- From: Antoine Zimmermann <antoine.zimmermann@emse.fr>
- Date: Tue, 23 Feb 2016 14:26:02 +0100
- To: ross.horne@gmail.com
- CC: Semantic Web <semantic-web@w3.org>, Pat Hayes <phayes@ihmc.us>
Ross, A few comments below. On 23/02/2016 10:24, Ross Horne wrote: > Hi AZ, > > I agree with you analysis of Bioportal. So would the official line in > this situation be to encourage the style in the provenance ontology, and > exert caution when performing RDFS inference in Bioportal? > > Let me follow up your nice example in the hypothetical "Bioportal" style > definition: >> "Where P has more than one rdfs:range property, then the resources >> denoted by the objects of triples with predicate P are instances of >> *some* class stated by the rdfs:range properties." > > You are exactly right with your subclass example. Under the hypothetical > *some* definition above, properties become more accommodating as more > schema information is discovered rather than more restrictive. > Discovering ex:myProperty rdfs:domain ex:Female would just confirm > something that is already known i.e. that ex:myProperty can be use with > ex:Person including females. Notice that this analysis assumes that > ex:Female is a sub class of ex:Person is part of the ontology. There is no assumption about whether ex:Person and ex:Female are relate via a subclass relation. In fact, In my example, I expected that ex:Female be interpreted as the class of all individuals that have the female gender, such as a female dog, a female lizard, or a female human. The imaginary ex:myProperty would only apply to the females of the human species. > However, if ex:myProperty rdfs:domain ex:Place was discovered, where > ex:Place and ex:Person are not related by the sub class relation, then > we discover something new. In particular, we discover that instances of > both ex:Person and ex:myProperty may appear in the subject position of a > triple with property ex:myProperty. If we find that ex:Place is also a domain of myProperty, then anything that has this property is a person, is female, and is a place. That's all and that's unrelated to what may or may not appear in triples. To my point clearer, consider rdfs:range, rather than rdfs:domain. Take for example the character string "AZ". RDF 1.1 Semantics tells us that this string has the types rdfs:Literal and xsd:string. I may also find somewhere such things as: :i ex:prop "AZ" . ex:prop rdfs:range ex:TwoLettersString . from which I can infer "AZ" is a ex:TwoLetterString. However, this does not say that "AZ" may appear in the subject position of a triple because as a matter of fact, it cannot. --AZ > > My follow up question is: whether anyone knows whether the more > accommodating inference, as implied by Bioportal, was ever discussed > during the RDFS standardisation process; and if so, why the more > restrictive definition for multiple domains and ranges was chosen. > > I suspect this question has a simple explanation in model theory, which > is why I also copy Pat. > > Best regards, > > Ross > > > > On 23 February 2016 at 16:37, Antoine Zimmermann > <antoine.zimmermann@emse.fr <mailto:antoine.zimmermann@emse.fr>> wrote: > > Ross, > > > The conclusion here is that Bioportal wrongly uses rdfs:domain. The > provenance ontology uses it correctly, and if DBpedia does not have > multiple domains or ranges, then no problem. > > There are certainly many more mistaken datasets with this respect, > as there are many other kinds of errors in datasets. There are also > many misinterpretations of HTML markups, mistakes in CSS files, and > in fact, all Web standards are misused to some extent. If the wrong > use of multiple domains / ranges was largely predominant, it would > be a source of concern for the standardisation groups of future > versions of RDF. But your observations in your email are not > sufficient to indicate that. > > In any case, your suggestion: > > > "Where P has more than one rdfs:range property, then the resources > > denoted by the objects of triples with predicate P are instances of > > *some* class stated by the rdfs:range properties." > > would not work well with the inherent incompleteness of knowledge on > the Web and with the distributed nature of Web data. If I see: > > ex:myProperty rdfs:domain ex:Person . > > somewhere on the Web, I would like to conclude something about those > individuals who have the property ex:myProperty. Then I may find the > following: > > ex:myProperty rdfs:domain ex:Female . > > Now I know more than before, so I should infer more about those who > have the property. With your suggestion, every time I would know > more about the domain of a property, I would know less about those > who have the property. > > > Best, > AZ > > > On 23/02/2016 03:36, Ross Horne wrote: > > Hi All, > > I'm wondering if many people here use multiple > rdfs:domain/rdfs:range > properties in RDF Schema? > > The RDF Schema spec is clearly worded: "Where P has more than one > rdfs:range property, then the resources denoted by the objects of > triples with predicate P are instances of *all* the classes > stated by > the rdfs:range properties." [similarly for rdfs:domain] > > However, this doesn't quite match the usage of multiple > rdfs:domain/rdfs:range properties in several popular datasets. > > For example, in Bioportal, the property bpo:has_event has three > classes > indicated as its domain: bpo:person, bpo:event and > bpo:disease_or_disorder. Following the wording of the spec, it would > appear that any resource that appears in the subject position of a > triple with property bpo:has_event is an instance of all three types > bpo:person, bpo:event and bpo:disease_or_disorder. However, > common sense > says that the resource cannot simultaneously be a person, event > and disease. > > Elsewhere, the provenance ontology avoids the problem by explicitly > using owl:unionOf. For example, prov:wasInfluencedBy has > rdfs:range such > that it is the owl:unionOf the classes prov:Activity, prov:Agent and > prov:Entity. DBpedia avoids the problem entirely, since I cannot > find > any multiple rdfs:domain/rdfs:range properties in their ontologies. > > The interpretation of multiple rdfs:range properties in the above > datasets, either implicitly or explicitly imply an alternative > spec such as: > > "Where P has more than one rdfs:range property, then the > resources > denoted by the objects of triples with predicate P are instances of > *some* class stated by the rdfs:range properties." > > I'm wondering whether anyone else has observed this mismatch > between the > spec and real world datasets; and what the official line would be on > avoiding this conflict? > > Regards, > > Ross > > > Note I'm using the following prefixes in examples: > bpo: <http://www.semanticweb.org/ontologies/2010/10/BPO.owl#> > prov: <http://www.w3.org/ns/prov#> > rdfs: <http://www.w3.org/2000/01/rdf-schema#> > > >
Received on Tuesday, 23 February 2016 13:26:28 UTC