Re: [BIONT-DSE] Inclusion versus exclusion criteria

On Sep 13, 2007, at 6:25 PM, Chimezie Ogbuji wrote:

> On Thu, 2007-09-13 at 16:44 +0100, Bijan Parsia wrote:
>> I trimmed the ccs since I get scared if I have to scroll a cc list.
>
> Thanks, I have a bad habit of not doing that :)
>
>> Eh. Not really. First of all, "domain of discourse" has a couple of
>> specific technical meanings so we should be a bit wary about it.
>
> What I meant by "domain of discourse" was essentially the subject  
> domain
> of the assertions (in this case: assertions made as a result of the
> patient care process)

Yes, see how that conflicts with "domain of discourse" meaning the  
universe that is the object of the interpretation function.

>>>   Those two papers address a different (non-trivial)
>>> question: whether the semantic web stack is best built on top of
>>> Description Logic or Logic Programming
>
>> DLP isn't a mapping. For mapping from, e.g.,
>
> I don't follow how DLP is not a mapping (unless you are specifically
> referring to the 'concerns' mentioned above about if it is indeed  
> lossy
> i.e., 1-1).

Yes. It's not a mapping from DL to LP. But from DL to Horn Logic.

> It is defined as a function whose input are DL expressions
> (an expressive subset: DHL) and the output is "equivalent" Horn Logic.
>
>> LP to FOL, look at
>> things like clark's completion.
>
> DHL is in the other direction, no?

DHL? DLP? For them to be meaning preserving they need to be  
bilateral, eh?

But I don't know of a way (at least a standard way) to go from first  
order horn (without function symbols) to Datalog under LP semantics  
for beyond ground entailments. My intuition sez it'd be pretty bad.

[snip]
>> Note that DLP-LP and DLP-FOL do coincide for certain classes of
>> answer for certain classes of query (since their ground entailments
>> coincide).
>
> Yes, this is clearly stated in the DLP paper and it is *this*
> intersection that is of primary interest because the LP evaluation
> complexity is significantly less (at least that is my understanding).

Your understanding is wrong. Think about it: If a certain class of  
entailments *are the same* under both semantics, then any technique  
that can compute the entailments under the first semantic can do the  
same under the second semantics. That is, the complexity of that  
problem is *exactly the same*.

There are a variety of algorithms for computing answers to that  
problem and they vary both in their worst case complexity (i.e., they  
can be in worse complexity classes, or have worse constant factors)  
and in their observed behavior.

>> This has gotten a bit too deeply technical perhaps.
>
> Yes, I would agree.  However, I think further research into this
> intersection

The research is active and ongoing.

> is a very important contribution towards reducing the
> (currently) intractable nature of inference over large ontologies
> (especially EL++ ontologies

It's not intractable currently. See CEL and Hermit.

> which are very common in this domain:
> SNOMED-CT, GALEN, FMA, etc..)

DLP is not really very helpful for dealing with EL++. EL++ is already  
well known to be horn and there are several techniques for dealing  
with it. The latest is hypertableau. See the link from my blog post:
	<http://clarkparsia.com/weblog/2007/06/25/scaling-owl-two-new-ways/>

I know you're really into it, but DLP is quite the uninteresting Horn  
Description Logic, esp. for Life Sciences. Check out the HornSHIQ  
papers, the DL Lite papers (which is relational!), and various EL  
papers.

(I think the worst thing about the DLP paper is that damn shield  
diagram which is one of the top 15 misleading diagrams out there.  
It's very seductive but really overall hurts ones understanding of  
the relationship between the various logics.)

>> From a language/
>> infrastructure design perspective, it seems clear that having some
>> sorts of non-monotonic features are quite useful in a number of
>> circumstances.
>
> Absolutely, this was really the main point I was trying to make. It
> just didn't seem right to suggest a framework that excludes certain
> non-monotonic features that make perfect sense for the very  
> controlled /
> focused nature of clinical & biomedical research.

Now here, *we* coincide :)

Cheers,
Bijan.

Received on Thursday, 13 September 2007 20:54:42 UTC