Assessing Integrative Complexity as a predictor of morphological learning using neural networks and artificial language learning

Morphological paradigms differ widely across languages: some feature relatively few contrasts, and others, dozens. Under the view that languages are under pressure to be learnable and that the distribution of languages in the world reflects biases in language learning, this diversity is surprising – how could paradigms which apparently differ so markedly be similarly learnable? Recent work on morphological complexity has argued that even very large paradigms are designed such that they are easy to learn and use. Specifically, Ackerman & Malouf, 2013 propose an information-theoretic measure, i-complexity, which captures the extent to which forms in one part of a paradigm predict forms elsewhere in the paradigm, and contrast this measure with e-complexity, which captures the number of distinctions made by the language and the different ways to mark each grammatical function. They show that languages which differ widely in e-complexity exhibit similar i-complexity; in other words, morphological paradigms with many contrasts reduce the learnability challenge for learners by having predictive relationships between forms. Here, we test whether i-complexity predicts the learnability of inflectional paradigms using both recurrent neural networks and human participants trained on an artificial language. Furthermore, we compare the effect of i-complexity on learning with that of e-complexity. We find that in RNNs both i- and e-complexity have an effect on learning: paradigms with lower i- and e-complexity are easier to learn, although the effect of e-complexity is larger. However, for human learners, we find only weak evidence (if any) that low i-complexity paradigms are easier to learn; in contrast, low e-complexity is clearly beneficial for learning. This suggests that i-complexity might have relatively little influence on the learnability of inflectional paradigms, with other factors, such as the e-complexity having a greater effect. These results suggest that appealing to i-complexity does not fully resolve the paradox of cross-linguistic variation in morphological systems.