Para2S: Towards an architecture for generating novel forms of known words using discriminatively learned associations

Kapatsinski, V.

University of Oregon

Speakers often face the task of computing a form of a known word appropriate for the current context (e.g., plural). The present paper develops a computational model of this task using associations learned by error-driven associative learning (Baayen et al., 2011). To compute the novel form, the speaker activates known forms and the target meaning. Paradigmatic associations (k->ch) link cues from the activated known forms of the word to output segments. A paradigmatic association can be strengthened by increasing its frequency in the input or making corresponding forms temporally adjacent. Schematic associations between the target meaning and form features (plural->k#) are strengthened by making certain segments overrepresented in forms with that meaning, e.g. if most plural forms end in [k], learners often add [k] to singulars that lack it, even if addition is never witnessed in training. Paradigmatic and schematic associations compete to fill out the output form in parallel. Syntagmatic associations come into play when some part of the output is selected before others. For example, if singular-final [p] changes to [m] before -i, and =i is favored by a singular-final [p], -i also becomes favored by singular-final [m], even if no [m]-final singulars are presented in training.