- From: Aldo Gangemi <a.gangemi@istc.cnr.it>
- Date: Sat, 27 Mar 2004 19:54:47 +0100
- To: Ian Horrocks <horrocks@cs.man.ac.uk>
- Cc: Guus Schreiber <schreiber@cs.vu.nl>, SWBPD list <public-swbp-wg@w3.org>, Nicola Guarino <guarino@loa-cnr.it>, claudio.masolo@ladseb.pd.cnr.it
My two cents ... At 12:57 +0000 27-03-2004, Ian Horrocks wrote: >On March 26, Guus Schreiber writes: >> Ian Horrocks wrote: > >[snip] > >> After reading this I was about to put my chai hat off and send a flame. >> But it generated such well-founded responses that I want to thank for >> starting this dialectic debate. > >Glad to be of service. > >Perhaps I could push things along a bit further by refining my initial >contention vis-a-vis meta classes. > >Regarding correctness of usage, I speak as I find. I have had the >discussion on numerous occasions, and have almost always found that >suggested requirement examples are "ill conceived". Examples include >four-engined-aeroplane as a meta-class, with Boeing-747 being an >instance (this came up at one of the WebOnt f2f meetings), an instance >of the patient class who is related via an instance of the "takesDrug" >property to the class "Aspirin" (I think that this may also have come >from a WebOnt f2f meeting) and "Boeing" as a subClass of aeroplane and >an instance of Company (this came up at a talk at WWW in >Budapest). Note that these examples were carefully crafted by KR >experts to illustrate the requirement for meta-classes. In all cases >it was subsequently agreed, even by the proposers, that these are >examples of poor modelling, i.e., they didn't capture the intended >meaning of the modeller (Boeing-747 is better modelled as a subClass >of four-engined-aeroplane, the patient should take an instance of >Aspirin, and the two Boeings are just different objects). I agree, from my experience most uses of metaclasses result from insufficient domain analysis. I think that insufficient analysis is responsible for many other simplifications, even in the use of the subClass relation or in the representation of qualities in flat value spaces (cf. Guus' message). Low-detailed domain analysis is sometimes due to a desire of economicity, for example if we want to talk about legal regulations in the domain of navigation, two or three colors might even happen to fall into the discussion, but a complete color hierarchy is useless. Similarly, if we talk of fishery commodities (only), having a complete taxonomy of aquatic organisms paired to a taxonomy of ways of post-harvesting those organism, and to a taxonomy of packaging them, can be overspecified, and a simple taxonomy of commodities fits perfectly the task. But often economicity leads to underspecification wrt to the task at hand, and in those cases metaclasses and multihierarchies are sometimes used to overcome the lack of primitives (Ian's examples suffice). >Having said all that, I am not asserting that no convincing use case >exists. The classic example from WebOnt was Guus's >Species-Elephant-Jumbo. In these cases, however, the Meta-Class >generally seems to be serving a mainly organisational or documentary >function (I think that someone already made a similar point in this >thread), and has very little impact on the semantics of the classes >and instances in the ontology. In fact I don't know of *any* example >where the Meta-Classes use constraints to specify logical >characteristics that are inherited by the classes that that >instantiate them. It therefore seems to make sense to use a *very* >much simpler language to describe meta-classes, one that would have >little if any impact on the problem of reasoning about classes and >individuals, or perhaps even to treat meta-classes as a form of >documentation. > >Ian I agree also on this. So-called "taxa" in Linnean taxonomies serve an organizational function, although they also have proper constraints (a species cannot be a species unless it has a superclass that is a genus, etc.). But once we try to analyze those constraints, they can be found to be approximations. In fact, modern classification of organisms has many classes that fall in between Linnean metaclasses, originating the need for e.g. subfamily, superfamily, suborder, etc. (for a sample resume of the criteria used in different classifications: http://www.science.siu.edu/plant-biology/PLB449/DLN.449lectures/Phylocode.htm). Another motivation for metaclasses has been advocated for properties that hold of a class in general, but not necessarily of any instance of that class. For example, "nursing newborns" is a property of mammals, but not of each instance of mammals (e.g. of males or young females). In my opinion, this is also a shortcut for something that in principle requires a more detailed analysis, leading in this case to a more complex restriction, stating that "for all newborns of mammals, there is a need to be nursed with a proper substance, and a subclass of this proper substance is secreted by certain glands found in grown females that have just delivered". Then, such organizational notions could well be implemented as annotations, like documentation, comments, encodings, etc., or may require a complex modelling such as the one I have mentioned (but staying at the object level). At this point one could wonder if there are real logical metaclasses, which need reasoning facilities. There are some cases that obviously require them, and they concern modelling primitives, for example the OWL namespace, the RDFS namespace, the built-in theories of most knowledge represenation systems, etc. In these cases, it's the structure of the language that is exploited by the reasoner to draw inferences, then one may be intersted in modifying that structure, at the cost of losing compatibility with other users, or of losing some functionalities tout court. But I think that these metaclasses have been discussed enough e.g. in the WebOnt WG, and are only indirectly of interest here. In some cases, we need to produce new schemas including metaclasses, e.g. in order to produce a schema for WordNet (see related thread in this WG), but the resulting ontology will not be used to reason in the same problem space as the ontologies depending on that schema. Another set of metaclasses that maybe requires reasoning in the same problem space as the classes that are instances of them are the metaclasses defined in e.g. the OntoClean papers by Nicola and Chris ("role", "type", "non-rigid", etc.). These have been implemented in various systems (Protégé, OntoEdit, WebOde, Web-KB) just as annotations, following the suggestion given above. On the other hand, those metaclasses have originally rich axiomatizations involving classes and relations at the object level. E.g., a "type" class implies that all its instances have that property for all their lifecycle, thus it could be interesting to check if there is any individual of a "type" class that is not preserving that property across time. But even in the last case, it is possible to reduce the problem to a matter of relations between individuals and roles that they play, putting both individuals and (reified) roles in the same domain. See for this solution our recent paper to appear in the next KR04. A report derived from it will be available at the LOA site (http://www.loa-cnr.it) in the next weeks. Bottom-line: let Natasha, Bernard, or anyone else produce cases of reasonable use of metaclasses, and let's try to find an alternative way to model them. If that way is unnecessarily complicated for the use case we are considering, then we can consider that use an example of best annotation practice. IMO, it remains to be demonstrated that there exists a case in which a reasoner must be able to reason on classes and metaclasses in the same problem space, without alternative solutions. Aldo PS oops, my two cents have become twenty cents ... -- *;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;*;* Aldo Gangemi Research Scientist Laboratory for Applied Ontology, ISTC-CNR Institute of Cognitive Sciences and Technologies (Laboratorio di Ontologia Applicata, Istituto di Scienze e Tecnologie della Cognizione, Consiglio Nazionale delle Ricerche) Viale Marx 15, 00137 Roma Italy +3906.86090249 +3906.824737 (fax) mailto://a.gangemi@istc.cnr.it mailto://gangemi@acm.org http://www.loa-cnr.it
Received on Saturday, 27 March 2004 14:02:29 UTC