Error-correction learning of second language verbal morphology: associating imperfect contingencies in naturalistic frequency distributions

We investigate what is learned from exposure to usage in verbal morphology using an error correction mechanism within an associative learning framework. We computationally simulated how L2 learners would respond to naturalistic input of aspectual usage, characterised by „imperfect contingencies”, given two types of instructions: grammatical vs lexical. To test these predictions, English L1 speakers (N=80) completed three online training sessions in two conditions (grammatical vs lexical) over three days, learning 21 Polish verbs across 189 exposures; and a 63-item post-test on day 4 (50% seen, 50% grammatical). The results confirmed the simulation predictions: the grammatical group performed better through stronger performance in contexts that allow only one aspect while the lexical group was slightly better in contexts where both aspects were possible. Rules offer some advantage early on, especially when the exemplars are already unambiguous, whereas an exemplar-based approach promises a more flexible system in the longer run.