Probabilistic Generation of Sequences Under Constraints

There is growing interest in the ability to generate natural and meaningful sequences (e.g., in domains such as language or music). Many existing sequence generation models, including Markov and neural algorithms, capture local coherence, but have no mechanism for applying the structural constraints that are so often essential for the development of meaning. We describe a novel solution to this problem which combines hidden Markov models with constraints, allowing sequences which obey user-defined constraints to be generated according to data-driven probability distributions. Compared to other constrained probabilistic solutions, our Constrained Hidden Markov Process (CHiMP) has significantly greater expressivity, allowing the user to generate constrained sequences that are longer and which have more numerous structural constraints.