Sequence-to-Lattice Models for Fast Translation