Structural Automata for Neural Semantic Parsing
Accepting-path decoding guarantees valid semantic parses and lifts Smatch++ under tight output-length limits.
The paper studies structural automata for neural semantic parsing at decoding, post-processing, and training, with additional AMR decoding tests. Accepting-path decoding maintains a verified completion within the remaining length and follows the model's preferred token when that token can still be completed, without retracting tokens. Under tight length limits it improves Smatch++ over local masking by 5.1 and 2.4 points on DRS and by 6.4 and 3.2 on AMR. Penalizing rejected tokens helps early training, but the gain shrinks once cross-entropy itself suppresses invalid tokens.
- Accepting-path decoding keeps a valid completion inside the length limit.
- Under tight limits, Smatch++ rises 5.1 and 2.4 on DRS and 6.4 and 3.2 on AMR.
- Invalid outputs cluster: 10.2% of limited-data inputs never yield a valid sample.
- Automaton penalties help early training but fade as cross-entropy learns constraints.
From the original article
1 min read at arxiv.orgNeural semantic parsers generate meaning representations token by token, yet well-formedness depends on references, scope, and graph relations that unfold during generation. Constrained decoding prevents locally invalid tokens but does not guarantee that a valid prefix can be completed within a fixed output-length limit.
By Xiao Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.