology mapping needs
background knowledge
Ph.D. student
?
=
AIO
Ontology mapping needs approximation
young
researcher
?
≈
post-doc
9
This work with
Zharko Aleksovski &
Michel Klein
Does context knowledge help
mapping?
The general idea
background
knowledge
anchoring
anchoring
inference
source
target
mapping
12
a realistic example
 Two Amsterdam hospitals (OLVG, AMC)
 Two Intensive Care Units, different vocab’s
 Want to compare quality of care
 OLVG-1400:




1400 terms in a flat list
used in the first 24 hour of stay
some implicit hierarchy e.g.6 types of Diabetes
Mellitus)
some reduncy (spelling mistakes)
 AMC: similar list, but from different hospital
13
Context ontology used
DICE:



2500 concepts (5000 terms), 4500 links
Formalised in DL
five main categories:
• tractus (e.g. nervous_system, respiratory_system)
• aetiology (e.g. virus, poising)
• abnormality (e.g. fracture, tumor)
• action (e.g. biopsy, observation, removal)
• anatomic_location (e.g. lungs, skin)
14
Baseline: Linguistic methods
 Combine lexical analysis with hierarchical structure
 313 suggested matches, around 70 % correct
 209 suggested matches, around 90 % correct
 High precision, low recall (“the easy cases”)
15
Now use background knowledge
DICE
(2500 concepts,
4500 links)
anchoring
anchoring
inference
OLVG
AMC
(1400, flat)
(1400, flat)
mapping
16
Example found with context
knowledge (beyond lexical)
17
Example 2
18
Anchoring strength
Anchoring = substring + trivial morphology
anchored on N aspects
N=5
N=4
N=3
N=2
N=1
total nr. of anchored terms
total nr. of anchorings
OLVG
0
0
4
144
401
AMC
2
198
711
285
208
549 39% 1404 96%
1298
5816
19
Experimental results
 Source & target =
flat lists of ±1400 ICU terms each
 Background = DICE (2300 concepts in DL)
 Manual Gold Standard (n=200)
21
Does more context
knowledge help?
Adding more context
 Only lexical
 DICE (2500 concepts)
 MeSH (22000 concepts)
 ICD-10 (11000 concepts)
 Anchoring strength:
DICE
MeSH
ICD10
4 aspects
0
8
0
3 aspects
0
89
0
2 aspects
135
201
0
1 aspect
413
694
80
total
548
992
80
23
Results with multiple ontologies
Separate
Recall
Precision
Lexical ICD-10
64%
64%
95%
95%
Joint
 Monotonic improvement
 Independent of order
 Linear increase of cost
100
90
80
70
60
50
40
30
20
10
0
Lexical
DICE
MeSH
76%
88%
94%
89%
ICD-10
DICE
MeSH
24
does structured context
knowledge help?
Exploiting structure
 CRISP: 700 concepts, broader-than
 MeSH: 1475 concepts, broader-than
 FMA: 75.000 concepts, 160 relation-types
(we used: is-a & part-of)
FMA
(75.000)
anchoring
anchoring
inference
CRISP
(738)
MeSH
(1475)
mapping
26
Using the structure or not ?
(S <a B) & (B < B’) & (B’ <a T) ! (S <i T)
a
a
i
27
Using the structure or not ?
a
(S < B) & (B < B’) & (B’
a
i
< T) ! (S < T)
No use of structure
Only stated is-a & part-of
Transitive chains of is-a, and
transitive chains of part-of
Transitive chains of is-a and part-of
One chain of part-of before
one chain of is-a
28
Examples
29
Examples
30
Matching results (CRISP to MeSH)
(Golden Standard n=30)
Recall
Exp.1:Direct
Exp.2:Indir. is-a + part-of
Exp.3:Indir. separate closures
Exp.4:Indir. mixed closures
Exp.5:Indir. part-of before is-a
=
·
¸
448 417 156
395 516 405
395 933 1402
395 1511 2228
395 972 1800
Precision
=
·
¸
Exp.1:Direct
Exp.4:Indir. mixed closures
Exp.5:Indir. part-of before is-a
17
14
14
18
39
37
3
59
50
total
incr.
1021
1316
2730
4143
3167
29%
167%
306%
210%
total correct
38
112
101
100%
94%
100%
31
Three obvious intuitions
The Semantic Web needs ontology mapping
Ontology mapping needs
background knowledge
Ontology mapping needs
approximation
young
researcher
?
≈
post-doc
32
This work with
Zharko Aleksovski
Risto Gligorov
Warner ten Kate
Approximating subsumptions
(and hence mappings)
query: A v B ?
B = B1 u B2 u B3 A v B1, A v B2, A v B3 ?
B2
B
B1
A
B3
34
Approximating subsumptions
Use “Google distance” to decide which
subproblems are reasonable to focus on
Google distance
max{lo
f
(
x
),
log
f
(
y
)}

log
f
(
x
,
y
)
NGD
(
x
,
y
)

log
M

min{l
f
(
x
),
log
f
(
y
)}
where
f(x) is the number of Google hits for x
f(x,y) is the number of Google hits for
the tuple of search items x and y
M is the number of web pages indexed by Google
35
Google distance
animal
sheep
cow
plant
vegeterian
madcow
37
Google for sloppy matching
 Algorithm for A v B
(B=B1 u B2 u B3)
 determine NGD(B, Bi)=i, i=1,2,3
 incrementally:
• increase sloppyness threshold 
• allow to ignore A v Bi with  i · 
 match if remaining A v Bj hold
38
Properties of sloppy matching
When sloppyness threshold  goes up,
set of matches grows monotonically
=0: classical matching
=1: trivial matching
?
Ideally: compute i such that:
 desirable matches
become true at low 
 undesirable matches
become true only at high 
Use random selection of B as baseline 39
Experiments in music domain
CDNow (Amazon.com)
Size: 2410 classes
Depth: 5 levels
ArtistGigs
Size: 382 classes
Depth: 4 levels
Artist Direct Network
Size: 465 classes
Depth: 2 levels
very sloppy terms
 good
CD baby
Size: 222 classes
Depth: 2 levels
Yahoo
All Music Guide
Size: 96 classes
Depth: 2 levels
Size: 403 classes
Depth: 3 levels
MusicMoz
Size: 1073 classes
Depth: 7 levels
40
Experiment
Manual Gold Standard, N=50, random pairs
 =0.53
97
 =0.5
precision
60
16-05-2006
classical
20
recall
random
NGD
7
wrapping up
Three obvious intuitions
The Semantic Web needs
ontology mapping
Ontology mapping needs
background knowledge
Ontology mapping needs approximation
43
So that
shared context & approximation
make ontology-mapping
a bit more like a social conversation
44
Future: Distributed/P2P setting
background
knowledge
anchoring
anchoring
inference
source
target
mapping
45
Vragen & discussie
[email protected]
http://www.cs.vu.nl/~frankh
46