Domain-general categorisation explains constrained cross-linguistic variation in noun classification

Languages appear limited in the range of concepts that are grammatically encoded. For example, person, number, and animacy distinctions are regularly found in e.g., grammatical agreement systems. But, despite their visual salience, colour distinctions are completely absent from such systems. Some have taken this to indicate domain-specific constraints on what can and cannot be part of grammars. Here, we test an alternative possibility, that domain-general cognitive capacities can explain these regularities. Using animacy- and colour-based agreement as our test cases, we show that a bias for animacy over colour indeed exists during learning of a miniature artificial agreement system. We then show that a parallel animacy-over-colour bias is found in a non-linguistic sorting task. Finally, we explore the cognitive roots of the animacy bias. Specifically, we ask whether it is driven by a domain-general categorisation principle favouring categorisation based on features that are highly predictive of other features. Using natural language corpus data, we find that animacy-based classification produces distinct and more compact categories, which are more easily learnable. We also find preliminary causal evidence for this explanation: when animacy is less predictive of other object features than colour, learners who notice this novel predictive structure learn animacy-based noun classes worse. Taken together, our results support the idea that domain-general principles may be responsible for the prevalence of certain semantic distinctions over others in grammar.