Logic and probability in grounded semantics

Natural language meanings allow speakers to encode important real-world distinctions, but corpora of grounded language use also reveal that speakers categorize the world in different ways and describe situations with different terminology. We can't account for this complexity by deriving one definitive mapping between words and the world.

 In this talk, I explore techniques that capture the vagueness and flexibility of grounded meaning with semantic representations that treat meaning as uncertain, with case studies from descriptions of color in English. The key idea is to represent a color description with a distribution over color categories, which weights possible meanings by the relative likelihood of a speaker using this meaning on any particular occasion. This representation allows us to learn accurate corpus-based models of the descriptions speakers choose and the information speakers provide in simple linguistic tasks. But it also allows us to explain linguistic and philosophical intuitions, to formalize the logical relationships among multiple uses of descriptions in discourse, and to predict interlocutors’ behavior in more complex communicative tasks. Joint work with Brian McMahan and Timothy Meo.