Does a Neural Network Inflect Spanish Verbs in a Human-Like Way?

How do children learn that a verb like negar-‘to deny’ / niego - ‘I deny’ diphthongizes the 1st person present indicative, while it’s near minimal pair, pegar-‘to stick’ / pego - ‘I stick’, does not? To pursue this question, three small experiments were conducted in Spanish, given its complex and rich morphology. The first experiment was conducted to determine whether the Tolerance Principle of the Rules & Competition model (Yang, 2018) can accurately determine the productivity of verb classes in Spanish, and whether any irregular verb classes are exhibit subregular productivity. The second experiment was conducted to determine how native Spanish speaking adults apply the rules of productivity to nonce words; and, by extension, whether they attend to the same rules of productivity that children do. And finally, the third experiment was used to determine whether a neural network makes human-like errors that resemble those made by human subjects in the second experiment. The results of the first experiment reveal that the Rules & Competition model accurately predicts that the regular inflection in the 1st person present singular indicative is productive, while e-i metaphony exhibits subregular productivity. The results of the second show that human subjects only apply regular inflections to nonce words, to the total exclusion of irregular inflections. Finally, the results of the third experiment show that the neural network, unlike human subjects, do apply irregular inflections to unseen words, showing a preference for e-i metaphony over the regular inflection, in particular. Ultimately, this suggests that the neural network’s domain general learning mechanism is not human-like, that morphological errors and intuitions are not generalizable based on local segmental environments, and that future exploration of this issue should be consider the extent to which the principles governing these learning mechanisms are abstract in nature