SPM Categories or Subclassing - again.
Hi All, I'd like to re-ignite a debate we had over the summer in the light of what was discussed at Bratislava and things I am thinking about now. The original way the Species Profile Model was structure was to have something like the attached UML diagram tagging.png We didn't define a class "Category" but just left it as being any valid URI in the RDF style of doing things. Markus and several others pointed out that this would not work with TAPIR very well as all the InfoItems will look the same. Something like this: <SpeciesProfileModel> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/123" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/124" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/125" /> <....> </InfoItem> </hasInformation> </SpeciesProfileModel> The xpaths to the data represented by <...> would all be the same. They suggested something like the next diagram: Where info item is subclassed. This allows RDF to be serialized like this: <SpeciesProfileModel> <hasInformation> <Ecology> <....> </Ecology > </hasInformation> <hasInformation> <Behaviour> <....> </Behaviour> </hasInformation> <hasInformation> <BehaviouralEcology> <....> </BehaviouralEcology> </hasInformation> </SpeciesProfileModel> The xpaths to the <...> are all different and the model becomes TAPIR friendly. At the time I wasn't too bothered either way because I saw this as an artifact of serialization. The same thing could be written. <SpeciesProfileModel> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/Ecology" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/Behaviour" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/BehaviouralEcology" /> <....> </InfoItem> </hasInformation> </SpeciesProfileModel> With more or less that same meaning and looking just like a tagging example. This was a mistake on my part because of course it isn't the same meaning just a similar serialization - an RDF type really needs to point to a class and can't point to any old vocabulary that it could if it was treated like a tag. I have just added the category property back into InfoItem for discussion purposes. I am concerned about going down the subclassing route for a few reasons. 1) Building a class hierarchy is difficult. The discussions we had in Bratislava highlighted this. The examples put together in the current vocabulary sparked a debate as to how things should be organized. There are issues with the diamond problem in multiple inheritance - if contradictory semantics occur in multiple inheritance routes to a class which has priority? If we don't have multiple inheritance how do we have InfoItems that are about both Ecology and Behaviour for example or any other two subjects such as Dispersal and Asexual reproduction. 2) If a class hierarchy is required for analysis/inference it can be superimposed on a tag based transfer protocol using OWL necessary and sufficient properties. In fact this is arguably a better way of approaching the situation than building a class hierarchy a priori. 3) The motivation for going down the subclassing route is largely so it will work with the transport protocol - which sets alarm bells ringing for me. 4) It precludes us from using something like SKOS to build our categories. "SKOS or Simple Knowledge Organisation System is a family of formal languages designed for representation of thesauri, classification schemes, taxonomies, subject-heading systems, or any other type of structured controlled vocabulary. SKOS is built upon RDF and RDFS, and its main objective is to enable easy publication of controlled structured vocabularies for the Semantic Web. SKOS is currently developed within the W3C framework." (http://en.wikipedia.org/wiki/SKOS) I would really like to be able to say to "domain experts" that they "just" need to build a SKOS vocabulary for the terms used in their domain and then use the URIs of these terms for tagging information that is passed between providers than specify that they need to construct a more formal ontology of classes. I appreciate that one could argue that SKOS terms are like classes but they form part of a structure that is specifically designed for having the debates that TDWG needs to have around the vocabulary part of the standards (rather than the exchange parts). I would be grateful for peoples thoughts and criticisms. All the best, Roger
In SDD we used what is here called the "tagging approach", i.e. for descriptive concepts and characters we use <Text ref="some.resource.org/123"><Content>Free form text</Content></Text> In SDD, all element/attribute names refer to classes/attributes in UML or Java, whereas the xml values are corresponding values. Expressing concepts and characters as classes was considered impractical if the aim is generic software (classes are creatable at runtime, but at a price) because the number of characters is often very large. My experience shows that the number of character in identification sets varies between 50 and over 1000. Some studies in manual data integration (in LIAS) convinced me that the number of "integratable" characters is often less than imagined. A similar case study may also be the FRIDA data sets, which limit the number of common characters (common to all plant families) to 200, and from thereon use separate