Structural Automata for Neural Semantic Parsing is on arXiv

Accepting-path decoding keeps a verified completion within the length limit, so a neural semantic parser always returns a well-formed meaning representation without retracting tokens.