Databases Reference
In-Depth Information
2 RL/RDF rules: the intersection of
AL
and
RL
. One of the most prominent fea-
tures we lose is the ability to reason over
owl:sameAs
relations; both to infer
such relations through, e.g., functional properties and inverse-functional proper-
ties, and to support the semantics of equality as per the rules in Table 3 (only
eq-sym
is A-Linear).
In terms of completeness with respect to standard bottom-up rule-evaluation
(i.e., without any distinction between T-Box or A-Box), the main limitation of
considering a separate static T-Box while reasoning over the A-Box is that it
can lead to incompleteness if new T-Box triples are found while reasoning over
the A-Box [41] (these triples will not be reflected in the T-Box). Inference of
T-Box triples during A-Box reasoning can occur due to non-standard use of the
core RDFS or OWL vocabulary (see Section 3). Workarounds for this problem
are possible: for example to recursively identify and reason over non-standard
triples in a pre-processing step, etc. However, non-standard use of the RDF(S)
and OWL vocabulary is not found often in Linked Data, with notable exceptions
being, e.g., the RDFS axiomatic triples and the documents dereferenced by the
RDF, RDFS and OWL terms themselves.
In the SAOR system, following previous papers [72,69], we also perform A-Box
reasoning in a distributed setting. We have evaluated the applicability of SAOR
over 1 billion Linked Data triples taken from 4 million Linked Data documents.
Using a variety of optimisations for our A-Linear profile of OWL 2 RL/RDF, on
a cluster of nine machines with 4GB of RAM and 2.2 GHz single-core processors,
we computed 1 billion unique and authoritative inferences in about 3.5 hours [41],
roughly doubling the input size. Without considering the authority of inferences,
we estimated that the volume of materialisation would increase by 55
×
,evenfor
the lightweight reasoning profile being considered [12].
5.3 Comparison and Open Issues
Meeting the Challenges
Tackling C1 (scalability) in the list of challenges enumer-
ated in Section 3, both Context-Dependent reasoning and Authoritative reason-
ing use distributed computing and partitioning techniques and various rule-based
optimisations to enable high levels of scale.
Tackling C2 (impure and fallible OWL), both approaches analyse the source
of input axioms and apply cautious materialisation, where incompleteness with
respect to standard OWL profiles is thus a feature, not a “bug”.
28
Both ap-
proaches can use rule-based inferencing to support an incomplete RDF-Based
semantics, which does not require input graphs to conform to OWL 2 DL re-
strictions enforced by OWL's Direct Semantics.
Regarding C3 (inconsistencies), both approaches use monotonic rule-based
reasoning techniques that do not reduce the deductive reasoning process to un-
satisfiability checking, and thus do not fall into “ex falso quod libet”. Inconsis-
tencies can be ignored. However, in the case of SAOR, we have also looked at
resolving the contradictions presented by inconsistencies: we investigated using
28
Importantly, a
non-standard
version of completeness can be rigorously defined in
both cases. See, e.g., [41] for details in the SAOR case.
