Abstract

Categorical compositional distributional model (Clark, Coecke, MS) provides a theoretical setting where one can combine the grammatical structures of phrases and sentences with the vector semantics of the words within them and build a vector semantics for the phrase and sentence. The theory has been applied to a number of linguistic tasks under a number of concrete instantiations and assumptions. The experimental results have been competitive in some cases and better in others. One potential of this model is that it provides a tool for building vectors and linear maps based on the roles of the words in propagating information around in the sentence.

In this talk, I aim to review this model in layman terms. I will show how one can build vector space semantics for simper functional words such as verbs and adjectives, this is previous joint work with E. Grefenstette. I will then review more recent work where one obtains vector space semantics for logical words such as relative pronouns, this is joint work with Kartsaklis, Clark, and Coecke. Finally, I will present ongoing work on how to extend this setting to quantifiers, or at least one kind of it. This is joint work with O. Rypacek.