The role of prediction error in linguistic generalization and item learning: evidence from an online eye-tracking study

The discriminative learning framework, grounded in principles of learning theory, was recently revitalized as a specific mechanism to carry out language acquisition within a domain-general, input-driven constructivist account. It specifies that language learning, rather than being guided by an innate linguistic system like Universal Grammar, could be constrained by obtaining negative evidence using prediction errors. Previous studies have tried to indirectly show the effect of prediction making on word-referent mapping by comparing the learning performance of two groups of learners who experienced language input with different input order conditions, where only one condition allowed cue competition. The current study applied eye tracking to the learning process and demonstrated direct evidence for the predictive behaviors during discriminative learning. Two learning experiments with a learning phase where cues predicted visual linguistic outcomes (labels) were conducted. Experiment 1 was a replication of an earlier word learning study – with eye tracking added – and showed that participants made predictive saccades towards positions of labels before they appeared for both high and low frequency conditions. Experiment 2 used a new cross-situational learning paradigm and did not demonstrate default fixations. Offline results showed weaker learning than previous studies and did not show above-chance accuracy for the low frequency trials. However, plot analysis of the individual variation for low frequency trials suggested that participants who failed to learn low frequency cues could have been persistently overgeneralizing in training. This overgeneralization pattern is discussed in the context of adult L2 learning