re-opening the discussion of the FMA-to-OWL conversion

[Todd Detwiler]

After several years without significant forward momentum, we are revisiting the conversion of the FMA into a proper OWL 2.0. As such I will begin using this list again to discuss transformation issues with anyone who feels inclined to participate. Let us start with one that is presently on my mind, reified (or attributed) relationships in frames. Please note, I may refer to approaches that one or more members of this group have suggested in the past. I am simply presenting these as possibilities that have been suggested. I am not, at least at present, attempting to endorse, nor object to any. And if I misrepresent your position, it is not intentional.

Reified relations are used to attach attributes to a relationship instance. For example, in the FMA we have a property called attributed_continuous_with. With this property we can say not only that A is continuous with B, but that A is continuous with B in a superior direction. "Superior" is not an attribute of A, nor of B, but rather of the relationship between A and B. So, in frames this is handled by creating a value type class that itself contains two properties. So, it would look something like this (if we have a class called Attributed_continuous_with_value):

_:A0 has_type Attributed_continuous_with_value
_:A0 related_object B
_:A0 direction "Superior"

A attributed_continuous_with _:A0

The above was written in a sort of triple notation, but note that it is supposed to be frames.

Here is how Christine Golbreich approached this transformation (for attributed_part), using nested class descriptions (from The Foundational Model of Anatomy in OWL: Experiences and Perspectives 2006):



In a recent conversation that I had with Alan Ruttenberg, he opposed the above approach for its indirect connection between between the source structure and its related part. The suggestion was to use axiom annotations as is supported by OWL2. However, I see a couple of possible issues with this. It is clear how to use this approach with annotation properties. For example, I can say that the class fma:Heart fma:Preferred_name "Heart". And I could then attach further properties to that whole statement. Here is an example in rdf/xml for the class Stapedius:

 <owl:Axiom>
        <fma:modification>Fri Jan 14 15:55:18 PST 2000</fma:modification>
        <fma:Date_entered_modified>Thu May 12 14:30:22 GMT-08:00 2005</fma:Date_entered_modified>
        <owl:annotatedTarget>Stapedius</owl:annotatedTarget>
        <fma:Latin_name__TA_>Musculus stapedius</fma:Latin_name__TA_>
        <fma:TA_ID>A15.3.02.062</fma:TA_ID>
        <fma:FMAID>144661</fma:FMAID>
        <fma:author>AUGUSTO V. AGONCILLO, MD</fma:author>
        <fma:authority>Terminologia Anatomica 1998</fma:authority>
        <owl:annotatedProperty rdf:resource="http://sig.uw.edu/fma#Preferred_name"/>
        <owl:annotatedSource rdf:resource="http://sig.uw.edu/fma#Stapedius"/>
    </owl:Axiom>


But it is not clear to me how to attach such annotations to a restriction, as is the case in our A subclassOf (continuous_with some B). Further, I don't see that this would allow me to specify any restrictions on properties other than related_part (i.e. that direction is required when used here, for example).

Any thoughts or past experiences with such conversions would be helpful to me.

Thanks,
Todd

[Chris Mungall]

On Tue, Oct 29, 2013 at 1:53 PM, Todd Detwiler <detw...@gmail.com> wrote:

After several years without significant forward momentum, we are revisiting the conversion of the FMA into a proper OWL 2.0. As such I will begin using this list again to discuss transformation issues with anyone who feels inclined to participate.

Great!

There are probably a lot of people not on this list who can be a massive help to you. I suggest you send a summary to obo-anatomy, together with instructions on how to subscribe.

 

Let us start with one that is presently on my mind, reified (or attributed) relationships in frames. Please note, I may refer to approaches that one or more members of this group have suggested in the past. I am simply presenting these as possibilities that have been suggested. I am not, at least at present, attempting to endorse, nor object to any. And if I misrepresent your position, it is not intentional.

Reified relations are used to attach attributes to a relationship instance. For example, in the FMA we have a property called attributed_continuous_with. With this property we can say not only that A is continuous with B, but that A is continuous with B in a superior direction. "Superior" is not an attribute of A, nor of B, but rather of the relationship between A and B. So, in frames this is handled by creating a value type class that itself contains two properties. So, it would look something like this (if we have a class called Attributed_continuous_with_value):

_:A0 has_type Attributed_continuous_with_value
_:A0 related_object B
_:A0 direction "Superior"

A attributed_continuous_with _:A0

The above was written in a sort of triple notation, but note that it is supposed to be frames.

As a general rule, if you expect some kind of semantics, don't use RDF-style reification. How are you expecting to define this attribution so a reasoner can understand it?

presumably if A is superior to and continuous with B, and B is superior to and continuous with C, then A is superior to C (but not continuous with)

Why not just create sub-relations? Sorry if this has been addressed before..

We do this in uberon by creating relation cross-products, e.g. connected_to x spatial_relation. This makes it very natural to write property chain axioms.

