The paper presents SERIMI, a class-based disambiguation approach for instance matching over heterogeneous web data. SERIMI uses a four step process: 1) clustering source instances into classes, 2) selecting blocking keys, 3) building pseudo-homonym sets, and 4) class-based disambiguation. It was evaluated on two datasets containing millions of triples and outperformed other approaches in precision, recall, and F1 score for instance matching when schemas have limited or no overlap.

1. SERIMI: Class-based
Disambiguation for Effective
Instance Matching over
Heterogeneous Web Data
Samur Araujo, DucThanh Tran, Arjen de Vries,
Jan Hidders, Daniel Schwabe
Delft University of Technology
WebDB 2012
Delft
University of
Technology

2. Me You
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 2

3. Apple
Me
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 3

4. You
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 4

5. ?
You
Ambiguous
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 5

6. Me
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 6

7. Me
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 7

8. Me
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 8

9. You
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 9

10. My Apple Your Apple
Spherical Shape Round Shape
Red Color Green Color
Eatable Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 10

11. My Apple Your Apple
Shape Shape
Color Color
Eatable Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 11

12. My Apple Your Apple
Spherical Shape Round Shape
Red Color Fruit Green Color
Eatable Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 12

13. My Apple Your Apple
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 13

14. Instance Matching
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 14

15. “Instance matching uses a
direct comparison paradigm”.
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 15

16. SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 16

17. Is your Apple like my
Apple?
Source
Humm..
Maybe!
Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 17

18. Homogenous data and schema.
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 18

19. The source and target
descriptions overlap.
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 19

20. Syntactic Overlap
Population = TotalPopulation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 20

21. Semantic Overlap
Population = Num_Inhabitants
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 21

22. Web of Data: heterogeneous
data and schema
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 22

23. None or limited overlap
between schemas
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 23

24. Instances do not instantiate
the schema, properly.
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 24

25. Apple
Nutritional Botanical
Information Information
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 25

26. “Direct comparison paradigm
does not apply”.
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 26

27. Apple
Me
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 27

28. Apple
Orange
Me
Pineapple
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 28

29. Apple
Me Orange
You
Pineapple
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 29

30. You
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 32

31. Food
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 34

32. SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 35

33. Eatable
Food
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 36

34. Source
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 37

35. My Apple Your Apple
My Orange Your Orange
My Pineapple Your Pineapple
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 38

36. SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 39

37. “We use a class-based
disambiguation paradigm …”
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 40

38. “We use a class-based
disambiguation paradigm …”
“… when there is no overlap
between schemas.”
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 41

39. Instance Matching with SERIMI
Source Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 42

40. Instance Representation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 43

41. Instance Representation
Predicate
Instance Value
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 44

42. Instance Representation
shape
Apple1 Round
title
Apple1 Apple
color
Apple1 Red
category
Apple1 Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 45

43. Instance Representation
shape
Apple1 Round
title
Apple1 Apple
color
Apple1 Red
category
Apple1 Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 46

44. Instance Representation
shape
Apple1 Round
title
Apple1 Apple
color
Apple1 Red
category
Apple1 Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 47

45. Instance Representation
[P(hi), D(hi), O(hi), T(hi)]
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 48

46. Step 1: Cluster the source
Cars
Source
Fruits
Companies
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 49

47. Step 2: Blocking Key Selection
Key
Selection
Source
instances
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 51

48. Step 2: Blocking Key Selection
key
Key
key
Selection
key
Source
instances
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 52

49. Step 2: Blocking Key Selection
e.g.Title
key
Key
key
Selection
key
Source
instances
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 53

50. Step 3: Pseudo-Homonyms Builder
Title=apple Pseudo-
Title=orange Homonyms
Title=pineapple Builder
Target
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 54

51. Step 3: Pseudo-Homonyms Builder
Everything
Target called Apple
Pseudo-
Homonyms
Builder
Source
instances
Target
Pseudo-homonyms
sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 55

52. Step 4: Class-based disambiguation
Target
Disambiguation
Class-based
Disambiguato
r
Pseudo-homonyms
sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 56

