Learning to Reason Pragmatically with Cognitive Limitations

Adam Vogel, Andrés Goméz Emilsson, Michael C. Frank, Dan Jurafsky, Christopher Potts

Abstract: Recursive Bayesian models of linguistic communication capture a variety of intricate kinds of pragmatic enrichment, but they tend to depend on the unrealistic assumption that agents are invariably optimal reasoners. We present a discriminative model that seeks to capitalize on the insights of such approaches while addressing these concerns about inferential power. The model relies on only approximate representations of language and context, and its recursive properties are limited to the training phase. The resulting behavior is often not optimal, but we present experimental evidence that this suboptimal behavior is closely aligned with human performance on both simple and complex reference games.