Efficient Learning of Regular Expressions from Good Examples

Abstract

We consider the problem of restoring regular expressions from expressive examples. We define the class of unambiguous regular expressions, the notion of the union number of an expression showing how many union operations can occur directly under any single iteration, and the notion of an expressive example. We present a polynomial time algorithm which tries to restore an unambiguous regular expression from one expressive example. We prove that if the union number of the expression is 0 or 1 and the example is long enough, then the algorithm correctly restores the original expression from one good example. The proof relies on original investigations in theory of covering symbol sequences (words) by different sets of generators. The algorithm has been implemented and we also report computer experiments which show that the proposed method is quite practical.