Statistical learning refers to a wide variety of phenomena, many of which have been argued to be related to language acquisition. However, there is little agreement on the process that underlies statistical learning. Several different accounts have been proposed, but it has been difficult to differentiate between these accounts as many converge on the same end result of learning. To identify the process responsible for statistical learning, it is necessary to more closely examine behavioral data that can characterize the dynamic characteristics of learning over the course of exposure. The objective of the current project is to develop and apply a novel method for examining statistical learning of linguistic materials. This method will provide more comprehensive and sensitive results than prior methods, which will enable these experiments to distinguish between theories of statistical learning in a way that has not been previously possible. In these experiments, participants will be exposed to a stream of syllables made up for nonsense words. Within this stream, words consistently co-occur, while syllable conjunctions formed across word boundaries are less predictable. Participants will be asked to listen for a particular syllable within the speech stream, and respond with a button press when they hear it. For some of these participants, the syllable will occur in an unpredictable location (for example, go in golabu is relatively unpredictable, because it can occur after the end of any word in the speech stream). For other participants, the syllable will occur in a predictable location (for example, bu in golabu is consistently signaled by the presence of both go and la). The experiments outlined in this proposal are a first step towards a process-based, mechanistic account of statistical learning. The first experiment demonstrates that this novel methodology is feasible, and will assess the extent to which the serial reaction time measure correlates with more standard post-test measures. The second and third experiments seek to test process-level predictions of a theory of statistical learning. Experiment 2 assesses the extent to which working memory is related to performance in the task, especially on different word lengths. Experiment 3 assesses a proposal about how chunking might be supplemented by processes of comparison to make learning of non-adjacent regularities possible. Finally, Experiment 4 asks how multiple cues to segmentation are integrated while learning is occurring in real time. By identifying the dynamic characteristics of learning over the course of exposure to the input, this research will test and refine theories of statistical learning in ways that have not previously been possible.
This research proposal will provide insight into the mechanisms underlying language learning. Understanding these mechanisms is critically important for designing interventions that improve language acquisition in both atypically developing children, and adults struggling to acquire a second language.
