Less is More in Bayesian Word Segmentation: When cognitively plausible learners outperform the ideal
Lawrence Phillips, Lisa Pearl · eScholarship (California Digital Library) · 2012
Purely statistical models have accounted for infants' early ability to segment words out of fluent speech, with Bayesian models performing best (Goldwater et al. 2009).Yet these models often incorporate unlikely assumptions, such as infants having unlimited processing and memory resources and knowing the full inventory of phonemes in their native language.Following Pearl, et al. (2011), we explore the impact of these assumptions on Bayesian learners by utilizing syllables as the basic unit of representation.We find a significant -Less is More‖ effect (Pearl et al 2011;Newport 1990) where memory and processing constraints appear to help, rather than hinder, performance.Further, this effect is more robust than earlier results and we suggest this is due a relaxing of the assumption of phonemic knowledge, demonstrating the importance of basic assumptions such as unit of representation.We argue that more cognitively plausible assumptions improve our understanding of language acquisition.