Abstract Neural network models of morphological inflection (NNMIs) have seen impressive improvements over the last several decades. These improvements, however, lie in raw accuracy and other practical issues. When the performance of these models is instead compared to that of child learners, many of the shortcomings of the classic models persist in today’s NNMIs: (1) they struggle with generalization across sparse paradigms, (2) they are prone to over-irregularization, and (3) they do not follow child acquisition trajectories. The persistence of these issues suggests that they reflect inherent or “innate” characteristics of NNMIs as a class which are distinct from the characteristics of human language learners. As such, even substantial practical improvements do not necessarily entail increased cognitive plausibility.
Payne et al. (Thu,) studied this question.