Schema Design Questions

John Cowan

Sure. externalRef can point at any pattern whatsoever. It's just a straight inclusion.

Eric van der Vlist

If you mean a pattern element, it's an error, but you can explicitly combine two name name classes through a "choice", for instance:

 <element>
  <choice>
   <name>foo</name>
   <name>bar</name>
  </choice>
  <empty/>
 </element>	    

John Cowan

Sure. Because attribute ordering doesn't matter in XML, RNG always treats attribute patterns in a group pattern (in this case, an implicit one) as if they were interleaved. Therefore, they can appear in any position when they are children of an element pattern.

John Cowan

Yes. The only time whitespace is not ignored is in data or value patterns where the datatype preserves it, which in practice is only in string and xsd:string.

James Clark

Right. But DTDs don't have real wildcards, so if a DTD author wants a real wildcard, the best they can do is approximate it with ANY. Most of the uses of ANY I've come across are cases of this, rather than people actually wanting the DTD semantics.

In any case, if you want exactly the DTD semantics, trang can give it to you. See the -i strict-any option: thaiopensource.com

John Cowan

Yes. The only time whitespace is not ignored is in data or value patterns where the datatype preserves it, which in practice is only in string and xsd:string.

Eric van der Vlist

If you include the schema for docbook, you need to either redefine the start pattern defined in the schema for docbook or combine its content with a new one. Since here you want to replace its definition, the only option left is to redefine it:

include "docbook.rnc" {
  start = document
}
include "mods3.rnc"
document = element document { meta, content }
meta = ModsSchema
content = { "article" | "book" | "chapter" }

BTW, this last pattern means that content is a value equal to "article", "book" or "chapter". This is illegal to include that after "meta" if it's an element and, in any case, probably not what you mean. What about:

content = { article | book | chapter }

I think that this is already possible.

Consider:

inc1.rnc
start = element foo{empty}

inc2.rnc
start = element bar{empty}

If you want to include inc1.rnc and keep only the definition of the start element in inc2.rnc, you can write:

include "inc1.rnc" {
	start |= notAllowed
}
include "inc2.rnc"

James Clark

They are allowed. Each file in a compact syntax schema is translated individually to XML syntax. Each file is translated the same way whether its the top level file, referenced in include or referenced in external. The compact syntax schema is correct if the requirements of Appendix A of the Compact Syntax spec are met and if the resulting XML syntax schema is correct.

James Clark

Right. This ensures that a user can always replace any literal character in the original source by an escape.

Right. There didn't seem any reason why it should be whitespace. This is covered by: relaxng.org

David Tolpin

Suppose a RELAX NG schema containing a definition for references:

<define name="reference">
  <attribute name="ref">
    <data type="IDREF"/>
  </attribute>
</define>

Now, I want to change this definition in a derived schema, such that a reference now refers either to an element within in the XML document or to an element within an other XML document. Thus, I enhance the definition with a further attribute transitivly referring to the other XML document. I would like to have the same attribute name for the reference to the element because the element is of the same kind as the original one.

<define name="reference">
  <choice>
    <group>
      <attribute name="instance">
	  <data type="IDREF"/>
      </attribute>
      <attribute name="ref">
	  <data type="NCName"/>
      </attribute>
    </group>
    <attribute name="ref">
      <data type="IDREF"/>
    </attribute>
  </choice>
</define>

But, jing returns an error message:

Is it not possible to define context-sensitive ID-types?

David answers:

these are two separate problems, I think. One is whether duplicate attributes are allowed -- they are not, as stated in 7.3.

The other is whether you need it; NCName and IDREF are the same type in Relax NG. Your grammar is equivalent to

<define name="reference">
  <optional>
    <attribute name="instance">
	<data type="IDREF"/>
    </attribute>
  </optional>
  <attribute name="ref">
    <data type="IDREF"/>
  </attribute>
</define>

or, since you've mentioned the compact syntax

reference=attribute instance {xsd:IDREF}?, attribute ref {xsd:IDREF}

Referential semantics is beyond the scope of Relax NG; IDREF are only checked syntacically, which is, in my opinion is good thing.

Eric van der Vlist

Also, be aware that there are several definitions of what an "ambiguous grammar" is and that this tool is only checking the definition described in its documentation. In other words, it's a tool for a specific range of applications that may not be meeting your requirements.

Kohsuke Kawaguchi

I currently have a solution :

ATTR_type  = attribute type {xsd:token {pattern =
"[Ss][Yy][Ss][Tt][Ee][Mm]"} |

  xsd:token {pattern =
"[Mm][Ee][Nn][Uu]"}}

Have you another solution?.

Define your own case insensitive datatype library. See thaiopensource.com

John Cowan

You may find my non-normative XLink schema at xml-dev archive useful.

