Semantic Universals in Bayesian Learning of Quantifiers

Abstract

Languages undoubtedly exhibit many surface differences; However, past works such as Goddard and Wierzbicka [1994] and von Fintel and Matthewson [2008] have identified semantic properties that are evident in a vast number of languages, i.e. semantic universals. This thesis concerns itself with universal properties of quantifiers (words such as “some”, “few”, etc.). Although there are many possible explanations for universal properties of quantifiers,I work off of the claim that quantifier universals are explained by ease of learning, i.e. that universal properties are universal in natural language quantifiers precisely because they result in quantifiers which are more easily learnable. In this thesis, I investigate the claim that representation length in a language of thought Fodor [1975], together with a degree of universality of a quantifier’s meaning, can serve as a predictor for ease of learning and thus provide more explanation for quantifier universals. Through use of an artificial quantifier gleeb whose meaning is expressed in a language of thought and varied over separate experiments, I study the ease with which both human participants and a Bayesian model learn its meaning through observation of usage contexts. In the end, this method of Bayesian learning far outperforms humans at the same tasks. Results nevertheless exhibit a human and model preference for non-order-based quantifiers. In addition, although models tend toward shorter meaning representations, a degree of universalityseems not to significantly affect model learning.