Of course, the resulting ontology can be a bit overwhelming, some people seem to have a kind of threshold of number of relations they deem acceptable, but it's possible to use standard OWL tools to create simplified views for users.

The other form of reification you can take is to reify the boundaries as classes, e.g. there is a "smooth boundary" AB between A and B. A is superior to AB, and AB is superior to B. From this you get inferences such as A is continuous with B. But this turns out to be tricky in practice. And you have an inflation of classes.

btw, I suggest a change in terminology - attributed relationship sounds like a relationship that has attribution - e.g. a PMID supporting a develops_from relation. It sounds like an axiom annotation.

 

Here is how Christine Golbreich approached this transformation (for attributed_part), using nested class descriptions (from The Foundational Model of Anatomy in OWL: Experiences and Perspectives 2006):

<img src="data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABBUAAAPCCAYAAAAnHeAeAABAAElEQVR4AezdCbx+1bw/8H3NrvEaMqtkjJC5TL9kTlIiQ/RTGSOlIkS/QmSMqEj5ZY5UqAgpDYZkTGYVpStcZbimy3//13td67nr2Wc/z3nO9Dvn/H7f7+t1zvM8e6+99lqfvfZa33n9W5uoCQoEAoFAIBAIBAKBQCAQCAQCgUAgEAgEAoFAYIYIXGWG

oops, attachment fail!

I think it's also worth restating your goals: is it to faithfully recapitulate the frames modeling in OWL, or is it to capture the anatomical-relationships-in-reality in OWL? This will influence your approach.

[Todd Detwiler]

As a bit of a follow-up I noticed two things about my last post.

1. If you view it on the web, the image from Christine's paper does not appear. Here is a link: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2270940/bin/nihms15020f8.jpg

2. I didn't mean to paint the input that I've received from Christine and Alan as oppositional. Just two varied approaches. I'm not trying to start a list feud, just a pros and cons discussion.

Thanks for any input you have.

Cheers,
Todd

--
 
---
You received this message because you are subscribed to the Google Groups "fma-owl-2009" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fma-owl-2009...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Todd Detwiler]

Hey Chris,
Thanks for the input. It took me a bit to digest what you meant in a couple of places, particularly "if you expect some kind of semantics" and "anatomical-relationships-in-reality", but I think I understand your meaning. Very useful input that is helping me frame my thinking on this matter.  I have a few brief responses that I've inserted below.
Cheers,
Todd

On 10/29/13 5:02 PM, Chris Mungall wrote:

As a general rule, if you expect some kind of semantics, don't use RDF-style reification. How are you expecting to define this attribution so a reasoner can understand it?

presumably if A is superior to and continuous with B, and B is superior to and continuous with C, then A is superior to C (but not continuous with)

In general, superior_to may not be transitive either (or it depends on how it is defined), but that is of little consequence here as your meaning is clear. By separating those properties they can have their own logical traits. I've had a lengthy discussion with Onard about breaking up the reified relationships in frames into many non-reified relationships. We are in favor, though we recognize a few potential problems;

1. It is clear that my example could be expanded to A continuous_with B, A superior_to B. But that is made possible by the fact that the direction slot (actually called coordinate in the FMA) has a small subset of possible values. Had that been a free-text field with infinite value possibilities, then this transformation would not have worked. Fortunately, I don't think that we have any such cases in the FMA at present.

2. We have identified one interesting reified slot in frames called orientation. Basically it says that the Esophagus has orientation Plane of pharyngoesophageal junction, superior and Plane of esophagogastric junction, inferior. It would not be possible for someone inspecting the esophagus to see these relationships directly with individual properties as they are indirect parts and you need to see both values to get an overall mental map of the orientation.

But, we are going to set #2 aside for now (and another case that I haven't mentioned yet, but I'm sure we can find a solution for, stay tuned) and pursue the notion, for now, of doing away with reification.

Why not just create sub-relations? Sorry if this has been addressed before..

We do this in uberon by creating relation cross-products, e.g. connected_to x spatial_relation. This makes it very natural to write property chain axioms.

We used to do this in the FMA, things like continuous_with_superior. But it lacks the ability, as you pointed out earlier, to attach semantics to the "superior" part separate from the "continuous_with" part. Also, in some cases, the explosion in the number of properties just got ridiculous.

Of course, the resulting ontology can be a bit overwhelming, some people seem to have a kind of threshold of number of relations they deem acceptable, but it's possible to use standard OWL tools to create simplified views for users.

The other form of reification you can take is to reify the boundaries as classes, e.g. there is a "smooth boundary" AB between A and B. A is superior to AB, and AB is superior to B. From this you get inferences such as A is continuous with B. But this turns out to be tricky in practice. And you have an inflation of classes.

btw, I suggest a change in terminology - attributed relationship sounds like a relationship that has attribution - e.g. a PMID supporting a develops_from relation. It sounds like an axiom annotation.

