No. DTDs and XSD are designed to be LL(k) (presumably LL(1) or have I fogotten it) so no backtracking is required for element content models. (For keyrefs and assertions, the grammar characterisation really does not fit. No-one would attempt to implement XPaths as a grammar, anyway, I think.) A simple recursive descent parser is all that is needed.
For DTDs, the ambiguity rule prevents the need for backtracking. For XSD, the Unique Particle Attribution rule does similar. If you grammar ever presents you with a choice of
But RELAX NG schemas are not necessarily ll(k) so they may indeed need backtracking. However, an alternative technique called Partial Evaluation (which continuously rewrites the grammar as you go) is what everyone uses AFAIK.
For XML attribute validation, and for XSD ALL content groupings, I think implementers choose some evaluation methods that go outside ll(k) to prevent combinatorial explosion: but this is an optimization technique rather than a necessary class of grammar.
There has been a lot of developments in grammar theory and parser implementation in the last 20 years. For example, PEG grammars are a class that is quite convenient for expressing XML with less lexical handwaving.
Rick
