I contemplated the same problem and concluded the thing to try first would be probabilistic clustering of snippets, but I haven’t actually tried it yet.