Toward the Emergence of Quantifiers

Abstract

This thesis explores factors influencing the emergence of quantifiers in a signaling game. It includes a few novel contributions. First is a new signaling game, called the Quantifier Game, designed to provide a setting where quantifier emergence is one natural possibility. It also introduces a new metric designed to measure how well a certain attribute has emerged in a given language, which is applied to measure the emergence of quantifiers. A third contribution is the introduction of a novel loss designed to encourage shorter messages in emergent communication systems. I explore the effects of message compression, curriculum learning, and iterated learning on quantifier emergence. I find that quantifiers do not convincingly emerge in any experiments, but that select implementations of iterated learning do have a small but significant positive effect on quantifier score. A secondary result is that I find no positive effect of iterated learning on topographic similarity, an effect that seemingly contradicts many recent results showing such a positive effect. I suggest as an explanation that iterated learning's effect on languages could actually be to minimize the number of tokens per relevant feature, something that also results in increased compositionality in some but not all cases.