Ewan Dunbar

The Acquisition of Morphophonology Under a Derivational Theory: A Basic Framework and Simulation Results Ewan Dunbar Department of Linguistics University of Toronto 2008 Since Rumelhart and McClelland 1986, the alternations found in the simple morphophonology of the English past tense have been run through numerous learning algorithms. These algorithms generally perform well at learning these very easy alternations, but fail to reflect a basic assumption about how grammar works held by many morphologists, namely, that affixation has a special status, different from other kinds of morphological changes. They also often do not address the question of how the learner restricts alternations like lie–lay to a single sense of lie (“lie down,” and not “prevaricate”) while still learning general patterns that apply to novel words regardless of their meaning. More generally, the morphophonological mappings deduced by previous morphological learning systems have never much resembled the kinds of grammars proposed by morphologists. In this paper I remedy this situation by constructing a system that deduces two kinds of rules—internal changes and suffixes—and applies them to words derivationally; it automatically finds rules that are not generally applicable and marks them as only applying to an arbitrary class of words. The model is largely compatible with developmental and psycholinguistic evidence; to the extent to which it is not compatible with this evidence, it raises interesting questions about how human morphophonological generalizations are best characterised.