|Title:||Patterned exceptions in phonology|
|Abstract:||Standard Optimality-Theoretic grammars contain only the information necessary to transform inputs into outputs; regularities among inputs are not accounted for. Using the example of Tagalog nasal substitution, this dissertation presents a model of how lexical regularities could be learned, represented in the grammar, used by speakers and listeners, and perpetuated over time.
Lexical regularities are represented as low-ranking constraints, their rankings learned through exposure to the lexicon using Boersma's Gradual Learning Algorithm. High-ranked constraints ensure the primacy of listed pronunciations; but when a speaker produces a novel word, these high-ranking constraints are irrelevant and the constraints that encode lexical regularities take over. The subterranean constraints are stochastically ranked; speakers' behavior on novel words probabilistically reflect the lexical regularities. The listener uses the same grammar to produce well-formedness judgments for novel words and to reconstruct inputs from an interlocutors' outputs. The model's well-formedness judgments reproduce the experimental result that although the productivity of nasal substitution on novel words is low, nasal-substituted novel words are judged more acceptable than non-substituted words in certain cases.
Bayesian reasoning by the listener favors novel nasal-substituted words--they are disproportionately likely to become listed. A computer simulation of the speech community confirms that although nasal substitution is the minority pronunciation for novel words, a word may eventually enter the lexicon as nasal-substituted.
Tagalog vowel raising under suffixation is close to exceptionless in the native vocabulary but quite exceptionful among loanwords. A loan stem's probability of resisting raising is highly influenced by its degree of internal similarity. I propose that internal similarity encourages speakers to construe a word as reduplicated, even without morphosyntactic motivation; raising is blocked because it would disrupt base-reduplicant identity.
Alternatives to encoding lexical regularities in the grammar are considered. It is argued that the vowel raising facts are not amenable to an associative-memory account. The qualitative difference between "regulars" and "exceptions" cited by proponents of the Dual-Mechanism model as evidence for leaving lexical regularities out of the grammar reduces to a difference between listed words and synthesized words; this difference can arise through listener reasoning, without a prior qualitative difference.