Poster D66, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Neural sensitivity to speech distributional information underlies statistical learning
Julie Schneider1, Yi-Lun Weng1, Violet Kozloff1, An Nguyen2, Zhenghan Qi1;1The University of Delaware, 2John Hopkins University
Human minds are apt to detect and extract statistical regularities from the environment (Saffran et al., 1996; Conway & Christiansen, 2005). The ability to rapidly learn frequencies, variabilities, and co-occurring information embedded in the inputs, known as statistical learning (SL), is foundational for various aspects of cognition including language development (Newport & Aslin, 2004). Conditional and distributional statistical regularities are the two types of information encoded by learners in numerous SL tasks. Conditional statistics refer to how frequently adjacent or non-adjacent elements co-occur in the inputs, while distributional statistics refer to the bare frequency of the occurrence of exemplars (Thiesson, 2017). The current study seeks to understand the neural processes underlying speech distributional SL and how it relates to the behavior of conditional SL. Using event-related potentials (ERPs), we measured participants’ neural sensitivity to speech distributional statistics in a passive auditory oddball task (N = 27). We manipulated the types of deviant stimuli (syllable or voice), as well as the global probability (rare or frequent) and the local probability (preceded by a long or short sequence of standard stimuli) of deviant stimuli. Participants also completed a target detection task while being exposed to a continuous stream of speech stimuli containing conditional statistical information (i.e. triplet patterns; Saffran et al., 1996). The linear acceleration slope of participants’ response time was computed as an index of conditional SL performance. The cluster-based mass univariate analysis between all deviants and standards resulted in two significant negativity effects: 22-180 msec (early window) at central electrodes and 324-500 msec (late window) at fronto-central electrodes. A repeated-measure ANOVA on the negativity amplitude in the late window yielded a significant main effect of deviant type (F(1,26) = 67.40, p = 0.02), with syllable deviants eliciting a greater negativity than voice deviants, and a significant interaction between local and global probability (F(1,26) = 82.74, p = .005). Post-hoc pairwise comparisons revealed a significant global probability effect (rare vs. frequent) when deviants were preceded by a short sequence of standards (t(53) = -2.27, p = .03), and a significant local probability effect (long vs. short distance) when deviants occurred frequently (t(53) = -1.95, p = .05). Individual participants’ ERP effects were then extracted to correlate with their performance in the conditional SL task. A greater sensitivity to global probability was associated with steeper acceleration of response time and faster response time in the conditional SL task (RT slope, rs = .58, one-tail p = .003; RT mean, rs = .55, one-tail p = .006). Greater sensitivity to the syllable deviants (as opposed to the voice deviants) was moderately related to steeper acceleration of response time (RT slope, rs = .43, one-tail p = .03). These findings suggest that adults are sensitive to distributional statistical information embedded at both the local and global contexts. The relationship between neural sensitivity to deviant type, local/global probability, and conditional SL behavior provide new evidence highlighting the important role of distributional statistical processing in learning conditional statistics embedded in speech.
Themes: Perception: Auditory, Speech Perception
Method: Electrophysiology (MEG/EEG/ECOG)