Relax NG specification oddity

[withtape]

Somebody please, please help me. I'm getting started using Relax NG to
describe document schemas. After reading the grammar summarized by
http://www.relaxng.org/spec-20011203.html (Section 5), I conclude that
the following is a valid Relax NG schema:

<grammar xmlns="http://relaxng.org/ns/structure/1.0">
<start>
<attribute name="wtf"/>
</start>
</grammar>

That is, the grammar defines no document element, and that document
element must have a "wtf" attribute.

Elsewhere in the specification (Appendix A) is a Relax NG schema for
Relax NG schemas, and validating the above XML against the schema for
Relax NG (using Tenuto) indeed declares it a valid Relax NG schema.
Yet, there is no possible XML document that could validate against it!

This doesn't present any problem when validating the eventual documents
that must match the "WTF" schema: no document will ever validate
against it. But (as I understand) it means that anyone writing Relax NG
schemas must take care above and beyond what is demanded by the Relax
NG specification to avoid creating such a "WTF" schema. Otherwise, he
confronts a paradox of having a schema vouch-safed (by the Relax NG
spec) as suitable for validating XML documents, that cannot possibly
validate any XML document.

Why do I care? Well, I am writing a tool that assists in the creation
of valid XML documents by "suggesting" appropriate elements and values
at various points in a document according to a given schema. Included
within will be a tool to assist creating schemas, based on the Relax NG
meta-schema. Am I to understand that, if I desire to avoid allowing the
creation of "impossible" schemas like above, it will be insufficient to
simply implement the Relax NG meta-schema? Must I include additional,
Relax NG-specific fixes to check for such things (I wish to avoid this
because it would not be standard)?

More generally, is this a flaw in Relax NG? Does it make sense for a
schema-definition language to allow schemas that cannot validate any
possible document? Was Relax NG deliberately designed this way, or is
this an oversight? Do other schema-definition languages allow this
situation?

 

[Joe Kesselman]

More generally, is this a flaw in Relax NG?

I don't know enough about RELAX NG to answer this, but I would suggest
contacting the folks responsible for its design and posing this question
directly to them. They're the ones who can say for sure whether it's
design or accident, and whether it will be declared acceptable or an
erratum.

http://www.relaxng.org/ points to a mailing list for RELAX NG, as well
as a long list of documents and tools. (I presume you've already looked
at that list, both to make sure your tool isn't redundant and to see
what they do when presented with this schema...)

Good luck. Let us know what you find out.

 

[withtape]

Thanks for the pointer! Actually, I had missed that in the sea of links
on that page. I would say that's exactly what I'm looking for, so I'll
check that out and followup if I get new info.

 

[George Bina]

Hi,

The Relax NG grammar for Relax NG does not check all the rules for
having a valid grammar. This applies also for XML Schema where if a
document is valid against the schema for schemas that does not mean it
is a valid XML Schema.

The grammar from our example is not valid, Jing reports an error
because the start element contains attribute.
See this also in the specification:
http://www.relaxng.org/spec-20011203.html#context-start
***
7.1.5. start element
The following paths are prohibited:

* start//attribute
[...]
***

A sample schema that will check only that the root element contains a
wtf attribute is:

<?xml version="1.0" encoding="UTF-8"?>
<grammar xmlns="http://relaxng.org/ns/structure/1.0" ns="">
<start>
<element>
<anyName/>
<attribute name="wtf"/>
<zeroOrMore>
<choice>
<attribute>
<anyName>
<except>
<name>wtf</name>
</except>
</anyName>
</attribute>
<text/>
<ref name="any"/>
</choice>
</zeroOrMore>
</element>
</start>
<define name="any">
<element>
<anyName/>
<zeroOrMore>
<choice>
<attribute>
<anyName/>
</attribute>
<text/>
<ref name="any"/>
</choice>
</zeroOrMore>
</element>
</define>
</grammar>

or in compact syntax

start =
element * {
attribute wtf { text },
(attribute * - wtf { text }
| text
| any)*
}
any =
element * {
(attribute * { text }
| text
| any)*
}

Best Regards,
George

 

[withtape]

George,
Thanks so much for your help! I confess to being overwhelmed by the
specification, so I did not consider those separate prohibitions.

To be clear, your example demonstrating how to ensure that, whatever
the root element of a document is, it must have a "wtf" attribute, is
good reference, but not applicable to my situation. I contrived the
example of a "headless" attribute to point out what I took to be the
deficiency of the grammar, but I thank you for the demonstration.

The main point I wanted to address is whether it is possible to devise
a Relax NG schema that, in itself, ensures valid schemas. I see now
that the grammar given in the spec, expressed as a schema, fails at
this task, for good reasons (for example, I came across a post on the
Relax NG mailing list that, I believe, explains how partial schemas
like my example could be referenced by other schemas, and thus the
grammar allows them). I wonder if that schema/grammar could be refined
so that a schema-producing program would have no external requirements
(such as, "start//attribute is invalid"). Perhaps by adding
restrictions on allowed elements after <start>, at the expense of
allowing external schema references (a trade I would be willing to make
for my purposes)?

Really, I just wanted reassurance that I wasn't crazy in thinking that
the Relax NG schema for Relax NG schemas does something other than what
I initially expected, and you have provided it very nicely.

-Jason Chang