Paper Overview
Field: NLP Author: Xiao Zhang Posted: 2026-10-09 arXiv: 2610.13896
Summary
Neural 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. The paper investigates structural automata at three stages of DRS parsing (decoding, post hoc processing, and training) and additionally evaluates decoding on AMR.
Decoding: Accepting-Path Decoding (APD)
APD maintains a verified completion within the remaining length limit and follows the model's preferred token whenever that token can also be completed. APD guarantees an accepted output without retracting tokens.
Under tight length limits:
- DRS: improves Smatch++ by 5.1 and 2.4 percentage points over local masking
- AMR: improves Smatch++ by 6.4 and 3.2 percentage points
- Comparable to rollback methods that require retracting tokens under the same constraints
Post-processing
Invalid outputs are concentrated on the same inputs: with limited training data, eight unconstrained samples fail to produce valid outputs for 10.2% of inputs, versus only 0.4% with full training data.
Training
Penalizing automaton-rejected tokens improves development-set Smatch++ by 3.6-11.0 percentage points at 10% of training. However, this advantage fades as cross-entropy loss alone reduces the probability of rejected tokens from 14% to under 2% within the first sixth of training (with limited data).
Original Abstract (excerpt)
> Neural 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. We investigate structural automata at three stages of DRS parsing (decoding, post hoc processing, and training) and additionally evaluate decoding on AMR. For decoding, we introduce accepting-path decoding (APD), which maintains a verified completion within the remaining length limit and follows the model's preferred token whenever that token can also be completed. APD guarantees an accepted output without retracting tokens. Under tight length limits, it improves Smatch++ ...
---
*Collected automatically on 2026-10-14*