Superior Statistical Learning Relies on Rejecting Partwords

Speech segmentation in statistical learning is typically measured by the 2-alternative-forced-choice (2AFC) task. Although previous analysis has found performance differences among learners and better performance might stems from the larger representational differences between target words and partwords given the memory-based model, the underlying learning mechanism that drives these differences remains unclear. In the current study, seventy-four participants listened to a novel language and were then asked to complete a 2AFC task and a 7-point Likert scale familiarity rating task. On the basis of participants' performance on the 2AFC task, we identified a real-learning criterion and divided participants into a superior-level group (38% of participants) and a regular-level group. Though both groups exhibited above-chance learning in the 2AFC and familiar rating tasks, superior learners performed significantly better than regular learners. In addition, a series of linear mixed effect models showed that superior- and regular-level groups produced comparable ratings on target words and nonwords; however, superior-level participants rated partwords as less familiar than regular-level participants. These patterns suggest that the only difference that contributed to their overall performance on the SL task was their perceived level of familiarity of partwords. This study highlights the importance of investigating SL mechanism from a memory-based model and provides a nuanced method for examining individual differences in SL tasks.