Current work in lexical distributed representations maps each word to a point
vector in low-dimensional space. Mapping instead to a density provides many
interesting advantages, including better capturing uncertainty about a
representation and its relationships, expressing asymmetries