Statistical Learning

Statistical learning (SL) refers to the ability of the human mind to extract patterns from complex input based on regularities. The concept originally derived from studies in infant language learning. Prior to the 1990s, there was a widespread belief among linguists that only inborn knowledge could explain how infants learn language so readily. A major tenet of these nativist views of language development is the impossibility of infants having any cognitive endowment that would be equal to the task of discovering language rules. In this view, the language heard by the child is impoverished because it lacks key information needed to unlock the rules of grammar. There are two important questions raised by the nativist view. First, what is the nature of the information in the input? That is, is it truly impoverished or is there a way for children to over time extract the rules and make [Page 1834]correct conclusions about grammar? The second question is what capacities does the mind of the infant have to store and analyze the information in the input? If children lack the capacity to store and process the vast amounts of data, clearly, it would be impossible to work out rules on their own without innate structures to guide their learning. Answers to these two questions became much clearer when Jenny Saffran, Robert Aslin, and Elissa Newport published their groundbreaking paper demonstrating the human infant’s ability to notice and recall patterns in complex linguistic input. This entry discusses the basic concept of SL, how it was established experimentally, differences between child and adult SL abilities, and the implications of SL for theories of language development.

The debate over information in the input continues but the novel experiments in SL in infants carried out by Saffran and colleagues clearly demonstrated that children have enormous capabilities, far more than anyone previously suspected. In this work, babies were exposed to strings of nonsense syllables that had predictable phonological patterns, constructed specifically for the purpose of finding out when and how infants might be able to detect patterns. Based on very brief exposure, infants were able to detect the patterns. They were able to recall these patterns when tested a short while later. This SL has been demonstrated repeatedly in infants and adults. To grasp what is statistical about this type of learning, it is important to note that the strings of sounds played to the infants were deliberately constructed so that the only information in the input was the artificially created probabilities of sound patterns occurring in certain sequences. No other information that normally would be present in language, such as pausing, pitch changes, or linguistic context, was available in these miniature languages.

Later research has shown evidence that SL is present in other modalities, including nonlinguistic auditory input, recognizing faces, and learning other visual patterns. These findings support the view of SL as a domain general process. That is, it is not specific to language but part of the human learning ability across many domains. Linguists working in the Chomskian generative tradition have argued against domain general learning abilities being adequate for language learning, so this aspect of the SL literature also is at odds with strict nativism.

...