characters for each family. Clearly, when limiting the approach to just "very high level concepts", both class and value approaches are possible. The benefit of the SPM 0.2 approach is that it does enable reasoning (as well as may simplify tapir usage). However, as presented in Bratislava, I am doubtful that this limit holds. In my experience the limit between free form text and categorical data tends be diffuse rather than sharp and my intuition is that the SPM mechanism, being extensible, will be extended. I would like to note that the current TAG strategy is to provide both RDF and xml schema. The schema approach to SPM, however, seems to be more and more undesirable as the number of descriptive concepts grows. I have a question for TAPIR: Does TAPIR find it easier to search elements having specific attribute values than elements having child elements with specific value: <hasInformation> <InfoItem category="voc.x.org/DescriptiveConcepts/Size<http://some.resource.org/123>"> <....> </InfoItem> I might be useful to know. Although RDF allows this for literals, it seems to require the rdf:resource="" step for categorical data, so this may or may not help. ---------- PS: SDD calls the concepts "characters" based on the tradition and character is clearly inappropriate. However, Category to me seems to express nothing other than the type of the value. All contextValue and Value in SPM in version 0.2 refer to categories. Can anyone suggest a better term? What is the opposite of "value"? I can think of "Class" (value = instance), or - but perhaps too much limited to descriptions - "Feature" (also in GML). Anything else? Gregor ----------------------------------------------------- Gregor Hagedorn (G.M.Hagedorn@gmail.com) Institute for Plant Virology, Microbiology, and Biosafety Federal Research Center for Agriculture and Forestry (BBA) Königin-Luise-Str. 19 Tel: +49-30-8304-2220 14195 Berlin, Germany Fax: +49-30-8304-2203
Donald, Roger, The problem we saw with TAPIR is that there are only 2 new concepts defined by SPM - category and value. If you do not want to map anything more than this in TAPIR models, you probably can produce SPM RDF. But if you want to only map the descriptions or distributions you are in trouble. Imagine a simple database table with 3 columns: taxon, category and value. You could map them to 3 concepts namely SPM/taxon, SPM/category and SPM/value. You can then create an output model similar to this one here: <outputModel <structure> <schema location="http://rs.tdwg.org/tapir/rs/SPM.xsd"/> </structure> <indexingElement path="/SpeciesProfileModel/hasInformation/InfoItem"/> <mapping> <node path="/SpeciesProfileModel/aboutTaxon/@rdf:resource"> <concept id="SPM/taxon"/> </node> <node path="/SpeciesProfileModel/hasInformation/InfoItem/category/@rdf:resource"> <concept id="SPM/category"/> </node> <node path="/SpeciesProfileModel/hasInformation/InfoItem/value"> <concept id="SPM/value"/> </node> </mapping> </outputModel> With current provider implementations you will get a new SpeciesProfileModel element for every record in your database, i.e. repeating taxa for every InfoItem. This is valid SPM and you can actually search on it too with TAPIR like Donald wanted to. So here you go. You would just need to know (and understand) what categories this provider is using. You can get those through an inventory instead of the capabilities, not a real problem. If someone has a table with a taxon and a colum for description, one for economic use and one for something else they would have to transform it into a simple 3 column table. My understanding of SPM in the beginning was just that I thought an initial broad vocabulary should be part of SPM. I still think this is very much needed, otherwise you all "mashup" will be done by the human enduser who is presented with a nice long list of facts, sorted alphabetically. Personally I am reluctant which approach to take. The tagging seems more flexible, so yes, lets go for it! But I cant see why OWL classes are not suited for defining a reusable, external vocabulary. Isn't this what they are made for and what GO and all the other ontologies use them for rather than just defining data exchange formats? At least the tagging approach would free us from imposing a certain definition language. I guess some can use SKOS, some OWL and some others just RDFS. And of course there will be many different semantic definitions in all those languages. Markus "Donald Hobern" wrote on 08.10.2007 13:53 Uhr:
Roger,
Your analysis (and the four points) seem perfectly correct to me. I prefer a plain tagging approach and am particularly dubious about the long term benefits of subclasses which represent multiple inheritance because this would seem to place a greater burden on consumers.
However I am not so sure that the tagging approach does prevent the use of TAPIR. Surely the user is expected to be searching for InfoItems and each InfoItem will include a relative path "./category/@rdf:resource" which could be searchable. In other words you could search for InfoItems where the concept identified by "./category/@rdf:resource" has the value "some.resource.org/123". Won't that work? Won't it return any InfoItem which has any combination of category elements provided at least one is identified to the given resource?
