Implicit Learning Of Chinese Tonal Symmetry Rules And The Neural Network Simulations

Humans are equipped with powerful learning mechanisms for acquiring unconscious knowledge of structural regularities from the environment. This form of learning is known as implicit learning, which was proposed by Reber in1967. Over the last several decades, researches into implicit learning generally address two major questions:(1) what kind of knowledge can be acquired implicitly:either the specific surface features or the underlying abstract rules?(2) how such knowledge is implicitly learned? What kind of the computational mechanism is responsible for this type of learning?Since most traditional researches into implicit learning have concentrated on the learning of finite-state grammars with almost entirely local dependencies, it is difficult to eliminate the effect of surface features such as chunks of local elements. Finite-state grammars are thus limited in demonstrating that whether implicit learning can acquire surface features or abstract rules. The present study used nonlocal dependencies generated by higher than finite-state grammars, specifically, Chinese tonal symmetries, i.e., inversion symmetry rule and retrograde symmetry rule, controlled the surface features such as chunks, and attempted to provide new evidence that abstract rules could be acquired implicitly. Moreover, differing from finite-state grammars, nonlocal symmetry rules generally require a memory buffer to be implicitly learned. The present study used human experiment and neural network simulations, investigated the nature of the memory buffer involved in implicit learning of symmetry rules by contrasting the implicit learning of the retrograde and inversion rules, and attempted to reveal the mechanisms of implicit learning further.Results of the present study showed that:(1) people could go beyond the learning of surface features such as chunks and learned abstract Chinese tonal symmetry rules, providing new evidence for the chunks learning vs. rules learning dispute;(2) implicit learning of Chinese tonal symmetry rules only occurred when the context of rules was with a large variability, demonstrating that implicit learning of abstract rules was modulated by the context variability, and hence providing new evidence that abstract rules could be acquired implicitly;(3) implicit learning was easier for inversions than retrogrades, suggesting that the memory buffer used in implicit learning mechanism is functionally more like a first in-first out memory buffer than a last in-first out memory buffer;(4) Simple recurrent networks (SRN) could match the implicit learning of human subjects, demonstrating that the memory buffer and computational mechanism used by SRN models could simulate the learning mechanism that human subjects used for learning symmetry rules. Furthermore, consistent with the results of human experiments, SRN simulations replicated an advantage of inversions over retrogrades, thus demonstrated that functionally a first in-first out memory buffer is more likely to be involved in implicit learning.