Revamped lemma matching code to make it much more efficient. It has now been split into 3 XSLTs, chained together in a build process. The lemma match function has also been rewritten so that it reports errors more accurately.