It seems to me that, at least for e-bidness, a very acceptable mechanism
would be to create 'canonical' representations for all the major data types
used in business. Since the documents are likely to be generated
mechanically and processed mechanically, translation from the user document
to the canonical representation for transmission, and the redigesting on
the receiving end, would be a none trivial amount of code but not horrible
relative to the benefits gained.
So, choose "[+/-]x.y" for floating point and say that's the canonical
format, period. If you want to be interchangeable for e-bidness, you use
the canonical format in your serialization of the data.
I know that doesn't do anything for the more document oriented folks, who
have a much harder row to hoe. However, dealing with the issue for e-biz
would be a big step and wouldn't be horrendous to do. It would be a matter
of buying a DLL/Java zip file for your locale that knows how to do the
translation to/from the canonical format.
There would need to be a little glue that knows how to take type
representation names from your particular Schema, and convert them to the
name of the canonical type so that the in/out processing knows how to deal
with your type names. But that wouldn't be a huge amount of work, and could
be packaged with a Schema implementation once the canonical format names
were known.