- From: Christopher Welty <welty@us.ibm.com>
- Date: Thu, 25 Mar 2004 08:09:36 -0500
- To: Ian Horrocks <horrocks@cs.man.ac.uk>
- Cc: Bernard Vatant <bernard.vatant@mondeca.com>, Jeremy Carroll <jjc@hplb.hpl.hp.com>, SWBPD <public-swbp-wg@w3.org>, public-swbp-wg-request@w3.org
- Message-ID: <OFF99F8BE5.478AA692-ON85256E62.0046C8AA-85256E62.00484A32@us.ibm.com>
Ian Horrocks <horrocks@cs.man.ac.uk> wrote on 03/24/2004 04:05:53 PM: > > Well, "mismodelling their world" is not limited to classes as instances. I > > find it rather dangerous to make such statements. People use subclass > > incorrectly, too, but that wasn't a reason to remove that axiom from OWL > > DL. > > I would say that there is a big difference. Like any part of the > language, subClass may occasionally be abused, but it is used very > widely and most people seem able to use it more or less > correctly. It is very widely used, yes, but people get it wrong about as much as they get it right. Since it's transitive, get one wrong near the "top" and your whole ontology will be off. > Moreover, subClass fits into a family of logics which are > theoretically well understood and for which there is considerable > implementation experience. Yes! THAT is the reason the choice was made for OWL DL, not: > In contrast, classes as instances are relatively rarely used (most > forms of conceptual modelling, databases etc., seem to have managed > perfectly well without them), and hardly ever used > "correctly". Well, I'm not surprised that a DL person would have this view. However the statement is simply false. Databases have had this since at least the mid-80s (see "Two Stage ER Diagrams" e.g.), and possibly earlier. Of course they don't do reasoning so they don't have the kinds of problems with their implementations that we do. THe issue was also recognized quite early on in object-oriented modeling, as well, and both Smalltalk and CLOS had "meta object" facilities for handling it. Outside the DL world, people building Knowledge-based systems use it all the time, and I would say that the per-capita misuse is about the same as it is for any other part of the language. Clearly if you go counting relations, subclass wins in sheer volume, but I can't say I have ever done a serious ontology without encountering a need to model this correctly, and in fact my experience is that systems like DLs actually encourage people to get it wrong by forcing them to ignore the distinction. > Moreover, the resulting logics are much less well > understood That's not true at all, unless you are conflating "understood" with: > and there is little implementation experience. There is less, for sure, I'm not sure I would say "litte". There has also been far less formal study of the complexity of reasoning algorithms for languages that have this expressiveness but not others. Peter and I looked at this in the early 90s for neo-classic, but we never followed through. The question was, like any language feature: what would you have to give up to have complete and decidable reasoning if you allowed instances to have instances. This is an open problem. Many people just throw up their hands prima facie because "2cnd order logic is incomplete and undecidable". However, it's not that simple - you can subset it just as we have been doing for FOL to find decidable fragments. > > Ian > > > > > > People just mismodel their worlds, I hope we can offer some advice on > > both how to do some of these things and how NOT to do it. > > > > Anyway, your analysis exposed some important misconceptions, espcially > > regarding so-called "subject hierarchies" and class hierarchies. I've > > written a paper or two about the problem, in this one: > > [http://dx.doi.org/10.1016/S0169-023X(99)90021-6] I basically show that > > subject taxonomies are actually "part"onomies, or more precisely spatial > > containment, not subclass (in fact, etymologically, "subject" means to > > throw under, "topic" is a region, and "about" means near). Some of the > > initial problems of representing subject taxonomies in DLs are discussed > > in a paper in the first FOIS conference, which may be hard to find. I > > can't seem to find a softcopy myself. > > > > > The class hierarchy in RDFS/OWL is there to describe hierarchies of > > classes > > > of resources. Just because you have a hierarchy of subject descriptors > > > doesn't make it a class hierarchy. > > > > > > It seems to be confusing the human way of thinking of analogy and > > metaphor > > > (any hierarchy can act as a metaphor for any other hierarchy) with what > > is > > > a logical and implementation issue about how to say what we want to say > > > about our knowledge of our world in a way that machines can process it. > > > > > > Thus if PhDThesis is an owl:Class what are the resources that we intend > > to > > > belong to it? Probably my PhD Thesis with title "Graph Grammars: an > > > approach to transfer based MT; exemplified by a Turkish-English system" > > is > > > one such resource, but the copy sitting on my bookshelf is probably not. > > > > > > Then if that is the case what would we mean by dc:subject linking the > > > resource of my thesis with this class .... hmmm ... we mean my thisis > > > belongs to that class, i.e. rdf:type. > > > So if we want to treat this subject hierachy as classes we really also > > want > > > > > > dc:creator rdf:subPropertyOf rdf:type . > > > > > > or perhaps > > > > > > eg:creator rdf:subPropertyOf rdf:type . > > > eg:creator rdf:subPropertyOf dc:creator . > > > > > > But if we click on dc:creator we get to: > > > http://purl.org/dc/elements/1.1/subject > > > > > > <rdf:Property rdf:about="http://purl.org/dc/elements/1.1/subject"> > > > <rdfs:label xml:lang="en-US">Subject and Keywords</rdfs:label> > > > <rdfs:comment xml:lang="en-US">The topic of the content of the > > > resource.</rdfs:comment> > > > <dc:description xml:lang="en-US"> > > > Typically, a Subject will be expressed as keywords, > > > key phrases or classification codes that describe a topic > > > of the resource. Recommended best practice is to select > > > a value from a controlled vocabulary or formal > > > classification scheme.</dc:description> > > > <rdfs:isDefinedBy rdf:resource="http://purl.org/dc/elements/1.1/"/> > > > <dcterms:issued>1999-07-02</dcterms:issued> > > > <dcterms:modified>2002-10-04</dcterms:modified> > > > <dc:type > > > > > rdf:resource="http://dublincore.org/usage/documents/principles/#element"/> > > > <dcterms:hasVersion > > > rdf:resource="http://dublincore.org/usage/terms/history/#subject-004"/> > > > </rdf:Property> > > > > > > and we see that dc:subject should typically be a string from a > > controlled > > > vocabulary. Thus it seems particularly poor practice to deviate from the > > > > > preferred usage of dc:subject in order to (over-)simplify our model. > > > > > > This points to the solution I was earlier advocating of using such > > strings, > > > using hasValue restrictions to map the strings into classes and then > > using > > > the class hierachy on those restrictions to show the hierarchical > > > relationships between the subject vocab terms. To do this well, we > > probably > > > want to specialise the dc:subject property with a subproperty > > eg:subject, > > > specify its range with an owl:Datarange explicitly enumerating the > > > controlled vocabulary, and for each term create a class using a hasValue > > > > > restriction. > > > For further clarity and usablility we might want to create two related > > > properties, one indicating the (single) intended subject code, and the > > > other indicating all implicit subject codes formed from the class > > hierachy. > > > The former would be a subproperty of both the latter and dc:subject; the > > > > > latter would be used to create the hasValue restrictions. > > > > > > Hmmm ... quite a lot of work initially, but the end result is that the > > > subject indicators are marked up using text strings from an explicit > > > controlled vocab; we conform with the defn of dc:subject, even with the > > > advertised best practice; we fall within OWL DL with the expectation > > that > > > this will give us better reasoning performance, and we have been clearer > > > > > about we are trying to say. I think the complexity can be hidden from > > the > > > end users. > > > > > > Jeremy > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Bernard Vatant wrote: > > > > > > > > *BV > > > > > > > >>>- Is it worth the trade-off to switch one's ontology (otherwise DL) > > > >>>to OWL-Full, just to > > > >>>allow its classes to be used as objects in 'dc:subject' predicates? > > > > > > > > > > > > *Jim > > > > > > > >>That's a weird way to ask the question. You mean, is it worth doing > > > >>the extra work to break your naturally occuring model just so that > > > >>you can be in DL? > > > > > > > > > > > > The way I put it might seem weird indeed, but it's the way it was set > > in the real project > > > > context (real world is weird). We had an OWL-DL ontology, and wanted > > to keep it so, and > > > > suddenly after six months or so some user wants to be able to use a > > class as a subject of > > > > a document ... which is one case out of one thousand, the 999 others > > using 'regular' > > > > subjects. So using a class as subject of a document is not exactly > > 'naturally occuring'. > > > > It's a borderline case - not to say a weird one :)) > > > > > > > > *Jim > > > > > > > >>I would argue this is indeed a BP issue, but probably for WORLD not > > > >>for OPEN... we need to explain why and when you would do the extra > > > >>work (and in every case we have explored it is extra work) to make > > > >>sure your ontology is in the DL profile of OWL. > > > > > > > > > > > > I suggested it might be in PORT scope, because it deals with the > > terminology vs ontology > > > > general issue. For me the heart of the question is to know what it > > means to 'use a > > > > concept' defined in a terminology (glossary, thesaurus, subject > > headings, index...) as a > > > > class (or a property) in an ontology. > > > > > > > > Is 'PhD Thesis' class the same 'subject' (using TM language here, > > sorry) or 'resource' > > > > than the original concept? The more I think about it, the more I have > > to deal with it, and > > > > the more I tend to say that they are distinct animals. Jim's PhD > > Thesis is an instance of > > > > the class, but not of the concept. One subject of 'Social Functions of > > PhD Thesis in > > > > Occidental University during 20th century', is the concept of PhD > > Thesis, not the class. > > > > > > > > So it's not just an issue of OWL-DL vs OWL-Full, it's also an issue of > > making distinct or > > > > not those two 'things'. This is a core issue in porting thesaurus to > > the SW, related to > > > > others of the same kind, like if concepts A and B are interpreted as > > classes, and there is > > > > a Broader-Narrower relationship between A and B in the Thesaurus, has > > it to be interpreted > > > > as a class-subclass relationship in the ontology etc. > > > > > > > > So I think in that case a BP definition would be two-fold > > > > > > > > 1. Is it generally a BP to make terminology concepts distinct from > > ontology classes (and > > > > properties)? > > > > 2. If agnostic about 1, what is the trade-off when choosing to make > > them distinct or to > > > > merge them ? > > > > > > > > FWIW, having tried both terms of the alternative in the course of > > time, my personal view, > > > > for above quoted reasons, is that they shoud be kept separate, and > > it's worth the extra > > > > work (even before being aware of the DL vs Full issue) > > > > > > > > Are there other concrete experience on that, not only theoretical > > considerations? Seems > > > > like there are not so many people exploring the terminology-ontology > > interoperability. Or > > > > are they? > > > > > > > > Bernard Vatant > > > > Senior Consultant > > > > Knowledge Engineering > > > > Mondeca - www.mondeca.com > > > > bernard.vatant@mondeca.com > > > > > > > > > > > > > > > <br><font size=2><tt>Jeremy wrote on 03/24/2004 04:24:16 AM:<br> > > <br> > > > <br> > > > Yes, like Bernard, I have been thinking more about this, and about > > Ian's <br> > > > insistence in WebOnt that classes-and-instances was almost always > > raised by <br> > > > people wanting to mismodel their world. (cc Ian, wondering if I have > > learnt <br> > > > my lessons well!, or misrepresented him)<br> > > </tt></font> > > <br><font size=2><tt>Well, "mismodelling their world" is not > > limited to classes as instances. I find it rather dangerous to make such > > statements. People use subclass incorrectly, too, but that wasn't > > a reason to remove that axiom from OWL DL. People just mismodel their > > worlds, I hope we can offer some advice on both how to do some of these > > things and how NOT to do it.</tt></font> > > <br> > > <br><font size=2><tt>Anyway, your analysis exposed some important misconceptions, > > espcially regarding so-called "subject hierarchies" and class > > hierarchies. I've written a paper or two about the problem, in this > > one: [http://dx.doi.org/10.1016/S0169-023X(99)90021-6] I basically show > > that subject taxonomies are actually "part"onomies, or more precisely > > spatial containment, not subclass (in fact, etymologically, "subject" > > means to throw under, "topic" is a region, and "about" > > means near). Some of the initial problems of representing subject > > taxonomies in DLs are discussed in a paper in the first FOIS conference, > > which may be hard to find. I can't seem to find a softcopy myself.