Extracting regularities from noise: Do infants encode patterns based on same and different relations?

A fundamental task of the young learner is to extract adjacent and distant dependency relations from the linguistic signal. Previous research suggests that infants successfully learn regularities from mini-grammars that contain a single consistent pattern. However, outside laboratory infants are exposed to a ‘noisy’ linguistic signal that contains multiple regularities and patterns that infants cannot yet interpret. In four experiments we explore how infants extract regularities from a 50 % ‘noisy’ input. Using an eye-tracker methodology we investigate how infants integrate patterns of varying complexity (e.g., adjacent and nonadjacent identity relations, or identity- and diversity-based relations) into differential anticipatory eye-movements. In Experiment 1, 7- and 12-month-olds were simultaneously exposed to AA and AB patterns (where As and Bs stand for syllables, such as in vava, valu) and they showed successful generalization for AA, but not for AB tokens. In Experiment 2, 7-month-olds heard AAB and ABA patterns and generalized only the AAB patterns. However, in Experiment 3 they could learn the ABA patterns when these were paired with ABC structures. Experiment 4 asked whether identity-based generalizations are restricted to exact physical identity. Infants generalized the AhAlBh patterns (where h stands for high pitch and l for low pitch), but no the AhBlAh ones, although in these the A syllables were physically identical. The results show that preverbal infants posses powerful abilities to extract regularities from a noisy signal, suggesting a possible hierarchy in encoding. Adjacent repetitions were computationally preferred over nonadjacent ones, even when adjacent repetitions were based on phonological identity, while nonadjacent ones on physical identity. While infants readily generalized the bi- or tri-syllabic structures based on identity relations they did not do so with the non-identity relations. Conceivably, encoding structures based on 'same’ relations are easier for infants than encoding diversity, suggesting that generalizations based on ‘different’ relations pose a challenge for the developing cognitive system.