I think the proper way to control this sort of operation -- and keep it from screwing up other entity references (like è and <, etc) is something like this:
s/&(?!\w+;)/&/;
This is just assuming that every valid entity reference that might exist in the original text is limited to alphanumerics and underscores between the initial ampersand and the final semi-colon, which is probably a safe-enough assumption.
But keep a backup of the original. If the data still causes parse errors after this simple edit, they might be different problems you haven't fixed yet, or they might be problems created by this simple edit. Careful diagnosis would be needed in that case.
