Abstract This article investigates phonological complexity by using artificial neural network and Bayesian structural equation models to derive representations of phonological complexity from counts of the segments associated with particular features in languages’ phonemic inventories. These latent representations can then be used alongside principal component analysis to further analyse how interactions between phonological features affect overall complexity, and what phonological complexity patterns a model can detect in a phonological feature data set. The results indicate that the per-feature segment counts investigated tend to contribute positively to a language’s complexity, and that the latent complexity variables approximate a log-normal distribution. This implies that phonological complexifications co-occur with other complexifications diachronically while tending to be more constrained at the upper and lower ends of the complexity range.
Frederik Hartmann (Thu,) studied this question.