Re: deterministic naming of blank nodes

On 09/26/2014 06:20 PM, Tim Berners-Lee wrote:
> On 2014-09 -24, at 00:37, Sandro Hawke <
> <>> wrote:
>> On 09/23/2014 05:34 PM, David Booth wrote:
>>> BTW, I want to draw attention to the fact that the need for defining
>>> an RDF-specific PATCH operation is *entirely* a consequence of RDF's
>>> allowance of unrestricted blank nodes.  I do not think that blank
>>> nodes should be eliminated from RDF, but I am convinced that RDF's
>>> current treatment of blank nodes is a significant design flaw that
>>> has *many* downstream effects that are ultimately detrimental to
>>> RDF's adoption. The need for RDF PATCH is another example.
>>> Unix/linux diff and patch utilities have been used successfully for
>>> *decades*, with many other information representations. Imagine how
>>> simple and easy it would be if we could just generate canonical
>>> N-Triples and use standard diff and patch against that! But we can't,
>>> because blank nodes are unstable across RDF serializations and no
>>> canonical way to generate them has been standardized.  This, in turn
>>> is because generating a canonical form of unrestricted RDF is a hard
>>> problem (NP-complete), because of blank nodes.  The problem is *much*
>>> easier if the use of blank nodes is limited to *implicit* blank nodes
>>> -- those that are generated implicitly by the use of square brackets
>>> "[]" or parentheses "()" for lists in Turtle -- and indeed this is
>>> the vast majority of blank node use.  (See "Everything You Always
>>> Wanted to Know About Blank Nodes", by Hogan, Arenas, Mallea and
>>> Polleres:
>>> )
>>> For this reason the use of "Well Behaved RDF" was proposed, which
>>> limits the use of blank nodes to implicit blank nodes:
>>> I don't know if Well Behaved RDF is the best solution to this
>>> problem. Maybe someone will come along with a better idea.  But I am
>>> convinced that the current treatment of blank nodes in RDF is a
>>> serious problem that we should fix in order to make RDF simpler to
>>> use, understand and adopt.
>>> I really don't like having to make excuses for RDF when it cannot be
>>> used in a similar way as nearly every other information
>>> representation -- such as being able to easily compare two RDF
>>> documents for "equality" (which in RDF becomes a complex graph
>>> isomorphism problem) or generate a simple diff and patch -- all
>>> because of RDF's unrestricted treatment of blank nodes.
>>> Clearly this is not something that the Linked Data Platform working
>>> group can fix.  But I think it is important to bring it to people's
>>> attention, in the hope that we will someday soon have the creativity
>>> and gumption to fix it.
>>> I should also acknowledge that there are some who do not feel that
>>> RDF's treatment of blank nodes is a problem.  Fine.  It may not be a
>>> problem to an elite few who are well steeped in the subtleties of
>>> description logic, model theory and RDF Semantics, and who don't mind
>>> having to use RDF-specific tools instead of generic tools.  But
>>> having tried for over 10 years to explain RDF to a wider audience of
>>> regular software developers, I am convinced that subtleties like
>>> RDF's treatment of blank nodes *are* a problem to a much wider
>>> audience of *potential* RDF users who would be more inclined to adopt
>>> RDF if it didn't have complexities like this. As it is they are more
>>> likely to stick with JSON or XML, whose complexities they already
>>> know, rather than venturing into the obscure and esoteric world of RDF.
>>> RDF tools are not as mature as those for XML or even JSON, which is
>>> much younger than RDF.  I believe blank nodes are one specific reason
>>> they're not.  The fact that we still don't even have a simple,
>>> standard way to compare RDF documents and compute diffs and patches,
>>> is a perfect example.
>>> David
>> I agree that it makes sense to have good terminology for graphs that
>> can be serialized in Turtle without blank node labels, and perhaps to
>> focus diff on these nice graphs.    (When I clicked on your
>> Well-Behaved RDF link, my PDF viewer remembered I was on page six. :-)   )
>> How would you name the blank nodes when serializing this in n-triples
>> or n-quads?   I guess given triple S P O, where O is a blank node,
>> you'd name O based on the hash of the names of S and P?   Something like
>>     o.asNTriplesTerm = "_:a"+sha1hex(s.asNTriplesTerm+"
>>     "+p.asNTriplesTerm)?
>> Or maybe leave off the NTriples baggage, the "<"...">" and "_:".    Maybe:
>> = sha1hex(" "
>> You'd have to evaluate these in the right order, but that would be the
>> order you parsed them from Turtle, so it should be fast and
>> deterministic.   Hm.   Can we avoid the repeated hashing?   I suppose
>> so, with some kind of path expression:
>>     Given:  <s1> <p1> [ <p2> ( 1 2 3 [ <p3> ( 11 12 [ <p4> 444 ] ) ] ) ]
>>     The subject of <p4> 444 would have its id determined by hashing
>>     (s1, p1, p2, 4, p3, 3).
>> (I'm treating integer list positions as if they were properties.) With
>> a little cleverness, caching those intermediate hashes, I think that
>> could run almost as fast as computing a hash of the turtle file (ie
>> very fast).
>> Wow, this might work.    Very good idea, David.    Beyond the fact
>> that it only works on this kind of graph, does anyone see any problem
>> with it?
> Hmmm , suppose a blank node can be identified by more than one path,
> then the hashed path would not be unique and so you couldn't use them to
> compare blank nodes

If the algorithm deterministically chose which path to use in generating 
the hash -- say, the first alphabetically -- then it would always 
generate the same hashes for the same (isomorphic) graphs.  Even if the 
graph changed slightly, such an algorithm could often choose the same 
path, so it would be fairly robust to minor changes in the graph, which 
is a very desirable property.  And in the cases where it did choose a 
different path, you would end up with two URIs for what should have been 
considered the same node, and owl:sameAs could be used to indicate their 
equivalence, once their equivalence is detected through 
application-specific means.


Received on Friday, 26 September 2014 22:48:57 UTC