Proceedings of the Annual Meeting of the Cognitive Science Society

A key challenge in language acquisition is learning morpho-logical transforms relating word roots to derived forms. Tra-ditional unsupervised algorithms find morphological patternsin sequences of phonemes, but struggle to distinguish validsegmentations from spurious ones because they ignore mean-ing. For example, a system that correctly discovers ”add /z/”as a valid morphological transform (song-songs, year-years)might incorrectly infer that ”add /ah.t/” is also valid (mark-market, spear-spirit). We propose that learners could avoidthese errors with a simple semantic assumption: morpholog-ical transforms approximately preserve meaning. We extendan algorithm from Chan and Yang (2008) by integrating prox-imity in vector-space word embeddings as a criterion for validtransforms. On a corpus of child-directed speech, we achieveboth higher accuracy and broader coverage than the purelyphonemic approach, even in more developmentally plausiblelearning paradigms. Finally, we consider a deeper semanticassumption that could guide the acquisition of more abstract,human-like morphological understanding.