Constrained sampling of Markov chains

David Bulger, Timothy Powles

    Research output: Chapter in Book/Report/Conference proceedingConference proceeding contributionpeer-review

    Abstract

    Homogeneous Markov chains with discrete state-space and time are very straightforward to simulate, due to the memorylessness property. In this talk, we consider sampling from the normalised restriction of a Markov chains joint distribution to a subset defined by two kinds of constraints: univariate constraints, restricting individual chain variables to belong to certain sets, and multivariate constraints, restricting non-adjacent groups of chain variables to be unequal but equivalent under a predefined equivalence relation. The multivariate constraints, in particular, present some challenges. An efficient sampling method involving Metropolis-Hastings with some pre-tabulated distributions is described.

    We demonstrate an application to text generation, and in particular the random generation of rhyming, scanning lyrics, as a component of an algorithmic songwriting project. In this setting, the states of the Markov chain are syllables (that know which words they are from, so that for instance the final syllables of perilous and marvellous are two different states). The univariate constraints enforce the metre of the text, by indicating which syllables must be stressed or unstressed (for correct scansion) and which syllables must be word-final (so that words do not straddle lines). The multivariate constraints enforce a rhyming scheme, by requiring certain groups of syllables to rhyme without being equal.

    This work relies on the Carnegie Mellon Pronouncing Dictionary (CMUdict; Weide, 1998) and sample text. CMUdict contains over 130000 entries, comprising words, proper nouns et cetera as used in spoken American English. This dictionary indicates the phonemic and stress pronunciation of each entry.

    A Markov chain including the entire dictionary would require over 330000 syllable states. We have used sample texts, both to reduce the state space, and to build a first-order Markov model of word sequence, or equivalently, syllable sequence. The problem of interest is to sample from the Markov chain, subject to the constraints imposed by a given rhyming scheme.

    Our algorithm firstly samples the multivariately constrained chain variables. After location of a feasible subsequence of rhymed syllables, ratios of joint probabilities are calculated and compared in a Metropolis-Hastings algorithm, where each candidate arises from the incumbent by replacing one set of rhyming syllables with another drawn from a pre-tabulated distribution. After a burn-in period, the skeleton provided by the rhymed syllables is fleshed out by sampling the unrhymed syllables according to the Markov chain, subject to metrical and boundary conditions.
    Original languageEnglish
    Title of host publication22nd International Congress on Modelling and Simulation
    EditorsG. Syme, D. Hatton MacDonald, B. Fulton, J. Piantodosi
    Place of PublicationHobart
    PublisherModelling & Simulation Society Australia & New Zealand
    Pages466-472
    Number of pages7
    ISBN (Electronic)9780987214379
    Publication statusPublished - Dec 2017
    EventInternational Congress on Modelling and Simulation (22nd : 2017) - Hobart, Australia
    Duration: 3 Dec 20178 Dec 2017

    Conference

    ConferenceInternational Congress on Modelling and Simulation (22nd : 2017)
    CountryAustralia
    CityHobart
    Period3/12/178/12/17

    Keywords

    • Constrained Markov chain
    • Metropolis
    • rhyme
    • text generation

    Fingerprint Dive into the research topics of 'Constrained sampling of Markov chains'. Together they form a unique fingerprint.

    Cite this