Constructive processes in immediate serial recall: A recurrent network model of the bigram frequency effect

Bovinick and Plaut Constructive processes in serial recall 1 Abstract Short-term memory for serial order, like other domains of memory, is subject to constructive effects. In particular, background knowledge concerning regularities in sequential structure can affect serial recall. A clear example of this is the bigram frequency effect, whereby letter strings conforming to the transitional probabilities of English are better recalled than random strings. Such effects present difficulty for most current models of immediate serial recall, which rely on associations between individual items and context representations. At the same time, approaches that have proven effective in modeling sequence processing in other contexts, such as recurrent neural networks, have been viewed as inapplicable to short-term serial memory, since they are thought to depend fundamentally on inter-item associations or chaining. We present a recurrent neural network model of immediate serial recall that overcomes this apparent dilemma. The model is used to simulate data from Baddeley (1968) that are widely agreed to rule out chaining models. The model is then applied to the bigram frequency effect, as reported in Baddeley (1964) and Kantowitz, Ornstein and Schwartz (1972). The simulation results dispel the notion that recurrent neural networks are simply chaining models, and provide what appears to be the only available computational account of the bigram frequency effect.