Can input statistics over-ride a prior bias in morpheme ordering? A test case with gender and number

In languages which mark both gender and number as distinct morphemes, there is a tendency to place gender closer to the noun stem than number. However, the typological data on this is sparse. Moreover, linguistic theories differ in how they explain ordering patterns of gender and number morphology: some theories focus on the structure of the representations of features in the speakers' minds, and other focus on the role of co-occurrence statistics. In a recent study, Saldana, Kanampiu, and Culbertson (2025) use artificial language learning to show that learners with a diverse range of language experience with grammatical gender and number exhibit a consistent bias for orders with gender closer to the noun stem than number. This order reflects the ordering in which most linguistic theories assume number and gender features are derived in word formation. Here, we build on this study to investigate how this bias interacts with the statistics of the linguistic input. In particular, we manipulate co-occurrence between stems and affixes so that learners are exposed to combinations of stems and number morphology more often than to stems and gender morphology. We test whether input statistics can push learners to reverse their natural preference, leading them to place number closer to the noun than gender. We find that our manipulation reduces, but does not eliminate or reverse, the preference for gender-closest order. However, our study also highlights some difficulties learners have in acquiring novel features from sparse data. Ultimately, our findings highlight the dynamic interplay between representations of meaning and input-based learning mechanisms.