Evaluating learning trajectories of neural morphology acquisition models

Computational models of morphology acquisition have played a central role in debates over the nature of morphological representations since the origin of the "past tense debate"in the 1980s. The apparent success of recent artificial neural network architectures for morphological inflection in natural language processing has revitalized this debate. However, despite their often good performance, the actual suitability of these advanced neural networks as models of human morphology acquisition remains uncertain. We argue that much of this confusion stems from inconsistent methods of training and evaluation. In this work, we demonstrate that more careful dataset creation and an evaluation combining quantitative analysis and comparison with human development puts the evaluation of neural models on firmer ground.