In Linked Data, the use of owl:sameAs is ubiquitous in 'inter-linking' data-sets. However, there is a lurking suspicion within the Linked Data community that this use of owl:sameAs may be somehow incorrect, in particular with regards to its interactions with inference. In fact, owl:sameAs can be considered just one type of 'identity link,' a link that declares two items to be identical in some fashion. After reviewing the definitions and history of the problem of identity in philosophy and knowledge representation, we outline four alternative readings of owl:sameAs, showing with examples how it is being (ab)used on the Web of data. Then we present possible solutions to this problem by introducing alternative identity links that rely on named graphs.
As large numbers of independently developed data-sets have been introduced to the Web as Linked Data, the vexing problem of identity has returned with a vengeance to the Semantic Web. As the ubiquitous owl:sameAs property is used as the RDF property to connect these data-sets, it has been dubbed the owl:sameAs problem. However, the problem of identity lies not within Linked Data or within the Semantic Web languages, but is an outstanding and wellknown – if sometimes not precisely labeled – issue in preSemantic Web knowledge representation languages in artificial intelligence. What precisely is new in its latest guise of this problem on the Web of Linked Data is that this is the first time the problem is being encountered by different individuals attempting to independently knit their knowledge representations together using the same standardized language. Much of the supposed “crisis” over the proliferation of owl:sameAs in Linked Data can be traced to the fact that these uses of owl:sameAs tend to be mutually incompatible, and almost always violate the rather strict logical semantics of identity demanded by owl:sameAs. However, the exact types of distinctions made by these individuals are important, even if they contradict the relevant specification of owl:sameAs. First, these uses and abuses of owl:sameAs demonstrate for the first time in the history of knowledge representation how precisely these problems play out in the wild. Second, as the Semantic Web is a project in development, it is always possible to specify anew different and new kinds of language constructs and more clearly specified best practices to align the specifications with the actual empirical use of the Semantic Web in the wild.
First, we will give an overview of the problem of identity and its somewhat dusty lineage in artificial intelligence, if only to show how what was already a known issue for knowledge representation becomes even more exacerbated when knowledge representation goes global for the Semantic Web. Then, four distinct uses of owl:sameAs are discussed in addition to the precise idea of “same thing as,” namely:
Finally, a number of suggestions for how the current situation can be improved are sketched. The necessity of both semantic and theoretical work is given as well. 2.
The problem of identity has a long and chequered history, spreading from philosophy and mathematics to linguistics and knowledge representation. In each of these fields, what it means for two things to be identical goes straight to the heart of semantics. 2.1
The father of knowledge representation, Leibnitz, was not
surprisingly also the first to phrase a coherent and
formalizable definition of identity, often called ‘Leibnitz’s Law’ or
the ‘The Identity of Indiscernables,’ namely that for every x
and every y, whatever is true of x is true of y, then x is
identical to y [
A number of classical problems already crop up in this analysis of identity. First, exactly what properties are being counted? Obviously, we can imagine worlds where things have the exact same properties but are nonetheless not identical, such as an exact clone (which is all too easy with digital objects). There are two obvious escape hatches: a thing may have a property of “being=itself” (an [haecceity) or things having different temporal-spatial co-ordinates could be counted as different, even if they share the rest of their properties in common. While that sounds like a common-sense distinction, is it true that Tim Berners-Lee is the same person he was when he was a child? Or if he lost his arm? This leads straight into arguments about perdurance and endurance in philosophy. Are there two different kinds of properties, properties that are somehow intrinsic to identity and others that are extrinsic, i.e. purely relative to other things? Lastly, perhaps the real question is who or what determines the conditions of identity, namely that identity can only be made in context of an explicit theory of identity criteria. These theories can be formalized in terms of the semantic interpretation of sentences to referents. However, this does not mean that these theories are compatible. If one has a theory of identity where one is talking only about people as employees in a particular role, then two different people who have the same job will be identical, but if one has a more fine-grained interpretation the very same people would be different. One can even imagine theories of identity based on different criteria, where some theories of identity subsume weaker or stronger ones. There is also the problem of vagueness, and the inability to specify precise properties (such as the exact latitude and longitude boundaries of Mount Everest). Regardless of the problems, the point of Leibnitz’s Law is clear: When someone says two things are the same, they mean they share all the same properties.
