Pragmatic Reasoning and the Evolution of Adjectival Monotonicity: an Experimental Approach

Abstract

How do humans learn to categorise concepts so quickly? Finding out which factors influence human learning can help bring AI and human learning closer together. Research has suggested the biases for simplicity and informativeness play an important role in how categories are learned. This in turn influences the evolution of these categories over generations, which leads to more efficient and learnable categories. However, previous studies have also shown the importance of pragmatic reasoning, in particular leading to monotonic categorisations of scalar adjectives. Monotonic scalar adjectives are widespread in natural language. Therefore, pragmatic reasoning should be implemented in AI, to make AI reasoning more like human reasoning and aid human-AI communication. In this thesis, I aim to test the importance of pragmatic reasoning in human category learning. In a (human) behavioural Iterated Learning experiment, I explore which factors and biases are most important in learning categories, and how they influence the evolution of categorisations. The results show a tendency towards monotonic categorisations of scalar adjectives, but a bias for informativeness was not properly induced in the experiment and participants behaved in unexpected ways. The results are statistically inconclusive, and cannot be used to back up my statement that pragmatic reasoning should be implemented in AI. However, there are some interesting findings which ask to be explored in more detail.