Representing and learning stress in a MaxEnt framework

We present a new freely available software package `SoftStress' that learns and solves weighted constraint grammars with hidden structure. The package is equipped with a Maximum Entropy Grammar learner that gradually updates the constraint weights based on the probability distribution over all possible hidden representations, and a Linear Programming solver that can check all possible analyses involving hidden representations given a constraint set, without having to generate the factorial typology for the constraint set. We demonstrate the learner and the solver can be used to compare the representational capacity and the learnability of foot-based and grid-based theories of word stress, for a set of attested stress patterns.