The limits of statistical learning in word segmentation: Accumulation of predictive information from unstructured input in the absence of (declarative) memory

Learning statistical regularities from the environment is ubiquitous across do- mains and species. It might support the earliest stages of language acquisi- tion, especially identifying and learning words from fluent speech (i.e., word- segmentation). But how do the statistical learning mechanisms involved in word-segmentation interact with the memory mechanisms needed to actually remember words as well as with the learning situations where words actually need to be learned? We show that, in a memory recall task after exposure to continuous, statistically structured speech sequences, participants track the statistical structure of the speech sequences and are thus sensitive to probable syllable transitions, but hardly remember any items at all. Analysis of their productions suggests that they are unable to identify probable word bound- aries. As a result, they tend to produce low-probability items even while pre- ferring high-probability items in a recognition test. Only discrete familiariza- tion sequences with isolated words yield memories of actual items. Through computational modeling, we show that earlier results purportedly supporting memory-based theories of statistical learning can be reproduced by memory- less Hebbian learning mechanisms. Turning to how specific learning situ- ations affect statistical learning, we show that it predominantly operates in continuous speech sequences like those used in earlier experiments, but not in discrete chunk sequences likely encountered during language acquisition. Taken together, these results suggest that statistical learning might be special- ized to accumulate distributional information, but that it is dissociable from the (declarative) memory mechanisms needed to acquire words and does not allow learners to identify probably word boundaries.