Predictive structure and the learnability of inflectional paradigms: investigating whether low i-complexity benefits human learners and neural networks
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. This typological variation is surprising within the approach that languages evolve to maximise learnability (e.g., Christiansen and Chater 2008; Deacon 1997; Kirby 2002). Ackerman and 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, sometimes called i-complexity. Ackerman and Malouf (2013) show that although languages differ ac- cording to surface paradigm complexity measures, called e-complexity, they tend to have low i-complexity. While it has been suggested that i-complexity affects the task of producing unknown forms (the Paradigm Cell Filling Problem, Ackerman, James P. Blevins, et al. 2009; Ackerman and Malouf 2015), its effect on the learnability of morphological paradigms has not been tested. In a series of artificial language learning tasks both with human learners and LSTM neural networks, I evaluate the hypothesis that learners are sensitive to i-complexity by testing how well paradigms which differ on this dimension are learned. In Part 1, I test whether learners are sensitive to i-complexity when learning inflected forms in a miniature language. In Part 2, I compare the effect of i-complexity on learning with that of e-complexity and assess the relationship between these two measures, using randomly con- structed paradigms. In Part 3, I test the effect of i-complexity on learning and generalisation tasks, manipulating the presence of extra-morphological cues for class membership. Results show weak evidence for an effect of i-complexity on learning, with evidence for greater effects of e-complexity in both human and neural network learners. A strong nega- tive correlation was found between i-complexity and e-complexity, suggesting that paradigms with higher surface paradigm complexity tend to have more predictive structure, as mea- sured by i-complexity. There is no evidence for an interaction between i-complexity and extra-morphological cues on learning and generalisation. This suggests that semantic or phonological cues for class membership, which are common in natural languages, do not enhance the effect of i-complexity on learning and generalisation. Finally, i-complexity was found to affect generalisation in both human and neural network learners, suggesting that i-complexity could, in principle, shape languages through the process of generalisation to unknown forms. I discuss the difference in the effects of i-complexity on learning and generalisation, the similarities between the effect of i-complexity in human learners and neural networks, and cases the two types of learner differed. Finally, I discuss the role that i-complexity is likely to have in language change based on the results.