Novel formulation of relative positional encoding which adds an element-wise multiplication of the relative positional values with a projection of the embedding of the corresponding entity. Seems to match or exceed performance of per-entity-type position values, while being much more general since it allows relative positional values to be a function of any (learned) feature of an entity rather than just it's type.
Novel formulation of relative positional encoding which adds an element-wise multiplication of the relative positional values with a projection of the embedding of the corresponding entity. Seems to match or exceed performance of per-entity-type position values, while being much more general since it allows relative positional values to be a function of any (learned) feature of an entity rather than just it's type.
Resolves https://github.com/entity-neural-network/incubator/issues/141