Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

SUO: RE: About SUMO's Mereotopology



Hi Pierre,
 
    Thanks a lot for your detailed message.  Many interesting and insightful comments here.  See my responses below.
 
-Ian
-----Original Message-----
From: Pierregrenon@aol.com [mailto:Pierregrenon@aol.com]
Sent: Wednesday, March 13, 2002 5:59 PM
To: apease@ks.teknowledge.com; iniles@teknowledge.com
Cc: standard-upper-ontology@ieee.org
Subject: About SUMO's Mereotopology

FAIW, here is a list of questions and comments regarding the first part of the module 'Mereotopology' of the SUMO (not including Varzi & Casati's hole story). I'd appreciate any comments, feel free to reply to me personally.

This really accounts for a first thorough pass over this material and aims mainly at making sure I'm not tripping too badly before going further. Thanks for bringing me kindly back to reason as needed. Another preliminary remark, I tried to play the game and to worry, at the level of my humble understanding, about the quality of this module within the context of the SUMO; in other words, trying to comment without arguing too much over the material of the base and structural modules. After the preliminaries, it was easier to go through the file and make comments and suggestions as they occur. I hope that this won't be too tedious to read and might be somewhat constructive at least as a starter for a few rounds of discussion.


Preliminary comments regarding &%part:
Something needs to be done with &%part, there are serious problems, even if trying to cope with the top level 'concepts' of the SUMO. The central problematic thing is the domain of this relation, currently &%SelfConnectedObject. IMHO, this is absurd enough but, even worse, the SUMO is not consistent with respect to this choice.

Is the rationale that disconnected objects do not form an object? Are not part of an object? It seems that they form &%Collections, I confess I have a hard time with this. Collections have members that are informally described as parts but, god forbid, collections are not mereological sums of their members, rather some sort-of intensional sort of set theoretical weirdos. Anyway, no need to be controversial here.

One of the first things to do is to decide which of &%part's argument constraints or of &%MereologicalSumFn's and others need to be modified. So, either generalize to (domain part 1 [and 2] Object) or alternatively restrain the domain of the mereological operators, which currently do not require self connected objects. The first solution would have important consequences for SUMO regarding its mereology. The second solution is unfortunate (the only correct way I could find to qualify this). The current nonsense comes from the fact that you could form the sum of two members of a collection while neither of these members would be 'parts' of this sum! Or am I missing something?

I suggest the first solution of course and moreover to then question the notion of &%member and think about what kind of species &%Object is lacking of. I seem to remember ranting about this awhile ago, I'll pass on this for this time around. However, let me add that, imho, the use of &%Collections conceived as disconnected objects would be sounder if &%member were a sub-relation of &%part, same thing for &%subCollection, assuming the broadening of the domain of &%part in each argument place.


Still about &%part, for the &%LinguisticObject guys… there are a few axioms related to linguistic stuff that are using &%part, see &%SymbolicString, &%Morpheme and &%Word for instance. Those aren't well formed (or are they?) given the domain of &%part (which doesn't subsume &%ContentBearingObject if I'm not mistaken). Three possible solutions to this, maybe more, maybe with combination: i) generalize the use of &%part (I would tend to prefer that anyway, although I'm not sure about the particular of the linguistic stuff), ii) subsume &%Morpheme and others under self connected object (I don't know what kind of sense that makes for the various specialization of &%ContentBearingObject), iii) use another relation where part was intended (to me this almost sounds like a gratuitous step toward i) though). What d'y'all say?  
 
As I read you, you have found the following problems with the mereotopological component of the SUMO:
 
1.  Given the current arg type restrictions on 'part' and 'MereologicalSumFn', it is possible to form mereosums from two 'Objects' and yet not to have those 'Objects' be 'parts' of the meresum.  This is, as you point out, extremely counterintuitive.
 
2.  The use of 'part' with instances of 'LinguisticExpression' should not be allowed given the current arg type restrictions on 'part'.
 
3.  There is currently no relationship between 'part' and 'member', even though these are very closely related predicates.
 
As you suggest, we can resolve these three problems by generalizing the arg type restrictions on 'part' to 'Object' and by making 'member' a subrelation of 'part'.  Both of these suggestions sound good to me.  If no one objects, I'll incorporate them into the SUMO. 



From now on, to follow these comments, I suggest holding a print out of the part of the file on mereotopology up to Varzi and Casati's stuff.


1. The document mentions papers from Smith and from Guarino. Any paper in particular ? Which? I'd be interested in this precision.  
 
Fair enough.  I'll try to track down the references. 


