Re: ISSUE: Classes as instances

[chair hat off]

Encl. is a detailed example of Raphael's scenario 1.

Thesaurus-ontology mapping

1. WordNet

Melnik and Decker have developed an (unofficial) RDF representation of
WordNet [1].  They provide an RDF Schema with (amongst others)
the following definitions:

  wn:LexicalConcept rdf:type rdfs:Class

  wn:wordForm rdf:type rdfs:Property
  wn:wordForm rdfs:domain wn:LexicalConcept
  wn:wordForm rdfs:range  rdf:literal

  wn:glossaryEntry rdf:type rdfs:Property
  wn:glossaryEntry rdfs:domain rdf:LexicalConcept
  wn:glossaryEntry rdfs:range  rdf:literal

  wn:hyponymOf rdf:type rdfs:Property
  wn:hyponymOf rdfs:domain rdf:LexicalConcept
  wn:hyponymOf rdfs:range  rdf:LexicalConcept

The class "LexicalConcept" denotes a WordNet synset.
The property "wordForm" links a synset to a term.
The property "glossaryEntry" links a synset to a textual description.
The property "hyponymOf" denotes a hyponym relation between two synsets.

[Actually, OWL would have been useful to make this schema more
precise. For example, the hyponymOf only holds between two subclasses
of LexicalConcept, namely noun and verb.]

The WordNet corpus itself is represented as a large (30+Mb) RDF database
with instances of these schema definitions. A few example instances:

  100001740 rdf:type wn:LexicalConcept
  100001740 wn:glossaryEntry "anything having existence (living or 
  100001740 wn:wordForm "entity"
  100001740 wn:wordForm "something"
  100002086 rdf:type wn:LexicalConcept
  100002086 wn:glossaryEntry "any living entity"
  100002086 wn:wordForm "being"
  100002086 wn:wordForm "life form"
  100002086 wn:wordForm "living thing"
  100002086 wn:wordForm "organism"
  100002086 wn:hyponymOf 100001740

2. WordNet interpretation

One disadvantage of the above representation is that the WordNet
hyponym hierarchy is now "hidden" in the hyponymOf triples. One could
therefore argue that this representation is wrong. However, if the
semantic web really becomes a reality, different representation
choices will be a fact of life. So, we looked at a way of
defining our interpretation of "WordNet as a class hierarchy" as an
add-on RDF Schema. It turned out that with two definitions we could
solve this problem:

  wn:LexicalConcept rdfs:subClassOf rdfs:Class
  wn:hyponymOf rdfs:subPropertyOf rdfs:subClassOf

The first definition makes it possible to treat a synset as a class,
which is a requirement to able to treat hyponymOf as a sort of
subclass relation (i.e. the second definition).

In addition, we included:

  wn:wordForm rdfs:subPropertyOf rdfs:label
  wn:glossaryEntry rdfs:subPropertyOf rdfs:comment

which also seems to make sense. See [2] for the RDF Schema file.

With these four definitions the RDF parser and browser used by our semantic
annotation & search tool is perfectly happy. The tool uses the
hierarchy in several ways. e.g.:
- for making it easy for users to understand the intended
  meaning of a term (e.g. "Venus" as a subclass of "Roman deity" or as a
  subclass of "planet"). This is used in annotation/search term
- for query generalization/specialization.
[See [3] for more info]

Although it first seemed like a hack, on second thought this might
actually be a decent way to do this kind of ontology/representation
mapping. Again, in the semantic web we will have to live with
representation choices made by others.

3. Implications for OWL

OWL should not disallow this type of mapping.

Comments/suggestions are very welcome,



Received on Friday, 12 July 2002 09:16:46 UTC