Many Linked Data applications have come to rely on owl:sameAs for linking datasets. However, the current semantics for owl:sameAs assert that identity entails isomorphism, or that if a=b, then all statements of a and b are shared by both. This becomes problematic when dealing with provenance, context, and imperfect representations, all of which are endemic issues in Linked Data. Merging provenance can be problematic or even catastrophic in biomedical applications that demand access to provenance information. We use examples in biospecimen management, experimental metadata representations, and personal identity in Friend-of-a-Friend (FOAF) to demonstrate some of the problems that can arise with the use of owl:sameAs. We also show that the existence of an isomorphic owl:sameAs can be inconsistent with current expectations in a number of our use cases. We present a solution that allows the extraction of isomorphic statements without requiring their direct assertion. We also introduce a set of identity properties that can be extended for domain-speci c purposes while maintaining clarity of de nition within each property.
The adoption of Linked Data by many communities has resulted in a wealth of
useful data that can be combined in many ways. One current practice in Linked
Data, and therefore in many resources on the Semantic Web, is to link datasets
using owl:sameAs to de ne that two entities between those datasets are the
same. On its face, this is a good idea because a higher level best practice is to
re-use properties when possible and appropriate and owl:sameAs is a property
that seems easy to understand and is part of OWL [
The property owl:sameAs is a very strict notion of identity. Its de nition
stems from that of mathematical identity, most speci cally, from the
Indiscernibility of Identicals [
We argue that one major issue with the use of owl:sameAs in Linked Data and the Semantic Web is related to the inference of isomorphism. We present some biomedical examples that help to describe some problems with owl:sameAs and isomorphic inference. Finally, we discuss a set of possible identity properties and show how the current uses outlined by Halpin and Hayes and other examples that we describe t into the new framework. 2
We present some examples that help to describe problems with owl:sameAs and
isomorphic inferencing. For instance, Ja ri et al. discusses a number of erroneous
mappings of owl:sameAs in DBpedia and DBLP [
There are also issues with creating Linked Data while maintaining provenance information. An example that we have found is in biomedical data. In Fig. 1, we show a common derivation graph of a tumor T that was removed from a patient A. A cell line was derived from the tumor which resulted in two specimens, LB and LA, where LA was the original cell colony and LB was derived from it by taking a part of LA and growing it on its own. There are datasets (D; E) that use (LB; LA) respectively, and a scientist would like to integrate those datasets. Through discussions with practicing scientists, we became aware of situations where owl:sameAs is used to accomplish this. For example, the use of owl:sameAs in Fig. 2 allows a scientist to integrate the datasets shown. However, isomorphism is applied to all other properties of the specimens, resulting in ambiguous or contradictory information. For instance, both specimens are inferred to have been created on both 8/31 and 9/20, and have a quantity of both 5 and 10 grams. Also, specimen LA now seems to be derived from itself. It has been noted that these properties could be set as annotations, which would mean that they are not subject to isomorphism. However, most of the properties speci ed here are rst-class data as speci ed by biospecimen management systems and would not be considered annotations by the originating system. There are other issues with this sort of inference. It is now very di cult to address the following potential problems that can arise: { The data doesn't look right. What were the methods and protocols, and how consistent were they, going back to surgical resection? { Did the \same cell line" actually come from the same tumor, or just from the same patient? Or even di erent patients? { What originally seemed to be a primary breast cancer or lung cancer is now a metastasized melanoma. How do we sort this out? { Is a histology slide made from a tumor the same as the tumor? What about the tissue microarray, the cell culture, or the isolated molecular material? None of these issues are problems that always come up, but when they do, it is critical that the provenance of biospecimens remains distinguishable.
Some have argued that these problems can be surmounted by putting the o ending statements in a named graph, or by attaching provenance statements as OWL 2 annotations on individuals. These solutions may work for limited contexts, but there are issues with both. Relying on named graphs is problematic because the statements may be embedded in other datasets which are already enclosed in named graphs. The user would have to extract identity statements from the original named graph and move them to another. Using annotations would work well if the problem were limited to provenance, and provenance were easily segregated into annotation properties. However, one data model's provenance is another's data. Biospecimen management systems concern themselves almost exclusively with what would be considered provenance to the experimentalist and analyst. It must be possible to realize ex post facto that large sections of a data model are considered to be provenance by other users. As we will show, this problem is not con ned to provenance, but applies more generally to statements of knowledge. Neither of these solutions address these issues. Fig. 2. As in Fig. 1, but with the assertion owl:sameAs(LA,LB), then D and E can be integrated because they can refer to the same specimens. However, doing so means that there are now multiple values for some important properties and LA appears to have been derived from itself. 3
There are some additional problems with using the indiscernibility of identicals
in knowledge representation discussed by Saul [
We can then infer that Lois Lane claims that Superman cannot y, and that Clark Kent can. These statements in combination with the assertions generate contradictions. Similar problems arise even without belief statements: 4. Clark Kent is Superman's secret identity. 5. Therefore, Superman is Superman's secret identity.
