W3C home > Mailing lists > Public > public-semweb-lifesci@w3.org > August 2011

Re: What OWL cant do

From: John Madden <john.madden@duke.edu>
Date: Fri, 26 Aug 2011 16:23:47 -0400
Cc: "public-semweb-lifesci@w3.org" <public-semweb-lifesci@w3.org>
Message-Id: <FA29B8D7-9166-4AC6-994E-F667B471B899@duke.edu>
To: "Hau, Dave (NIH/NCI) [E]" <haudt@mail.nih.gov>
I believe Robert correctly analyzed the situation, and I am not sanguine about prospects for "determining equivalence" of any "expressions" that use RIM or SNOMED "codes" inside some kind of OWL-like syntax.

Two expressions "are equivalent" in OWL/RDF if they denote sets that have the same members (same extension) in all models.

SNOMED is not designed and built to support especially informative answers in reasoning of that kind "out-of-the box", i.e. it is underspecified  for that kind of reasoning. I don't know as much about the RIM, but I suspect the same is true of it.

This relates to several features of SNOMED. 

One is that SNOMED (intentionally) does not specify a particular "information model". What that means (in part) is that SNOMED does not take an "official" stance as to whether a SNOMED "concept" denotes a set (RDF/OWL-speak: rdfs/owl:Class) of individuals (OWL-speak: instances), and (if so) what those instances might be. Rather, SNOMED intentionally leaves that decision up to implementers. Some implementers might choose to interpret any particular SNOMED concept as not being a class/set at all; some might interpret it as being an rdfs:Class; some might interpret as being an owl:Class; some might interpret it as being a class/set in some non-OWL language (e.g. in some particular relational database algebra). Put differently, the meaning of SNOMED is_A is (intentionally) underspecified from the RDF/OWL perspective.

A second feature is that even if two implementers agree on the interpretation of each and every SNOMED concept with respect to whether they are in fact owl:Classes, and what the OWL meaning of their is_A relations to each other are, this still doesn't mean that they necessarily agree what the real-world nature of the instances of these classes are. For example, take the SNOMED code for "influenza". If that code be identified with an owl:Class of "somethings", what are those "somethings"? Clinically-identified cases of influenza? Variants of the disease "influenza"? All cases of influenza virus infection whether diagnosed or not? There is actually no single "correct" answer. To get an answer, you propose an information model, and you specify it. SNOMED intentionally leaves that part up to you.

Third, OWL allows a rich set of logical axioms to be specified on classes, including disjointness relations, restrictions on properties of various kinds, etc. SNOMED intentionally does not come with very many such class relations modeled "out-of-the-box". Therefore, even if #1 and #2 were out of the way, without such explicitly specified class relations, an OWL reasoner wouldn't have a lot to work with. Dean Allmenang likes to say that "it's hard to say anything *false* in RDFS". Out-of-the-box, SNOMED is modeled at a level of axiomatic explicitness that somewhat resembles typical RDFS. From the perspective of an OWL reasoner, there just isn't a lot there to get traction on. I venture to say that many, many, many "expressions" comprised of SNOMED codes embedded inside some sort of OWL-like syntax -- expressions whose lexical representations (using SNOMED preferred labels, say) you and I might subjectively respond to as expressing "clinically quite different" things -- would be logically indistinguishable to an OWL reasoner provided just with a T-box consisting of some or another version of some kind of "OWL-SNOMED out-of-the-box".


