ABSTRACT
This paper takes as its starting point a theory that views the learner of a phonology or syntax as actively constructing rules and grammars by forming hypotheses and abstract rules over available data. The evidence for this theory is only briefly given here. Rather the focus in on the critical questions, under such a theory, of (i) why the learner succeeds so quickly, given the standard Chomskyan assumption that any grammar is radically underdetermined by the data, and on the related question of (ii) how the learner unlearns overly-general or incorrect rules. Let's call this the learning riddle and acknowledge at the outset that it is deeply connected to the theory's hypothesis testing mechanism, its abstract representational structure, a powerful and expressive formalism for that representation, and the assumption of an extensionally complex and variable world.
