Structural Automata for Neural Semantic Parsing
Accepting-path decoding keeps semantic parser outputs valid under tight length limits without backing up.
The paper introduces APD, a decoding method that only follows a model’s preferred next token when the remaining budget still contains a verified path to an accepted output. On DRS and AMR tests with tight length limits, it beats local masking by up to 6.4 Smatch++ points and performs comparably to rollback methods that retract tokens. The author also finds that invalid outputs cluster on the same inputs, and that automaton-based training penalties help early but fade as cross-entropy training already drives rejected-token probability down.
ArXiv · AI/CL/LG's note
The paper introduces APD, a decoding method that only follows a model’s preferred next token when the remaining budget still contains a verified path to an accepted output. On DRS and AMR tests with tight length limits, it beats local masking by up to 6.4 Smatch++ points and performs comparably to rollback methods that retract tokens. The author also finds that invalid outputs cluster on the same inputs, and that automaton-based training penalties help early but fade as cross-entropy training already drives rejected-token probability down.
ArXiv · AI/CL/LG's note
score 4