Combining Logical and Distributional Models of Meaning

Words are building blocks of sentences, yet the meaning of a sentence goes well beyond meanings of its words. Formalizing the process of meaning assignment has been a foundational quest for computational and mathematical linguistics; the two most successful approaches there each miss on a key aspect: the 'logical' one misses on the meanings of words, the vector space one on the grammar. I will present a setting where we can have both! This is based on recent advances in grammar by Lambek pregroups and in vector spaces of physics by Abramsky and Coecke, both using compact categories! I will work through a concrete
application built from a toy document corpus, where for the first time in the field we are able to compute and compare meanings of  sentences compositionally. The fun part is the use of diagrammatic calculus, which visualizes and simplifies the computations to a great extent. This is collaborative work with  E. Greffenstete, C. Clark, B. Coecke, S. Pulman.