A computational model of quantification in natural language

Abstract

Natural languages have various ways of expressing quantification, such as the English words "some" and "all." Different such words exist in different languages, and the same word can communicate quite different quantities depending on the context. This thesis presents a computational framework for modeling quantificational meanings and their use in communication. The model can represent meanings that depend on absolute amounts (e.g., two) as well as relative amounts (e.g., half of the total) and context-dependent amounts. It can also represent meanings with presuppositions. Communication between a speaker and a listener is modeled as single exchanges in which both participants have noisy perception of the actual state of the world, the speaker tries to communicate some quantity to the listener by using some word chosen to be informative, and the listener tries to infer the quantity using the word and the assumption that the speaker was being informative. The usage patterns predicted by the model are qualitatively similar to how the words are actually used. The model also shows that the sets of words in real languages result in more efficient communication than randomly selected sets of words with comparable meanings.