Structural Automata for Neural Semantic Parsing
What it is
This paper proposes using structural automata to enforce well-formedness in neural semantic parsers for DRS and AMR. It introduces Accepting-Path Decoding (APD), which guarantees valid outputs by tracking completion states during generation, and evaluates automaton-based filtering and training penalties.
The problem it solves
Neural parsers often generate locally valid tokens that fail to form a complete, well-formed structure (due to scope or reference errors) within a fixed length limit, leading to invalid or empty outputs.
What you could use it for
- Improving the validity of semantic parsing models for Discourse Representation Structures (DRS)
- Generating well-formed Abstract Meaning Representation (AMR) graphs under length constraints
- Integrating structural constraints into the decoding or training loops of sequence-to-structure models
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.