Dynamics of Retrieval from Lexical Memory (Information-Processing, Spreading Activation, Reaction Time).

Abstract

A visually presented word can be recognized faster if it appears shortly after a semantically related context word than if it appears in isolation. This effect can be explained by models of memory that postulate a network of interconnected concept nodes in which activation spreads from a source node (corresponding to the semantic context) to a recipient node (corresponding to the word to be recognized). If the recipient node has benefitted from contextual activation, little additional activation from sensory channels will be needed to achieve a recognition activation threshold. A fundamental question about such a structure concerns the nature of activation growth at a node in the network. Two mutually exclusive and exhaustive possibilities are (1) that activation grows in a discrete fashion, making abrupt transitions between two or more distinct states, and (2) that activation grows continuously from a resting level to some asymptotic level. In the dissertation, this dichotomy is characterized with theoretical examples from the literature, and an adaptive priming procedure is introduced that is used to test the two classes of models. The procedure involves the presentation of contextual (or priming) information immediately before the presentation of a target letter string; subjects are required to make a lexical (word/nonword) decision about the target. The interval between the presentation of the prime and target strings is varied adaptively (according to subject performance), and the latency and accuracy of responses as a function of this interval are recorded. According to models in the discrete class, there is a unique latency distribution associated with each state of activation. If the moment of transition between states varies stochastically, then for a fixed intermediate-duration priming interval, one or another of the states will produce the response on each trial. Over many trials, the latency distribution so produced will constitute a finite mixture of the basis distributions associated with the discrete activation states. Models of the continuous class assert instead that activation grows continuously with time and that there is a unique latency distribution associated with any given degree of intermediate priming. Such models predict that no finite mixture distribution will emerge from the adaptive priming procedure when the priming interval has an intermediate duration. Four experiments with the adaptive priming procedure are reported, all of which require the rejection of the mixture prediction. It is concluded that models requiring only two discrete states of activation do not accurately characterize the dynamics of retrieval from lexical memory. Continuous or higher-order discrete models better account for the results. Several current models of retrieval are consistent with this finding, including r and om-walk models, parallel distributed processing models, and some forms of spreading activation models.