Using Density Matrices in a Compositional Distributional Model of Meaning
Esma Balkr · 2014
In this dissertation I present a framework for using density matrices instead of vectors in a distributional model of meaning. I present an asymmetric similarity measure between two density matrix representations based on relative entropy and show that this measure can be used for hyponymy-hypernymy relations. It is possible to compose density matrix representations of words to get a density matrix representation of a sentence. This map respects the generality of individual words in a sentence, taking sentences with more general words to more general sentence representations.