The extraction and integration framework: A two-process account of statistical learning.

Abstract

The term statistical learning in infancy research originally referred to sensitivity to transitional probabilities. Subsequent research has demonstrated that statistical learning contributes to infant development in a wide array of domains. The range of statistical learning phenomena necessitates a broader view of the processes underlying statistical learning. Learners are sensitive to a much wider range of statistical information than the conditional relations indexed by transitional probabilities, including distributional and cue-based statistics. We propose a novel framework that unifies learning about all of these kinds of statistical structure. From our perspective, learning about conditional relations outputs discrete representations (such as words). Integration across these discrete representations yields sensitivity to cues and distributional information. To achieve sensitivity to all of these kinds of statistical structure, our framework combines processes that extract segments of the input with processes that compare across these extracted items. In this framework, the items extracted from the input serve as exemplars in long-term memory. The similarity structure of those exemplars in long-term memory leads to the discovery of cues and categorical structure, which guides subsequent extraction. The extraction and integration framework provides a way to explain sensitivity to both conditional statistical structure (such as transitional probabilities) and distributional statistical structure (such as item frequency and variability), and also a framework for thinking about how these different aspects of statistical learning influence each other.Keywords: statistical learning, language development, implicit learning, word learning Humans live in a world filled with statistical regularities. Balls thrown into the air typically fall back to earth; nouns such as dog or boy are typically preceded by articles such as a or the. There is no doubt that learners are sensitive to these statistical regularities. One term to describe the ability to detect and use statistical structure is statistical learning. Saffran, Aslin, and Newport (1996) proposed this term to describe infants' ability to identify word boundaries solely from the statistical relation between sounds in the input. It is now widely acknowledged that infants and adults encode the statistical structure of their environment in a variety of tasks, including sequence learning (e.g., Haith, Wentworth, & Canfield, 1993;Stadler, 1992), category boundary detection (e.g., Maye, Werker, & Gerken, 2002), word-object association (Smith & Yu, 2008), cue-category association (Thiessen & Saffran, 2007), and causal learning (Sobel & Kirkham, 2007). Statistical learning likely plays a role in many different aspects of development, but it is thought to play an especially crucial role in language development. The discovery that infants are capable of benefiting from statistical structure in the input led to a reevaluation of the role of learning in language acquisition, ...

Individual Differences in Statistical Learning: Conceptual and Measurement Issues

“…We focus on word segmentation tasks for several reasons: there are several different "languages" available for use (so that participants can be trained on multiple stimulus sets, each with a different set of training items), these "languages" and tasks are all quite similar in nature (short training sets, followed by an assessment), and these tasks all tapped into the same dimension of SL (which we describe as "chunking" or "extraction"; [9]), yet are distinct enough to minimize potential confusion or interference. Because these languages have been used previously, their selection allows for continuity with and comparison to prior research.…”

“…This uncertainty can be seen at both a theoretical and a methodological level. On a theoretical level, different proposals carve up SL in different ways: some suggest that it is a unitary process (e.g., [40]), whereas others suggest it is an umbrella term for multiple processes, but differ on the number and nature of these underlying processes (e.g., [9,34]). On a methodological level, different measures of SL are quite different, and not necessarily correlated (e.g., [1,41,42]).…”

“…Within the domain of language, the acquisition of probabilistic structure has often been termed statistical learning (SL) 1 . SL has been demonstrated in learning and use of phonemic categories [13], lexical forms [14,15], phonotactic and phonological regularities characterizing words [16,17,18], word meaning [19], and syntactic structure [20,21,22,23,24].…”

“…From this perspective, a straightforward prediction is that individual differences in SL ability should be related to variation in language outcomes, with the specific expectation that individuals who are better at SL should show better language learning and processing (e.g., [1,9,25,26]). Although this prediction flows naturally from a number of extant psycholinguistic theories (e.g., [10,12]), it has received comparatively little attention in the literature.…”

“…This adaptability plays a role in many domains of formal and informal learning, including language (e.g., [3]), mathematics [4], locomotion (e.g., [5]), and perception (e.g., [6]). Recent linguistic theories suggest that adaptation to probabilistic information characterizes both language acquisition in infancy [7,8,9], and language comprehension and production in adulthood (e.g., [10,11,12]). Within the domain of language, the acquisition of probabilistic structure has often been termed statistical learning (SL) 1 .…”