Infants can learn nonadjacent dependencies easily with correlated semantic cues

Willits, J. 1 & Saffran, J. 2

1 Indiana University
2 University of Wisconsin Madison

Nonadjacent dependencies are ubiquitous in language. They contribute to its complexity, and to the difficulty of learning a language. Nonadjacent dependencies are also, on some views, difficult to learn using association-based mechanisms. In the present research, we show that adding real world complexity to the learning situation - in the form of correlated semantic cues - greatly enhances the learnability of nonadjacent dependencies. In three experiments, 24-MO infants were presented with a familiarization corpus containing 64 three-word strings, containing nonadjacent dependencies such that the first word perfectly predicted the third word in each string. In Experiment 1, infants heard a consistently semantically related corpus, where the nonadjacently related words were also always members of the same semantic category (e.g. doggy-x-kitty; bird-x-fishy, cookie-x-banana, toast-x-cheese). In Experiment 2 infants heard a consistently unrelated corpus, where the nonadjacently related items were consistently members of opposite categories, preserving the relation between the two categories (e.g. doggy-x-banana; bird-x-cheese, cookie-x-kitty, toast-x-fishy). Experiment 3 was a control experiment, with no consistent semantic relation between the nonadjacently related items (e.g. doggy-x-kitty; bird-x-cheese, cookie-x-banana, toast-x-fishy). Infants in all three experiments were then given 16 test strings, half from familiarization and half that violated the nonadjacent dependencies on which they were trained. Infants in Experiment 1 and 2 listened significantly longer to the novel violation items, with no significant difference in Experiment 3. Further, the size of each infant's novelty preference was significantly correlated with their MCDI score. These experiments show that learning nonadjacent dependencies is in fact quite tractable. Many cues (such as meaning) that learners might use are stripped away in overly controlled experiments, making the problem harder than it is in the real world. Complexity is not the same thing as noise, if that complexity provides learners with useful cues to the structure of the world.