Abstract: Learning multiple non-adjacent dependencies is fundamental to processing the hierarchical structures of natural language and constitutes one of the most challenging aspects of language acquisition. Such dependencies are typically acquired automatically and implicitly; however, how to effectively facilitate their learning remains unresolved. A particularly powerful factor in improving learning and generalization is variability in the input examples. However, its role in the implicit learning of multiple non-adjacent dependencies remains unclear, particularly whether such variability enables learners to move beyond value-value mappings to operations over variables, thereby acquiring transferable abstract rules. Accordingly, this study first examines how different levels of example variability influence implicit learning performance (Experiment 1) and then explores whether the effect of variability operates at the level of abstract variable-variable mappings (Experiment 2).
The study employed multiple non-adjacent rules defined over the tones with which Chinese syllables were spoken. Specifically, the tone types (ping or ze) of the first five syllables predicted those of the final five by an inversion relation in strings of length ten. In Experiment 1, three conditions of variability (low, medium, and high) were established by manipulating the number of syllable types used in the strings. The experiment comprised a training phase and a test phase. During the training phase, participants listened to a series of tone-syllable strings that followed the multiple non-adjacent rules and repeated them. In the test phase, participants were told that all strings they heard during the training phase followed a grammatical rule, and were asked to judge whether a set of new strings conformed to the learned rule. In Experiment 2, to further examine the level at which variability facilitates learning, two types of materials were constructed. One included only partial tone mappings (e.g., Tone 1-3 or Tone 2-4), while the other included complete tone mappings between level tones (Tones 1/2) and contour tones (Tones 3/4). Unlike in Experiment 1, during the training phase of Experiment 2, participants were exposed only to strings containing partial mappings. They were then tested in a non-transfer test (strings with the same partial mappings as in training), and either a mixed transfer test (strings with complete mappings; Experiment 2a) or a novel transfer test (strings with novel partial mappings that conflicted with training; Experiment 2b).
The results revealed a U-shaped relationship between example variability and implicit learning performance in Experiment 1: participants showed significant implicit learning effects in both the low- and high-variability conditions, whereas no learning effect was observed in the medium-variability condition. Performance was greater in the low- and high-variability conditions than in the medium-variability condition. In Experiment 2, the participants demonstrated implicit learning in the non-transfer test (Experiment 2a) across all variability conditions. However, their ability to transfer this learning to new materials varied significantly by condition. Participants in the low- and high-variability conditions were able to implicitly extract abstract rules from partial mappings and generalize them to complete mappings (Experiment 2a). Remarkably, this transfer ability was maintained even when participants were confronted with novel partial mappings that directly contradicted the training material (Experiment 2b). In contrast, participants in the medium-variability condition failed to transfer their knowledge to new materials in either transfer tests. This indicates that both low and high variability facilitated the implicit learning of transferable abstract rules (variable-variable mapping level), whereas participants in the medium-variability condition implicitly learned only surface-level value-value mappings.
Taken together, this study provides a comprehensive account of how example variability influences the implicit learning of multiple non-adjacent dependencies. It reveals a nonlinear relationship between example variability and implicit learning and identifies that this effect operates primarily at the level of abstract variable-variable mappings. From the perspective of example variability, the findings further indicate that the implicit acquisition of abstract rules is constrained by specific boundary conditions. The findings not only deepen the understanding of the mechanisms underlying implicit learning, but also provide theoretical implications for optimizing instructional design.

Key words: variability, implicit learning, multiple non-adjacent dependencies, transfer, U-shaped relationship