My opinion is that the best way is to code all xml-files in utf-8, and output html too :>)
Some years ago I solved very similar job, so I used DTD's as dictionary. My convention was: If some file.xml have to be localized, it's locale-specific content must be realized by entities, defined in $locale/file.dtd . Consequently, I had en/file.dtd and cz/file.dtd and softlinks file.xml => ../file.xml in both locale specific directories. Finally, when I parsed cz/file.xml, I had czech xml, when I parsed en/file.xml, I had english xml. Just idea...