A Connectionist Model of Phonological Working Memory

Abstract

We propose a connectionist model that can learn the pronunciation of words by imitation. The model “hears” the pronunciation of words (sequences of phonemes) and develops a static internal representation of phonology. It then tries to reproduce the pronunciation from the internal representation. We trained the model to imitate the pronunciation of 3,684 English monosyllabic words. After 15 million training trials, the network achieved a 99.2% correct response. To simulate the process of the serial recall of words or nonwords, we added a subsystem called a temporary maintenance module. This module preserves the phonological internal representation by using a Hebbian one-shot algorithm to transform the activation pattern to a connection pattern. The model could reproduce the effect of the phonotactic probability on a serial recall task done by human children. The result of the simulation implies that our model can explain the effect of long-term memory on the phonological working memory.