53. Step 4: Class-based disambiguation
Target
Target
Disambiguation
Class-based
Disambiguato
r Source
instances
Pseudo-homonyms
Pseudo-homonyms sets
sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 57

54. Step 4: Class-based disambiguation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 58

55. Step 4: Class-based disambiguation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 59

56. Step 4: Class-based disambiguation
h11 h21 h31
instances
h12 h22 h32
h13 h33
h14
H1 H2 H3
pseudo-homonym sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 60

57. Step 4: Class-based disambiguation
[P(hi11), D(hi11), O(hi11), T(hi11)]
h11 h21 h31
instances
h12 h22 h32
h13 h33
h14
H1 H2 H3
pseudo-homonym sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 61

58. Instance Representation
shape
Apple1 Round
title
Apple1 Apple
color
Apple1 Red
category
Apple1 Eatable
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 62

59. Step 4: Class-based disambiguation
h11 h21 h31 0.98 0.95 0.94
instances
h12 h22 h32 0.32 0.53 0.91
h13 h33 0.32 0.87
h14 0.76
H1 H2 H3 H1 H2 H3
pseudo-homonym sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 63

60. Step 4: Class-based disambiguation
h11 h21 h31
h12 h22 h32
h13 h33
h14
H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 64

61. Step 4: Class-based disambiguation
[P(hi11), D(hi11), O(hi11), T(hi11)]
h11 h21 h31
h12 h22 h32
h13 h33
h14
H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 65

62. Step 4: Class-based disambiguation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 66

63. Step 4: Class-based disambiguation
h11 h21 h31 0.98 h21 h31
h12 h22 h32 h12 h22 h32
h13 h33 h13 h33
h14 h14
H1 H2 H3 H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 67

64. Step 4: Class-based disambiguation
0.98 0.95 0.94
0.32 0.53 0.91
0.32 0.87
0.76
H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 68

65. Step 4: Class-based disambiguation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 69

66. Step 4: Class-based disambiguation
0.98 0.95 0.94
TOP-K or Threshold
H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 70

67. Step 4: Class-based disambiguation
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 71

68. Step 4: Class-based disambiguation
Target
Disambiguation
Class-based
Disambiguato
r Source
instances
Pseudo-homonyms
sets
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 72

69. Experiment
• Ontology AlignmentEvaluation Initiative (OAEI 2010)
• Collections: the life science (LS) collection (DBPedia, Sider,
Drugbank, LinkedCT, Dailymed, TCM, and Diseasome) and the
Person-Restaurant (PR)
• 20 gigabytes of data, millions of triples.
• We compared SERIMI to ObjectCoref and RiMON
• Precision, Recall and F1
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 73

70. Results
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 74

71. Results
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 75

72. Results
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 76

73. SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 77

74. Results
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 80

75. Step 4: Class-based disambiguation
0.98 0.95 0.94
TOP-K or Threshold
H1 H2 H3
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 81

76. Results for Top-K
1.00
0.90
0.80
0.70
0.60
0.50 Top-1
F1
0.40 Top-2
0.30 Top-5
0.20 Top-10
0.10
0.00
Sider-Daily. Sider-Drug. Drug.-Sider P11-P12
Dataset Pair
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 82

77. Results for δ threshold
1.00
0.90
0.80
0.70
0.60
δ >= δm
0.50 δ = 1.0
F1
0.40 δ >= 0.95
δ >= 0.90
0.30
δ >= 0.85
0.20
0.10
0.00
Sider-Daily. Sider-Drug. Drug.-Sider P11-P12
Dataset Pair
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 83

78. Conclusion
• SERIMI is complementary approach to direct-match based
instance matching tools.
• SERIMI is recommended for heterogeneous data where there is
no overlap between schemas.
• It is recommended for multi-class disambiguation.
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 84

79. THANK YOU!
• Samur Araujo
s.f.cardosodearaujo@tudelft.nl
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data
SERIMI: Class-based Disambiguation for Effective
Instance Matching over Heterogeneous Web Data 85