Token-level Sequence Labeling for Spoken Language Understanding using Compositional End-to-End Models