Networks that Learn about Phonological Feature Persistence

Two experiments investigating the power of simple recurrent networks to acquire phonological regularities that can be described in terms of the persistence of particular features show that an SRN is capable of learning patterns of feature persistence across morpheme boundaries.

Abstract

Natural language phonology presents a challenge to connectionists because it is an example of apparently symbolic, rule-governed behavior. This paper describes two experiments investigating the power of simple recurrent networks (SRNs) to acquire phonological regularities that can be described in terms of the persistence of particular features. The first experiment demonstrates the ability of an SRN to learn simple harmony constraints within morphemes; that is, restrictions on the co-occurrence of particular types of segments. The second experiment shows that an SRN is capable of learning patterns of feature persistence across morpheme boundaries, in this case, the alterations occurring in the regular plural forms of English nouns. This behavior is usually characterized in terms of a derivation from a more to a less abstract level; here it takes the form of the more psychologically plausible processes of the production of a sequence of segments given a meaning or the generation of a meaning given a sequence of segments. This is accomplished by having both segmental and semantic inputs and outputs in the network. Trained on the task of auto-associating the current segment and meaning and predicting the next segment, the network learns to produce the appropriate plural suffix on nouns it has seen in only the singular and to recognize novel plural forms as plural.