Language learning as uncertainty reduction: effects of prediction error and entropy on generalization and item-learning

Vujovic, M. ¹ , Ramscar, M. . ² & Wonnacott, E. ¹

1 University College London
2 University of Tubingen

Learning theory frames learning as uncertainty reduction (Rescorla & Wagner, 1972) and makes key predictions about the role of linear order in language learning. We test this by comparing suffixing and prefixing artificial languages, predicting: (i) Greater generalization for suffixes, since this allows cue competition and prediction error over the preceding lexical items (Ramscar et al 2010); (ii) Greater item-learning with prefixing, since hearing a prefix smooths entropy for the subsequent noun (Dye et al., 2017).

Experiment1 tested (i) by exposing adults (N=120) to languages where nouns co-occurred with either prefixes or suffixes, with the choice of affix depending on semantic cues. As predicted, at test, generalization over semantic cues was better for suffixes. Experiment2 tested (ii) in a vocabulary learning experiment (N=40 adults). Nouns again co-occurred with either prefixes or suffixes, and had visual referents. Importantly, however, we first exposed participants to the prefix-noun or noun-suffix bigrams (Day1), only adding in the visual referents in a second exposure stage (Day2). As predicted, at test, participants were better able to map particular nouns to their objects in the prefix-condition.

Results demonstrate the crucial role of prediction error and entropy, and the importance of linear order in human language learning.