Morpheme learning in the noisy landscape of natural text

Most words in English and other languages are formed by combining smaller meaningful units called morphemes. Understanding morphology allows us to generalise from known elements; for instance, we can interpret “quickify” because we know how the affix “-ify” changes the meanings of stems. This talk examines how we learn affix morphemes through reading experience. Because these units rarely appear on their own, their functions must be inferred from exposure to whole words. Theories propose that we learn affixes because they provide reliable information about meaning, but evidence has come largely from small laboratory studies or simulations. I present findings from corpus linguistics, computational modelling, and behavioural research investigating how affixes are learned from large-scale text input. The results show that affixes vary substantially in the ease with which they can be identified as combinatorial units in different words and the consistency with which they signal meaning. Results further show that these properties influence learning in both models and humans, such that acquired morpheme knowledge becomes a mirror of morpheme distribution defined orthographically in large-scale text. Overall, this work advances understanding of how statistical learning operates in naturalistic, noisy environments over the extended timescale of reading acquisition.