comments on http://www.w3.org/TR/2006/NOTE-swbp-n-aryRelations-20060412/

This email conveys some comments on the note "Defining N-ary Relations
on the Semantic Web" (12 April 2006).  I realize that this publication
is intended to represent finished work and that no revision is
currently intended.  If you do ever revise the document, however, you
may wish to consider these reactions of an interested reader.

Thank you for addressing this topic, which is an important one to some
prospective users of RDF.  On the whole, I found the paper useful and
interesting; I do have some comments on details.

1) I was surprised that the document appears nowhere to address what
seems to me to be one of the pre-eminent issues with the use of
Pattern 1: it involves ontological commitments that may be
uncongenial.

It may well be the considered opinion of the WG that the engineering
arguments in favor of reifying relations as Pattern 1 does must
outweigh the philosophical arguments that might be brought to bear
against attributing existence to, or making individuals out of,
events, actions, or anything else one might typically describe with an
n-ary relation.  But if so, that fact is worth recording.

Since bad philosophy so frequently makes for bad engineering, I'm not
sure whether the engineering convenience does outweigh the
philosophical objections Quine or others might feel to treating
diagnoses, or even purchases, as individuals.

2) A third pattern may also merit mention, namely currying. An n-ary
relation may be replaced by a binary relation between argument one and
a relation of arity n - 1, which may itself be replaced by a binary
relation between *its* first argument and ...  and so on.  An example
is given in [1].

[1]
http://www.mulberrytech.com/Extreme/Proceedings/html/2004/Sperberg-McQueen01/EML2004Sperberg-McQueen01.html

You may wish not to recommend this pattern, but it seems to me to be
worth mentioning, if only because some readers will think of it and
wonder whether it is to be preferred or not to the Pattern 1 and
Pattern 2 identified in the document.  (I note that it does appear to
avoid requiring anything more in the way of ontological commitment
than the reification of statements.)

3) The note numbers in the document seem to be garbled: in order, they
appear to be "3", "3", "4", "5", "6".  At the end of the document,
there are five notes.

4) What on earth is the motivation of modeling a diagnosis by clumping
Christine and her breast cancer into a single chimerical object?

  :_Diagnosis_relation_1
      a       :Diagnosis_Relation ;
      :diagnosis_probability :HIGH;
      :diagnosis_value :Breast_Tumor_Christine .

Instead of modeling it as involving a patient, a disease, and a
probability?

  :_Diagnosis_relation_1
      a       :Diagnosis_Relation ;
      :diagnosis_probability :HIGH;
      :diagnosis_patient :Christine;
      :diagnosis_disease :Breast_Tumor .


If there is some best-practices point being made by introducing the
individual Breast_Tumor_Christine, you should know that it went clean
over the head of this reader.  If there isn't, perhaps it would lower
the astonishment factor for some readers if the example just relied on
the individuals Christine and Breast_Tumor.  Not everyone will be
pleased, I suppose, to treat diseases as individuals.  But will anyone
not a follower of Meinong really wish to treat Breast_Tumor_Christine
as an individual?  (The footnote numbered 2 suggests that Christine
and her tumor were merged "for simplicity", or may be so read, but I
have not yet been able to think of a way that a universe including
Christine and Breast_Tumor_Christine is simpler than one including
Christine and Breast_Tumor.)

I am similarly puzzled by the restriction of diagnoses to single
diseases and single probabilities; verisimilitude might be better
served by allowing multiple disease/probability pairs.  (Or else say
explicitly that the definition of diagnosis_value and
diagnosis_probability as functions is a didactic simplification.)

Am I right to think that the definitions given for Diagnosis_Relation
do not restrict Christine to having diagnoses whose diagnosis_value
involves Christine?  That is, the definitions given don't seem to rule
out a graph of the form

  :Christine
      a       :Person ;
      :has_diagnosis _:Diagnosis_Relation_1 .

  :_Diagnosis_relation_1
      a       :Diagnosis_Relation ;
      :diagnosis_probability :HIGH;
      :diagnosis_value :Breast_Tumor_Alison .

5) The rising-temperature example also seemed to me to have
verisimilitude problems.  Surely 'trend' is a secondary property,
which can and should be calculated by comparing a temperature reading
with its predecessor in a time series?  I quite understand wishing to
stay away from the problems involved with representations of time
series data, but that seems to suggest it might be desirable to find a
different example which doesn't seem to cry out for time series for
its proper treatment.

6) While on the subject of verisimilitude, I was troubled by the
example given for Pattern 2.  LAX/DFW/JFK doesn't seem a terribly
likely itinerary for a UAL flight with a four-digit number (which
normally means it's a United Express flight or a code-share), and as
far as I can tell United has no flight 3177 and no single flight
numbers which follow the itinerary mentioned.  It's not important to
the example, I admit.  But by the same token, it's not an important
part of your goal to distract the reader by distracting him from your
content by making him wonder how a long-haul service to such important
cities got a four-digit number.

--C. M. Sperberg-McQueen
  World Wide Web Consortium
  MIT CSAIL

Received on Saturday, 22 April 2006 01:44:42 UTC