Nonparametric Learning of Phonological Constraints in Optimality Theory

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1094–1103, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Nonparametric Learning of Phonological Constraints in Optimality Theory Gabriel Doyle Department of Linguistics UC San Diego La Jolla, CA, USA 92093 gdoyle@ucsd.edu Klinton Bicknell Department of Linguistics Northwestern University Evanston, IL, USA 60208 kbicknell@northwestern.edu Roger Levy Department of Linguistics UC San Diego La Jolla, CA, USA 92093 rlevy@ucsd.edu Abstract We present a method to jointly learn fea- tures and weights directly from distri- butional data in a log-linear framework. Specifically, we propose a non-parametric Bayesian model for learning phonologi- cal markedness constraints directly from the distribution of input-output mappings in an Optimality Theory (OT) setting. The model uses an Indian Buffet Process prior to learn the feature values used in the log- linear method, and is the first algorithm for learning phonological constraints with- out presupposing constraint structure. The model learns a system of constraints that explains observed data as well as the phonologically-grounded constraints of a standard analysis, with a violation struc- ture corresponding to the standard con- straints. These results suggest an alterna- tive data-driven source for constraints in- stead of a fully innate constraint set. 1 Introduction Many aspects of human cognition involve the in- teraction of constraints that push a decision-maker toward different options, whether in something so trivial as choosing a movie or so important as a fight-or-flight response. These constraint-driven decisions can be modeled with a log-linear system. In these models, a set of constraints is weighted and their violations are used to determine a prob- ability distribution over outcomes. But where do these constraints come from? We consider this question by examining the dominant framework in modern phonology, Opti- mality Theory (Prince and Smolensky, 1993, OT), implemented in a log-linear framework, MaxEnt OT (Goldwater and Johnson, 2003), with output forms’ probabilities based on a weighted sum of constraint violations. OT analyses generally as- sume that the constraints are innate and univer- sal, both to obviate the problem of learning con- straints’ identities and to limit the set of possible languages. We propose a new approach: to learn con- straints with limited innate phonological knowl- edge by identifying sets of constraint violations that explain the observed distributional data, in- stead of selecting constraints from an innate set of constraint definitions. Because the constraints are identified as sets of violations, this also per- mits constraints specific to a given language to be learned. This method, which we call IBPOT, uses an Indian Buffet Process (IBP) prior to define the space of possible constraint violation matri- ces, and uses Bayesian reasoning to identify con- straint matrices likely to have generated the ob- served data. In identifying constraints solely by their extensional violation profiles, this method does not directly identify the intensional defini- tions of the identified constraints, but to the extent that the resulting violation profiles are phonologi- cally interpretable, we may conclude that the data themselves guide constraint identification. We …

Content truncated to 500 words