Semantic Representations in Working Memory: A Computational Model

Verbal working memory is supported by semantic knowledge. One manifestation of this is the rich pattern of semantic similarity effects found in immediate serial recall tasks. These effects differ from the effects of similarity on other dimensions (e.g., phonological similarity), which renders them difficult to explain. We propose a comprehensive mechanistic explanation of semantic similarity effects by extending standard connectionist architecture for modeling immediate serial recall to incorporate semantic representations. Central to our proposal is the selective encoding of categorical features shared among multiple list items. The selective encoding of shared semantic features is made possible via a tagging mechanism that enables the model to encode shared feature retrospectively. Through this mechanism, our model accounts for the majority of semantic similarity effects. Our results imply that working memory represents semantic information in a more restricted way than phonological information.