Predicting semi-regular patterns in morphologically complex words

We expect generative models of language to correctly predict surface forms from underlying forms, but morphologically complex words, especially compounds, can exhibit idiosyncratic outputs, which require an extra lexical listing. This results in (a) a poorer Minimum Description Length of our model (Goldsmith 2011) and (b) failure of a grammar to capture patterning among exceptions. To solve an instance of this problem, we examine pitch-accent patterns of 2-mora-2-mora Japanese Yamato (native) noun-noun compounds, hitherto considered semi-predictable but which show gradient tendencies among constituents to trigger a particular accent pattern. In the framework of Gradient Symbolic Computation (Smolensky and Goldrick 2015), a type of harmonic grammar which allows partially activated feature values and weighted constraints, such gradient patterns can be captured through the additive combination of coalescing features on each conjunct, which results in a pitch accent when the summed activations surpass a threshold determined by the grammar. The ability of this framework to completely predict these semi-regular patterns holds promise that it can also explain similar kinds of patterns in other languages.