Re: RIF-in-RDF: Requirement 4 [Switching to KR Argument]

On Sun, 2010-07-25 at 22:30 -0400, Michael Kifer wrote: 
> OK, I think I am getting closer. You should have started with this.
> So, it is not a matter of inference and its completeness but rather
> of which rule exactly a particular RDF graph encodes, right?
> 
> That is, if we do not use lists explicitly in RDF then there is no way to encode
> 
>     a&b&c=>d
> 
> correctly? That it will be instead be encoding a set of rules that in addition
> includes  a&b=>d, a&c=>d, a=>d, etc.

Depends what we are trying to do here.

Up until now I have assumed we are doing a pure syntactic encoding,
there's no notion that RIF semantics are an extension of the RDF
semantics. This is not an embedding. The encoding is not required to be
monotonic. It is not the same situation as OWL extending RDF.

We do need the encoding to be robust against things that RDF processors
might do. So if D is the decode transformation that maps an RDF graph to
a set of RIF rules (or rejects it). Then we require that D be invariant
under RDF simple entailment:

    if    D(G) = R 
    and   Gr is equivalent to G under RDF simple entailment 
    then  D(Gr) = R

If we have an RDFS "ontology" O for the encoding then we also require
invariance under the additional entailments of that ontology:

    if    D(G) = R 
    and   (G + O) entails Go by RDFS entailment 
    then  D(G + Go) = R

We probably also want additivity:

    if    D(G) = R   
    and   D(G') = R'
    then  D(G + G') = R + R'  (for a suitable interpretation of +)

But we don't require this for an arbitrary G' which does not itself
encode a RIF rule set, you can't add random triples in safely.
Conversely, you can't take arbitrary subgraphs of G drop random triples
and expect to get a subset of R, even if that subgraph looks well
formed.

Sandro seems to be asking for a stronger monotonicity requirement that
any subgraph that looks like a RIF rule set should be safe to decode:
     
    if    D(G) = R   
    and   G' \subgraph G and G' is well-formed as a RIF encoding
    then  D(G') \subset R

I guess don't object to having such a requirement, and in that case you
would indeed need to uses lists instead of repeated properties in order
to have a suitable test of well-formedness of the encoding. Though I
still don't buy the notion of incomplete transformations as being a
motivation for this requirement.

Dave

Received on Monday, 26 July 2010 09:32:43 UTC