The Role of Pauses and Entropy in Learning Nonadjacent Dependencies

Acquiring language entails inferring rules from limited input. The Entropy Model (EM) states that abstract generalizations are gradually attained as entropy (input-complexity) exceeds channel capacity (CC; brain’s encoding power). Radulescu et al. (2019) confirmed this prediction using an AAB rule. We tested EM’s predictions on nonadjacent-dependencies (NADs), which frequently reflect morphosyntactic rules.
Adults were exposed online to a NAD-artificial-language (aXb: a predicts b; X varies), in two entropy conditions: Low/High. Presence/absence of pauses segmenting the stream into aXb-items was manipulated. Pauses reduce entropy but also guide attention to individual aXb-items, decreasing attentional resources and thus available CC for encoding distributional/structural information. A grammaticality-judgment task including grammatical/ungrammatical test-items with familiar or unfamiliar X-elements, tested NAD-detection or NAD-generalization over novel X-elements, respectively. Higher entropy and lower CC were predicted to promote NAD-generalization, whereas lower entropy NAD-detection.
NAD-detection and NAD-generalization (preference-rate for grammatical-over-ungrammatical test-items) were higher in pause vs no-pause conditions, as predicted. NAD-detection was higher than NAD-generalization, showing the graduality of generalization. Entropy however did not affect generalization. This may be due to the lack of cues enhancing generalization used in studies that found an entropy-effect. Furthermore, online testing allowed limited control over participants. Results are discussed in relation to the EM.