Frege was the first to note what Galois called the
“linguistic counterpart” to Leibnitz’s Law: the Principle of
Substitution, which states that if x and y are identical, then x
may be substituted for y without changing the truth-value
of the sentence in which the substitution is made [
There are two important consequences. The first is the
classic division between use and mentioning of a name.
Even if “Mount Everest” in English and “Qomolangma”
in Tibetan mean the same thing, names from different
languages cannot be substituted for each other often. Sentences
like “Qomolangma is a word in Tibetan” mention a name,
while the sentence “Mount Everest is the highest mountain
in the world” uses the name. Obviously, “Mount Everest is
a word in Tibetan” is false. Is “Qomolangma is the highest
mountain in the world” true? This fact was not necessarily
known in Tibet before the era of global geological surveys.
So one could easily have a case of a geoscientist who has
never visited the mountain knowing it is the highest
mountain in the world and a Tibetan monk who lives not too far
from the mountain not knowing - or caring - if it is the
highest mountain in the world. This distinction was called the
distinction between two names having the same referent but
different senses, i.e. contexts that do or do not share certain
information [
It would be easy to dismiss these arguments over identity
as being mandarin philosophical questions, until the ‘pedal
hits the metal’ in the the world of knowledge representation.
This is precisely what happened to semantic networks, a
predecessor of the Semantic Web in knowledge representation.
Semantic networks, as pioneered by Quillian, were viewed
as an alternative knowledge representation scheme to
firstorder logic in the early days of artificial intelligence [
Semantic networks, by relying on words from natural
language or pseudo-words to label their constructs, whose
meanings were somehow supposed to be simply obvious, actually
led to these constructs being ambiguous. The classic
example was the infamous IS-A label used by Brachman in his
What IS-A is and isn’t. An Analysis of Taxonomic Links in
Semantic Networks [
In an effort to remedy this crisis, Brachman and others
decided to split what they called the “epistemological level”
- the kinds of nodes and edges that remained neutral to the
underlying primitives yet could be given a specific semantics
from the rest of the semantic network whose meaning could
only be grounded in some linguistic convention [
The Semantic Web seems to have learned from semantic
networks. The formal semantics of RDF are important
precisely because RDF statements can be given the same logical
meaning uniformly across a distributed network, even if the
semantics of RDF have relatively light inferential power and
do not constrain the semantic interpretations very tightly
[
Contrary to popular belief in some circles, formal semantics are not a silver bullet. Just because a construct in a knowledge representation language is prescribed a behavior using formal semantics does not necessarily mean that people will follow those semantics when actually using that language “in the wild.” This can be laid down to a wide variety of reasons. In particular, the language may not provide the facilities needed by people as they actually try to encode knowledge, so they may use a construct that seems close enough to their desired one. A combination of not reading specifications - especially formal semantics, which even most software developers and engineers lack training in - and the labeling of constructs with “English-like” mnemonics naturally will lead to the use of a knowledge representation language by actual users that varies from what its designers intended. In decentralized systems like the Semantic Web, this problem is naturally exacerbated. However, far from being a sign of abuse, it is a sign of success, as it means that the Semantic Web is actually being deployed outside academia and research labs.
The problem has definitely arisen on the Semantic Web
in terms of the use of owl:sameAs in Linked Data. In
Linked Data, each item of interest is given a URI, that in
turn redirects to either human-readable HTML or
machinereadable RDF depending on content negotiation. The URI
for the item itself, which is called rather confusingly a
“noninformation resource” in Linked Data circles, as a web-page
or RDF graph would be an information resource, as the
“ distinguishing characteristic of these resources is that all
of their essential characteristics can be conveyed in a
message” [
The most typical link used is owl:sameAs, which is in
general used to to say “that two URI references actually refer
to the same thing” [
At first, this use of owl:sameAs seems to be harmless.
