Re: Information Preserving and ISSUE-42

On Wed, 2011-05-18 at 08:26 -0400, Marcelo Arenas wrote:
> On Wed, May 18, 2011 at 7:51 AM, Alexandre Bertails <bertails@w3.org> wrote:
> > On Wed, 2011-05-18 at 12:07 +0100, Richard Cyganiak wrote:
> >> Hi Juan,
> >>
> >> On 18 May 2011, at 05:44, Juan Sequeda wrote:
> >> > IF the direct mapping has knowledge of the schema then translating NULLs is not necessary for information preserving.
> >>
> >> Yes.
> >
> > What do you guys mean by "the direct mapping has knowledge of the
> > schema"?
> 
> It means that the mapping also produces some triples to represent the
> relational schema (for example, to say what are the tables in the
> relational schema, and what are the attributes in these tables).
> Notice that the current version of the direct mapping does not
> implement this.

That's true, but we've always said that was an "easy" task that was
postponed until we would have solved all the data mapping itself.

In any case, the Direct Mapping *does* have "knowledge of the schema",
the same way it has knowledge of the data, this is part of the RDB ADT,
already defined and used. It can already generate a description of the
database, given the right ontology.

The Direct Mapping is intended to denote the information provided by
RDB, by mapping its data structure (not the data model) to RDF. That
explains why:
* NULLs don't produce triples;
* values which are foreign keys produce triples with an URI referencing
another row, instead of the value for the field.

Denoting the absence of value is a *lot harder* than it seems when it
comes to denotational semantics techniques... So please come back to the
initial goal of the Direct Mapping and don't add new requirements that
can invalidate all the choices that were made so far.

More explicitly: don't produce triples for NULL values.

Alexandre.


> 
> Cheers,
> 
> Marcelo
> 
> 
> 
> 
> > Alexandre.
> >
> >
> >
> >
> >>
> >> > However, the direct mapping as it is in its current version does not consider the schema at all.
> >>
> >> Correct.
> >>
> >> > It would be information preserving as-is, if we were to also translate NULLs.
> >>
> >> And this is wrong. For the direct mapping to be information preserving, we'd have to be able to reconstruct the schema of an EMPTY TABLE after the table is translated to RDF via the direct mapping. But an empty table produces NO TRIPLES, and from no triples you cannot reconstruct the original relational table!
> >>
> >> > My proposal would be to extend the direct mapping to consider the schema and translate it to RDFS/OWL. But I would like to know what other think.
> >>
> >> But can you capture all of the semantics of the SQL model? PKs, FKs, data types, nullability,
> >> multiset semantics and so on? Or are you suggesting to do just the minimal RDFS domain/range thing?
> >>
> >> Best,
> >> Richard
> >>
> >>
> >>
> >> >
> >> >
> >> > Best,
> >> > Richard
> >> >
> >> >
> >> > On 17 May 2011, at 19:01, Juan Sequeda wrote:
> >> >
> >> > > Group,
> >> > >
> >> > > By information preserving, I mean that given the RDF data, I can reconstruct the relational table with all its values. Informally, given an identity SQL query (a query that outputs the whole table: SELECT * FROM table), there exist a SPARQL query which is executed on the RDF data and will return the same results of the identity SQL query.
> >> > >
> >> > > There are two cases for information preserving
> >> > >
> >> > > 1) We have knowledge the schema
> >> > >
> >> > > If the relational schema is directly mapped to RDFS/OWL, then we DO NOT need to translate nulls in order to preserve information. For example, consider the table R with attributes A and B and instances:
> >> > >
> >> > > R(Bob, NULL)
> >> > > R(Alice, 25)
> >> > >
> >> > >
> >> > > The ontology from this schema is
> >> > >
> >> > > <R> <type> <class>
> >> > > <A> <type> <property>
> >> > > <A> <domain> <R>
> >> > > <A> <range> <whatever datatype>
> >> > > <B> <type> <property>
> >> > > <B> <domain> <R>
> >> > > <B> <range> <whatever datatype>
> >> > >
> >> > > And the RDF data, without translating nulls:
> >> > >
> >> > > <row1> <R#A> "Bob"
> >> > > <row2> <R#A> "Alice"
> >> > > <row2> <R#B> "25"
> >> > >
> >> > > The identity SQL query is
> >> > >
> >> > > SELECT A, B FROM R
> >> > >
> >> > > Given that we know the schema, we can construct a SPARQL query:
> >> > >
> >> > > SELECT ?a ?b
> >> > > WHERE{
> >> > > ?x <R#A> ?a
> >> > > OPTIONAL{
> >> > >  ?x <R#B> ?B
> >> > > }
> >> > > }
> >> > >
> >> > > There we go... with that SPARQL query, we can reconstruct the the original relational table. No need of nulls. If we did triples for NULL values, then the SPARQL query wouldn't have OPTIONALS. The issue here is that we don't need triples for NULL values.
> >> > >
> >> > > 2) We don't have knowledge of the schema
> >> > >
> >> > > If we do not have knowledge of the schema, then we can't create a SPARQL query like the previous example. Just imagine that you can only look at the RDF data. For example, consider the following RDF:
> >> > >
> >> > > <row1> <R#A> "Bob"
> >> > > <row2> <R#A> "Alice"
> >> > > <row2> <R#B> "25"
> >> > >
> >> > >
> >> > > Given that one of the row 2 has <R#B> and row 1 doesn't, I could guess that the value of row 1 for attribute B is null. But what if the original table has a column C and every single row has a NULL value for that column. In this case, it would be necessary to explicitly translate NULL values into an RDF triple. Otherwise, then the mapping would not be information preserving.
> >> > >
> >> > >
> >> > > CONCLUSION:
> >> > >
> >> > > - At this moment, neither the Direct Mapping or R2RML consider the schema, therefore in order for the mappings to be Information Preserving we must explicitly translate NULL values to an RDF triple.
> >> > > - We need to figure out how is this triple going to show up?
> >> > > - From a theoretical side, if we do not generate triples for NULL values, them mapping monotonic. On the other hand, generating triples for NULL values will make the mapping non-monotonic. Do we care? Not really. But implementation and performance-wise, there can be some overhead when dealing with non-monotonicity
> >> > >
> >> > >
> >> > > Juan Sequeda
> >> > > +1-575-SEQ-UEDA
> >> > > www.juansequeda.com
> >> >
> >> >
> >>
> >>
> >>
> >
> >
> >
> >
>

Received on Wednesday, 18 May 2011 12:46:09 UTC