Modeling how suffixes are learned in infancy

Recent experimental work offers evidence that infants become aware of suffixes at a remarkably early age, as early as 6 months for the English suffix -s. Here, we seek to understand this ability though the strategy of computational modeling. We evaluate a set of distributional learning models for their ability to mimic the observed acquisition order for various suffixes when trained on a corpus of child-directed speech. Our best-performing model first segments utterances of the corpus into candidate words, thus populating a proto-lexicon. It then searches the proto-lexicon to discover affixes, making use of two distributional heuristics that we call Terminus Frequency and Parse Reliability. With suitable parameter settings, this model is able to mimic the order of acquisition of several suffixes, as established in experimental work. In contrast, models that attempt to spot affixes within utterances, without reference to words, consistently fail. Specifically, they fail to match acquisition order, and they extract implausible pseudo-affixes from single words of high token frequency, as in [pi-] from peekaboo. Our modeling results thus suggest that affix learning proceeds hierarchically, with word discovery providing the essential basis for affix discovery.