Its actual definition is that “the built-in OWL property
owl:sameAs links an individual to an individual” and “Such
an owl:sameAs statement indicates that two URI references
actually refer to the same thing: the individuals have the
same identity” [
However, owl:sameAs does have a particular semantics of individual identity, namely that the two individuals are exactly the same and so share all the same properties, and thus are equivalent in terms of Leibnitz’s identity of indiscernables. Given that OWL has no unique name assumption, once there is an application of owl:sameAs to two different URIs, then any statement that is given to a single URI is true for every other URI that has an owl:sameAs link. Furthermore, while in OWL Full owl:sameAs can be considered to be the same as between any URIs as classes can be considered “individual” instances of other classes and properties can be considered individuals, in OWL DL in order to preserve decidability individuals are strictly separated from classes, and so one should use OWL DL equivalentClass and equivalentProperty instead. Therefore, quick-and-dirty use of owl:sameAs will almost always lead to OWL Full, which very little work has been done on in terms of efficient implementations of inference. The real trick with owl:sameAs is that it works both ways: as it is both symmetric and transitive, so that anyone can link to your data-set with owl:sameAs from anywhere else on the Web without your permission, and any statement they make about their own URI will immediately apply to yours. As imaginable, such transitive closures can immediately get very large. There have been considerable rumors in the Linked Data community that such use of owl:sameAs is somehow “wrong” with regards to the formal semantics of OWL. It does seem intuitively that the use of owl:sameAs may be the logical equivalent of a bulldozer. Since inference is rarely used on the Linked Data, these possible side-effects have not been noticed. Does this really matter? Is the use of owl:sameAs an exploding time-bomb for Linked Data, or a harmless convention? What exactly is the point of linking data if nobody is going to draw any conclusions which use the links? 5.
The first case is when the two URIs do refer to the same thing, but all the properties ascribed to one URI are not necessarily accepted by the other. This means that the use of the URI is referentially opaque, which means that one URI cannot be substituted for another (the Principle of Substitution is violated), i.e. the context is intensional. A classic example of this would be the the concept of sodium in DBpedia, which has an owl:sameAs link to the concept of sodium in OpenCyc. The OpenCyc ontology says that an element is the set (class) of all pieces of the pure element, so that for example sodium in Cyc has a member which is the lump of pure metallic sodium. On the other hand, sodium as defined by DBPedia is used to also include isotopes, which have different number of neutrons than “standard” sodium. So, one should not state the number of neutrons in DBPedia’s use of sodium, but one can with OpenCyc. Therefore, owl:sameAs here is in error, as it does not allow mutual substitutivity. Indeed, this use of URIs in an opaque referential context is likely what most uses of owl:sameAs actually are for, as it is unlikely that most deployers of Linked Data actually check whether or not all the properties and their associated inferences are shared amongst linked data-sets. This property is exceedingly important for Linked Data, as contrary to popular doctrine, it is possible that the Web is full of referentially opaque contexts. The problem is there is no way to get a handle on contexts informally without descending into non-logical reasoning currently. 5.2
In this case, two URIs do refer to the same thing and all properties do hold of both URIs, but that we cannot re-use the URI in a different context. The central intuition here is there are ’forms of reference’ appropriate to a context, especially in social contexts. To use an example from Lynn Stein, when at a meeting of the PTA (Parent-Teacher Association) she is Ms. Stein, Rachel’s mum, not Professor Stein of MIT. This does not mean that in the PTA meeting Ms. Stein is somehow not a professor, but that within that context those properties do not matter. At first, this distinction may not seem directly relevant to linked data, provided we keep ’name’ in the social sense distinct from ’identifier’ in the Web sense. However, this distinction raises other issues about what kind of ’names’ URIs really are and precisely why certain properties for linked data are given in the RDF description of a certain URI and others are not. 5.3
Often identity is conflated with representation. While the term “representation” is often very contentious, its intuitive definition is that, just as a picture of something depicts something, a URI can be for a representation of a thing rather than the thing itself. Intuitively, there seems to be a clear-cut line between that which represents something (the representation) and that which is represented (the referent), sometimes called the relationship between a “sign” and a “signifier.” However, the relationship is often not as clear-cut as we would lead ourselves to believe. In fact, in human natural language use-mention confusions are ubiquitous and often useful. For example, often a web-page or an e-mail address are used to refer to a person. Rather than yell at the world to get an education in philosophical logic, it may be better to clarify this relationship. It also might be worth distinguishing between using a representation to refer to the represented, such as using a picture of Berners-Lee to refer to Tim Berners-Lee himself, using something accidentally or contextually to refer to something, a phenomenon called displaced reference. The example of using an e-mail box to refer to a person is not an error but rather more a displaced reference. 5.4
Sometimes its clear that two things are not identical but simply closely related in some manner. This, for example, is the relationship between the district of Paris and the Department of Paris in Cyc. Furthermore, there are often complex, structured, yet hard-to-specify relationships between things, such as the relationship between isotopes and elements, the quantity and a measurement of a quantity, and an image and a facsimile of that image. In web architecture, it is clear there is a close relationship These relationships that are ‘very similar to’ seem to deserve their own property, but are often currently lumped together in Linked Data under the all-encompassing use of owl:sameAs. Most of the more noticeable errors of owl:sameAs seem to come from this category, and it is likely that examples such as the relationship of sodium within DBPedia to sodium in OpenCyc are of this kind as well.
