Abstract

My research has a foot in two worlds: the logical world of formal semantics, and the data-driven world of distributional semantics. The aim of formal semantics is to develop mathematical models of meaning, with a particular focus on semantic composition and logical inference. The aim of distributional semantics is to develop computational models of meaning, using algorithms that can be run on a corpus of text. In this talk, I will present Functional Distributional Semantics, a framework which embeds a formal semantic model inside a probabilistic computational model, with benefits in both directions. From the computational perspective, I will show how formal semantic structure enables the model to learn information not captured by mainstream distributional semantic models (vector space models). From the formal perspective, I will show how a probabilistic approach can shed light on phenomena such as vagueness, generic sentences, and donkey anaphora.