Nearly Zero-Shot Learning for Semantic Decoding in Spoken Dialogue Systems