Over the last decade, iterated learning studies have provided compelling evidence for the claim that linguistic structure can emerge from non‐structured input, through the process of transmission. However, it is unclear whether individuals differ in their tendency to add structure, an issue with implications for understanding who are the agents of change. Here, we identify and test two contrasting predictions: The first sees learning as a pre‐requisite for structure addition, and predicts a positive correlation between learning accuracy and structure addition, whereas the second maintains that it is those learners who struggle with learning and reproducing their input who add structure to it. This prediction is hard to test in standard iterated learning paradigms since each learner is exposed to a different input, and since structure and accuracy are computed using the same test items. Here, we test these contrasting predictions in two experiments using a one‐generation artificial language learning paradigm designed to provide independent measures of learning accuracy and structure addition. Adults (N = 48 in each study) were exposed to a semi‐regular language (with probabilistic structure) and had to learn it: Learning was assessed using seen items, whereas structure addition was calculated over unseen items. In both studies, we find a strong positive correlation between individuals' ability to learn the language and their tendency to add structure to it: Better learners also produced more structured languages. These findings suggest a strong link between learning and generalization. We discuss the implications of these findings for iterated language models and theories of language change more generally.

Research on cross-linguistic differences in morphological paradigms reveals a wide range of variation on many dimensions, including the number of categories expressed, the number of unique forms, and the number of inflectional classes. However, in an influential paper, Ackerman & Malouf (2013) argue that there is one dimension on which languages do not differ widely: in predictive structure. Predictive structure in a paradigm describes the extent to which forms predict each other, called i-complexity. Ackerman & Malouf (2013) show that although languages differ according to measure of surface paradigm complexity, called e-complexity, they tend to have low i-complexity. They conclude that morphological paradigms have evolved under a pressure for low i-complexity, such that even paradigms with very high e-complexity are relatively easy to learn so long as they have low i-complexity. While this would potentially explain why languages are able to maintain large paradigms, recent work by Johnson et al. (submitted) suggests that both neural networks and human learners may actually be more sensitive to e-complexity than i-complexity. Here we will build on this work, reporting a series of experiments under more realistic learning conditions which confirm that indeed, across a range of paradigms that vary in either e- or i-complexity, neural networks (LSTMs) are sensitive to both, but show a larger effect of e-complexity (and other measures associated with size and diversity of forms). In human learners, we fail to find any effect of i-complexity at all. Further, analysis of a large number of randomly generated paradigms show that e- and i-complexity are negatively correlated: paradigms with high e-complexity necessarily show low i-complexity.These findings suggest that the observations made by Ackerman & Malouf (2013) for natural language paradigms may stem from the nature of these measures rather than learning pressures specially attuned to i-complexity.

Research on cross-linguistic differences in morphological paradigms reveals a wide range of variation on many dimensions, including the number of categories expressed, the number of unique forms, and the number of inflectional classes. However, in an influential paper, Ackerman & Malouf (2013) argue that there is one dimension on which languages do not differ widely: in predictive structure. Predictive structure in a paradigm describes the extent to which forms predict each other, called i-complexity. Ackerman & Malouf (2013) show that although languages differ according to measure of surface paradigm complexity, called e-complexity, they tend to have low i-complexity. They conclude that morphological paradigms have evolved under a pressure for low i-complexity, such that even paradigms with very high e-complexity are relatively easy to learn so long as they have low i-complexity. While this would potentially explain why languages are able to maintain large paradigms, recent work by Johnson et al. (submitted) suggests that both neural networks and human learners may actually be more sensitive to e-complexity than i-complexity. Here we will build on this work, reporting a series of experiments under more realistic learning conditions which confirm that indeed, across a range of paradigms that vary in either e- or i-complexity, neural networks (LSTMs) are sensitive to both, but show a larger effect of e-complexity (and other measures associated with size and diversity of forms). In human learners, we fail to find any effect of i-complexity at all. Further, analysis of a large number of randomly generated paradigms show that e- and i-complexity are negatively correlated: paradigms with high e-complexity necessarily show low i-complexity. These findings suggest that the observations made by Ackerman & Malouf (2013) for natural language paradigms may stem from the nature of these measures rather than learning pressures specially attuned to i-complexity.

No abstract