Surprisingly, most of the time people use owl:sameAs they
are accidentally doing what is sort of an implicit import of
statements of the subject of the owl:sameAs statement.
Obviously, to address the weaker identity implied by the
referentially opaque use of identity, a weaker version of owl:sameAs
should be specified that does not import all the properties in
a full transitive closure.Somewhat similar predicates already
exist in SKOS as skos:exactMatch and skos:closeMatch, but
their use seems rare in Linked Data [
There is already a notion of context built into RDF, namely
named graphs [
The use of owl:sameAs is already a sort of statement of this kind in the FOAF vocabulary, the foaf:isPrimaryTopicOf statement. One possible solution to this problem would be to wrap such a property into some core W3C approved standard. However, the problem is that it is unclear if a strict separation between mention and use is necessary or even desirable. In many contexts, as relevant experience in OpenID deployment shows, using an e-mail as an identifier for a person is often more natural than the URI of a home-page, or even a “non-information resource.” What is needed however, is a way to make the distinctions that either conflate or separate mention and use or on the fly. The use of weak identity statements - and in this case, a “represents” statement - and explicit importing and de-importing of properties within the context of particular named graphs would allow us to do state things like “Within this named graph and only within this named graph, the e-mail address URI is identical to the person and shares their properties” and “Within this other named graph, the e-mail address represents the person, but does not have all the properties of that person.” 6.4
Again, the tempting easy solution is simply to introduce a new predicate for “very similar to.” The SKOS vocabulary has a number of “matching” predicates that are close in meaning to this, ranging from hierarchically structured skos:broadMatch and skos:narrowMatch to the more suitable skos:closeMatch. However, the main issue with these predicates is that again, their use may be a matter of opinion, as someone’s close match may be another person’s identical match. One is also tempted to engage with some sort of “fuzzy” or numerically weighted uncertainty measure between one and zero of identity, but then the real hard questions of where precisely will these real values come from and their relationship to actual probability theory muddies these conceptual waters quickly. It seems that beneath this apparently simple property is likely a whole family of heterogeneous and semi-structured identity relationships that should be studied more carefully and empirically observed before any hasty judgments are made.
It is possible to do empirical studies of exactly how people use owl:sameAs in the wild. Examples of owl:sameAs can be taken from the Linked Data Web in the wild in order to determine how experimentally robust these distinctions are would be, i.e. do people actually use owl:sameAs in the four ways that are outlined above, and are there more possible ways that we are not aware of? In fact, even the ability to recognize these kinds of distinctions may vary quite wildly by background and training. Lastly, if a number of empirical distinctions between identity links that are currently conflated by owl:sameAs can be made in a robust manner, then there is considerable formal semantic work to be done. Giving the Linked Data community well-defined (both formally and informally) predicates should be done even when one does think of the properties given to URIs as absolute truths given by Linked Data publishers or W3C specifications, but as functions of their actual use. The (ab)use of owl:sameAs in Linked Data is not a threat, it’s an opportunity.