[Fwd: [seek-kr-sms] Modeling large scale ontologies in OWL: Unmet needs (fwd)]
Several of you have seen this interesting piece at its point of origin
-------- Original Message -------- Subject: [seek-kr-sms] Modeling large scale ontologies in OWL: Unmet needs (fwd) Date: Thu, 21 Sep 2006 09:23:43 -0700 (PDT) From: Shawn Bowers sbowers@ucdavis.edu To: seek-kr-sms@ecoinformatics.org
This is a good explanation of a current problem in OWL, and relevant to SEEK KR/SMS. This problem haso been addressed in OWL 1.1 via qualified cardinality restrictions. Interestingly, I believe OBOE was recently criticized for using the same property name with two distinct domains (the "hasSubject" property), however, this email argues that this is often a useful approach.
-shawn
---------- Forwarded message ---------- Date: Wed, 20 Sep 2006 16:25:33 +0200 From: David Decraene David@landcglobal.com To: public-semweb-lifesci@w3.org Subject: Modeling large scale ontologies in OWL: Unmet needs Resent-Date: Thu, 21 Sep 2006 04:57:52 +0000 Resent-From: public-semweb-lifesci@w3.org
I would like to bring to discussion/debate what I believe is a flaw in the design of the OWL web ontology language, at least if OWL will also serve a purpose in representing large scale ontologies. I'll start the discussion by listing some general examples of OWL 'properties' from the OWL tutorials: HasFinger, HasSpouse, HasTopping, HasVintageYear, HasBrother, HasParent, HasUncle, HasMaker, HasSugar, etc ... We have a problem with this, seeing that these are not properties, but concatenated Property - Target combinations. For a large scale (formal) ontology, imagine the amount of duplication would occur if we would treat every parthood relation as such; HasFinger, HasToe, HasNail, HasHair, HasSkin, HasVein, HasCell, HasBone, HasMuscle, etc... In large scale ontologies, one link should suffice, HasPart, and whether the part is a finger, toe, nail, muscle or anything else is not a task for the property to describe, but for the target, e.g. the ISA hierarchy where the respective parts are situated in. Now this might seem trivial, and it might seem that OWL allows the second approach as well, but the fact is that OWL propagates this line of thought in how properties should be constructed and restricted: Some of the existing expressivity of OWL (mainly: assigning cardinalities) can not be applied if we decide to generalize property types (formal modeling) to represent the nature of the relation, e.g. without a description of what the target is (which is a task for the target hierarchy).
Lost expressivity: cardinality. In OWL optional properties can be used by setting mincardinality to 0: <Hand> Has6thFinger mincardinality 0 See how in the above example we do NOT state a target. As a matter of fact the above named property only has a meaning when the range is specified, in this case the range might be <6th finger>. So we can use optional properties, but only if we create a unique property for every property - target combination. In formal ontology you could express this relation on a general level of parthood: <Hand> HasPart <6thfinger>, cardinality 0. This is not possible in OWL.
In OWL absence of properties can be modeled by setting cardinality to 0: <Undiagnosed hyperbilirubinemia> IsDiagnosisParticipant cardinality 0 <Untreated hyperbilirubinemia> IsTreatmentParticipant cardinality 0 Again in these examples we do NOT state the target, and again these above stated examples might be referring to anything, the only way we (and a reasoner) are able to figure that the properties are about <Diagnosis> and <Treatment> is if we set the range of those properties as such. In formal ontology you would express this relation on a general level of participation: <Undiagnosed hyperbilirubinemia> IsParticipantOf <Diagnosis> Cardinality 0 <Untreated hyperbilirubinemia> IsParticipantOf <Treatment> Cardinality 0
Conclusion: OWL offers great expressivity, but only IF we use the above mentioned approach of proliferating redundant properties at a large scale. This is not a luxury we have in (formal) modeling of large scale ontologies. As an example, the core domain of LinkBase (our inhouse formal, medicine-oriented ontology) exists of about 700.000 concepts (classes) which are related to each other with more than 700.000 relation instantiations. The OWL approach would lead to hundred thousands of properties, which in many ways would duplicate the classhierarchy.
It would be great to receive some feedback on this matter.
David Decraene PH.D. BioMedical sciences/Ontology modeller. Language and Computing
_______________________________________________ Seek-kr-sms mailing list Seek-kr-sms@ecoinformatics.org http://mercury.nceas.ucsb.edu/ecoinformatics/mailman/listinfo/seek-kr-sms
Thanks for pointing this out Bob,
I find it hard to follow without knowing what LinkBase is for. Are they trying to create an ontology that 'does anything'. The definition on the Language and Communication website says:
/ LinKBase® is an expansive medical knowledge base that provides the "intelligence" used by all L&C applications./
I read this as "an ontology that will do anything our customers need".
He is arguing between two approaches. One may be suitable for one purpose and another suitable for another purpose - even though it is less elegant.
There is no 'correct' model to be discovered here only a model that is useful for doing X.
