Key points are not available for this paper at this time.
Statistical learning (SL) is considered a cornerstone of cognition. While decades of research have unveiled the remarkable breadth of structures that participants can learn from statistical patterns in experimental contexts, how this ability interfaces with real-world cognitive phenomena remains inconclusive. These mixed results may arise from the fact that SL is often treated as a general ability that operates uniformly across all domains, typically assuming that sensitivity to one kind of regularity implies equal sensitivity to others. In a preregistered study, we sought to clarify the link between SL and language by aligning the type of structure being processed in each task. We focused on the learning of trigram patterns using artificial and natural language statistics, to evaluate whether SL predicts sensitivity to comparable structures in natural speech. Adults were trained and tested on an artificial language incorporating statistically-defined syllable trigrams. We then evaluated their sensitivity to similar statistical structures in natural language using a multiword chunking task, which examines serial recall of high-frequency word trigrams-one of the building blocks of language. Participants' aptitude in learning artificial syllable trigrams positively correlated with their sensitivity to high-frequency word trigrams in natural language, suggesting that similar computations span learning across both tasks. Short-term SL taps into key aspects of long-term language acquisition when the statistical structures-and the computations used to process them-are comparable. Better aligning the specific statistical patterning across tasks may therefore provide an important steppingstone toward elucidating the relationship between SL and cognition at large.
Isbilen et al. (Wed,) studied this question.