Attributed relationship and reified relationship are terms that are used interchangeably in frames. I'll stick to using the latter from now on when communicating with OWL folks.

 

Here is how Christine Golbreich approached this transformation (for attributed_part), using nested class descriptions (from The Foundational Model of Anatomy in OWL: Experiences and Perspectives 2006):

<img src="data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABBUAAAPCCAYAAAAnHeAeAABAAElEQVR4AezdCbx+1bw/8H3NrvEaMqtkjJC5TL9kTlIiQ/RTGSOlIkS/QmSMqEj5ZY5UqAgpDYZkTGYVpStcZbimy3//13td67nr2Wc/z3nO9Dvn/H7f7+t1zvM8e6+99lqfvfZa33n9W5uoCQoEAoFAIBAIBAKBQCAQCAQCgUAgEAgEAoFAYIYIXGWG

oops, attachment fail!

Yes, I noticed this when looking at the Google group on the web (though interestingly the image came through to my email). Anyway, I sent a subsequent email with a link to the image. Sorry about that.

I think it's also worth restating your goals: is it to faithfully recapitulate the frames modeling in OWL, or is it to capture the anatomical-relationships-in-reality in OWL? This will influence your approach.

Our goal is that latter. We already have the former, and it is no longer sufficient. However, having said that, on first pass we aren't going to do a lot of addition of knowledge that isn't present in the frames model. I know that this would be useful, but it will be left for future augmentations. But we will try and model the knowledge that is already present in the sort of patterns that will have the most utility in a description logic modeling paradigm.

Cheers,

[Chris Mungall]

On Wed, Oct 30, 2013 at 11:43 AM, Todd Detwiler <detw...@gmail.com> wrote:

1. It is clear that my example could be expanded to A continuous_with B, A superior_to B.

This isn't quite the same. Remember, you won't be working with instances any more, so you can't say A continuous_with B. It's worth getting in the habit of writing and thinking in OWL (or classes in general)

You would say

A SubClassOf continuous_with some B
A SubClassOf superior_to some B

This is entailed by but weaker than saying there is some B that A is both continuous with and superior to. The above admits models where you are talking about two separate instances of B in each sentence.

This may not may not matter depending on what your use cases for these relations are

But it is worth bringing up, as I think this switch to thinking in a more class-oriented way will be a radical change for the FMA. In fact it's even worth going back and re-examining the assumptions underlying this fma-owl-2009 project. Looking at the FMA literature, the talk of canonicality, the way many structures are modeled in the FMA (with axioms pushed down to the leaf nodes of the partonomy), an argument can be made that the natural fit for the FMA is an instance-based representation of an abstract adult human.

[Todd Detwiler]

On 10/30/13 12:54 PM, Chris Mungall wrote:

On Wed, Oct 30, 2013 at 11:43 AM, Todd Detwiler <detw...@gmail.com> wrote:

1. It is clear that my example could be expanded to A continuous_with B, A superior_to B.

This isn't quite the same. Remember, you won't be working with instances any more, so you can't say A continuous_with B. It's worth getting in the habit of writing and thinking in OWL (or classes in general)

You would say

A SubClassOf continuous_with some B
A SubClassOf superior_to some B

That is what I meant. Or rather, I was being vague about whether or not A and B represented classes or individuals. Since there are not statements about individuals in the FMA, it would naturally follow that I meant classes. Sorry for creating that confusion.

This is entailed by but weaker than saying there is some B that A is both continuous with and superior to. The above admits models where you are talking about two separate instances of B in each sentence.

This may not may not matter depending on what your use cases for these relations are

But it is worth bringing up, as I think this switch to thinking in a more class-oriented way will be a radical change for the FMA. In fact it's even worth going back and re-examining the assumptions underlying this fma-owl-2009 project. Looking at the FMA literature, the talk of canonicality, the way many structures are modeled in the FMA (with axioms pushed down to the leaf nodes of the partonomy), an argument can be made that the natural fit for the FMA is an instance-based representation of an abstract adult human.

Onard and I were just discussing questions like, can the FMA be instantiated? Is there a canonical human instance? Does that require all canonical parts? Is there a canonical heart? Tough questions for sure. But right now I am primarily concerned with transferring the knowledge that exists in the FMA, not re-imagining it. So, is your argument that Heart, Right atrium, etc. should be instances (Organ, Artery, etc. should be classes)? Or perhaps there should be both a class Heart and an instance heart? If so, what does this buy us?

I've previously pondered the generation of such an artifact, from the class structure, for the purpose of running queries (SPARQL type queries). But that was a pragmatic, use driven reason, to fit a particular need I had but did not necessarily fit the needs of the community. I'd be curious to hear your rational for such a model.

Thanks again,
Todd

 

But that is made possible by the fact that the direction slot (actually called coordinate in the FMA) has a small subset of possible values. Had that been a free-text field with infinite value possibilities, then this transformation would not have worked. Fortunately, I don't think that we have any such cases in the FMA at present.