Cultural evolution of scalar categorization: how cognition and communication affect the structure of categories on scalar conceptual domains

Abstract

Concepts can be thought of as regions of geometrically structured conceptual domains. Of all such possible regions, only very few are lexicalized, i.e. expressed by natural language in a morphologically simple fashion. In the thesis, I discuss lexicalized concepts on conceptual domains that are scalar, more specifically the concepts expressed by gradable adjectives and quantifiers. I consider two generalizations about such concepts. The first generalization is that lexicalized scalar concepts are monotonic, i.e. they can be defined in terms of a single threshold on the scale. The second is that if the conceptual domain has a maximum or a minimum, the threshold is often positioned at one of the extrema. I show that these two properties are non-trivial, in the sense that some scalar concepts, while semantically coherent and cognitively plausible, fail to have these properties. The main of this thesis is to develop an account of how these two properties of monotonicity and extremeness evolve. I focus first on monotonicity, and show with a computational model that its emergence can be explained as an adaptation of language to two pressures, namely a pressure favouring languages that are easy to learn and a pressure on languages to be useful in communication. This explanation of monotonicity relies on the assumption that language users are pragmatically skilful. Moreover, the model makes assumptions about the cognitive biases of the language users. These assumptions are tested in a series of six category learning experiments. The results of three of these experiments are analysed with a Bayesian cognitive model. Overall, the experimental results are inconclusive. I present an agent-based model where learners are neural networks, which provides evidence that monotonic categories are easier to learn than non-monotonic categories. Finally, I turn to the evolution of extremeness. Previous literature has focussed on the role that communicative accuracy plays in the evolution of extremeness. In contrast to previous approaches, I study the role of learning. I show with an evolutionary computational model that extreme categories evolve more often than chance even under a pressure from learning alone, as long as the language teachers and learners are pragmatically skilful.