Bayesian Inference Semantics for Natural Language

Abstract

In this chapter we present Bayesian Inference Semantics (BIS). This system assigns probability conditions to inferences, and it defines functions for the typed constituents of sentences that generate these conditions compositionally. This framework permits us to capture vagueness through probability distributions for predicates, and the sentential assertions that are constructed from them. Vagueness is, then, a core property of expressions in our account. This allows us to provide natural representations of scalar adjectives and vague classifier terms, while these are problematic for classical semantic theories. Using probability distributions over the definitions of predicates also permits us to handle the sorites paradox in a straightforward way. We sustain the fuzzy boundaries of classifiers through these distributions, without invoking sharp borders between objects to which classifier terms apply, and those to which they do not.