An Effective Non-Autoregressive Model for Spoken Language Understanding