</tt></font> > > <br> > > <br><font size=2><tt>> The class hierarchy in RDFS/OWL is there to describe > > hierarchies of classes <br> > > > of resources. Just because you have a hierarchy of subject descriptors > > <br> > > > doesn't make it a class hierarchy.<br> > > > <br> > > > It seems to be confusing the human way of thinking of analogy and > > metaphor <br> > > > (any hierarchy can act as a metaphor for any other hierarchy) with > > what is <br> > > > a logical and implementation issue about how to say what we want to > > say <br> > > > about our knowledge of our world in a way that machines can process > > it.<br> > > > <br> > > > Thus if PhDThesis is an owl:Class what are the resources that we intend > > to <br> > > > belong to it? Probably my PhD Thesis with title "Graph Grammars: > > an <br> > > > approach to transfer based MT; exemplified by a Turkish-English system" > > is <br> > > > one such resource, but the copy sitting on my bookshelf is probably > > not.<br> > > > <br> > > > Then if that is the case what would we mean by dc:subject linking > > the <br> > > > resource of my thesis with this class .... hmmm ... we mean my thisis > > <br> > > > belongs to that class, i.e. rdf:type.<br> > > > So if we want to treat this subject hierachy as classes we really > > also want<br> > > > <br> > > > dc:creator rdf:subPropertyOf rdf:type .<br> > > > <br> > > > or perhaps<br> > > > <br> > > > eg:creator rdf:subPropertyOf rdf:type .<br> > > > eg:creator rdf:subPropertyOf dc:creator .<br> > > > <br> > > > But if we click on dc:creator we get to:<br> > > > http://purl.org/dc/elements/1.1/subject<br> > > > <br> > > > <rdf:Property rdf:about="http://purl.org/dc/elements/1.1/subject"><br> > > > <rdfs:label xml:lang="en-US">Subject and Keywords</rdfs: > label><br> > > > <rdfs:comment xml:lang="en-US">The topic of the content > > of the <br> > > > resource.</rdfs:comment><br> > > > <dc:description xml:lang="en-US"><br> > > > Typically, a Subject will be expressed as keywords,<br> > > > key phrases or classification codes that describe a topic<br> > > > of the resource. Recommended best practice is to select<br> > > > a value from a controlled vocabulary or formal<br> > > > classification scheme.</dc:description><br> > > > <rdfs:isDefinedBy rdf:resource="http://purl.org/dc/elements/1. > 1/"/><br> > > > <dcterms:issued>1999-07-02</dcterms:issued><br> > > > <dcterms:modified>2002-10-04</dcterms:modified><br> > > > <dc:type <br> > > > rdf:resource="http://dublincore. > org/usage/documents/principles/#element"/><br> > > > <dcterms:hasVersion <br> > > > rdf:resource="http://dublincore. > org/usage/terms/history/#subject-004"/><br> > > > </rdf:Property><br> > > > <br> > > > and we see that dc:subject should typically be a string from a controlled > > <br> > > > vocabulary. Thus it seems particularly poor practice to deviate from > > the <br> > > > preferred usage of dc:subject in order to (over-)simplify our model.<br> > > > <br> > > > This points to the solution I was earlier advocating of using such > > strings, <br> > > > using hasValue restrictions to map the strings into classes and then > > using <br> > > > the class hierachy on those restrictions to show the hierarchical > > <br> > > > relationships between the subject vocab terms. To do this well, we > > probably <br> > > > want to specialise the dc:subject property with a subproperty eg:subject, > > <br> > > > specify its range with an owl:Datarange explicitly enumerating the > > <br> > > > controlled vocabulary, and for each term create a class using a hasValue > > <br> > > > restriction.<br> > > > For further clarity and usablility we might want to create two related > > <br> > > > properties, one indicating the (single) intended subject code, and > > the <br> > > > other indicating all implicit subject codes formed from the class > > hierachy.<br> > > > The former would be a subproperty of both the latter and dc:subject; > > the <br> > > > latter would be used to create the hasValue restrictions.<br> > > > <br> > > > Hmmm ... quite a lot of work initially, but the end result is that > > the <br> > > > subject indicators are marked up using text strings from an explicit > > <br> > > > controlled vocab; we conform with the defn of dc:subject, even with > > the <br> > > > advertised best practice; we fall within OWL DL with the expectation > > that <br> > > > this will give us better reasoning performance, and we have been clearer > > <br> > > > about we are trying to say. I think the complexity can be hidden from > > the <br> > > > end users.