A common method for contextualizing facts in knowledge graph formalisms is by adding property-value pairs, qualifiers, to the facts in the knowledge graph. Qualifiers work well for information that is additional to the base fact but pose an unwarranted burden on consumers of the information in knowledge graphs when the qualifier instead contextualizes the base fact, as in limiting the applicability of the fact to some time period or providing a confidence level for the fact. Contextualization should instead by done in a more principled manner and accompanied by tools that lessen the burden on consumers of knowledge graphs.
Consider first using qualifiers to represent the role played by a cast member of a film. This information is somewhat awkward to represent in triple-based formalisms because it is a three-place relationship. So Wikidata and schema.org use qualifiers to attach the role played to the base information that an actor is in the cast of a film.
RDF reification and singleton properties can use qualifiers for this as well. In RDF
reification [
_:rt rdf:object :ClarkGable . where _:rt is a statement node that represents the triple
:GWTW :castMember :ClarkGable .
The statement node can then have the other information associated with it, as in _:rt :characterRole :RhettButler .
It is important to note that the reification of a triple in RDF does not entail the triple, i.e., it is possible to have the reification of a triple in RDF without implying that the triple itself is true so there is no implication from the above that Gone with the Wind has Clark Gable as a cast member, nor that he plays Rhett Butler in the movie.
Wikidata [
Gone with the Wind instance of: film cast member: Clark Gable
character role: Rhett Butler The intended meaning of this is that Gable is a cast member of Gone with the Wind and that the character role of this cast membership is Rhett Butler. Here the underlying “triple” is indeed a fact, as opposed to the situation with RDF reification, but, as Wikidata does not have a formal semantics, there is no formal statement of this intended meaning. As qualifiers are placed directly on regular statements any tools that use Wikidata information don’t need any special code to handle the statements that have attached qualifiers.
Singleton properties [
:castMember_1 :characterRole :RhettButler This looks like a syntactic abbreviation for RDF reification but it has one important difference—there is an entailment in singleton properties from the singleton property to its regular property. That is, in the singleton properties semantics the above triples entail
Because singleton properties uses a special data structure (the singleton property) a tool for singleton properties needs to have special code that makes the inference from triples with singleton properties to the triple with the correct regular property.
schema.org contains a special class Role (https://schema.org/Role) that is used for additional information about a relationship.6 Instead of a property directly linking from a subject to an object, the property links from the subject to a role item and then again from the role item to the object, as in (again recast slightly) :GWTW :castMember _:r . _:r rdf:type schema:Role . _:r :castMember :ClarkGable .
_:r :characterName :RhettButler .
The intended meaning of this is again that Gone with the Wind has Clark Gable as a cast member, and that the character role of this cast membership is Rhett Butler. (Indeed schema.org has a specialization of Role, PerformanceRole, that is designed specifically for this particular kind of extra information.) So here again the underlying “triple” is indeed a fact, as opposed to the situation with RDF reification. As schema.org does not have a formal semantics, there is no formal statement of this intended meaning. Any tool for schema.org data needs to handle Role items specially, short-circuiting the double link to recover the underlying normal meaning
From this example it appears that each of Wikidata, singleton properties, and schema.org have qualifiers working very nicely. They all permit additional information to be associated with their analogue of triples while retaining the underlying truth of the triple. The special processing required to recover this underlying truth is not onerous. RDF reification appears the loser here as it does not directly retain the underlying truth while associating extra information with a triple. If the triple itself is supposed to be true then it has to be asserted independently, and connecting the additional information to this triple also has to be done. 6 http://blog.schema.org/2014/06/introducing-role.html In the above example the qualifer is adding extra information about the fact, namely that the cast member is playing a particular role. This extra information does not interfere in any way with the cast membership, so tools (and queries) that are interested in cast membership can safely ignore the qualifier so long as they handle the modifications to the underlying data structure to handle qualifiers. This is true even in RDF, so long as the base triple is also present.
However, many, perhaps most, uses of qualifiers in Wikidata7 and schema.org contextualize the underlying fact, i.e., they limit the contexts in which the underlying fact is true. One of the most important contextualizations is temporal contextualization, which is generally handled in these schemes via a start and end qualifier.
