Towards Learning Lattice Orderings for Distributional Semantics
Daoud Clarke ( University of Sussex )
Methods of representing meanings of words based on the contexts in which they occur have been very successful in many natural language processing applications. This "distributional" approach to semantics typically represents word meanings using vectors. Recently, researchers have begun investigating how we can compose these vectors to represent meanings of phrases and sentences. In this talk I will propose a new approach to tackling this problem based on learning a lattice order on the vector space. I will introduce the mathematics behind the resulting structures, vector lattices or Riesz spaces, and discuss some potential applications to natural language processing. I will also discuss the problem of how to learn a vector lattice order from data.