Structural Automata for Neural Semantic Parsing
arXiv cs.CL · 10 Oct 2026, 03:42 AEST
Structural automata enhance neural semantic parsing by ensuring well-formedness during token generation. The introduction of accepting-path decoding (APD) significantly improves output validity under length constraints.
Key facts
- Neural semantic parsers generate meaning representations token by token, yet well-formedness depends on references, scope, and graph relations that unfold during generation.
- 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.
- 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 points on AMR, comparable to rollback under the same limit, which retracts tokens.
- For training, penalizing automaton-rejected tokens improves development Smatch++ by 3.6-11.0 points after 10% of training, but this advantage diminishes as cross-entropy alone reduces the probability assigned to rejected tokens from 14% to below 2% within the first sixth of training with limited data.
Read aloud
Structural automata enhance neural semantic parsing by ensuring well-formedness during token generation. The introduction of accepting-path decoding APD significantly improves output validity under length constraints. Neural semantic parsers generate meaning representations token by token, yet well-formedness depends on references, scope, and graph relations that unfold during generation. 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. Under tight length limits, it improves Smatch++ over local masking by five point one and two point four points on DRS and by six point four and three point two points on AMR, comparable to rollback under the same limit, which retracts tokens. For training, penalizing automaton-rejected tokens improves development Smatch++ by three point six-eleven point zero points after ten percent of training, but this advantage diminishes as cross-entropy alone reduces the probability assigned to rejected tokens from fourteen percent to below two percent within the first sixth of training with limited data.
Sources in this cluster
- Structural Automata for Neural Semantic Parsing
arXiv cs.CL · 10 Oct 2026, 03:42 AEST