Rule Reliability in Natural and Artificial Grammar: The Case of Velar Palatalization

Russian velar palatalization changes velars into alveopalatals before certain suffixes, including the stem extension –i and the diminutive suffixes –ok and –ek/ik. While velar palatalization always applies before the relevant suffixes in the established lexicon, as depicted by dictionaries, it often fails with nonce loanwords before –i and –ik but not before –ok or –ek. A model of rule induction and weighting (the Rule-based Learner, developed by Albright and Hayes 2003) is trained on the established lexicon of Russian, in which velar palatalization is exceptionless, and tested on new borrowings. Despite the fact that velar palatalization is exceptionless in the training set for every suffix, it is correctly predicted to often fail with novel words before –i and –ik but not before –ek or –ok based on information in the lexicon. This success can be traced to the model‟s weighting of competing rules according to their reliability. Reliability-driven competition between rules is shown to predict that a morphophonological rule will fail if the triggering suffix comes to attach to inputs that are not eligible to undergo the rule. This prediction is confirmed in an artificial grammar learning experiment. A method for distinguishing between sourceand product-oriented mental grammars is developed and product-oriented generalizations are shown to be unable to account for the data from the examined artificial grammar learning paradigm (Bybee & Newman 1995). The influence of the learning paradigm on the shape of the learned grammar is discussed. Finally, the winning model (the Rule-based Learner) is shown to succeed only if the suffix and the stem shape are chosen simultaneously, as opposed to the suffix being chosen first and then triggering or failing to trigger a stem change and if the choice between competing rules is stochastic.