<br> > > > <br> > > > Jeremy<br> > > > <br> > > > <br> > > > <br> > > > <br> > > > <br> > > > <br> > > > <br> > > > <br> > > > <br> > > > Bernard Vatant wrote:<br> > > > > <br> > > > > *BV<br> > > > > <br> > > > >>>- Is it worth the trade-off to switch one's ontology (otherwise > > DL)<br> > > > >>>to OWL-Full, just to<br> > > > >>>allow its classes to be used as objects in 'dc:subject' > > predicates?<br> > > > > <br> > > > > <br> > > > > *Jim<br> > > > > <br> > > > >>That's a weird way to ask the question. You mean, is > > it worth doing<br> > > > >>the extra work to break your naturally occuring model just > > so that<br> > > > >>you can be in DL?<br> > > > > <br> > > > > <br> > > > > The way I put it might seem weird indeed, but it's the way it > > was set in the real project<br> > > > > context (real world is weird). We had an OWL-DL ontology, and > > wanted to keep it so, and<br> > > > > suddenly after six months or so some user wants to be able to > > use a class as a subject of<br> > > > > a document ... which is one case out of one thousand, the 999 > > others using 'regular'<br> > > > > subjects. So using a class as subject of a document is not exactly > > 'naturally occuring'.<br> > > > > It's a borderline case - not to say a weird one :))<br> > > > > <br> > > > > *Jim<br> > > > > <br> > > > >>I would argue this is indeed a BP issue, but probably for > > WORLD not<br> > > > >>for OPEN... we need to explain why and when you would do the > > extra<br> > > > >>work (and in every case we have explored it is extra work) > > to make<br> > > > >>sure your ontology is in the DL profile of OWL.<br> > > > > <br> > > > > <br> > > > > I suggested it might be in PORT scope, because it deals with > > the terminology vs ontology<br> > > > > general issue. For me the heart of the question is to know what > > it means to 'use a<br> > > > > concept' defined in a terminology (glossary, thesaurus, subject > > headings, index...) as a<br> > > > > class (or a property) in an ontology.<br> > > > > <br> > > > > Is 'PhD Thesis' class the same 'subject' (using TM language here, > > sorry) or 'resource'<br> > > > > than the original concept? The more I think about it, the more > > I have to deal with it, and<br> > > > > the more I tend to say that they are distinct animals. Jim's > > PhD Thesis is an instance of<br> > > > > the class, but not of the concept. One subject of 'Social Functions > > of PhD Thesis in<br> > > > > Occidental University during 20th century', is the concept of > > PhD Thesis, not the class.<br> > > > > <br> > > > > So it's not just an issue of OWL-DL vs OWL-Full, it's also an > > issue of making distinct or<br> > > > > not those two 'things'. This is a core issue in porting thesaurus > > to the SW, related to<br> > > > > others of the same kind, like if concepts A and B are interpreted > > as classes, and there is<br> > > > > a Broader-Narrower relationship between A and B in the Thesaurus, > > has it to be interpreted<br> > > > > as a class-subclass relationship in the ontology etc.<br> > > > > <br> > > > > So I think in that case a BP definition would be two-fold<br> > > > > <br> > > > > 1. Is it generally a BP to make terminology concepts distinct > > from ontology classes (and<br> > > > > properties)?<br> > > > > 2. If agnostic about 1, what is the trade-off when choosing to > > make them distinct or to<br> > > > > merge them ?<br> > > > > <br> > > > > FWIW, having tried both terms of the alternative in the course > > of time, my personal view,<br> > > > > for above quoted reasons, is that they shoud be kept separate, > > and it's worth the extra<br> > > > > work (even before being aware of the DL vs Full issue)<br> > > > > <br> > > > > Are there other concrete experience on that, not only theoretical > > considerations? Seems<br> > > > > like there are not so many people exploring the terminology-ontology > > interoperability. Or<br> > > > > are they?<br> > > > > <br> > > > > Bernard Vatant<br> > > > > Senior Consultant<br> > > > > Knowledge Engineering<br> > > > > Mondeca - www.mondeca.com<br> > > > > bernard.vatant@mondeca.com<br> > > > > <br> > > > > <br> > > > <br> > > </tt></font>
Received on Thursday, 25 March 2004 08:13:06 UTC