Learning Grammar via Statistical Mechanism

A novel word-by-word prediction paradigm, which employs a task similar to the one given to Elman's (1990, 1991, 1993) simple recurrent networks, is used in the current study to investigate whether adults exhibit learning higher-order contingencies corresponding to structural contingencies. The results indicate that token and category contingencies are learnable regardless of training conditions, but higher-order constituent contingency was only learnable in conditions that facilitated initial learning of bigram and trigram word frequencies. Our results are consistent with previous studies that involve both human subjects and computer simulations, and provide a guide toward identifying factors that may enhance learning of higher-order statistical relations in human subjects.