This is nonsensical, as a secret identity must be di erent from a public identity. This is also true of changes in identity over time: 6. I never made it to Constantinople, but I visited Istanbul last week. 7. Istanbul was Constantinople. 8. I never made it to Istanbul, but I visited Istanbul last week.
These examples all violate Leibniz's law regarding the identity of indiscernibles
[
A New
The examples above indicate a need for an identity (or similar) notion that is
not inherently isomorphic. We also have a need to be able to nd all entities
that are inferred (or stated) to be identical to a designated entity. The modeling
we need to do that is driven by these and other use cases, lead us to desire
more options related to the concept of identity. As mentioned above, Halpin and
Hayes [
The property owl:sameAs is, in addition to being isomorphic, Transitive, Symmetric, and Re exive. Each permutation of those meta-properties can be viewed as a new kind of identity. We have de ned a new ontology called the Identity Ontology (IO).1 The properties of the IO are shown in Table 1. Domain-speci c properties can be created as sub-properties of one of the eight IO properties in order to maximize interoperability while maintaining distinctions among future concepts of identity. We have also de ned a mapping ontology that shows examples of mappings with existing identity properties from RDFS, OWL, and SKOS2 and show the subproperty relationship among the new and existing identity properties in Fig. 3.
id:identical This is the most restrictive property of identity in IO. It follows the
same de nition as owl:sameAs, which \indicates" that two URI references
actually refer to the same thing: the individuals have the same `identity'."
[
The usages \Same Thing As But Di erent Context" and \Same Thing As But
Referentially Opaque" from Halpin and Hayes [
Symmetric, it serves as a way for one entity to claim the identity of another,
without the inverse needing to be true. As a super property of id:identical,
everything that is actually identical makes the claim of identity, with both
sides of the claim being made due to the symmetry of id:identical. This
property is transitive because if entity a claims to be entity b and b claims
to be entity c, then a cannot deny that it is claiming to be c as well.
The usage \Represents" can be supported using id:claimsIdentical using a
subproperty \representedBy". Since id:claimsIdentical suggests that a can be
replaced by b if claimsIdentical(a; b), then it can be said that b represents a, or
to t into our usage more clearly, representedBy(a; b).
id:matches This property is re exive and symmetric. It is inspired by
skos:exactMatch, which \indicates a high degree of con dence that two concepts can
be used interchangeably across a wide range of information retrieval
applications." [
For extremely strong assertions of \Very Similar To" from Halpin and Hayes
[
This property is closely related to and is a super property of skos:related [
An example subproperty of id:claimsRelated is a depiction. Since a photograph or a illustration of a person or thing is not the thing itself, but a representation of the thing, this is a kind of identity that is not symmetric (the photograph is not depicted by the person), not transitive (a depiction of the depiction may not depict the original subject), and not re exive (does a person depict themselves?).
These properties cover the wide range of identity relationships from \a is the same thing as b" to \b has more information about a" and allow the expression of precise concepts of identity while also leaving room for domain-speci c concepts as well. 4.2
For any re exive statement of identity, it is possible to recover isomorphic statements using the following SPARQL query snippet: select ?s, ?p, ?o where { ?s id:identical ?x.
?x ?p ?o.
}
A major bene t of this formulation is that any property can be used in place
of id:identical and can be used for domain-speci c concepts of identity.
Additionally, property chains in OWL 2 [
speci c properties where that behavior is warranted. The speci c pattern would be:
SubObjectP ropertyOf (ObjectP ropertyChain(identical; p); p) It is therefore possible to construct properties that are isomorphic across speci c concepts of identity and allows users to query for values for any other property that would have been isomorphic if the identity had been asserted using owl:sameAs. 5
We and others [
The Identity Ontology is a starting point for developing a more nuanced approach to identity in the semantic web. IO addresses numerous challenges in our biomedical examples, and we have begun to use IO to represent concepts of identity in biomedical datasets. Speci cally, we integrated two experiments from Array Express: E-TABM-65 and E-MEXP-1029. Both of these experiments use the NCI-60, a panel of cell lines used for cancer research. We converted the two experiments to RDF using MAGETAB2RDF5 and aligned the biological sources using biomedidentity:sameAsBioSource.6 In that ontology, we also make mage:has derivative,7 and mged:has biomaterial characteristics 8 isomorphic across our identity property using property chains. We plan to continue to investigate the properties of identity in relation to constructs included in IO as well as owl:sameAs. 6
We have elaborated on problems with isomorphism in the current use of owl:sameAs in Linked Data on the Semantic Web. We have provided more options for representing identity using our Identity Ontology and have initial work supporting its usage in a number of use cases. We also show how isomorphic statements can be queried, and how particular properties can be made to be isomorphic using property chains in OWL 2. We have also successfully used the Identity
Ontology to enable more granular control over inference. We have found that this additional control over inferred information is a better match to our biomedical application needs than what we previously had access to using owl:sameAs alone.