So in singleton properties representing the temporal aspects of Bob Dylan’s marital
information would be (adapted from [
Bob Dylan spouse: Sara Dylan end time: 29 June 1977 start time: 22 November 1965 spouse: Carolyn Dennis end time: October 1992 start time: 4 June 1986 In singleton properties there is the entailment from the above triples to :BobDylan :isMarriedTo :SaraLownds .
:BobDylan :isMarriedTo :CarolDennis . i.e., that Bob Dylan is a bigamist. In Wikidata there is also the implied intent that there are two spouses of Bob Dylan, but although this is somewhat misleading there is perhaps not quite the same force to the conclusion of bigamy because of the name of the 7 Of the ten most-used qualifier properties in Wikidata as of July 2018. (See https:// tools.wmflabs.org/sqid/#/browse?type=properties.) four provide temporal context (point in time, start time, valid in period, and end time), one provides spatial context (chromosome), one can be considered to provide a certainty context (determination method), three are ”additive” (taxon author, points scored, and matches played), and one provides no information content (stated as). relationship. As there is no formal semantics for Wikidata there is no formal support for drawing both conclusions. The qualifier methodology of schema.org would act the same as Wikidata.
RDF reification, on the other hand, can represent this information as in :rt1 rdf:subject :BobDylan . :rt1 rdf:predicate :isMarriedTo . :rt1 rdf:object :SaraLownds . :rt1 :hasStart "1965-11-22"ˆˆxsd:date . :rt1 :hasEnd "1977-06-29"ˆˆxsd:date . :rt1 rdf:subject :BobDylan . :rt1 rdf:predicate :isMarriedTo . :rt1 rdf:object :CarolDennis . :rt1 :hasStart "1986-06-04"ˆˆxsd:date .
:rt1 :hasEnd "1992-10-01"ˆˆxsd:date . which does not imply that Bob Dylan is a bigamist. 3
The underlying problem is that contextual information does not add to the base information but instead modifies the it, stating in which context (temporal or otherwise) the information is true or providing information as to how likely the base information is to be true. Consumers of the information need to always be aware of whether the context that they are (perhaps implicitly) working in is not one to which the qualifiers attached to a particular piece of information apply.
This might not be so hard if the only contextual information is temporal and that information is carried only by start and end dates. Tools (and queries against the underlying data) working in a particular time point can explicitly exclude facts with temporal qualifiers that do not cover that time point. Tools (and queries) working in an implicit now can exclude any fact with an end time qualifier (assuming of course that no end date is in the future, which could be a requirement for temporal qualifiers).
However, each and every contextual qualifier has to be considered by every tool (and query against the underlying data). So there can be multiple exclusions, such as for contextual location, confidence, and certainty, and tools (and queries) will have to be updated whenever new such qualifiers are added. Wikidata has over one hundred and fifty qualifiers so to determine whether a fact is true in a context in Wikidata each and every one has to be examined to see if it is a contextual qualifier and for those which are their interaction with the current context or the implicit context has to be determined. This is a high bar indeed, made even worse as new qualifiers are added on an ongoing basis. 4
The right solution is to not show contextual qualifiers to consumers. Ideally, instead replace contextual qualifiers with a formal theory of the contexts so that basic tools (contextualizers, reasoners) can be written that correctly take context into account. Alternatively, create low-level tools that remove facts that are not valid in the contexts that a consumer wants to use.
The formal approach requires building logics and theories for the contextual constructs. Building and implementing these formal theories is certainly not a trivial task— handling temporal contexts, for example, requires a temporal logic8 —but the result will provide firm underpinnings of the constructs, avoiding problems with differing views of their meaning by different consumers of the information. Making the complexities inherent in contexts part of the base capabilities of knowledge graphs also means that more of the complexities are handled by the producers of contextualized knowledge graphs and fewer are inflicted on the consumers of information from knowledge graphs.
The non-formal approach requires less initial effort. For example, instead of being
given full temporal reasoning tools, consumers would be given an interface that
produces only the information valid in a particular context or a particular range of contexts.
This is somewhat similar to RDF extracts produced from Wikidata [
It is quite reasonable to combine the two approaches, providing extraction tools that are based on a formal theory of the contexts. Consumers that only need simple contextual processing can use the extraction tools; consumers that need more complex contextual processing can use the full reasoning tools. 8 If all time points are constants on a single time line, the required temporal reasoning might be simple, but still a temporal logic should be used to underpin the temporal reasoning algorithm.