Sequence-to-Sequence Learning via Shared Latent Representation