The role of generalisation in an Adaptive Resonance Theory model of learning inflection classes

Language users can generalise encountered linguistic rules and apply them in new contexts, for example to conjugate unseen words. Inflection classes, groups of words that are inflected in the same way, help language users to deduce unseen word forms based on the patterns characteristic to the class. To simulate the evolution of inflection classes, one needs a component for their acquisition on the individual level. We model this individual learning task by using an unsupervised Adaptive Resonance Theory 1 (ART1) model, whose level of generalisation can be adjusted with a single parameter. We find a range of generalisation values for which ART1 is able to incrementally learn inflection classes for the Latin present tense. Analysis of the clusters shows a good match between the learned categories and the attested inflection classes. This method could eventually be used as a component in a diachronic model, studying the evolution of inflection classes.