Underphonologization and modularity bias

Elliott Moreton (submitted)

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 (19 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 24 studies of 17 languages).  A learning
algorithm is proposed in which the learner chooses between 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 (ones that
interact with few other constraints).  Its application to the tone
case is illustrated by simulation.