Without a constrained definition of what X is there are no criteria to say whether what is proposed is good. "Anything" is not a good definition of X. Once you have an ontology for doing X you may discover it is pretty good at doing Y but it will never be able to do everything. Nothing can do everything - which is a pretty good argument for very small ontologies :)
A model is merely a view onto reality from a particular perspective - the perspective that seems important right now.
Large ontologies are things that people have spent a great deal of time and money on and are therefore expected to be applicably in many ways. This doesn't mean that they actually are applicable in many ways it just means it would be nice if they were.
Just my thoughts and probably off the topic of OWL usage. I am sure, as OWL goes, what is discussed is all technically true.
All the best,
Roger
Bob Morris wrote:
Several of you have seen this interesting piece at its point of origin
-------- Original Message -------- Subject: [seek-kr-sms] Modeling large scale ontologies in OWL: Unmet needs (fwd) Date: Thu, 21 Sep 2006 09:23:43 -0700 (PDT) From: Shawn Bowers sbowers@ucdavis.edu To: seek-kr-sms@ecoinformatics.org
This is a good explanation of a current problem in OWL, and relevant to SEEK KR/SMS. This problem haso been addressed in OWL 1.1 via qualified cardinality restrictions. Interestingly, I believe OBOE was recently criticized for using the same property name with two distinct domains (the "hasSubject" property), however, this email argues that this is often a useful approach.
-shawn
---------- Forwarded message ---------- Date: Wed, 20 Sep 2006 16:25:33 +0200 From: David Decraene David@landcglobal.com To: public-semweb-lifesci@w3.org Subject: Modeling large scale ontologies in OWL: Unmet needs Resent-Date: Thu, 21 Sep 2006 04:57:52 +0000 Resent-From: public-semweb-lifesci@w3.org
I would like to bring to discussion/debate what I believe is a flaw in the design of the OWL web ontology language, at least if OWL will also serve a purpose in representing large scale ontologies. I'll start the discussion by listing some general examples of OWL 'properties' from the OWL tutorials: HasFinger, HasSpouse, HasTopping, HasVintageYear, HasBrother, HasParent, HasUncle, HasMaker, HasSugar, etc ... We have a problem with this, seeing that these are not properties, but concatenated Property - Target combinations. For a large scale (formal) ontology, imagine the amount of duplication would occur if we would treat every parthood relation as such; HasFinger, HasToe, HasNail, HasHair, HasSkin, HasVein, HasCell, HasBone, HasMuscle, etc... In large scale ontologies, one link should suffice, HasPart, and whether the part is a finger, toe, nail, muscle or anything else is not a task for the property to describe, but for the target, e.g. the ISA hierarchy where the respective parts are situated in. Now this might seem trivial, and it might seem that OWL allows the second approach as well, but the fact is that OWL propagates this line of thought in how properties should be constructed and restricted: Some of the existing expressivity of OWL (mainly: assigning cardinalities) can not be applied if we decide to generalize property types (formal modeling) to represent the nature of the relation, e.g. without a description of what the target is (which is a task for the target hierarchy).
Lost expressivity: cardinality. In OWL optional properties can be used by setting mincardinality to 0: <Hand> Has6thFinger mincardinality 0 See how in the above example we do NOT state a target. As a matter of fact the above named property only has a meaning when the range is specified, in this case the range might be <6th finger>. So we can use optional properties, but only if we create a unique property for every property - target combination. In formal ontology you could express this relation on a general level of parthood: <Hand> HasPart <6thfinger>, cardinality 0. This is not possible in OWL.
In OWL absence of properties can be modeled by setting cardinality to 0: <Undiagnosed hyperbilirubinemia> IsDiagnosisParticipant cardinality 0 <Untreated hyperbilirubinemia> IsTreatmentParticipant cardinality 0 Again in these examples we do NOT state the target, and again these above stated examples might be referring to anything, the only way we (and a reasoner) are able to figure that the properties are about <Diagnosis> and <Treatment> is if we set the range of those properties as such. In formal ontology you would express this relation on a general level of participation: <Undiagnosed hyperbilirubinemia> IsParticipantOf <Diagnosis> Cardinality 0 <Untreated hyperbilirubinemia> IsParticipantOf <Treatment> Cardinality 0
Conclusion: OWL offers great expressivity, but only IF we use the above mentioned approach of proliferating redundant properties at a large scale. This is not a luxury we have in (formal) modeling of large scale ontologies. As an example, the core domain of LinkBase (our inhouse formal, medicine-oriented ontology) exists of about 700.000 concepts (classes) which are related to each other with more than 700.000 relation instantiations. The OWL approach would lead to hundred thousands of properties, which in many ways would duplicate the classhierarchy.
It would be great to receive some feedback on this matter.
David Decraene PH.D. BioMedical sciences/Ontology modeller. Language and Computing
Seek-kr-sms mailing list Seek-kr-sms@ecoinformatics.org http://mercury.nceas.ucsb.edu/ecoinformatics/mailman/listinfo/seek-kr-sms
participants (2)
-
Bob Morris -
Roger Hyam