Simple Recurrent Networks and Competition Effects in Spoken Word Recognition

Continuous mapping models of spoken word recognition such as TRACE (McClelland and Elman, 1986) make robust predictions about a wide variety of phenomena. However, most of these models are interactive activation models with preset weights, and do not provide an account of learning. Simple recurrent networks (SRNs, e.g., Elman, 1990) are continuous mapping models that can process sequential patterns and learn representations, and thus may provide an alternative to TRACE. However, it has been suggested that the features that allow SRNs to learn temporal dependencies lead them to work much like the Cohort model (e.g., Marslen-Wilson, 1987), such that items are activated by onset similarity to an input, but not by offset similarity (Norris, 1990). This would make them incompatible with TRACE and with recent results indicating that words that rhyme compete during spoken word recognition (Allopenna, Magnuson and Tanenhaus, 1998). We present simulations demonstrating that rhyme effects do emerge in SRNs, but this depends on how the training is carried out. We also find that SRN predictions provide a good fit to a series of recent studies of the time course of competition effects in spoken word recognition, including cohort, rhyme, and neighborhood density effects.