Vujovic, M. 1 , Ramscar, M. 2 & Wonnacott, E. . 1
1 Department of Language and Cognition, University College London
2 Department of Linguistics, University of Tübingen
Discriminative-learning frames language learning as a process by which prediction error is used to discriminate uninformative cues and to reinforce informative cues (Ramscar et al., 2010). This approach successfully predicts that word learning is facilitated when learners (adults and 2-year-olds) view referents before hearing their labels, since this allows opportunity for prediction error to discriminate the appropriate set of semantic features for each label. Ramscar (2013) applied this approach to learning at morphological/syntactic level as an explanation of the cross-linguistic preference for suffixing over prefixing (Greenberg, 1963). Specifically, as a consequence of linear order in discriminative-learning, suffixing benefits learning of abstract common dimensions across preceding lexical items (generalization), whilst prefixing benefits item-based learning (Arnon & Ramscar 2012; Ramscar 2013).
In the current work, adult participants were exposed to an artificial language with two noun categories (i) marked by phonological and semantic cues and (ii) accompanied by an affix (category1_affix: ge, category2_affix: ma) which either preceded (prefixing-condition) or followed (suffixing-condition) the noun. Training was followed by tests of vocabulary and generalization. Computational models (using a discriminative implementation of the delta rule (Widrow & Hoff 1960)) trained on each version of the language predict that participants in the prefixing-condition should show better vocabulary learning, whereas the suffixing-condition should be better at generalizing the correct affix to new category members. In Experiment 1 (N=84), we found better generalization in the suffix condition but no evidence of vocabulary learning in either condition, possibly due to vocabulary size (16 items). In Experiment 2 (data collection ongoing), we halved the vocabulary. Preliminary analyses suggest that again we see better generalization with suffixing, but now some vocabulary learning, which is better in the prefix condition. The results demonstrate the crucial role of prediction error in linguistic generalization, and the importance of linear order in human language learning.