An Approach to the Automatic Acquisition of Phonotactic Constraints

This paper describes a formal approach and a practical learning method for automatically acquiring phonotactic constraints encoded as-nite automata. It is proposed that the use of diierent classes of syllables with class-speciic intra-syllabic phonotactics results in a more accurate hypothesis of a language's phonological grammar than the single syllable class traditionally used. Intra-syllabic constraints are encoded as acyclic nite automata with input alphabets of phonemic symbols. These automata in turn form the transitions in cyclic nite automata that encode the inter-syllabic constraints of word-level phonology. A genetic algorithm is used to automatically construct nite automata from training sets of symbol strings. Results are reported for a set of German syllables and a set of Russian bisyllabic feminine nouns. 1 Background 1.1 Phonotactic Description In recent years, phonology | partly under the innuence of computational models | has moved away from procedural, rule-based approaches towards explicitly declarative statements of the constraints that hold on possible phonological forms. Such statements form sets of constraints that apply at a given level of description, and ill-formedness is often deened as constraint violation. Phonotactic descriptions state the constraints that hold for possible sequences of phonetic or phonemic features or symbols, usually at the level of the syllable (more rarely for onset, peak and coda separately). Phonological words are deened as sequences of at least one syllable. A phonotactic description is typically thought to be adequate only if it generalises beyond the set of phonological forms that exist in a language to a superset of possible forms that also includes forms that could exist but do not. This distinction between non-existent but possible forms on the one hand, and non-existent and impossible forms on the other, is often described in terms of accidental vs. systematic gaps (e.g. lists ve encoding schemes for phonotactic description at the syllable level found in the literature: templates that merely state the number of consonants permitted in the onset and coda, distribution matrices with a separate matrix for each type of consonant cluster, enhanced templates which add the notion of phoneme classes, feature-based phono-tactic networks using feature bundles, natural classes, variables, defaults and underspeci-cation, and phrase-structure rules which have the same (potential) representative power as feature-based phonotactic networks. In the approach presented here, nite state automata (FSA) encoding is preferred over these other schemes, since they can all equivalently be represented by FSAs | what is described is always a regular …