Processing Ambiguous Words: Are Blends Necessary for Lexical Decision?

A previous computational model (Joordens & Besner, 1994) has suggested that during lexical access, ambiguous words tend toward a blend state; that is, network activations settle into an incorrect state that is a mixture of the multiple representations of the ambiguous item. It has been suggested that this blend state actually aids lexical decision (LD) for ambiguous items as the blend state creates a larger “feeling of familiarity” which lexical decision may exploit. This theory, however, is based on the results of a computational model (a simple Hopfield network) in which multiple representations cannot be learned. Here we use a Symmetric Diffusion Network (SDN) to effectively learn and retrieve multiple mappings for a single input (i.e., ambiguous items). The model consists of three main processing regions– orthographics, phonology, and semantics–and is trained on a corpus of unambiguous items and ambiguous items that range in their degree of balance (probability distribution) between the multiple meanings. Following training, the SDN is able to reproduce the correct probability distributions for the ambiguous items; that is, it does not produce blend states. Furthermore, the model qualitatively captures the processing advantage for ambiguous items. Consequently, the notion of a blend state being used for LD is re-evaluated, and further assumptions about semantic processing are explored.