Structuring relations for the R.O.

Fritz Lehmann (
Fri, 20 Sep 1996 14:48-0500

Date: Thu, 19 Sep 1996 12:56 CDT
From: Kavi Mahesh <>
Hi All,
A couple of questions on the structuring relations we talked
- are we going to use "subrelation" to taxonomize qualities (aka
attributes) or do we need a separate "subquality"? E.g., how do
we connect length and height to size?
- are taxonomic links going to be bidirectional? For example, in the
case of instance-of, are we going to list all the instances in
the parent? Or, is this something that is to be derived by the
API/other programs?
Kavi Mahesh
PS: I really enjoyed the meetings. They were more exciting and useful
than most of the ontology workshops I have been to.

Dear Kavi,

The structuring relations decided on at the Stanford standards meeting
were: instance, subclass, and subrelation (the latter for relations of
arity greater than or equal to two). Although we discussed the notion
of "qualities" (so-called attributes as opposed to classes), we didn't
discuss the subsumption relationship among these, or among their values.

Cyc uses the collection AttributeValue for the general "qualities" notion and
more specific attributes are specializations of AttributeValue -- their
instances are the actual values. Subsumption between the attributes is
stated with genls relations between specializations of the
collection AttributeValue (e.g. (genls Height Size)); subsumption between
the values themselves is stated with the genlAttributes relation (e.g.
(genlAttributes NavyBlueColor BlueColor) -- both are instances of Color).

When you speak of taxonomizing qualities, you may wish to distinguish
subsumption of the height-to-size variety from subsumption of the
NavyBlue-to-Blue variety.

Incidentally, the "subrelation" structuring relation needs to handle
inverses (if there are any) consistently. If REL1, REL2, REL3 and REL4
are dyadic relations, and (inverse REL1 REL2) and (inverse REL3 REL4),
then: if (subrelation REL3 REL1) then (subrelation REL4 REL2). Etc.
When inverses are dropped, as in Cyc, it may be useful to have variants
of "subrelation" that allow for permuting of the argument-places.
(Keeping in mind that triadic relation has 5 "inverses", a tetradic
relation has 23, etc.)

On whether to have inverses: as I recall the discussion, I myself ranked
the possibilities as 1. (best) Having no inverses, and generating
inverses at the input/output phase, as in Cyc. 2. Giving all relations
all their inverses. 3. (worst) Have inverses for some relations but not
others. But most other people who spoke out at the meeting seemed to
favor 3. (I opposed 3. for the reason that users accustomed to having
inverses on the familiar relations won't know what to do when they need
an inverse for an unfamiliar relation that doesn't come with one. But
my preference is only mild.) I presume that, in policy 3., the
taxonomy-structuring relations would be among those for which inverses
are supplied.

Yours truly, Fritz Lehmann