Discovering underlying forms: Contrast pairs and ranking

Phonological learners must acquire a lexicon of underlying forms and a constraint ranking. These must be acquired simultaneously, as the ranking and the underlying forms are interdependent. Exhaustive search of all possible lexica is intractable; the space of lexica is simply too large. Searching the underlying forms for each overt form in isolation poses other problems. A single overt form is often highly ambiguous among both underlying forms and rankings. In this dissertation I propose a learning algorithm that attends to pairs of overt forms that differ in exactly one morpheme. These pairs can exhibit less ambiguity than the isolated overt forms, while still providing a reduced search space. The algorithm first assigns underlying values to occurrences of features whose surface realization never alternates; the other underlying features are left initially unset (Tesar et al., 2003). Pairs of overt forms that differ in one morpheme are then constructed. The algorithm then considers the possible values of unset features for each pair, processing pairs with the fewest unset features first. It uses inconsistency detection (Tesar, 1997) to test sets of values of unset features for viability. A set of values for the unset features is viable if it produces the correct overt forms under some ranking. Those feature values which are common across all viable solutions are then set. In the process of testing for inconsistency for each set of values of unset features a set of winner-loser pairs is generated. The learner determines the ranking restrictions jointly entailed by these sets of winner-loser pairs. These ranking restrictions are then maintained while processing all further contrast pairs. After all pairs have been processed, any still unset feature values are assigned default values. The general success of the algorithm depends upon these features being fully predictable in the output. A ranking is then obtained from this lexicon using Biased Constraint Demotion (Prince and Tesar, 2004). Fixing all non-alternating features reduces the effective lexical search space. The algorithm further reduces the lexical search space by breaking up the search into tractable local pair searches. Extracting shared ranking information from winner-loser pairs generated from inconsistency detection restricts which featural combinations for future contrast pairs will be viable providing information that is otherwise unavailable to the learner.%%%%