Of course searching for InfoItems like this may prevent us from using search elements outside the scope of the InfoItem - would this stop us from using a ScientificName specified against the SpeciesProfileModel element? In other words, would it be better to move the aboutTaxon down to the InfoItem?
Donald
Roger Hyam wrote:
Hi All,
I'd like to re-ignite a debate we had over the summer in the light of what was discussed at Bratislava and things I am thinking about now.
The original way the Species Profile Model was structure was to have something like the attached UML diagram tagging.png
We didn't define a class "Category" but just left it as being any valid URI in the RDF style of doing things.
Markus and several others pointed out that this would not work with TAPIR very well as all the InfoItems will look the same. Something like this:
<SpeciesProfileModel> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/123" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/124" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <category rdf:resource="some.resource.org/125" /> <....> </InfoItem> </hasInformation> </SpeciesProfileModel>
The xpaths to the data represented by <...> would all be the same.
They suggested something like the next diagram:
Where info item is subclassed. This allows RDF to be serialized like this:
<SpeciesProfileModel> <hasInformation> <Ecology> <....> </Ecology > </hasInformation> <hasInformation> <Behaviour> <....> </Behaviour> </hasInformation> <hasInformation> <BehaviouralEcology> <....> </BehaviouralEcology> </hasInformation> </SpeciesProfileModel>
The xpaths to the <...> are all different and the model becomes TAPIR friendly.
At the time I wasn't too bothered either way because I saw this as an artifact of serialization. The same thing could be written.
<SpeciesProfileModel> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/Ecology" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/Behaviour" /> <....> </InfoItem> </hasInformation> <hasInformation> <InfoItem> <rdf:type rdf:resource="some.resource.org/BehaviouralEcology" /> <....> </InfoItem> </hasInformation> </SpeciesProfileModel>
With more or less that same meaning and looking just like a tagging example.
This was a mistake on my part because of course it isn't the same meaning just a similar serialization - an RDF type really needs to point to a class and can't point to any old vocabulary that it could if it was treated like a tag.
I have just added the category property back into InfoItem for discussion purposes.
I am concerned about going down the subclassing route for a few reasons.
1) Building a class hierarchy is difficult. The discussions we had in Bratislava highlighted this. The examples put together in the current vocabulary sparked a debate as to how things should be organized. There are issues with the diamond problem in multiple inheritance - if contradictory semantics occur in multiple inheritance routes to a class which has priority? If we don't have multiple inheritance how do we have InfoItems that are about both Ecology and Behaviour for example or any other two subjects such as Dispersal and Asexual reproduction.
2) If a class hierarchy is required for analysis/inference it can be superimposed on a tag based transfer protocol using OWL necessary and sufficient properties. In fact this is arguably a better way of approaching the situation than building a class hierarchy a priori.
3) The motivation for going down the subclassing route is largely so it will work with the transport protocol - which sets alarm bells ringing for me.
4) It precludes us from using something like SKOS to build our categories.
"SKOS or Simple Knowledge Organisation System is a family of formal languages designed for representation of thesauri, classification schemes, taxonomies, subject-heading systems, or any other type of structured controlled vocabulary. SKOS is built upon RDF and RDFS, and its main objective is to enable easy publication of controlled structured vocabularies for the Semantic Web. SKOS is currently developed within the W3C framework." (http://en.wikipedia.org/wiki/SKOS)
I would really like to be able to say to "domain experts" that they "just" need to build a SKOS vocabulary for the terms used in their domain and then use the URIs of these terms for tagging information that is passed between providers than specify that they need to construct a more formal ontology of classes. I appreciate that one could argue that SKOS terms are like classes but they form part of a structure that is specifically designed for having the debates that TDWG needs to have around the vocabulary part of the standards (rather than the exchange parts).
I would be grateful for peoples thoughts and criticisms.
All the best,
Roger
_______________________________________________ tdwg-tag mailing list tdwg-tag@lists.tdwg.org http://lists.tdwg.org/mailman/listinfo/tdwg-tag
participants (4)
-
Donald Hobern -
Döring, Markus -
Gregor Hagedorn -
Roger Hyam