Algorithmic Learnability of Phonologies

Rutgers University New Brunswick, New Brunswick NJ

Investigators

Abstract

This project will develop an explicit formal system for the learning of phonologies, capable of inferring lexical representations, deducing constituency and other nonovert structure from overt observables, and constructing the mapping between lexical forms and the grammar's output. The primary formal problem to be attacked is the mutual entanglement of constituent structure, phonological mappings and phonological lexical representations. A language learner has direct access to none of these; they must be inferred from positive overt data. However, the three are tightly interrelated. The correct phonological mapping depends upon the structural analysis assigned to overt forms by the target grammar as well as on the lexical representations taken to underlie the overt forms. Assigning the correct structural analysis to an ambiguous overt form requires some grasp of what the correct phonological mapping is like, as does the induction of underlying forms. In the face of these entanglements, a successful solution to the full learning problem must work on all three simultaneously, using progress on one to achieve further progress on the others, ultimately arriving at the correct conclusion for each. The research will build on existing work solving important subproblems within phonological learning under Optimality Theory (OT). In prior work, learning algorithms have been developed for OT systems in which there are mutual entanglements between constituent structural analysis and phonological mapping, but in which lexical representations need not be learned. The new research proposed here will extend and generalize these approaches so that they may apply to systems which require non-trivial learning of lexical representations. This requires the addition of significant further structure to the learning and processing algorithms: a lexicon must be constructed and maintained, the parsing algorithms which assign analyses to overt data must be expanded to make use of the lexical representations, and the learner must have procedures for hypothesizing and adjusting lexical representations. The basic algorithms for learning mappings will also be modified to ensure learning of phonotactic distributions as a preliminary to the analysis of lexical relations. The investigations will begin with metrical stress grammars, including those with rich morphophonemic relations dependent on underlying contrasts in stress and quantity. An important part of the project will be the construction, as targets for learning, of constraint systems that plausibly capture phenomena requiring the nontrivial interaction between lexical representations and phonological mappings. Extensive survey and analysis of the targeted linguistic generalizations will be required to establish the empirical basis of the learner's goals. The proposed learning algorithms will be tested and evaluated via both formal analysis and computer simulations. Given that the property of mutual entanglement of analysis and mapping is not particular to phonology, but is endemic to the problem of learning from observable data in all linguistic domains, the results of this research are expected to provide insight into how language learning must proceed in general.