When native and non-native English speakers inflect novel verb forms for the past tense, non-natives are more likely toproduce irregular (non -ed) forms than natives (Cuskley et al., 2015). We test whether participants can reverse-engineerthe correct present tense stem from regular and irregular past tense forms of novel verbs. All participants are better able toidentify the stem of regularly inflected forms than irregular forms, but we find no difference between native and non-nativespeakers. Phonological similarity to existing irregulars interferes with recognition of regularly inflected non-verbs (e.g.,proximity of sleened to sling/slung makes it more difficult than drocked). While non-natives are more likely to produceirregular past tense forms, they are not better than native speakers at interpreting them. Non-native over-production ofirregulars may reflect statistical patterns in their more limited input, but these factors do not seem to affect the process ofinferring stems.

A recent large-scale wug-task study found that non-nativespeakers of English tend to produce fewer regular past-tense-ed inflections than native speakers (Cuskley et al., 2015). Inthis paper we present a model that can account for this dif-ference in behaviour as resulting from a difference in inputamounts and distributions. This model attends to both fre-quency, using Bayesian non-parametric methods, and phono-logical similarity between words, using a neural model of wordforms, and unifies these factors within a single probabilisticframework. We show that the general pattern of over-use ofirregular inflections in non-native speakers can result simplyfrom exposure to a smaller amount of input and does not re-quire any model-internal distinction of native and non-nativespeakers. Our model also captures the interaction betweenclass frequency and phonological similarity that was evidentacross all participant productions.