On learning distributed representations of semantics