I like the idea of comparing each pair of sentences for similarity. There are several metrics for sentence similarity that come to mind:
in reply to Brainstorming session: detecting plagiarism
Edit distance & longest-common subsequence. These two are pretty similar: look for an edit distance less than a certain percent of the sentence length, or an LCS larger than a certain percent.
As you are doing now, I would do this on a word-by-word basis, and not character-by-character. However, these algorithms can be generalized a bit, so that instead of each pair of words either agreeing or disagreeing, each pair can have a fractional level of agreement between 0 and 1. If you implemented a "synonym measurer", you could easily plug this into such generalized LCS/Levenshtein algorithms. (This could also quite easily encompass changes in word stemming as well as synonimity.)
Another metric you can use for sentence similarity is Zaxo's favorite method: using compression & information theory, although you may not be able to pull out as much information about *how* the sentences are similar as in the algorithms above.