=====cut here=====

# RELAX NG schema for W3C XLink

datatypes xsd = "http://www.w3.org/2001/XMLSchema-datatypes";
namespace xlink = "http://www.w3.org/1999/xlink";

href = attribute xlink:href {xsd:anyURI}
role = attribute xlink:role {xsd:anyURI}
arcrole = attribute xlink:arcrole {xsd:anyURI}
title.att = attribute xlink:title {text}
show = attribute xlink:show {"new"|"replace"|"embed"|"other"|"none"}
actuate = attribute xlink:actuate {"onLoad"|"onRequest"|"other"|"none"}
label = attribute xlink:label {xsd:NMTOKEN}
from = attribute xlink:from {xsd:NMTOKEN}
to = attribute xlink:to {xsd:NMTOKEN}

simple = element * {
		attribute xlink:type {"simple"}, anyAttr*,
		href?, role?, arcrole?, title.att?, show?, actuate?,
		(anyElem | text)*
		}

extended = element * {
		attribute xlink:type {"extended"}, anyAttr*,
		role?, title.att?,
		(title | resource | locator | arc | anyElem | text)*
		}

title = element * {
		attribute xlink:type {"title"}, anyAttr*,
		(anyElem | text)*
		}

resource = element * {
		attribute xlink:type {"resource"}, anyAttr*,
		role?, title.att?, label?,
		(anyElem | text)*
		}

locator = element * {
		attribute xlink:type {"locator"}, anyAttr*,
		href, role?, title.att?, label?,
		(title | anyElem | text)*
		}

arc = element * {
		attribute xlink:type {"arc"}, anyAttr*,
		arcrole?, title.att?, show?, actuate?, from?, to?,
		(title | anyElem | text)*
		}

start = element * { anyAttr*, (simple | extended | anyElem)* }

anyElem = element * {anyAttr*, (anyElem | text)*}

anyAttr = attribute * - xlink:*  {text}

John Cowan et al

You bumped up against one of the few restrictions of RELAX NG: you can't have an iteration (either + or *, they produce the same message) that contains a sequence (or interleave, same thing in this case) of attributes.

What you have written is tantamount to requesting an attribute sequence like attribute foo, attribute bar, attribute foo, attribute bar .... be allowed. This makes no sense, given that there can only be one attribute of a given name in a given element.

Fortunately, the problem is easily cured by removing the first reference to NamedElementAttribute.

You can't expect correct results when applying Trang to a schema that Jing rejects. The corrected version does convert to a plausible XSD schema.

Try to avoid creating sequences/interleaves that contain both attributes and data. They tend to be problematic.

Bob Foster goes into a deeper explanation

ConfigGroupEntry* where ConfigGroupEntry has as an alternative ConfigValue, a group that contains a group, NamedElementAttributes, that contains two attributes.

The "oneOrMore" reference is because * (zeroOrMore) is defined in RELAX NG as +? (optional oneOrMore).

NamedElementAttributes is obviously redundant as it is used, so a simple fix would be to take it out of ConfigValue.

Trang can't convert everything, esp. an invalid schema, and even if it did the conversion could not be faithful where there is a choice between an attribute and an element. XML Schema can't describe this.

Don't give up. Granted the error messages from jing can be pretty cryptic, but in this case you had all the info you needed, you just didn't understand it. OTOH, if you had spent the same several hours trying to write an XML Schema with no previous knowledge of that language, I wager you wouldn't be anywhere near as far along as you are. ;-}

First, a grammar isn't thoroughly checked until you try to validate with it, because some grammars are perfectly ok as fragments but not acceptable for validation. So exercise the grammar on simple test cases to make sure it is fully checked and gives you the result you wanted.

Second, compact syntax error messages from jing come in two flavors: "syntax error," for which the only cure is to stare at the location identified and consult the syntax until you see the error, and substantive error messages, like the one you describe, that usually refer to constraints described in RELAX NG (non-compact) oasis.org. In this case, you could have found the restriction by searching the restriction section for "attribute".

MURATA Makoto takes up the spec references

If you check the RELAX NG specification, you'll find, in section 7.1.2, a prohibition of the path:

Consider an illegal pattern

(element * {text}, attribute * {text})*

This means that the number of attributes and that of elements are equal.

Next, consider another illegal pattern

(attribute a:* {text}, attribute b:* {text})*

This means that the number of attributes in the namespace for the prefix a is the same as the number of attributes in the namespace for the prefix b.

To implement such patterns, we cannot use abstract machines having *finite* states, since we need one state for each natural number. Then, Bali and Miaou (which I should have finished long time ago) will become impossible.

Jing constructs patterns (which are "states") lazily. Thus, the above patterns can be implemented actually. However, there are some malicious patterns which cause (an earlier version of) Jing to explode, because so many patterns have to be created during evaluation.

MURATA Makoto

This error happens when (1) some attribute declaration has a wildcard matching absolutely any namespace and (2) you have an ID attribute somewhere else in your schema.

There are two solutions. One is to make the wildcard match all namespaces EXCEPT the namespace of the element declaration having that ID attribute. The other is to convert the ID attribute to a string attribute.

In your case, I would rewrite

A fundamendal reasons is as follows. In DTD, an application program can determine whether an attribute is ID by examining the attribute name and the parent element name. The DTD Compatibility specification of RELAX NG is intended to mimic DTDs and ensures this simplicity.

David Carlisle

Simplification is a transformation defined in the RELAX NG specification: relaxng.org The simple syntax (not to be confused with the compact syntax) is a subset of the full syntax which you normally author schemas in. The semantics of RELAX NG validation are entirely described in terms of the simple syntax.

Simply put, the simplification transformation flattens a schema into a more DTD like structure where the only remaining top level constructs are definitions containing a single element pattern. Additionally, there is a single start pattern. All other named patterns, includes, overrides, combination attributes and nested grammars go away. Also, inherited attributes are resolved so you don't have to look upwards to determine the value of for example a ns attribute.

So, if you want to process a schema and don't care how the originial schema was structured and modularized, the simplified version is just much simpler to use.

rng2srng is a tool which will perform the simplification transformation described above.