2. About &%connects. Shouldn't the domain of the first argument be &%SelfConnectedObject? Maybe that's just a need for clarifying the notion of &%SelfConnectedObject or maybe the relation &%connects itself. It looks like some configurations are allowed as examples by the present definition that seem not to have been intended.

Let OBJ2 and OBJ3 be two non connected pieces of ceramic on Mr. Bell's workbench (here, I guess it suffices to say that OBJ2 does not touch OBJ3). Let OBJ1 be the mereological sum of two layers of glue such that LG1 (respectively LG2) is spread over OBJ2 (respectively OBJ3) at time t by Mr. Bell. Let's not argue over glue's physical properties, unless that really is something decisive I'm missing, just take the layer of glue as two &%(SelfConnected?)Objects. At t, &%connects does not hold of LG1 and LG2, that is OBJ1 is not self connected. At time t+s, Mr. Bell assembles OBJ2 and OBJ3, thus fixing Odette's saucer. LG1 and LG2 are connected (say, &%meetsSpatially holds). It seems to me that the current definition allows &%connects to hold of OBJ1, OBJ2 and OBJ3 in this order at both time t and t+s. What's puzzling being that it could hold at t.  
 
I agree that this is counterintuitive and that the best way of resolving the problem is to narrow the arg type restrictions on 'connects' to 'SelfConnectedObject'. 


3.
Isn't

(=>
(partlyLocated ?OBJ ?REGION)
(overlapsSpatially ?OBJ ?REGION))

better replaced with,

(subrelation partlyLocated overlapsSpatially)?  
 
 This seems better to me too.  If no one objects, I'll make the replacement in the SUMO source file. 

4. About &%overlapsPartially: this is symmetric, right? Then, couldn't the current implication be simplified?

For instance,

(not
(and
(overlapsPartially ?OBJ1 ?OBJ2)
(part ?OBJ1 ?OBJ2)))

Also, isn't this relation irreflexive?  
 
Right, the predicate is both symmetric and irreflexive, and I agree that the axiom can be simplified as you indicate. 


5. A question about &%superficialPart. Why is it so that an appeal to the geometric complement of an object should be taken as an intuitive recourse? Why isn't this notion in the SUMO and why not formalizing the parenthetic criterion in the documentation of &%superficialPart? It seems that an object has a single geometric complement, is that correct? Is the 'geometric complement' something potentially as general as an &%Object or more specifically a &%Region? Or it could a &%Collection maybe? Of regions only? 
 
I think the appeal to "geometric complement" is probably not very helpful.  I'll reformulate the documentation string. 

6. Isn't there a problem with the axiom for &%surface:

(=>
(surface ?OBJ1 ?OBJ2)
(forall (?PART)
  (and
(instance ?OBJ1 SelfConnetcedObject)
           (=>
(superficialPart ?PART ?OBJ2)
(part ?OBJ3 ?OBJ1)))))?

Assume (surface ?OBJ1 ?OBJ2). On one hand, a surface of an object is self connected, (instance ?OBJ1 SelfConnectedObject). (Note that given the present setting, this is trivial given the domain specification of the relation.) On the other hand, any superficial part of an object is a part of this self connected object, (forAll (?PART) (=> (superficialPart ?OBJ3 ?OBJ2) (part ?OBJ3 ?OBJ1))).  
 
I agree that the clause about 'SelfConnectedObject' is redundant given what the arg type restrictions should be on 'surface'.  I'll add these restrictions and delete the clause in the source file.  As for the rest, note that you have not copied the axiom correctly. 

That just can't be, that is, unless ?OBJ2 has only 1 surface, i.e. only one maximally connected superficial part. Take an hollow object such as a tennis ball. The geometric complement is disconnected, that is an indication that there are at least two superficial parts. There's an 'exterior' and an 'interior' surface in other words.
The implication in the consequent of the present axiom could hold in case the object has a certain property of 'density', maybe this can be indexed to the number of maximally connected superficial parts (the present form of the axiom is adequate if the object has no 'interior superficial part'). Note that 'interior superficial parts' are allowed, those are not '&%interiorParts'!

This could suggest introducing terms for relation for maximally connected part, vocabulary for density or alternatively flesh out the hierarchy of self connected objects according to their topology, and accordingly, sub-relations of &%superficialPart, etc.  
 
Well, I think an easier route would be just to allow that hollow objects like tennis balls can have two disconnected surfaces, the inner surface and the outer surface, and to change the documentation string of 'surface' so that readers aren't left with the incorrect impression that all 'SelfConnectedObjects' have a single surface. 



I have to stop there, apologies for the messy presentation, poor language and any confusion.

Thanks for your time,
Pierre