Blurb:

The most straightforward theory of how phonologization interacts with Universal Grammar to determine typology is that UG defines the cognitively possible grammars (‘hard’ typology), while phonologization determines how frequent they are (‘soft’ typology). This paper argues instead that some soft typology has a cognitive source, and proposes a formal explanation. Phonological patterns relating tone to tone are shown to be more common than those relating tone to voicing and aspiration (20 families on 5 continents versus 8 families on 4 continents). This soft typological fact cannot be derived from differential robustness of the phonetic precursors, which have similar magnitude (survey of 26 studies of 17 languages). A learning algorithm is proposed in which the learner chooses between Optimality-Theoretic constraint sets based on how probable they make the training data (‘Bayesian Constraint Addition’). This biases the learner towards phonologizing processes driven by ‘modular’ markedness constraints, i.e., ones that interact with few other constraints. Its application to the tone case is illustrated by simulation, and compared with alternatives.