-
Most alignment research focuses on one-to-one correspondences between named
ontology elements [
1
], but these are not always sufficient for performing many
integration tasks [
2
]. Data values, for example, may need some form of translation, or
some form of condition may be required to scope a broader concept to correspond
with a narrower one. These correspondences, which contain conditions or
transformations, are known as complex correspondences.
There are many known patterns of complex correspondence [
2
]. Conditional
correspondences – where instances of a concept in one ontology are related to a
corresponding concept in the other ontology only if they have a particular value for a given
attribute – include Class by Attribute Type (CAT), Class by Attribute Value (CAV),
and Class by Attribute Existence (CAE). Similarly, Class by Attribute Path
Correspondences (PATH) occur when some path of attributes must be followed before the
scope of the more general concept can be narrowed. Correspondences where the value
of an attribute must be altered in some way are called Attribute Transformation
Correspondences (ATC).
In a sample of 50 concepts from YAGO2 [
3
], six of these concepts corresponded to
equivalent concepts in the DBpedia [
4
] ontology, and 14 concepts required a Class by
Attribute Value correspondence. Twenty-one concepts from YAGO2 corresponded
with DBpedia concepts with broader scope which could not be narrowed with a
correspondence pattern. Six YAGO2 concepts were aligned with DBpedia instances. We
found no cases of CAT or PATH correspondences.
Pattern Fitting
MRDM
Model Fitting
Boolean values
No
Numerical
Detecting Complex Correspondences
Approaches to detecting complex correspondences include a pattern based approach
[
5
], multi relational data mining (MRDM) [
6
] and our model based approach [
7
].
Each approach differs in the particular types of correspondence it can detect, and
these differences are outlined in table 1. The pattern based approach is the least
flexible. For attribute value based patterns it is only capable of detecting cases where
attributes have Boolean values. Each of the complex correspondences we found
between DBpedia and YAGO2 use non-Boolean attributes, and so it could not detect
these. The MRDM approach is more flexible, and is theoretically capable of finding
most correspondence patterns listed in section 1, except value transformation patterns.
Only the model fitting approach is capable of detecting value transformation
correspondences. The current implementation can detect numerical transformations, but the
approach could be extended to also detect transformations such as string splitting.
Acknowledgement: This research is supported by the Science Foundation Ireland
(Grant 08/SRC/I1403) as part of the FAME Strategic Research Cluster.