Could always rely on the Levenshtein distance. You have to be careful with similarity approaches though as you may end up filtering important messages because they are structurally similar to the unimportant message.