On Aug 26, 2011, at 2:16 PM, Hau, Dave (NIH/NCI) [E] wrote:

> Clarification...
> 
> The first one means:
> 
> Act.participant[@typeCode='verifier']
> 
> The second one means:
> 
> Act.relationship[@typeCode='verifier'].Act[@code='verification', @moodCode='EVN'].participant[@typeCode='performer']
> 
> Is there any reasoning that can be done to determine equivalence between the two?
> 
> If not, is there any way to modify the modeling framework to allow that reasoning to happen?
> 
> - Dave
> 
> 
> 
> 
> -----Original Message-----
> From: Hau, Dave (NIH/NCI) [E] 
> Sent: Friday, August 26, 2011 2:09 PM
> To: public-semweb-lifesci@w3.org
> Subject: FW: What OWL cant do
> 
>> <<One common one is substituting a participation for a full-blown Act. For example, VRF (verifier) hanging off an Act, vs. Act being subject of a Verification act with a performer of the verifier. >>
> 
> Is there any way of telling these two things are the same in OWL, short of using a rule?
> 
> "VRF (verifier) hanging off an Act", in OWL speak, means you have an Act with a property which is a child class (verifier) of Participation.
> 
> "Act being subject of a Verification act with a performer of the verifier" means you have a child class (Verification) of Act with a property of Participation that then has a property of a child class (verifier) of role.
> 
> That's my preliminary interpretation.  Corrections welcome from those who are more familiar with RIM and/or OWL.
> 
> - Dave
> 
> 
> 
> -----Original Message-----
> From: owner-rimbaa@lists.hl7.org [mailto:owner-rimbaa@lists.hl7.org] On Behalf Of peter
> Sent: Friday, August 26, 2011 12:57 PM
> To: robert worden; 'Yeb Havinga'; 'Lloyd McKenzie'
> Cc: 'Andrew McIntyre'; 'Grahame Grieve'; 'Eliot Muir'; 'Zel, M van der'; 'HL7-MnM'; 'RIMBAA'; 'HL7 ITS'
> Subject: What OWL cant do
> 
> Figure it's time to change the subject line.
> 
> Thinking about one of Lloyd's examples, it makes me think there are certain kinds of semantic non equivalence that 
> an OWL reasoner would not think are the same, but a clinician would consider for all practical purposes exactly the 
> same.
> 
> One of Lloyd's examples is this.
> 
> <<One common one is substituting a participation for a full-blown Act.  For example, VRF (verifier) hanging off an 
> Act, vs. Act being subject of a Verification act with a performer of the verifier. >>
> 
> If I go into reasoner mode these are different. They can not be normalized to the same graph.
> 
> If I go into human clinician mode these are the same.
> 
> Possibly a reasoner which then feeds a rules engine could handle this, but now we're getting way to complex.
> 
> Therefore, I'm not convinced we could use OWL for this task (showing the example cases above to be the same).  
> It would seem we'd have to come up with strict rules like "don't ever express and Act as the subject of a verification 
> Act, you must always hang the verifier of the act itself."   Would there end up being hundreds of rules, or are there 
> only a finite relatively small number of problems like this?
> 
> RFH would solve this problem. There'd be only one Resource available to express this in one way.
> But I think Lloyd's example above disproves the idea that OWL alone could do this. 
> 
> I'd say it this way.
> 
> A case of mathematical semantic inequality may in fact still be a case of clinical equality.
> 
> 
> 
> 
>> -------Original Message-------
>> From: robert worden <rpworden@me.com>
>> To: 'Yeb Havinga' <yhavinga@gmail.com>, 'Lloyd McKenzie' <lloyd@lmckenzie.com>
>> Cc: 'Andrew McIntyre' <andrew@medical-objects.com.au>, 'Grahame Grieve' <grahame@kestral.com.au>, 'Eliot 
> Muir' <eliot.muir@interfaceware.com>, 'Zel, M van der' <m.van.der.zel@umcg.nl>, 'HL7-MnM' <mnm@lists.hl7.org>, 
> 'RIMBAA' <rimbaa@lists.hl7.org>, 'HL7 ITS' <its@lists.hl7.org>
>> Subject: RE: RFH - What can Owl do?
>> Sent: 26 Aug '11 02:57
>> 
>> Hi All -
>> 
>> I'm with Yeb on this one - although I can't express it very articulately,
>> having been away from it for many years. Here's my take:
>> 
>> Any system of symbols  - like the RIM, or SNOMED CT, or a natural language -
