Natural languages exhibit properties that are difficult to explain from a purely functional perspective. One of these properties is the systematic lack of upper-bounds in the literal meaning of scalar expressions. This investigation addresses the development and selection of such semantics from a space of possible alternatives. To do so we put forward a model that integrates Bayesian learning into the replicator-mutator dynamics commonly used in evolutionary game theory. We argue this synthesis to provide a suitable and general model to analyze the dynamics involved in the use and transmission of language. Our results shed light on the semantics-pragmatics divide and show how a learning bias in tandem with functional pressure may prevent the lexicalization of pragmatic inferences.