Fwd: Parallel Reasoning over large datasets

[Regis Pires Magalhães]

Paper from ISWC 2009: "Parallel Materialization of the Finite RDFS Closure for Hundreds of Millions of Triples" [1].

[1] http://data.semanticweb.org/pdfs/iswc/2009/paper241.pdf

The implementation is written in C using Redland RDF Libraries (http://librdf.org/).

"This approach may be useful in a MapReduce implementation of an RDFS reasoner."

---------- Forwarded message ----------
From: Stephen Allen <sal...@apache.org>
Date: 2011/11/3
Subject: Re: Giraph: anyone?
To: jena-...@incubator.apache.org


Hi Paulo,

You may be interested this paper from ISWC 2009: "Parallel
Materialization of the Finite RDFS Closure for Hundreds of Millions of
Triples" [1].

-Stephen

[1] http://data.semanticweb.org/pdfs/iswc/2009/paper241.pdf

On Thu, Nov 3, 2011 at 11:52 AM, Paolo Castagna
<castagn...@googlemail.com> wrote:
> Hi,
> I wonder if anyone here had a look at Apache Giraph (in the incubator)
> (i.e. a Google Pregel clone), here: http://incubator.apache.org/giraph/
>
> I am curious to know if and how this could be used to implement an
> rule based inference engine. :-)
>
> Pregel seems to me a better model/architecture (than MapReduce) for
> things such as the RETE algorithm.
>
> Having said that, a think you can easily do with MapReduce is to
> distribute all your vocabularies/ontologies to all your nodes (via
> DistributedCache) load them in RAM (they are typically not that
> big) and then perform inference in parallel.
>
> An example of this is here: [1,2]. This is using RIOT infer approach,
> but I was wondering if I could just use Jena inference engine there
> and if that would work on one triple at the time. I don't think
> everything works in this case.
>
> However, perhaps there is a way to group the RDF data so that
> inference would work as if all the instance data were available.
> Any idea/suggestion?
>
> Since I am here and I was looking at RIOT infer command implementation
> yesterday, I found this comment here [3]:
>
>  * TDB Infer
>  *   RDFS
>  *   owl:sameAs (in T-Box, not A-Box)
>  *   owl:equivalentClass, owl:equivalentProperty
>  *   owl:TransitiveProperty, owl:SymmetricProperty
>
> (by the way, interesting comments in the RIOT infer's command ;-))
>
> I see how RDFS is implemented (very elegant and compact).
>
> I can see how we could do similar things for:
>
>  - owl:equivalentClass
>  - owl:equivalentProperty
>  - owl:SymmetricProperty
>
> I think I know how to do:
>
>  - owl:sameAs (in T-Box, not A-Box)
>
> But, I do not really see how we could do owl:TransitiveProperty in a
> similar streaming fashion. Is that possible?
>
> I was also thinking if those "fragments" of OWL should be added to the
> existing Java classes or we should keep RDFS and OWL separate and have
> InferenceProcessorOWL.java.
>
> Cheers,
> Paolo
>
>  [1]
> https://github.com/castagna/tdbloader3/blob/master/src/main/java/org/apache/jena/tdbloader3/InferDriver.java
>  [2]
> https://github.com/castagna/tdbloader3/blob/master/src/main/java/org/apache/jena/tdbloader3/InferMapper.java
>  [3]
> http://svn.apache.org/repos/asf/incubator/jena/Jena2/ARQ/trunk/src/main/java/riotcmd/infer.java
>

[Fabio Porto]

Oi Regis,
 obrigado.
aperece interessante,
abcs,
Fabio.

2011/11/4 Regis Pires Magalhães <regisp...@gmail.com>

--
Dr. Fabio Porto
LNCC - Computer Science Department