>> can be helped by having a Model-theoretic Semantics. Here, the 'model' is
>> basically just set theory , and not high-powered maths at all. But for any
>> expression made up of the symbols, there are rules (compositional rules) to
>> say what the expression means, in terms of the set-theoretic objects. This
>> is the 'Intepretation' of the expression, in the model.  Having these rules,
>> and so having a set-theory interpretation of any expression, brings immense
>> clarity to any debate about 'what do the symbols mean?' For any expression,
>> you can point to a set-theoretic object, and poke it to see how it behaves.
>> 
>> This works well in natural language semantics (e.g Montague, Situation
>> semantics) and computer language semantics. I suspect that some of the long
>> and inconclusive debates about RIM semantics (negationInd, etc.) have arisen
>> precisely because there is not a model-based semantics for the RIM. With a
>> model-based semantics, they might have been closed down rather quickly.
>> Without it, everybody has their own model in their head.
>> 
>> But Fresh Look can't stop and do this before it does anything else - and in
>> any case, who has got time to do it, and what benefit would it bring? I
>> think it would bring clarity, and tell us just what we can't do safely with
>> the RIM. This includes knowing whether two RIM-based models are equivalent -
>> the set-theoretic model is the test.  Could we spin off an academic research
>> project?
>> 
>> Cheers
>> 
>> Robert
>> 
>> 
>> 
>> mobile: 07970 197968
>> landline: 01353 777668
>> 
>> -----Original Message-----
>> From: Yeb Havinga [mailto:yhavinga@gmail.com]
>> Sent: 26 August 2011 10:24
>> To: Lloyd McKenzie
>> Cc: Yeb Havinga; robert worden; Andrew McIntyre; Grahame Grieve; Eliot Muir;
>> Zel, M van der; HL7-MnM; RIMBAA; HL7 ITS
>> Subject: Re: RFH - What can Owl do?
>> 
>> Hi Lloyd, list
>> 
>> On 2011-08-25 18:18, Lloyd McKenzie wrote:
>>> Hi Yeb,
>>> 
>>> Everything's on the table.  If there are changes to RIM structure that
>>> would improve its implementability or usefulness in analysis, those
>>> can be considered.  However, you're going to need to provide some
>>> additional guidance on what changes you'd like to see because I'm not
>>> quite following.
>> 
>>> 
>>> "X observes that Y orders that Z take medication" would be X as data
>>> enterer, Y as author and Z as subject.
>> Hmm this was just an example to illustrate the RIMs lack of being able to
>> express stacked modalities - without wanting to start a hl7 modeling
>> discussion, if there's a situation in real life where no computers are
>> around for data entry; it's still possible that x observers that y orders
>> that z take medication (or any other substance).
>>>   If you make it more complex, you may need to introduce ControlActs.
>> 
>> I don't understand how; ControlAct seems to be meant for something else,
>> looking at it's associations and textual definition.
>>> 
>>> In terms of having a mathematical representation of the world, I'm not
>>> sure I understand.  An example would definitely help.
>> 
>> I browsed a little for how this was described for OWL: see section 3.1 from
>> http://www.w3.org/TR/owl-semantics/direct.html are the kinds of things I'm
>> talking about - IMO HL7 would benefit from something similar for both the
>> datatypes and the RIM.
>> 
>> Also, I once gave a presentation for a RIMBAA meeting, that among other
>> things talks about these things - added as attachment.
>> 
>> regards,
>> Yeb Havinga
>> 
>> 
>> 
>> 
>> 
>> ************************************************
>> To access the Archives of this or other lists or change your list settings and information, go to: 
> http://www.hl7.org/listservice
>> 
> 
> 
> ************************************************
> To access the Archives of this or other lists or change your list settings and information, go to: http://www.hl7.org/listservice
> 
> 
Received on Friday, 26 August 2011 20:24:16 UTC

This archive was generated by hypermail 2.3.1 : Wednesday, 7 January 2015 14:52:48 UTC