A Multi-Agent, Connectionist Model of Metonymy Production

Metonymy, like metaphor, has received much attention in cognitive linguistics literature (Croft, 1993; Kövecses & Radden, 1998; Panther & Radden, 1999). Most experimental work focuses on comprehension. However, why a speaker would choose to produce a metonym in some cases and not others is not fully understood. Connectionist models are well-suited to deal with the partialsemantic/partial-syntactic information which is characteristic of metonymy. Moreover, these models can capture some of the complex, time-varying dynamics of language development. Here, a model is presented in which dyads of artificial agents (each a recurrent neural network based on Rogers, et al., 2004) were trained in a structured environment to develop a language and use it to coordinate in a simple task. The model was informed by an experiment with human subjects based on existing work (Clark & Wilkes-Gibbs, 1986; Selten & Warglien, 2007). While inconclusive, the study offers some insight into the strengths and weaknesses of this particular approach to modeling metonymy production.